Root/fs/ocfs2/file.c

1/* -*- mode: c; c-basic-offset: 8; -*-
2 * vim: noexpandtab sw=8 ts=8 sts=0:
3 *
4 * file.c
5 *
6 * File open, close, extend, truncate
7 *
8 * Copyright (C) 2002, 2004 Oracle. All rights reserved.
9 *
10 * This program is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU General Public
12 * License as published by the Free Software Foundation; either
13 * version 2 of the License, or (at your option) any later version.
14 *
15 * This program is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18 * General Public License for more details.
19 *
20 * You should have received a copy of the GNU General Public
21 * License along with this program; if not, write to the
22 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
23 * Boston, MA 021110-1307, USA.
24 */
25
26#include <linux/capability.h>
27#include <linux/fs.h>
28#include <linux/types.h>
29#include <linux/slab.h>
30#include <linux/highmem.h>
31#include <linux/pagemap.h>
32#include <linux/uio.h>
33#include <linux/sched.h>
34#include <linux/splice.h>
35#include <linux/mount.h>
36#include <linux/writeback.h>
37#include <linux/falloc.h>
38#include <linux/quotaops.h>
39#include <linux/blkdev.h>
40
41#include <cluster/masklog.h>
42
43#include "ocfs2.h"
44
45#include "alloc.h"
46#include "aops.h"
47#include "dir.h"
48#include "dlmglue.h"
49#include "extent_map.h"
50#include "file.h"
51#include "sysfile.h"
52#include "inode.h"
53#include "ioctl.h"
54#include "journal.h"
55#include "locks.h"
56#include "mmap.h"
57#include "suballoc.h"
58#include "super.h"
59#include "xattr.h"
60#include "acl.h"
61#include "quota.h"
62#include "refcounttree.h"
63#include "ocfs2_trace.h"
64
65#include "buffer_head_io.h"
66
67static int ocfs2_init_file_private(struct inode *inode, struct file *file)
68{
69    struct ocfs2_file_private *fp;
70
71    fp = kzalloc(sizeof(struct ocfs2_file_private), GFP_KERNEL);
72    if (!fp)
73        return -ENOMEM;
74
75    fp->fp_file = file;
76    mutex_init(&fp->fp_mutex);
77    ocfs2_file_lock_res_init(&fp->fp_flock, fp);
78    file->private_data = fp;
79
80    return 0;
81}
82
83static void ocfs2_free_file_private(struct inode *inode, struct file *file)
84{
85    struct ocfs2_file_private *fp = file->private_data;
86    struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
87
88    if (fp) {
89        ocfs2_simple_drop_lockres(osb, &fp->fp_flock);
90        ocfs2_lock_res_free(&fp->fp_flock);
91        kfree(fp);
92        file->private_data = NULL;
93    }
94}
95
96static int ocfs2_file_open(struct inode *inode, struct file *file)
97{
98    int status;
99    int mode = file->f_flags;
100    struct ocfs2_inode_info *oi = OCFS2_I(inode);
101
102    trace_ocfs2_file_open(inode, file, file->f_path.dentry,
103                  (unsigned long long)OCFS2_I(inode)->ip_blkno,
104                  file->f_path.dentry->d_name.len,
105                  file->f_path.dentry->d_name.name, mode);
106
107    if (file->f_mode & FMODE_WRITE)
108        dquot_initialize(inode);
109
110    spin_lock(&oi->ip_lock);
111
112    /* Check that the inode hasn't been wiped from disk by another
113     * node. If it hasn't then we're safe as long as we hold the
114     * spin lock until our increment of open count. */
115    if (OCFS2_I(inode)->ip_flags & OCFS2_INODE_DELETED) {
116        spin_unlock(&oi->ip_lock);
117
118        status = -ENOENT;
119        goto leave;
120    }
121
122    if (mode & O_DIRECT)
123        oi->ip_flags |= OCFS2_INODE_OPEN_DIRECT;
124
125    oi->ip_open_count++;
126    spin_unlock(&oi->ip_lock);
127
128    status = ocfs2_init_file_private(inode, file);
129    if (status) {
130        /*
131         * We want to set open count back if we're failing the
132         * open.
133         */
134        spin_lock(&oi->ip_lock);
135        oi->ip_open_count--;
136        spin_unlock(&oi->ip_lock);
137    }
138
139leave:
140    return status;
141}
142
143static int ocfs2_file_release(struct inode *inode, struct file *file)
144{
145    struct ocfs2_inode_info *oi = OCFS2_I(inode);
146
147    spin_lock(&oi->ip_lock);
148    if (!--oi->ip_open_count)
149        oi->ip_flags &= ~OCFS2_INODE_OPEN_DIRECT;
150
151    trace_ocfs2_file_release(inode, file, file->f_path.dentry,
152                 oi->ip_blkno,
153                 file->f_path.dentry->d_name.len,
154                 file->f_path.dentry->d_name.name,
155                 oi->ip_open_count);
156    spin_unlock(&oi->ip_lock);
157
158    ocfs2_free_file_private(inode, file);
159
160    return 0;
161}
162
163static int ocfs2_dir_open(struct inode *inode, struct file *file)
164{
165    return ocfs2_init_file_private(inode, file);
166}
167
168static int ocfs2_dir_release(struct inode *inode, struct file *file)
169{
170    ocfs2_free_file_private(inode, file);
171    return 0;
172}
173
174static int ocfs2_sync_file(struct file *file, int datasync)
175{
176    int err = 0;
177    journal_t *journal;
178    struct inode *inode = file->f_mapping->host;
179    struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
180
181    trace_ocfs2_sync_file(inode, file, file->f_path.dentry,
182                  OCFS2_I(inode)->ip_blkno,
183                  file->f_path.dentry->d_name.len,
184                  file->f_path.dentry->d_name.name,
185                  (unsigned long long)datasync);
186
187    if (datasync && !(inode->i_state & I_DIRTY_DATASYNC)) {
188        /*
189         * We still have to flush drive's caches to get data to the
190         * platter
191         */
192        if (osb->s_mount_opt & OCFS2_MOUNT_BARRIER)
193            blkdev_issue_flush(inode->i_sb->s_bdev, GFP_KERNEL, NULL);
194        goto bail;
195    }
196
197    journal = osb->journal->j_journal;
198    err = jbd2_journal_force_commit(journal);
199
200bail:
201    if (err)
202        mlog_errno(err);
203
204    return (err < 0) ? -EIO : 0;
205}
206
207int ocfs2_should_update_atime(struct inode *inode,
208                  struct vfsmount *vfsmnt)
209{
210    struct timespec now;
211    struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
212
213    if (ocfs2_is_hard_readonly(osb) || ocfs2_is_soft_readonly(osb))
214        return 0;
215
216    if ((inode->i_flags & S_NOATIME) ||
217        ((inode->i_sb->s_flags & MS_NODIRATIME) && S_ISDIR(inode->i_mode)))
218        return 0;
219
220    /*
221     * We can be called with no vfsmnt structure - NFSD will
222     * sometimes do this.
223     *
224     * Note that our action here is different than touch_atime() -
225     * if we can't tell whether this is a noatime mount, then we
226     * don't know whether to trust the value of s_atime_quantum.
227     */
228    if (vfsmnt == NULL)
229        return 0;
230
231    if ((vfsmnt->mnt_flags & MNT_NOATIME) ||
232        ((vfsmnt->mnt_flags & MNT_NODIRATIME) && S_ISDIR(inode->i_mode)))
233        return 0;
234
235    if (vfsmnt->mnt_flags & MNT_RELATIME) {
236        if ((timespec_compare(&inode->i_atime, &inode->i_mtime) <= 0) ||
237            (timespec_compare(&inode->i_atime, &inode->i_ctime) <= 0))
238            return 1;
239
240        return 0;
241    }
242
243    now = CURRENT_TIME;
244    if ((now.tv_sec - inode->i_atime.tv_sec <= osb->s_atime_quantum))
245        return 0;
246    else
247        return 1;
248}
249
250int ocfs2_update_inode_atime(struct inode *inode,
251                 struct buffer_head *bh)
252{
253    int ret;
254    struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
255    handle_t *handle;
256    struct ocfs2_dinode *di = (struct ocfs2_dinode *) bh->b_data;
257
258    handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
259    if (IS_ERR(handle)) {
260        ret = PTR_ERR(handle);
261        mlog_errno(ret);
262        goto out;
263    }
264
265    ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), bh,
266                      OCFS2_JOURNAL_ACCESS_WRITE);
267    if (ret) {
268        mlog_errno(ret);
269        goto out_commit;
270    }
271
272    /*
273     * Don't use ocfs2_mark_inode_dirty() here as we don't always
274     * have i_mutex to guard against concurrent changes to other
275     * inode fields.
276     */
277    inode->i_atime = CURRENT_TIME;
278    di->i_atime = cpu_to_le64(inode->i_atime.tv_sec);
279    di->i_atime_nsec = cpu_to_le32(inode->i_atime.tv_nsec);
280    ocfs2_journal_dirty(handle, bh);
281
282out_commit:
283    ocfs2_commit_trans(OCFS2_SB(inode->i_sb), handle);
284out:
285    return ret;
286}
287
288static int ocfs2_set_inode_size(handle_t *handle,
289                struct inode *inode,
290                struct buffer_head *fe_bh,
291                u64 new_i_size)
292{
293    int status;
294
295    i_size_write(inode, new_i_size);
296    inode->i_blocks = ocfs2_inode_sector_count(inode);
297    inode->i_ctime = inode->i_mtime = CURRENT_TIME;
298
299    status = ocfs2_mark_inode_dirty(handle, inode, fe_bh);
300    if (status < 0) {
301        mlog_errno(status);
302        goto bail;
303    }
304
305bail:
306    return status;
307}
308
309int ocfs2_simple_size_update(struct inode *inode,
310                 struct buffer_head *di_bh,
311                 u64 new_i_size)
312{
313    int ret;
314    struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
315    handle_t *handle = NULL;
316
317    handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
318    if (IS_ERR(handle)) {
319        ret = PTR_ERR(handle);
320        mlog_errno(ret);
321        goto out;
322    }
323
324    ret = ocfs2_set_inode_size(handle, inode, di_bh,
325                   new_i_size);
326    if (ret < 0)
327        mlog_errno(ret);
328
329    ocfs2_commit_trans(osb, handle);
330out:
331    return ret;
332}
333
334static int ocfs2_cow_file_pos(struct inode *inode,
335                  struct buffer_head *fe_bh,
336                  u64 offset)
337{
338    int status;
339    u32 phys, cpos = offset >> OCFS2_SB(inode->i_sb)->s_clustersize_bits;
340    unsigned int num_clusters = 0;
341    unsigned int ext_flags = 0;
342
343    /*
344     * If the new offset is aligned to the range of the cluster, there is
345     * no space for ocfs2_zero_range_for_truncate to fill, so no need to
346     * CoW either.
347     */
348    if ((offset & (OCFS2_SB(inode->i_sb)->s_clustersize - 1)) == 0)
349        return 0;
350
351    status = ocfs2_get_clusters(inode, cpos, &phys,
352                    &num_clusters, &ext_flags);
353    if (status) {
354        mlog_errno(status);
355        goto out;
356    }
357
358    if (!(ext_flags & OCFS2_EXT_REFCOUNTED))
359        goto out;
360
361    return ocfs2_refcount_cow(inode, NULL, fe_bh, cpos, 1, cpos+1);
362
363out:
364    return status;
365}
366
367static int ocfs2_orphan_for_truncate(struct ocfs2_super *osb,
368                     struct inode *inode,
369                     struct buffer_head *fe_bh,
370                     u64 new_i_size)
371{
372    int status;
373    handle_t *handle;
374    struct ocfs2_dinode *di;
375    u64 cluster_bytes;
376
377    /*
378     * We need to CoW the cluster contains the offset if it is reflinked
379     * since we will call ocfs2_zero_range_for_truncate later which will
380     * write "0" from offset to the end of the cluster.
381     */
382    status = ocfs2_cow_file_pos(inode, fe_bh, new_i_size);
383    if (status) {
384        mlog_errno(status);
385        return status;
386    }
387
388    /* TODO: This needs to actually orphan the inode in this
389     * transaction. */
390
391    handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
392    if (IS_ERR(handle)) {
393        status = PTR_ERR(handle);
394        mlog_errno(status);
395        goto out;
396    }
397
398    status = ocfs2_journal_access_di(handle, INODE_CACHE(inode), fe_bh,
399                     OCFS2_JOURNAL_ACCESS_WRITE);
400    if (status < 0) {
401        mlog_errno(status);
402        goto out_commit;
403    }
404
405    /*
406     * Do this before setting i_size.
407     */
408    cluster_bytes = ocfs2_align_bytes_to_clusters(inode->i_sb, new_i_size);
409    status = ocfs2_zero_range_for_truncate(inode, handle, new_i_size,
410                           cluster_bytes);
411    if (status) {
412        mlog_errno(status);
413        goto out_commit;
414    }
415
416    i_size_write(inode, new_i_size);
417    inode->i_ctime = inode->i_mtime = CURRENT_TIME;
418
419    di = (struct ocfs2_dinode *) fe_bh->b_data;
420    di->i_size = cpu_to_le64(new_i_size);
421    di->i_ctime = di->i_mtime = cpu_to_le64(inode->i_ctime.tv_sec);
422    di->i_ctime_nsec = di->i_mtime_nsec = cpu_to_le32(inode->i_ctime.tv_nsec);
423
424    ocfs2_journal_dirty(handle, fe_bh);
425
426out_commit:
427    ocfs2_commit_trans(osb, handle);
428out:
429    return status;
430}
431
432static int ocfs2_truncate_file(struct inode *inode,
433                   struct buffer_head *di_bh,
434                   u64 new_i_size)
435{
436    int status = 0;
437    struct ocfs2_dinode *fe = NULL;
438    struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
439
440    /* We trust di_bh because it comes from ocfs2_inode_lock(), which
441     * already validated it */
442    fe = (struct ocfs2_dinode *) di_bh->b_data;
443
444    trace_ocfs2_truncate_file((unsigned long long)OCFS2_I(inode)->ip_blkno,
445                  (unsigned long long)le64_to_cpu(fe->i_size),
446                  (unsigned long long)new_i_size);
447
448    mlog_bug_on_msg(le64_to_cpu(fe->i_size) != i_size_read(inode),
449            "Inode %llu, inode i_size = %lld != di "
450            "i_size = %llu, i_flags = 0x%x\n",
451            (unsigned long long)OCFS2_I(inode)->ip_blkno,
452            i_size_read(inode),
453            (unsigned long long)le64_to_cpu(fe->i_size),
454            le32_to_cpu(fe->i_flags));
455
456    if (new_i_size > le64_to_cpu(fe->i_size)) {
457        trace_ocfs2_truncate_file_error(
458            (unsigned long long)le64_to_cpu(fe->i_size),
459            (unsigned long long)new_i_size);
460        status = -EINVAL;
461        mlog_errno(status);
462        goto bail;
463    }
464
465    /* lets handle the simple truncate cases before doing any more
466     * cluster locking. */
467    if (new_i_size == le64_to_cpu(fe->i_size))
468        goto bail;
469
470    down_write(&OCFS2_I(inode)->ip_alloc_sem);
471
472    ocfs2_resv_discard(&osb->osb_la_resmap,
473               &OCFS2_I(inode)->ip_la_data_resv);
474
475    /*
476     * The inode lock forced other nodes to sync and drop their
477     * pages, which (correctly) happens even if we have a truncate
478     * without allocation change - ocfs2 cluster sizes can be much
479     * greater than page size, so we have to truncate them
480     * anyway.
481     */
482    unmap_mapping_range(inode->i_mapping, new_i_size + PAGE_SIZE - 1, 0, 1);
483    truncate_inode_pages(inode->i_mapping, new_i_size);
484
485    if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
486        status = ocfs2_truncate_inline(inode, di_bh, new_i_size,
487                           i_size_read(inode), 1);
488        if (status)
489            mlog_errno(status);
490
491        goto bail_unlock_sem;
492    }
493
494    /* alright, we're going to need to do a full blown alloc size
495     * change. Orphan the inode so that recovery can complete the
496     * truncate if necessary. This does the task of marking
497     * i_size. */
498    status = ocfs2_orphan_for_truncate(osb, inode, di_bh, new_i_size);
499    if (status < 0) {
500        mlog_errno(status);
501        goto bail_unlock_sem;
502    }
503
504    status = ocfs2_commit_truncate(osb, inode, di_bh);
505    if (status < 0) {
506        mlog_errno(status);
507        goto bail_unlock_sem;
508    }
509
510    /* TODO: orphan dir cleanup here. */
511bail_unlock_sem:
512    up_write(&OCFS2_I(inode)->ip_alloc_sem);
513
514bail:
515    if (!status && OCFS2_I(inode)->ip_clusters == 0)
516        status = ocfs2_try_remove_refcount_tree(inode, di_bh);
517
518    return status;
519}
520
521/*
522 * extend file allocation only here.
523 * we'll update all the disk stuff, and oip->alloc_size
524 *
525 * expect stuff to be locked, a transaction started and enough data /
526 * metadata reservations in the contexts.
527 *
528 * Will return -EAGAIN, and a reason if a restart is needed.
529 * If passed in, *reason will always be set, even in error.
530 */
531int ocfs2_add_inode_data(struct ocfs2_super *osb,
532             struct inode *inode,
533             u32 *logical_offset,
534             u32 clusters_to_add,
535             int mark_unwritten,
536             struct buffer_head *fe_bh,
537             handle_t *handle,
538             struct ocfs2_alloc_context *data_ac,
539             struct ocfs2_alloc_context *meta_ac,
540             enum ocfs2_alloc_restarted *reason_ret)
541{
542    int ret;
543    struct ocfs2_extent_tree et;
544
545    ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), fe_bh);
546    ret = ocfs2_add_clusters_in_btree(handle, &et, logical_offset,
547                      clusters_to_add, mark_unwritten,
548                      data_ac, meta_ac, reason_ret);
549
550    return ret;
551}
552
553static int __ocfs2_extend_allocation(struct inode *inode, u32 logical_start,
554                     u32 clusters_to_add, int mark_unwritten)
555{
556    int status = 0;
557    int restart_func = 0;
558    int credits;
559    u32 prev_clusters;
560    struct buffer_head *bh = NULL;
561    struct ocfs2_dinode *fe = NULL;
562    handle_t *handle = NULL;
563    struct ocfs2_alloc_context *data_ac = NULL;
564    struct ocfs2_alloc_context *meta_ac = NULL;
565    enum ocfs2_alloc_restarted why;
566    struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
567    struct ocfs2_extent_tree et;
568    int did_quota = 0;
569
570    /*
571     * This function only exists for file systems which don't
572     * support holes.
573     */
574    BUG_ON(mark_unwritten && !ocfs2_sparse_alloc(osb));
575
576    status = ocfs2_read_inode_block(inode, &bh);
577    if (status < 0) {
578        mlog_errno(status);
579        goto leave;
580    }
581    fe = (struct ocfs2_dinode *) bh->b_data;
582
583restart_all:
584    BUG_ON(le32_to_cpu(fe->i_clusters) != OCFS2_I(inode)->ip_clusters);
585
586    ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), bh);
587    status = ocfs2_lock_allocators(inode, &et, clusters_to_add, 0,
588                       &data_ac, &meta_ac);
589    if (status) {
590        mlog_errno(status);
591        goto leave;
592    }
593
594    credits = ocfs2_calc_extend_credits(osb->sb, &fe->id2.i_list,
595                        clusters_to_add);
596    handle = ocfs2_start_trans(osb, credits);
597    if (IS_ERR(handle)) {
598        status = PTR_ERR(handle);
599        handle = NULL;
600        mlog_errno(status);
601        goto leave;
602    }
603
604restarted_transaction:
605    trace_ocfs2_extend_allocation(
606        (unsigned long long)OCFS2_I(inode)->ip_blkno,
607        (unsigned long long)i_size_read(inode),
608        le32_to_cpu(fe->i_clusters), clusters_to_add,
609        why, restart_func);
610
611    status = dquot_alloc_space_nodirty(inode,
612            ocfs2_clusters_to_bytes(osb->sb, clusters_to_add));
613    if (status)
614        goto leave;
615    did_quota = 1;
616
617    /* reserve a write to the file entry early on - that we if we
618     * run out of credits in the allocation path, we can still
619     * update i_size. */
620    status = ocfs2_journal_access_di(handle, INODE_CACHE(inode), bh,
621                     OCFS2_JOURNAL_ACCESS_WRITE);
622    if (status < 0) {
623        mlog_errno(status);
624        goto leave;
625    }
626
627    prev_clusters = OCFS2_I(inode)->ip_clusters;
628
629    status = ocfs2_add_inode_data(osb,
630                      inode,
631                      &logical_start,
632                      clusters_to_add,
633                      mark_unwritten,
634                      bh,
635                      handle,
636                      data_ac,
637                      meta_ac,
638                      &why);
639    if ((status < 0) && (status != -EAGAIN)) {
640        if (status != -ENOSPC)
641            mlog_errno(status);
642        goto leave;
643    }
644
645    ocfs2_journal_dirty(handle, bh);
646
647    spin_lock(&OCFS2_I(inode)->ip_lock);
648    clusters_to_add -= (OCFS2_I(inode)->ip_clusters - prev_clusters);
649    spin_unlock(&OCFS2_I(inode)->ip_lock);
650    /* Release unused quota reservation */
651    dquot_free_space(inode,
652            ocfs2_clusters_to_bytes(osb->sb, clusters_to_add));
653    did_quota = 0;
654
655    if (why != RESTART_NONE && clusters_to_add) {
656        if (why == RESTART_META) {
657            restart_func = 1;
658            status = 0;
659        } else {
660            BUG_ON(why != RESTART_TRANS);
661
662            /* TODO: This can be more intelligent. */
663            credits = ocfs2_calc_extend_credits(osb->sb,
664                                &fe->id2.i_list,
665                                clusters_to_add);
666            status = ocfs2_extend_trans(handle, credits);
667            if (status < 0) {
668                /* handle still has to be committed at
669                 * this point. */
670                status = -ENOMEM;
671                mlog_errno(status);
672                goto leave;
673            }
674            goto restarted_transaction;
675        }
676    }
677
678    trace_ocfs2_extend_allocation_end(OCFS2_I(inode)->ip_blkno,
679         le32_to_cpu(fe->i_clusters),
680         (unsigned long long)le64_to_cpu(fe->i_size),
681         OCFS2_I(inode)->ip_clusters,
682         (unsigned long long)i_size_read(inode));
683
684leave:
685    if (status < 0 && did_quota)
686        dquot_free_space(inode,
687            ocfs2_clusters_to_bytes(osb->sb, clusters_to_add));
688    if (handle) {
689        ocfs2_commit_trans(osb, handle);
690        handle = NULL;
691    }
692    if (data_ac) {
693        ocfs2_free_alloc_context(data_ac);
694        data_ac = NULL;
695    }
696    if (meta_ac) {
697        ocfs2_free_alloc_context(meta_ac);
698        meta_ac = NULL;
699    }
700    if ((!status) && restart_func) {
701        restart_func = 0;
702        goto restart_all;
703    }
704    brelse(bh);
705    bh = NULL;
706
707    return status;
708}
709
710/*
711 * While a write will already be ordering the data, a truncate will not.
712 * Thus, we need to explicitly order the zeroed pages.
713 */
714static handle_t *ocfs2_zero_start_ordered_transaction(struct inode *inode)
715{
716    struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
717    handle_t *handle = NULL;
718    int ret = 0;
719
720    if (!ocfs2_should_order_data(inode))
721        goto out;
722
723    handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
724    if (IS_ERR(handle)) {
725        ret = -ENOMEM;
726        mlog_errno(ret);
727        goto out;
728    }
729
730    ret = ocfs2_jbd2_file_inode(handle, inode);
731    if (ret < 0)
732        mlog_errno(ret);
733
734out:
735    if (ret) {
736        if (!IS_ERR(handle))
737            ocfs2_commit_trans(osb, handle);
738        handle = ERR_PTR(ret);
739    }
740    return handle;
741}
742
743/* Some parts of this taken from generic_cont_expand, which turned out
744 * to be too fragile to do exactly what we need without us having to
745 * worry about recursive locking in ->write_begin() and ->write_end(). */
746static int ocfs2_write_zero_page(struct inode *inode, u64 abs_from,
747                 u64 abs_to)
748{
749    struct address_space *mapping = inode->i_mapping;
750    struct page *page;
751    unsigned long index = abs_from >> PAGE_CACHE_SHIFT;
752    handle_t *handle = NULL;
753    int ret = 0;
754    unsigned zero_from, zero_to, block_start, block_end;
755
756    BUG_ON(abs_from >= abs_to);
757    BUG_ON(abs_to > (((u64)index + 1) << PAGE_CACHE_SHIFT));
758    BUG_ON(abs_from & (inode->i_blkbits - 1));
759
760    page = find_or_create_page(mapping, index, GFP_NOFS);
761    if (!page) {
762        ret = -ENOMEM;
763        mlog_errno(ret);
764        goto out;
765    }
766
767    /* Get the offsets within the page that we want to zero */
768    zero_from = abs_from & (PAGE_CACHE_SIZE - 1);
769    zero_to = abs_to & (PAGE_CACHE_SIZE - 1);
770    if (!zero_to)
771        zero_to = PAGE_CACHE_SIZE;
772
773    trace_ocfs2_write_zero_page(
774            (unsigned long long)OCFS2_I(inode)->ip_blkno,
775            (unsigned long long)abs_from,
776            (unsigned long long)abs_to,
777            index, zero_from, zero_to);
778
779    /* We know that zero_from is block aligned */
780    for (block_start = zero_from; block_start < zero_to;
781         block_start = block_end) {
782        block_end = block_start + (1 << inode->i_blkbits);
783
784        /*
785         * block_start is block-aligned. Bump it by one to force
786         * __block_write_begin and block_commit_write to zero the
787         * whole block.
788         */
789        ret = __block_write_begin(page, block_start + 1, 0,
790                      ocfs2_get_block);
791        if (ret < 0) {
792            mlog_errno(ret);
793            goto out_unlock;
794        }
795
796        if (!handle) {
797            handle = ocfs2_zero_start_ordered_transaction(inode);
798            if (IS_ERR(handle)) {
799                ret = PTR_ERR(handle);
800                handle = NULL;
801                break;
802            }
803        }
804
805        /* must not update i_size! */
806        ret = block_commit_write(page, block_start + 1,
807                     block_start + 1);
808        if (ret < 0)
809            mlog_errno(ret);
810        else
811            ret = 0;
812    }
813
814    if (handle)
815        ocfs2_commit_trans(OCFS2_SB(inode->i_sb), handle);
816
817out_unlock:
818    unlock_page(page);
819    page_cache_release(page);
820out:
821    return ret;
822}
823
824/*
825 * Find the next range to zero. We do this in terms of bytes because
826 * that's what ocfs2_zero_extend() wants, and it is dealing with the
827 * pagecache. We may return multiple extents.
828 *
829 * zero_start and zero_end are ocfs2_zero_extend()s current idea of what
830 * needs to be zeroed. range_start and range_end return the next zeroing
831 * range. A subsequent call should pass the previous range_end as its
832 * zero_start. If range_end is 0, there's nothing to do.
833 *
834 * Unwritten extents are skipped over. Refcounted extents are CoWd.
835 */
836static int ocfs2_zero_extend_get_range(struct inode *inode,
837                       struct buffer_head *di_bh,
838                       u64 zero_start, u64 zero_end,
839                       u64 *range_start, u64 *range_end)
840{
841    int rc = 0, needs_cow = 0;
842    u32 p_cpos, zero_clusters = 0;
843    u32 zero_cpos =
844        zero_start >> OCFS2_SB(inode->i_sb)->s_clustersize_bits;
845    u32 last_cpos = ocfs2_clusters_for_bytes(inode->i_sb, zero_end);
846    unsigned int num_clusters = 0;
847    unsigned int ext_flags = 0;
848
849    while (zero_cpos < last_cpos) {
850        rc = ocfs2_get_clusters(inode, zero_cpos, &p_cpos,
851                    &num_clusters, &ext_flags);
852        if (rc) {
853            mlog_errno(rc);
854            goto out;
855        }
856
857        if (p_cpos && !(ext_flags & OCFS2_EXT_UNWRITTEN)) {
858            zero_clusters = num_clusters;
859            if (ext_flags & OCFS2_EXT_REFCOUNTED)
860                needs_cow = 1;
861            break;
862        }
863
864        zero_cpos += num_clusters;
865    }
866    if (!zero_clusters) {
867        *range_end = 0;
868        goto out;
869    }
870
871    while ((zero_cpos + zero_clusters) < last_cpos) {
872        rc = ocfs2_get_clusters(inode, zero_cpos + zero_clusters,
873                    &p_cpos, &num_clusters,
874                    &ext_flags);
875        if (rc) {
876            mlog_errno(rc);
877            goto out;
878        }
879
880        if (!p_cpos || (ext_flags & OCFS2_EXT_UNWRITTEN))
881            break;
882        if (ext_flags & OCFS2_EXT_REFCOUNTED)
883            needs_cow = 1;
884        zero_clusters += num_clusters;
885    }
886    if ((zero_cpos + zero_clusters) > last_cpos)
887        zero_clusters = last_cpos - zero_cpos;
888
889    if (needs_cow) {
890        rc = ocfs2_refcount_cow(inode, NULL, di_bh, zero_cpos,
891                    zero_clusters, UINT_MAX);
892        if (rc) {
893            mlog_errno(rc);
894            goto out;
895        }
896    }
897
898    *range_start = ocfs2_clusters_to_bytes(inode->i_sb, zero_cpos);
899    *range_end = ocfs2_clusters_to_bytes(inode->i_sb,
900                         zero_cpos + zero_clusters);
901
902out:
903    return rc;
904}
905
906/*
907 * Zero one range returned from ocfs2_zero_extend_get_range(). The caller
908 * has made sure that the entire range needs zeroing.
909 */
910static int ocfs2_zero_extend_range(struct inode *inode, u64 range_start,
911                   u64 range_end)
912{
913    int rc = 0;
914    u64 next_pos;
915    u64 zero_pos = range_start;
916
917    trace_ocfs2_zero_extend_range(
918            (unsigned long long)OCFS2_I(inode)->ip_blkno,
919            (unsigned long long)range_start,
920            (unsigned long long)range_end);
921    BUG_ON(range_start >= range_end);
922
923    while (zero_pos < range_end) {
924        next_pos = (zero_pos & PAGE_CACHE_MASK) + PAGE_CACHE_SIZE;
925        if (next_pos > range_end)
926            next_pos = range_end;
927        rc = ocfs2_write_zero_page(inode, zero_pos, next_pos);
928        if (rc < 0) {
929            mlog_errno(rc);
930            break;
931        }
932        zero_pos = next_pos;
933
934        /*
935         * Very large extends have the potential to lock up
936         * the cpu for extended periods of time.
937         */
938        cond_resched();
939    }
940
941    return rc;
942}
943
944int ocfs2_zero_extend(struct inode *inode, struct buffer_head *di_bh,
945              loff_t zero_to_size)
946{
947    int ret = 0;
948    u64 zero_start, range_start = 0, range_end = 0;
949    struct super_block *sb = inode->i_sb;
950
951    zero_start = ocfs2_align_bytes_to_blocks(sb, i_size_read(inode));
952    trace_ocfs2_zero_extend((unsigned long long)OCFS2_I(inode)->ip_blkno,
953                (unsigned long long)zero_start,
954                (unsigned long long)i_size_read(inode));
955    while (zero_start < zero_to_size) {
956        ret = ocfs2_zero_extend_get_range(inode, di_bh, zero_start,
957                          zero_to_size,
958                          &range_start,
959                          &range_end);
960        if (ret) {
961            mlog_errno(ret);
962            break;
963        }
964        if (!range_end)
965            break;
966        /* Trim the ends */
967        if (range_start < zero_start)
968            range_start = zero_start;
969        if (range_end > zero_to_size)
970            range_end = zero_to_size;
971
972        ret = ocfs2_zero_extend_range(inode, range_start,
973                          range_end);
974        if (ret) {
975            mlog_errno(ret);
976            break;
977        }
978        zero_start = range_end;
979    }
980
981    return ret;
982}
983
984int ocfs2_extend_no_holes(struct inode *inode, struct buffer_head *di_bh,
985              u64 new_i_size, u64 zero_to)
986{
987    int ret;
988    u32 clusters_to_add;
989    struct ocfs2_inode_info *oi = OCFS2_I(inode);
990
991    /*
992     * Only quota files call this without a bh, and they can't be
993     * refcounted.
994     */
995    BUG_ON(!di_bh && (oi->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL));
996    BUG_ON(!di_bh && !(oi->ip_flags & OCFS2_INODE_SYSTEM_FILE));
997
998    clusters_to_add = ocfs2_clusters_for_bytes(inode->i_sb, new_i_size);
999    if (clusters_to_add < oi->ip_clusters)
1000        clusters_to_add = 0;
1001    else
1002        clusters_to_add -= oi->ip_clusters;
1003
1004    if (clusters_to_add) {
1005        ret = __ocfs2_extend_allocation(inode, oi->ip_clusters,
1006                        clusters_to_add, 0);
1007        if (ret) {
1008            mlog_errno(ret);
1009            goto out;
1010        }
1011    }
1012
1013    /*
1014     * Call this even if we don't add any clusters to the tree. We
1015     * still need to zero the area between the old i_size and the
1016     * new i_size.
1017     */
1018    ret = ocfs2_zero_extend(inode, di_bh, zero_to);
1019    if (ret < 0)
1020        mlog_errno(ret);
1021
1022out:
1023    return ret;
1024}
1025
1026static int ocfs2_extend_file(struct inode *inode,
1027                 struct buffer_head *di_bh,
1028                 u64 new_i_size)
1029{
1030    int ret = 0;
1031    struct ocfs2_inode_info *oi = OCFS2_I(inode);
1032
1033    BUG_ON(!di_bh);
1034
1035    /* setattr sometimes calls us like this. */
1036    if (new_i_size == 0)
1037        goto out;
1038
1039    if (i_size_read(inode) == new_i_size)
1040        goto out;
1041    BUG_ON(new_i_size < i_size_read(inode));
1042
1043    /*
1044     * The alloc sem blocks people in read/write from reading our
1045     * allocation until we're done changing it. We depend on
1046     * i_mutex to block other extend/truncate calls while we're
1047     * here. We even have to hold it for sparse files because there
1048     * might be some tail zeroing.
1049     */
1050    down_write(&oi->ip_alloc_sem);
1051
1052    if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1053        /*
1054         * We can optimize small extends by keeping the inodes
1055         * inline data.
1056         */
1057        if (ocfs2_size_fits_inline_data(di_bh, new_i_size)) {
1058            up_write(&oi->ip_alloc_sem);
1059            goto out_update_size;
1060        }
1061
1062        ret = ocfs2_convert_inline_data_to_extents(inode, di_bh);
1063        if (ret) {
1064            up_write(&oi->ip_alloc_sem);
1065            mlog_errno(ret);
1066            goto out;
1067        }
1068    }
1069
1070    if (ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
1071        ret = ocfs2_zero_extend(inode, di_bh, new_i_size);
1072    else
1073        ret = ocfs2_extend_no_holes(inode, di_bh, new_i_size,
1074                        new_i_size);
1075
1076    up_write(&oi->ip_alloc_sem);
1077
1078    if (ret < 0) {
1079        mlog_errno(ret);
1080        goto out;
1081    }
1082
1083out_update_size:
1084    ret = ocfs2_simple_size_update(inode, di_bh, new_i_size);
1085    if (ret < 0)
1086        mlog_errno(ret);
1087
1088out:
1089    return ret;
1090}
1091
1092int ocfs2_setattr(struct dentry *dentry, struct iattr *attr)
1093{
1094    int status = 0, size_change;
1095    struct inode *inode = dentry->d_inode;
1096    struct super_block *sb = inode->i_sb;
1097    struct ocfs2_super *osb = OCFS2_SB(sb);
1098    struct buffer_head *bh = NULL;
1099    handle_t *handle = NULL;
1100    struct dquot *transfer_to[MAXQUOTAS] = { };
1101    int qtype;
1102
1103    trace_ocfs2_setattr(inode, dentry,
1104                (unsigned long long)OCFS2_I(inode)->ip_blkno,
1105                dentry->d_name.len, dentry->d_name.name,
1106                attr->ia_valid, attr->ia_mode,
1107                attr->ia_uid, attr->ia_gid);
1108
1109    /* ensuring we don't even attempt to truncate a symlink */
1110    if (S_ISLNK(inode->i_mode))
1111        attr->ia_valid &= ~ATTR_SIZE;
1112
1113#define OCFS2_VALID_ATTRS (ATTR_ATIME | ATTR_MTIME | ATTR_CTIME | ATTR_SIZE \
1114               | ATTR_GID | ATTR_UID | ATTR_MODE)
1115    if (!(attr->ia_valid & OCFS2_VALID_ATTRS))
1116        return 0;
1117
1118    status = inode_change_ok(inode, attr);
1119    if (status)
1120        return status;
1121
1122    if (is_quota_modification(inode, attr))
1123        dquot_initialize(inode);
1124    size_change = S_ISREG(inode->i_mode) && attr->ia_valid & ATTR_SIZE;
1125    if (size_change) {
1126        status = ocfs2_rw_lock(inode, 1);
1127        if (status < 0) {
1128            mlog_errno(status);
1129            goto bail;
1130        }
1131    }
1132
1133    status = ocfs2_inode_lock(inode, &bh, 1);
1134    if (status < 0) {
1135        if (status != -ENOENT)
1136            mlog_errno(status);
1137        goto bail_unlock_rw;
1138    }
1139
1140    if (size_change && attr->ia_size != i_size_read(inode)) {
1141        status = inode_newsize_ok(inode, attr->ia_size);
1142        if (status)
1143            goto bail_unlock;
1144
1145        if (i_size_read(inode) > attr->ia_size) {
1146            if (ocfs2_should_order_data(inode)) {
1147                status = ocfs2_begin_ordered_truncate(inode,
1148                                      attr->ia_size);
1149                if (status)
1150                    goto bail_unlock;
1151            }
1152            status = ocfs2_truncate_file(inode, bh, attr->ia_size);
1153        } else
1154            status = ocfs2_extend_file(inode, bh, attr->ia_size);
1155        if (status < 0) {
1156            if (status != -ENOSPC)
1157                mlog_errno(status);
1158            status = -ENOSPC;
1159            goto bail_unlock;
1160        }
1161    }
1162
1163    if ((attr->ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid) ||
1164        (attr->ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid)) {
1165        /*
1166         * Gather pointers to quota structures so that allocation /
1167         * freeing of quota structures happens here and not inside
1168         * dquot_transfer() where we have problems with lock ordering
1169         */
1170        if (attr->ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid
1171            && OCFS2_HAS_RO_COMPAT_FEATURE(sb,
1172            OCFS2_FEATURE_RO_COMPAT_USRQUOTA)) {
1173            transfer_to[USRQUOTA] = dqget(sb, attr->ia_uid,
1174                              USRQUOTA);
1175            if (!transfer_to[USRQUOTA]) {
1176                status = -ESRCH;
1177                goto bail_unlock;
1178            }
1179        }
1180        if (attr->ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid
1181            && OCFS2_HAS_RO_COMPAT_FEATURE(sb,
1182            OCFS2_FEATURE_RO_COMPAT_GRPQUOTA)) {
1183            transfer_to[GRPQUOTA] = dqget(sb, attr->ia_gid,
1184                              GRPQUOTA);
1185            if (!transfer_to[GRPQUOTA]) {
1186                status = -ESRCH;
1187                goto bail_unlock;
1188            }
1189        }
1190        handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS +
1191                       2 * ocfs2_quota_trans_credits(sb));
1192        if (IS_ERR(handle)) {
1193            status = PTR_ERR(handle);
1194            mlog_errno(status);
1195            goto bail_unlock;
1196        }
1197        status = __dquot_transfer(inode, transfer_to);
1198        if (status < 0)
1199            goto bail_commit;
1200    } else {
1201        handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1202        if (IS_ERR(handle)) {
1203            status = PTR_ERR(handle);
1204            mlog_errno(status);
1205            goto bail_unlock;
1206        }
1207    }
1208
1209    /*
1210     * This will intentionally not wind up calling truncate_setsize(),
1211     * since all the work for a size change has been done above.
1212     * Otherwise, we could get into problems with truncate as
1213     * ip_alloc_sem is used there to protect against i_size
1214     * changes.
1215     *
1216     * XXX: this means the conditional below can probably be removed.
1217     */
1218    if ((attr->ia_valid & ATTR_SIZE) &&
1219        attr->ia_size != i_size_read(inode)) {
1220        status = vmtruncate(inode, attr->ia_size);
1221        if (status) {
1222            mlog_errno(status);
1223            goto bail_commit;
1224        }
1225    }
1226
1227    setattr_copy(inode, attr);
1228    mark_inode_dirty(inode);
1229
1230    status = ocfs2_mark_inode_dirty(handle, inode, bh);
1231    if (status < 0)
1232        mlog_errno(status);
1233
1234bail_commit:
1235    ocfs2_commit_trans(osb, handle);
1236bail_unlock:
1237    ocfs2_inode_unlock(inode, 1);
1238bail_unlock_rw:
1239    if (size_change)
1240        ocfs2_rw_unlock(inode, 1);
1241bail:
1242    brelse(bh);
1243
1244    /* Release quota pointers in case we acquired them */
1245    for (qtype = 0; qtype < MAXQUOTAS; qtype++)
1246        dqput(transfer_to[qtype]);
1247
1248    if (!status && attr->ia_valid & ATTR_MODE) {
1249        status = ocfs2_acl_chmod(inode);
1250        if (status < 0)
1251            mlog_errno(status);
1252    }
1253
1254    return status;
1255}
1256
1257int ocfs2_getattr(struct vfsmount *mnt,
1258          struct dentry *dentry,
1259          struct kstat *stat)
1260{
1261    struct inode *inode = dentry->d_inode;
1262    struct super_block *sb = dentry->d_inode->i_sb;
1263    struct ocfs2_super *osb = sb->s_fs_info;
1264    int err;
1265
1266    err = ocfs2_inode_revalidate(dentry);
1267    if (err) {
1268        if (err != -ENOENT)
1269            mlog_errno(err);
1270        goto bail;
1271    }
1272
1273    generic_fillattr(inode, stat);
1274
1275    /* We set the blksize from the cluster size for performance */
1276    stat->blksize = osb->s_clustersize;
1277
1278bail:
1279    return err;
1280}
1281
1282int ocfs2_permission(struct inode *inode, int mask, unsigned int flags)
1283{
1284    int ret;
1285
1286    if (flags & IPERM_FLAG_RCU)
1287        return -ECHILD;
1288
1289    ret = ocfs2_inode_lock(inode, NULL, 0);
1290    if (ret) {
1291        if (ret != -ENOENT)
1292            mlog_errno(ret);
1293        goto out;
1294    }
1295
1296    ret = generic_permission(inode, mask, flags, ocfs2_check_acl);
1297
1298    ocfs2_inode_unlock(inode, 0);
1299out:
1300    return ret;
1301}
1302
1303static int __ocfs2_write_remove_suid(struct inode *inode,
1304                     struct buffer_head *bh)
1305{
1306    int ret;
1307    handle_t *handle;
1308    struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1309    struct ocfs2_dinode *di;
1310
1311    trace_ocfs2_write_remove_suid(
1312            (unsigned long long)OCFS2_I(inode)->ip_blkno,
1313            inode->i_mode);
1314
1315    handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1316    if (IS_ERR(handle)) {
1317        ret = PTR_ERR(handle);
1318        mlog_errno(ret);
1319        goto out;
1320    }
1321
1322    ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), bh,
1323                      OCFS2_JOURNAL_ACCESS_WRITE);
1324    if (ret < 0) {
1325        mlog_errno(ret);
1326        goto out_trans;
1327    }
1328
1329    inode->i_mode &= ~S_ISUID;
1330    if ((inode->i_mode & S_ISGID) && (inode->i_mode & S_IXGRP))
1331        inode->i_mode &= ~S_ISGID;
1332
1333    di = (struct ocfs2_dinode *) bh->b_data;
1334    di->i_mode = cpu_to_le16(inode->i_mode);
1335
1336    ocfs2_journal_dirty(handle, bh);
1337
1338out_trans:
1339    ocfs2_commit_trans(osb, handle);
1340out:
1341    return ret;
1342}
1343
1344/*
1345 * Will look for holes and unwritten extents in the range starting at
1346 * pos for count bytes (inclusive).
1347 */
1348static int ocfs2_check_range_for_holes(struct inode *inode, loff_t pos,
1349                       size_t count)
1350{
1351    int ret = 0;
1352    unsigned int extent_flags;
1353    u32 cpos, clusters, extent_len, phys_cpos;
1354    struct super_block *sb = inode->i_sb;
1355
1356    cpos = pos >> OCFS2_SB(sb)->s_clustersize_bits;
1357    clusters = ocfs2_clusters_for_bytes(sb, pos + count) - cpos;
1358
1359    while (clusters) {
1360        ret = ocfs2_get_clusters(inode, cpos, &phys_cpos, &extent_len,
1361                     &extent_flags);
1362        if (ret < 0) {
1363            mlog_errno(ret);
1364            goto out;
1365        }
1366
1367        if (phys_cpos == 0 || (extent_flags & OCFS2_EXT_UNWRITTEN)) {
1368            ret = 1;
1369            break;
1370        }
1371
1372        if (extent_len > clusters)
1373            extent_len = clusters;
1374
1375        clusters -= extent_len;
1376        cpos += extent_len;
1377    }
1378out:
1379    return ret;
1380}
1381
1382static int ocfs2_write_remove_suid(struct inode *inode)
1383{
1384    int ret;
1385    struct buffer_head *bh = NULL;
1386
1387    ret = ocfs2_read_inode_block(inode, &bh);
1388    if (ret < 0) {
1389        mlog_errno(ret);
1390        goto out;
1391    }
1392
1393    ret = __ocfs2_write_remove_suid(inode, bh);
1394out:
1395    brelse(bh);
1396    return ret;
1397}
1398
1399/*
1400 * Allocate enough extents to cover the region starting at byte offset
1401 * start for len bytes. Existing extents are skipped, any extents
1402 * added are marked as "unwritten".
1403 */
1404static int ocfs2_allocate_unwritten_extents(struct inode *inode,
1405                        u64 start, u64 len)
1406{
1407    int ret;
1408    u32 cpos, phys_cpos, clusters, alloc_size;
1409    u64 end = start + len;
1410    struct buffer_head *di_bh = NULL;
1411
1412    if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1413        ret = ocfs2_read_inode_block(inode, &di_bh);
1414        if (ret) {
1415            mlog_errno(ret);
1416            goto out;
1417        }
1418
1419        /*
1420         * Nothing to do if the requested reservation range
1421         * fits within the inode.
1422         */
1423        if (ocfs2_size_fits_inline_data(di_bh, end))
1424            goto out;
1425
1426        ret = ocfs2_convert_inline_data_to_extents(inode, di_bh);
1427        if (ret) {
1428            mlog_errno(ret);
1429            goto out;
1430        }
1431    }
1432
1433    /*
1434     * We consider both start and len to be inclusive.
1435     */
1436    cpos = start >> OCFS2_SB(inode->i_sb)->s_clustersize_bits;
1437    clusters = ocfs2_clusters_for_bytes(inode->i_sb, start + len);
1438    clusters -= cpos;
1439
1440    while (clusters) {
1441        ret = ocfs2_get_clusters(inode, cpos, &phys_cpos,
1442                     &alloc_size, NULL);
1443        if (ret) {
1444            mlog_errno(ret);
1445            goto out;
1446        }
1447
1448        /*
1449         * Hole or existing extent len can be arbitrary, so
1450         * cap it to our own allocation request.
1451         */
1452        if (alloc_size > clusters)
1453            alloc_size = clusters;
1454
1455        if (phys_cpos) {
1456            /*
1457             * We already have an allocation at this
1458             * region so we can safely skip it.
1459             */
1460            goto next;
1461        }
1462
1463        ret = __ocfs2_extend_allocation(inode, cpos, alloc_size, 1);
1464        if (ret) {
1465            if (ret != -ENOSPC)
1466                mlog_errno(ret);
1467            goto out;
1468        }
1469
1470next:
1471        cpos += alloc_size;
1472        clusters -= alloc_size;
1473    }
1474
1475    ret = 0;
1476out:
1477
1478    brelse(di_bh);
1479    return ret;
1480}
1481
1482/*
1483 * Truncate a byte range, avoiding pages within partial clusters. This
1484 * preserves those pages for the zeroing code to write to.
1485 */
1486static void ocfs2_truncate_cluster_pages(struct inode *inode, u64 byte_start,
1487                     u64 byte_len)
1488{
1489    struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1490    loff_t start, end;
1491    struct address_space *mapping = inode->i_mapping;
1492
1493    start = (loff_t)ocfs2_align_bytes_to_clusters(inode->i_sb, byte_start);
1494    end = byte_start + byte_len;
1495    end = end & ~(osb->s_clustersize - 1);
1496
1497    if (start < end) {
1498        unmap_mapping_range(mapping, start, end - start, 0);
1499        truncate_inode_pages_range(mapping, start, end - 1);
1500    }
1501}
1502
1503static int ocfs2_zero_partial_clusters(struct inode *inode,
1504                       u64 start, u64 len)
1505{
1506    int ret = 0;
1507    u64 tmpend, end = start + len;
1508    struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1509    unsigned int csize = osb->s_clustersize;
1510    handle_t *handle;
1511
1512    /*
1513     * The "start" and "end" values are NOT necessarily part of
1514     * the range whose allocation is being deleted. Rather, this
1515     * is what the user passed in with the request. We must zero
1516     * partial clusters here. There's no need to worry about
1517     * physical allocation - the zeroing code knows to skip holes.
1518     */
1519    trace_ocfs2_zero_partial_clusters(
1520        (unsigned long long)OCFS2_I(inode)->ip_blkno,
1521        (unsigned long long)start, (unsigned long long)end);
1522
1523    /*
1524     * If both edges are on a cluster boundary then there's no
1525     * zeroing required as the region is part of the allocation to
1526     * be truncated.
1527     */
1528    if ((start & (csize - 1)) == 0 && (end & (csize - 1)) == 0)
1529        goto out;
1530
1531    handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1532    if (IS_ERR(handle)) {
1533        ret = PTR_ERR(handle);
1534        mlog_errno(ret);
1535        goto out;
1536    }
1537
1538    /*
1539     * We want to get the byte offset of the end of the 1st cluster.
1540     */
1541    tmpend = (u64)osb->s_clustersize + (start & ~(osb->s_clustersize - 1));
1542    if (tmpend > end)
1543        tmpend = end;
1544
1545    trace_ocfs2_zero_partial_clusters_range1((unsigned long long)start,
1546                         (unsigned long long)tmpend);
1547
1548    ret = ocfs2_zero_range_for_truncate(inode, handle, start, tmpend);
1549    if (ret)
1550        mlog_errno(ret);
1551
1552    if (tmpend < end) {
1553        /*
1554         * This may make start and end equal, but the zeroing
1555         * code will skip any work in that case so there's no
1556         * need to catch it up here.
1557         */
1558        start = end & ~(osb->s_clustersize - 1);
1559
1560        trace_ocfs2_zero_partial_clusters_range2(
1561            (unsigned long long)start, (unsigned long long)end);
1562
1563        ret = ocfs2_zero_range_for_truncate(inode, handle, start, end);
1564        if (ret)
1565            mlog_errno(ret);
1566    }
1567
1568    ocfs2_commit_trans(osb, handle);
1569out:
1570    return ret;
1571}
1572
1573static int ocfs2_find_rec(struct ocfs2_extent_list *el, u32 pos)
1574{
1575    int i;
1576    struct ocfs2_extent_rec *rec = NULL;
1577
1578    for (i = le16_to_cpu(el->l_next_free_rec) - 1; i >= 0; i--) {
1579
1580        rec = &el->l_recs[i];
1581
1582        if (le32_to_cpu(rec->e_cpos) < pos)
1583            break;
1584    }
1585
1586    return i;
1587}
1588
1589/*
1590 * Helper to calculate the punching pos and length in one run, we handle the
1591 * following three cases in order:
1592 *
1593 * - remove the entire record
1594 * - remove a partial record
1595 * - no record needs to be removed (hole-punching completed)
1596*/
1597static void ocfs2_calc_trunc_pos(struct inode *inode,
1598                 struct ocfs2_extent_list *el,
1599                 struct ocfs2_extent_rec *rec,
1600                 u32 trunc_start, u32 *trunc_cpos,
1601                 u32 *trunc_len, u32 *trunc_end,
1602                 u64 *blkno, int *done)
1603{
1604    int ret = 0;
1605    u32 coff, range;
1606
1607    range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
1608
1609    if (le32_to_cpu(rec->e_cpos) >= trunc_start) {
1610        /*
1611         * remove an entire extent record.
1612         */
1613        *trunc_cpos = le32_to_cpu(rec->e_cpos);
1614        /*
1615         * Skip holes if any.
1616         */
1617        if (range < *trunc_end)
1618            *trunc_end = range;
1619        *trunc_len = *trunc_end - le32_to_cpu(rec->e_cpos);
1620        *blkno = le64_to_cpu(rec->e_blkno);
1621        *trunc_end = le32_to_cpu(rec->e_cpos);
1622    } else if (range > trunc_start) {
1623        /*
1624         * remove a partial extent record, which means we're
1625         * removing the last extent record.
1626         */
1627        *trunc_cpos = trunc_start;
1628        /*
1629         * skip hole if any.
1630         */
1631        if (range < *trunc_end)
1632            *trunc_end = range;
1633        *trunc_len = *trunc_end - trunc_start;
1634        coff = trunc_start - le32_to_cpu(rec->e_cpos);
1635        *blkno = le64_to_cpu(rec->e_blkno) +
1636                ocfs2_clusters_to_blocks(inode->i_sb, coff);
1637        *trunc_end = trunc_start;
1638    } else {
1639        /*
1640         * It may have two following possibilities:
1641         *
1642         * - last record has been removed
1643         * - trunc_start was within a hole
1644         *
1645         * both two cases mean the completion of hole punching.
1646         */
1647        ret = 1;
1648    }
1649
1650    *done = ret;
1651}
1652
1653static int ocfs2_remove_inode_range(struct inode *inode,
1654                    struct buffer_head *di_bh, u64 byte_start,
1655                    u64 byte_len)
1656{
1657    int ret = 0, flags = 0, done = 0, i;
1658    u32 trunc_start, trunc_len, trunc_end, trunc_cpos, phys_cpos;
1659    u32 cluster_in_el;
1660    struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1661    struct ocfs2_cached_dealloc_ctxt dealloc;
1662    struct address_space *mapping = inode->i_mapping;
1663    struct ocfs2_extent_tree et;
1664    struct ocfs2_path *path = NULL;
1665    struct ocfs2_extent_list *el = NULL;
1666    struct ocfs2_extent_rec *rec = NULL;
1667    struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
1668    u64 blkno, refcount_loc = le64_to_cpu(di->i_refcount_loc);
1669
1670    ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), di_bh);
1671    ocfs2_init_dealloc_ctxt(&dealloc);
1672
1673    trace_ocfs2_remove_inode_range(
1674            (unsigned long long)OCFS2_I(inode)->ip_blkno,
1675            (unsigned long long)byte_start,
1676            (unsigned long long)byte_len);
1677
1678    if (byte_len == 0)
1679        return 0;
1680
1681    if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1682        ret = ocfs2_truncate_inline(inode, di_bh, byte_start,
1683                        byte_start + byte_len, 0);
1684        if (ret) {
1685            mlog_errno(ret);
1686            goto out;
1687        }
1688        /*
1689         * There's no need to get fancy with the page cache
1690         * truncate of an inline-data inode. We're talking
1691         * about less than a page here, which will be cached
1692         * in the dinode buffer anyway.
1693         */
1694        unmap_mapping_range(mapping, 0, 0, 0);
1695        truncate_inode_pages(mapping, 0);
1696        goto out;
1697    }
1698
1699    /*
1700     * For reflinks, we may need to CoW 2 clusters which might be
1701     * partially zero'd later, if hole's start and end offset were
1702     * within one cluster(means is not exactly aligned to clustersize).
1703     */
1704
1705    if (OCFS2_I(inode)->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL) {
1706
1707        ret = ocfs2_cow_file_pos(inode, di_bh, byte_start);
1708        if (ret) {
1709            mlog_errno(ret);
1710            goto out;
1711        }
1712
1713        ret = ocfs2_cow_file_pos(inode, di_bh, byte_start + byte_len);
1714        if (ret) {
1715            mlog_errno(ret);
1716            goto out;
1717        }
1718    }
1719
1720    trunc_start = ocfs2_clusters_for_bytes(osb->sb, byte_start);
1721    trunc_end = (byte_start + byte_len) >> osb->s_clustersize_bits;
1722    cluster_in_el = trunc_end;
1723
1724    ret = ocfs2_zero_partial_clusters(inode, byte_start, byte_len);
1725    if (ret) {
1726        mlog_errno(ret);
1727        goto out;
1728    }
1729
1730    path = ocfs2_new_path_from_et(&et);
1731    if (!path) {
1732        ret = -ENOMEM;
1733        mlog_errno(ret);
1734        goto out;
1735    }
1736
1737    while (trunc_end > trunc_start) {
1738
1739        ret = ocfs2_find_path(INODE_CACHE(inode), path,
1740                      cluster_in_el);
1741        if (ret) {
1742            mlog_errno(ret);
1743            goto out;
1744        }
1745
1746        el = path_leaf_el(path);
1747
1748        i = ocfs2_find_rec(el, trunc_end);
1749        /*
1750         * Need to go to previous extent block.
1751         */
1752        if (i < 0) {
1753            if (path->p_tree_depth == 0)
1754                break;
1755
1756            ret = ocfs2_find_cpos_for_left_leaf(inode->i_sb,
1757                                path,
1758                                &cluster_in_el);
1759            if (ret) {
1760                mlog_errno(ret);
1761                goto out;
1762            }
1763
1764            /*
1765             * We've reached the leftmost extent block,
1766             * it's safe to leave.
1767             */
1768            if (cluster_in_el == 0)
1769                break;
1770
1771            /*
1772             * The 'pos' searched for previous extent block is
1773             * always one cluster less than actual trunc_end.
1774             */
1775            trunc_end = cluster_in_el + 1;
1776
1777            ocfs2_reinit_path(path, 1);
1778
1779            continue;
1780
1781        } else
1782            rec = &el->l_recs[i];
1783
1784        ocfs2_calc_trunc_pos(inode, el, rec, trunc_start, &trunc_cpos,
1785                     &trunc_len, &trunc_end, &blkno, &done);
1786        if (done)
1787            break;
1788
1789        flags = rec->e_flags;
1790        phys_cpos = ocfs2_blocks_to_clusters(inode->i_sb, blkno);
1791
1792        ret = ocfs2_remove_btree_range(inode, &et, trunc_cpos,
1793                           phys_cpos, trunc_len, flags,
1794                           &dealloc, refcount_loc);
1795        if (ret < 0) {
1796            mlog_errno(ret);
1797            goto out;
1798        }
1799
1800        cluster_in_el = trunc_end;
1801
1802        ocfs2_reinit_path(path, 1);
1803    }
1804
1805    ocfs2_truncate_cluster_pages(inode, byte_start, byte_len);
1806
1807out:
1808    ocfs2_schedule_truncate_log_flush(osb, 1);
1809    ocfs2_run_deallocs(osb, &dealloc);
1810
1811    return ret;
1812}
1813
1814/*
1815 * Parts of this function taken from xfs_change_file_space()
1816 */
1817static int __ocfs2_change_file_space(struct file *file, struct inode *inode,
1818                     loff_t f_pos, unsigned int cmd,
1819                     struct ocfs2_space_resv *sr,
1820                     int change_size)
1821{
1822    int ret;
1823    s64 llen;
1824    loff_t size;
1825    struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1826    struct buffer_head *di_bh = NULL;
1827    handle_t *handle;
1828    unsigned long long max_off = inode->i_sb->s_maxbytes;
1829
1830    if (ocfs2_is_hard_readonly(osb) || ocfs2_is_soft_readonly(osb))
1831        return -EROFS;
1832
1833    mutex_lock(&inode->i_mutex);
1834
1835    /*
1836     * This prevents concurrent writes on other nodes
1837     */
1838    ret = ocfs2_rw_lock(inode, 1);
1839    if (ret) {
1840        mlog_errno(ret);
1841        goto out;
1842    }
1843
1844    ret = ocfs2_inode_lock(inode, &di_bh, 1);
1845    if (ret) {
1846        mlog_errno(ret);
1847        goto out_rw_unlock;
1848    }
1849
1850    if (inode->i_flags & (S_IMMUTABLE|S_APPEND)) {
1851        ret = -EPERM;
1852        goto out_inode_unlock;
1853    }
1854
1855    switch (sr->l_whence) {
1856    case 0: /*SEEK_SET*/
1857        break;
1858    case 1: /*SEEK_CUR*/
1859        sr->l_start += f_pos;
1860        break;
1861    case 2: /*SEEK_END*/
1862        sr->l_start += i_size_read(inode);
1863        break;
1864    default:
1865        ret = -EINVAL;
1866        goto out_inode_unlock;
1867    }
1868    sr->l_whence = 0;
1869
1870    llen = sr->l_len > 0 ? sr->l_len - 1 : sr->l_len;
1871
1872    if (sr->l_start < 0
1873        || sr->l_start > max_off
1874        || (sr->l_start + llen) < 0
1875        || (sr->l_start + llen) > max_off) {
1876        ret = -EINVAL;
1877        goto out_inode_unlock;
1878    }
1879    size = sr->l_start + sr->l_len;
1880
1881    if (cmd == OCFS2_IOC_RESVSP || cmd == OCFS2_IOC_RESVSP64) {
1882        if (sr->l_len <= 0) {
1883            ret = -EINVAL;
1884            goto out_inode_unlock;
1885        }
1886    }
1887
1888    if (file && should_remove_suid(file->f_path.dentry)) {
1889        ret = __ocfs2_write_remove_suid(inode, di_bh);
1890        if (ret) {
1891            mlog_errno(ret);
1892            goto out_inode_unlock;
1893        }
1894    }
1895
1896    down_write(&OCFS2_I(inode)->ip_alloc_sem);
1897    switch (cmd) {
1898    case OCFS2_IOC_RESVSP:
1899    case OCFS2_IOC_RESVSP64:
1900        /*
1901         * This takes unsigned offsets, but the signed ones we
1902         * pass have been checked against overflow above.
1903         */
1904        ret = ocfs2_allocate_unwritten_extents(inode, sr->l_start,
1905                               sr->l_len);
1906        break;
1907    case OCFS2_IOC_UNRESVSP:
1908    case OCFS2_IOC_UNRESVSP64:
1909        ret = ocfs2_remove_inode_range(inode, di_bh, sr->l_start,
1910                           sr->l_len);
1911        break;
1912    default:
1913        ret = -EINVAL;
1914    }
1915    up_write(&OCFS2_I(inode)->ip_alloc_sem);
1916    if (ret) {
1917        mlog_errno(ret);
1918        goto out_inode_unlock;
1919    }
1920
1921    /*
1922     * We update c/mtime for these changes
1923     */
1924    handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1925    if (IS_ERR(handle)) {
1926        ret = PTR_ERR(handle);
1927        mlog_errno(ret);
1928        goto out_inode_unlock;
1929    }
1930
1931    if (change_size && i_size_read(inode) < size)
1932        i_size_write(inode, size);
1933
1934    inode->i_ctime = inode->i_mtime = CURRENT_TIME;
1935    ret = ocfs2_mark_inode_dirty(handle, inode, di_bh);
1936    if (ret < 0)
1937        mlog_errno(ret);
1938
1939    ocfs2_commit_trans(osb, handle);
1940
1941out_inode_unlock:
1942    brelse(di_bh);
1943    ocfs2_inode_unlock(inode, 1);
1944out_rw_unlock:
1945    ocfs2_rw_unlock(inode, 1);
1946
1947out:
1948    mutex_unlock(&inode->i_mutex);
1949    return ret;
1950}
1951
1952int ocfs2_change_file_space(struct file *file, unsigned int cmd,
1953                struct ocfs2_space_resv *sr)
1954{
1955    struct inode *inode = file->f_path.dentry->d_inode;
1956    struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1957
1958    if ((cmd == OCFS2_IOC_RESVSP || cmd == OCFS2_IOC_RESVSP64) &&
1959        !ocfs2_writes_unwritten_extents(osb))
1960        return -ENOTTY;
1961    else if ((cmd == OCFS2_IOC_UNRESVSP || cmd == OCFS2_IOC_UNRESVSP64) &&
1962         !ocfs2_sparse_alloc(osb))
1963        return -ENOTTY;
1964
1965    if (!S_ISREG(inode->i_mode))
1966        return -EINVAL;
1967
1968    if (!(file->f_mode & FMODE_WRITE))
1969        return -EBADF;
1970
1971    return __ocfs2_change_file_space(file, inode, file->f_pos, cmd, sr, 0);
1972}
1973
1974static long ocfs2_fallocate(struct file *file, int mode, loff_t offset,
1975                loff_t len)
1976{
1977    struct inode *inode = file->f_path.dentry->d_inode;
1978    struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1979    struct ocfs2_space_resv sr;
1980    int change_size = 1;
1981    int cmd = OCFS2_IOC_RESVSP64;
1982
1983    if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
1984        return -EOPNOTSUPP;
1985    if (!ocfs2_writes_unwritten_extents(osb))
1986        return -EOPNOTSUPP;
1987
1988    if (mode & FALLOC_FL_KEEP_SIZE)
1989        change_size = 0;
1990
1991    if (mode & FALLOC_FL_PUNCH_HOLE)
1992        cmd = OCFS2_IOC_UNRESVSP64;
1993
1994    sr.l_whence = 0;
1995    sr.l_start = (s64)offset;
1996    sr.l_len = (s64)len;
1997
1998    return __ocfs2_change_file_space(NULL, inode, offset, cmd, &sr,
1999                     change_size);
2000}
2001
2002int ocfs2_check_range_for_refcount(struct inode *inode, loff_t pos,
2003                   size_t count)
2004{
2005    int ret = 0;
2006    unsigned int extent_flags;
2007    u32 cpos, clusters, extent_len, phys_cpos;
2008    struct super_block *sb = inode->i_sb;
2009
2010    if (!ocfs2_refcount_tree(OCFS2_SB(inode->i_sb)) ||
2011        !(OCFS2_I(inode)->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL) ||
2012        OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
2013        return 0;
2014
2015    cpos = pos >> OCFS2_SB(sb)->s_clustersize_bits;
2016    clusters = ocfs2_clusters_for_bytes(sb, pos + count) - cpos;
2017
2018    while (clusters) {
2019        ret = ocfs2_get_clusters(inode, cpos, &phys_cpos, &extent_len,
2020                     &extent_flags);
2021        if (ret < 0) {
2022            mlog_errno(ret);
2023            goto out;
2024        }
2025
2026        if (phys_cpos && (extent_flags & OCFS2_EXT_REFCOUNTED)) {
2027            ret = 1;
2028            break;
2029        }
2030
2031        if (extent_len > clusters)
2032            extent_len = clusters;
2033
2034        clusters -= extent_len;
2035        cpos += extent_len;
2036    }
2037out:
2038    return ret;
2039}
2040
2041static int ocfs2_prepare_inode_for_refcount(struct inode *inode,
2042                        struct file *file,
2043                        loff_t pos, size_t count,
2044                        int *meta_level)
2045{
2046    int ret;
2047    struct buffer_head *di_bh = NULL;
2048    u32 cpos = pos >> OCFS2_SB(inode->i_sb)->s_clustersize_bits;
2049    u32 clusters =
2050        ocfs2_clusters_for_bytes(inode->i_sb, pos + count) - cpos;
2051
2052    ret = ocfs2_inode_lock(inode, &di_bh, 1);
2053    if (ret) {
2054        mlog_errno(ret);
2055        goto out;
2056    }
2057
2058    *meta_level = 1;
2059
2060    ret = ocfs2_refcount_cow(inode, file, di_bh, cpos, clusters, UINT_MAX);
2061    if (ret)
2062        mlog_errno(ret);
2063out:
2064    brelse(di_bh);
2065    return ret;
2066}
2067
2068static int ocfs2_prepare_inode_for_write(struct file *file,
2069                     loff_t *ppos,
2070                     size_t count,
2071                     int appending,
2072                     int *direct_io,
2073                     int *has_refcount)
2074{
2075    int ret = 0, meta_level = 0;
2076    struct dentry *dentry = file->f_path.dentry;
2077    struct inode *inode = dentry->d_inode;
2078    loff_t saved_pos = 0, end;
2079
2080    /*
2081     * We start with a read level meta lock and only jump to an ex
2082     * if we need to make modifications here.
2083     */
2084    for(;;) {
2085        ret = ocfs2_inode_lock(inode, NULL, meta_level);
2086        if (ret < 0) {
2087            meta_level = -1;
2088            mlog_errno(ret);
2089            goto out;
2090        }
2091
2092        /* Clear suid / sgid if necessary. We do this here
2093         * instead of later in the write path because
2094         * remove_suid() calls ->setattr without any hint that
2095         * we may have already done our cluster locking. Since
2096         * ocfs2_setattr() *must* take cluster locks to
2097         * proceeed, this will lead us to recursively lock the
2098         * inode. There's also the dinode i_size state which
2099         * can be lost via setattr during extending writes (we
2100         * set inode->i_size at the end of a write. */
2101        if (should_remove_suid(dentry)) {
2102            if (meta_level == 0) {
2103                ocfs2_inode_unlock(inode, meta_level);
2104                meta_level = 1;
2105                continue;
2106            }
2107
2108            ret = ocfs2_write_remove_suid(inode);
2109            if (ret < 0) {
2110                mlog_errno(ret);
2111                goto out_unlock;
2112            }
2113        }
2114
2115        /* work on a copy of ppos until we're sure that we won't have
2116         * to recalculate it due to relocking. */
2117        if (appending)
2118            saved_pos = i_size_read(inode);
2119        else
2120            saved_pos = *ppos;
2121
2122        end = saved_pos + count;
2123
2124        ret = ocfs2_check_range_for_refcount(inode, saved_pos, count);
2125        if (ret == 1) {
2126            ocfs2_inode_unlock(inode, meta_level);
2127            meta_level = -1;
2128
2129            ret = ocfs2_prepare_inode_for_refcount(inode,
2130                                   file,
2131                                   saved_pos,
2132                                   count,
2133                                   &meta_level);
2134            if (has_refcount)
2135                *has_refcount = 1;
2136            if (direct_io)
2137                *direct_io = 0;
2138        }
2139
2140        if (ret < 0) {
2141            mlog_errno(ret);
2142            goto out_unlock;
2143        }
2144
2145        /*
2146         * Skip the O_DIRECT checks if we don't need
2147         * them.
2148         */
2149        if (!direct_io || !(*direct_io))
2150            break;
2151
2152        /*
2153         * There's no sane way to do direct writes to an inode
2154         * with inline data.
2155         */
2156        if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
2157            *direct_io = 0;
2158            break;
2159        }
2160
2161        /*
2162         * Allowing concurrent direct writes means
2163         * i_size changes wouldn't be synchronized, so
2164         * one node could wind up truncating another
2165         * nodes writes.
2166         */
2167        if (end > i_size_read(inode)) {
2168            *direct_io = 0;
2169            break;
2170        }
2171
2172        /*
2173         * We don't fill holes during direct io, so
2174         * check for them here. If any are found, the
2175         * caller will have to retake some cluster
2176         * locks and initiate the io as buffered.
2177         */
2178        ret = ocfs2_check_range_for_holes(inode, saved_pos, count);
2179        if (ret == 1) {
2180            *direct_io = 0;
2181            ret = 0;
2182        } else if (ret < 0)
2183            mlog_errno(ret);
2184        break;
2185    }
2186
2187    if (appending)
2188        *ppos = saved_pos;
2189
2190out_unlock:
2191    trace_ocfs2_prepare_inode_for_write(OCFS2_I(inode)->ip_blkno,
2192                        saved_pos, appending, count,
2193                        direct_io, has_refcount);
2194
2195    if (meta_level >= 0)
2196        ocfs2_inode_unlock(inode, meta_level);
2197
2198out:
2199    return ret;
2200}
2201
2202static ssize_t ocfs2_file_aio_write(struct kiocb *iocb,
2203                    const struct iovec *iov,
2204                    unsigned long nr_segs,
2205                    loff_t pos)
2206{
2207    int ret, direct_io, appending, rw_level, have_alloc_sem = 0;
2208    int can_do_direct, has_refcount = 0;
2209    ssize_t written = 0;
2210    size_t ocount; /* original count */
2211    size_t count; /* after file limit checks */
2212    loff_t old_size, *ppos = &iocb->ki_pos;
2213    u32 old_clusters;
2214    struct file *file = iocb->ki_filp;
2215    struct inode *inode = file->f_path.dentry->d_inode;
2216    struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2217    int full_coherency = !(osb->s_mount_opt &
2218                   OCFS2_MOUNT_COHERENCY_BUFFERED);
2219
2220    trace_ocfs2_file_aio_write(inode, file, file->f_path.dentry,
2221        (unsigned long long)OCFS2_I(inode)->ip_blkno,
2222        file->f_path.dentry->d_name.len,
2223        file->f_path.dentry->d_name.name,
2224        (unsigned int)nr_segs);
2225
2226    if (iocb->ki_left == 0)
2227        return 0;
2228
2229    vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
2230
2231    appending = file->f_flags & O_APPEND ? 1 : 0;
2232    direct_io = file->f_flags & O_DIRECT ? 1 : 0;
2233
2234    mutex_lock(&inode->i_mutex);
2235
2236    ocfs2_iocb_clear_sem_locked(iocb);
2237
2238relock:
2239    /* to match setattr's i_mutex -> i_alloc_sem -> rw_lock ordering */
2240    if (direct_io) {
2241        down_read(&inode->i_alloc_sem);
2242        have_alloc_sem = 1;
2243        /* communicate with ocfs2_dio_end_io */
2244        ocfs2_iocb_set_sem_locked(iocb);
2245    }
2246
2247    /*
2248     * Concurrent O_DIRECT writes are allowed with
2249     * mount_option "coherency=buffered".
2250     */
2251    rw_level = (!direct_io || full_coherency);
2252
2253    ret = ocfs2_rw_lock(inode, rw_level);
2254    if (ret < 0) {
2255        mlog_errno(ret);
2256        goto out_sems;
2257    }
2258
2259    /*
2260     * O_DIRECT writes with "coherency=full" need to take EX cluster
2261     * inode_lock to guarantee coherency.
2262     */
2263    if (direct_io && full_coherency) {
2264        /*
2265         * We need to take and drop the inode lock to force
2266         * other nodes to drop their caches. Buffered I/O
2267         * already does this in write_begin().
2268         */
2269        ret = ocfs2_inode_lock(inode, NULL, 1);
2270        if (ret < 0) {
2271            mlog_errno(ret);
2272            goto out_sems;
2273        }
2274
2275        ocfs2_inode_unlock(inode, 1);
2276    }
2277
2278    can_do_direct = direct_io;
2279    ret = ocfs2_prepare_inode_for_write(file, ppos,
2280                        iocb->ki_left, appending,
2281                        &can_do_direct, &has_refcount);
2282    if (ret < 0) {
2283        mlog_errno(ret);
2284        goto out;
2285    }
2286
2287    /*
2288     * We can't complete the direct I/O as requested, fall back to
2289     * buffered I/O.
2290     */
2291    if (direct_io && !can_do_direct) {
2292        ocfs2_rw_unlock(inode, rw_level);
2293        up_read(&inode->i_alloc_sem);
2294
2295        have_alloc_sem = 0;
2296        rw_level = -1;
2297
2298        direct_io = 0;
2299        goto relock;
2300    }
2301
2302    /*
2303     * To later detect whether a journal commit for sync writes is
2304     * necessary, we sample i_size, and cluster count here.
2305     */
2306    old_size = i_size_read(inode);
2307    old_clusters = OCFS2_I(inode)->ip_clusters;
2308
2309    /* communicate with ocfs2_dio_end_io */
2310    ocfs2_iocb_set_rw_locked(iocb, rw_level);
2311
2312    ret = generic_segment_checks(iov, &nr_segs, &ocount,
2313                     VERIFY_READ);
2314    if (ret)
2315        goto out_dio;
2316
2317    count = ocount;
2318    ret = generic_write_checks(file, ppos, &count,
2319                   S_ISBLK(inode->i_mode));
2320    if (ret)
2321        goto out_dio;
2322
2323    if (direct_io) {
2324        written = generic_file_direct_write(iocb, iov, &nr_segs, *ppos,
2325                            ppos, count, ocount);
2326        if (written < 0) {
2327            ret = written;
2328            goto out_dio;
2329        }
2330    } else {
2331        current->backing_dev_info = file->f_mapping->backing_dev_info;
2332        written = generic_file_buffered_write(iocb, iov, nr_segs, *ppos,
2333                              ppos, count, 0);
2334        current->backing_dev_info = NULL;
2335    }
2336
2337out_dio:
2338    /* buffered aio wouldn't have proper lock coverage today */
2339    BUG_ON(ret == -EIOCBQUEUED && !(file->f_flags & O_DIRECT));
2340
2341    if (((file->f_flags & O_DSYNC) && !direct_io) || IS_SYNC(inode) ||
2342        ((file->f_flags & O_DIRECT) && !direct_io)) {
2343        ret = filemap_fdatawrite_range(file->f_mapping, pos,
2344                           pos + count - 1);
2345        if (ret < 0)
2346            written = ret;
2347
2348        if (!ret && ((old_size != i_size_read(inode)) ||
2349                 (old_clusters != OCFS2_I(inode)->ip_clusters) ||
2350                 has_refcount)) {
2351            ret = jbd2_journal_force_commit(osb->journal->j_journal);
2352            if (ret < 0)
2353                written = ret;
2354        }
2355
2356        if (!ret)
2357            ret = filemap_fdatawait_range(file->f_mapping, pos,
2358                              pos + count - 1);
2359    }
2360
2361    /*
2362     * deep in g_f_a_w_n()->ocfs2_direct_IO we pass in a ocfs2_dio_end_io
2363     * function pointer which is called when o_direct io completes so that
2364     * it can unlock our rw lock. (it's the clustered equivalent of
2365     * i_alloc_sem; protects truncate from racing with pending ios).
2366     * Unfortunately there are error cases which call end_io and others
2367     * that don't. so we don't have to unlock the rw_lock if either an
2368     * async dio is going to do it in the future or an end_io after an
2369     * error has already done it.
2370     */
2371    if ((ret == -EIOCBQUEUED) || (!ocfs2_iocb_is_rw_locked(iocb))) {
2372        rw_level = -1;
2373        have_alloc_sem = 0;
2374    }
2375
2376out:
2377    if (rw_level != -1)
2378        ocfs2_rw_unlock(inode, rw_level);
2379
2380out_sems:
2381    if (have_alloc_sem) {
2382        up_read(&inode->i_alloc_sem);
2383        ocfs2_iocb_clear_sem_locked(iocb);
2384    }
2385
2386    mutex_unlock(&inode->i_mutex);
2387
2388    if (written)
2389        ret = written;
2390    return ret;
2391}
2392
2393static int ocfs2_splice_to_file(struct pipe_inode_info *pipe,
2394                struct file *out,
2395                struct splice_desc *sd)
2396{
2397    int ret;
2398
2399    ret = ocfs2_prepare_inode_for_write(out, &sd->pos,
2400                        sd->total_len, 0, NULL, NULL);
2401    if (ret < 0) {
2402        mlog_errno(ret);
2403        return ret;
2404    }
2405
2406    return splice_from_pipe_feed(pipe, sd, pipe_to_file);
2407}
2408
2409static ssize_t ocfs2_file_splice_write(struct pipe_inode_info *pipe,
2410                       struct file *out,
2411                       loff_t *ppos,
2412                       size_t len,
2413                       unsigned int flags)
2414{
2415    int ret;
2416    struct address_space *mapping = out->f_mapping;
2417    struct inode *inode = mapping->host;
2418    struct splice_desc sd = {
2419        .total_len = len,
2420        .flags = flags,
2421        .pos = *ppos,
2422        .u.file = out,
2423    };
2424
2425
2426    trace_ocfs2_file_splice_write(inode, out, out->f_path.dentry,
2427            (unsigned long long)OCFS2_I(inode)->ip_blkno,
2428            out->f_path.dentry->d_name.len,
2429            out->f_path.dentry->d_name.name, len);
2430
2431    if (pipe->inode)
2432        mutex_lock_nested(&pipe->inode->i_mutex, I_MUTEX_PARENT);
2433
2434    splice_from_pipe_begin(&sd);
2435    do {
2436        ret = splice_from_pipe_next(pipe, &sd);
2437        if (ret <= 0)
2438            break;
2439
2440        mutex_lock_nested(&inode->i_mutex, I_MUTEX_CHILD);
2441        ret = ocfs2_rw_lock(inode, 1);
2442        if (ret < 0)
2443            mlog_errno(ret);
2444        else {
2445            ret = ocfs2_splice_to_file(pipe, out, &sd);
2446            ocfs2_rw_unlock(inode, 1);
2447        }
2448        mutex_unlock(&inode->i_mutex);
2449    } while (ret > 0);
2450    splice_from_pipe_end(pipe, &sd);
2451
2452    if (pipe->inode)
2453        mutex_unlock(&pipe->inode->i_mutex);
2454
2455    if (sd.num_spliced)
2456        ret = sd.num_spliced;
2457
2458    if (ret > 0) {
2459        unsigned long nr_pages;
2460        int err;
2461
2462        nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
2463
2464        err = generic_write_sync(out, *ppos, ret);
2465        if (err)
2466            ret = err;
2467        else
2468            *ppos += ret;
2469
2470        balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
2471    }
2472
2473    return ret;
2474}
2475
2476static ssize_t ocfs2_file_splice_read(struct file *in,
2477                      loff_t *ppos,
2478                      struct pipe_inode_info *pipe,
2479                      size_t len,
2480                      unsigned int flags)
2481{
2482    int ret = 0, lock_level = 0;
2483    struct inode *inode = in->f_path.dentry->d_inode;
2484
2485    trace_ocfs2_file_splice_read(inode, in, in->f_path.dentry,
2486            (unsigned long long)OCFS2_I(inode)->ip_blkno,
2487            in->f_path.dentry->d_name.len,
2488            in->f_path.dentry->d_name.name, len);
2489
2490    /*
2491     * See the comment in ocfs2_file_aio_read()
2492     */
2493    ret = ocfs2_inode_lock_atime(inode, in->f_vfsmnt, &lock_level);
2494    if (ret < 0) {
2495        mlog_errno(ret);
2496        goto bail;
2497    }
2498    ocfs2_inode_unlock(inode, lock_level);
2499
2500    ret = generic_file_splice_read(in, ppos, pipe, len, flags);
2501
2502bail:
2503    return ret;
2504}
2505
2506static ssize_t ocfs2_file_aio_read(struct kiocb *iocb,
2507                   const struct iovec *iov,
2508                   unsigned long nr_segs,
2509                   loff_t pos)
2510{
2511    int ret = 0, rw_level = -1, have_alloc_sem = 0, lock_level = 0;
2512    struct file *filp = iocb->ki_filp;
2513    struct inode *inode = filp->f_path.dentry->d_inode;
2514
2515    trace_ocfs2_file_aio_read(inode, filp, filp->f_path.dentry,
2516            (unsigned long long)OCFS2_I(inode)->ip_blkno,
2517            filp->f_path.dentry->d_name.len,
2518            filp->f_path.dentry->d_name.name, nr_segs);
2519
2520
2521    if (!inode) {
2522        ret = -EINVAL;
2523        mlog_errno(ret);
2524        goto bail;
2525    }
2526
2527    ocfs2_iocb_clear_sem_locked(iocb);
2528
2529    /*
2530     * buffered reads protect themselves in ->readpage(). O_DIRECT reads
2531     * need locks to protect pending reads from racing with truncate.
2532     */
2533    if (filp->f_flags & O_DIRECT) {
2534        down_read(&inode->i_alloc_sem);
2535        have_alloc_sem = 1;
2536        ocfs2_iocb_set_sem_locked(iocb);
2537
2538        ret = ocfs2_rw_lock(inode, 0);
2539        if (ret < 0) {
2540            mlog_errno(ret);
2541            goto bail;
2542        }
2543        rw_level = 0;
2544        /* communicate with ocfs2_dio_end_io */
2545        ocfs2_iocb_set_rw_locked(iocb, rw_level);
2546    }
2547
2548    /*
2549     * We're fine letting folks race truncates and extending
2550     * writes with read across the cluster, just like they can
2551     * locally. Hence no rw_lock during read.
2552     *
2553     * Take and drop the meta data lock to update inode fields
2554     * like i_size. This allows the checks down below
2555     * generic_file_aio_read() a chance of actually working.
2556     */
2557    ret = ocfs2_inode_lock_atime(inode, filp->f_vfsmnt, &lock_level);
2558    if (ret < 0) {
2559        mlog_errno(ret);
2560        goto bail;
2561    }
2562    ocfs2_inode_unlock(inode, lock_level);
2563
2564    ret = generic_file_aio_read(iocb, iov, nr_segs, iocb->ki_pos);
2565    trace_generic_file_aio_read_ret(ret);
2566
2567    /* buffered aio wouldn't have proper lock coverage today */
2568    BUG_ON(ret == -EIOCBQUEUED && !(filp->f_flags & O_DIRECT));
2569
2570    /* see ocfs2_file_aio_write */
2571    if (ret == -EIOCBQUEUED || !ocfs2_iocb_is_rw_locked(iocb)) {
2572        rw_level = -1;
2573        have_alloc_sem = 0;
2574    }
2575
2576bail:
2577    if (have_alloc_sem) {
2578        up_read(&inode->i_alloc_sem);
2579        ocfs2_iocb_clear_sem_locked(iocb);
2580    }
2581    if (rw_level != -1)
2582        ocfs2_rw_unlock(inode, rw_level);
2583
2584    return ret;
2585}
2586
2587const struct inode_operations ocfs2_file_iops = {
2588    .setattr = ocfs2_setattr,
2589    .getattr = ocfs2_getattr,
2590    .permission = ocfs2_permission,
2591    .setxattr = generic_setxattr,
2592    .getxattr = generic_getxattr,
2593    .listxattr = ocfs2_listxattr,
2594    .removexattr = generic_removexattr,
2595    .fiemap = ocfs2_fiemap,
2596};
2597
2598const struct inode_operations ocfs2_special_file_iops = {
2599    .setattr = ocfs2_setattr,
2600    .getattr = ocfs2_getattr,
2601    .permission = ocfs2_permission,
2602};
2603
2604/*
2605 * Other than ->lock, keep ocfs2_fops and ocfs2_dops in sync with
2606 * ocfs2_fops_no_plocks and ocfs2_dops_no_plocks!
2607 */
2608const struct file_operations ocfs2_fops = {
2609    .llseek = generic_file_llseek,
2610    .read = do_sync_read,
2611    .write = do_sync_write,
2612    .mmap = ocfs2_mmap,
2613    .fsync = ocfs2_sync_file,
2614    .release = ocfs2_file_release,
2615    .open = ocfs2_file_open,
2616    .aio_read = ocfs2_file_aio_read,
2617    .aio_write = ocfs2_file_aio_write,
2618    .unlocked_ioctl = ocfs2_ioctl,
2619#ifdef CONFIG_COMPAT
2620    .compat_ioctl = ocfs2_compat_ioctl,
2621#endif
2622    .lock = ocfs2_lock,
2623    .flock = ocfs2_flock,
2624    .splice_read = ocfs2_file_splice_read,
2625    .splice_write = ocfs2_file_splice_write,
2626    .fallocate = ocfs2_fallocate,
2627};
2628
2629const struct file_operations ocfs2_dops = {
2630    .llseek = generic_file_llseek,
2631    .read = generic_read_dir,
2632    .readdir = ocfs2_readdir,
2633    .fsync = ocfs2_sync_file,
2634    .release = ocfs2_dir_release,
2635    .open = ocfs2_dir_open,
2636    .unlocked_ioctl = ocfs2_ioctl,
2637#ifdef CONFIG_COMPAT
2638    .compat_ioctl = ocfs2_compat_ioctl,
2639#endif
2640    .lock = ocfs2_lock,
2641    .flock = ocfs2_flock,
2642};
2643
2644/*
2645 * POSIX-lockless variants of our file_operations.
2646 *
2647 * These will be used if the underlying cluster stack does not support
2648 * posix file locking, if the user passes the "localflocks" mount
2649 * option, or if we have a local-only fs.
2650 *
2651 * ocfs2_flock is in here because all stacks handle UNIX file locks,
2652 * so we still want it in the case of no stack support for
2653 * plocks. Internally, it will do the right thing when asked to ignore
2654 * the cluster.
2655 */
2656const struct file_operations ocfs2_fops_no_plocks = {
2657    .llseek = generic_file_llseek,
2658    .read = do_sync_read,
2659    .write = do_sync_write,
2660    .mmap = ocfs2_mmap,
2661    .fsync = ocfs2_sync_file,
2662    .release = ocfs2_file_release,
2663    .open = ocfs2_file_open,
2664    .aio_read = ocfs2_file_aio_read,
2665    .aio_write = ocfs2_file_aio_write,
2666    .unlocked_ioctl = ocfs2_ioctl,
2667#ifdef CONFIG_COMPAT
2668    .compat_ioctl = ocfs2_compat_ioctl,
2669#endif
2670    .flock = ocfs2_flock,
2671    .splice_read = ocfs2_file_splice_read,
2672    .splice_write = ocfs2_file_splice_write,
2673    .fallocate = ocfs2_fallocate,
2674};
2675
2676const struct file_operations ocfs2_dops_no_plocks = {
2677    .llseek = generic_file_llseek,
2678    .read = generic_read_dir,
2679    .readdir = ocfs2_readdir,
2680    .fsync = ocfs2_sync_file,
2681    .release = ocfs2_dir_release,
2682    .open = ocfs2_dir_open,
2683    .unlocked_ioctl = ocfs2_ioctl,
2684#ifdef CONFIG_COMPAT
2685    .compat_ioctl = ocfs2_compat_ioctl,
2686#endif
2687    .flock = ocfs2_flock,
2688};
2689

Archive Download this file



interactive