fs/ocfs2/aops.c

   1 /* -*- mode: c; c-basic-offset: 8; -*-
   2  * vim: noexpandtab sw=8 ts=8 sts=0:
   3  *
   4  * Copyright (C) 2002, 2004 Oracle.  All rights reserved.
   5  *
   6  * This program is free software; you can redistribute it and/or
   7  * modify it under the terms of the GNU General Public
   8  * License as published by the Free Software Foundation; either
   9  * version 2 of the License, or (at your option) any later version.
  10  *
  11  * This program is distributed in the hope that it will be useful,
  12  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  13  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  14  * General Public License for more details.
  15  *
  16  * You should have received a copy of the GNU General Public
  17  * License along with this program; if not, write to the
  18  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  19  * Boston, MA 021110-1307, USA.
  20  */
  21
  22 #include <linux/fs.h>
  23 #include <linux/slab.h>
  24 #include <linux/highmem.h>
  25 #include <linux/pagemap.h>
  26 #include <asm/byteorder.h>
  27 #include <linux/swap.h>
  28 #include <linux/pipe_fs_i.h>
  29 #include <linux/mpage.h>
  30 #include <linux/quotaops.h>
  31 #include <linux/blkdev.h>
  32 #include <linux/uio.h>
  33
  34 #include <cluster/masklog.h>
  35
  36 #include "ocfs2.h"
  37
  38 #include "alloc.h"
  39 #include "aops.h"
  40 #include "dlmglue.h"
  41 #include "extent_map.h"
  42 #include "file.h"
  43 #include "inode.h"
  44 #include "journal.h"
  45 #include "suballoc.h"
  46 #include "super.h"
  47 #include "symlink.h"
  48 #include "refcounttree.h"
  49 #include "ocfs2_trace.h"
  50
  51 #include "buffer_head_io.h"
  52 #include "dir.h"
  53 #include "namei.h"
  54 #include "sysfile.h"
  55
  56 static int ocfs2_symlink_get_block(struct inode *inode, sector_t iblock,
  57                                    struct buffer_head *bh_result, int create)
  58 {
  59         int err = -EIO;
  60         int status;
  61         struct ocfs2_dinode *fe = NULL;
  62         struct buffer_head *bh = NULL;
  63         struct buffer_head *buffer_cache_bh = NULL;
  64         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
  65         void *kaddr;
  66
  67         trace_ocfs2_symlink_get_block(
  68                         (unsigned long long)OCFS2_I(inode)->ip_blkno,
  69                         (unsigned long long)iblock, bh_result, create);
  70
  71         BUG_ON(ocfs2_inode_is_fast_symlink(inode));
  72
  73         if ((iblock << inode->i_sb->s_blocksize_bits) > PATH_MAX + 1) {
  74                 mlog(ML_ERROR, "block offset > PATH_MAX: %llu",
  75                      (unsigned long long)iblock);
  76                 goto bail;
  77         }
  78
  79         status = ocfs2_read_inode_block(inode, &bh);
  80         if (status < 0) {
  81                 mlog_errno(status);
  82                 goto bail;
  83         }
  84         fe = (struct ocfs2_dinode *) bh->b_data;
  85
  86         if ((u64)iblock >= ocfs2_clusters_to_blocks(inode->i_sb,
  87                                                     le32_to_cpu(fe->i_clusters))) {
  88                 err = -ENOMEM;
  89                 mlog(ML_ERROR, "block offset is outside the allocated size: "
  90                      "%llu\n", (unsigned long long)iblock);
  91                 goto bail;
  92         }
  93
  94         /* We don't use the page cache to create symlink data, so if
  95          * need be, copy it over from the buffer cache. */
  96         if (!buffer_uptodate(bh_result) && ocfs2_inode_is_new(inode)) {
  97                 u64 blkno = le64_to_cpu(fe->id2.i_list.l_recs[0].e_blkno) +
  98                             iblock;
  99                 buffer_cache_bh = sb_getblk(osb->sb, blkno);
 100                 if (!buffer_cache_bh) {
 101                         err = -ENOMEM;
 102                         mlog(ML_ERROR, "couldn't getblock for symlink!\n");
 103                         goto bail;
 104                 }
 105
 106                 /* we haven't locked out transactions, so a commit
 107                  * could've happened. Since we've got a reference on
 108                  * the bh, even if it commits while we're doing the
 109                  * copy, the data is still good. */
 110                 if (buffer_jbd(buffer_cache_bh)
 111                     && ocfs2_inode_is_new(inode)) {
 112                         kaddr = kmap_atomic(bh_result->b_page);
 113                         if (!kaddr) {
 114                                 mlog(ML_ERROR, "couldn't kmap!\n");
 115                                 goto bail;
 116                         }
 117                         memcpy(kaddr + (bh_result->b_size * iblock),
 118                                buffer_cache_bh->b_data,
 119                                bh_result->b_size);
 120                         kunmap_atomic(kaddr);
 121                         set_buffer_uptodate(bh_result);
 122                 }
 123                 brelse(buffer_cache_bh);
 124         }
 125
 126         map_bh(bh_result, inode->i_sb,
 127                le64_to_cpu(fe->id2.i_list.l_recs[0].e_blkno) + iblock);
 128
 129         err = 0;
 130
 131 bail:
 132         brelse(bh);
 133
 134         return err;
 135 }
 136
 137 int ocfs2_get_block(struct inode *inode, sector_t iblock,
 138                     struct buffer_head *bh_result, int create)
 139 {
 140         int err = 0;
 141         unsigned int ext_flags;
 142         u64 max_blocks = bh_result->b_size >> inode->i_blkbits;
 143         u64 p_blkno, count, past_eof;
 144         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 145
 146         trace_ocfs2_get_block((unsigned long long)OCFS2_I(inode)->ip_blkno,
 147                               (unsigned long long)iblock, bh_result, create);
 148
 149         if (OCFS2_I(inode)->ip_flags & OCFS2_INODE_SYSTEM_FILE)
 150                 mlog(ML_NOTICE, "get_block on system inode 0x%p (%lu)\n",
 151                      inode, inode->i_ino);
 152
 153         if (S_ISLNK(inode->i_mode)) {
 154                 /* this always does I/O for some reason. */
 155                 err = ocfs2_symlink_get_block(inode, iblock, bh_result, create);
 156                 goto bail;
 157         }
 158
 159         err = ocfs2_extent_map_get_blocks(inode, iblock, &p_blkno, &count,
 160                                           &ext_flags);
 161         if (err) {
 162                 mlog(ML_ERROR, "Error %d from get_blocks(0x%p, %llu, 1, "
 163                      "%llu, NULL)\n", err, inode, (unsigned long long)iblock,
 164                      (unsigned long long)p_blkno);
 165                 goto bail;
 166         }
 167
 168         if (max_blocks < count)
 169                 count = max_blocks;
 170
 171         /*
 172          * ocfs2 never allocates in this function - the only time we
 173          * need to use BH_New is when we're extending i_size on a file
 174          * system which doesn't support holes, in which case BH_New
 175          * allows __block_write_begin() to zero.
 176          *
 177          * If we see this on a sparse file system, then a truncate has
 178          * raced us and removed the cluster. In this case, we clear
 179          * the buffers dirty and uptodate bits and let the buffer code
 180          * ignore it as a hole.
 181          */
 182         if (create && p_blkno == 0 && ocfs2_sparse_alloc(osb)) {
 183                 clear_buffer_dirty(bh_result);
 184                 clear_buffer_uptodate(bh_result);
 185                 goto bail;
 186         }
 187
 188         /* Treat the unwritten extent as a hole for zeroing purposes. */
 189         if (p_blkno && !(ext_flags & OCFS2_EXT_UNWRITTEN))
 190                 map_bh(bh_result, inode->i_sb, p_blkno);
 191
 192         bh_result->b_size = count << inode->i_blkbits;
 193
 194         if (!ocfs2_sparse_alloc(osb)) {
 195                 if (p_blkno == 0) {
 196                         err = -EIO;
 197                         mlog(ML_ERROR,
 198                              "iblock = %llu p_blkno = %llu blkno=(%llu)\n",
 199                              (unsigned long long)iblock,
 200                              (unsigned long long)p_blkno,
 201                              (unsigned long long)OCFS2_I(inode)->ip_blkno);
 202                         mlog(ML_ERROR, "Size %llu, clusters %u\n", (unsigned long long)i_size_read(inode), OCFS2_I(inode)->ip_clusters);
 203                         dump_stack();
 204                         goto bail;
 205                 }
 206         }
 207
 208         past_eof = ocfs2_blocks_for_bytes(inode->i_sb, i_size_read(inode));
 209
 210         trace_ocfs2_get_block_end((unsigned long long)OCFS2_I(inode)->ip_blkno,
 211                                   (unsigned long long)past_eof);
 212         if (create && (iblock >= past_eof))
 213                 set_buffer_new(bh_result);
 214
 215 bail:
 216         if (err < 0)
 217                 err = -EIO;
 218
 219         return err;
 220 }
 221
 222 int ocfs2_read_inline_data(struct inode *inode, struct page *page,
 223                            struct buffer_head *di_bh)
 224 {
 225         void *kaddr;
 226         loff_t size;
 227         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
 228
 229         if (!(le16_to_cpu(di->i_dyn_features) & OCFS2_INLINE_DATA_FL)) {
 230                 ocfs2_error(inode->i_sb, "Inode %llu lost inline data flag\n",
 231                             (unsigned long long)OCFS2_I(inode)->ip_blkno);
 232                 return -EROFS;
 233         }
 234
 235         size = i_size_read(inode);
 236
 237         if (size > PAGE_CACHE_SIZE ||
 238             size > ocfs2_max_inline_data_with_xattr(inode->i_sb, di)) {
 239                 ocfs2_error(inode->i_sb,
 240                             "Inode %llu has with inline data has bad size: %Lu\n",
 241                             (unsigned long long)OCFS2_I(inode)->ip_blkno,
 242                             (unsigned long long)size);
 243                 return -EROFS;
 244         }
 245
 246         kaddr = kmap_atomic(page);
 247         if (size)
 248                 memcpy(kaddr, di->id2.i_data.id_data, size);
 249         /* Clear the remaining part of the page */
 250         memset(kaddr + size, 0, PAGE_CACHE_SIZE - size);
 251         flush_dcache_page(page);
 252         kunmap_atomic(kaddr);
 253
 254         SetPageUptodate(page);
 255
 256         return 0;
 257 }
 258
 259 static int ocfs2_readpage_inline(struct inode *inode, struct page *page)
 260 {
 261         int ret;
 262         struct buffer_head *di_bh = NULL;
 263
 264         BUG_ON(!PageLocked(page));
 265         BUG_ON(!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL));
 266
 267         ret = ocfs2_read_inode_block(inode, &di_bh);
 268         if (ret) {
 269                 mlog_errno(ret);
 270                 goto out;
 271         }
 272
 273         ret = ocfs2_read_inline_data(inode, page, di_bh);
 274 out:
 275         unlock_page(page);
 276
 277         brelse(di_bh);
 278         return ret;
 279 }
 280
 281 static int ocfs2_readpage(struct file *file, struct page *page)
 282 {
 283         struct inode *inode = page->mapping->host;
 284         struct ocfs2_inode_info *oi = OCFS2_I(inode);
 285         loff_t start = (loff_t)page->index << PAGE_CACHE_SHIFT;
 286         int ret, unlock = 1;
 287
 288         trace_ocfs2_readpage((unsigned long long)oi->ip_blkno,
 289                              (page ? page->index : 0));
 290
 291         ret = ocfs2_inode_lock_with_page(inode, NULL, 0, page);
 292         if (ret != 0) {
 293                 if (ret == AOP_TRUNCATED_PAGE)
 294                         unlock = 0;
 295                 mlog_errno(ret);
 296                 goto out;
 297         }
 298
 299         if (down_read_trylock(&oi->ip_alloc_sem) == 0) {
 300                 /*
 301                  * Unlock the page and cycle ip_alloc_sem so that we don't
 302                  * busyloop waiting for ip_alloc_sem to unlock
 303                  */
 304                 ret = AOP_TRUNCATED_PAGE;
 305                 unlock_page(page);
 306                 unlock = 0;
 307                 down_read(&oi->ip_alloc_sem);
 308                 up_read(&oi->ip_alloc_sem);
 309                 goto out_inode_unlock;
 310         }
 311
 312         /*
 313          * i_size might have just been updated as we grabed the meta lock.  We
 314          * might now be discovering a truncate that hit on another node.
 315          * block_read_full_page->get_block freaks out if it is asked to read
 316          * beyond the end of a file, so we check here.  Callers
 317          * (generic_file_read, vm_ops->fault) are clever enough to check i_size
 318          * and notice that the page they just read isn't needed.
 319          *
 320          * XXX sys_readahead() seems to get that wrong?
 321          */
 322         if (start >= i_size_read(inode)) {
 323                 zero_user(page, 0, PAGE_SIZE);
 324                 SetPageUptodate(page);
 325                 ret = 0;
 326                 goto out_alloc;
 327         }
 328
 329         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL)
 330                 ret = ocfs2_readpage_inline(inode, page);
 331         else
 332                 ret = block_read_full_page(page, ocfs2_get_block);
 333         unlock = 0;
 334
 335 out_alloc:
 336         up_read(&OCFS2_I(inode)->ip_alloc_sem);
 337 out_inode_unlock:
 338         ocfs2_inode_unlock(inode, 0);
 339 out:
 340         if (unlock)
 341                 unlock_page(page);
 342         return ret;
 343 }
 344
 345 /*
 346  * This is used only for read-ahead. Failures or difficult to handle
 347  * situations are safe to ignore.
 348  *
 349  * Right now, we don't bother with BH_Boundary - in-inode extent lists
 350  * are quite large (243 extents on 4k blocks), so most inodes don't
 351  * grow out to a tree. If need be, detecting boundary extents could
 352  * trivially be added in a future version of ocfs2_get_block().
 353  */
 354 static int ocfs2_readpages(struct file *filp, struct address_space *mapping,
 355                            struct list_head *pages, unsigned nr_pages)
 356 {
 357         int ret, err = -EIO;
 358         struct inode *inode = mapping->host;
 359         struct ocfs2_inode_info *oi = OCFS2_I(inode);
 360         loff_t start;
 361         struct page *last;
 362
 363         /*
 364          * Use the nonblocking flag for the dlm code to avoid page
 365          * lock inversion, but don't bother with retrying.
 366          */
 367         ret = ocfs2_inode_lock_full(inode, NULL, 0, OCFS2_LOCK_NONBLOCK);
 368         if (ret)
 369                 return err;
 370
 371         if (down_read_trylock(&oi->ip_alloc_sem) == 0) {
 372                 ocfs2_inode_unlock(inode, 0);
 373                 return err;
 374         }
 375
 376         /*
 377          * Don't bother with inline-data. There isn't anything
 378          * to read-ahead in that case anyway...
 379          */
 380         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL)
 381                 goto out_unlock;
 382
 383         /*
 384          * Check whether a remote node truncated this file - we just
 385          * drop out in that case as it's not worth handling here.
 386          */
 387         last = list_entry(pages->prev, struct page, lru);
 388         start = (loff_t)last->index << PAGE_CACHE_SHIFT;
 389         if (start >= i_size_read(inode))
 390                 goto out_unlock;
 391
 392         err = mpage_readpages(mapping, pages, nr_pages, ocfs2_get_block);
 393
 394 out_unlock:
 395         up_read(&oi->ip_alloc_sem);
 396         ocfs2_inode_unlock(inode, 0);
 397
 398         return err;
 399 }
 400
 401 /* Note: Because we don't support holes, our allocation has
 402  * already happened (allocation writes zeros to the file data)
 403  * so we don't have to worry about ordered writes in
 404  * ocfs2_writepage.
 405  *
 406  * ->writepage is called during the process of invalidating the page cache
 407  * during blocked lock processing.  It can't block on any cluster locks
 408  * to during block mapping.  It's relying on the fact that the block
 409  * mapping can't have disappeared under the dirty pages that it is
 410  * being asked to write back.
 411  */
 412 static int ocfs2_writepage(struct page *page, struct writeback_control *wbc)
 413 {
 414         trace_ocfs2_writepage(
 415                 (unsigned long long)OCFS2_I(page->mapping->host)->ip_blkno,
 416                 page->index);
 417
 418         return block_write_full_page(page, ocfs2_get_block, wbc);
 419 }
 420
 421 /* Taken from ext3. We don't necessarily need the full blown
 422  * functionality yet, but IMHO it's better to cut and paste the whole
 423  * thing so we can avoid introducing our own bugs (and easily pick up
 424  * their fixes when they happen) --Mark */
 425 int walk_page_buffers(  handle_t *handle,
 426                         struct buffer_head *head,
 427                         unsigned from,
 428                         unsigned to,
 429                         int *partial,
 430                         int (*fn)(      handle_t *handle,
 431                                         struct buffer_head *bh))
 432 {
 433         struct buffer_head *bh;
 434         unsigned block_start, block_end;
 435         unsigned blocksize = head->b_size;
 436         int err, ret = 0;
 437         struct buffer_head *next;
 438
 439         for (   bh = head, block_start = 0;
 440                 ret == 0 && (bh != head || !block_start);
 441                 block_start = block_end, bh = next)
 442         {
 443                 next = bh->b_this_page;
 444                 block_end = block_start + blocksize;
 445                 if (block_end <= from || block_start >= to) {
 446                         if (partial && !buffer_uptodate(bh))
 447                                 *partial = 1;
 448                         continue;
 449                 }
 450                 err = (*fn)(handle, bh);
 451                 if (!ret)
 452                         ret = err;
 453         }
 454         return ret;
 455 }
 456
 457 static sector_t ocfs2_bmap(struct address_space *mapping, sector_t block)
 458 {
 459         sector_t status;
 460         u64 p_blkno = 0;
 461         int err = 0;
 462         struct inode *inode = mapping->host;
 463
 464         trace_ocfs2_bmap((unsigned long long)OCFS2_I(inode)->ip_blkno,
 465                          (unsigned long long)block);
 466
 467         /* We don't need to lock journal system files, since they aren't
 468          * accessed concurrently from multiple nodes.
 469          */
 470         if (!INODE_JOURNAL(inode)) {
 471                 err = ocfs2_inode_lock(inode, NULL, 0);
 472                 if (err) {
 473                         if (err != -ENOENT)
 474                                 mlog_errno(err);
 475                         goto bail;
 476                 }
 477                 down_read(&OCFS2_I(inode)->ip_alloc_sem);
 478         }
 479
 480         if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL))
 481                 err = ocfs2_extent_map_get_blocks(inode, block, &p_blkno, NULL,
 482                                                   NULL);
 483
 484         if (!INODE_JOURNAL(inode)) {
 485                 up_read(&OCFS2_I(inode)->ip_alloc_sem);
 486                 ocfs2_inode_unlock(inode, 0);
 487         }
 488
 489         if (err) {
 490                 mlog(ML_ERROR, "get_blocks() failed, block = %llu\n",
 491                      (unsigned long long)block);
 492                 mlog_errno(err);
 493                 goto bail;
 494         }
 495
 496 bail:
 497         status = err ? 0 : p_blkno;
 498
 499         return status;
 500 }
 501
 502 /*
 503  * TODO: Make this into a generic get_blocks function.
 504  *
 505  * From do_direct_io in direct-io.c:
 506  *  "So what we do is to permit the ->get_blocks function to populate
 507  *   bh.b_size with the size of IO which is permitted at this offset and
 508  *   this i_blkbits."
 509  *
 510  * This function is called directly from get_more_blocks in direct-io.c.
 511  *
 512  * called like this: dio->get_blocks(dio->inode, fs_startblk,
 513  *                                      fs_count, map_bh, dio->rw == WRITE);
 514  */
 515 static int ocfs2_direct_IO_get_blocks(struct inode *inode, sector_t iblock,
 516                                      struct buffer_head *bh_result, int create)
 517 {
 518         int ret;
 519         u32 cpos = 0;
 520         int alloc_locked = 0;
 521         u64 p_blkno, inode_blocks, contig_blocks;
 522         unsigned int ext_flags;
 523         unsigned char blocksize_bits = inode->i_sb->s_blocksize_bits;
 524         unsigned long max_blocks = bh_result->b_size >> inode->i_blkbits;
 525         unsigned long len = bh_result->b_size;
 526         unsigned int clusters_to_alloc = 0, contig_clusters = 0;
 527
 528         cpos = ocfs2_blocks_to_clusters(inode->i_sb, iblock);
 529
 530         /* This function won't even be called if the request isn't all
 531          * nicely aligned and of the right size, so there's no need
 532          * for us to check any of that. */
 533
 534         inode_blocks = ocfs2_blocks_for_bytes(inode->i_sb, i_size_read(inode));
 535
 536         down_read(&OCFS2_I(inode)->ip_alloc_sem);
 537
 538         /* This figures out the size of the next contiguous block, and
 539          * our logical offset */
 540         ret = ocfs2_extent_map_get_blocks(inode, iblock, &p_blkno,
 541                                           &contig_blocks, &ext_flags);
 542         up_read(&OCFS2_I(inode)->ip_alloc_sem);
 543
 544         if (ret) {
 545                 mlog(ML_ERROR, "get_blocks() failed iblock=%llu\n",
 546                      (unsigned long long)iblock);
 547                 ret = -EIO;
 548                 goto bail;
 549         }
 550
 551         /* We should already CoW the refcounted extent in case of create. */
 552         BUG_ON(create && (ext_flags & OCFS2_EXT_REFCOUNTED));
 553
 554         /* allocate blocks if no p_blkno is found, and create == 1 */
 555         if (!p_blkno && create) {
 556                 ret = ocfs2_inode_lock(inode, NULL, 1);
 557                 if (ret < 0) {
 558                         mlog_errno(ret);
 559                         goto bail;
 560                 }
 561
 562                 alloc_locked = 1;
 563
 564                 down_write(&OCFS2_I(inode)->ip_alloc_sem);
 565
 566                 /* fill hole, allocate blocks can't be larger than the size
 567                  * of the hole */
 568                 clusters_to_alloc = ocfs2_clusters_for_bytes(inode->i_sb, len);
 569                 contig_clusters = ocfs2_clusters_for_blocks(inode->i_sb,
 570                                 contig_blocks);
 571                 if (clusters_to_alloc > contig_clusters)
 572                         clusters_to_alloc = contig_clusters;
 573
 574                 /* allocate extent and insert them into the extent tree */
 575                 ret = ocfs2_extend_allocation(inode, cpos,
 576                                 clusters_to_alloc, 0);
 577                 if (ret < 0) {
 578                         up_write(&OCFS2_I(inode)->ip_alloc_sem);
 579                         mlog_errno(ret);
 580                         goto bail;
 581                 }
 582
 583                 ret = ocfs2_extent_map_get_blocks(inode, iblock, &p_blkno,
 584                                 &contig_blocks, &ext_flags);
 585                 if (ret < 0) {
 586                         up_write(&OCFS2_I(inode)->ip_alloc_sem);
 587                         mlog(ML_ERROR, "get_blocks() failed iblock=%llu\n",
 588                                         (unsigned long long)iblock);
 589                         ret = -EIO;
 590                         goto bail;
 591                 }
 592                 set_buffer_new(bh_result);
 593                 up_write(&OCFS2_I(inode)->ip_alloc_sem);
 594         }
 595
 596         /*
 597          * get_more_blocks() expects us to describe a hole by clearing
 598          * the mapped bit on bh_result().
 599          *
 600          * Consider an unwritten extent as a hole.
 601          */
 602         if (p_blkno && !(ext_flags & OCFS2_EXT_UNWRITTEN))
 603                 map_bh(bh_result, inode->i_sb, p_blkno);
 604         else
 605                 clear_buffer_mapped(bh_result);
 606
 607         /* make sure we don't map more than max_blocks blocks here as
 608            that's all the kernel will handle at this point. */
 609         if (max_blocks < contig_blocks)
 610                 contig_blocks = max_blocks;
 611         bh_result->b_size = contig_blocks << blocksize_bits;
 612 bail:
 613         if (alloc_locked)
 614                 ocfs2_inode_unlock(inode, 1);
 615         return ret;
 616 }
 617
 618 /*
 619  * ocfs2_dio_end_io is called by the dio core when a dio is finished.  We're
 620  * particularly interested in the aio/dio case.  We use the rw_lock DLM lock
 621  * to protect io on one node from truncation on another.
 622  */
 623 static int ocfs2_dio_end_io(struct kiocb *iocb,
 624                              loff_t offset,
 625                              ssize_t bytes,
 626                              void *private)
 627 {
 628         struct inode *inode = file_inode(iocb->ki_filp);
 629         int level;
 630
 631         if (bytes <= 0)
 632                 return 0;
 633
 634         /* this io's submitter should not have unlocked this before we could */
 635         BUG_ON(!ocfs2_iocb_is_rw_locked(iocb));
 636
 637         if (ocfs2_iocb_is_unaligned_aio(iocb)) {
 638                 ocfs2_iocb_clear_unaligned_aio(iocb);
 639
 640                 mutex_unlock(&OCFS2_I(inode)->ip_unaligned_aio);
 641         }
 642
 643         /* Let rw unlock to be done later to protect append direct io write */
 644         if (offset + bytes <= i_size_read(inode)) {
 645                 ocfs2_iocb_clear_rw_locked(iocb);
 646
 647                 level = ocfs2_iocb_rw_locked_level(iocb);
 648                 ocfs2_rw_unlock(inode, level);
 649         }
 650
 651         return 0;
 652 }
 653
 654 static int ocfs2_releasepage(struct page *page, gfp_t wait)
 655 {
 656         if (!page_has_buffers(page))
 657                 return 0;
 658         return try_to_free_buffers(page);
 659 }
 660
 661 static int ocfs2_is_overwrite(struct ocfs2_super *osb,
 662                 struct inode *inode, loff_t offset)
 663 {
 664         int ret = 0;
 665         u32 v_cpos = 0;
 666         u32 p_cpos = 0;
 667         unsigned int num_clusters = 0;
 668         unsigned int ext_flags = 0;
 669
 670         v_cpos = ocfs2_bytes_to_clusters(osb->sb, offset);
 671         ret = ocfs2_get_clusters(inode, v_cpos, &p_cpos,
 672                         &num_clusters, &ext_flags);
 673         if (ret < 0) {
 674                 mlog_errno(ret);
 675                 return ret;
 676         }
 677
 678         if (p_cpos && !(ext_flags & OCFS2_EXT_UNWRITTEN))
 679                 return 1;
 680
 681         return 0;
 682 }
 683
 684 static int ocfs2_direct_IO_zero_extend(struct ocfs2_super *osb,
 685                 struct inode *inode, loff_t offset,
 686                 u64 zero_len, int cluster_align)
 687 {
 688         u32 p_cpos = 0;
 689         u32 v_cpos = ocfs2_bytes_to_clusters(osb->sb, i_size_read(inode));
 690         unsigned int num_clusters = 0;
 691         unsigned int ext_flags = 0;
 692         int ret = 0;
 693
 694         if (offset <= i_size_read(inode) || cluster_align)
 695                 return 0;
 696
 697         ret = ocfs2_get_clusters(inode, v_cpos, &p_cpos, &num_clusters,
 698                         &ext_flags);
 699         if (ret < 0) {
 700                 mlog_errno(ret);
 701                 return ret;
 702         }
 703
 704         if (p_cpos && !(ext_flags & OCFS2_EXT_UNWRITTEN)) {
 705                 u64 s = i_size_read(inode);
 706                 sector_t sector = ((u64)p_cpos << (osb->s_clustersize_bits - 9)) +
 707                         (do_div(s, osb->s_clustersize) >> 9);
 708
 709                 ret = blkdev_issue_zeroout(osb->sb->s_bdev, sector,
 710                                 zero_len >> 9, GFP_NOFS, false);
 711                 if (ret < 0)
 712                         mlog_errno(ret);
 713         }
 714
 715         return ret;
 716 }
 717
 718 static int ocfs2_direct_IO_extend_no_holes(struct ocfs2_super *osb,
 719                 struct inode *inode, loff_t offset)
 720 {
 721         u64 zero_start, zero_len, total_zero_len;
 722         u32 p_cpos = 0, clusters_to_add;
 723         u32 v_cpos = ocfs2_bytes_to_clusters(osb->sb, i_size_read(inode));
 724         unsigned int num_clusters = 0;
 725         unsigned int ext_flags = 0;
 726         u32 size_div, offset_div;
 727         int ret = 0;
 728
 729         {
 730                 u64 o = offset;
 731                 u64 s = i_size_read(inode);
 732
 733                 offset_div = do_div(o, osb->s_clustersize);
 734                 size_div = do_div(s, osb->s_clustersize);
 735         }
 736
 737         if (offset <= i_size_read(inode))
 738                 return 0;
 739
 740         clusters_to_add = ocfs2_bytes_to_clusters(inode->i_sb, offset) -
 741                 ocfs2_bytes_to_clusters(inode->i_sb, i_size_read(inode));
 742         total_zero_len = offset - i_size_read(inode);
 743         if (clusters_to_add)
 744                 total_zero_len -= offset_div;
 745
 746         /* Allocate clusters to fill out holes, and this is only needed
 747          * when we add more than one clusters. Otherwise the cluster will
 748          * be allocated during direct IO */
 749         if (clusters_to_add > 1) {
 750                 ret = ocfs2_extend_allocation(inode,
 751                                 OCFS2_I(inode)->ip_clusters,
 752                                 clusters_to_add - 1, 0);
 753                 if (ret) {
 754                         mlog_errno(ret);
 755                         goto out;
 756                 }
 757         }
 758
 759         while (total_zero_len) {
 760                 ret = ocfs2_get_clusters(inode, v_cpos, &p_cpos, &num_clusters,
 761                                 &ext_flags);
 762                 if (ret < 0) {
 763                         mlog_errno(ret);
 764                         goto out;
 765                 }
 766
 767                 zero_start = ocfs2_clusters_to_bytes(osb->sb, p_cpos) +
 768                         size_div;
 769                 zero_len = ocfs2_clusters_to_bytes(osb->sb, num_clusters) -
 770                         size_div;
 771                 zero_len = min(total_zero_len, zero_len);
 772
 773                 if (p_cpos && !(ext_flags & OCFS2_EXT_UNWRITTEN)) {
 774                         ret = blkdev_issue_zeroout(osb->sb->s_bdev,
 775                                         zero_start >> 9, zero_len >> 9,
 776                                         GFP_NOFS, false);
 777                         if (ret < 0) {
 778                                 mlog_errno(ret);
 779                                 goto out;
 780                         }
 781                 }
 782
 783                 total_zero_len -= zero_len;
 784                 v_cpos += ocfs2_bytes_to_clusters(osb->sb, zero_len + size_div);
 785
 786                 /* Only at first iteration can be cluster not aligned.
 787                  * So set size_div to 0 for the rest */
 788                 size_div = 0;
 789         }
 790
 791 out:
 792         return ret;
 793 }
 794
 795 static ssize_t ocfs2_direct_IO_write(struct kiocb *iocb,
 796                 struct iov_iter *iter,
 797                 loff_t offset)
 798 {
 799         ssize_t ret = 0;
 800         ssize_t written = 0;
 801         bool orphaned = false;
 802         int is_overwrite = 0;
 803         struct file *file = iocb->ki_filp;
 804         struct inode *inode = file_inode(file)->i_mapping->host;
 805         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 806         struct buffer_head *di_bh = NULL;
 807         size_t count = iter->count;
 808         journal_t *journal = osb->journal->j_journal;
 809         u64 zero_len_head, zero_len_tail;
 810         int cluster_align_head, cluster_align_tail;
 811         loff_t final_size = offset + count;
 812         int append_write = offset >= i_size_read(inode) ? 1 : 0;
 813         unsigned int num_clusters = 0;
 814         unsigned int ext_flags = 0;
 815
 816         {
 817                 u64 o = offset;
 818                 u64 s = i_size_read(inode);
 819
 820                 zero_len_head = do_div(o, 1 << osb->s_clustersize_bits);
 821                 cluster_align_head = !zero_len_head;
 822
 823                 zero_len_tail = osb->s_clustersize -
 824                         do_div(s, osb->s_clustersize);
 825                 if ((offset - i_size_read(inode)) < zero_len_tail)
 826                         zero_len_tail = offset - i_size_read(inode);
 827                 cluster_align_tail = !zero_len_tail;
 828         }
 829
 830         /*
 831          * when final_size > inode->i_size, inode->i_size will be
 832          * updated after direct write, so add the inode to orphan
 833          * dir first.
 834          */
 835         if (final_size > i_size_read(inode)) {
 836                 ret = ocfs2_add_inode_to_orphan(osb, inode);
 837                 if (ret < 0) {
 838                         mlog_errno(ret);
 839                         goto out;
 840                 }
 841                 orphaned = true;
 842         }
 843
 844         if (append_write) {
 845                 ret = ocfs2_inode_lock(inode, NULL, 1);
 846                 if (ret < 0) {
 847                         mlog_errno(ret);
 848                         goto clean_orphan;
 849                 }
 850
 851                 /* zeroing out the previously allocated cluster tail
 852                  * that but not zeroed */
 853                 if (ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb))) {
 854                         down_read(&OCFS2_I(inode)->ip_alloc_sem);
 855                         ret = ocfs2_direct_IO_zero_extend(osb, inode, offset,
 856                                         zero_len_tail, cluster_align_tail);
 857                         up_read(&OCFS2_I(inode)->ip_alloc_sem);
 858                 } else {
 859                         down_write(&OCFS2_I(inode)->ip_alloc_sem);
 860                         ret = ocfs2_direct_IO_extend_no_holes(osb, inode,
 861                                         offset);
 862                         up_write(&OCFS2_I(inode)->ip_alloc_sem);
 863                 }
 864                 if (ret < 0) {
 865                         mlog_errno(ret);
 866                         ocfs2_inode_unlock(inode, 1);
 867                         goto clean_orphan;
 868                 }
 869
 870                 is_overwrite = ocfs2_is_overwrite(osb, inode, offset);
 871                 if (is_overwrite < 0) {
 872                         mlog_errno(is_overwrite);
 873                         ret = is_overwrite;
 874                         ocfs2_inode_unlock(inode, 1);
 875                         goto clean_orphan;
 876                 }
 877
 878                 ocfs2_inode_unlock(inode, 1);
 879         }
 880
 881         written = __blockdev_direct_IO(iocb, inode, inode->i_sb->s_bdev, iter,
 882                                        offset, ocfs2_direct_IO_get_blocks,
 883                                        ocfs2_dio_end_io, NULL, 0);
 884         /* overwrite aio may return -EIOCBQUEUED, and it is not an error */
 885         if ((written < 0) && (written != -EIOCBQUEUED)) {
 886                 loff_t i_size = i_size_read(inode);
 887
 888                 if (offset + count > i_size) {
 889                         ret = ocfs2_inode_lock(inode, &di_bh, 1);
 890                         if (ret < 0) {
 891                                 mlog_errno(ret);
 892                                 goto clean_orphan;
 893                         }
 894
 895                         if (i_size == i_size_read(inode)) {
 896                                 ret = ocfs2_truncate_file(inode, di_bh,
 897                                                 i_size);
 898                                 if (ret < 0) {
 899                                         if (ret != -ENOSPC)
 900                                                 mlog_errno(ret);
 901
 902                                         ocfs2_inode_unlock(inode, 1);
 903                                         brelse(di_bh);
 904                                         di_bh = NULL;
 905                                         goto clean_orphan;
 906                                 }
 907                         }
 908
 909                         ocfs2_inode_unlock(inode, 1);
 910                         brelse(di_bh);
 911                         di_bh = NULL;
 912
 913                         ret = jbd2_journal_force_commit(journal);
 914                         if (ret < 0)
 915                                 mlog_errno(ret);
 916                 }
 917         } else if (written > 0 && append_write && !is_overwrite &&
 918                         !cluster_align_head) {
 919                 /* zeroing out the allocated cluster head */
 920                 u32 p_cpos = 0;
 921                 u32 v_cpos = ocfs2_bytes_to_clusters(osb->sb, offset);
 922
 923                 ret = ocfs2_inode_lock(inode, NULL, 0);
 924                 if (ret < 0) {
 925                         mlog_errno(ret);
 926                         goto clean_orphan;
 927                 }
 928
 929                 ret = ocfs2_get_clusters(inode, v_cpos, &p_cpos,
 930                                 &num_clusters, &ext_flags);
 931                 if (ret < 0) {
 932                         mlog_errno(ret);
 933                         ocfs2_inode_unlock(inode, 0);
 934                         goto clean_orphan;
 935                 }
 936
 937                 BUG_ON(!p_cpos || (ext_flags & OCFS2_EXT_UNWRITTEN));
 938
 939                 ret = blkdev_issue_zeroout(osb->sb->s_bdev,
 940                                 (u64)p_cpos << (osb->s_clustersize_bits - 9),
 941                                 zero_len_head >> 9, GFP_NOFS, false);
 942                 if (ret < 0)
 943                         mlog_errno(ret);
 944
 945                 ocfs2_inode_unlock(inode, 0);
 946         }
 947
 948 clean_orphan:
 949         if (orphaned) {
 950                 int tmp_ret;
 951                 int update_isize = written > 0 ? 1 : 0;
 952                 loff_t end = update_isize ? offset + written : 0;
 953
 954                 tmp_ret = ocfs2_inode_lock(inode, &di_bh, 1);
 955                 if (tmp_ret < 0) {
 956                         ret = tmp_ret;
 957                         mlog_errno(ret);
 958                         goto out;
 959                 }
 960
 961                 tmp_ret = ocfs2_del_inode_from_orphan(osb, inode, di_bh,
 962                                 update_isize, end);
 963                 if (tmp_ret < 0) {
 964                         ocfs2_inode_unlock(inode, 1);
 965                         ret = tmp_ret;
 966                         mlog_errno(ret);
 967                         brelse(di_bh);
 968                         goto out;
 969                 }
 970
 971                 ocfs2_inode_unlock(inode, 1);
 972                 brelse(di_bh);
 973
 974                 tmp_ret = jbd2_journal_force_commit(journal);
 975                 if (tmp_ret < 0) {
 976                         ret = tmp_ret;
 977                         mlog_errno(tmp_ret);
 978                 }
 979         }
 980
 981 out:
 982         if (ret >= 0)
 983                 ret = written;
 984         return ret;
 985 }
 986
 987 static ssize_t ocfs2_direct_IO(struct kiocb *iocb, struct iov_iter *iter,
 988                                loff_t offset)
 989 {
 990         struct file *file = iocb->ki_filp;
 991         struct inode *inode = file_inode(file)->i_mapping->host;
 992         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 993         int full_coherency = !(osb->s_mount_opt &
 994                         OCFS2_MOUNT_COHERENCY_BUFFERED);
 995
 996         /*
 997          * Fallback to buffered I/O if we see an inode without
 998          * extents.
 999          */
1000         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
1001                 return 0;
1002
1003         /* Fallback to buffered I/O if we are appending and
1004          * concurrent O_DIRECT writes are allowed.
1005          */
1006         if (i_size_read(inode) <= offset && !full_coherency)
1007                 return 0;
1008
1009         if (iov_iter_rw(iter) == READ)
1010                 return __blockdev_direct_IO(iocb, inode, inode->i_sb->s_bdev,
1011                                             iter, offset,
1012                                             ocfs2_direct_IO_get_blocks,
1013                                             ocfs2_dio_end_io, NULL, 0);
1014         else
1015                 return ocfs2_direct_IO_write(iocb, iter, offset);
1016 }
1017
1018 static void ocfs2_figure_cluster_boundaries(struct ocfs2_super *osb,
1019                                             u32 cpos,
1020                                             unsigned int *start,
1021                                             unsigned int *end)
1022 {
1023         unsigned int cluster_start = 0, cluster_end = PAGE_CACHE_SIZE;
1024
1025         if (unlikely(PAGE_CACHE_SHIFT > osb->s_clustersize_bits)) {
1026                 unsigned int cpp;
1027
1028                 cpp = 1 << (PAGE_CACHE_SHIFT - osb->s_clustersize_bits);
1029
1030                 cluster_start = cpos % cpp;
1031                 cluster_start = cluster_start << osb->s_clustersize_bits;
1032
1033                 cluster_end = cluster_start + osb->s_clustersize;
1034         }
1035
1036         BUG_ON(cluster_start > PAGE_SIZE);
1037         BUG_ON(cluster_end > PAGE_SIZE);
1038
1039         if (start)
1040                 *start = cluster_start;
1041         if (end)
1042                 *end = cluster_end;
1043 }
1044
1045 /*
1046  * 'from' and 'to' are the region in the page to avoid zeroing.
1047  *
1048  * If pagesize > clustersize, this function will avoid zeroing outside
1049  * of the cluster boundary.
1050  *
1051  * from == to == 0 is code for "zero the entire cluster region"
1052  */
1053 static void ocfs2_clear_page_regions(struct page *page,
1054                                      struct ocfs2_super *osb, u32 cpos,
1055                                      unsigned from, unsigned to)
1056 {
1057         void *kaddr;
1058         unsigned int cluster_start, cluster_end;
1059
1060         ocfs2_figure_cluster_boundaries(osb, cpos, &cluster_start, &cluster_end);
1061
1062         kaddr = kmap_atomic(page);
1063
1064         if (from || to) {
1065                 if (from > cluster_start)
1066                         memset(kaddr + cluster_start, 0, from - cluster_start);
1067                 if (to < cluster_end)
1068                         memset(kaddr + to, 0, cluster_end - to);
1069         } else {
1070                 memset(kaddr + cluster_start, 0, cluster_end - cluster_start);
1071         }
1072
1073         kunmap_atomic(kaddr);
1074 }
1075
1076 /*
1077  * Nonsparse file systems fully allocate before we get to the write
1078  * code. This prevents ocfs2_write() from tagging the write as an
1079  * allocating one, which means ocfs2_map_page_blocks() might try to
1080  * read-in the blocks at the tail of our file. Avoid reading them by
1081  * testing i_size against each block offset.
1082  */
1083 static int ocfs2_should_read_blk(struct inode *inode, struct page *page,
1084                                  unsigned int block_start)
1085 {
1086         u64 offset = page_offset(page) + block_start;
1087
1088         if (ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
1089                 return 1;
1090
1091         if (i_size_read(inode) > offset)
1092                 return 1;
1093
1094         return 0;
1095 }
1096
1097 /*
1098  * Some of this taken from __block_write_begin(). We already have our
1099  * mapping by now though, and the entire write will be allocating or
1100  * it won't, so not much need to use BH_New.
1101  *
1102  * This will also skip zeroing, which is handled externally.
1103  */
1104 int ocfs2_map_page_blocks(struct page *page, u64 *p_blkno,
1105                           struct inode *inode, unsigned int from,
1106                           unsigned int to, int new)
1107 {
1108         int ret = 0;
1109         struct buffer_head *head, *bh, *wait[2], **wait_bh = wait;
1110         unsigned int block_end, block_start;
1111         unsigned int bsize = 1 << inode->i_blkbits;
1112
1113         if (!page_has_buffers(page))
1114                 create_empty_buffers(page, bsize, 0);
1115
1116         head = page_buffers(page);
1117         for (bh = head, block_start = 0; bh != head || !block_start;
1118              bh = bh->b_this_page, block_start += bsize) {
1119                 block_end = block_start + bsize;
1120
1121                 clear_buffer_new(bh);
1122
1123                 /*
1124                  * Ignore blocks outside of our i/o range -
1125                  * they may belong to unallocated clusters.
1126                  */
1127                 if (block_start >= to || block_end <= from) {
1128                         if (PageUptodate(page))
1129                                 set_buffer_uptodate(bh);
1130                         continue;
1131                 }
1132
1133                 /*
1134                  * For an allocating write with cluster size >= page
1135                  * size, we always write the entire page.
1136                  */
1137                 if (new)
1138                         set_buffer_new(bh);
1139
1140                 if (!buffer_mapped(bh)) {
1141                         map_bh(bh, inode->i_sb, *p_blkno);
1142                         unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
1143                 }
1144
1145                 if (PageUptodate(page)) {
1146                         if (!buffer_uptodate(bh))
1147                                 set_buffer_uptodate(bh);
1148                 } else if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
1149                            !buffer_new(bh) &&
1150                            ocfs2_should_read_blk(inode, page, block_start) &&
1151                            (block_start < from || block_end > to)) {
1152                         ll_rw_block(READ, 1, &bh);
1153                         *wait_bh++=bh;
1154                 }
1155
1156                 *p_blkno = *p_blkno + 1;
1157         }
1158
1159         /*
1160          * If we issued read requests - let them complete.
1161          */
1162         while(wait_bh > wait) {
1163                 wait_on_buffer(*--wait_bh);
1164                 if (!buffer_uptodate(*wait_bh))
1165                         ret = -EIO;
1166         }
1167
1168         if (ret == 0 || !new)
1169                 return ret;
1170
1171         /*
1172          * If we get -EIO above, zero out any newly allocated blocks
1173          * to avoid exposing stale data.
1174          */
1175         bh = head;
1176         block_start = 0;
1177         do {
1178                 block_end = block_start + bsize;
1179                 if (block_end <= from)
1180                         goto next_bh;
1181                 if (block_start >= to)
1182                         break;
1183
1184                 zero_user(page, block_start, bh->b_size);
1185                 set_buffer_uptodate(bh);
1186                 mark_buffer_dirty(bh);
1187
1188 next_bh:
1189                 block_start = block_end;
1190                 bh = bh->b_this_page;
1191         } while (bh != head);
1192
1193         return ret;
1194 }
1195
1196 #if (PAGE_CACHE_SIZE >= OCFS2_MAX_CLUSTERSIZE)
1197 #define OCFS2_MAX_CTXT_PAGES    1
1198 #else
1199 #define OCFS2_MAX_CTXT_PAGES    (OCFS2_MAX_CLUSTERSIZE / PAGE_CACHE_SIZE)
1200 #endif
1201
1202 #define OCFS2_MAX_CLUSTERS_PER_PAGE     (PAGE_CACHE_SIZE / OCFS2_MIN_CLUSTERSIZE)
1203
1204 /*
1205  * Describe the state of a single cluster to be written to.
1206  */
1207 struct ocfs2_write_cluster_desc {
1208         u32             c_cpos;
1209         u32             c_phys;
1210         /*
1211          * Give this a unique field because c_phys eventually gets
1212          * filled.
1213          */
1214         unsigned        c_new;
1215         unsigned        c_clear_unwritten;
1216         unsigned        c_needs_zero;
1217 };
1218
1219 struct ocfs2_write_ctxt {
1220         /* Logical cluster position / len of write */
1221         u32                             w_cpos;
1222         u32                             w_clen;
1223
1224         /* First cluster allocated in a nonsparse extend */
1225         u32                             w_first_new_cpos;
1226
1227         /* Type of caller. Must be one of buffer, mmap, direct.  */
1228         ocfs2_write_type_t              w_type;
1229
1230         struct ocfs2_write_cluster_desc w_desc[OCFS2_MAX_CLUSTERS_PER_PAGE];
1231
1232         /*
1233          * This is true if page_size > cluster_size.
1234          *
1235          * It triggers a set of special cases during write which might
1236          * have to deal with allocating writes to partial pages.
1237          */
1238         unsigned int                    w_large_pages;
1239
1240         /*
1241          * Pages involved in this write.
1242          *
1243          * w_target_page is the page being written to by the user.
1244          *
1245          * w_pages is an array of pages which always contains
1246          * w_target_page, and in the case of an allocating write with
1247          * page_size < cluster size, it will contain zero'd and mapped
1248          * pages adjacent to w_target_page which need to be written
1249          * out in so that future reads from that region will get
1250          * zero's.
1251          */
1252         unsigned int                    w_num_pages;
1253         struct page                     *w_pages[OCFS2_MAX_CTXT_PAGES];
1254         struct page                     *w_target_page;
1255
1256         /*
1257          * w_target_locked is used for page_mkwrite path indicating no unlocking
1258          * against w_target_page in ocfs2_write_end_nolock.
1259          */
1260         unsigned int                    w_target_locked:1;
1261
1262         /*
1263          * ocfs2_write_end() uses this to know what the real range to
1264          * write in the target should be.
1265          */
1266         unsigned int                    w_target_from;
1267         unsigned int                    w_target_to;
1268
1269         /*
1270          * We could use journal_current_handle() but this is cleaner,
1271          * IMHO -Mark
1272          */
1273         handle_t                        *w_handle;
1274
1275         struct buffer_head              *w_di_bh;
1276
1277         struct ocfs2_cached_dealloc_ctxt w_dealloc;
1278 };
1279
1280 void ocfs2_unlock_and_free_pages(struct page **pages, int num_pages)
1281 {
1282         int i;
1283
1284         for(i = 0; i < num_pages; i++) {
1285                 if (pages[i]) {
1286                         unlock_page(pages[i]);
1287                         mark_page_accessed(pages[i]);
1288                         page_cache_release(pages[i]);
1289                 }
1290         }
1291 }
1292
1293 static void ocfs2_unlock_pages(struct ocfs2_write_ctxt *wc)
1294 {
1295         int i;
1296
1297         /*
1298          * w_target_locked is only set to true in the page_mkwrite() case.
1299          * The intent is to allow us to lock the target page from write_begin()
1300          * to write_end(). The caller must hold a ref on w_target_page.
1301          */
1302         if (wc->w_target_locked) {
1303                 BUG_ON(!wc->w_target_page);
1304                 for (i = 0; i < wc->w_num_pages; i++) {
1305                         if (wc->w_target_page == wc->w_pages[i]) {
1306                                 wc->w_pages[i] = NULL;
1307                                 break;
1308                         }
1309                 }
1310                 mark_page_accessed(wc->w_target_page);
1311                 page_cache_release(wc->w_target_page);
1312         }
1313         ocfs2_unlock_and_free_pages(wc->w_pages, wc->w_num_pages);
1314 }
1315
1316 static void ocfs2_free_write_ctxt(struct ocfs2_write_ctxt *wc)
1317 {
1318         ocfs2_unlock_pages(wc);
1319         brelse(wc->w_di_bh);
1320         kfree(wc);
1321 }
1322
1323 static int ocfs2_alloc_write_ctxt(struct ocfs2_write_ctxt **wcp,
1324                                   struct ocfs2_super *osb, loff_t pos,
1325                                   unsigned len, ocfs2_write_type_t type,
1326                                   struct buffer_head *di_bh)
1327 {
1328         u32 cend;
1329         struct ocfs2_write_ctxt *wc;
1330
1331         wc = kzalloc(sizeof(struct ocfs2_write_ctxt), GFP_NOFS);
1332         if (!wc)
1333                 return -ENOMEM;
1334
1335         wc->w_cpos = pos >> osb->s_clustersize_bits;
1336         wc->w_first_new_cpos = UINT_MAX;
1337         cend = (pos + len - 1) >> osb->s_clustersize_bits;
1338         wc->w_clen = cend - wc->w_cpos + 1;
1339         get_bh(di_bh);
1340         wc->w_di_bh = di_bh;
1341         wc->w_type = type;
1342
1343         if (unlikely(PAGE_CACHE_SHIFT > osb->s_clustersize_bits))
1344                 wc->w_large_pages = 1;
1345         else
1346                 wc->w_large_pages = 0;
1347
1348         ocfs2_init_dealloc_ctxt(&wc->w_dealloc);
1349
1350         *wcp = wc;
1351
1352         return 0;
1353 }
1354
1355 /*
1356  * If a page has any new buffers, zero them out here, and mark them uptodate
1357  * and dirty so they'll be written out (in order to prevent uninitialised
1358  * block data from leaking). And clear the new bit.
1359  */
1360 static void ocfs2_zero_new_buffers(struct page *page, unsigned from, unsigned to)
1361 {
1362         unsigned int block_start, block_end;
1363         struct buffer_head *head, *bh;
1364
1365         BUG_ON(!PageLocked(page));
1366         if (!page_has_buffers(page))
1367                 return;
1368
1369         bh = head = page_buffers(page);
1370         block_start = 0;
1371         do {
1372                 block_end = block_start + bh->b_size;
1373
1374                 if (buffer_new(bh)) {
1375                         if (block_end > from && block_start < to) {
1376                                 if (!PageUptodate(page)) {
1377                                         unsigned start, end;
1378
1379                                         start = max(from, block_start);
1380                                         end = min(to, block_end);
1381
1382                                         zero_user_segment(page, start, end);
1383                                         set_buffer_uptodate(bh);
1384                                 }
1385
1386                                 clear_buffer_new(bh);
1387                                 mark_buffer_dirty(bh);
1388                         }
1389                 }
1390
1391                 block_start = block_end;
1392                 bh = bh->b_this_page;
1393         } while (bh != head);
1394 }
1395
1396 /*
1397  * Only called when we have a failure during allocating write to write
1398  * zero's to the newly allocated region.
1399  */
1400 static void ocfs2_write_failure(struct inode *inode,
1401                                 struct ocfs2_write_ctxt *wc,
1402                                 loff_t user_pos, unsigned user_len)
1403 {
1404         int i;
1405         unsigned from = user_pos & (PAGE_CACHE_SIZE - 1),
1406                 to = user_pos + user_len;
1407         struct page *tmppage;
1408
1409         if (wc->w_target_page)
1410                 ocfs2_zero_new_buffers(wc->w_target_page, from, to);
1411
1412         for(i = 0; i < wc->w_num_pages; i++) {
1413                 tmppage = wc->w_pages[i];
1414
1415                 if (tmppage && page_has_buffers(tmppage)) {
1416                         if (ocfs2_should_order_data(inode))
1417                                 ocfs2_jbd2_file_inode(wc->w_handle, inode);
1418
1419                         block_commit_write(tmppage, from, to);
1420                 }
1421         }
1422 }
1423
1424 static int ocfs2_prepare_page_for_write(struct inode *inode, u64 *p_blkno,
1425                                         struct ocfs2_write_ctxt *wc,
1426                                         struct page *page, u32 cpos,
1427                                         loff_t user_pos, unsigned user_len,
1428                                         int new)
1429 {
1430         int ret;
1431         unsigned int map_from = 0, map_to = 0;
1432         unsigned int cluster_start, cluster_end;
1433         unsigned int user_data_from = 0, user_data_to = 0;
1434
1435         ocfs2_figure_cluster_boundaries(OCFS2_SB(inode->i_sb), cpos,
1436                                         &cluster_start, &cluster_end);
1437
1438         /* treat the write as new if the a hole/lseek spanned across
1439          * the page boundary.
1440          */
1441         new = new | ((i_size_read(inode) <= page_offset(page)) &&
1442                         (page_offset(page) <= user_pos));
1443
1444         if (page == wc->w_target_page) {
1445                 map_from = user_pos & (PAGE_CACHE_SIZE - 1);
1446                 map_to = map_from + user_len;
1447
1448                 if (new)
1449                         ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1450                                                     cluster_start, cluster_end,
1451                                                     new);
1452                 else
1453                         ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1454                                                     map_from, map_to, new);
1455                 if (ret) {
1456                         mlog_errno(ret);
1457                         goto out;
1458                 }
1459
1460                 user_data_from = map_from;
1461                 user_data_to = map_to;
1462                 if (new) {
1463                         map_from = cluster_start;
1464                         map_to = cluster_end;
1465                 }
1466         } else {
1467                 /*
1468                  * If we haven't allocated the new page yet, we
1469                  * shouldn't be writing it out without copying user
1470                  * data. This is likely a math error from the caller.
1471                  */
1472                 BUG_ON(!new);
1473
1474                 map_from = cluster_start;
1475                 map_to = cluster_end;
1476
1477                 ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1478                                             cluster_start, cluster_end, new);
1479                 if (ret) {
1480                         mlog_errno(ret);
1481                         goto out;
1482                 }
1483         }
1484
1485         /*
1486          * Parts of newly allocated pages need to be zero'd.
1487          *
1488          * Above, we have also rewritten 'to' and 'from' - as far as
1489          * the rest of the function is concerned, the entire cluster
1490          * range inside of a page needs to be written.
1491          *
1492          * We can skip this if the page is up to date - it's already
1493          * been zero'd from being read in as a hole.
1494          */
1495         if (new && !PageUptodate(page))
1496                 ocfs2_clear_page_regions(page, OCFS2_SB(inode->i_sb),
1497                                          cpos, user_data_from, user_data_to);
1498
1499         flush_dcache_page(page);
1500
1501 out:
1502         return ret;
1503 }
1504
1505 /*
1506  * This function will only grab one clusters worth of pages.
1507  */
1508 static int ocfs2_grab_pages_for_write(struct address_space *mapping,
1509                                       struct ocfs2_write_ctxt *wc,
1510                                       u32 cpos, loff_t user_pos,
1511                                       unsigned user_len, int new,
1512                                       struct page *mmap_page)
1513 {
1514         int ret = 0, i;
1515         unsigned long start, target_index, end_index, index;
1516         struct inode *inode = mapping->host;
1517         loff_t last_byte;
1518
1519         target_index = user_pos >> PAGE_CACHE_SHIFT;
1520
1521         /*
1522          * Figure out how many pages we'll be manipulating here. For
1523          * non allocating write, we just change the one
1524          * page. Otherwise, we'll need a whole clusters worth.  If we're
1525          * writing past i_size, we only need enough pages to cover the
1526          * last page of the write.
1527          */
1528         if (new) {
1529                 wc->w_num_pages = ocfs2_pages_per_cluster(inode->i_sb);
1530                 start = ocfs2_align_clusters_to_page_index(inode->i_sb, cpos);
1531                 /*
1532                  * We need the index *past* the last page we could possibly
1533                  * touch.  This is the page past the end of the write or
1534                  * i_size, whichever is greater.
1535                  */
1536                 last_byte = max(user_pos + user_len, i_size_read(inode));
1537                 BUG_ON(last_byte < 1);
1538                 end_index = ((last_byte - 1) >> PAGE_CACHE_SHIFT) + 1;
1539                 if ((start + wc->w_num_pages) > end_index)
1540                         wc->w_num_pages = end_index - start;
1541         } else {
1542                 wc->w_num_pages = 1;
1543                 start = target_index;
1544         }
1545         end_index = (user_pos + user_len - 1) >> PAGE_CACHE_SHIFT;
1546
1547         for(i = 0; i < wc->w_num_pages; i++) {
1548                 index = start + i;
1549
1550                 if (index >= target_index && index <= end_index &&
1551                     wc->w_type == OCFS2_WRITE_MMAP) {
1552                         /*
1553                          * ocfs2_pagemkwrite() is a little different
1554                          * and wants us to directly use the page
1555                          * passed in.
1556                          */
1557                         lock_page(mmap_page);
1558
1559                         /* Exit and let the caller retry */
1560                         if (mmap_page->mapping != mapping) {
1561                                 WARN_ON(mmap_page->mapping);
1562                                 unlock_page(mmap_page);
1563                                 ret = -EAGAIN;
1564                                 goto out;
1565                         }
1566
1567                         page_cache_get(mmap_page);
1568                         wc->w_pages[i] = mmap_page;
1569                         wc->w_target_locked = true;
1570                 } else if (index >= target_index && index <= end_index &&
1571                            wc->w_type == OCFS2_WRITE_DIRECT) {
1572                         /* Direct write has no mapping page. */
1573                         wc->w_pages[i] = NULL;
1574                         continue;
1575                 } else {
1576                         wc->w_pages[i] = find_or_create_page(mapping, index,
1577                                                              GFP_NOFS);
1578                         if (!wc->w_pages[i]) {
1579                                 ret = -ENOMEM;
1580                                 mlog_errno(ret);
1581                                 goto out;
1582                         }
1583                 }
1584                 wait_for_stable_page(wc->w_pages[i]);
1585
1586                 if (index == target_index)
1587                         wc->w_target_page = wc->w_pages[i];
1588         }
1589 out:
1590         if (ret)
1591                 wc->w_target_locked = false;
1592         return ret;
1593 }
1594
1595 /*
1596  * Prepare a single cluster for write one cluster into the file.
1597  */
1598 static int ocfs2_write_cluster(struct address_space *mapping,
1599                                u32 *phys, unsigned int new,
1600                                unsigned int clear_unwritten,
1601                                unsigned int should_zero,
1602                                struct ocfs2_alloc_context *data_ac,
1603                                struct ocfs2_alloc_context *meta_ac,
1604                                struct ocfs2_write_ctxt *wc, u32 cpos,
1605                                loff_t user_pos, unsigned user_len)
1606 {
1607         int ret, i;
1608         u64 p_blkno;
1609         struct inode *inode = mapping->host;
1610         struct ocfs2_extent_tree et;
1611         int bpc = ocfs2_clusters_to_blocks(inode->i_sb, 1);
1612
1613         if (new) {
1614                 u32 tmp_pos;
1615
1616                 /*
1617                  * This is safe to call with the page locks - it won't take
1618                  * any additional semaphores or cluster locks.
1619                  */
1620                 tmp_pos = cpos;
1621                 ret = ocfs2_add_inode_data(OCFS2_SB(inode->i_sb), inode,
1622                                            &tmp_pos, 1, !clear_unwritten,
1623                                            wc->w_di_bh, wc->w_handle,
1624                                            data_ac, meta_ac, NULL);
1625                 /*
1626                  * This shouldn't happen because we must have already
1627                  * calculated the correct meta data allocation required. The
1628                  * internal tree allocation code should know how to increase
1629                  * transaction credits itself.
1630                  *
1631                  * If need be, we could handle -EAGAIN for a
1632                  * RESTART_TRANS here.
1633                  */
1634                 mlog_bug_on_msg(ret == -EAGAIN,
1635                                 "Inode %llu: EAGAIN return during allocation.\n",
1636                                 (unsigned long long)OCFS2_I(inode)->ip_blkno);
1637                 if (ret < 0) {
1638                         mlog_errno(ret);
1639                         goto out;
1640                 }
1641         } else if (clear_unwritten) {
1642                 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode),
1643                                               wc->w_di_bh);
1644                 ret = ocfs2_mark_extent_written(inode, &et,
1645                                                 wc->w_handle, cpos, 1, *phys,
1646                                                 meta_ac, &wc->w_dealloc);
1647                 if (ret < 0) {
1648                         mlog_errno(ret);
1649                         goto out;
1650                 }
1651         }
1652
1653         /*
1654          * The only reason this should fail is due to an inability to
1655          * find the extent added.
1656          */
1657         ret = ocfs2_get_clusters(inode, cpos, phys, NULL, NULL);
1658         if (ret < 0) {
1659                 mlog(ML_ERROR, "Get physical blkno failed for inode %llu, "
1660                             "at logical cluster %u",
1661                             (unsigned long long)OCFS2_I(inode)->ip_blkno, cpos);
1662                 goto out;
1663         }
1664
1665         BUG_ON(*phys == 0);
1666
1667         p_blkno = ocfs2_clusters_to_blocks(inode->i_sb, *phys);
1668         if (!should_zero)
1669                 p_blkno += (user_pos >> inode->i_sb->s_blocksize_bits) & (u64)(bpc - 1);
1670
1671         for(i = 0; i < wc->w_num_pages; i++) {
1672                 int tmpret;
1673
1674                 /* This is the direct io target page. */
1675                 if (wc->w_pages[i] == NULL) {
1676                         p_blkno++;
1677                         continue;
1678                 }
1679
1680                 tmpret = ocfs2_prepare_page_for_write(inode, &p_blkno, wc,
1681                                                       wc->w_pages[i], cpos,
1682                                                       user_pos, user_len,
1683                                                       should_zero);
1684                 if (tmpret) {
1685                         mlog_errno(tmpret);
1686                         if (ret == 0)
1687                                 ret = tmpret;
1688                 }
1689         }
1690
1691         /*
1692          * We only have cleanup to do in case of allocating write.
1693          */
1694         if (ret && new)
1695                 ocfs2_write_failure(inode, wc, user_pos, user_len);
1696
1697 out:
1698
1699         return ret;
1700 }
1701
1702 static int ocfs2_write_cluster_by_desc(struct address_space *mapping,
1703                                        struct ocfs2_alloc_context *data_ac,
1704                                        struct ocfs2_alloc_context *meta_ac,
1705                                        struct ocfs2_write_ctxt *wc,
1706                                        loff_t pos, unsigned len)
1707 {
1708         int ret, i;
1709         loff_t cluster_off;
1710         unsigned int local_len = len;
1711         struct ocfs2_write_cluster_desc *desc;
1712         struct ocfs2_super *osb = OCFS2_SB(mapping->host->i_sb);
1713
1714         for (i = 0; i < wc->w_clen; i++) {
1715                 desc = &wc->w_desc[i];
1716
1717                 /*
1718                  * We have to make sure that the total write passed in
1719                  * doesn't extend past a single cluster.
1720                  */
1721                 local_len = len;
1722                 cluster_off = pos & (osb->s_clustersize - 1);
1723                 if ((cluster_off + local_len) > osb->s_clustersize)
1724                         local_len = osb->s_clustersize - cluster_off;
1725
1726                 ret = ocfs2_write_cluster(mapping, &desc->c_phys,
1727                                           desc->c_new,
1728                                           desc->c_clear_unwritten,
1729                                           desc->c_needs_zero,
1730                                           data_ac, meta_ac,
1731                                           wc, desc->c_cpos, pos, local_len);
1732                 if (ret) {
1733                         mlog_errno(ret);
1734                         goto out;
1735                 }
1736
1737                 len -= local_len;
1738                 pos += local_len;
1739         }
1740
1741         ret = 0;
1742 out:
1743         return ret;
1744 }
1745
1746 /*
1747  * ocfs2_write_end() wants to know which parts of the target page it
1748  * should complete the write on. It's easiest to compute them ahead of
1749  * time when a more complete view of the write is available.
1750  */
1751 static void ocfs2_set_target_boundaries(struct ocfs2_super *osb,
1752                                         struct ocfs2_write_ctxt *wc,
1753                                         loff_t pos, unsigned len, int alloc)
1754 {
1755         struct ocfs2_write_cluster_desc *desc;
1756
1757         wc->w_target_from = pos & (PAGE_CACHE_SIZE - 1);
1758         wc->w_target_to = wc->w_target_from + len;
1759
1760         if (alloc == 0)
1761                 return;
1762
1763         /*
1764          * Allocating write - we may have different boundaries based
1765          * on page size and cluster size.
1766          *
1767          * NOTE: We can no longer compute one value from the other as
1768          * the actual write length and user provided length may be
1769          * different.
1770          */
1771
1772         if (wc->w_large_pages) {
1773                 /*
1774                  * We only care about the 1st and last cluster within
1775                  * our range and whether they should be zero'd or not. Either
1776                  * value may be extended out to the start/end of a
1777                  * newly allocated cluster.
1778                  */
1779                 desc = &wc->w_desc[0];
1780                 if (desc->c_needs_zero)
1781                         ocfs2_figure_cluster_boundaries(osb,
1782                                                         desc->c_cpos,
1783                                                         &wc->w_target_from,
1784                                                         NULL);
1785
1786                 desc = &wc->w_desc[wc->w_clen - 1];
1787                 if (desc->c_needs_zero)
1788                         ocfs2_figure_cluster_boundaries(osb,
1789                                                         desc->c_cpos,
1790                                                         NULL,
1791                                                         &wc->w_target_to);
1792         } else {
1793                 wc->w_target_from = 0;
1794                 wc->w_target_to = PAGE_CACHE_SIZE;
1795         }
1796 }
1797
1798 /*
1799  * Populate each single-cluster write descriptor in the write context
1800  * with information about the i/o to be done.
1801  *
1802  * Returns the number of clusters that will have to be allocated, as
1803  * well as a worst case estimate of the number of extent records that
1804  * would have to be created during a write to an unwritten region.
1805  */
1806 static int ocfs2_populate_write_desc(struct inode *inode,
1807                                      struct ocfs2_write_ctxt *wc,
1808                                      unsigned int *clusters_to_alloc,
1809                                      unsigned int *extents_to_split)
1810 {
1811         int ret;
1812         struct ocfs2_write_cluster_desc *desc;
1813         unsigned int num_clusters = 0;
1814         unsigned int ext_flags = 0;
1815         u32 phys = 0;
1816         int i;
1817
1818         *clusters_to_alloc = 0;
1819         *extents_to_split = 0;
1820
1821         for (i = 0; i < wc->w_clen; i++) {
1822                 desc = &wc->w_desc[i];
1823                 desc->c_cpos = wc->w_cpos + i;
1824
1825                 if (num_clusters == 0) {
1826                         /*
1827                          * Need to look up the next extent record.
1828                          */
1829                         ret = ocfs2_get_clusters(inode, desc->c_cpos, &phys,
1830                                                  &num_clusters, &ext_flags);
1831                         if (ret) {
1832                                 mlog_errno(ret);
1833                                 goto out;
1834                         }
1835
1836                         /* We should already CoW the refcountd extent. */
1837                         BUG_ON(ext_flags & OCFS2_EXT_REFCOUNTED);
1838
1839                         /*
1840                          * Assume worst case - that we're writing in
1841                          * the middle of the extent.
1842                          *
1843                          * We can assume that the write proceeds from
1844                          * left to right, in which case the extent
1845                          * insert code is smart enough to coalesce the
1846                          * next splits into the previous records created.
1847                          */
1848                         if (ext_flags & OCFS2_EXT_UNWRITTEN)
1849                                 *extents_to_split = *extents_to_split + 2;
1850                 } else if (phys) {
1851                         /*
1852                          * Only increment phys if it doesn't describe
1853                          * a hole.
1854                          */
1855                         phys++;
1856                 }
1857
1858                 /*
1859                  * If w_first_new_cpos is < UINT_MAX, we have a non-sparse
1860                  * file that got extended.  w_first_new_cpos tells us
1861                  * where the newly allocated clusters are so we can
1862                  * zero them.
1863                  */
1864                 if (desc->c_cpos >= wc->w_first_new_cpos) {
1865                         BUG_ON(phys == 0);
1866                         desc->c_needs_zero = 1;
1867                 }
1868
1869                 desc->c_phys = phys;
1870                 if (phys == 0) {
1871                         desc->c_new = 1;
1872                         desc->c_needs_zero = 1;
1873                         desc->c_clear_unwritten = 1;
1874                         *clusters_to_alloc = *clusters_to_alloc + 1;
1875                 }
1876
1877                 if (ext_flags & OCFS2_EXT_UNWRITTEN) {
1878                         desc->c_clear_unwritten = 1;
1879                         desc->c_needs_zero = 1;
1880                 }
1881
1882                 num_clusters--;
1883         }
1884
1885         ret = 0;
1886 out:
1887         return ret;
1888 }
1889
1890 static int ocfs2_write_begin_inline(struct address_space *mapping,
1891                                     struct inode *inode,
1892                                     struct ocfs2_write_ctxt *wc)
1893 {
1894         int ret;
1895         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1896         struct page *page;
1897         handle_t *handle;
1898         struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1899
1900         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1901         if (IS_ERR(handle)) {
1902                 ret = PTR_ERR(handle);
1903                 mlog_errno(ret);
1904                 goto out;
1905         }
1906
1907         page = find_or_create_page(mapping, 0, GFP_NOFS);
1908         if (!page) {
1909                 ocfs2_commit_trans(osb, handle);
1910                 ret = -ENOMEM;
1911                 mlog_errno(ret);
1912                 goto out;
1913         }
1914         /*
1915          * If we don't set w_num_pages then this page won't get unlocked
1916          * and freed on cleanup of the write context.
1917          */
1918         wc->w_pages[0] = wc->w_target_page = page;
1919         wc->w_num_pages = 1;
1920
1921         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), wc->w_di_bh,
1922                                       OCFS2_JOURNAL_ACCESS_WRITE);
1923         if (ret) {
1924                 ocfs2_commit_trans(osb, handle);
1925
1926                 mlog_errno(ret);
1927                 goto out;
1928         }
1929
1930         if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL))
1931                 ocfs2_set_inode_data_inline(inode, di);
1932
1933         if (!PageUptodate(page)) {
1934                 ret = ocfs2_read_inline_data(inode, page, wc->w_di_bh);
1935                 if (ret) {
1936                         ocfs2_commit_trans(osb, handle);
1937
1938                         goto out;
1939                 }
1940         }
1941
1942         wc->w_handle = handle;
1943 out:
1944         return ret;
1945 }
1946
1947 int ocfs2_size_fits_inline_data(struct buffer_head *di_bh, u64 new_size)
1948 {
1949         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
1950
1951         if (new_size <= le16_to_cpu(di->id2.i_data.id_count))
1952                 return 1;
1953         return 0;
1954 }
1955
1956 static int ocfs2_try_to_write_inline_data(struct address_space *mapping,
1957                                           struct inode *inode, loff_t pos,
1958                                           unsigned len, struct page *mmap_page,
1959                                           struct ocfs2_write_ctxt *wc)
1960 {
1961         int ret, written = 0;
1962         loff_t end = pos + len;
1963         struct ocfs2_inode_info *oi = OCFS2_I(inode);
1964         struct ocfs2_dinode *di = NULL;
1965
1966         trace_ocfs2_try_to_write_inline_data((unsigned long long)oi->ip_blkno,
1967                                              len, (unsigned long long)pos,
1968                                              oi->ip_dyn_features);
1969
1970         /*
1971          * Handle inodes which already have inline data 1st.
1972          */
1973         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1974                 if (mmap_page == NULL &&
1975                     ocfs2_size_fits_inline_data(wc->w_di_bh, end))
1976                         goto do_inline_write;
1977
1978                 /*
1979                  * The write won't fit - we have to give this inode an
1980                  * inline extent list now.
1981                  */
1982                 ret = ocfs2_convert_inline_data_to_extents(inode, wc->w_di_bh);
1983                 if (ret)
1984                         mlog_errno(ret);
1985                 goto out;
1986         }
1987
1988         /*
1989          * Check whether the inode can accept inline data.
1990          */
1991         if (oi->ip_clusters != 0 || i_size_read(inode) != 0)
1992                 return 0;
1993
1994         /*
1995          * Check whether the write can fit.
1996          */
1997         di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1998         if (mmap_page ||
1999             end > ocfs2_max_inline_data_with_xattr(inode->i_sb, di))
2000                 return 0;
2001
2002 do_inline_write:
2003         ret = ocfs2_write_begin_inline(mapping, inode, wc);
2004         if (ret) {
2005                 mlog_errno(ret);
2006                 goto out;
2007         }
2008
2009         /*
2010          * This signals to the caller that the data can be written
2011          * inline.
2012          */
2013         written = 1;
2014 out:
2015         return written ? written : ret;
2016 }
2017
2018 /*
2019  * This function only does anything for file systems which can't
2020  * handle sparse files.
2021  *
2022  * What we want to do here is fill in any hole between the current end
2023  * of allocation and the end of our write. That way the rest of the
2024  * write path can treat it as an non-allocating write, which has no
2025  * special case code for sparse/nonsparse files.
2026  */
2027 static int ocfs2_expand_nonsparse_inode(struct inode *inode,
2028                                         struct buffer_head *di_bh,
2029                                         loff_t pos, unsigned len,
2030                                         struct ocfs2_write_ctxt *wc)
2031 {
2032         int ret;
2033         loff_t newsize = pos + len;
2034
2035         BUG_ON(ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)));
2036
2037         if (newsize <= i_size_read(inode))
2038                 return 0;
2039
2040         ret = ocfs2_extend_no_holes(inode, di_bh, newsize, pos);
2041         if (ret)
2042                 mlog_errno(ret);
2043
2044         /* There is no wc if this is call from direct. */
2045         if (wc)
2046                 wc->w_first_new_cpos =
2047                         ocfs2_clusters_for_bytes(inode->i_sb, i_size_read(inode));
2048
2049         return ret;
2050 }
2051
2052 static int ocfs2_zero_tail(struct inode *inode, struct buffer_head *di_bh,
2053                            loff_t pos)
2054 {
2055         int ret = 0;
2056
2057         BUG_ON(!ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)));
2058         if (pos > i_size_read(inode))
2059                 ret = ocfs2_zero_extend(inode, di_bh, pos);
2060
2061         return ret;
2062 }
2063
2064 /*
2065  * Try to flush truncate logs if we can free enough clusters from it.
2066  * As for return value, "< 0" means error, "0" no space and "1" means
2067  * we have freed enough spaces and let the caller try to allocate again.
2068  */
2069 static int ocfs2_try_to_free_truncate_log(struct ocfs2_super *osb,
2070                                           unsigned int needed)
2071 {
2072         tid_t target;
2073         int ret = 0;
2074         unsigned int truncated_clusters;
2075
2076         inode_lock(osb->osb_tl_inode);
2077         truncated_clusters = osb->truncated_clusters;
2078         inode_unlock(osb->osb_tl_inode);
2079
2080         /*
2081          * Check whether we can succeed in allocating if we free
2082          * the truncate log.
2083          */
2084         if (truncated_clusters < needed)
2085                 goto out;
2086
2087         ret = ocfs2_flush_truncate_log(osb);
2088         if (ret) {
2089                 mlog_errno(ret);
2090                 goto out;
2091         }
2092
2093         if (jbd2_journal_start_commit(osb->journal->j_journal, &target)) {
2094                 jbd2_log_wait_commit(osb->journal->j_journal, target);
2095                 ret = 1;
2096         }
2097 out:
2098         return ret;
2099 }
2100
2101 int ocfs2_write_begin_nolock(struct address_space *mapping,
2102                              loff_t pos, unsigned len, ocfs2_write_type_t type,
2103                              struct page **pagep, void **fsdata,
2104                              struct buffer_head *di_bh, struct page *mmap_page)
2105 {
2106         int ret, cluster_of_pages, credits = OCFS2_INODE_UPDATE_CREDITS;
2107         unsigned int clusters_to_alloc, extents_to_split, clusters_need = 0;
2108         struct ocfs2_write_ctxt *wc;
2109         struct inode *inode = mapping->host;
2110         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2111         struct ocfs2_dinode *di;
2112         struct ocfs2_alloc_context *data_ac = NULL;
2113         struct ocfs2_alloc_context *meta_ac = NULL;
2114         handle_t *handle;
2115         struct ocfs2_extent_tree et;
2116         int try_free = 1, ret1;
2117
2118 try_again:
2119         ret = ocfs2_alloc_write_ctxt(&wc, osb, pos, len, type, di_bh);
2120         if (ret) {
2121                 mlog_errno(ret);
2122                 return ret;
2123         }
2124
2125         if (ocfs2_supports_inline_data(osb)) {
2126                 ret = ocfs2_try_to_write_inline_data(mapping, inode, pos, len,
2127                                                      mmap_page, wc);
2128                 if (ret == 1) {
2129                         ret = 0;
2130                         goto success;
2131                 }
2132                 if (ret < 0) {
2133                         mlog_errno(ret);
2134                         goto out;
2135                 }
2136         }
2137
2138         /* Direct io change i_size late, should not zero tail here. */
2139         if (type != OCFS2_WRITE_DIRECT) {
2140                 if (ocfs2_sparse_alloc(osb))
2141                         ret = ocfs2_zero_tail(inode, di_bh, pos);
2142                 else
2143                         ret = ocfs2_expand_nonsparse_inode(inode, di_bh, pos,
2144                                                            len, wc);
2145                 if (ret) {
2146                         mlog_errno(ret);
2147                         goto out;
2148                 }
2149         }
2150
2151         ret = ocfs2_check_range_for_refcount(inode, pos, len);
2152         if (ret < 0) {
2153                 mlog_errno(ret);
2154                 goto out;
2155         } else if (ret == 1) {
2156                 clusters_need = wc->w_clen;
2157                 ret = ocfs2_refcount_cow(inode, di_bh,
2158                                          wc->w_cpos, wc->w_clen, UINT_MAX);
2159                 if (ret) {
2160                         mlog_errno(ret);
2161                         goto out;
2162                 }
2163         }
2164
2165         ret = ocfs2_populate_write_desc(inode, wc, &clusters_to_alloc,
2166                                         &extents_to_split);
2167         if (ret) {
2168                 mlog_errno(ret);
2169                 goto out;
2170         }
2171         clusters_need += clusters_to_alloc;
2172
2173         di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
2174
2175         trace_ocfs2_write_begin_nolock(
2176                         (unsigned long long)OCFS2_I(inode)->ip_blkno,
2177                         (long long)i_size_read(inode),
2178                         le32_to_cpu(di->i_clusters),
2179                         pos, len, type, mmap_page,
2180                         clusters_to_alloc, extents_to_split);
2181
2182         /*
2183          * We set w_target_from, w_target_to here so that
2184          * ocfs2_write_end() knows which range in the target page to
2185          * write out. An allocation requires that we write the entire
2186          * cluster range.
2187          */
2188         if (clusters_to_alloc || extents_to_split) {
2189                 /*
2190                  * XXX: We are stretching the limits of
2191                  * ocfs2_lock_allocators(). It greatly over-estimates
2192                  * the work to be done.
2193                  */
2194                 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode),
2195                                               wc->w_di_bh);
2196                 ret = ocfs2_lock_allocators(inode, &et,
2197                                             clusters_to_alloc, extents_to_split,
2198                                             &data_ac, &meta_ac);
2199                 if (ret) {
2200                         mlog_errno(ret);
2201                         goto out;
2202                 }
2203
2204                 if (data_ac)
2205                         data_ac->ac_resv = &OCFS2_I(inode)->ip_la_data_resv;
2206
2207                 credits = ocfs2_calc_extend_credits(inode->i_sb,
2208                                                     &di->id2.i_list);
2209         } else if (type == OCFS2_WRITE_DIRECT)
2210                 /* direct write needs not to start trans if no extents alloc. */
2211                 goto success;
2212
2213         /*
2214          * We have to zero sparse allocated clusters, unwritten extent clusters,
2215          * and non-sparse clusters we just extended.  For non-sparse writes,
2216          * we know zeros will only be needed in the first and/or last cluster.
2217          */
2218         if (clusters_to_alloc || extents_to_split ||
2219             (wc->w_clen && (wc->w_desc[0].c_needs_zero ||
2220                             wc->w_desc[wc->w_clen - 1].c_needs_zero)))
2221                 cluster_of_pages = 1;
2222         else
2223                 cluster_of_pages = 0;
2224
2225         ocfs2_set_target_boundaries(osb, wc, pos, len, cluster_of_pages);
2226
2227         handle = ocfs2_start_trans(osb, credits);
2228         if (IS_ERR(handle)) {
2229                 ret = PTR_ERR(handle);
2230                 mlog_errno(ret);
2231                 goto out;
2232         }
2233
2234         wc->w_handle = handle;
2235
2236         if (clusters_to_alloc) {
2237                 ret = dquot_alloc_space_nodirty(inode,
2238                         ocfs2_clusters_to_bytes(osb->sb, clusters_to_alloc));
2239                 if (ret)
2240                         goto out_commit;
2241         }
2242
2243         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), wc->w_di_bh,
2244                                       OCFS2_JOURNAL_ACCESS_WRITE);
2245         if (ret) {
2246                 mlog_errno(ret);
2247                 goto out_quota;
2248         }
2249
2250         /*
2251          * Fill our page array first. That way we've grabbed enough so
2252          * that we can zero and flush if we error after adding the
2253          * extent.
2254          */
2255         ret = ocfs2_grab_pages_for_write(mapping, wc, wc->w_cpos, pos, len,
2256                                          cluster_of_pages, mmap_page);
2257         if (ret && ret != -EAGAIN) {
2258                 mlog_errno(ret);
2259                 goto out_quota;
2260         }
2261
2262         /*
2263          * ocfs2_grab_pages_for_write() returns -EAGAIN if it could not lock
2264          * the target page. In this case, we exit with no error and no target
2265          * page. This will trigger the caller, page_mkwrite(), to re-try
2266          * the operation.
2267          */
2268         if (ret == -EAGAIN) {
2269                 BUG_ON(wc->w_target_page);
2270                 ret = 0;
2271                 goto out_quota;
2272         }
2273
2274         ret = ocfs2_write_cluster_by_desc(mapping, data_ac, meta_ac, wc, pos,
2275                                           len);
2276         if (ret) {
2277                 mlog_errno(ret);
2278                 goto out_quota;
2279         }
2280
2281         if (data_ac)
2282                 ocfs2_free_alloc_context(data_ac);
2283         if (meta_ac)
2284                 ocfs2_free_alloc_context(meta_ac);
2285
2286 success:
2287         if (pagep)
2288                 *pagep = wc->w_target_page;
2289         *fsdata = wc;
2290         return 0;
2291 out_quota:
2292         if (clusters_to_alloc)
2293                 dquot_free_space(inode,
2294                           ocfs2_clusters_to_bytes(osb->sb, clusters_to_alloc));
2295 out_commit:
2296         ocfs2_commit_trans(osb, handle);
2297
2298 out:
2299         ocfs2_free_write_ctxt(wc);
2300
2301         if (data_ac) {
2302                 ocfs2_free_alloc_context(data_ac);
2303                 data_ac = NULL;
2304         }
2305         if (meta_ac) {
2306                 ocfs2_free_alloc_context(meta_ac);
2307                 meta_ac = NULL;
2308         }
2309
2310         if (ret == -ENOSPC && try_free) {
2311                 /*
2312                  * Try to free some truncate log so that we can have enough
2313                  * clusters to allocate.
2314                  */
2315                 try_free = 0;
2316
2317                 ret1 = ocfs2_try_to_free_truncate_log(osb, clusters_need);
2318                 if (ret1 == 1)
2319                         goto try_again;
2320
2321                 if (ret1 < 0)
2322                         mlog_errno(ret1);
2323         }
2324
2325         return ret;
2326 }
2327
2328 static int ocfs2_write_begin(struct file *file, struct address_space *mapping,
2329                              loff_t pos, unsigned len, unsigned flags,
2330                              struct page **pagep, void **fsdata)
2331 {
2332         int ret;
2333         struct buffer_head *di_bh = NULL;
2334         struct inode *inode = mapping->host;
2335
2336         ret = ocfs2_inode_lock(inode, &di_bh, 1);
2337         if (ret) {
2338                 mlog_errno(ret);
2339                 return ret;
2340         }
2341
2342         /*
2343          * Take alloc sem here to prevent concurrent lookups. That way
2344          * the mapping, zeroing and tree manipulation within
2345          * ocfs2_write() will be safe against ->readpage(). This
2346          * should also serve to lock out allocation from a shared
2347          * writeable region.
2348          */
2349         down_write(&OCFS2_I(inode)->ip_alloc_sem);
2350
2351         ret = ocfs2_write_begin_nolock(mapping, pos, len, OCFS2_WRITE_BUFFER,
2352                                        pagep, fsdata, di_bh, NULL);
2353         if (ret) {
2354                 mlog_errno(ret);
2355                 goto out_fail;
2356         }
2357
2358         brelse(di_bh);
2359
2360         return 0;
2361
2362 out_fail:
2363         up_write(&OCFS2_I(inode)->ip_alloc_sem);
2364
2365         brelse(di_bh);
2366         ocfs2_inode_unlock(inode, 1);
2367
2368         return ret;
2369 }
2370
2371 static void ocfs2_write_end_inline(struct inode *inode, loff_t pos,
2372                                    unsigned len, unsigned *copied,
2373                                    struct ocfs2_dinode *di,
2374                                    struct ocfs2_write_ctxt *wc)
2375 {
2376         void *kaddr;
2377
2378         if (unlikely(*copied < len)) {
2379                 if (!PageUptodate(wc->w_target_page)) {
2380                         *copied = 0;
2381                         return;
2382                 }
2383         }
2384
2385         kaddr = kmap_atomic(wc->w_target_page);
2386         memcpy(di->id2.i_data.id_data + pos, kaddr + pos, *copied);
2387         kunmap_atomic(kaddr);
2388
2389         trace_ocfs2_write_end_inline(
2390              (unsigned long long)OCFS2_I(inode)->ip_blkno,
2391              (unsigned long long)pos, *copied,
2392              le16_to_cpu(di->id2.i_data.id_count),
2393              le16_to_cpu(di->i_dyn_features));
2394 }
2395
2396 int ocfs2_write_end_nolock(struct address_space *mapping,
2397                            loff_t pos, unsigned len, unsigned copied,
2398                            struct page *page, void *fsdata)
2399 {
2400         int i, ret;
2401         unsigned from, to, start = pos & (PAGE_CACHE_SIZE - 1);
2402         struct inode *inode = mapping->host;
2403         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2404         struct ocfs2_write_ctxt *wc = fsdata;
2405         struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
2406         handle_t *handle = wc->w_handle;
2407         struct page *tmppage;
2408
2409         if (handle) {
2410                 ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode),
2411                                 wc->w_di_bh, OCFS2_JOURNAL_ACCESS_WRITE);
2412                 if (ret) {
2413                         copied = ret;
2414                         mlog_errno(ret);
2415                         goto out;
2416                 }
2417         }
2418
2419         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
2420                 ocfs2_write_end_inline(inode, pos, len, &copied, di, wc);
2421                 goto out_write_size;
2422         }
2423
2424         if (unlikely(copied < len) && wc->w_target_page) {
2425                 if (!PageUptodate(wc->w_target_page))
2426                         copied = 0;
2427
2428                 ocfs2_zero_new_buffers(wc->w_target_page, start+copied,
2429                                        start+len);
2430         }
2431         if (wc->w_target_page)
2432                 flush_dcache_page(wc->w_target_page);
2433
2434         for(i = 0; i < wc->w_num_pages; i++) {
2435                 tmppage = wc->w_pages[i];
2436
2437                 /* This is the direct io target page. */
2438                 if (tmppage == NULL)
2439                         continue;
2440
2441                 if (tmppage == wc->w_target_page) {
2442                         from = wc->w_target_from;
2443                         to = wc->w_target_to;
2444
2445                         BUG_ON(from > PAGE_CACHE_SIZE ||
2446                                to > PAGE_CACHE_SIZE ||
2447                                to < from);
2448                 } else {
2449                         /*
2450                          * Pages adjacent to the target (if any) imply
2451                          * a hole-filling write in which case we want
2452                          * to flush their entire range.
2453                          */
2454                         from = 0;
2455                         to = PAGE_CACHE_SIZE;
2456                 }
2457
2458                 if (page_has_buffers(tmppage)) {
2459                         if (handle && ocfs2_should_order_data(inode))
2460                                 ocfs2_jbd2_file_inode(handle, inode);
2461                         block_commit_write(tmppage, from, to);
2462                 }
2463         }
2464
2465 out_write_size:
2466         /* Direct io do not update i_size here. */
2467         if (wc->w_type != OCFS2_WRITE_DIRECT) {
2468                 pos += copied;
2469                 if (pos > i_size_read(inode)) {
2470                         i_size_write(inode, pos);
2471                         mark_inode_dirty(inode);
2472                 }
2473                 inode->i_blocks = ocfs2_inode_sector_count(inode);
2474                 di->i_size = cpu_to_le64((u64)i_size_read(inode));
2475                 inode->i_mtime = inode->i_ctime = CURRENT_TIME;
2476                 di->i_mtime = di->i_ctime = cpu_to_le64(inode->i_mtime.tv_sec);
2477                 di->i_mtime_nsec = di->i_ctime_nsec = cpu_to_le32(inode->i_mtime.tv_nsec);
2478                 ocfs2_update_inode_fsync_trans(handle, inode, 1);
2479         }
2480         if (handle)
2481                 ocfs2_journal_dirty(handle, wc->w_di_bh);
2482
2483 out:
2484         /* unlock pages before dealloc since it needs acquiring j_trans_barrier
2485          * lock, or it will cause a deadlock since journal commit threads holds
2486          * this lock and will ask for the page lock when flushing the data.
2487          * put it here to preserve the unlock order.
2488          */
2489         ocfs2_unlock_pages(wc);
2490
2491         if (handle)
2492                 ocfs2_commit_trans(osb, handle);
2493
2494         ocfs2_run_deallocs(osb, &wc->w_dealloc);
2495
2496         brelse(wc->w_di_bh);
2497         kfree(wc);
2498
2499         return copied;
2500 }
2501
2502 static int ocfs2_write_end(struct file *file, struct address_space *mapping,
2503                            loff_t pos, unsigned len, unsigned copied,
2504                            struct page *page, void *fsdata)
2505 {
2506         int ret;
2507         struct inode *inode = mapping->host;
2508
2509         ret = ocfs2_write_end_nolock(mapping, pos, len, copied, page, fsdata);
2510
2511         up_write(&OCFS2_I(inode)->ip_alloc_sem);
2512         ocfs2_inode_unlock(inode, 1);
2513
2514         return ret;
2515 }
2516
2517 const struct address_space_operations ocfs2_aops = {
2518         .readpage               = ocfs2_readpage,
2519         .readpages              = ocfs2_readpages,
2520         .writepage              = ocfs2_writepage,
2521         .write_begin            = ocfs2_write_begin,
2522         .write_end              = ocfs2_write_end,
2523         .bmap                   = ocfs2_bmap,
2524         .direct_IO              = ocfs2_direct_IO,
2525         .invalidatepage         = block_invalidatepage,
2526         .releasepage            = ocfs2_releasepage,
2527         .migratepage            = buffer_migrate_page,
2528         .is_partially_uptodate  = block_is_partially_uptodate,
2529         .error_remove_page      = generic_error_remove_page,
2530 };