Root/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
32#define MLOG_MASK_PREFIX ML_FILE_IO
33#include <cluster/masklog.h>
34
35#include "ocfs2.h"
36
37#include "alloc.h"
38#include "aops.h"
39#include "dlmglue.h"
40#include "extent_map.h"
41#include "file.h"
42#include "inode.h"
43#include "journal.h"
44#include "suballoc.h"
45#include "super.h"
46#include "symlink.h"
47#include "refcounttree.h"
48
49#include "buffer_head_io.h"
50
51static int ocfs2_symlink_get_block(struct inode *inode, sector_t iblock,
52                   struct buffer_head *bh_result, int create)
53{
54    int err = -EIO;
55    int status;
56    struct ocfs2_dinode *fe = NULL;
57    struct buffer_head *bh = NULL;
58    struct buffer_head *buffer_cache_bh = NULL;
59    struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
60    void *kaddr;
61
62    mlog_entry("(0x%p, %llu, 0x%p, %d)\n", inode,
63           (unsigned long long)iblock, bh_result, create);
64
65    BUG_ON(ocfs2_inode_is_fast_symlink(inode));
66
67    if ((iblock << inode->i_sb->s_blocksize_bits) > PATH_MAX + 1) {
68        mlog(ML_ERROR, "block offset > PATH_MAX: %llu",
69             (unsigned long long)iblock);
70        goto bail;
71    }
72
73    status = ocfs2_read_inode_block(inode, &bh);
74    if (status < 0) {
75        mlog_errno(status);
76        goto bail;
77    }
78    fe = (struct ocfs2_dinode *) bh->b_data;
79
80    if ((u64)iblock >= ocfs2_clusters_to_blocks(inode->i_sb,
81                            le32_to_cpu(fe->i_clusters))) {
82        mlog(ML_ERROR, "block offset is outside the allocated size: "
83             "%llu\n", (unsigned long long)iblock);
84        goto bail;
85    }
86
87    /* We don't use the page cache to create symlink data, so if
88     * need be, copy it over from the buffer cache. */
89    if (!buffer_uptodate(bh_result) && ocfs2_inode_is_new(inode)) {
90        u64 blkno = le64_to_cpu(fe->id2.i_list.l_recs[0].e_blkno) +
91                iblock;
92        buffer_cache_bh = sb_getblk(osb->sb, blkno);
93        if (!buffer_cache_bh) {
94            mlog(ML_ERROR, "couldn't getblock for symlink!\n");
95            goto bail;
96        }
97
98        /* we haven't locked out transactions, so a commit
99         * could've happened. Since we've got a reference on
100         * the bh, even if it commits while we're doing the
101         * copy, the data is still good. */
102        if (buffer_jbd(buffer_cache_bh)
103            && ocfs2_inode_is_new(inode)) {
104            kaddr = kmap_atomic(bh_result->b_page, KM_USER0);
105            if (!kaddr) {
106                mlog(ML_ERROR, "couldn't kmap!\n");
107                goto bail;
108            }
109            memcpy(kaddr + (bh_result->b_size * iblock),
110                   buffer_cache_bh->b_data,
111                   bh_result->b_size);
112            kunmap_atomic(kaddr, KM_USER0);
113            set_buffer_uptodate(bh_result);
114        }
115        brelse(buffer_cache_bh);
116    }
117
118    map_bh(bh_result, inode->i_sb,
119           le64_to_cpu(fe->id2.i_list.l_recs[0].e_blkno) + iblock);
120
121    err = 0;
122
123bail:
124    brelse(bh);
125
126    mlog_exit(err);
127    return err;
128}
129
130int ocfs2_get_block(struct inode *inode, sector_t iblock,
131            struct buffer_head *bh_result, int create)
132{
133    int err = 0;
134    unsigned int ext_flags;
135    u64 max_blocks = bh_result->b_size >> inode->i_blkbits;
136    u64 p_blkno, count, past_eof;
137    struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
138
139    mlog_entry("(0x%p, %llu, 0x%p, %d)\n", inode,
140           (unsigned long long)iblock, bh_result, create);
141
142    if (OCFS2_I(inode)->ip_flags & OCFS2_INODE_SYSTEM_FILE)
143        mlog(ML_NOTICE, "get_block on system inode 0x%p (%lu)\n",
144             inode, inode->i_ino);
145
146    if (S_ISLNK(inode->i_mode)) {
147        /* this always does I/O for some reason. */
148        err = ocfs2_symlink_get_block(inode, iblock, bh_result, create);
149        goto bail;
150    }
151
152    err = ocfs2_extent_map_get_blocks(inode, iblock, &p_blkno, &count,
153                      &ext_flags);
154    if (err) {
155        mlog(ML_ERROR, "Error %d from get_blocks(0x%p, %llu, 1, "
156             "%llu, NULL)\n", err, inode, (unsigned long long)iblock,
157             (unsigned long long)p_blkno);
158        goto bail;
159    }
160
161    if (max_blocks < count)
162        count = max_blocks;
163
164    /*
165     * ocfs2 never allocates in this function - the only time we
166     * need to use BH_New is when we're extending i_size on a file
167     * system which doesn't support holes, in which case BH_New
168     * allows block_prepare_write() to zero.
169     *
170     * If we see this on a sparse file system, then a truncate has
171     * raced us and removed the cluster. In this case, we clear
172     * the buffers dirty and uptodate bits and let the buffer code
173     * ignore it as a hole.
174     */
175    if (create && p_blkno == 0 && ocfs2_sparse_alloc(osb)) {
176        clear_buffer_dirty(bh_result);
177        clear_buffer_uptodate(bh_result);
178        goto bail;
179    }
180
181    /* Treat the unwritten extent as a hole for zeroing purposes. */
182    if (p_blkno && !(ext_flags & OCFS2_EXT_UNWRITTEN))
183        map_bh(bh_result, inode->i_sb, p_blkno);
184
185    bh_result->b_size = count << inode->i_blkbits;
186
187    if (!ocfs2_sparse_alloc(osb)) {
188        if (p_blkno == 0) {
189            err = -EIO;
190            mlog(ML_ERROR,
191                 "iblock = %llu p_blkno = %llu blkno=(%llu)\n",
192                 (unsigned long long)iblock,
193                 (unsigned long long)p_blkno,
194                 (unsigned long long)OCFS2_I(inode)->ip_blkno);
195            mlog(ML_ERROR, "Size %llu, clusters %u\n", (unsigned long long)i_size_read(inode), OCFS2_I(inode)->ip_clusters);
196            dump_stack();
197            goto bail;
198        }
199
200        past_eof = ocfs2_blocks_for_bytes(inode->i_sb, i_size_read(inode));
201        mlog(0, "Inode %lu, past_eof = %llu\n", inode->i_ino,
202             (unsigned long long)past_eof);
203
204        if (create && (iblock >= past_eof))
205            set_buffer_new(bh_result);
206    }
207
208bail:
209    if (err < 0)
210        err = -EIO;
211
212    mlog_exit(err);
213    return err;
214}
215
216int ocfs2_read_inline_data(struct inode *inode, struct page *page,
217               struct buffer_head *di_bh)
218{
219    void *kaddr;
220    loff_t size;
221    struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
222
223    if (!(le16_to_cpu(di->i_dyn_features) & OCFS2_INLINE_DATA_FL)) {
224        ocfs2_error(inode->i_sb, "Inode %llu lost inline data flag",
225                (unsigned long long)OCFS2_I(inode)->ip_blkno);
226        return -EROFS;
227    }
228
229    size = i_size_read(inode);
230
231    if (size > PAGE_CACHE_SIZE ||
232        size > ocfs2_max_inline_data_with_xattr(inode->i_sb, di)) {
233        ocfs2_error(inode->i_sb,
234                "Inode %llu has with inline data has bad size: %Lu",
235                (unsigned long long)OCFS2_I(inode)->ip_blkno,
236                (unsigned long long)size);
237        return -EROFS;
238    }
239
240    kaddr = kmap_atomic(page, KM_USER0);
241    if (size)
242        memcpy(kaddr, di->id2.i_data.id_data, size);
243    /* Clear the remaining part of the page */
244    memset(kaddr + size, 0, PAGE_CACHE_SIZE - size);
245    flush_dcache_page(page);
246    kunmap_atomic(kaddr, KM_USER0);
247
248    SetPageUptodate(page);
249
250    return 0;
251}
252
253static int ocfs2_readpage_inline(struct inode *inode, struct page *page)
254{
255    int ret;
256    struct buffer_head *di_bh = NULL;
257
258    BUG_ON(!PageLocked(page));
259    BUG_ON(!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL));
260
261    ret = ocfs2_read_inode_block(inode, &di_bh);
262    if (ret) {
263        mlog_errno(ret);
264        goto out;
265    }
266
267    ret = ocfs2_read_inline_data(inode, page, di_bh);
268out:
269    unlock_page(page);
270
271    brelse(di_bh);
272    return ret;
273}
274
275static int ocfs2_readpage(struct file *file, struct page *page)
276{
277    struct inode *inode = page->mapping->host;
278    struct ocfs2_inode_info *oi = OCFS2_I(inode);
279    loff_t start = (loff_t)page->index << PAGE_CACHE_SHIFT;
280    int ret, unlock = 1;
281
282    mlog_entry("(0x%p, %lu)\n", file, (page ? page->index : 0));
283
284    ret = ocfs2_inode_lock_with_page(inode, NULL, 0, page);
285    if (ret != 0) {
286        if (ret == AOP_TRUNCATED_PAGE)
287            unlock = 0;
288        mlog_errno(ret);
289        goto out;
290    }
291
292    if (down_read_trylock(&oi->ip_alloc_sem) == 0) {
293        ret = AOP_TRUNCATED_PAGE;
294        goto out_inode_unlock;
295    }
296
297    /*
298     * i_size might have just been updated as we grabed the meta lock. We
299     * might now be discovering a truncate that hit on another node.
300     * block_read_full_page->get_block freaks out if it is asked to read
301     * beyond the end of a file, so we check here. Callers
302     * (generic_file_read, vm_ops->fault) are clever enough to check i_size
303     * and notice that the page they just read isn't needed.
304     *
305     * XXX sys_readahead() seems to get that wrong?
306     */
307    if (start >= i_size_read(inode)) {
308        zero_user(page, 0, PAGE_SIZE);
309        SetPageUptodate(page);
310        ret = 0;
311        goto out_alloc;
312    }
313
314    if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL)
315        ret = ocfs2_readpage_inline(inode, page);
316    else
317        ret = block_read_full_page(page, ocfs2_get_block);
318    unlock = 0;
319
320out_alloc:
321    up_read(&OCFS2_I(inode)->ip_alloc_sem);
322out_inode_unlock:
323    ocfs2_inode_unlock(inode, 0);
324out:
325    if (unlock)
326        unlock_page(page);
327    mlog_exit(ret);
328    return ret;
329}
330
331/*
332 * This is used only for read-ahead. Failures or difficult to handle
333 * situations are safe to ignore.
334 *
335 * Right now, we don't bother with BH_Boundary - in-inode extent lists
336 * are quite large (243 extents on 4k blocks), so most inodes don't
337 * grow out to a tree. If need be, detecting boundary extents could
338 * trivially be added in a future version of ocfs2_get_block().
339 */
340static int ocfs2_readpages(struct file *filp, struct address_space *mapping,
341               struct list_head *pages, unsigned nr_pages)
342{
343    int ret, err = -EIO;
344    struct inode *inode = mapping->host;
345    struct ocfs2_inode_info *oi = OCFS2_I(inode);
346    loff_t start;
347    struct page *last;
348
349    /*
350     * Use the nonblocking flag for the dlm code to avoid page
351     * lock inversion, but don't bother with retrying.
352     */
353    ret = ocfs2_inode_lock_full(inode, NULL, 0, OCFS2_LOCK_NONBLOCK);
354    if (ret)
355        return err;
356
357    if (down_read_trylock(&oi->ip_alloc_sem) == 0) {
358        ocfs2_inode_unlock(inode, 0);
359        return err;
360    }
361
362    /*
363     * Don't bother with inline-data. There isn't anything
364     * to read-ahead in that case anyway...
365     */
366    if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL)
367        goto out_unlock;
368
369    /*
370     * Check whether a remote node truncated this file - we just
371     * drop out in that case as it's not worth handling here.
372     */
373    last = list_entry(pages->prev, struct page, lru);
374    start = (loff_t)last->index << PAGE_CACHE_SHIFT;
375    if (start >= i_size_read(inode))
376        goto out_unlock;
377
378    err = mpage_readpages(mapping, pages, nr_pages, ocfs2_get_block);
379
380out_unlock:
381    up_read(&oi->ip_alloc_sem);
382    ocfs2_inode_unlock(inode, 0);
383
384    return err;
385}
386
387/* Note: Because we don't support holes, our allocation has
388 * already happened (allocation writes zeros to the file data)
389 * so we don't have to worry about ordered writes in
390 * ocfs2_writepage.
391 *
392 * ->writepage is called during the process of invalidating the page cache
393 * during blocked lock processing. It can't block on any cluster locks
394 * to during block mapping. It's relying on the fact that the block
395 * mapping can't have disappeared under the dirty pages that it is
396 * being asked to write back.
397 */
398static int ocfs2_writepage(struct page *page, struct writeback_control *wbc)
399{
400    int ret;
401
402    mlog_entry("(0x%p)\n", page);
403
404    ret = block_write_full_page(page, ocfs2_get_block, wbc);
405
406    mlog_exit(ret);
407
408    return ret;
409}
410
411/*
412 * This is called from ocfs2_write_zero_page() which has handled it's
413 * own cluster locking and has ensured allocation exists for those
414 * blocks to be written.
415 */
416int ocfs2_prepare_write_nolock(struct inode *inode, struct page *page,
417                   unsigned from, unsigned to)
418{
419    int ret;
420
421    ret = block_prepare_write(page, from, to, ocfs2_get_block);
422
423    return ret;
424}
425
426/* Taken from ext3. We don't necessarily need the full blown
427 * functionality yet, but IMHO it's better to cut and paste the whole
428 * thing so we can avoid introducing our own bugs (and easily pick up
429 * their fixes when they happen) --Mark */
430int walk_page_buffers( handle_t *handle,
431            struct buffer_head *head,
432            unsigned from,
433            unsigned to,
434            int *partial,
435            int (*fn)( handle_t *handle,
436                    struct buffer_head *bh))
437{
438    struct buffer_head *bh;
439    unsigned block_start, block_end;
440    unsigned blocksize = head->b_size;
441    int err, ret = 0;
442    struct buffer_head *next;
443
444    for ( bh = head, block_start = 0;
445        ret == 0 && (bh != head || !block_start);
446            block_start = block_end, bh = next)
447    {
448        next = bh->b_this_page;
449        block_end = block_start + blocksize;
450        if (block_end <= from || block_start >= to) {
451            if (partial && !buffer_uptodate(bh))
452                *partial = 1;
453            continue;
454        }
455        err = (*fn)(handle, bh);
456        if (!ret)
457            ret = err;
458    }
459    return ret;
460}
461
462handle_t *ocfs2_start_walk_page_trans(struct inode *inode,
463                             struct page *page,
464                             unsigned from,
465                             unsigned to)
466{
467    struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
468    handle_t *handle;
469    int ret = 0;
470
471    handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
472    if (IS_ERR(handle)) {
473        ret = -ENOMEM;
474        mlog_errno(ret);
475        goto out;
476    }
477
478    if (ocfs2_should_order_data(inode)) {
479        ret = ocfs2_jbd2_file_inode(handle, inode);
480        if (ret < 0)
481            mlog_errno(ret);
482    }
483out:
484    if (ret) {
485        if (!IS_ERR(handle))
486            ocfs2_commit_trans(osb, handle);
487        handle = ERR_PTR(ret);
488    }
489    return handle;
490}
491
492static sector_t ocfs2_bmap(struct address_space *mapping, sector_t block)
493{
494    sector_t status;
495    u64 p_blkno = 0;
496    int err = 0;
497    struct inode *inode = mapping->host;
498
499    mlog_entry("(block = %llu)\n", (unsigned long long)block);
500
501    /* We don't need to lock journal system files, since they aren't
502     * accessed concurrently from multiple nodes.
503     */
504    if (!INODE_JOURNAL(inode)) {
505        err = ocfs2_inode_lock(inode, NULL, 0);
506        if (err) {
507            if (err != -ENOENT)
508                mlog_errno(err);
509            goto bail;
510        }
511        down_read(&OCFS2_I(inode)->ip_alloc_sem);
512    }
513
514    if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL))
515        err = ocfs2_extent_map_get_blocks(inode, block, &p_blkno, NULL,
516                          NULL);
517
518    if (!INODE_JOURNAL(inode)) {
519        up_read(&OCFS2_I(inode)->ip_alloc_sem);
520        ocfs2_inode_unlock(inode, 0);
521    }
522
523    if (err) {
524        mlog(ML_ERROR, "get_blocks() failed, block = %llu\n",
525             (unsigned long long)block);
526        mlog_errno(err);
527        goto bail;
528    }
529
530bail:
531    status = err ? 0 : p_blkno;
532
533    mlog_exit((int)status);
534
535    return status;
536}
537
538/*
539 * TODO: Make this into a generic get_blocks function.
540 *
541 * From do_direct_io in direct-io.c:
542 * "So what we do is to permit the ->get_blocks function to populate
543 * bh.b_size with the size of IO which is permitted at this offset and
544 * this i_blkbits."
545 *
546 * This function is called directly from get_more_blocks in direct-io.c.
547 *
548 * called like this: dio->get_blocks(dio->inode, fs_startblk,
549 * fs_count, map_bh, dio->rw == WRITE);
550 *
551 * Note that we never bother to allocate blocks here, and thus ignore the
552 * create argument.
553 */
554static int ocfs2_direct_IO_get_blocks(struct inode *inode, sector_t iblock,
555                     struct buffer_head *bh_result, int create)
556{
557    int ret;
558    u64 p_blkno, inode_blocks, contig_blocks;
559    unsigned int ext_flags;
560    unsigned char blocksize_bits = inode->i_sb->s_blocksize_bits;
561    unsigned long max_blocks = bh_result->b_size >> inode->i_blkbits;
562
563    /* This function won't even be called if the request isn't all
564     * nicely aligned and of the right size, so there's no need
565     * for us to check any of that. */
566
567    inode_blocks = ocfs2_blocks_for_bytes(inode->i_sb, i_size_read(inode));
568
569    /* This figures out the size of the next contiguous block, and
570     * our logical offset */
571    ret = ocfs2_extent_map_get_blocks(inode, iblock, &p_blkno,
572                      &contig_blocks, &ext_flags);
573    if (ret) {
574        mlog(ML_ERROR, "get_blocks() failed iblock=%llu\n",
575             (unsigned long long)iblock);
576        ret = -EIO;
577        goto bail;
578    }
579
580    /* We should already CoW the refcounted extent in case of create. */
581    BUG_ON(create && (ext_flags & OCFS2_EXT_REFCOUNTED));
582
583    /*
584     * get_more_blocks() expects us to describe a hole by clearing
585     * the mapped bit on bh_result().
586     *
587     * Consider an unwritten extent as a hole.
588     */
589    if (p_blkno && !(ext_flags & OCFS2_EXT_UNWRITTEN))
590        map_bh(bh_result, inode->i_sb, p_blkno);
591    else
592        clear_buffer_mapped(bh_result);
593
594    /* make sure we don't map more than max_blocks blocks here as
595       that's all the kernel will handle at this point. */
596    if (max_blocks < contig_blocks)
597        contig_blocks = max_blocks;
598    bh_result->b_size = contig_blocks << blocksize_bits;
599bail:
600    return ret;
601}
602
603/*
604 * ocfs2_dio_end_io is called by the dio core when a dio is finished. We're
605 * particularly interested in the aio/dio case. Like the core uses
606 * i_alloc_sem, we use the rw_lock DLM lock to protect io on one node from
607 * truncation on another.
608 */
609static void ocfs2_dio_end_io(struct kiocb *iocb,
610                 loff_t offset,
611                 ssize_t bytes,
612                 void *private)
613{
614    struct inode *inode = iocb->ki_filp->f_path.dentry->d_inode;
615    int level;
616
617    /* this io's submitter should not have unlocked this before we could */
618    BUG_ON(!ocfs2_iocb_is_rw_locked(iocb));
619
620    ocfs2_iocb_clear_rw_locked(iocb);
621
622    level = ocfs2_iocb_rw_locked_level(iocb);
623    if (!level)
624        up_read(&inode->i_alloc_sem);
625    ocfs2_rw_unlock(inode, level);
626}
627
628/*
629 * ocfs2_invalidatepage() and ocfs2_releasepage() are shamelessly stolen
630 * from ext3. PageChecked() bits have been removed as OCFS2 does not
631 * do journalled data.
632 */
633static void ocfs2_invalidatepage(struct page *page, unsigned long offset)
634{
635    journal_t *journal = OCFS2_SB(page->mapping->host->i_sb)->journal->j_journal;
636
637    jbd2_journal_invalidatepage(journal, page, offset);
638}
639
640static int ocfs2_releasepage(struct page *page, gfp_t wait)
641{
642    journal_t *journal = OCFS2_SB(page->mapping->host->i_sb)->journal->j_journal;
643
644    if (!page_has_buffers(page))
645        return 0;
646    return jbd2_journal_try_to_free_buffers(journal, page, wait);
647}
648
649static ssize_t ocfs2_direct_IO(int rw,
650                   struct kiocb *iocb,
651                   const struct iovec *iov,
652                   loff_t offset,
653                   unsigned long nr_segs)
654{
655    struct file *file = iocb->ki_filp;
656    struct inode *inode = file->f_path.dentry->d_inode->i_mapping->host;
657    int ret;
658
659    mlog_entry_void();
660
661    /*
662     * Fallback to buffered I/O if we see an inode without
663     * extents.
664     */
665    if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
666        return 0;
667
668    /* Fallback to buffered I/O if we are appending. */
669    if (i_size_read(inode) <= offset)
670        return 0;
671
672    ret = blockdev_direct_IO_no_locking(rw, iocb, inode,
673                        inode->i_sb->s_bdev, iov, offset,
674                        nr_segs,
675                        ocfs2_direct_IO_get_blocks,
676                        ocfs2_dio_end_io);
677
678    mlog_exit(ret);
679    return ret;
680}
681
682static void ocfs2_figure_cluster_boundaries(struct ocfs2_super *osb,
683                        u32 cpos,
684                        unsigned int *start,
685                        unsigned int *end)
686{
687    unsigned int cluster_start = 0, cluster_end = PAGE_CACHE_SIZE;
688
689    if (unlikely(PAGE_CACHE_SHIFT > osb->s_clustersize_bits)) {
690        unsigned int cpp;
691
692        cpp = 1 << (PAGE_CACHE_SHIFT - osb->s_clustersize_bits);
693
694        cluster_start = cpos % cpp;
695        cluster_start = cluster_start << osb->s_clustersize_bits;
696
697        cluster_end = cluster_start + osb->s_clustersize;
698    }
699
700    BUG_ON(cluster_start > PAGE_SIZE);
701    BUG_ON(cluster_end > PAGE_SIZE);
702
703    if (start)
704        *start = cluster_start;
705    if (end)
706        *end = cluster_end;
707}
708
709/*
710 * 'from' and 'to' are the region in the page to avoid zeroing.
711 *
712 * If pagesize > clustersize, this function will avoid zeroing outside
713 * of the cluster boundary.
714 *
715 * from == to == 0 is code for "zero the entire cluster region"
716 */
717static void ocfs2_clear_page_regions(struct page *page,
718                     struct ocfs2_super *osb, u32 cpos,
719                     unsigned from, unsigned to)
720{
721    void *kaddr;
722    unsigned int cluster_start, cluster_end;
723
724    ocfs2_figure_cluster_boundaries(osb, cpos, &cluster_start, &cluster_end);
725
726    kaddr = kmap_atomic(page, KM_USER0);
727
728    if (from || to) {
729        if (from > cluster_start)
730            memset(kaddr + cluster_start, 0, from - cluster_start);
731        if (to < cluster_end)
732            memset(kaddr + to, 0, cluster_end - to);
733    } else {
734        memset(kaddr + cluster_start, 0, cluster_end - cluster_start);
735    }
736
737    kunmap_atomic(kaddr, KM_USER0);
738}
739
740/*
741 * Nonsparse file systems fully allocate before we get to the write
742 * code. This prevents ocfs2_write() from tagging the write as an
743 * allocating one, which means ocfs2_map_page_blocks() might try to
744 * read-in the blocks at the tail of our file. Avoid reading them by
745 * testing i_size against each block offset.
746 */
747static int ocfs2_should_read_blk(struct inode *inode, struct page *page,
748                 unsigned int block_start)
749{
750    u64 offset = page_offset(page) + block_start;
751
752    if (ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
753        return 1;
754
755    if (i_size_read(inode) > offset)
756        return 1;
757
758    return 0;
759}
760
761/*
762 * Some of this taken from block_prepare_write(). We already have our
763 * mapping by now though, and the entire write will be allocating or
764 * it won't, so not much need to use BH_New.
765 *
766 * This will also skip zeroing, which is handled externally.
767 */
768int ocfs2_map_page_blocks(struct page *page, u64 *p_blkno,
769              struct inode *inode, unsigned int from,
770              unsigned int to, int new)
771{
772    int ret = 0;
773    struct buffer_head *head, *bh, *wait[2], **wait_bh = wait;
774    unsigned int block_end, block_start;
775    unsigned int bsize = 1 << inode->i_blkbits;
776
777    if (!page_has_buffers(page))
778        create_empty_buffers(page, bsize, 0);
779
780    head = page_buffers(page);
781    for (bh = head, block_start = 0; bh != head || !block_start;
782         bh = bh->b_this_page, block_start += bsize) {
783        block_end = block_start + bsize;
784
785        clear_buffer_new(bh);
786
787        /*
788         * Ignore blocks outside of our i/o range -
789         * they may belong to unallocated clusters.
790         */
791        if (block_start >= to || block_end <= from) {
792            if (PageUptodate(page))
793                set_buffer_uptodate(bh);
794            continue;
795        }
796
797        /*
798         * For an allocating write with cluster size >= page
799         * size, we always write the entire page.
800         */
801        if (new)
802            set_buffer_new(bh);
803
804        if (!buffer_mapped(bh)) {
805            map_bh(bh, inode->i_sb, *p_blkno);
806            unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
807        }
808
809        if (PageUptodate(page)) {
810            if (!buffer_uptodate(bh))
811                set_buffer_uptodate(bh);
812        } else if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
813               !buffer_new(bh) &&
814               ocfs2_should_read_blk(inode, page, block_start) &&
815               (block_start < from || block_end > to)) {
816            ll_rw_block(READ, 1, &bh);
817            *wait_bh++=bh;
818        }
819
820        *p_blkno = *p_blkno + 1;
821    }
822
823    /*
824     * If we issued read requests - let them complete.
825     */
826    while(wait_bh > wait) {
827        wait_on_buffer(*--wait_bh);
828        if (!buffer_uptodate(*wait_bh))
829            ret = -EIO;
830    }
831
832    if (ret == 0 || !new)
833        return ret;
834
835    /*
836     * If we get -EIO above, zero out any newly allocated blocks
837     * to avoid exposing stale data.
838     */
839    bh = head;
840    block_start = 0;
841    do {
842        block_end = block_start + bsize;
843        if (block_end <= from)
844            goto next_bh;
845        if (block_start >= to)
846            break;
847
848        zero_user(page, block_start, bh->b_size);
849        set_buffer_uptodate(bh);
850        mark_buffer_dirty(bh);
851
852next_bh:
853        block_start = block_end;
854        bh = bh->b_this_page;
855    } while (bh != head);
856
857    return ret;
858}
859
860#if (PAGE_CACHE_SIZE >= OCFS2_MAX_CLUSTERSIZE)
861#define OCFS2_MAX_CTXT_PAGES 1
862#else
863#define OCFS2_MAX_CTXT_PAGES (OCFS2_MAX_CLUSTERSIZE / PAGE_CACHE_SIZE)
864#endif
865
866#define OCFS2_MAX_CLUSTERS_PER_PAGE (PAGE_CACHE_SIZE / OCFS2_MIN_CLUSTERSIZE)
867
868/*
869 * Describe the state of a single cluster to be written to.
870 */
871struct ocfs2_write_cluster_desc {
872    u32 c_cpos;
873    u32 c_phys;
874    /*
875     * Give this a unique field because c_phys eventually gets
876     * filled.
877     */
878    unsigned c_new;
879    unsigned c_unwritten;
880    unsigned c_needs_zero;
881};
882
883struct ocfs2_write_ctxt {
884    /* Logical cluster position / len of write */
885    u32 w_cpos;
886    u32 w_clen;
887
888    /* First cluster allocated in a nonsparse extend */
889    u32 w_first_new_cpos;
890
891    struct ocfs2_write_cluster_desc w_desc[OCFS2_MAX_CLUSTERS_PER_PAGE];
892
893    /*
894     * This is true if page_size > cluster_size.
895     *
896     * It triggers a set of special cases during write which might
897     * have to deal with allocating writes to partial pages.
898     */
899    unsigned int w_large_pages;
900
901    /*
902     * Pages involved in this write.
903     *
904     * w_target_page is the page being written to by the user.
905     *
906     * w_pages is an array of pages which always contains
907     * w_target_page, and in the case of an allocating write with
908     * page_size < cluster size, it will contain zero'd and mapped
909     * pages adjacent to w_target_page which need to be written
910     * out in so that future reads from that region will get
911     * zero's.
912     */
913    struct page *w_pages[OCFS2_MAX_CTXT_PAGES];
914    unsigned int w_num_pages;
915    struct page *w_target_page;
916
917    /*
918     * ocfs2_write_end() uses this to know what the real range to
919     * write in the target should be.
920     */
921    unsigned int w_target_from;
922    unsigned int w_target_to;
923
924    /*
925     * We could use journal_current_handle() but this is cleaner,
926     * IMHO -Mark
927     */
928    handle_t *w_handle;
929
930    struct buffer_head *w_di_bh;
931
932    struct ocfs2_cached_dealloc_ctxt w_dealloc;
933};
934
935void ocfs2_unlock_and_free_pages(struct page **pages, int num_pages)
936{
937    int i;
938
939    for(i = 0; i < num_pages; i++) {
940        if (pages[i]) {
941            unlock_page(pages[i]);
942            mark_page_accessed(pages[i]);
943            page_cache_release(pages[i]);
944        }
945    }
946}
947
948static void ocfs2_free_write_ctxt(struct ocfs2_write_ctxt *wc)
949{
950    ocfs2_unlock_and_free_pages(wc->w_pages, wc->w_num_pages);
951
952    brelse(wc->w_di_bh);
953    kfree(wc);
954}
955
956static int ocfs2_alloc_write_ctxt(struct ocfs2_write_ctxt **wcp,
957                  struct ocfs2_super *osb, loff_t pos,
958                  unsigned len, struct buffer_head *di_bh)
959{
960    u32 cend;
961    struct ocfs2_write_ctxt *wc;
962
963    wc = kzalloc(sizeof(struct ocfs2_write_ctxt), GFP_NOFS);
964    if (!wc)
965        return -ENOMEM;
966
967    wc->w_cpos = pos >> osb->s_clustersize_bits;
968    wc->w_first_new_cpos = UINT_MAX;
969    cend = (pos + len - 1) >> osb->s_clustersize_bits;
970    wc->w_clen = cend - wc->w_cpos + 1;
971    get_bh(di_bh);
972    wc->w_di_bh = di_bh;
973
974    if (unlikely(PAGE_CACHE_SHIFT > osb->s_clustersize_bits))
975        wc->w_large_pages = 1;
976    else
977        wc->w_large_pages = 0;
978
979    ocfs2_init_dealloc_ctxt(&wc->w_dealloc);
980
981    *wcp = wc;
982
983    return 0;
984}
985
986/*
987 * If a page has any new buffers, zero them out here, and mark them uptodate
988 * and dirty so they'll be written out (in order to prevent uninitialised
989 * block data from leaking). And clear the new bit.
990 */
991static void ocfs2_zero_new_buffers(struct page *page, unsigned from, unsigned to)
992{
993    unsigned int block_start, block_end;
994    struct buffer_head *head, *bh;
995
996    BUG_ON(!PageLocked(page));
997    if (!page_has_buffers(page))
998        return;
999
1000    bh = head = page_buffers(page);
1001    block_start = 0;
1002    do {
1003        block_end = block_start + bh->b_size;
1004
1005        if (buffer_new(bh)) {
1006            if (block_end > from && block_start < to) {
1007                if (!PageUptodate(page)) {
1008                    unsigned start, end;
1009
1010                    start = max(from, block_start);
1011                    end = min(to, block_end);
1012
1013                    zero_user_segment(page, start, end);
1014                    set_buffer_uptodate(bh);
1015                }
1016
1017                clear_buffer_new(bh);
1018                mark_buffer_dirty(bh);
1019            }
1020        }
1021
1022        block_start = block_end;
1023        bh = bh->b_this_page;
1024    } while (bh != head);
1025}
1026
1027/*
1028 * Only called when we have a failure during allocating write to write
1029 * zero's to the newly allocated region.
1030 */
1031static void ocfs2_write_failure(struct inode *inode,
1032                struct ocfs2_write_ctxt *wc,
1033                loff_t user_pos, unsigned user_len)
1034{
1035    int i;
1036    unsigned from = user_pos & (PAGE_CACHE_SIZE - 1),
1037        to = user_pos + user_len;
1038    struct page *tmppage;
1039
1040    ocfs2_zero_new_buffers(wc->w_target_page, from, to);
1041
1042    for(i = 0; i < wc->w_num_pages; i++) {
1043        tmppage = wc->w_pages[i];
1044
1045        if (page_has_buffers(tmppage)) {
1046            if (ocfs2_should_order_data(inode))
1047                ocfs2_jbd2_file_inode(wc->w_handle, inode);
1048
1049            block_commit_write(tmppage, from, to);
1050        }
1051    }
1052}
1053
1054static int ocfs2_prepare_page_for_write(struct inode *inode, u64 *p_blkno,
1055                    struct ocfs2_write_ctxt *wc,
1056                    struct page *page, u32 cpos,
1057                    loff_t user_pos, unsigned user_len,
1058                    int new)
1059{
1060    int ret;
1061    unsigned int map_from = 0, map_to = 0;
1062    unsigned int cluster_start, cluster_end;
1063    unsigned int user_data_from = 0, user_data_to = 0;
1064
1065    ocfs2_figure_cluster_boundaries(OCFS2_SB(inode->i_sb), cpos,
1066                    &cluster_start, &cluster_end);
1067
1068    if (page == wc->w_target_page) {
1069        map_from = user_pos & (PAGE_CACHE_SIZE - 1);
1070        map_to = map_from + user_len;
1071
1072        if (new)
1073            ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1074                            cluster_start, cluster_end,
1075                            new);
1076        else
1077            ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1078                            map_from, map_to, new);
1079        if (ret) {
1080            mlog_errno(ret);
1081            goto out;
1082        }
1083
1084        user_data_from = map_from;
1085        user_data_to = map_to;
1086        if (new) {
1087            map_from = cluster_start;
1088            map_to = cluster_end;
1089        }
1090    } else {
1091        /*
1092         * If we haven't allocated the new page yet, we
1093         * shouldn't be writing it out without copying user
1094         * data. This is likely a math error from the caller.
1095         */
1096        BUG_ON(!new);
1097
1098        map_from = cluster_start;
1099        map_to = cluster_end;
1100
1101        ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1102                        cluster_start, cluster_end, new);
1103        if (ret) {
1104            mlog_errno(ret);
1105            goto out;
1106        }
1107    }
1108
1109    /*
1110     * Parts of newly allocated pages need to be zero'd.
1111     *
1112     * Above, we have also rewritten 'to' and 'from' - as far as
1113     * the rest of the function is concerned, the entire cluster
1114     * range inside of a page needs to be written.
1115     *
1116     * We can skip this if the page is up to date - it's already
1117     * been zero'd from being read in as a hole.
1118     */
1119    if (new && !PageUptodate(page))
1120        ocfs2_clear_page_regions(page, OCFS2_SB(inode->i_sb),
1121                     cpos, user_data_from, user_data_to);
1122
1123    flush_dcache_page(page);
1124
1125out:
1126    return ret;
1127}
1128
1129/*
1130 * This function will only grab one clusters worth of pages.
1131 */
1132static int ocfs2_grab_pages_for_write(struct address_space *mapping,
1133                      struct ocfs2_write_ctxt *wc,
1134                      u32 cpos, loff_t user_pos, int new,
1135                      struct page *mmap_page)
1136{
1137    int ret = 0, i;
1138    unsigned long start, target_index, index;
1139    struct inode *inode = mapping->host;
1140
1141    target_index = user_pos >> PAGE_CACHE_SHIFT;
1142
1143    /*
1144     * Figure out how many pages we'll be manipulating here. For
1145     * non allocating write, we just change the one
1146     * page. Otherwise, we'll need a whole clusters worth.
1147     */
1148    if (new) {
1149        wc->w_num_pages = ocfs2_pages_per_cluster(inode->i_sb);
1150        start = ocfs2_align_clusters_to_page_index(inode->i_sb, cpos);
1151    } else {
1152        wc->w_num_pages = 1;
1153        start = target_index;
1154    }
1155
1156    for(i = 0; i < wc->w_num_pages; i++) {
1157        index = start + i;
1158
1159        if (index == target_index && mmap_page) {
1160            /*
1161             * ocfs2_pagemkwrite() is a little different
1162             * and wants us to directly use the page
1163             * passed in.
1164             */
1165            lock_page(mmap_page);
1166
1167            if (mmap_page->mapping != mapping) {
1168                unlock_page(mmap_page);
1169                /*
1170                 * Sanity check - the locking in
1171                 * ocfs2_pagemkwrite() should ensure
1172                 * that this code doesn't trigger.
1173                 */
1174                ret = -EINVAL;
1175                mlog_errno(ret);
1176                goto out;
1177            }
1178
1179            page_cache_get(mmap_page);
1180            wc->w_pages[i] = mmap_page;
1181        } else {
1182            wc->w_pages[i] = find_or_create_page(mapping, index,
1183                                 GFP_NOFS);
1184            if (!wc->w_pages[i]) {
1185                ret = -ENOMEM;
1186                mlog_errno(ret);
1187                goto out;
1188            }
1189        }
1190
1191        if (index == target_index)
1192            wc->w_target_page = wc->w_pages[i];
1193    }
1194out:
1195    return ret;
1196}
1197
1198/*
1199 * Prepare a single cluster for write one cluster into the file.
1200 */
1201static int ocfs2_write_cluster(struct address_space *mapping,
1202                   u32 phys, unsigned int unwritten,
1203                   unsigned int should_zero,
1204                   struct ocfs2_alloc_context *data_ac,
1205                   struct ocfs2_alloc_context *meta_ac,
1206                   struct ocfs2_write_ctxt *wc, u32 cpos,
1207                   loff_t user_pos, unsigned user_len)
1208{
1209    int ret, i, new;
1210    u64 v_blkno, p_blkno;
1211    struct inode *inode = mapping->host;
1212    struct ocfs2_extent_tree et;
1213
1214    new = phys == 0 ? 1 : 0;
1215    if (new) {
1216        u32 tmp_pos;
1217
1218        /*
1219         * This is safe to call with the page locks - it won't take
1220         * any additional semaphores or cluster locks.
1221         */
1222        tmp_pos = cpos;
1223        ret = ocfs2_add_inode_data(OCFS2_SB(inode->i_sb), inode,
1224                       &tmp_pos, 1, 0, wc->w_di_bh,
1225                       wc->w_handle, data_ac,
1226                       meta_ac, NULL);
1227        /*
1228         * This shouldn't happen because we must have already
1229         * calculated the correct meta data allocation required. The
1230         * internal tree allocation code should know how to increase
1231         * transaction credits itself.
1232         *
1233         * If need be, we could handle -EAGAIN for a
1234         * RESTART_TRANS here.
1235         */
1236        mlog_bug_on_msg(ret == -EAGAIN,
1237                "Inode %llu: EAGAIN return during allocation.\n",
1238                (unsigned long long)OCFS2_I(inode)->ip_blkno);
1239        if (ret < 0) {
1240            mlog_errno(ret);
1241            goto out;
1242        }
1243    } else if (unwritten) {
1244        ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode),
1245                          wc->w_di_bh);
1246        ret = ocfs2_mark_extent_written(inode, &et,
1247                        wc->w_handle, cpos, 1, phys,
1248                        meta_ac, &wc->w_dealloc);
1249        if (ret < 0) {
1250            mlog_errno(ret);
1251            goto out;
1252        }
1253    }
1254
1255    if (should_zero)
1256        v_blkno = ocfs2_clusters_to_blocks(inode->i_sb, cpos);
1257    else
1258        v_blkno = user_pos >> inode->i_sb->s_blocksize_bits;
1259
1260    /*
1261     * The only reason this should fail is due to an inability to
1262     * find the extent added.
1263     */
1264    ret = ocfs2_extent_map_get_blocks(inode, v_blkno, &p_blkno, NULL,
1265                      NULL);
1266    if (ret < 0) {
1267        ocfs2_error(inode->i_sb, "Corrupting extend for inode %llu, "
1268                "at logical block %llu",
1269                (unsigned long long)OCFS2_I(inode)->ip_blkno,
1270                (unsigned long long)v_blkno);
1271        goto out;
1272    }
1273
1274    BUG_ON(p_blkno == 0);
1275
1276    for(i = 0; i < wc->w_num_pages; i++) {
1277        int tmpret;
1278
1279        tmpret = ocfs2_prepare_page_for_write(inode, &p_blkno, wc,
1280                              wc->w_pages[i], cpos,
1281                              user_pos, user_len,
1282                              should_zero);
1283        if (tmpret) {
1284            mlog_errno(tmpret);
1285            if (ret == 0)
1286                ret = tmpret;
1287        }
1288    }
1289
1290    /*
1291     * We only have cleanup to do in case of allocating write.
1292     */
1293    if (ret && new)
1294        ocfs2_write_failure(inode, wc, user_pos, user_len);
1295
1296out:
1297
1298    return ret;
1299}
1300
1301static int ocfs2_write_cluster_by_desc(struct address_space *mapping,
1302                       struct ocfs2_alloc_context *data_ac,
1303                       struct ocfs2_alloc_context *meta_ac,
1304                       struct ocfs2_write_ctxt *wc,
1305                       loff_t pos, unsigned len)
1306{
1307    int ret, i;
1308    loff_t cluster_off;
1309    unsigned int local_len = len;
1310    struct ocfs2_write_cluster_desc *desc;
1311    struct ocfs2_super *osb = OCFS2_SB(mapping->host->i_sb);
1312
1313    for (i = 0; i < wc->w_clen; i++) {
1314        desc = &wc->w_desc[i];
1315
1316        /*
1317         * We have to make sure that the total write passed in
1318         * doesn't extend past a single cluster.
1319         */
1320        local_len = len;
1321        cluster_off = pos & (osb->s_clustersize - 1);
1322        if ((cluster_off + local_len) > osb->s_clustersize)
1323            local_len = osb->s_clustersize - cluster_off;
1324
1325        ret = ocfs2_write_cluster(mapping, desc->c_phys,
1326                      desc->c_unwritten,
1327                      desc->c_needs_zero,
1328                      data_ac, meta_ac,
1329                      wc, desc->c_cpos, pos, local_len);
1330        if (ret) {
1331            mlog_errno(ret);
1332            goto out;
1333        }
1334
1335        len -= local_len;
1336        pos += local_len;
1337    }
1338
1339    ret = 0;
1340out:
1341    return ret;
1342}
1343
1344/*
1345 * ocfs2_write_end() wants to know which parts of the target page it
1346 * should complete the write on. It's easiest to compute them ahead of
1347 * time when a more complete view of the write is available.
1348 */
1349static void ocfs2_set_target_boundaries(struct ocfs2_super *osb,
1350                    struct ocfs2_write_ctxt *wc,
1351                    loff_t pos, unsigned len, int alloc)
1352{
1353    struct ocfs2_write_cluster_desc *desc;
1354
1355    wc->w_target_from = pos & (PAGE_CACHE_SIZE - 1);
1356    wc->w_target_to = wc->w_target_from + len;
1357
1358    if (alloc == 0)
1359        return;
1360
1361    /*
1362     * Allocating write - we may have different boundaries based
1363     * on page size and cluster size.
1364     *
1365     * NOTE: We can no longer compute one value from the other as
1366     * the actual write length and user provided length may be
1367     * different.
1368     */
1369
1370    if (wc->w_large_pages) {
1371        /*
1372         * We only care about the 1st and last cluster within
1373         * our range and whether they should be zero'd or not. Either
1374         * value may be extended out to the start/end of a
1375         * newly allocated cluster.
1376         */
1377        desc = &wc->w_desc[0];
1378        if (desc->c_needs_zero)
1379            ocfs2_figure_cluster_boundaries(osb,
1380                            desc->c_cpos,
1381                            &wc->w_target_from,
1382                            NULL);
1383
1384        desc = &wc->w_desc[wc->w_clen - 1];
1385        if (desc->c_needs_zero)
1386            ocfs2_figure_cluster_boundaries(osb,
1387                            desc->c_cpos,
1388                            NULL,
1389                            &wc->w_target_to);
1390    } else {
1391        wc->w_target_from = 0;
1392        wc->w_target_to = PAGE_CACHE_SIZE;
1393    }
1394}
1395
1396/*
1397 * Populate each single-cluster write descriptor in the write context
1398 * with information about the i/o to be done.
1399 *
1400 * Returns the number of clusters that will have to be allocated, as
1401 * well as a worst case estimate of the number of extent records that
1402 * would have to be created during a write to an unwritten region.
1403 */
1404static int ocfs2_populate_write_desc(struct inode *inode,
1405                     struct ocfs2_write_ctxt *wc,
1406                     unsigned int *clusters_to_alloc,
1407                     unsigned int *extents_to_split)
1408{
1409    int ret;
1410    struct ocfs2_write_cluster_desc *desc;
1411    unsigned int num_clusters = 0;
1412    unsigned int ext_flags = 0;
1413    u32 phys = 0;
1414    int i;
1415
1416    *clusters_to_alloc = 0;
1417    *extents_to_split = 0;
1418
1419    for (i = 0; i < wc->w_clen; i++) {
1420        desc = &wc->w_desc[i];
1421        desc->c_cpos = wc->w_cpos + i;
1422
1423        if (num_clusters == 0) {
1424            /*
1425             * Need to look up the next extent record.
1426             */
1427            ret = ocfs2_get_clusters(inode, desc->c_cpos, &phys,
1428                         &num_clusters, &ext_flags);
1429            if (ret) {
1430                mlog_errno(ret);
1431                goto out;
1432            }
1433
1434            /* We should already CoW the refcountd extent. */
1435            BUG_ON(ext_flags & OCFS2_EXT_REFCOUNTED);
1436
1437            /*
1438             * Assume worst case - that we're writing in
1439             * the middle of the extent.
1440             *
1441             * We can assume that the write proceeds from
1442             * left to right, in which case the extent
1443             * insert code is smart enough to coalesce the
1444             * next splits into the previous records created.
1445             */
1446            if (ext_flags & OCFS2_EXT_UNWRITTEN)
1447                *extents_to_split = *extents_to_split + 2;
1448        } else if (phys) {
1449            /*
1450             * Only increment phys if it doesn't describe
1451             * a hole.
1452             */
1453            phys++;
1454        }
1455
1456        /*
1457         * If w_first_new_cpos is < UINT_MAX, we have a non-sparse
1458         * file that got extended. w_first_new_cpos tells us
1459         * where the newly allocated clusters are so we can
1460         * zero them.
1461         */
1462        if (desc->c_cpos >= wc->w_first_new_cpos) {
1463            BUG_ON(phys == 0);
1464            desc->c_needs_zero = 1;
1465        }
1466
1467        desc->c_phys = phys;
1468        if (phys == 0) {
1469            desc->c_new = 1;
1470            desc->c_needs_zero = 1;
1471            *clusters_to_alloc = *clusters_to_alloc + 1;
1472        }
1473
1474        if (ext_flags & OCFS2_EXT_UNWRITTEN) {
1475            desc->c_unwritten = 1;
1476            desc->c_needs_zero = 1;
1477        }
1478
1479        num_clusters--;
1480    }
1481
1482    ret = 0;
1483out:
1484    return ret;
1485}
1486
1487static int ocfs2_write_begin_inline(struct address_space *mapping,
1488                    struct inode *inode,
1489                    struct ocfs2_write_ctxt *wc)
1490{
1491    int ret;
1492    struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1493    struct page *page;
1494    handle_t *handle;
1495    struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1496
1497    page = find_or_create_page(mapping, 0, GFP_NOFS);
1498    if (!page) {
1499        ret = -ENOMEM;
1500        mlog_errno(ret);
1501        goto out;
1502    }
1503    /*
1504     * If we don't set w_num_pages then this page won't get unlocked
1505     * and freed on cleanup of the write context.
1506     */
1507    wc->w_pages[0] = wc->w_target_page = page;
1508    wc->w_num_pages = 1;
1509
1510    handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1511    if (IS_ERR(handle)) {
1512        ret = PTR_ERR(handle);
1513        mlog_errno(ret);
1514        goto out;
1515    }
1516
1517    ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), wc->w_di_bh,
1518                      OCFS2_JOURNAL_ACCESS_WRITE);
1519    if (ret) {
1520        ocfs2_commit_trans(osb, handle);
1521
1522        mlog_errno(ret);
1523        goto out;
1524    }
1525
1526    if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL))
1527        ocfs2_set_inode_data_inline(inode, di);
1528
1529    if (!PageUptodate(page)) {
1530        ret = ocfs2_read_inline_data(inode, page, wc->w_di_bh);
1531        if (ret) {
1532            ocfs2_commit_trans(osb, handle);
1533
1534            goto out;
1535        }
1536    }
1537
1538    wc->w_handle = handle;
1539out:
1540    return ret;
1541}
1542
1543int ocfs2_size_fits_inline_data(struct buffer_head *di_bh, u64 new_size)
1544{
1545    struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
1546
1547    if (new_size <= le16_to_cpu(di->id2.i_data.id_count))
1548        return 1;
1549    return 0;
1550}
1551
1552static int ocfs2_try_to_write_inline_data(struct address_space *mapping,
1553                      struct inode *inode, loff_t pos,
1554                      unsigned len, struct page *mmap_page,
1555                      struct ocfs2_write_ctxt *wc)
1556{
1557    int ret, written = 0;
1558    loff_t end = pos + len;
1559    struct ocfs2_inode_info *oi = OCFS2_I(inode);
1560    struct ocfs2_dinode *di = NULL;
1561
1562    mlog(0, "Inode %llu, write of %u bytes at off %llu. features: 0x%x\n",
1563         (unsigned long long)oi->ip_blkno, len, (unsigned long long)pos,
1564         oi->ip_dyn_features);
1565
1566    /*
1567     * Handle inodes which already have inline data 1st.
1568     */
1569    if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1570        if (mmap_page == NULL &&
1571            ocfs2_size_fits_inline_data(wc->w_di_bh, end))
1572            goto do_inline_write;
1573
1574        /*
1575         * The write won't fit - we have to give this inode an
1576         * inline extent list now.
1577         */
1578        ret = ocfs2_convert_inline_data_to_extents(inode, wc->w_di_bh);
1579        if (ret)
1580            mlog_errno(ret);
1581        goto out;
1582    }
1583
1584    /*
1585     * Check whether the inode can accept inline data.
1586     */
1587    if (oi->ip_clusters != 0 || i_size_read(inode) != 0)
1588        return 0;
1589
1590    /*
1591     * Check whether the write can fit.
1592     */
1593    di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1594    if (mmap_page ||
1595        end > ocfs2_max_inline_data_with_xattr(inode->i_sb, di))
1596        return 0;
1597
1598do_inline_write:
1599    ret = ocfs2_write_begin_inline(mapping, inode, wc);
1600    if (ret) {
1601        mlog_errno(ret);
1602        goto out;
1603    }
1604
1605    /*
1606     * This signals to the caller that the data can be written
1607     * inline.
1608     */
1609    written = 1;
1610out:
1611    return written ? written : ret;
1612}
1613
1614/*
1615 * This function only does anything for file systems which can't
1616 * handle sparse files.
1617 *
1618 * What we want to do here is fill in any hole between the current end
1619 * of allocation and the end of our write. That way the rest of the
1620 * write path can treat it as an non-allocating write, which has no
1621 * special case code for sparse/nonsparse files.
1622 */
1623static int ocfs2_expand_nonsparse_inode(struct inode *inode, loff_t pos,
1624                    unsigned len,
1625                    struct ocfs2_write_ctxt *wc)
1626{
1627    int ret;
1628    struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1629    loff_t newsize = pos + len;
1630
1631    if (ocfs2_sparse_alloc(osb))
1632        return 0;
1633
1634    if (newsize <= i_size_read(inode))
1635        return 0;
1636
1637    ret = ocfs2_extend_no_holes(inode, newsize, pos);
1638    if (ret)
1639        mlog_errno(ret);
1640
1641    wc->w_first_new_cpos =
1642        ocfs2_clusters_for_bytes(inode->i_sb, i_size_read(inode));
1643
1644    return ret;
1645}
1646
1647int ocfs2_write_begin_nolock(struct address_space *mapping,
1648                 loff_t pos, unsigned len, unsigned flags,
1649                 struct page **pagep, void **fsdata,
1650                 struct buffer_head *di_bh, struct page *mmap_page)
1651{
1652    int ret, cluster_of_pages, credits = OCFS2_INODE_UPDATE_CREDITS;
1653    unsigned int clusters_to_alloc, extents_to_split;
1654    struct ocfs2_write_ctxt *wc;
1655    struct inode *inode = mapping->host;
1656    struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1657    struct ocfs2_dinode *di;
1658    struct ocfs2_alloc_context *data_ac = NULL;
1659    struct ocfs2_alloc_context *meta_ac = NULL;
1660    handle_t *handle;
1661    struct ocfs2_extent_tree et;
1662
1663    ret = ocfs2_alloc_write_ctxt(&wc, osb, pos, len, di_bh);
1664    if (ret) {
1665        mlog_errno(ret);
1666        return ret;
1667    }
1668
1669    if (ocfs2_supports_inline_data(osb)) {
1670        ret = ocfs2_try_to_write_inline_data(mapping, inode, pos, len,
1671                             mmap_page, wc);
1672        if (ret == 1) {
1673            ret = 0;
1674            goto success;
1675        }
1676        if (ret < 0) {
1677            mlog_errno(ret);
1678            goto out;
1679        }
1680    }
1681
1682    ret = ocfs2_expand_nonsparse_inode(inode, pos, len, wc);
1683    if (ret) {
1684        mlog_errno(ret);
1685        goto out;
1686    }
1687
1688    ret = ocfs2_check_range_for_refcount(inode, pos, len);
1689    if (ret < 0) {
1690        mlog_errno(ret);
1691        goto out;
1692    } else if (ret == 1) {
1693        ret = ocfs2_refcount_cow(inode, di_bh,
1694                     wc->w_cpos, wc->w_clen, UINT_MAX);
1695        if (ret) {
1696            mlog_errno(ret);
1697            goto out;
1698        }
1699    }
1700
1701    ret = ocfs2_populate_write_desc(inode, wc, &clusters_to_alloc,
1702                    &extents_to_split);
1703    if (ret) {
1704        mlog_errno(ret);
1705        goto out;
1706    }
1707
1708    di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1709
1710    /*
1711     * We set w_target_from, w_target_to here so that
1712     * ocfs2_write_end() knows which range in the target page to
1713     * write out. An allocation requires that we write the entire
1714     * cluster range.
1715     */
1716    if (clusters_to_alloc || extents_to_split) {
1717        /*
1718         * XXX: We are stretching the limits of
1719         * ocfs2_lock_allocators(). It greatly over-estimates
1720         * the work to be done.
1721         */
1722        mlog(0, "extend inode %llu, i_size = %lld, di->i_clusters = %u,"
1723             " clusters_to_add = %u, extents_to_split = %u\n",
1724             (unsigned long long)OCFS2_I(inode)->ip_blkno,
1725             (long long)i_size_read(inode), le32_to_cpu(di->i_clusters),
1726             clusters_to_alloc, extents_to_split);
1727
1728        ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode),
1729                          wc->w_di_bh);
1730        ret = ocfs2_lock_allocators(inode, &et,
1731                        clusters_to_alloc, extents_to_split,
1732                        &data_ac, &meta_ac);
1733        if (ret) {
1734            mlog_errno(ret);
1735            goto out;
1736        }
1737
1738        credits = ocfs2_calc_extend_credits(inode->i_sb,
1739                            &di->id2.i_list,
1740                            clusters_to_alloc);
1741
1742    }
1743
1744    /*
1745     * We have to zero sparse allocated clusters, unwritten extent clusters,
1746     * and non-sparse clusters we just extended. For non-sparse writes,
1747     * we know zeros will only be needed in the first and/or last cluster.
1748     */
1749    if (clusters_to_alloc || extents_to_split ||
1750        (wc->w_clen && (wc->w_desc[0].c_needs_zero ||
1751                wc->w_desc[wc->w_clen - 1].c_needs_zero)))
1752        cluster_of_pages = 1;
1753    else
1754        cluster_of_pages = 0;
1755
1756    ocfs2_set_target_boundaries(osb, wc, pos, len, cluster_of_pages);
1757
1758    handle = ocfs2_start_trans(osb, credits);
1759    if (IS_ERR(handle)) {
1760        ret = PTR_ERR(handle);
1761        mlog_errno(ret);
1762        goto out;
1763    }
1764
1765    wc->w_handle = handle;
1766
1767    if (clusters_to_alloc) {
1768        ret = dquot_alloc_space_nodirty(inode,
1769            ocfs2_clusters_to_bytes(osb->sb, clusters_to_alloc));
1770        if (ret)
1771            goto out_commit;
1772    }
1773    /*
1774     * We don't want this to fail in ocfs2_write_end(), so do it
1775     * here.
1776     */
1777    ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), wc->w_di_bh,
1778                      OCFS2_JOURNAL_ACCESS_WRITE);
1779    if (ret) {
1780        mlog_errno(ret);
1781        goto out_quota;
1782    }
1783
1784    /*
1785     * Fill our page array first. That way we've grabbed enough so
1786     * that we can zero and flush if we error after adding the
1787     * extent.
1788     */
1789    ret = ocfs2_grab_pages_for_write(mapping, wc, wc->w_cpos, pos,
1790                     cluster_of_pages, mmap_page);
1791    if (ret) {
1792        mlog_errno(ret);
1793        goto out_quota;
1794    }
1795
1796    ret = ocfs2_write_cluster_by_desc(mapping, data_ac, meta_ac, wc, pos,
1797                      len);
1798    if (ret) {
1799        mlog_errno(ret);
1800        goto out_quota;
1801    }
1802
1803    if (data_ac)
1804        ocfs2_free_alloc_context(data_ac);
1805    if (meta_ac)
1806        ocfs2_free_alloc_context(meta_ac);
1807
1808success:
1809    *pagep = wc->w_target_page;
1810    *fsdata = wc;
1811    return 0;
1812out_quota:
1813    if (clusters_to_alloc)
1814        dquot_free_space(inode,
1815              ocfs2_clusters_to_bytes(osb->sb, clusters_to_alloc));
1816out_commit:
1817    ocfs2_commit_trans(osb, handle);
1818
1819out:
1820    ocfs2_free_write_ctxt(wc);
1821
1822    if (data_ac)
1823        ocfs2_free_alloc_context(data_ac);
1824    if (meta_ac)
1825        ocfs2_free_alloc_context(meta_ac);
1826    return ret;
1827}
1828
1829static int ocfs2_write_begin(struct file *file, struct address_space *mapping,
1830                 loff_t pos, unsigned len, unsigned flags,
1831                 struct page **pagep, void **fsdata)
1832{
1833    int ret;
1834    struct buffer_head *di_bh = NULL;
1835    struct inode *inode = mapping->host;
1836
1837    ret = ocfs2_inode_lock(inode, &di_bh, 1);
1838    if (ret) {
1839        mlog_errno(ret);
1840        return ret;
1841    }
1842
1843    /*
1844     * Take alloc sem here to prevent concurrent lookups. That way
1845     * the mapping, zeroing and tree manipulation within
1846     * ocfs2_write() will be safe against ->readpage(). This
1847     * should also serve to lock out allocation from a shared
1848     * writeable region.
1849     */
1850    down_write(&OCFS2_I(inode)->ip_alloc_sem);
1851
1852    ret = ocfs2_write_begin_nolock(mapping, pos, len, flags, pagep,
1853                       fsdata, di_bh, NULL);
1854    if (ret) {
1855        mlog_errno(ret);
1856        goto out_fail;
1857    }
1858
1859    brelse(di_bh);
1860
1861    return 0;
1862
1863out_fail:
1864    up_write(&OCFS2_I(inode)->ip_alloc_sem);
1865
1866    brelse(di_bh);
1867    ocfs2_inode_unlock(inode, 1);
1868
1869    return ret;
1870}
1871
1872static void ocfs2_write_end_inline(struct inode *inode, loff_t pos,
1873                   unsigned len, unsigned *copied,
1874                   struct ocfs2_dinode *di,
1875                   struct ocfs2_write_ctxt *wc)
1876{
1877    void *kaddr;
1878
1879    if (unlikely(*copied < len)) {
1880        if (!PageUptodate(wc->w_target_page)) {
1881            *copied = 0;
1882            return;
1883        }
1884    }
1885
1886    kaddr = kmap_atomic(wc->w_target_page, KM_USER0);
1887    memcpy(di->id2.i_data.id_data + pos, kaddr + pos, *copied);
1888    kunmap_atomic(kaddr, KM_USER0);
1889
1890    mlog(0, "Data written to inode at offset %llu. "
1891         "id_count = %u, copied = %u, i_dyn_features = 0x%x\n",
1892         (unsigned long long)pos, *copied,
1893         le16_to_cpu(di->id2.i_data.id_count),
1894         le16_to_cpu(di->i_dyn_features));
1895}
1896
1897int ocfs2_write_end_nolock(struct address_space *mapping,
1898               loff_t pos, unsigned len, unsigned copied,
1899               struct page *page, void *fsdata)
1900{
1901    int i;
1902    unsigned from, to, start = pos & (PAGE_CACHE_SIZE - 1);
1903    struct inode *inode = mapping->host;
1904    struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1905    struct ocfs2_write_ctxt *wc = fsdata;
1906    struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1907    handle_t *handle = wc->w_handle;
1908    struct page *tmppage;
1909
1910    if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1911        ocfs2_write_end_inline(inode, pos, len, &copied, di, wc);
1912        goto out_write_size;
1913    }
1914
1915    if (unlikely(copied < len)) {
1916        if (!PageUptodate(wc->w_target_page))
1917            copied = 0;
1918
1919        ocfs2_zero_new_buffers(wc->w_target_page, start+copied,
1920                       start+len);
1921    }
1922    flush_dcache_page(wc->w_target_page);
1923
1924    for(i = 0; i < wc->w_num_pages; i++) {
1925        tmppage = wc->w_pages[i];
1926
1927        if (tmppage == wc->w_target_page) {
1928            from = wc->w_target_from;
1929            to = wc->w_target_to;
1930
1931            BUG_ON(from > PAGE_CACHE_SIZE ||
1932                   to > PAGE_CACHE_SIZE ||
1933                   to < from);
1934        } else {
1935            /*
1936             * Pages adjacent to the target (if any) imply
1937             * a hole-filling write in which case we want
1938             * to flush their entire range.
1939             */
1940            from = 0;
1941            to = PAGE_CACHE_SIZE;
1942        }
1943
1944        if (page_has_buffers(tmppage)) {
1945            if (ocfs2_should_order_data(inode))
1946                ocfs2_jbd2_file_inode(wc->w_handle, inode);
1947            block_commit_write(tmppage, from, to);
1948        }
1949    }
1950
1951out_write_size:
1952    pos += copied;
1953    if (pos > inode->i_size) {
1954        i_size_write(inode, pos);
1955        mark_inode_dirty(inode);
1956    }
1957    inode->i_blocks = ocfs2_inode_sector_count(inode);
1958    di->i_size = cpu_to_le64((u64)i_size_read(inode));
1959    inode->i_mtime = inode->i_ctime = CURRENT_TIME;
1960    di->i_mtime = di->i_ctime = cpu_to_le64(inode->i_mtime.tv_sec);
1961    di->i_mtime_nsec = di->i_ctime_nsec = cpu_to_le32(inode->i_mtime.tv_nsec);
1962    ocfs2_journal_dirty(handle, wc->w_di_bh);
1963
1964    ocfs2_commit_trans(osb, handle);
1965
1966    ocfs2_run_deallocs(osb, &wc->w_dealloc);
1967
1968    ocfs2_free_write_ctxt(wc);
1969
1970    return copied;
1971}
1972
1973static int ocfs2_write_end(struct file *file, struct address_space *mapping,
1974               loff_t pos, unsigned len, unsigned copied,
1975               struct page *page, void *fsdata)
1976{
1977    int ret;
1978    struct inode *inode = mapping->host;
1979
1980    ret = ocfs2_write_end_nolock(mapping, pos, len, copied, page, fsdata);
1981
1982    up_write(&OCFS2_I(inode)->ip_alloc_sem);
1983    ocfs2_inode_unlock(inode, 1);
1984
1985    return ret;
1986}
1987
1988const struct address_space_operations ocfs2_aops = {
1989    .readpage = ocfs2_readpage,
1990    .readpages = ocfs2_readpages,
1991    .writepage = ocfs2_writepage,
1992    .write_begin = ocfs2_write_begin,
1993    .write_end = ocfs2_write_end,
1994    .bmap = ocfs2_bmap,
1995    .sync_page = block_sync_page,
1996    .direct_IO = ocfs2_direct_IO,
1997    .invalidatepage = ocfs2_invalidatepage,
1998    .releasepage = ocfs2_releasepage,
1999    .migratepage = buffer_migrate_page,
2000    .is_partially_uptodate = block_is_partially_uptodate,
2001    .error_remove_page = generic_error_remove_page,
2002};
2003

Archive Download this file



interactive