Merge git://git.kernel.org/pub/scm/linux/kernel/git/sfrench/cifs-2.6
[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_dir2.h"
27 #include "xfs_alloc.h"
28 #include "xfs_dmapi.h"
29 #include "xfs_quota.h"
30 #include "xfs_mount.h"
31 #include "xfs_bmap_btree.h"
32 #include "xfs_alloc_btree.h"
33 #include "xfs_ialloc_btree.h"
34 #include "xfs_dir2_sf.h"
35 #include "xfs_attr_sf.h"
36 #include "xfs_dinode.h"
37 #include "xfs_inode.h"
38 #include "xfs_bmap.h"
39 #include "xfs_btree.h"
40 #include "xfs_ialloc.h"
41 #include "xfs_rtalloc.h"
42 #include "xfs_error.h"
43 #include "xfs_itable.h"
44 #include "xfs_rw.h"
45 #include "xfs_acl.h"
46 #include "xfs_attr.h"
47 #include "xfs_inode_item.h"
48 #include "xfs_buf_item.h"
49 #include "xfs_utils.h"
50 #include "xfs_iomap.h"
51 #include "xfs_vnodeops.h"
52
53 #include <linux/capability.h>
54 #include <linux/mount.h>
55 #include <linux/writeback.h>
56
57
58 #if defined(XFS_RW_TRACE)
59 void
60 xfs_rw_enter_trace(
61         int                     tag,
62         xfs_inode_t             *ip,
63         void                    *data,
64         size_t                  segs,
65         loff_t                  offset,
66         int                     ioflags)
67 {
68         if (ip->i_rwtrace == NULL)
69                 return;
70         ktrace_enter(ip->i_rwtrace,
71                 (void *)(unsigned long)tag,
72                 (void *)ip,
73                 (void *)((unsigned long)((ip->i_d.di_size >> 32) & 0xffffffff)),
74                 (void *)((unsigned long)(ip->i_d.di_size & 0xffffffff)),
75                 (void *)data,
76                 (void *)((unsigned long)segs),
77                 (void *)((unsigned long)((offset >> 32) & 0xffffffff)),
78                 (void *)((unsigned long)(offset & 0xffffffff)),
79                 (void *)((unsigned long)ioflags),
80                 (void *)((unsigned long)((ip->i_new_size >> 32) & 0xffffffff)),
81                 (void *)((unsigned long)(ip->i_new_size & 0xffffffff)),
82                 (void *)((unsigned long)current_pid()),
83                 (void *)NULL,
84                 (void *)NULL,
85                 (void *)NULL,
86                 (void *)NULL);
87 }
88
89 void
90 xfs_inval_cached_trace(
91         xfs_inode_t     *ip,
92         xfs_off_t       offset,
93         xfs_off_t       len,
94         xfs_off_t       first,
95         xfs_off_t       last)
96 {
97
98         if (ip->i_rwtrace == NULL)
99                 return;
100         ktrace_enter(ip->i_rwtrace,
101                 (void *)(__psint_t)XFS_INVAL_CACHED,
102                 (void *)ip,
103                 (void *)((unsigned long)((offset >> 32) & 0xffffffff)),
104                 (void *)((unsigned long)(offset & 0xffffffff)),
105                 (void *)((unsigned long)((len >> 32) & 0xffffffff)),
106                 (void *)((unsigned long)(len & 0xffffffff)),
107                 (void *)((unsigned long)((first >> 32) & 0xffffffff)),
108                 (void *)((unsigned long)(first & 0xffffffff)),
109                 (void *)((unsigned long)((last >> 32) & 0xffffffff)),
110                 (void *)((unsigned long)(last & 0xffffffff)),
111                 (void *)((unsigned long)current_pid()),
112                 (void *)NULL,
113                 (void *)NULL,
114                 (void *)NULL,
115                 (void *)NULL,
116                 (void *)NULL);
117 }
118 #endif
119
120 /*
121  *      xfs_iozero
122  *
123  *      xfs_iozero clears the specified range of buffer supplied,
124  *      and marks all the affected blocks as valid and modified.  If
125  *      an affected block is not allocated, it will be allocated.  If
126  *      an affected block is not completely overwritten, and is not
127  *      valid before the operation, it will be read from disk before
128  *      being partially zeroed.
129  */
130 STATIC int
131 xfs_iozero(
132         struct xfs_inode        *ip,    /* inode                        */
133         loff_t                  pos,    /* offset in file               */
134         size_t                  count)  /* size of data to zero         */
135 {
136         struct page             *page;
137         struct address_space    *mapping;
138         int                     status;
139
140         mapping = VFS_I(ip)->i_mapping;
141         do {
142                 unsigned offset, bytes;
143                 void *fsdata;
144
145                 offset = (pos & (PAGE_CACHE_SIZE -1)); /* Within page */
146                 bytes = PAGE_CACHE_SIZE - offset;
147                 if (bytes > count)
148                         bytes = count;
149
150                 status = pagecache_write_begin(NULL, mapping, pos, bytes,
151                                         AOP_FLAG_UNINTERRUPTIBLE,
152                                         &page, &fsdata);
153                 if (status)
154                         break;
155
156                 zero_user(page, offset, bytes);
157
158                 status = pagecache_write_end(NULL, mapping, pos, bytes, bytes,
159                                         page, fsdata);
160                 WARN_ON(status <= 0); /* can't return less than zero! */
161                 pos += bytes;
162                 count -= bytes;
163                 status = 0;
164         } while (count);
165
166         return (-status);
167 }
168
169 ssize_t                 /* bytes read, or (-)  error */
170 xfs_read(
171         xfs_inode_t             *ip,
172         struct kiocb            *iocb,
173         const struct iovec      *iovp,
174         unsigned int            segs,
175         loff_t                  *offset,
176         int                     ioflags)
177 {
178         struct file             *file = iocb->ki_filp;
179         struct inode            *inode = file->f_mapping->host;
180         xfs_mount_t             *mp = ip->i_mount;
181         size_t                  size = 0;
182         ssize_t                 ret = 0;
183         xfs_fsize_t             n;
184         unsigned long           seg;
185
186
187         XFS_STATS_INC(xs_read_calls);
188
189         /* START copy & waste from filemap.c */
190         for (seg = 0; seg < segs; seg++) {
191                 const struct iovec *iv = &iovp[seg];
192
193                 /*
194                  * If any segment has a negative length, or the cumulative
195                  * length ever wraps negative then return -EINVAL.
196                  */
197                 size += iv->iov_len;
198                 if (unlikely((ssize_t)(size|iv->iov_len) < 0))
199                         return XFS_ERROR(-EINVAL);
200         }
201         /* END copy & waste from filemap.c */
202
203         if (unlikely(ioflags & IO_ISDIRECT)) {
204                 xfs_buftarg_t   *target =
205                         XFS_IS_REALTIME_INODE(ip) ?
206                                 mp->m_rtdev_targp : mp->m_ddev_targp;
207                 if ((*offset & target->bt_smask) ||
208                     (size & target->bt_smask)) {
209                         if (*offset == ip->i_size) {
210                                 return (0);
211                         }
212                         return -XFS_ERROR(EINVAL);
213                 }
214         }
215
216         n = XFS_MAXIOFFSET(mp) - *offset;
217         if ((n <= 0) || (size == 0))
218                 return 0;
219
220         if (n < size)
221                 size = n;
222
223         if (XFS_FORCED_SHUTDOWN(mp))
224                 return -EIO;
225
226         if (unlikely(ioflags & IO_ISDIRECT))
227                 mutex_lock(&inode->i_mutex);
228         xfs_ilock(ip, XFS_IOLOCK_SHARED);
229
230         if (DM_EVENT_ENABLED(ip, DM_EVENT_READ) && !(ioflags & IO_INVIS)) {
231                 int dmflags = FILP_DELAY_FLAG(file) | DM_SEM_FLAG_RD(ioflags);
232                 int iolock = XFS_IOLOCK_SHARED;
233
234                 ret = -XFS_SEND_DATA(mp, DM_EVENT_READ, ip, *offset, size,
235                                         dmflags, &iolock);
236                 if (ret) {
237                         xfs_iunlock(ip, XFS_IOLOCK_SHARED);
238                         if (unlikely(ioflags & IO_ISDIRECT))
239                                 mutex_unlock(&inode->i_mutex);
240                         return ret;
241                 }
242         }
243
244         if (unlikely(ioflags & IO_ISDIRECT)) {
245                 if (inode->i_mapping->nrpages)
246                         ret = xfs_flushinval_pages(ip, (*offset & PAGE_CACHE_MASK),
247                                                     -1, FI_REMAPF_LOCKED);
248                 mutex_unlock(&inode->i_mutex);
249                 if (ret) {
250                         xfs_iunlock(ip, XFS_IOLOCK_SHARED);
251                         return ret;
252                 }
253         }
254
255         xfs_rw_enter_trace(XFS_READ_ENTER, ip,
256                                 (void *)iovp, segs, *offset, ioflags);
257
258         iocb->ki_pos = *offset;
259         ret = generic_file_aio_read(iocb, iovp, segs, *offset);
260         if (ret == -EIOCBQUEUED && !(ioflags & IO_ISAIO))
261                 ret = wait_on_sync_kiocb(iocb);
262         if (ret > 0)
263                 XFS_STATS_ADD(xs_read_bytes, ret);
264
265         xfs_iunlock(ip, XFS_IOLOCK_SHARED);
266         return ret;
267 }
268
269 ssize_t
270 xfs_splice_read(
271         xfs_inode_t             *ip,
272         struct file             *infilp,
273         loff_t                  *ppos,
274         struct pipe_inode_info  *pipe,
275         size_t                  count,
276         int                     flags,
277         int                     ioflags)
278 {
279         xfs_mount_t             *mp = ip->i_mount;
280         ssize_t                 ret;
281
282         XFS_STATS_INC(xs_read_calls);
283         if (XFS_FORCED_SHUTDOWN(ip->i_mount))
284                 return -EIO;
285
286         xfs_ilock(ip, XFS_IOLOCK_SHARED);
287
288         if (DM_EVENT_ENABLED(ip, DM_EVENT_READ) && !(ioflags & IO_INVIS)) {
289                 int iolock = XFS_IOLOCK_SHARED;
290                 int error;
291
292                 error = XFS_SEND_DATA(mp, DM_EVENT_READ, ip, *ppos, count,
293                                         FILP_DELAY_FLAG(infilp), &iolock);
294                 if (error) {
295                         xfs_iunlock(ip, XFS_IOLOCK_SHARED);
296                         return -error;
297                 }
298         }
299         xfs_rw_enter_trace(XFS_SPLICE_READ_ENTER, ip,
300                            pipe, count, *ppos, ioflags);
301         ret = generic_file_splice_read(infilp, ppos, pipe, count, flags);
302         if (ret > 0)
303                 XFS_STATS_ADD(xs_read_bytes, ret);
304
305         xfs_iunlock(ip, XFS_IOLOCK_SHARED);
306         return ret;
307 }
308
309 ssize_t
310 xfs_splice_write(
311         xfs_inode_t             *ip,
312         struct pipe_inode_info  *pipe,
313         struct file             *outfilp,
314         loff_t                  *ppos,
315         size_t                  count,
316         int                     flags,
317         int                     ioflags)
318 {
319         xfs_mount_t             *mp = ip->i_mount;
320         ssize_t                 ret;
321         struct inode            *inode = outfilp->f_mapping->host;
322         xfs_fsize_t             isize, new_size;
323
324         XFS_STATS_INC(xs_write_calls);
325         if (XFS_FORCED_SHUTDOWN(ip->i_mount))
326                 return -EIO;
327
328         xfs_ilock(ip, XFS_IOLOCK_EXCL);
329
330         if (DM_EVENT_ENABLED(ip, DM_EVENT_WRITE) && !(ioflags & IO_INVIS)) {
331                 int iolock = XFS_IOLOCK_EXCL;
332                 int error;
333
334                 error = XFS_SEND_DATA(mp, DM_EVENT_WRITE, ip, *ppos, count,
335                                         FILP_DELAY_FLAG(outfilp), &iolock);
336                 if (error) {
337                         xfs_iunlock(ip, XFS_IOLOCK_EXCL);
338                         return -error;
339                 }
340         }
341
342         new_size = *ppos + count;
343
344         xfs_ilock(ip, XFS_ILOCK_EXCL);
345         if (new_size > ip->i_size)
346                 ip->i_new_size = new_size;
347         xfs_iunlock(ip, XFS_ILOCK_EXCL);
348
349         xfs_rw_enter_trace(XFS_SPLICE_WRITE_ENTER, ip,
350                            pipe, count, *ppos, ioflags);
351         ret = generic_file_splice_write(pipe, outfilp, ppos, count, flags);
352         if (ret > 0)
353                 XFS_STATS_ADD(xs_write_bytes, ret);
354
355         isize = i_size_read(inode);
356         if (unlikely(ret < 0 && ret != -EFAULT && *ppos > isize))
357                 *ppos = isize;
358
359         if (*ppos > ip->i_size) {
360                 xfs_ilock(ip, XFS_ILOCK_EXCL);
361                 if (*ppos > ip->i_size)
362                         ip->i_size = *ppos;
363                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
364         }
365
366         if (ip->i_new_size) {
367                 xfs_ilock(ip, XFS_ILOCK_EXCL);
368                 ip->i_new_size = 0;
369                 if (ip->i_d.di_size > ip->i_size)
370                         ip->i_d.di_size = ip->i_size;
371                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
372         }
373         xfs_iunlock(ip, XFS_IOLOCK_EXCL);
374         return ret;
375 }
376
377 /*
378  * This routine is called to handle zeroing any space in the last
379  * block of the file that is beyond the EOF.  We do this since the
380  * size is being increased without writing anything to that block
381  * and we don't want anyone to read the garbage on the disk.
382  */
383 STATIC int                              /* error (positive) */
384 xfs_zero_last_block(
385         xfs_inode_t     *ip,
386         xfs_fsize_t     offset,
387         xfs_fsize_t     isize)
388 {
389         xfs_fileoff_t   last_fsb;
390         xfs_mount_t     *mp = ip->i_mount;
391         int             nimaps;
392         int             zero_offset;
393         int             zero_len;
394         int             error = 0;
395         xfs_bmbt_irec_t imap;
396
397         ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
398
399         zero_offset = XFS_B_FSB_OFFSET(mp, isize);
400         if (zero_offset == 0) {
401                 /*
402                  * There are no extra bytes in the last block on disk to
403                  * zero, so return.
404                  */
405                 return 0;
406         }
407
408         last_fsb = XFS_B_TO_FSBT(mp, isize);
409         nimaps = 1;
410         error = xfs_bmapi(NULL, ip, last_fsb, 1, 0, NULL, 0, &imap,
411                           &nimaps, NULL, NULL);
412         if (error) {
413                 return error;
414         }
415         ASSERT(nimaps > 0);
416         /*
417          * If the block underlying isize is just a hole, then there
418          * is nothing to zero.
419          */
420         if (imap.br_startblock == HOLESTARTBLOCK) {
421                 return 0;
422         }
423         /*
424          * Zero the part of the last block beyond the EOF, and write it
425          * out sync.  We need to drop the ilock while we do this so we
426          * don't deadlock when the buffer cache calls back to us.
427          */
428         xfs_iunlock(ip, XFS_ILOCK_EXCL);
429
430         zero_len = mp->m_sb.sb_blocksize - zero_offset;
431         if (isize + zero_len > offset)
432                 zero_len = offset - isize;
433         error = xfs_iozero(ip, isize, zero_len);
434
435         xfs_ilock(ip, XFS_ILOCK_EXCL);
436         ASSERT(error >= 0);
437         return error;
438 }
439
440 /*
441  * Zero any on disk space between the current EOF and the new,
442  * larger EOF.  This handles the normal case of zeroing the remainder
443  * of the last block in the file and the unusual case of zeroing blocks
444  * out beyond the size of the file.  This second case only happens
445  * with fixed size extents and when the system crashes before the inode
446  * size was updated but after blocks were allocated.  If fill is set,
447  * then any holes in the range are filled and zeroed.  If not, the holes
448  * are left alone as holes.
449  */
450
451 int                                     /* error (positive) */
452 xfs_zero_eof(
453         xfs_inode_t     *ip,
454         xfs_off_t       offset,         /* starting I/O offset */
455         xfs_fsize_t     isize)          /* current inode size */
456 {
457         xfs_mount_t     *mp = ip->i_mount;
458         xfs_fileoff_t   start_zero_fsb;
459         xfs_fileoff_t   end_zero_fsb;
460         xfs_fileoff_t   zero_count_fsb;
461         xfs_fileoff_t   last_fsb;
462         xfs_fileoff_t   zero_off;
463         xfs_fsize_t     zero_len;
464         int             nimaps;
465         int             error = 0;
466         xfs_bmbt_irec_t imap;
467
468         ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL|XFS_IOLOCK_EXCL));
469         ASSERT(offset > isize);
470
471         /*
472          * First handle zeroing the block on which isize resides.
473          * We only zero a part of that block so it is handled specially.
474          */
475         error = xfs_zero_last_block(ip, offset, isize);
476         if (error) {
477                 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL|XFS_IOLOCK_EXCL));
478                 return error;
479         }
480
481         /*
482          * Calculate the range between the new size and the old
483          * where blocks needing to be zeroed may exist.  To get the
484          * block where the last byte in the file currently resides,
485          * we need to subtract one from the size and truncate back
486          * to a block boundary.  We subtract 1 in case the size is
487          * exactly on a block boundary.
488          */
489         last_fsb = isize ? XFS_B_TO_FSBT(mp, isize - 1) : (xfs_fileoff_t)-1;
490         start_zero_fsb = XFS_B_TO_FSB(mp, (xfs_ufsize_t)isize);
491         end_zero_fsb = XFS_B_TO_FSBT(mp, offset - 1);
492         ASSERT((xfs_sfiloff_t)last_fsb < (xfs_sfiloff_t)start_zero_fsb);
493         if (last_fsb == end_zero_fsb) {
494                 /*
495                  * The size was only incremented on its last block.
496                  * We took care of that above, so just return.
497                  */
498                 return 0;
499         }
500
501         ASSERT(start_zero_fsb <= end_zero_fsb);
502         while (start_zero_fsb <= end_zero_fsb) {
503                 nimaps = 1;
504                 zero_count_fsb = end_zero_fsb - start_zero_fsb + 1;
505                 error = xfs_bmapi(NULL, ip, start_zero_fsb, zero_count_fsb,
506                                   0, NULL, 0, &imap, &nimaps, NULL, NULL);
507                 if (error) {
508                         ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL|XFS_IOLOCK_EXCL));
509                         return error;
510                 }
511                 ASSERT(nimaps > 0);
512
513                 if (imap.br_state == XFS_EXT_UNWRITTEN ||
514                     imap.br_startblock == HOLESTARTBLOCK) {
515                         /*
516                          * This loop handles initializing pages that were
517                          * partially initialized by the code below this
518                          * loop. It basically zeroes the part of the page
519                          * that sits on a hole and sets the page as P_HOLE
520                          * and calls remapf if it is a mapped file.
521                          */
522                         start_zero_fsb = imap.br_startoff + imap.br_blockcount;
523                         ASSERT(start_zero_fsb <= (end_zero_fsb + 1));
524                         continue;
525                 }
526
527                 /*
528                  * There are blocks we need to zero.
529                  * Drop the inode lock while we're doing the I/O.
530                  * We'll still have the iolock to protect us.
531                  */
532                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
533
534                 zero_off = XFS_FSB_TO_B(mp, start_zero_fsb);
535                 zero_len = XFS_FSB_TO_B(mp, imap.br_blockcount);
536
537                 if ((zero_off + zero_len) > offset)
538                         zero_len = offset - zero_off;
539
540                 error = xfs_iozero(ip, zero_off, zero_len);
541                 if (error) {
542                         goto out_lock;
543                 }
544
545                 start_zero_fsb = imap.br_startoff + imap.br_blockcount;
546                 ASSERT(start_zero_fsb <= (end_zero_fsb + 1));
547
548                 xfs_ilock(ip, XFS_ILOCK_EXCL);
549         }
550
551         return 0;
552
553 out_lock:
554         xfs_ilock(ip, XFS_ILOCK_EXCL);
555         ASSERT(error >= 0);
556         return error;
557 }
558
559 ssize_t                         /* bytes written, or (-) error */
560 xfs_write(
561         struct xfs_inode        *xip,
562         struct kiocb            *iocb,
563         const struct iovec      *iovp,
564         unsigned int            nsegs,
565         loff_t                  *offset,
566         int                     ioflags)
567 {
568         struct file             *file = iocb->ki_filp;
569         struct address_space    *mapping = file->f_mapping;
570         struct inode            *inode = mapping->host;
571         unsigned long           segs = nsegs;
572         xfs_mount_t             *mp;
573         ssize_t                 ret = 0, error = 0;
574         xfs_fsize_t             isize, new_size;
575         int                     iolock;
576         int                     eventsent = 0;
577         size_t                  ocount = 0, count;
578         loff_t                  pos;
579         int                     need_i_mutex;
580
581         XFS_STATS_INC(xs_write_calls);
582
583         error = generic_segment_checks(iovp, &segs, &ocount, VERIFY_READ);
584         if (error)
585                 return error;
586
587         count = ocount;
588         pos = *offset;
589
590         if (count == 0)
591                 return 0;
592
593         mp = xip->i_mount;
594
595         xfs_wait_for_freeze(mp, SB_FREEZE_WRITE);
596
597         if (XFS_FORCED_SHUTDOWN(mp))
598                 return -EIO;
599
600 relock:
601         if (ioflags & IO_ISDIRECT) {
602                 iolock = XFS_IOLOCK_SHARED;
603                 need_i_mutex = 0;
604         } else {
605                 iolock = XFS_IOLOCK_EXCL;
606                 need_i_mutex = 1;
607                 mutex_lock(&inode->i_mutex);
608         }
609
610         xfs_ilock(xip, XFS_ILOCK_EXCL|iolock);
611
612 start:
613         error = -generic_write_checks(file, &pos, &count,
614                                         S_ISBLK(inode->i_mode));
615         if (error) {
616                 xfs_iunlock(xip, XFS_ILOCK_EXCL|iolock);
617                 goto out_unlock_mutex;
618         }
619
620         if ((DM_EVENT_ENABLED(xip, DM_EVENT_WRITE) &&
621             !(ioflags & IO_INVIS) && !eventsent)) {
622                 int             dmflags = FILP_DELAY_FLAG(file);
623
624                 if (need_i_mutex)
625                         dmflags |= DM_FLAGS_IMUX;
626
627                 xfs_iunlock(xip, XFS_ILOCK_EXCL);
628                 error = XFS_SEND_DATA(xip->i_mount, DM_EVENT_WRITE, xip,
629                                       pos, count, dmflags, &iolock);
630                 if (error) {
631                         goto out_unlock_internal;
632                 }
633                 xfs_ilock(xip, XFS_ILOCK_EXCL);
634                 eventsent = 1;
635
636                 /*
637                  * The iolock was dropped and reacquired in XFS_SEND_DATA
638                  * so we have to recheck the size when appending.
639                  * We will only "goto start;" once, since having sent the
640                  * event prevents another call to XFS_SEND_DATA, which is
641                  * what allows the size to change in the first place.
642                  */
643                 if ((file->f_flags & O_APPEND) && pos != xip->i_size)
644                         goto start;
645         }
646
647         if (ioflags & IO_ISDIRECT) {
648                 xfs_buftarg_t   *target =
649                         XFS_IS_REALTIME_INODE(xip) ?
650                                 mp->m_rtdev_targp : mp->m_ddev_targp;
651
652                 if ((pos & target->bt_smask) || (count & target->bt_smask)) {
653                         xfs_iunlock(xip, XFS_ILOCK_EXCL|iolock);
654                         return XFS_ERROR(-EINVAL);
655                 }
656
657                 if (!need_i_mutex && (mapping->nrpages || pos > xip->i_size)) {
658                         xfs_iunlock(xip, XFS_ILOCK_EXCL|iolock);
659                         iolock = XFS_IOLOCK_EXCL;
660                         need_i_mutex = 1;
661                         mutex_lock(&inode->i_mutex);
662                         xfs_ilock(xip, XFS_ILOCK_EXCL|iolock);
663                         goto start;
664                 }
665         }
666
667         new_size = pos + count;
668         if (new_size > xip->i_size)
669                 xip->i_new_size = new_size;
670
671         /*
672          * We're not supposed to change timestamps in readonly-mounted
673          * filesystems.  Throw it away if anyone asks us.
674          */
675         if (likely(!(ioflags & IO_INVIS) &&
676                    !mnt_want_write(file->f_path.mnt))) {
677                 xfs_ichgtime(xip, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
678                 mnt_drop_write(file->f_path.mnt);
679         }
680
681         /*
682          * If the offset is beyond the size of the file, we have a couple
683          * of things to do. First, if there is already space allocated
684          * we need to either create holes or zero the disk or ...
685          *
686          * If there is a page where the previous size lands, we need
687          * to zero it out up to the new size.
688          */
689
690         if (pos > xip->i_size) {
691                 error = xfs_zero_eof(xip, pos, xip->i_size);
692                 if (error) {
693                         xfs_iunlock(xip, XFS_ILOCK_EXCL);
694                         goto out_unlock_internal;
695                 }
696         }
697         xfs_iunlock(xip, XFS_ILOCK_EXCL);
698
699         /*
700          * If we're writing the file then make sure to clear the
701          * setuid and setgid bits if the process is not being run
702          * by root.  This keeps people from modifying setuid and
703          * setgid binaries.
704          */
705
706         if (((xip->i_d.di_mode & S_ISUID) ||
707             ((xip->i_d.di_mode & (S_ISGID | S_IXGRP)) ==
708                 (S_ISGID | S_IXGRP))) &&
709              !capable(CAP_FSETID)) {
710                 error = xfs_write_clear_setuid(xip);
711                 if (likely(!error))
712                         error = -file_remove_suid(file);
713                 if (unlikely(error)) {
714                         goto out_unlock_internal;
715                 }
716         }
717
718 retry:
719         /* We can write back this queue in page reclaim */
720         current->backing_dev_info = mapping->backing_dev_info;
721
722         if ((ioflags & IO_ISDIRECT)) {
723                 if (mapping->nrpages) {
724                         WARN_ON(need_i_mutex == 0);
725                         xfs_inval_cached_trace(xip, pos, -1,
726                                         (pos & PAGE_CACHE_MASK), -1);
727                         error = xfs_flushinval_pages(xip,
728                                         (pos & PAGE_CACHE_MASK),
729                                         -1, FI_REMAPF_LOCKED);
730                         if (error)
731                                 goto out_unlock_internal;
732                 }
733
734                 if (need_i_mutex) {
735                         /* demote the lock now the cached pages are gone */
736                         xfs_ilock_demote(xip, XFS_IOLOCK_EXCL);
737                         mutex_unlock(&inode->i_mutex);
738
739                         iolock = XFS_IOLOCK_SHARED;
740                         need_i_mutex = 0;
741                 }
742
743                 xfs_rw_enter_trace(XFS_DIOWR_ENTER, xip, (void *)iovp, segs,
744                                 *offset, ioflags);
745                 ret = generic_file_direct_write(iocb, iovp,
746                                 &segs, pos, offset, count, ocount);
747
748                 /*
749                  * direct-io write to a hole: fall through to buffered I/O
750                  * for completing the rest of the request.
751                  */
752                 if (ret >= 0 && ret != count) {
753                         XFS_STATS_ADD(xs_write_bytes, ret);
754
755                         pos += ret;
756                         count -= ret;
757
758                         ioflags &= ~IO_ISDIRECT;
759                         xfs_iunlock(xip, iolock);
760                         goto relock;
761                 }
762         } else {
763                 xfs_rw_enter_trace(XFS_WRITE_ENTER, xip, (void *)iovp, segs,
764                                 *offset, ioflags);
765                 ret = generic_file_buffered_write(iocb, iovp, segs,
766                                 pos, offset, count, ret);
767         }
768
769         current->backing_dev_info = NULL;
770
771         if (ret == -EIOCBQUEUED && !(ioflags & IO_ISAIO))
772                 ret = wait_on_sync_kiocb(iocb);
773
774         if (ret == -ENOSPC &&
775             DM_EVENT_ENABLED(xip, DM_EVENT_NOSPACE) && !(ioflags & IO_INVIS)) {
776                 xfs_iunlock(xip, iolock);
777                 if (need_i_mutex)
778                         mutex_unlock(&inode->i_mutex);
779                 error = XFS_SEND_NAMESP(xip->i_mount, DM_EVENT_NOSPACE, xip,
780                                 DM_RIGHT_NULL, xip, DM_RIGHT_NULL, NULL, NULL,
781                                 0, 0, 0); /* Delay flag intentionally  unused */
782                 if (need_i_mutex)
783                         mutex_lock(&inode->i_mutex);
784                 xfs_ilock(xip, iolock);
785                 if (error)
786                         goto out_unlock_internal;
787                 pos = xip->i_size;
788                 ret = 0;
789                 goto retry;
790         }
791
792         isize = i_size_read(inode);
793         if (unlikely(ret < 0 && ret != -EFAULT && *offset > isize))
794                 *offset = isize;
795
796         if (*offset > xip->i_size) {
797                 xfs_ilock(xip, XFS_ILOCK_EXCL);
798                 if (*offset > xip->i_size)
799                         xip->i_size = *offset;
800                 xfs_iunlock(xip, XFS_ILOCK_EXCL);
801         }
802
803         error = -ret;
804         if (ret <= 0)
805                 goto out_unlock_internal;
806
807         XFS_STATS_ADD(xs_write_bytes, ret);
808
809         /* Handle various SYNC-type writes */
810         if ((file->f_flags & O_SYNC) || IS_SYNC(inode)) {
811                 int error2;
812
813                 xfs_iunlock(xip, iolock);
814                 if (need_i_mutex)
815                         mutex_unlock(&inode->i_mutex);
816                 error2 = sync_page_range(inode, mapping, pos, ret);
817                 if (!error)
818                         error = error2;
819                 if (need_i_mutex)
820                         mutex_lock(&inode->i_mutex);
821                 xfs_ilock(xip, iolock);
822                 error2 = xfs_write_sync_logforce(mp, xip);
823                 if (!error)
824                         error = error2;
825         }
826
827  out_unlock_internal:
828         if (xip->i_new_size) {
829                 xfs_ilock(xip, XFS_ILOCK_EXCL);
830                 xip->i_new_size = 0;
831                 /*
832                  * If this was a direct or synchronous I/O that failed (such
833                  * as ENOSPC) then part of the I/O may have been written to
834                  * disk before the error occured.  In this case the on-disk
835                  * file size may have been adjusted beyond the in-memory file
836                  * size and now needs to be truncated back.
837                  */
838                 if (xip->i_d.di_size > xip->i_size)
839                         xip->i_d.di_size = xip->i_size;
840                 xfs_iunlock(xip, XFS_ILOCK_EXCL);
841         }
842         xfs_iunlock(xip, iolock);
843  out_unlock_mutex:
844         if (need_i_mutex)
845                 mutex_unlock(&inode->i_mutex);
846         return -error;
847 }
848
849 /*
850  * All xfs metadata buffers except log state machine buffers
851  * get this attached as their b_bdstrat callback function.
852  * This is so that we can catch a buffer
853  * after prematurely unpinning it to forcibly shutdown the filesystem.
854  */
855 int
856 xfs_bdstrat_cb(struct xfs_buf *bp)
857 {
858         xfs_mount_t     *mp;
859
860         mp = XFS_BUF_FSPRIVATE3(bp, xfs_mount_t *);
861         if (!XFS_FORCED_SHUTDOWN(mp)) {
862                 xfs_buf_iorequest(bp);
863                 return 0;
864         } else {
865                 xfs_buftrace("XFS__BDSTRAT IOERROR", bp);
866                 /*
867                  * Metadata write that didn't get logged but
868                  * written delayed anyway. These aren't associated
869                  * with a transaction, and can be ignored.
870                  */
871                 if (XFS_BUF_IODONE_FUNC(bp) == NULL &&
872                     (XFS_BUF_ISREAD(bp)) == 0)
873                         return (xfs_bioerror_relse(bp));
874                 else
875                         return (xfs_bioerror(bp));
876         }
877 }
878
879 /*
880  * Wrapper around bdstrat so that we can stop data from going to disk in case
881  * we are shutting down the filesystem.  Typically user data goes thru this
882  * path; one of the exceptions is the superblock.
883  */
884 void
885 xfsbdstrat(
886         struct xfs_mount        *mp,
887         struct xfs_buf          *bp)
888 {
889         ASSERT(mp);
890         if (!XFS_FORCED_SHUTDOWN(mp)) {
891                 xfs_buf_iorequest(bp);
892                 return;
893         }
894
895         xfs_buftrace("XFSBDSTRAT IOERROR", bp);
896         xfs_bioerror_relse(bp);
897 }
898
899 /*
900  * If the underlying (data/log/rt) device is readonly, there are some
901  * operations that cannot proceed.
902  */
903 int
904 xfs_dev_is_read_only(
905         xfs_mount_t             *mp,
906         char                    *message)
907 {
908         if (xfs_readonly_buftarg(mp->m_ddev_targp) ||
909             xfs_readonly_buftarg(mp->m_logdev_targp) ||
910             (mp->m_rtdev_targp && xfs_readonly_buftarg(mp->m_rtdev_targp))) {
911                 cmn_err(CE_NOTE,
912                         "XFS: %s required on read-only device.", message);
913                 cmn_err(CE_NOTE,
914                         "XFS: write access unavailable, cannot proceed.");
915                 return EROFS;
916         }
917         return 0;
918 }