Merge master.kernel.org:/home/rmk/linux-2.6-arm
[linux-2.6] / fs / xfs / linux-2.6 / xfs_lrw.c
1 /*
2  * Copyright (c) 2000-2003,2005 Silicon Graphics, Inc.
3  * All Rights Reserved.
4  *
5  * This program is free software; you can redistribute it and/or
6  * modify it under the terms of the GNU General Public License as
7  * published by the Free Software Foundation.
8  *
9  * This program is distributed in the hope that it would be useful,
10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
12  * GNU General Public License for more details.
13  *
14  * You should have received a copy of the GNU General Public License
15  * along with this program; if not, write the Free Software Foundation,
16  * Inc.,  51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
17  */
18 #include "xfs.h"
19 #include "xfs_fs.h"
20 #include "xfs_bit.h"
21 #include "xfs_log.h"
22 #include "xfs_inum.h"
23 #include "xfs_trans.h"
24 #include "xfs_sb.h"
25 #include "xfs_ag.h"
26 #include "xfs_dir.h"
27 #include "xfs_dir2.h"
28 #include "xfs_alloc.h"
29 #include "xfs_dmapi.h"
30 #include "xfs_quota.h"
31 #include "xfs_mount.h"
32 #include "xfs_bmap_btree.h"
33 #include "xfs_alloc_btree.h"
34 #include "xfs_ialloc_btree.h"
35 #include "xfs_dir_sf.h"
36 #include "xfs_dir2_sf.h"
37 #include "xfs_attr_sf.h"
38 #include "xfs_dinode.h"
39 #include "xfs_inode.h"
40 #include "xfs_bmap.h"
41 #include "xfs_btree.h"
42 #include "xfs_ialloc.h"
43 #include "xfs_rtalloc.h"
44 #include "xfs_error.h"
45 #include "xfs_itable.h"
46 #include "xfs_rw.h"
47 #include "xfs_acl.h"
48 #include "xfs_cap.h"
49 #include "xfs_mac.h"
50 #include "xfs_attr.h"
51 #include "xfs_inode_item.h"
52 #include "xfs_buf_item.h"
53 #include "xfs_utils.h"
54 #include "xfs_iomap.h"
55
56 #include <linux/capability.h>
57 #include <linux/writeback.h>
58
59
60 #if defined(XFS_RW_TRACE)
61 void
62 xfs_rw_enter_trace(
63         int                     tag,
64         xfs_iocore_t            *io,
65         void                    *data,
66         size_t                  segs,
67         loff_t                  offset,
68         int                     ioflags)
69 {
70         xfs_inode_t     *ip = XFS_IO_INODE(io);
71
72         if (ip->i_rwtrace == NULL)
73                 return;
74         ktrace_enter(ip->i_rwtrace,
75                 (void *)(unsigned long)tag,
76                 (void *)ip,
77                 (void *)((unsigned long)((ip->i_d.di_size >> 32) & 0xffffffff)),
78                 (void *)((unsigned long)(ip->i_d.di_size & 0xffffffff)),
79                 (void *)data,
80                 (void *)((unsigned long)segs),
81                 (void *)((unsigned long)((offset >> 32) & 0xffffffff)),
82                 (void *)((unsigned long)(offset & 0xffffffff)),
83                 (void *)((unsigned long)ioflags),
84                 (void *)((unsigned long)((io->io_new_size >> 32) & 0xffffffff)),
85                 (void *)((unsigned long)(io->io_new_size & 0xffffffff)),
86                 (void *)((unsigned long)current_pid()),
87                 (void *)NULL,
88                 (void *)NULL,
89                 (void *)NULL,
90                 (void *)NULL);
91 }
92
93 void
94 xfs_inval_cached_trace(
95         xfs_iocore_t    *io,
96         xfs_off_t       offset,
97         xfs_off_t       len,
98         xfs_off_t       first,
99         xfs_off_t       last)
100 {
101         xfs_inode_t     *ip = XFS_IO_INODE(io);
102
103         if (ip->i_rwtrace == NULL)
104                 return;
105         ktrace_enter(ip->i_rwtrace,
106                 (void *)(__psint_t)XFS_INVAL_CACHED,
107                 (void *)ip,
108                 (void *)((unsigned long)((offset >> 32) & 0xffffffff)),
109                 (void *)((unsigned long)(offset & 0xffffffff)),
110                 (void *)((unsigned long)((len >> 32) & 0xffffffff)),
111                 (void *)((unsigned long)(len & 0xffffffff)),
112                 (void *)((unsigned long)((first >> 32) & 0xffffffff)),
113                 (void *)((unsigned long)(first & 0xffffffff)),
114                 (void *)((unsigned long)((last >> 32) & 0xffffffff)),
115                 (void *)((unsigned long)(last & 0xffffffff)),
116                 (void *)((unsigned long)current_pid()),
117                 (void *)NULL,
118                 (void *)NULL,
119                 (void *)NULL,
120                 (void *)NULL,
121                 (void *)NULL);
122 }
123 #endif
124
125 /*
126  *      xfs_iozero
127  *
128  *      xfs_iozero clears the specified range of buffer supplied,
129  *      and marks all the affected blocks as valid and modified.  If
130  *      an affected block is not allocated, it will be allocated.  If
131  *      an affected block is not completely overwritten, and is not
132  *      valid before the operation, it will be read from disk before
133  *      being partially zeroed.
134  */
135 STATIC int
136 xfs_iozero(
137         struct inode            *ip,    /* inode                        */
138         loff_t                  pos,    /* offset in file               */
139         size_t                  count,  /* size of data to zero         */
140         loff_t                  end_size)       /* max file size to set */
141 {
142         unsigned                bytes;
143         struct page             *page;
144         struct address_space    *mapping;
145         char                    *kaddr;
146         int                     status;
147
148         mapping = ip->i_mapping;
149         do {
150                 unsigned long index, offset;
151
152                 offset = (pos & (PAGE_CACHE_SIZE -1)); /* Within page */
153                 index = pos >> PAGE_CACHE_SHIFT;
154                 bytes = PAGE_CACHE_SIZE - offset;
155                 if (bytes > count)
156                         bytes = count;
157
158                 status = -ENOMEM;
159                 page = grab_cache_page(mapping, index);
160                 if (!page)
161                         break;
162
163                 kaddr = kmap(page);
164                 status = mapping->a_ops->prepare_write(NULL, page, offset,
165                                                         offset + bytes);
166                 if (status) {
167                         goto unlock;
168                 }
169
170                 memset((void *) (kaddr + offset), 0, bytes);
171                 flush_dcache_page(page);
172                 status = mapping->a_ops->commit_write(NULL, page, offset,
173                                                         offset + bytes);
174                 if (!status) {
175                         pos += bytes;
176                         count -= bytes;
177                         if (pos > i_size_read(ip))
178                                 i_size_write(ip, pos < end_size ? pos : end_size);
179                 }
180
181 unlock:
182                 kunmap(page);
183                 unlock_page(page);
184                 page_cache_release(page);
185                 if (status)
186                         break;
187         } while (count);
188
189         return (-status);
190 }
191
192 ssize_t                 /* bytes read, or (-)  error */
193 xfs_read(
194         bhv_desc_t              *bdp,
195         struct kiocb            *iocb,
196         const struct iovec      *iovp,
197         unsigned int            segs,
198         loff_t                  *offset,
199         int                     ioflags,
200         cred_t                  *credp)
201 {
202         struct file             *file = iocb->ki_filp;
203         struct inode            *inode = file->f_mapping->host;
204         size_t                  size = 0;
205         ssize_t                 ret;
206         xfs_fsize_t             n;
207         xfs_inode_t             *ip;
208         xfs_mount_t             *mp;
209         vnode_t                 *vp;
210         unsigned long           seg;
211
212         ip = XFS_BHVTOI(bdp);
213         vp = BHV_TO_VNODE(bdp);
214         mp = ip->i_mount;
215
216         XFS_STATS_INC(xs_read_calls);
217
218         /* START copy & waste from filemap.c */
219         for (seg = 0; seg < segs; seg++) {
220                 const struct iovec *iv = &iovp[seg];
221
222                 /*
223                  * If any segment has a negative length, or the cumulative
224                  * length ever wraps negative then return -EINVAL.
225                  */
226                 size += iv->iov_len;
227                 if (unlikely((ssize_t)(size|iv->iov_len) < 0))
228                         return XFS_ERROR(-EINVAL);
229         }
230         /* END copy & waste from filemap.c */
231
232         if (unlikely(ioflags & IO_ISDIRECT)) {
233                 xfs_buftarg_t   *target =
234                         (ip->i_d.di_flags & XFS_DIFLAG_REALTIME) ?
235                                 mp->m_rtdev_targp : mp->m_ddev_targp;
236                 if ((*offset & target->bt_smask) ||
237                     (size & target->bt_smask)) {
238                         if (*offset == ip->i_d.di_size) {
239                                 return (0);
240                         }
241                         return -XFS_ERROR(EINVAL);
242                 }
243         }
244
245         n = XFS_MAXIOFFSET(mp) - *offset;
246         if ((n <= 0) || (size == 0))
247                 return 0;
248
249         if (n < size)
250                 size = n;
251
252         if (XFS_FORCED_SHUTDOWN(mp))
253                 return -EIO;
254
255         if (unlikely(ioflags & IO_ISDIRECT))
256                 mutex_lock(&inode->i_mutex);
257         xfs_ilock(ip, XFS_IOLOCK_SHARED);
258
259         if (DM_EVENT_ENABLED(vp->v_vfsp, ip, DM_EVENT_READ) &&
260             !(ioflags & IO_INVIS)) {
261                 vrwlock_t locktype = VRWLOCK_READ;
262                 int dmflags = FILP_DELAY_FLAG(file) | DM_SEM_FLAG_RD(ioflags);
263
264                 ret = -XFS_SEND_DATA(mp, DM_EVENT_READ,
265                                         BHV_TO_VNODE(bdp), *offset, size,
266                                         dmflags, &locktype);
267                 if (ret) {
268                         xfs_iunlock(ip, XFS_IOLOCK_SHARED);
269                         goto unlock_mutex;
270                 }
271         }
272
273         if (unlikely((ioflags & IO_ISDIRECT) && VN_CACHED(vp)))
274                 VOP_FLUSHINVAL_PAGES(vp, ctooff(offtoct(*offset)),
275                                                 -1, FI_REMAPF_LOCKED);
276
277         xfs_rw_enter_trace(XFS_READ_ENTER, &ip->i_iocore,
278                                 (void *)iovp, segs, *offset, ioflags);
279         ret = __generic_file_aio_read(iocb, iovp, segs, offset);
280         if (ret == -EIOCBQUEUED && !(ioflags & IO_ISAIO))
281                 ret = wait_on_sync_kiocb(iocb);
282         if (ret > 0)
283                 XFS_STATS_ADD(xs_read_bytes, ret);
284
285         xfs_iunlock(ip, XFS_IOLOCK_SHARED);
286
287 unlock_mutex:
288         if (unlikely(ioflags & IO_ISDIRECT))
289                 mutex_unlock(&inode->i_mutex);
290         return ret;
291 }
292
293 ssize_t
294 xfs_sendfile(
295         bhv_desc_t              *bdp,
296         struct file             *filp,
297         loff_t                  *offset,
298         int                     ioflags,
299         size_t                  count,
300         read_actor_t            actor,
301         void                    *target,
302         cred_t                  *credp)
303 {
304         ssize_t                 ret;
305         xfs_fsize_t             n;
306         xfs_inode_t             *ip;
307         xfs_mount_t             *mp;
308         vnode_t                 *vp;
309
310         ip = XFS_BHVTOI(bdp);
311         vp = BHV_TO_VNODE(bdp);
312         mp = ip->i_mount;
313
314         XFS_STATS_INC(xs_read_calls);
315
316         n = XFS_MAXIOFFSET(mp) - *offset;
317         if ((n <= 0) || (count == 0))
318                 return 0;
319
320         if (n < count)
321                 count = n;
322
323         if (XFS_FORCED_SHUTDOWN(ip->i_mount))
324                 return -EIO;
325
326         xfs_ilock(ip, XFS_IOLOCK_SHARED);
327
328         if (DM_EVENT_ENABLED(vp->v_vfsp, ip, DM_EVENT_READ) &&
329             (!(ioflags & IO_INVIS))) {
330                 vrwlock_t locktype = VRWLOCK_READ;
331                 int error;
332
333                 error = XFS_SEND_DATA(mp, DM_EVENT_READ, BHV_TO_VNODE(bdp), *offset, count,
334                                       FILP_DELAY_FLAG(filp), &locktype);
335                 if (error) {
336                         xfs_iunlock(ip, XFS_IOLOCK_SHARED);
337                         return -error;
338                 }
339         }
340         xfs_rw_enter_trace(XFS_SENDFILE_ENTER, &ip->i_iocore,
341                    (void *)(unsigned long)target, count, *offset, ioflags);
342         ret = generic_file_sendfile(filp, offset, count, actor, target);
343
344         xfs_iunlock(ip, XFS_IOLOCK_SHARED);
345
346         if (ret > 0)
347                 XFS_STATS_ADD(xs_read_bytes, ret);
348
349         return ret;
350 }
351
352 /*
353  * This routine is called to handle zeroing any space in the last
354  * block of the file that is beyond the EOF.  We do this since the
355  * size is being increased without writing anything to that block
356  * and we don't want anyone to read the garbage on the disk.
357  */
358 STATIC int                              /* error (positive) */
359 xfs_zero_last_block(
360         struct inode    *ip,
361         xfs_iocore_t    *io,
362         xfs_fsize_t     isize,
363         xfs_fsize_t     end_size)
364 {
365         xfs_fileoff_t   last_fsb;
366         xfs_mount_t     *mp;
367         int             nimaps;
368         int             zero_offset;
369         int             zero_len;
370         int             error = 0;
371         xfs_bmbt_irec_t imap;
372         loff_t          loff;
373
374         ASSERT(ismrlocked(io->io_lock, MR_UPDATE) != 0);
375
376         mp = io->io_mount;
377
378         zero_offset = XFS_B_FSB_OFFSET(mp, isize);
379         if (zero_offset == 0) {
380                 /*
381                  * There are no extra bytes in the last block on disk to
382                  * zero, so return.
383                  */
384                 return 0;
385         }
386
387         last_fsb = XFS_B_TO_FSBT(mp, isize);
388         nimaps = 1;
389         error = XFS_BMAPI(mp, NULL, io, last_fsb, 1, 0, NULL, 0, &imap,
390                           &nimaps, NULL);
391         if (error) {
392                 return error;
393         }
394         ASSERT(nimaps > 0);
395         /*
396          * If the block underlying isize is just a hole, then there
397          * is nothing to zero.
398          */
399         if (imap.br_startblock == HOLESTARTBLOCK) {
400                 return 0;
401         }
402         /*
403          * Zero the part of the last block beyond the EOF, and write it
404          * out sync.  We need to drop the ilock while we do this so we
405          * don't deadlock when the buffer cache calls back to us.
406          */
407         XFS_IUNLOCK(mp, io, XFS_ILOCK_EXCL| XFS_EXTSIZE_RD);
408         loff = XFS_FSB_TO_B(mp, last_fsb);
409
410         zero_len = mp->m_sb.sb_blocksize - zero_offset;
411
412         error = xfs_iozero(ip, loff + zero_offset, zero_len, end_size);
413
414         XFS_ILOCK(mp, io, XFS_ILOCK_EXCL|XFS_EXTSIZE_RD);
415         ASSERT(error >= 0);
416         return error;
417 }
418
419 /*
420  * Zero any on disk space between the current EOF and the new,
421  * larger EOF.  This handles the normal case of zeroing the remainder
422  * of the last block in the file and the unusual case of zeroing blocks
423  * out beyond the size of the file.  This second case only happens
424  * with fixed size extents and when the system crashes before the inode
425  * size was updated but after blocks were allocated.  If fill is set,
426  * then any holes in the range are filled and zeroed.  If not, the holes
427  * are left alone as holes.
428  */
429
430 int                                     /* error (positive) */
431 xfs_zero_eof(
432         vnode_t         *vp,
433         xfs_iocore_t    *io,
434         xfs_off_t       offset,         /* starting I/O offset */
435         xfs_fsize_t     isize,          /* current inode size */
436         xfs_fsize_t     end_size)       /* terminal inode size */
437 {
438         struct inode    *ip = vn_to_inode(vp);
439         xfs_fileoff_t   start_zero_fsb;
440         xfs_fileoff_t   end_zero_fsb;
441         xfs_fileoff_t   zero_count_fsb;
442         xfs_fileoff_t   last_fsb;
443         xfs_extlen_t    buf_len_fsb;
444         xfs_mount_t     *mp;
445         int             nimaps;
446         int             error = 0;
447         xfs_bmbt_irec_t imap;
448
449         ASSERT(ismrlocked(io->io_lock, MR_UPDATE));
450         ASSERT(ismrlocked(io->io_iolock, MR_UPDATE));
451         ASSERT(offset > isize);
452
453         mp = io->io_mount;
454
455         /*
456          * First handle zeroing the block on which isize resides.
457          * We only zero a part of that block so it is handled specially.
458          */
459         error = xfs_zero_last_block(ip, io, isize, end_size);
460         if (error) {
461                 ASSERT(ismrlocked(io->io_lock, MR_UPDATE));
462                 ASSERT(ismrlocked(io->io_iolock, MR_UPDATE));
463                 return error;
464         }
465
466         /*
467          * Calculate the range between the new size and the old
468          * where blocks needing to be zeroed may exist.  To get the
469          * block where the last byte in the file currently resides,
470          * we need to subtract one from the size and truncate back
471          * to a block boundary.  We subtract 1 in case the size is
472          * exactly on a block boundary.
473          */
474         last_fsb = isize ? XFS_B_TO_FSBT(mp, isize - 1) : (xfs_fileoff_t)-1;
475         start_zero_fsb = XFS_B_TO_FSB(mp, (xfs_ufsize_t)isize);
476         end_zero_fsb = XFS_B_TO_FSBT(mp, offset - 1);
477         ASSERT((xfs_sfiloff_t)last_fsb < (xfs_sfiloff_t)start_zero_fsb);
478         if (last_fsb == end_zero_fsb) {
479                 /*
480                  * The size was only incremented on its last block.
481                  * We took care of that above, so just return.
482                  */
483                 return 0;
484         }
485
486         ASSERT(start_zero_fsb <= end_zero_fsb);
487         while (start_zero_fsb <= end_zero_fsb) {
488                 nimaps = 1;
489                 zero_count_fsb = end_zero_fsb - start_zero_fsb + 1;
490                 error = XFS_BMAPI(mp, NULL, io, start_zero_fsb, zero_count_fsb,
491                                   0, NULL, 0, &imap, &nimaps, NULL);
492                 if (error) {
493                         ASSERT(ismrlocked(io->io_lock, MR_UPDATE));
494                         ASSERT(ismrlocked(io->io_iolock, MR_UPDATE));
495                         return error;
496                 }
497                 ASSERT(nimaps > 0);
498
499                 if (imap.br_state == XFS_EXT_UNWRITTEN ||
500                     imap.br_startblock == HOLESTARTBLOCK) {
501                         /*
502                          * This loop handles initializing pages that were
503                          * partially initialized by the code below this
504                          * loop. It basically zeroes the part of the page
505                          * that sits on a hole and sets the page as P_HOLE
506                          * and calls remapf if it is a mapped file.
507                          */
508                         start_zero_fsb = imap.br_startoff + imap.br_blockcount;
509                         ASSERT(start_zero_fsb <= (end_zero_fsb + 1));
510                         continue;
511                 }
512
513                 /*
514                  * There are blocks in the range requested.
515                  * Zero them a single write at a time.  We actually
516                  * don't zero the entire range returned if it is
517                  * too big and simply loop around to get the rest.
518                  * That is not the most efficient thing to do, but it
519                  * is simple and this path should not be exercised often.
520                  */
521                 buf_len_fsb = XFS_FILBLKS_MIN(imap.br_blockcount,
522                                               mp->m_writeio_blocks << 8);
523                 /*
524                  * Drop the inode lock while we're doing the I/O.
525                  * We'll still have the iolock to protect us.
526                  */
527                 XFS_IUNLOCK(mp, io, XFS_ILOCK_EXCL|XFS_EXTSIZE_RD);
528
529                 error = xfs_iozero(ip,
530                                    XFS_FSB_TO_B(mp, start_zero_fsb),
531                                    XFS_FSB_TO_B(mp, buf_len_fsb),
532                                    end_size);
533
534                 if (error) {
535                         goto out_lock;
536                 }
537
538                 start_zero_fsb = imap.br_startoff + buf_len_fsb;
539                 ASSERT(start_zero_fsb <= (end_zero_fsb + 1));
540
541                 XFS_ILOCK(mp, io, XFS_ILOCK_EXCL|XFS_EXTSIZE_RD);
542         }
543
544         return 0;
545
546 out_lock:
547
548         XFS_ILOCK(mp, io, XFS_ILOCK_EXCL|XFS_EXTSIZE_RD);
549         ASSERT(error >= 0);
550         return error;
551 }
552
553 ssize_t                         /* bytes written, or (-) error */
554 xfs_write(
555         bhv_desc_t              *bdp,
556         struct kiocb            *iocb,
557         const struct iovec      *iovp,
558         unsigned int            nsegs,
559         loff_t                  *offset,
560         int                     ioflags,
561         cred_t                  *credp)
562 {
563         struct file             *file = iocb->ki_filp;
564         struct address_space    *mapping = file->f_mapping;
565         struct inode            *inode = mapping->host;
566         unsigned long           segs = nsegs;
567         xfs_inode_t             *xip;
568         xfs_mount_t             *mp;
569         ssize_t                 ret = 0, error = 0;
570         xfs_fsize_t             isize, new_size;
571         xfs_iocore_t            *io;
572         vnode_t                 *vp;
573         unsigned long           seg;
574         int                     iolock;
575         int                     eventsent = 0;
576         vrwlock_t               locktype;
577         size_t                  ocount = 0, count;
578         loff_t                  pos;
579         int                     need_i_mutex = 1, need_flush = 0;
580
581         XFS_STATS_INC(xs_write_calls);
582
583         vp = BHV_TO_VNODE(bdp);
584         xip = XFS_BHVTOI(bdp);
585
586         for (seg = 0; seg < segs; seg++) {
587                 const struct iovec *iv = &iovp[seg];
588
589                 /*
590                  * If any segment has a negative length, or the cumulative
591                  * length ever wraps negative then return -EINVAL.
592                  */
593                 ocount += iv->iov_len;
594                 if (unlikely((ssize_t)(ocount|iv->iov_len) < 0))
595                         return -EINVAL;
596                 if (access_ok(VERIFY_READ, iv->iov_base, iv->iov_len))
597                         continue;
598                 if (seg == 0)
599                         return -EFAULT;
600                 segs = seg;
601                 ocount -= iv->iov_len;  /* This segment is no good */
602                 break;
603         }
604
605         count = ocount;
606         pos = *offset;
607
608         if (count == 0)
609                 return 0;
610
611         io = &xip->i_iocore;
612         mp = io->io_mount;
613
614         if (XFS_FORCED_SHUTDOWN(mp))
615                 return -EIO;
616
617         fs_check_frozen(vp->v_vfsp, SB_FREEZE_WRITE);
618
619         if (ioflags & IO_ISDIRECT) {
620                 xfs_buftarg_t   *target =
621                         (xip->i_d.di_flags & XFS_DIFLAG_REALTIME) ?
622                                 mp->m_rtdev_targp : mp->m_ddev_targp;
623
624                 if ((pos & target->bt_smask) || (count & target->bt_smask))
625                         return XFS_ERROR(-EINVAL);
626
627                 if (!VN_CACHED(vp) && pos < i_size_read(inode))
628                         need_i_mutex = 0;
629
630                 if (VN_CACHED(vp))
631                         need_flush = 1;
632         }
633
634 relock:
635         if (need_i_mutex) {
636                 iolock = XFS_IOLOCK_EXCL;
637                 locktype = VRWLOCK_WRITE;
638
639                 mutex_lock(&inode->i_mutex);
640         } else {
641                 iolock = XFS_IOLOCK_SHARED;
642                 locktype = VRWLOCK_WRITE_DIRECT;
643         }
644
645         xfs_ilock(xip, XFS_ILOCK_EXCL|iolock);
646
647         isize = i_size_read(inode);
648
649         if (file->f_flags & O_APPEND)
650                 *offset = isize;
651
652 start:
653         error = -generic_write_checks(file, &pos, &count,
654                                         S_ISBLK(inode->i_mode));
655         if (error) {
656                 xfs_iunlock(xip, XFS_ILOCK_EXCL|iolock);
657                 goto out_unlock_mutex;
658         }
659
660         new_size = pos + count;
661         if (new_size > isize)
662                 io->io_new_size = new_size;
663
664         if ((DM_EVENT_ENABLED(vp->v_vfsp, xip, DM_EVENT_WRITE) &&
665             !(ioflags & IO_INVIS) && !eventsent)) {
666                 loff_t          savedsize = pos;
667                 int             dmflags = FILP_DELAY_FLAG(file);
668
669                 if (need_i_mutex)
670                         dmflags |= DM_FLAGS_IMUX;
671
672                 xfs_iunlock(xip, XFS_ILOCK_EXCL);
673                 error = XFS_SEND_DATA(xip->i_mount, DM_EVENT_WRITE, vp,
674                                       pos, count,
675                                       dmflags, &locktype);
676                 if (error) {
677                         xfs_iunlock(xip, iolock);
678                         goto out_unlock_mutex;
679                 }
680                 xfs_ilock(xip, XFS_ILOCK_EXCL);
681                 eventsent = 1;
682
683                 /*
684                  * The iolock was dropped and reaquired in XFS_SEND_DATA
685                  * so we have to recheck the size when appending.
686                  * We will only "goto start;" once, since having sent the
687                  * event prevents another call to XFS_SEND_DATA, which is
688                  * what allows the size to change in the first place.
689                  */
690                 if ((file->f_flags & O_APPEND) && savedsize != isize) {
691                         pos = isize = xip->i_d.di_size;
692                         goto start;
693                 }
694         }
695
696         if (likely(!(ioflags & IO_INVIS))) {
697                 file_update_time(file);
698                 xfs_ichgtime_fast(xip, inode,
699                                   XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
700         }
701
702         /*
703          * If the offset is beyond the size of the file, we have a couple
704          * of things to do. First, if there is already space allocated
705          * we need to either create holes or zero the disk or ...
706          *
707          * If there is a page where the previous size lands, we need
708          * to zero it out up to the new size.
709          */
710
711         if (pos > isize) {
712                 error = xfs_zero_eof(BHV_TO_VNODE(bdp), io, pos,
713                                         isize, pos + count);
714                 if (error) {
715                         xfs_iunlock(xip, XFS_ILOCK_EXCL|iolock);
716                         goto out_unlock_mutex;
717                 }
718         }
719         xfs_iunlock(xip, XFS_ILOCK_EXCL);
720
721         /*
722          * If we're writing the file then make sure to clear the
723          * setuid and setgid bits if the process is not being run
724          * by root.  This keeps people from modifying setuid and
725          * setgid binaries.
726          */
727
728         if (((xip->i_d.di_mode & S_ISUID) ||
729             ((xip->i_d.di_mode & (S_ISGID | S_IXGRP)) ==
730                 (S_ISGID | S_IXGRP))) &&
731              !capable(CAP_FSETID)) {
732                 error = xfs_write_clear_setuid(xip);
733                 if (likely(!error))
734                         error = -remove_suid(file->f_dentry);
735                 if (unlikely(error)) {
736                         xfs_iunlock(xip, iolock);
737                         goto out_unlock_mutex;
738                 }
739         }
740
741 retry:
742         /* We can write back this queue in page reclaim */
743         current->backing_dev_info = mapping->backing_dev_info;
744
745         if ((ioflags & IO_ISDIRECT)) {
746                 if (need_flush) {
747                         xfs_inval_cached_trace(io, pos, -1,
748                                         ctooff(offtoct(pos)), -1);
749                         VOP_FLUSHINVAL_PAGES(vp, ctooff(offtoct(pos)),
750                                         -1, FI_REMAPF_LOCKED);
751                 }
752
753                 if (need_i_mutex) {
754                         /* demote the lock now the cached pages are gone */
755                         XFS_ILOCK_DEMOTE(mp, io, XFS_IOLOCK_EXCL);
756                         mutex_unlock(&inode->i_mutex);
757
758                         iolock = XFS_IOLOCK_SHARED;
759                         locktype = VRWLOCK_WRITE_DIRECT;
760                         need_i_mutex = 0;
761                 }
762
763                 xfs_rw_enter_trace(XFS_DIOWR_ENTER, io, (void *)iovp, segs,
764                                 *offset, ioflags);
765                 ret = generic_file_direct_write(iocb, iovp,
766                                 &segs, pos, offset, count, ocount);
767
768                 /*
769                  * direct-io write to a hole: fall through to buffered I/O
770                  * for completing the rest of the request.
771                  */
772                 if (ret >= 0 && ret != count) {
773                         XFS_STATS_ADD(xs_write_bytes, ret);
774
775                         pos += ret;
776                         count -= ret;
777
778                         need_i_mutex = 1;
779                         ioflags &= ~IO_ISDIRECT;
780                         xfs_iunlock(xip, iolock);
781                         goto relock;
782                 }
783         } else {
784                 xfs_rw_enter_trace(XFS_WRITE_ENTER, io, (void *)iovp, segs,
785                                 *offset, ioflags);
786                 ret = generic_file_buffered_write(iocb, iovp, segs,
787                                 pos, offset, count, ret);
788         }
789
790         current->backing_dev_info = NULL;
791
792         if (ret == -EIOCBQUEUED && !(ioflags & IO_ISAIO))
793                 ret = wait_on_sync_kiocb(iocb);
794
795         if ((ret == -ENOSPC) &&
796             DM_EVENT_ENABLED(vp->v_vfsp, xip, DM_EVENT_NOSPACE) &&
797             !(ioflags & IO_INVIS)) {
798
799                 xfs_rwunlock(bdp, locktype);
800                 if (need_i_mutex)
801                         mutex_unlock(&inode->i_mutex);
802                 error = XFS_SEND_NAMESP(xip->i_mount, DM_EVENT_NOSPACE, vp,
803                                 DM_RIGHT_NULL, vp, DM_RIGHT_NULL, NULL, NULL,
804                                 0, 0, 0); /* Delay flag intentionally  unused */
805                 if (error)
806                         goto out_nounlocks;
807                 if (need_i_mutex)
808                         mutex_lock(&inode->i_mutex);
809                 xfs_rwlock(bdp, locktype);
810                 pos = xip->i_d.di_size;
811                 ret = 0;
812                 goto retry;
813         }
814
815         isize = i_size_read(inode);
816         if (unlikely(ret < 0 && ret != -EFAULT && *offset > isize))
817                 *offset = isize;
818
819         if (*offset > xip->i_d.di_size) {
820                 xfs_ilock(xip, XFS_ILOCK_EXCL);
821                 if (*offset > xip->i_d.di_size) {
822                         xip->i_d.di_size = *offset;
823                         i_size_write(inode, *offset);
824                         xip->i_update_core = 1;
825                         xip->i_update_size = 1;
826                 }
827                 xfs_iunlock(xip, XFS_ILOCK_EXCL);
828         }
829
830         error = -ret;
831         if (ret <= 0)
832                 goto out_unlock_internal;
833
834         XFS_STATS_ADD(xs_write_bytes, ret);
835
836         /* Handle various SYNC-type writes */
837         if ((file->f_flags & O_SYNC) || IS_SYNC(inode)) {
838                 /*
839                  * If we're treating this as O_DSYNC and we have not updated the
840                  * size, force the log.
841                  */
842                 if (!(mp->m_flags & XFS_MOUNT_OSYNCISOSYNC) &&
843                     !(xip->i_update_size)) {
844                         xfs_inode_log_item_t    *iip = xip->i_itemp;
845
846                         /*
847                          * If an allocation transaction occurred
848                          * without extending the size, then we have to force
849                          * the log up the proper point to ensure that the
850                          * allocation is permanent.  We can't count on
851                          * the fact that buffered writes lock out direct I/O
852                          * writes - the direct I/O write could have extended
853                          * the size nontransactionally, then finished before
854                          * we started.  xfs_write_file will think that the file
855                          * didn't grow but the update isn't safe unless the
856                          * size change is logged.
857                          *
858                          * Force the log if we've committed a transaction
859                          * against the inode or if someone else has and
860                          * the commit record hasn't gone to disk (e.g.
861                          * the inode is pinned).  This guarantees that
862                          * all changes affecting the inode are permanent
863                          * when we return.
864                          */
865                         if (iip && iip->ili_last_lsn) {
866                                 xfs_log_force(mp, iip->ili_last_lsn,
867                                                 XFS_LOG_FORCE | XFS_LOG_SYNC);
868                         } else if (xfs_ipincount(xip) > 0) {
869                                 xfs_log_force(mp, (xfs_lsn_t)0,
870                                                 XFS_LOG_FORCE | XFS_LOG_SYNC);
871                         }
872
873                 } else {
874                         xfs_trans_t     *tp;
875
876                         /*
877                          * O_SYNC or O_DSYNC _with_ a size update are handled
878                          * the same way.
879                          *
880                          * If the write was synchronous then we need to make
881                          * sure that the inode modification time is permanent.
882                          * We'll have updated the timestamp above, so here
883                          * we use a synchronous transaction to log the inode.
884                          * It's not fast, but it's necessary.
885                          *
886                          * If this a dsync write and the size got changed
887                          * non-transactionally, then we need to ensure that
888                          * the size change gets logged in a synchronous
889                          * transaction.
890                          */
891
892                         tp = xfs_trans_alloc(mp, XFS_TRANS_WRITE_SYNC);
893                         if ((error = xfs_trans_reserve(tp, 0,
894                                                       XFS_SWRITE_LOG_RES(mp),
895                                                       0, 0, 0))) {
896                                 /* Transaction reserve failed */
897                                 xfs_trans_cancel(tp, 0);
898                         } else {
899                                 /* Transaction reserve successful */
900                                 xfs_ilock(xip, XFS_ILOCK_EXCL);
901                                 xfs_trans_ijoin(tp, xip, XFS_ILOCK_EXCL);
902                                 xfs_trans_ihold(tp, xip);
903                                 xfs_trans_log_inode(tp, xip, XFS_ILOG_CORE);
904                                 xfs_trans_set_sync(tp);
905                                 error = xfs_trans_commit(tp, 0, NULL);
906                                 xfs_iunlock(xip, XFS_ILOCK_EXCL);
907                         }
908                         if (error)
909                                 goto out_unlock_internal;
910                 }
911
912                 xfs_rwunlock(bdp, locktype);
913                 if (need_i_mutex)
914                         mutex_unlock(&inode->i_mutex);
915
916                 error = sync_page_range(inode, mapping, pos, ret);
917                 if (!error)
918                         error = ret;
919                 return error;
920         }
921
922  out_unlock_internal:
923         xfs_rwunlock(bdp, locktype);
924  out_unlock_mutex:
925         if (need_i_mutex)
926                 mutex_unlock(&inode->i_mutex);
927  out_nounlocks:
928         return -error;
929 }
930
931 /*
932  * All xfs metadata buffers except log state machine buffers
933  * get this attached as their b_bdstrat callback function.
934  * This is so that we can catch a buffer
935  * after prematurely unpinning it to forcibly shutdown the filesystem.
936  */
937 int
938 xfs_bdstrat_cb(struct xfs_buf *bp)
939 {
940         xfs_mount_t     *mp;
941
942         mp = XFS_BUF_FSPRIVATE3(bp, xfs_mount_t *);
943         if (!XFS_FORCED_SHUTDOWN(mp)) {
944                 xfs_buf_iorequest(bp);
945                 return 0;
946         } else {
947                 xfs_buftrace("XFS__BDSTRAT IOERROR", bp);
948                 /*
949                  * Metadata write that didn't get logged but
950                  * written delayed anyway. These aren't associated
951                  * with a transaction, and can be ignored.
952                  */
953                 if (XFS_BUF_IODONE_FUNC(bp) == NULL &&
954                     (XFS_BUF_ISREAD(bp)) == 0)
955                         return (xfs_bioerror_relse(bp));
956                 else
957                         return (xfs_bioerror(bp));
958         }
959 }
960
961
962 int
963 xfs_bmap(bhv_desc_t     *bdp,
964         xfs_off_t       offset,
965         ssize_t         count,
966         int             flags,
967         xfs_iomap_t     *iomapp,
968         int             *niomaps)
969 {
970         xfs_inode_t     *ip = XFS_BHVTOI(bdp);
971         xfs_iocore_t    *io = &ip->i_iocore;
972
973         ASSERT((ip->i_d.di_mode & S_IFMT) == S_IFREG);
974         ASSERT(((ip->i_d.di_flags & XFS_DIFLAG_REALTIME) != 0) ==
975                ((ip->i_iocore.io_flags & XFS_IOCORE_RT) != 0));
976
977         return xfs_iomap(io, offset, count, flags, iomapp, niomaps);
978 }
979
980 /*
981  * Wrapper around bdstrat so that we can stop data
982  * from going to disk in case we are shutting down the filesystem.
983  * Typically user data goes thru this path; one of the exceptions
984  * is the superblock.
985  */
986 int
987 xfsbdstrat(
988         struct xfs_mount        *mp,
989         struct xfs_buf          *bp)
990 {
991         ASSERT(mp);
992         if (!XFS_FORCED_SHUTDOWN(mp)) {
993                 /* Grio redirection would go here
994                  * if (XFS_BUF_IS_GRIO(bp)) {
995                  */
996
997                 xfs_buf_iorequest(bp);
998                 return 0;
999         }
1000
1001         xfs_buftrace("XFSBDSTRAT IOERROR", bp);
1002         return (xfs_bioerror_relse(bp));
1003 }
1004
1005 /*
1006  * If the underlying (data/log/rt) device is readonly, there are some
1007  * operations that cannot proceed.
1008  */
1009 int
1010 xfs_dev_is_read_only(
1011         xfs_mount_t             *mp,
1012         char                    *message)
1013 {
1014         if (xfs_readonly_buftarg(mp->m_ddev_targp) ||
1015             xfs_readonly_buftarg(mp->m_logdev_targp) ||
1016             (mp->m_rtdev_targp && xfs_readonly_buftarg(mp->m_rtdev_targp))) {
1017                 cmn_err(CE_NOTE,
1018                         "XFS: %s required on read-only device.", message);
1019                 cmn_err(CE_NOTE,
1020                         "XFS: write access unavailable, cannot proceed.");
1021                 return EROFS;
1022         }
1023         return 0;
1024 }