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#include <cluster/masklog.h>
33
34#include "ocfs2.h"
35
36#include "alloc.h"
37#include "aops.h"
38#include "dlmglue.h"
39#include "extent_map.h"
40#include "file.h"
41#include "inode.h"
42#include "journal.h"
43#include "suballoc.h"
44#include "super.h"
45#include "symlink.h"
46#include "refcounttree.h"
47#include "ocfs2_trace.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    trace_ocfs2_symlink_get_block(
63            (unsigned long long)OCFS2_I(inode)->ip_blkno,
64            (unsigned long long)iblock, bh_result, create);
65
66    BUG_ON(ocfs2_inode_is_fast_symlink(inode));
67
68    if ((iblock << inode->i_sb->s_blocksize_bits) > PATH_MAX + 1) {
69        mlog(ML_ERROR, "block offset > PATH_MAX: %llu",
70             (unsigned long long)iblock);
71        goto bail;
72    }
73
74    status = ocfs2_read_inode_block(inode, &bh);
75    if (status < 0) {
76        mlog_errno(status);
77        goto bail;
78    }
79    fe = (struct ocfs2_dinode *) bh->b_data;
80
81    if ((u64)iblock >= ocfs2_clusters_to_blocks(inode->i_sb,
82                            le32_to_cpu(fe->i_clusters))) {
83        mlog(ML_ERROR, "block offset is outside the allocated size: "
84             "%llu\n", (unsigned long long)iblock);
85        goto bail;
86    }
87
88    /* We don't use the page cache to create symlink data, so if
89     * need be, copy it over from the buffer cache. */
90    if (!buffer_uptodate(bh_result) && ocfs2_inode_is_new(inode)) {
91        u64 blkno = le64_to_cpu(fe->id2.i_list.l_recs[0].e_blkno) +
92                iblock;
93        buffer_cache_bh = sb_getblk(osb->sb, blkno);
94        if (!buffer_cache_bh) {
95            mlog(ML_ERROR, "couldn't getblock for symlink!\n");
96            goto bail;
97        }
98
99        /* we haven't locked out transactions, so a commit
100         * could've happened. Since we've got a reference on
101         * the bh, even if it commits while we're doing the
102         * copy, the data is still good. */
103        if (buffer_jbd(buffer_cache_bh)
104            && ocfs2_inode_is_new(inode)) {
105            kaddr = kmap_atomic(bh_result->b_page, KM_USER0);
106            if (!kaddr) {
107                mlog(ML_ERROR, "couldn't kmap!\n");
108                goto bail;
109            }
110            memcpy(kaddr + (bh_result->b_size * iblock),
111                   buffer_cache_bh->b_data,
112                   bh_result->b_size);
113            kunmap_atomic(kaddr, KM_USER0);
114            set_buffer_uptodate(bh_result);
115        }
116        brelse(buffer_cache_bh);
117    }
118
119    map_bh(bh_result, inode->i_sb,
120           le64_to_cpu(fe->id2.i_list.l_recs[0].e_blkno) + iblock);
121
122    err = 0;
123
124bail:
125    brelse(bh);
126
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    trace_ocfs2_get_block((unsigned long long)OCFS2_I(inode)->ip_blkno,
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_write_begin() 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
201    past_eof = ocfs2_blocks_for_bytes(inode->i_sb, i_size_read(inode));
202
203    trace_ocfs2_get_block_end((unsigned long long)OCFS2_I(inode)->ip_blkno,
204                  (unsigned long long)past_eof);
205    if (create && (iblock >= past_eof))
206        set_buffer_new(bh_result);
207
208bail:
209    if (err < 0)
210        err = -EIO;
211
212    return err;
213}
214
215int ocfs2_read_inline_data(struct inode *inode, struct page *page,
216               struct buffer_head *di_bh)
217{
218    void *kaddr;
219    loff_t size;
220    struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
221
222    if (!(le16_to_cpu(di->i_dyn_features) & OCFS2_INLINE_DATA_FL)) {
223        ocfs2_error(inode->i_sb, "Inode %llu lost inline data flag",
224                (unsigned long long)OCFS2_I(inode)->ip_blkno);
225        return -EROFS;
226    }
227
228    size = i_size_read(inode);
229
230    if (size > PAGE_CACHE_SIZE ||
231        size > ocfs2_max_inline_data_with_xattr(inode->i_sb, di)) {
232        ocfs2_error(inode->i_sb,
233                "Inode %llu has with inline data has bad size: %Lu",
234                (unsigned long long)OCFS2_I(inode)->ip_blkno,
235                (unsigned long long)size);
236        return -EROFS;
237    }
238
239    kaddr = kmap_atomic(page, KM_USER0);
240    if (size)
241        memcpy(kaddr, di->id2.i_data.id_data, size);
242    /* Clear the remaining part of the page */
243    memset(kaddr + size, 0, PAGE_CACHE_SIZE - size);
244    flush_dcache_page(page);
245    kunmap_atomic(kaddr, KM_USER0);
246
247    SetPageUptodate(page);
248
249    return 0;
250}
251
252static int ocfs2_readpage_inline(struct inode *inode, struct page *page)
253{
254    int ret;
255    struct buffer_head *di_bh = NULL;
256
257    BUG_ON(!PageLocked(page));
258    BUG_ON(!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL));
259
260    ret = ocfs2_read_inode_block(inode, &di_bh);
261    if (ret) {
262        mlog_errno(ret);
263        goto out;
264    }
265
266    ret = ocfs2_read_inline_data(inode, page, di_bh);
267out:
268    unlock_page(page);
269
270    brelse(di_bh);
271    return ret;
272}
273
274static int ocfs2_readpage(struct file *file, struct page *page)
275{
276    struct inode *inode = page->mapping->host;
277    struct ocfs2_inode_info *oi = OCFS2_I(inode);
278    loff_t start = (loff_t)page->index << PAGE_CACHE_SHIFT;
279    int ret, unlock = 1;
280
281    trace_ocfs2_readpage((unsigned long long)oi->ip_blkno,
282                 (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    return ret;
328}
329
330/*
331 * This is used only for read-ahead. Failures or difficult to handle
332 * situations are safe to ignore.
333 *
334 * Right now, we don't bother with BH_Boundary - in-inode extent lists
335 * are quite large (243 extents on 4k blocks), so most inodes don't
336 * grow out to a tree. If need be, detecting boundary extents could
337 * trivially be added in a future version of ocfs2_get_block().
338 */
339static int ocfs2_readpages(struct file *filp, struct address_space *mapping,
340               struct list_head *pages, unsigned nr_pages)
341{
342    int ret, err = -EIO;
343    struct inode *inode = mapping->host;
344    struct ocfs2_inode_info *oi = OCFS2_I(inode);
345    loff_t start;
346    struct page *last;
347
348    /*
349     * Use the nonblocking flag for the dlm code to avoid page
350     * lock inversion, but don't bother with retrying.
351     */
352    ret = ocfs2_inode_lock_full(inode, NULL, 0, OCFS2_LOCK_NONBLOCK);
353    if (ret)
354        return err;
355
356    if (down_read_trylock(&oi->ip_alloc_sem) == 0) {
357        ocfs2_inode_unlock(inode, 0);
358        return err;
359    }
360
361    /*
362     * Don't bother with inline-data. There isn't anything
363     * to read-ahead in that case anyway...
364     */
365    if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL)
366        goto out_unlock;
367
368    /*
369     * Check whether a remote node truncated this file - we just
370     * drop out in that case as it's not worth handling here.
371     */
372    last = list_entry(pages->prev, struct page, lru);
373    start = (loff_t)last->index << PAGE_CACHE_SHIFT;
374    if (start >= i_size_read(inode))
375        goto out_unlock;
376
377    err = mpage_readpages(mapping, pages, nr_pages, ocfs2_get_block);
378
379out_unlock:
380    up_read(&oi->ip_alloc_sem);
381    ocfs2_inode_unlock(inode, 0);
382
383    return err;
384}
385
386/* Note: Because we don't support holes, our allocation has
387 * already happened (allocation writes zeros to the file data)
388 * so we don't have to worry about ordered writes in
389 * ocfs2_writepage.
390 *
391 * ->writepage is called during the process of invalidating the page cache
392 * during blocked lock processing. It can't block on any cluster locks
393 * to during block mapping. It's relying on the fact that the block
394 * mapping can't have disappeared under the dirty pages that it is
395 * being asked to write back.
396 */
397static int ocfs2_writepage(struct page *page, struct writeback_control *wbc)
398{
399    trace_ocfs2_writepage(
400        (unsigned long long)OCFS2_I(page->mapping->host)->ip_blkno,
401        page->index);
402
403    return block_write_full_page(page, ocfs2_get_block, wbc);
404}
405
406/* Taken from ext3. We don't necessarily need the full blown
407 * functionality yet, but IMHO it's better to cut and paste the whole
408 * thing so we can avoid introducing our own bugs (and easily pick up
409 * their fixes when they happen) --Mark */
410int walk_page_buffers( handle_t *handle,
411            struct buffer_head *head,
412            unsigned from,
413            unsigned to,
414            int *partial,
415            int (*fn)( handle_t *handle,
416                    struct buffer_head *bh))
417{
418    struct buffer_head *bh;
419    unsigned block_start, block_end;
420    unsigned blocksize = head->b_size;
421    int err, ret = 0;
422    struct buffer_head *next;
423
424    for ( bh = head, block_start = 0;
425        ret == 0 && (bh != head || !block_start);
426            block_start = block_end, bh = next)
427    {
428        next = bh->b_this_page;
429        block_end = block_start + blocksize;
430        if (block_end <= from || block_start >= to) {
431            if (partial && !buffer_uptodate(bh))
432                *partial = 1;
433            continue;
434        }
435        err = (*fn)(handle, bh);
436        if (!ret)
437            ret = err;
438    }
439    return ret;
440}
441
442static sector_t ocfs2_bmap(struct address_space *mapping, sector_t block)
443{
444    sector_t status;
445    u64 p_blkno = 0;
446    int err = 0;
447    struct inode *inode = mapping->host;
448
449    trace_ocfs2_bmap((unsigned long long)OCFS2_I(inode)->ip_blkno,
450             (unsigned long long)block);
451
452    /* We don't need to lock journal system files, since they aren't
453     * accessed concurrently from multiple nodes.
454     */
455    if (!INODE_JOURNAL(inode)) {
456        err = ocfs2_inode_lock(inode, NULL, 0);
457        if (err) {
458            if (err != -ENOENT)
459                mlog_errno(err);
460            goto bail;
461        }
462        down_read(&OCFS2_I(inode)->ip_alloc_sem);
463    }
464
465    if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL))
466        err = ocfs2_extent_map_get_blocks(inode, block, &p_blkno, NULL,
467                          NULL);
468
469    if (!INODE_JOURNAL(inode)) {
470        up_read(&OCFS2_I(inode)->ip_alloc_sem);
471        ocfs2_inode_unlock(inode, 0);
472    }
473
474    if (err) {
475        mlog(ML_ERROR, "get_blocks() failed, block = %llu\n",
476             (unsigned long long)block);
477        mlog_errno(err);
478        goto bail;
479    }
480
481bail:
482    status = err ? 0 : p_blkno;
483
484    return status;
485}
486
487/*
488 * TODO: Make this into a generic get_blocks function.
489 *
490 * From do_direct_io in direct-io.c:
491 * "So what we do is to permit the ->get_blocks function to populate
492 * bh.b_size with the size of IO which is permitted at this offset and
493 * this i_blkbits."
494 *
495 * This function is called directly from get_more_blocks in direct-io.c.
496 *
497 * called like this: dio->get_blocks(dio->inode, fs_startblk,
498 * fs_count, map_bh, dio->rw == WRITE);
499 *
500 * Note that we never bother to allocate blocks here, and thus ignore the
501 * create argument.
502 */
503static int ocfs2_direct_IO_get_blocks(struct inode *inode, sector_t iblock,
504                     struct buffer_head *bh_result, int create)
505{
506    int ret;
507    u64 p_blkno, inode_blocks, contig_blocks;
508    unsigned int ext_flags;
509    unsigned char blocksize_bits = inode->i_sb->s_blocksize_bits;
510    unsigned long max_blocks = bh_result->b_size >> inode->i_blkbits;
511
512    /* This function won't even be called if the request isn't all
513     * nicely aligned and of the right size, so there's no need
514     * for us to check any of that. */
515
516    inode_blocks = ocfs2_blocks_for_bytes(inode->i_sb, i_size_read(inode));
517
518    /* This figures out the size of the next contiguous block, and
519     * our logical offset */
520    ret = ocfs2_extent_map_get_blocks(inode, iblock, &p_blkno,
521                      &contig_blocks, &ext_flags);
522    if (ret) {
523        mlog(ML_ERROR, "get_blocks() failed iblock=%llu\n",
524             (unsigned long long)iblock);
525        ret = -EIO;
526        goto bail;
527    }
528
529    /* We should already CoW the refcounted extent in case of create. */
530    BUG_ON(create && (ext_flags & OCFS2_EXT_REFCOUNTED));
531
532    /*
533     * get_more_blocks() expects us to describe a hole by clearing
534     * the mapped bit on bh_result().
535     *
536     * Consider an unwritten extent as a hole.
537     */
538    if (p_blkno && !(ext_flags & OCFS2_EXT_UNWRITTEN))
539        map_bh(bh_result, inode->i_sb, p_blkno);
540    else
541        clear_buffer_mapped(bh_result);
542
543    /* make sure we don't map more than max_blocks blocks here as
544       that's all the kernel will handle at this point. */
545    if (max_blocks < contig_blocks)
546        contig_blocks = max_blocks;
547    bh_result->b_size = contig_blocks << blocksize_bits;
548bail:
549    return ret;
550}
551
552/*
553 * ocfs2_dio_end_io is called by the dio core when a dio is finished. We're
554 * particularly interested in the aio/dio case. Like the core uses
555 * i_alloc_sem, we use the rw_lock DLM lock to protect io on one node from
556 * truncation on another.
557 */
558static void ocfs2_dio_end_io(struct kiocb *iocb,
559                 loff_t offset,
560                 ssize_t bytes,
561                 void *private,
562                 int ret,
563                 bool is_async)
564{
565    struct inode *inode = iocb->ki_filp->f_path.dentry->d_inode;
566    int level;
567
568    /* this io's submitter should not have unlocked this before we could */
569    BUG_ON(!ocfs2_iocb_is_rw_locked(iocb));
570
571    if (ocfs2_iocb_is_sem_locked(iocb)) {
572        up_read(&inode->i_alloc_sem);
573        ocfs2_iocb_clear_sem_locked(iocb);
574    }
575
576    ocfs2_iocb_clear_rw_locked(iocb);
577
578    level = ocfs2_iocb_rw_locked_level(iocb);
579    ocfs2_rw_unlock(inode, level);
580
581    if (is_async)
582        aio_complete(iocb, ret, 0);
583}
584
585/*
586 * ocfs2_invalidatepage() and ocfs2_releasepage() are shamelessly stolen
587 * from ext3. PageChecked() bits have been removed as OCFS2 does not
588 * do journalled data.
589 */
590static void ocfs2_invalidatepage(struct page *page, unsigned long offset)
591{
592    journal_t *journal = OCFS2_SB(page->mapping->host->i_sb)->journal->j_journal;
593
594    jbd2_journal_invalidatepage(journal, page, offset);
595}
596
597static int ocfs2_releasepage(struct page *page, gfp_t wait)
598{
599    journal_t *journal = OCFS2_SB(page->mapping->host->i_sb)->journal->j_journal;
600
601    if (!page_has_buffers(page))
602        return 0;
603    return jbd2_journal_try_to_free_buffers(journal, page, wait);
604}
605
606static ssize_t ocfs2_direct_IO(int rw,
607                   struct kiocb *iocb,
608                   const struct iovec *iov,
609                   loff_t offset,
610                   unsigned long nr_segs)
611{
612    struct file *file = iocb->ki_filp;
613    struct inode *inode = file->f_path.dentry->d_inode->i_mapping->host;
614
615    /*
616     * Fallback to buffered I/O if we see an inode without
617     * extents.
618     */
619    if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
620        return 0;
621
622    /* Fallback to buffered I/O if we are appending. */
623    if (i_size_read(inode) <= offset)
624        return 0;
625
626    return __blockdev_direct_IO(rw, iocb, inode, inode->i_sb->s_bdev,
627                    iov, offset, nr_segs,
628                    ocfs2_direct_IO_get_blocks,
629                    ocfs2_dio_end_io, NULL, 0);
630}
631
632static void ocfs2_figure_cluster_boundaries(struct ocfs2_super *osb,
633                        u32 cpos,
634                        unsigned int *start,
635                        unsigned int *end)
636{
637    unsigned int cluster_start = 0, cluster_end = PAGE_CACHE_SIZE;
638
639    if (unlikely(PAGE_CACHE_SHIFT > osb->s_clustersize_bits)) {
640        unsigned int cpp;
641
642        cpp = 1 << (PAGE_CACHE_SHIFT - osb->s_clustersize_bits);
643
644        cluster_start = cpos % cpp;
645        cluster_start = cluster_start << osb->s_clustersize_bits;
646
647        cluster_end = cluster_start + osb->s_clustersize;
648    }
649
650    BUG_ON(cluster_start > PAGE_SIZE);
651    BUG_ON(cluster_end > PAGE_SIZE);
652
653    if (start)
654        *start = cluster_start;
655    if (end)
656        *end = cluster_end;
657}
658
659/*
660 * 'from' and 'to' are the region in the page to avoid zeroing.
661 *
662 * If pagesize > clustersize, this function will avoid zeroing outside
663 * of the cluster boundary.
664 *
665 * from == to == 0 is code for "zero the entire cluster region"
666 */
667static void ocfs2_clear_page_regions(struct page *page,
668                     struct ocfs2_super *osb, u32 cpos,
669                     unsigned from, unsigned to)
670{
671    void *kaddr;
672    unsigned int cluster_start, cluster_end;
673
674    ocfs2_figure_cluster_boundaries(osb, cpos, &cluster_start, &cluster_end);
675
676    kaddr = kmap_atomic(page, KM_USER0);
677
678    if (from || to) {
679        if (from > cluster_start)
680            memset(kaddr + cluster_start, 0, from - cluster_start);
681        if (to < cluster_end)
682            memset(kaddr + to, 0, cluster_end - to);
683    } else {
684        memset(kaddr + cluster_start, 0, cluster_end - cluster_start);
685    }
686
687    kunmap_atomic(kaddr, KM_USER0);
688}
689
690/*
691 * Nonsparse file systems fully allocate before we get to the write
692 * code. This prevents ocfs2_write() from tagging the write as an
693 * allocating one, which means ocfs2_map_page_blocks() might try to
694 * read-in the blocks at the tail of our file. Avoid reading them by
695 * testing i_size against each block offset.
696 */
697static int ocfs2_should_read_blk(struct inode *inode, struct page *page,
698                 unsigned int block_start)
699{
700    u64 offset = page_offset(page) + block_start;
701
702    if (ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
703        return 1;
704
705    if (i_size_read(inode) > offset)
706        return 1;
707
708    return 0;
709}
710
711/*
712 * Some of this taken from __block_write_begin(). We already have our
713 * mapping by now though, and the entire write will be allocating or
714 * it won't, so not much need to use BH_New.
715 *
716 * This will also skip zeroing, which is handled externally.
717 */
718int ocfs2_map_page_blocks(struct page *page, u64 *p_blkno,
719              struct inode *inode, unsigned int from,
720              unsigned int to, int new)
721{
722    int ret = 0;
723    struct buffer_head *head, *bh, *wait[2], **wait_bh = wait;
724    unsigned int block_end, block_start;
725    unsigned int bsize = 1 << inode->i_blkbits;
726
727    if (!page_has_buffers(page))
728        create_empty_buffers(page, bsize, 0);
729
730    head = page_buffers(page);
731    for (bh = head, block_start = 0; bh != head || !block_start;
732         bh = bh->b_this_page, block_start += bsize) {
733        block_end = block_start + bsize;
734
735        clear_buffer_new(bh);
736
737        /*
738         * Ignore blocks outside of our i/o range -
739         * they may belong to unallocated clusters.
740         */
741        if (block_start >= to || block_end <= from) {
742            if (PageUptodate(page))
743                set_buffer_uptodate(bh);
744            continue;
745        }
746
747        /*
748         * For an allocating write with cluster size >= page
749         * size, we always write the entire page.
750         */
751        if (new)
752            set_buffer_new(bh);
753
754        if (!buffer_mapped(bh)) {
755            map_bh(bh, inode->i_sb, *p_blkno);
756            unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
757        }
758
759        if (PageUptodate(page)) {
760            if (!buffer_uptodate(bh))
761                set_buffer_uptodate(bh);
762        } else if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
763               !buffer_new(bh) &&
764               ocfs2_should_read_blk(inode, page, block_start) &&
765               (block_start < from || block_end > to)) {
766            ll_rw_block(READ, 1, &bh);
767            *wait_bh++=bh;
768        }
769
770        *p_blkno = *p_blkno + 1;
771    }
772
773    /*
774     * If we issued read requests - let them complete.
775     */
776    while(wait_bh > wait) {
777        wait_on_buffer(*--wait_bh);
778        if (!buffer_uptodate(*wait_bh))
779            ret = -EIO;
780    }
781
782    if (ret == 0 || !new)
783        return ret;
784
785    /*
786     * If we get -EIO above, zero out any newly allocated blocks
787     * to avoid exposing stale data.
788     */
789    bh = head;
790    block_start = 0;
791    do {
792        block_end = block_start + bsize;
793        if (block_end <= from)
794            goto next_bh;
795        if (block_start >= to)
796            break;
797
798        zero_user(page, block_start, bh->b_size);
799        set_buffer_uptodate(bh);
800        mark_buffer_dirty(bh);
801
802next_bh:
803        block_start = block_end;
804        bh = bh->b_this_page;
805    } while (bh != head);
806
807    return ret;
808}
809
810#if (PAGE_CACHE_SIZE >= OCFS2_MAX_CLUSTERSIZE)
811#define OCFS2_MAX_CTXT_PAGES 1
812#else
813#define OCFS2_MAX_CTXT_PAGES (OCFS2_MAX_CLUSTERSIZE / PAGE_CACHE_SIZE)
814#endif
815
816#define OCFS2_MAX_CLUSTERS_PER_PAGE (PAGE_CACHE_SIZE / OCFS2_MIN_CLUSTERSIZE)
817
818/*
819 * Describe the state of a single cluster to be written to.
820 */
821struct ocfs2_write_cluster_desc {
822    u32 c_cpos;
823    u32 c_phys;
824    /*
825     * Give this a unique field because c_phys eventually gets
826     * filled.
827     */
828    unsigned c_new;
829    unsigned c_unwritten;
830    unsigned c_needs_zero;
831};
832
833struct ocfs2_write_ctxt {
834    /* Logical cluster position / len of write */
835    u32 w_cpos;
836    u32 w_clen;
837
838    /* First cluster allocated in a nonsparse extend */
839    u32 w_first_new_cpos;
840
841    struct ocfs2_write_cluster_desc w_desc[OCFS2_MAX_CLUSTERS_PER_PAGE];
842
843    /*
844     * This is true if page_size > cluster_size.
845     *
846     * It triggers a set of special cases during write which might
847     * have to deal with allocating writes to partial pages.
848     */
849    unsigned int w_large_pages;
850
851    /*
852     * Pages involved in this write.
853     *
854     * w_target_page is the page being written to by the user.
855     *
856     * w_pages is an array of pages which always contains
857     * w_target_page, and in the case of an allocating write with
858     * page_size < cluster size, it will contain zero'd and mapped
859     * pages adjacent to w_target_page which need to be written
860     * out in so that future reads from that region will get
861     * zero's.
862     */
863    unsigned int w_num_pages;
864    struct page *w_pages[OCFS2_MAX_CTXT_PAGES];
865    struct page *w_target_page;
866
867    /*
868     * ocfs2_write_end() uses this to know what the real range to
869     * write in the target should be.
870     */
871    unsigned int w_target_from;
872    unsigned int w_target_to;
873
874    /*
875     * We could use journal_current_handle() but this is cleaner,
876     * IMHO -Mark
877     */
878    handle_t *w_handle;
879
880    struct buffer_head *w_di_bh;
881
882    struct ocfs2_cached_dealloc_ctxt w_dealloc;
883};
884
885void ocfs2_unlock_and_free_pages(struct page **pages, int num_pages)
886{
887    int i;
888
889    for(i = 0; i < num_pages; i++) {
890        if (pages[i]) {
891            unlock_page(pages[i]);
892            mark_page_accessed(pages[i]);
893            page_cache_release(pages[i]);
894        }
895    }
896}
897
898static void ocfs2_free_write_ctxt(struct ocfs2_write_ctxt *wc)
899{
900    ocfs2_unlock_and_free_pages(wc->w_pages, wc->w_num_pages);
901
902    brelse(wc->w_di_bh);
903    kfree(wc);
904}
905
906static int ocfs2_alloc_write_ctxt(struct ocfs2_write_ctxt **wcp,
907                  struct ocfs2_super *osb, loff_t pos,
908                  unsigned len, struct buffer_head *di_bh)
909{
910    u32 cend;
911    struct ocfs2_write_ctxt *wc;
912
913    wc = kzalloc(sizeof(struct ocfs2_write_ctxt), GFP_NOFS);
914    if (!wc)
915        return -ENOMEM;
916
917    wc->w_cpos = pos >> osb->s_clustersize_bits;
918    wc->w_first_new_cpos = UINT_MAX;
919    cend = (pos + len - 1) >> osb->s_clustersize_bits;
920    wc->w_clen = cend - wc->w_cpos + 1;
921    get_bh(di_bh);
922    wc->w_di_bh = di_bh;
923
924    if (unlikely(PAGE_CACHE_SHIFT > osb->s_clustersize_bits))
925        wc->w_large_pages = 1;
926    else
927        wc->w_large_pages = 0;
928
929    ocfs2_init_dealloc_ctxt(&wc->w_dealloc);
930
931    *wcp = wc;
932
933    return 0;
934}
935
936/*
937 * If a page has any new buffers, zero them out here, and mark them uptodate
938 * and dirty so they'll be written out (in order to prevent uninitialised
939 * block data from leaking). And clear the new bit.
940 */
941static void ocfs2_zero_new_buffers(struct page *page, unsigned from, unsigned to)
942{
943    unsigned int block_start, block_end;
944    struct buffer_head *head, *bh;
945
946    BUG_ON(!PageLocked(page));
947    if (!page_has_buffers(page))
948        return;
949
950    bh = head = page_buffers(page);
951    block_start = 0;
952    do {
953        block_end = block_start + bh->b_size;
954
955        if (buffer_new(bh)) {
956            if (block_end > from && block_start < to) {
957                if (!PageUptodate(page)) {
958                    unsigned start, end;
959
960                    start = max(from, block_start);
961                    end = min(to, block_end);
962
963                    zero_user_segment(page, start, end);
964                    set_buffer_uptodate(bh);
965                }
966
967                clear_buffer_new(bh);
968                mark_buffer_dirty(bh);
969            }
970        }
971
972        block_start = block_end;
973        bh = bh->b_this_page;
974    } while (bh != head);
975}
976
977/*
978 * Only called when we have a failure during allocating write to write
979 * zero's to the newly allocated region.
980 */
981static void ocfs2_write_failure(struct inode *inode,
982                struct ocfs2_write_ctxt *wc,
983                loff_t user_pos, unsigned user_len)
984{
985    int i;
986    unsigned from = user_pos & (PAGE_CACHE_SIZE - 1),
987        to = user_pos + user_len;
988    struct page *tmppage;
989
990    ocfs2_zero_new_buffers(wc->w_target_page, from, to);
991
992    for(i = 0; i < wc->w_num_pages; i++) {
993        tmppage = wc->w_pages[i];
994
995        if (page_has_buffers(tmppage)) {
996            if (ocfs2_should_order_data(inode))
997                ocfs2_jbd2_file_inode(wc->w_handle, inode);
998
999            block_commit_write(tmppage, from, to);
1000        }
1001    }
1002}
1003
1004static int ocfs2_prepare_page_for_write(struct inode *inode, u64 *p_blkno,
1005                    struct ocfs2_write_ctxt *wc,
1006                    struct page *page, u32 cpos,
1007                    loff_t user_pos, unsigned user_len,
1008                    int new)
1009{
1010    int ret;
1011    unsigned int map_from = 0, map_to = 0;
1012    unsigned int cluster_start, cluster_end;
1013    unsigned int user_data_from = 0, user_data_to = 0;
1014
1015    ocfs2_figure_cluster_boundaries(OCFS2_SB(inode->i_sb), cpos,
1016                    &cluster_start, &cluster_end);
1017
1018    /* treat the write as new if the a hole/lseek spanned across
1019     * the page boundary.
1020     */
1021    new = new | ((i_size_read(inode) <= page_offset(page)) &&
1022            (page_offset(page) <= user_pos));
1023
1024    if (page == wc->w_target_page) {
1025        map_from = user_pos & (PAGE_CACHE_SIZE - 1);
1026        map_to = map_from + user_len;
1027
1028        if (new)
1029            ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1030                            cluster_start, cluster_end,
1031                            new);
1032        else
1033            ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1034                            map_from, map_to, new);
1035        if (ret) {
1036            mlog_errno(ret);
1037            goto out;
1038        }
1039
1040        user_data_from = map_from;
1041        user_data_to = map_to;
1042        if (new) {
1043            map_from = cluster_start;
1044            map_to = cluster_end;
1045        }
1046    } else {
1047        /*
1048         * If we haven't allocated the new page yet, we
1049         * shouldn't be writing it out without copying user
1050         * data. This is likely a math error from the caller.
1051         */
1052        BUG_ON(!new);
1053
1054        map_from = cluster_start;
1055        map_to = cluster_end;
1056
1057        ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1058                        cluster_start, cluster_end, new);
1059        if (ret) {
1060            mlog_errno(ret);
1061            goto out;
1062        }
1063    }
1064
1065    /*
1066     * Parts of newly allocated pages need to be zero'd.
1067     *
1068     * Above, we have also rewritten 'to' and 'from' - as far as
1069     * the rest of the function is concerned, the entire cluster
1070     * range inside of a page needs to be written.
1071     *
1072     * We can skip this if the page is up to date - it's already
1073     * been zero'd from being read in as a hole.
1074     */
1075    if (new && !PageUptodate(page))
1076        ocfs2_clear_page_regions(page, OCFS2_SB(inode->i_sb),
1077                     cpos, user_data_from, user_data_to);
1078
1079    flush_dcache_page(page);
1080
1081out:
1082    return ret;
1083}
1084
1085/*
1086 * This function will only grab one clusters worth of pages.
1087 */
1088static int ocfs2_grab_pages_for_write(struct address_space *mapping,
1089                      struct ocfs2_write_ctxt *wc,
1090                      u32 cpos, loff_t user_pos,
1091                      unsigned user_len, int new,
1092                      struct page *mmap_page)
1093{
1094    int ret = 0, i;
1095    unsigned long start, target_index, end_index, index;
1096    struct inode *inode = mapping->host;
1097    loff_t last_byte;
1098
1099    target_index = user_pos >> PAGE_CACHE_SHIFT;
1100
1101    /*
1102     * Figure out how many pages we'll be manipulating here. For
1103     * non allocating write, we just change the one
1104     * page. Otherwise, we'll need a whole clusters worth. If we're
1105     * writing past i_size, we only need enough pages to cover the
1106     * last page of the write.
1107     */
1108    if (new) {
1109        wc->w_num_pages = ocfs2_pages_per_cluster(inode->i_sb);
1110        start = ocfs2_align_clusters_to_page_index(inode->i_sb, cpos);
1111        /*
1112         * We need the index *past* the last page we could possibly
1113         * touch. This is the page past the end of the write or
1114         * i_size, whichever is greater.
1115         */
1116        last_byte = max(user_pos + user_len, i_size_read(inode));
1117        BUG_ON(last_byte < 1);
1118        end_index = ((last_byte - 1) >> PAGE_CACHE_SHIFT) + 1;
1119        if ((start + wc->w_num_pages) > end_index)
1120            wc->w_num_pages = end_index - start;
1121    } else {
1122        wc->w_num_pages = 1;
1123        start = target_index;
1124    }
1125
1126    for(i = 0; i < wc->w_num_pages; i++) {
1127        index = start + i;
1128
1129        if (index == target_index && mmap_page) {
1130            /*
1131             * ocfs2_pagemkwrite() is a little different
1132             * and wants us to directly use the page
1133             * passed in.
1134             */
1135            lock_page(mmap_page);
1136
1137            if (mmap_page->mapping != mapping) {
1138                unlock_page(mmap_page);
1139                /*
1140                 * Sanity check - the locking in
1141                 * ocfs2_pagemkwrite() should ensure
1142                 * that this code doesn't trigger.
1143                 */
1144                ret = -EINVAL;
1145                mlog_errno(ret);
1146                goto out;
1147            }
1148
1149            page_cache_get(mmap_page);
1150            wc->w_pages[i] = mmap_page;
1151        } else {
1152            wc->w_pages[i] = find_or_create_page(mapping, index,
1153                                 GFP_NOFS);
1154            if (!wc->w_pages[i]) {
1155                ret = -ENOMEM;
1156                mlog_errno(ret);
1157                goto out;
1158            }
1159        }
1160
1161        if (index == target_index)
1162            wc->w_target_page = wc->w_pages[i];
1163    }
1164out:
1165    return ret;
1166}
1167
1168/*
1169 * Prepare a single cluster for write one cluster into the file.
1170 */
1171static int ocfs2_write_cluster(struct address_space *mapping,
1172                   u32 phys, unsigned int unwritten,
1173                   unsigned int should_zero,
1174                   struct ocfs2_alloc_context *data_ac,
1175                   struct ocfs2_alloc_context *meta_ac,
1176                   struct ocfs2_write_ctxt *wc, u32 cpos,
1177                   loff_t user_pos, unsigned user_len)
1178{
1179    int ret, i, new;
1180    u64 v_blkno, p_blkno;
1181    struct inode *inode = mapping->host;
1182    struct ocfs2_extent_tree et;
1183
1184    new = phys == 0 ? 1 : 0;
1185    if (new) {
1186        u32 tmp_pos;
1187
1188        /*
1189         * This is safe to call with the page locks - it won't take
1190         * any additional semaphores or cluster locks.
1191         */
1192        tmp_pos = cpos;
1193        ret = ocfs2_add_inode_data(OCFS2_SB(inode->i_sb), inode,
1194                       &tmp_pos, 1, 0, wc->w_di_bh,
1195                       wc->w_handle, data_ac,
1196                       meta_ac, NULL);
1197        /*
1198         * This shouldn't happen because we must have already
1199         * calculated the correct meta data allocation required. The
1200         * internal tree allocation code should know how to increase
1201         * transaction credits itself.
1202         *
1203         * If need be, we could handle -EAGAIN for a
1204         * RESTART_TRANS here.
1205         */
1206        mlog_bug_on_msg(ret == -EAGAIN,
1207                "Inode %llu: EAGAIN return during allocation.\n",
1208                (unsigned long long)OCFS2_I(inode)->ip_blkno);
1209        if (ret < 0) {
1210            mlog_errno(ret);
1211            goto out;
1212        }
1213    } else if (unwritten) {
1214        ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode),
1215                          wc->w_di_bh);
1216        ret = ocfs2_mark_extent_written(inode, &et,
1217                        wc->w_handle, cpos, 1, phys,
1218                        meta_ac, &wc->w_dealloc);
1219        if (ret < 0) {
1220            mlog_errno(ret);
1221            goto out;
1222        }
1223    }
1224
1225    if (should_zero)
1226        v_blkno = ocfs2_clusters_to_blocks(inode->i_sb, cpos);
1227    else
1228        v_blkno = user_pos >> inode->i_sb->s_blocksize_bits;
1229
1230    /*
1231     * The only reason this should fail is due to an inability to
1232     * find the extent added.
1233     */
1234    ret = ocfs2_extent_map_get_blocks(inode, v_blkno, &p_blkno, NULL,
1235                      NULL);
1236    if (ret < 0) {
1237        ocfs2_error(inode->i_sb, "Corrupting extend for inode %llu, "
1238                "at logical block %llu",
1239                (unsigned long long)OCFS2_I(inode)->ip_blkno,
1240                (unsigned long long)v_blkno);
1241        goto out;
1242    }
1243
1244    BUG_ON(p_blkno == 0);
1245
1246    for(i = 0; i < wc->w_num_pages; i++) {
1247        int tmpret;
1248
1249        tmpret = ocfs2_prepare_page_for_write(inode, &p_blkno, wc,
1250                              wc->w_pages[i], cpos,
1251                              user_pos, user_len,
1252                              should_zero);
1253        if (tmpret) {
1254            mlog_errno(tmpret);
1255            if (ret == 0)
1256                ret = tmpret;
1257        }
1258    }
1259
1260    /*
1261     * We only have cleanup to do in case of allocating write.
1262     */
1263    if (ret && new)
1264        ocfs2_write_failure(inode, wc, user_pos, user_len);
1265
1266out:
1267
1268    return ret;
1269}
1270
1271static int ocfs2_write_cluster_by_desc(struct address_space *mapping,
1272                       struct ocfs2_alloc_context *data_ac,
1273                       struct ocfs2_alloc_context *meta_ac,
1274                       struct ocfs2_write_ctxt *wc,
1275                       loff_t pos, unsigned len)
1276{
1277    int ret, i;
1278    loff_t cluster_off;
1279    unsigned int local_len = len;
1280    struct ocfs2_write_cluster_desc *desc;
1281    struct ocfs2_super *osb = OCFS2_SB(mapping->host->i_sb);
1282
1283    for (i = 0; i < wc->w_clen; i++) {
1284        desc = &wc->w_desc[i];
1285
1286        /*
1287         * We have to make sure that the total write passed in
1288         * doesn't extend past a single cluster.
1289         */
1290        local_len = len;
1291        cluster_off = pos & (osb->s_clustersize - 1);
1292        if ((cluster_off + local_len) > osb->s_clustersize)
1293            local_len = osb->s_clustersize - cluster_off;
1294
1295        ret = ocfs2_write_cluster(mapping, desc->c_phys,
1296                      desc->c_unwritten,
1297                      desc->c_needs_zero,
1298                      data_ac, meta_ac,
1299                      wc, desc->c_cpos, pos, local_len);
1300        if (ret) {
1301            mlog_errno(ret);
1302            goto out;
1303        }
1304
1305        len -= local_len;
1306        pos += local_len;
1307    }
1308
1309    ret = 0;
1310out:
1311    return ret;
1312}
1313
1314/*
1315 * ocfs2_write_end() wants to know which parts of the target page it
1316 * should complete the write on. It's easiest to compute them ahead of
1317 * time when a more complete view of the write is available.
1318 */
1319static void ocfs2_set_target_boundaries(struct ocfs2_super *osb,
1320                    struct ocfs2_write_ctxt *wc,
1321                    loff_t pos, unsigned len, int alloc)
1322{
1323    struct ocfs2_write_cluster_desc *desc;
1324
1325    wc->w_target_from = pos & (PAGE_CACHE_SIZE - 1);
1326    wc->w_target_to = wc->w_target_from + len;
1327
1328    if (alloc == 0)
1329        return;
1330
1331    /*
1332     * Allocating write - we may have different boundaries based
1333     * on page size and cluster size.
1334     *
1335     * NOTE: We can no longer compute one value from the other as
1336     * the actual write length and user provided length may be
1337     * different.
1338     */
1339
1340    if (wc->w_large_pages) {
1341        /*
1342         * We only care about the 1st and last cluster within
1343         * our range and whether they should be zero'd or not. Either
1344         * value may be extended out to the start/end of a
1345         * newly allocated cluster.
1346         */
1347        desc = &wc->w_desc[0];
1348        if (desc->c_needs_zero)
1349            ocfs2_figure_cluster_boundaries(osb,
1350                            desc->c_cpos,
1351                            &wc->w_target_from,
1352                            NULL);
1353
1354        desc = &wc->w_desc[wc->w_clen - 1];
1355        if (desc->c_needs_zero)
1356            ocfs2_figure_cluster_boundaries(osb,
1357                            desc->c_cpos,
1358                            NULL,
1359                            &wc->w_target_to);
1360    } else {
1361        wc->w_target_from = 0;
1362        wc->w_target_to = PAGE_CACHE_SIZE;
1363    }
1364}
1365
1366/*
1367 * Populate each single-cluster write descriptor in the write context
1368 * with information about the i/o to be done.
1369 *
1370 * Returns the number of clusters that will have to be allocated, as
1371 * well as a worst case estimate of the number of extent records that
1372 * would have to be created during a write to an unwritten region.
1373 */
1374static int ocfs2_populate_write_desc(struct inode *inode,
1375                     struct ocfs2_write_ctxt *wc,
1376                     unsigned int *clusters_to_alloc,
1377                     unsigned int *extents_to_split)
1378{
1379    int ret;
1380    struct ocfs2_write_cluster_desc *desc;
1381    unsigned int num_clusters = 0;
1382    unsigned int ext_flags = 0;
1383    u32 phys = 0;
1384    int i;
1385
1386    *clusters_to_alloc = 0;
1387    *extents_to_split = 0;
1388
1389    for (i = 0; i < wc->w_clen; i++) {
1390        desc = &wc->w_desc[i];
1391        desc->c_cpos = wc->w_cpos + i;
1392
1393        if (num_clusters == 0) {
1394            /*
1395             * Need to look up the next extent record.
1396             */
1397            ret = ocfs2_get_clusters(inode, desc->c_cpos, &phys,
1398                         &num_clusters, &ext_flags);
1399            if (ret) {
1400                mlog_errno(ret);
1401                goto out;
1402            }
1403
1404            /* We should already CoW the refcountd extent. */
1405            BUG_ON(ext_flags & OCFS2_EXT_REFCOUNTED);
1406
1407            /*
1408             * Assume worst case - that we're writing in
1409             * the middle of the extent.
1410             *
1411             * We can assume that the write proceeds from
1412             * left to right, in which case the extent
1413             * insert code is smart enough to coalesce the
1414             * next splits into the previous records created.
1415             */
1416            if (ext_flags & OCFS2_EXT_UNWRITTEN)
1417                *extents_to_split = *extents_to_split + 2;
1418        } else if (phys) {
1419            /*
1420             * Only increment phys if it doesn't describe
1421             * a hole.
1422             */
1423            phys++;
1424        }
1425
1426        /*
1427         * If w_first_new_cpos is < UINT_MAX, we have a non-sparse
1428         * file that got extended. w_first_new_cpos tells us
1429         * where the newly allocated clusters are so we can
1430         * zero them.
1431         */
1432        if (desc->c_cpos >= wc->w_first_new_cpos) {
1433            BUG_ON(phys == 0);
1434            desc->c_needs_zero = 1;
1435        }
1436
1437        desc->c_phys = phys;
1438        if (phys == 0) {
1439            desc->c_new = 1;
1440            desc->c_needs_zero = 1;
1441            *clusters_to_alloc = *clusters_to_alloc + 1;
1442        }
1443
1444        if (ext_flags & OCFS2_EXT_UNWRITTEN) {
1445            desc->c_unwritten = 1;
1446            desc->c_needs_zero = 1;
1447        }
1448
1449        num_clusters--;
1450    }
1451
1452    ret = 0;
1453out:
1454    return ret;
1455}
1456
1457static int ocfs2_write_begin_inline(struct address_space *mapping,
1458                    struct inode *inode,
1459                    struct ocfs2_write_ctxt *wc)
1460{
1461    int ret;
1462    struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1463    struct page *page;
1464    handle_t *handle;
1465    struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1466
1467    page = find_or_create_page(mapping, 0, GFP_NOFS);
1468    if (!page) {
1469        ret = -ENOMEM;
1470        mlog_errno(ret);
1471        goto out;
1472    }
1473    /*
1474     * If we don't set w_num_pages then this page won't get unlocked
1475     * and freed on cleanup of the write context.
1476     */
1477    wc->w_pages[0] = wc->w_target_page = page;
1478    wc->w_num_pages = 1;
1479
1480    handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1481    if (IS_ERR(handle)) {
1482        ret = PTR_ERR(handle);
1483        mlog_errno(ret);
1484        goto out;
1485    }
1486
1487    ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), wc->w_di_bh,
1488                      OCFS2_JOURNAL_ACCESS_WRITE);
1489    if (ret) {
1490        ocfs2_commit_trans(osb, handle);
1491
1492        mlog_errno(ret);
1493        goto out;
1494    }
1495
1496    if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL))
1497        ocfs2_set_inode_data_inline(inode, di);
1498
1499    if (!PageUptodate(page)) {
1500        ret = ocfs2_read_inline_data(inode, page, wc->w_di_bh);
1501        if (ret) {
1502            ocfs2_commit_trans(osb, handle);
1503
1504            goto out;
1505        }
1506    }
1507
1508    wc->w_handle = handle;
1509out:
1510    return ret;
1511}
1512
1513int ocfs2_size_fits_inline_data(struct buffer_head *di_bh, u64 new_size)
1514{
1515    struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
1516
1517    if (new_size <= le16_to_cpu(di->id2.i_data.id_count))
1518        return 1;
1519    return 0;
1520}
1521
1522static int ocfs2_try_to_write_inline_data(struct address_space *mapping,
1523                      struct inode *inode, loff_t pos,
1524                      unsigned len, struct page *mmap_page,
1525                      struct ocfs2_write_ctxt *wc)
1526{
1527    int ret, written = 0;
1528    loff_t end = pos + len;
1529    struct ocfs2_inode_info *oi = OCFS2_I(inode);
1530    struct ocfs2_dinode *di = NULL;
1531
1532    trace_ocfs2_try_to_write_inline_data((unsigned long long)oi->ip_blkno,
1533                         len, (unsigned long long)pos,
1534                         oi->ip_dyn_features);
1535
1536    /*
1537     * Handle inodes which already have inline data 1st.
1538     */
1539    if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1540        if (mmap_page == NULL &&
1541            ocfs2_size_fits_inline_data(wc->w_di_bh, end))
1542            goto do_inline_write;
1543
1544        /*
1545         * The write won't fit - we have to give this inode an
1546         * inline extent list now.
1547         */
1548        ret = ocfs2_convert_inline_data_to_extents(inode, wc->w_di_bh);
1549        if (ret)
1550            mlog_errno(ret);
1551        goto out;
1552    }
1553
1554    /*
1555     * Check whether the inode can accept inline data.
1556     */
1557    if (oi->ip_clusters != 0 || i_size_read(inode) != 0)
1558        return 0;
1559
1560    /*
1561     * Check whether the write can fit.
1562     */
1563    di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1564    if (mmap_page ||
1565        end > ocfs2_max_inline_data_with_xattr(inode->i_sb, di))
1566        return 0;
1567
1568do_inline_write:
1569    ret = ocfs2_write_begin_inline(mapping, inode, wc);
1570    if (ret) {
1571        mlog_errno(ret);
1572        goto out;
1573    }
1574
1575    /*
1576     * This signals to the caller that the data can be written
1577     * inline.
1578     */
1579    written = 1;
1580out:
1581    return written ? written : ret;
1582}
1583
1584/*
1585 * This function only does anything for file systems which can't
1586 * handle sparse files.
1587 *
1588 * What we want to do here is fill in any hole between the current end
1589 * of allocation and the end of our write. That way the rest of the
1590 * write path can treat it as an non-allocating write, which has no
1591 * special case code for sparse/nonsparse files.
1592 */
1593static int ocfs2_expand_nonsparse_inode(struct inode *inode,
1594                    struct buffer_head *di_bh,
1595                    loff_t pos, unsigned len,
1596                    struct ocfs2_write_ctxt *wc)
1597{
1598    int ret;
1599    loff_t newsize = pos + len;
1600
1601    BUG_ON(ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)));
1602
1603    if (newsize <= i_size_read(inode))
1604        return 0;
1605
1606    ret = ocfs2_extend_no_holes(inode, di_bh, newsize, pos);
1607    if (ret)
1608        mlog_errno(ret);
1609
1610    wc->w_first_new_cpos =
1611        ocfs2_clusters_for_bytes(inode->i_sb, i_size_read(inode));
1612
1613    return ret;
1614}
1615
1616static int ocfs2_zero_tail(struct inode *inode, struct buffer_head *di_bh,
1617               loff_t pos)
1618{
1619    int ret = 0;
1620
1621    BUG_ON(!ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)));
1622    if (pos > i_size_read(inode))
1623        ret = ocfs2_zero_extend(inode, di_bh, pos);
1624
1625    return ret;
1626}
1627
1628/*
1629 * Try to flush truncate logs if we can free enough clusters from it.
1630 * As for return value, "< 0" means error, "0" no space and "1" means
1631 * we have freed enough spaces and let the caller try to allocate again.
1632 */
1633static int ocfs2_try_to_free_truncate_log(struct ocfs2_super *osb,
1634                      unsigned int needed)
1635{
1636    tid_t target;
1637    int ret = 0;
1638    unsigned int truncated_clusters;
1639
1640    mutex_lock(&osb->osb_tl_inode->i_mutex);
1641    truncated_clusters = osb->truncated_clusters;
1642    mutex_unlock(&osb->osb_tl_inode->i_mutex);
1643
1644    /*
1645     * Check whether we can succeed in allocating if we free
1646     * the truncate log.
1647     */
1648    if (truncated_clusters < needed)
1649        goto out;
1650
1651    ret = ocfs2_flush_truncate_log(osb);
1652    if (ret) {
1653        mlog_errno(ret);
1654        goto out;
1655    }
1656
1657    if (jbd2_journal_start_commit(osb->journal->j_journal, &target)) {
1658        jbd2_log_wait_commit(osb->journal->j_journal, target);
1659        ret = 1;
1660    }
1661out:
1662    return ret;
1663}
1664
1665int ocfs2_write_begin_nolock(struct file *filp,
1666                 struct address_space *mapping,
1667                 loff_t pos, unsigned len, unsigned flags,
1668                 struct page **pagep, void **fsdata,
1669                 struct buffer_head *di_bh, struct page *mmap_page)
1670{
1671    int ret, cluster_of_pages, credits = OCFS2_INODE_UPDATE_CREDITS;
1672    unsigned int clusters_to_alloc, extents_to_split, clusters_need = 0;
1673    struct ocfs2_write_ctxt *wc;
1674    struct inode *inode = mapping->host;
1675    struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1676    struct ocfs2_dinode *di;
1677    struct ocfs2_alloc_context *data_ac = NULL;
1678    struct ocfs2_alloc_context *meta_ac = NULL;
1679    handle_t *handle;
1680    struct ocfs2_extent_tree et;
1681    int try_free = 1, ret1;
1682
1683try_again:
1684    ret = ocfs2_alloc_write_ctxt(&wc, osb, pos, len, di_bh);
1685    if (ret) {
1686        mlog_errno(ret);
1687        return ret;
1688    }
1689
1690    if (ocfs2_supports_inline_data(osb)) {
1691        ret = ocfs2_try_to_write_inline_data(mapping, inode, pos, len,
1692                             mmap_page, wc);
1693        if (ret == 1) {
1694            ret = 0;
1695            goto success;
1696        }
1697        if (ret < 0) {
1698            mlog_errno(ret);
1699            goto out;
1700        }
1701    }
1702
1703    if (ocfs2_sparse_alloc(osb))
1704        ret = ocfs2_zero_tail(inode, di_bh, pos);
1705    else
1706        ret = ocfs2_expand_nonsparse_inode(inode, di_bh, pos, len,
1707                           wc);
1708    if (ret) {
1709        mlog_errno(ret);
1710        goto out;
1711    }
1712
1713    ret = ocfs2_check_range_for_refcount(inode, pos, len);
1714    if (ret < 0) {
1715        mlog_errno(ret);
1716        goto out;
1717    } else if (ret == 1) {
1718        clusters_need = wc->w_clen;
1719        ret = ocfs2_refcount_cow(inode, filp, di_bh,
1720                     wc->w_cpos, wc->w_clen, UINT_MAX);
1721        if (ret) {
1722            mlog_errno(ret);
1723            goto out;
1724        }
1725    }
1726
1727    ret = ocfs2_populate_write_desc(inode, wc, &clusters_to_alloc,
1728                    &extents_to_split);
1729    if (ret) {
1730        mlog_errno(ret);
1731        goto out;
1732    }
1733    clusters_need += clusters_to_alloc;
1734
1735    di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1736
1737    trace_ocfs2_write_begin_nolock(
1738            (unsigned long long)OCFS2_I(inode)->ip_blkno,
1739            (long long)i_size_read(inode),
1740            le32_to_cpu(di->i_clusters),
1741            pos, len, flags, mmap_page,
1742            clusters_to_alloc, extents_to_split);
1743
1744    /*
1745     * We set w_target_from, w_target_to here so that
1746     * ocfs2_write_end() knows which range in the target page to
1747     * write out. An allocation requires that we write the entire
1748     * cluster range.
1749     */
1750    if (clusters_to_alloc || extents_to_split) {
1751        /*
1752         * XXX: We are stretching the limits of
1753         * ocfs2_lock_allocators(). It greatly over-estimates
1754         * the work to be done.
1755         */
1756        ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode),
1757                          wc->w_di_bh);
1758        ret = ocfs2_lock_allocators(inode, &et,
1759                        clusters_to_alloc, extents_to_split,
1760                        &data_ac, &meta_ac);
1761        if (ret) {
1762            mlog_errno(ret);
1763            goto out;
1764        }
1765
1766        if (data_ac)
1767            data_ac->ac_resv = &OCFS2_I(inode)->ip_la_data_resv;
1768
1769        credits = ocfs2_calc_extend_credits(inode->i_sb,
1770                            &di->id2.i_list,
1771                            clusters_to_alloc);
1772
1773    }
1774
1775    /*
1776     * We have to zero sparse allocated clusters, unwritten extent clusters,
1777     * and non-sparse clusters we just extended. For non-sparse writes,
1778     * we know zeros will only be needed in the first and/or last cluster.
1779     */
1780    if (clusters_to_alloc || extents_to_split ||
1781        (wc->w_clen && (wc->w_desc[0].c_needs_zero ||
1782                wc->w_desc[wc->w_clen - 1].c_needs_zero)))
1783        cluster_of_pages = 1;
1784    else
1785        cluster_of_pages = 0;
1786
1787    ocfs2_set_target_boundaries(osb, wc, pos, len, cluster_of_pages);
1788
1789    handle = ocfs2_start_trans(osb, credits);
1790    if (IS_ERR(handle)) {
1791        ret = PTR_ERR(handle);
1792        mlog_errno(ret);
1793        goto out;
1794    }
1795
1796    wc->w_handle = handle;
1797
1798    if (clusters_to_alloc) {
1799        ret = dquot_alloc_space_nodirty(inode,
1800            ocfs2_clusters_to_bytes(osb->sb, clusters_to_alloc));
1801        if (ret)
1802            goto out_commit;
1803    }
1804    /*
1805     * We don't want this to fail in ocfs2_write_end(), so do it
1806     * here.
1807     */
1808    ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), wc->w_di_bh,
1809                      OCFS2_JOURNAL_ACCESS_WRITE);
1810    if (ret) {
1811        mlog_errno(ret);
1812        goto out_quota;
1813    }
1814
1815    /*
1816     * Fill our page array first. That way we've grabbed enough so
1817     * that we can zero and flush if we error after adding the
1818     * extent.
1819     */
1820    ret = ocfs2_grab_pages_for_write(mapping, wc, wc->w_cpos, pos, len,
1821                     cluster_of_pages, mmap_page);
1822    if (ret) {
1823        mlog_errno(ret);
1824        goto out_quota;
1825    }
1826
1827    ret = ocfs2_write_cluster_by_desc(mapping, data_ac, meta_ac, wc, pos,
1828                      len);
1829    if (ret) {
1830        mlog_errno(ret);
1831        goto out_quota;
1832    }
1833
1834    if (data_ac)
1835        ocfs2_free_alloc_context(data_ac);
1836    if (meta_ac)
1837        ocfs2_free_alloc_context(meta_ac);
1838
1839success:
1840    *pagep = wc->w_target_page;
1841    *fsdata = wc;
1842    return 0;
1843out_quota:
1844    if (clusters_to_alloc)
1845        dquot_free_space(inode,
1846              ocfs2_clusters_to_bytes(osb->sb, clusters_to_alloc));
1847out_commit:
1848    ocfs2_commit_trans(osb, handle);
1849
1850out:
1851    ocfs2_free_write_ctxt(wc);
1852
1853    if (data_ac)
1854        ocfs2_free_alloc_context(data_ac);
1855    if (meta_ac)
1856        ocfs2_free_alloc_context(meta_ac);
1857
1858    if (ret == -ENOSPC && try_free) {
1859        /*
1860         * Try to free some truncate log so that we can have enough
1861         * clusters to allocate.
1862         */
1863        try_free = 0;
1864
1865        ret1 = ocfs2_try_to_free_truncate_log(osb, clusters_need);
1866        if (ret1 == 1)
1867            goto try_again;
1868
1869        if (ret1 < 0)
1870            mlog_errno(ret1);
1871    }
1872
1873    return ret;
1874}
1875
1876static int ocfs2_write_begin(struct file *file, struct address_space *mapping,
1877                 loff_t pos, unsigned len, unsigned flags,
1878                 struct page **pagep, void **fsdata)
1879{
1880    int ret;
1881    struct buffer_head *di_bh = NULL;
1882    struct inode *inode = mapping->host;
1883
1884    ret = ocfs2_inode_lock(inode, &di_bh, 1);
1885    if (ret) {
1886        mlog_errno(ret);
1887        return ret;
1888    }
1889
1890    /*
1891     * Take alloc sem here to prevent concurrent lookups. That way
1892     * the mapping, zeroing and tree manipulation within
1893     * ocfs2_write() will be safe against ->readpage(). This
1894     * should also serve to lock out allocation from a shared
1895     * writeable region.
1896     */
1897    down_write(&OCFS2_I(inode)->ip_alloc_sem);
1898
1899    ret = ocfs2_write_begin_nolock(file, mapping, pos, len, flags, pagep,
1900                       fsdata, di_bh, NULL);
1901    if (ret) {
1902        mlog_errno(ret);
1903        goto out_fail;
1904    }
1905
1906    brelse(di_bh);
1907
1908    return 0;
1909
1910out_fail:
1911    up_write(&OCFS2_I(inode)->ip_alloc_sem);
1912
1913    brelse(di_bh);
1914    ocfs2_inode_unlock(inode, 1);
1915
1916    return ret;
1917}
1918
1919static void ocfs2_write_end_inline(struct inode *inode, loff_t pos,
1920                   unsigned len, unsigned *copied,
1921                   struct ocfs2_dinode *di,
1922                   struct ocfs2_write_ctxt *wc)
1923{
1924    void *kaddr;
1925
1926    if (unlikely(*copied < len)) {
1927        if (!PageUptodate(wc->w_target_page)) {
1928            *copied = 0;
1929            return;
1930        }
1931    }
1932
1933    kaddr = kmap_atomic(wc->w_target_page, KM_USER0);
1934    memcpy(di->id2.i_data.id_data + pos, kaddr + pos, *copied);
1935    kunmap_atomic(kaddr, KM_USER0);
1936
1937    trace_ocfs2_write_end_inline(
1938         (unsigned long long)OCFS2_I(inode)->ip_blkno,
1939         (unsigned long long)pos, *copied,
1940         le16_to_cpu(di->id2.i_data.id_count),
1941         le16_to_cpu(di->i_dyn_features));
1942}
1943
1944int ocfs2_write_end_nolock(struct address_space *mapping,
1945               loff_t pos, unsigned len, unsigned copied,
1946               struct page *page, void *fsdata)
1947{
1948    int i;
1949    unsigned from, to, start = pos & (PAGE_CACHE_SIZE - 1);
1950    struct inode *inode = mapping->host;
1951    struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1952    struct ocfs2_write_ctxt *wc = fsdata;
1953    struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1954    handle_t *handle = wc->w_handle;
1955    struct page *tmppage;
1956
1957    if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1958        ocfs2_write_end_inline(inode, pos, len, &copied, di, wc);
1959        goto out_write_size;
1960    }
1961
1962    if (unlikely(copied < len)) {
1963        if (!PageUptodate(wc->w_target_page))
1964            copied = 0;
1965
1966        ocfs2_zero_new_buffers(wc->w_target_page, start+copied,
1967                       start+len);
1968    }
1969    flush_dcache_page(wc->w_target_page);
1970
1971    for(i = 0; i < wc->w_num_pages; i++) {
1972        tmppage = wc->w_pages[i];
1973
1974        if (tmppage == wc->w_target_page) {
1975            from = wc->w_target_from;
1976            to = wc->w_target_to;
1977
1978            BUG_ON(from > PAGE_CACHE_SIZE ||
1979                   to > PAGE_CACHE_SIZE ||
1980                   to < from);
1981        } else {
1982            /*
1983             * Pages adjacent to the target (if any) imply
1984             * a hole-filling write in which case we want
1985             * to flush their entire range.
1986             */
1987            from = 0;
1988            to = PAGE_CACHE_SIZE;
1989        }
1990
1991        if (page_has_buffers(tmppage)) {
1992            if (ocfs2_should_order_data(inode))
1993                ocfs2_jbd2_file_inode(wc->w_handle, inode);
1994            block_commit_write(tmppage, from, to);
1995        }
1996    }
1997
1998out_write_size:
1999    pos += copied;
2000    if (pos > inode->i_size) {
2001        i_size_write(inode, pos);
2002        mark_inode_dirty(inode);
2003    }
2004    inode->i_blocks = ocfs2_inode_sector_count(inode);
2005    di->i_size = cpu_to_le64((u64)i_size_read(inode));
2006    inode->i_mtime = inode->i_ctime = CURRENT_TIME;
2007    di->i_mtime = di->i_ctime = cpu_to_le64(inode->i_mtime.tv_sec);
2008    di->i_mtime_nsec = di->i_ctime_nsec = cpu_to_le32(inode->i_mtime.tv_nsec);
2009    ocfs2_journal_dirty(handle, wc->w_di_bh);
2010
2011    ocfs2_commit_trans(osb, handle);
2012
2013    ocfs2_run_deallocs(osb, &wc->w_dealloc);
2014
2015    ocfs2_free_write_ctxt(wc);
2016
2017    return copied;
2018}
2019
2020static int ocfs2_write_end(struct file *file, struct address_space *mapping,
2021               loff_t pos, unsigned len, unsigned copied,
2022               struct page *page, void *fsdata)
2023{
2024    int ret;
2025    struct inode *inode = mapping->host;
2026
2027    ret = ocfs2_write_end_nolock(mapping, pos, len, copied, page, fsdata);
2028
2029    up_write(&OCFS2_I(inode)->ip_alloc_sem);
2030    ocfs2_inode_unlock(inode, 1);
2031
2032    return ret;
2033}
2034
2035const struct address_space_operations ocfs2_aops = {
2036    .readpage = ocfs2_readpage,
2037    .readpages = ocfs2_readpages,
2038    .writepage = ocfs2_writepage,
2039    .write_begin = ocfs2_write_begin,
2040    .write_end = ocfs2_write_end,
2041    .bmap = ocfs2_bmap,
2042    .direct_IO = ocfs2_direct_IO,
2043    .invalidatepage = ocfs2_invalidatepage,
2044    .releasepage = ocfs2_releasepage,
2045    .migratepage = buffer_migrate_page,
2046    .is_partially_uptodate = block_is_partially_uptodate,
2047    .error_remove_page = generic_error_remove_page,
2048};
2049

Archive Download this file



interactive