Merge branch 'master'
[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 *)NULL,
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 *)NULL,
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->pbr_smask) ||
237                     (size & target->pbr_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
256         if (unlikely(ioflags & IO_ISDIRECT))
257                 down(&inode->i_sem);
258         xfs_ilock(ip, XFS_IOLOCK_SHARED);
259
260         if (DM_EVENT_ENABLED(vp->v_vfsp, ip, DM_EVENT_READ) &&
261             !(ioflags & IO_INVIS)) {
262                 vrwlock_t locktype = VRWLOCK_READ;
263                 int dmflags = FILP_DELAY_FLAG(file) | DM_SEM_FLAG_RD(ioflags);
264
265                 ret = -XFS_SEND_DATA(mp, DM_EVENT_READ,
266                                         BHV_TO_VNODE(bdp), *offset, size,
267                                         dmflags, &locktype);
268                 if (ret) {
269                         xfs_iunlock(ip, XFS_IOLOCK_SHARED);
270                         goto unlock_isem;
271                 }
272         }
273
274         xfs_rw_enter_trace(XFS_READ_ENTER, &ip->i_iocore,
275                                 (void *)iovp, segs, *offset, ioflags);
276         ret = __generic_file_aio_read(iocb, iovp, segs, offset);
277         if (ret == -EIOCBQUEUED && !(ioflags & IO_ISAIO))
278                 ret = wait_on_sync_kiocb(iocb);
279         if (ret > 0)
280                 XFS_STATS_ADD(xs_read_bytes, ret);
281
282         xfs_iunlock(ip, XFS_IOLOCK_SHARED);
283
284         if (likely(!(ioflags & IO_INVIS)))
285                 xfs_ichgtime_fast(ip, inode, XFS_ICHGTIME_ACC);
286
287 unlock_isem:
288         if (unlikely(ioflags & IO_ISDIRECT))
289                 up(&inode->i_sem);
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         if (likely(!(ioflags & IO_INVIS)))
350                 xfs_ichgtime_fast(ip, LINVFS_GET_IP(vp), XFS_ICHGTIME_ACC);
351
352         return ret;
353 }
354
355 /*
356  * This routine is called to handle zeroing any space in the last
357  * block of the file that is beyond the EOF.  We do this since the
358  * size is being increased without writing anything to that block
359  * and we don't want anyone to read the garbage on the disk.
360  */
361 STATIC int                              /* error (positive) */
362 xfs_zero_last_block(
363         struct inode    *ip,
364         xfs_iocore_t    *io,
365         xfs_off_t       offset,
366         xfs_fsize_t     isize,
367         xfs_fsize_t     end_size)
368 {
369         xfs_fileoff_t   last_fsb;
370         xfs_mount_t     *mp;
371         int             nimaps;
372         int             zero_offset;
373         int             zero_len;
374         int             isize_fsb_offset;
375         int             error = 0;
376         xfs_bmbt_irec_t imap;
377         loff_t          loff;
378         size_t          lsize;
379
380         ASSERT(ismrlocked(io->io_lock, MR_UPDATE) != 0);
381         ASSERT(offset > isize);
382
383         mp = io->io_mount;
384
385         isize_fsb_offset = XFS_B_FSB_OFFSET(mp, isize);
386         if (isize_fsb_offset == 0) {
387                 /*
388                  * There are no extra bytes in the last block on disk to
389                  * zero, so return.
390                  */
391                 return 0;
392         }
393
394         last_fsb = XFS_B_TO_FSBT(mp, isize);
395         nimaps = 1;
396         error = XFS_BMAPI(mp, NULL, io, last_fsb, 1, 0, NULL, 0, &imap,
397                           &nimaps, NULL);
398         if (error) {
399                 return error;
400         }
401         ASSERT(nimaps > 0);
402         /*
403          * If the block underlying isize is just a hole, then there
404          * is nothing to zero.
405          */
406         if (imap.br_startblock == HOLESTARTBLOCK) {
407                 return 0;
408         }
409         /*
410          * Zero the part of the last block beyond the EOF, and write it
411          * out sync.  We need to drop the ilock while we do this so we
412          * don't deadlock when the buffer cache calls back to us.
413          */
414         XFS_IUNLOCK(mp, io, XFS_ILOCK_EXCL| XFS_EXTSIZE_RD);
415         loff = XFS_FSB_TO_B(mp, last_fsb);
416         lsize = XFS_FSB_TO_B(mp, 1);
417
418         zero_offset = isize_fsb_offset;
419         zero_len = mp->m_sb.sb_blocksize - isize_fsb_offset;
420
421         error = xfs_iozero(ip, loff + zero_offset, zero_len, end_size);
422
423         XFS_ILOCK(mp, io, XFS_ILOCK_EXCL|XFS_EXTSIZE_RD);
424         ASSERT(error >= 0);
425         return error;
426 }
427
428 /*
429  * Zero any on disk space between the current EOF and the new,
430  * larger EOF.  This handles the normal case of zeroing the remainder
431  * of the last block in the file and the unusual case of zeroing blocks
432  * out beyond the size of the file.  This second case only happens
433  * with fixed size extents and when the system crashes before the inode
434  * size was updated but after blocks were allocated.  If fill is set,
435  * then any holes in the range are filled and zeroed.  If not, the holes
436  * are left alone as holes.
437  */
438
439 int                                     /* error (positive) */
440 xfs_zero_eof(
441         vnode_t         *vp,
442         xfs_iocore_t    *io,
443         xfs_off_t       offset,         /* starting I/O offset */
444         xfs_fsize_t     isize,          /* current inode size */
445         xfs_fsize_t     end_size)       /* terminal inode size */
446 {
447         struct inode    *ip = LINVFS_GET_IP(vp);
448         xfs_fileoff_t   start_zero_fsb;
449         xfs_fileoff_t   end_zero_fsb;
450         xfs_fileoff_t   prev_zero_fsb;
451         xfs_fileoff_t   zero_count_fsb;
452         xfs_fileoff_t   last_fsb;
453         xfs_extlen_t    buf_len_fsb;
454         xfs_extlen_t    prev_zero_count;
455         xfs_mount_t     *mp;
456         int             nimaps;
457         int             error = 0;
458         xfs_bmbt_irec_t imap;
459         loff_t          loff;
460         size_t          lsize;
461
462         ASSERT(ismrlocked(io->io_lock, MR_UPDATE));
463         ASSERT(ismrlocked(io->io_iolock, MR_UPDATE));
464
465         mp = io->io_mount;
466
467         /*
468          * First handle zeroing the block on which isize resides.
469          * We only zero a part of that block so it is handled specially.
470          */
471         error = xfs_zero_last_block(ip, io, offset, isize, end_size);
472         if (error) {
473                 ASSERT(ismrlocked(io->io_lock, MR_UPDATE));
474                 ASSERT(ismrlocked(io->io_iolock, MR_UPDATE));
475                 return error;
476         }
477
478         /*
479          * Calculate the range between the new size and the old
480          * where blocks needing to be zeroed may exist.  To get the
481          * block where the last byte in the file currently resides,
482          * we need to subtract one from the size and truncate back
483          * to a block boundary.  We subtract 1 in case the size is
484          * exactly on a block boundary.
485          */
486         last_fsb = isize ? XFS_B_TO_FSBT(mp, isize - 1) : (xfs_fileoff_t)-1;
487         start_zero_fsb = XFS_B_TO_FSB(mp, (xfs_ufsize_t)isize);
488         end_zero_fsb = XFS_B_TO_FSBT(mp, offset - 1);
489         ASSERT((xfs_sfiloff_t)last_fsb < (xfs_sfiloff_t)start_zero_fsb);
490         if (last_fsb == end_zero_fsb) {
491                 /*
492                  * The size was only incremented on its last block.
493                  * We took care of that above, so just return.
494                  */
495                 return 0;
496         }
497
498         ASSERT(start_zero_fsb <= end_zero_fsb);
499         prev_zero_fsb = NULLFILEOFF;
500         prev_zero_count = 0;
501         while (start_zero_fsb <= end_zero_fsb) {
502                 nimaps = 1;
503                 zero_count_fsb = end_zero_fsb - start_zero_fsb + 1;
504                 error = XFS_BMAPI(mp, NULL, io, start_zero_fsb, zero_count_fsb,
505                                   0, NULL, 0, &imap, &nimaps, NULL);
506                 if (error) {
507                         ASSERT(ismrlocked(io->io_lock, MR_UPDATE));
508                         ASSERT(ismrlocked(io->io_iolock, MR_UPDATE));
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                         prev_zero_fsb = NULLFILEOFF;
523                         prev_zero_count = 0;
524                         start_zero_fsb = imap.br_startoff +
525                                          imap.br_blockcount;
526                         ASSERT(start_zero_fsb <= (end_zero_fsb + 1));
527                         continue;
528                 }
529
530                 /*
531                  * There are blocks in the range requested.
532                  * Zero them a single write at a time.  We actually
533                  * don't zero the entire range returned if it is
534                  * too big and simply loop around to get the rest.
535                  * That is not the most efficient thing to do, but it
536                  * is simple and this path should not be exercised often.
537                  */
538                 buf_len_fsb = XFS_FILBLKS_MIN(imap.br_blockcount,
539                                               mp->m_writeio_blocks << 8);
540                 /*
541                  * Drop the inode lock while we're doing the I/O.
542                  * We'll still have the iolock to protect us.
543                  */
544                 XFS_IUNLOCK(mp, io, XFS_ILOCK_EXCL|XFS_EXTSIZE_RD);
545
546                 loff = XFS_FSB_TO_B(mp, start_zero_fsb);
547                 lsize = XFS_FSB_TO_B(mp, buf_len_fsb);
548
549                 error = xfs_iozero(ip, loff, lsize, end_size);
550
551                 if (error) {
552                         goto out_lock;
553                 }
554
555                 prev_zero_fsb = start_zero_fsb;
556                 prev_zero_count = buf_len_fsb;
557                 start_zero_fsb = imap.br_startoff + buf_len_fsb;
558                 ASSERT(start_zero_fsb <= (end_zero_fsb + 1));
559
560                 XFS_ILOCK(mp, io, XFS_ILOCK_EXCL|XFS_EXTSIZE_RD);
561         }
562
563         return 0;
564
565 out_lock:
566
567         XFS_ILOCK(mp, io, XFS_ILOCK_EXCL|XFS_EXTSIZE_RD);
568         ASSERT(error >= 0);
569         return error;
570 }
571
572 ssize_t                         /* bytes written, or (-) error */
573 xfs_write(
574         bhv_desc_t              *bdp,
575         struct kiocb            *iocb,
576         const struct iovec      *iovp,
577         unsigned int            nsegs,
578         loff_t                  *offset,
579         int                     ioflags,
580         cred_t                  *credp)
581 {
582         struct file             *file = iocb->ki_filp;
583         struct address_space    *mapping = file->f_mapping;
584         struct inode            *inode = mapping->host;
585         unsigned long           segs = nsegs;
586         xfs_inode_t             *xip;
587         xfs_mount_t             *mp;
588         ssize_t                 ret = 0, error = 0;
589         xfs_fsize_t             isize, new_size;
590         xfs_iocore_t            *io;
591         vnode_t                 *vp;
592         unsigned long           seg;
593         int                     iolock;
594         int                     eventsent = 0;
595         vrwlock_t               locktype;
596         size_t                  ocount = 0, count;
597         loff_t                  pos;
598         int                     need_isem = 1, need_flush = 0;
599
600         XFS_STATS_INC(xs_write_calls);
601
602         vp = BHV_TO_VNODE(bdp);
603         xip = XFS_BHVTOI(bdp);
604
605         for (seg = 0; seg < segs; seg++) {
606                 const struct iovec *iv = &iovp[seg];
607
608                 /*
609                  * If any segment has a negative length, or the cumulative
610                  * length ever wraps negative then return -EINVAL.
611                  */
612                 ocount += iv->iov_len;
613                 if (unlikely((ssize_t)(ocount|iv->iov_len) < 0))
614                         return -EINVAL;
615                 if (access_ok(VERIFY_READ, iv->iov_base, iv->iov_len))
616                         continue;
617                 if (seg == 0)
618                         return -EFAULT;
619                 segs = seg;
620                 ocount -= iv->iov_len;  /* This segment is no good */
621                 break;
622         }
623
624         count = ocount;
625         pos = *offset;
626
627         if (count == 0)
628                 return 0;
629
630         io = &xip->i_iocore;
631         mp = io->io_mount;
632
633         if (XFS_FORCED_SHUTDOWN(mp))
634                 return -EIO;
635
636         fs_check_frozen(vp->v_vfsp, SB_FREEZE_WRITE);
637
638         if (ioflags & IO_ISDIRECT) {
639                 xfs_buftarg_t   *target =
640                         (xip->i_d.di_flags & XFS_DIFLAG_REALTIME) ?
641                                 mp->m_rtdev_targp : mp->m_ddev_targp;
642
643                 if ((pos & target->pbr_smask) || (count & target->pbr_smask))
644                         return XFS_ERROR(-EINVAL);
645
646                 if (!VN_CACHED(vp) && pos < i_size_read(inode))
647                         need_isem = 0;
648
649                 if (VN_CACHED(vp))
650                         need_flush = 1;
651         }
652
653 relock:
654         if (need_isem) {
655                 iolock = XFS_IOLOCK_EXCL;
656                 locktype = VRWLOCK_WRITE;
657
658                 down(&inode->i_sem);
659         } else {
660                 iolock = XFS_IOLOCK_SHARED;
661                 locktype = VRWLOCK_WRITE_DIRECT;
662         }
663
664         xfs_ilock(xip, XFS_ILOCK_EXCL|iolock);
665
666         isize = i_size_read(inode);
667
668         if (file->f_flags & O_APPEND)
669                 *offset = isize;
670
671 start:
672         error = -generic_write_checks(file, &pos, &count,
673                                         S_ISBLK(inode->i_mode));
674         if (error) {
675                 xfs_iunlock(xip, XFS_ILOCK_EXCL|iolock);
676                 goto out_unlock_isem;
677         }
678
679         new_size = pos + count;
680         if (new_size > isize)
681                 io->io_new_size = new_size;
682
683         if ((DM_EVENT_ENABLED(vp->v_vfsp, xip, DM_EVENT_WRITE) &&
684             !(ioflags & IO_INVIS) && !eventsent)) {
685                 loff_t          savedsize = pos;
686                 int             dmflags = FILP_DELAY_FLAG(file);
687
688                 if (need_isem)
689                         dmflags |= DM_FLAGS_ISEM;
690
691                 xfs_iunlock(xip, XFS_ILOCK_EXCL);
692                 error = XFS_SEND_DATA(xip->i_mount, DM_EVENT_WRITE, vp,
693                                       pos, count,
694                                       dmflags, &locktype);
695                 if (error) {
696                         xfs_iunlock(xip, iolock);
697                         goto out_unlock_isem;
698                 }
699                 xfs_ilock(xip, XFS_ILOCK_EXCL);
700                 eventsent = 1;
701
702                 /*
703                  * The iolock was dropped and reaquired in XFS_SEND_DATA
704                  * so we have to recheck the size when appending.
705                  * We will only "goto start;" once, since having sent the
706                  * event prevents another call to XFS_SEND_DATA, which is
707                  * what allows the size to change in the first place.
708                  */
709                 if ((file->f_flags & O_APPEND) && savedsize != isize) {
710                         pos = isize = xip->i_d.di_size;
711                         goto start;
712                 }
713         }
714
715         if (likely(!(ioflags & IO_INVIS))) {
716                 inode_update_time(inode, 1);
717                 xfs_ichgtime_fast(xip, inode,
718                                   XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
719         }
720
721         /*
722          * If the offset is beyond the size of the file, we have a couple
723          * of things to do. First, if there is already space allocated
724          * we need to either create holes or zero the disk or ...
725          *
726          * If there is a page where the previous size lands, we need
727          * to zero it out up to the new size.
728          */
729
730         if (pos > isize) {
731                 error = xfs_zero_eof(BHV_TO_VNODE(bdp), io, pos,
732                                         isize, pos + count);
733                 if (error) {
734                         xfs_iunlock(xip, XFS_ILOCK_EXCL|iolock);
735                         goto out_unlock_isem;
736                 }
737         }
738         xfs_iunlock(xip, XFS_ILOCK_EXCL);
739
740         /*
741          * If we're writing the file then make sure to clear the
742          * setuid and setgid bits if the process is not being run
743          * by root.  This keeps people from modifying setuid and
744          * setgid binaries.
745          */
746
747         if (((xip->i_d.di_mode & S_ISUID) ||
748             ((xip->i_d.di_mode & (S_ISGID | S_IXGRP)) ==
749                 (S_ISGID | S_IXGRP))) &&
750              !capable(CAP_FSETID)) {
751                 error = xfs_write_clear_setuid(xip);
752                 if (likely(!error))
753                         error = -remove_suid(file->f_dentry);
754                 if (unlikely(error)) {
755                         xfs_iunlock(xip, iolock);
756                         goto out_unlock_isem;
757                 }
758         }
759
760 retry:
761         /* We can write back this queue in page reclaim */
762         current->backing_dev_info = mapping->backing_dev_info;
763
764         if ((ioflags & IO_ISDIRECT)) {
765                 if (need_flush) {
766                         xfs_inval_cached_trace(io, pos, -1,
767                                         ctooff(offtoct(pos)), -1);
768                         VOP_FLUSHINVAL_PAGES(vp, ctooff(offtoct(pos)),
769                                         -1, FI_REMAPF_LOCKED);
770                 }
771
772                 if (need_isem) {
773                         /* demote the lock now the cached pages are gone */
774                         XFS_ILOCK_DEMOTE(mp, io, XFS_IOLOCK_EXCL);
775                         up(&inode->i_sem);
776
777                         iolock = XFS_IOLOCK_SHARED;
778                         locktype = VRWLOCK_WRITE_DIRECT;
779                         need_isem = 0;
780                 }
781
782                 xfs_rw_enter_trace(XFS_DIOWR_ENTER, io, (void *)iovp, segs,
783                                 *offset, ioflags);
784                 ret = generic_file_direct_write(iocb, iovp,
785                                 &segs, pos, offset, count, ocount);
786
787                 /*
788                  * direct-io write to a hole: fall through to buffered I/O
789                  * for completing the rest of the request.
790                  */
791                 if (ret >= 0 && ret != count) {
792                         XFS_STATS_ADD(xs_write_bytes, ret);
793
794                         pos += ret;
795                         count -= ret;
796
797                         need_isem = 1;
798                         ioflags &= ~IO_ISDIRECT;
799                         xfs_iunlock(xip, iolock);
800                         goto relock;
801                 }
802         } else {
803                 xfs_rw_enter_trace(XFS_WRITE_ENTER, io, (void *)iovp, segs,
804                                 *offset, ioflags);
805                 ret = generic_file_buffered_write(iocb, iovp, segs,
806                                 pos, offset, count, ret);
807         }
808
809         current->backing_dev_info = NULL;
810
811         if (ret == -EIOCBQUEUED && !(ioflags & IO_ISAIO))
812                 ret = wait_on_sync_kiocb(iocb);
813
814         if ((ret == -ENOSPC) &&
815             DM_EVENT_ENABLED(vp->v_vfsp, xip, DM_EVENT_NOSPACE) &&
816             !(ioflags & IO_INVIS)) {
817
818                 xfs_rwunlock(bdp, locktype);
819                 if (need_isem)
820                         up(&inode->i_sem);
821                 error = XFS_SEND_NAMESP(xip->i_mount, DM_EVENT_NOSPACE, vp,
822                                 DM_RIGHT_NULL, vp, DM_RIGHT_NULL, NULL, NULL,
823                                 0, 0, 0); /* Delay flag intentionally  unused */
824                 if (error)
825                         goto out_nounlocks;
826                 if (need_isem)
827                         down(&inode->i_sem);
828                 xfs_rwlock(bdp, locktype);
829                 pos = xip->i_d.di_size;
830                 ret = 0;
831                 goto retry;
832         }
833
834         if (*offset > xip->i_d.di_size) {
835                 xfs_ilock(xip, XFS_ILOCK_EXCL);
836                 if (*offset > xip->i_d.di_size) {
837                         xip->i_d.di_size = *offset;
838                         i_size_write(inode, *offset);
839                         xip->i_update_core = 1;
840                         xip->i_update_size = 1;
841                 }
842                 xfs_iunlock(xip, XFS_ILOCK_EXCL);
843         }
844
845         error = -ret;
846         if (ret <= 0)
847                 goto out_unlock_internal;
848
849         XFS_STATS_ADD(xs_write_bytes, ret);
850
851         /* Handle various SYNC-type writes */
852         if ((file->f_flags & O_SYNC) || IS_SYNC(inode)) {
853                 /*
854                  * If we're treating this as O_DSYNC and we have not updated the
855                  * size, force the log.
856                  */
857                 if (!(mp->m_flags & XFS_MOUNT_OSYNCISOSYNC) &&
858                     !(xip->i_update_size)) {
859                         xfs_inode_log_item_t    *iip = xip->i_itemp;
860
861                         /*
862                          * If an allocation transaction occurred
863                          * without extending the size, then we have to force
864                          * the log up the proper point to ensure that the
865                          * allocation is permanent.  We can't count on
866                          * the fact that buffered writes lock out direct I/O
867                          * writes - the direct I/O write could have extended
868                          * the size nontransactionally, then finished before
869                          * we started.  xfs_write_file will think that the file
870                          * didn't grow but the update isn't safe unless the
871                          * size change is logged.
872                          *
873                          * Force the log if we've committed a transaction
874                          * against the inode or if someone else has and
875                          * the commit record hasn't gone to disk (e.g.
876                          * the inode is pinned).  This guarantees that
877                          * all changes affecting the inode are permanent
878                          * when we return.
879                          */
880                         if (iip && iip->ili_last_lsn) {
881                                 xfs_log_force(mp, iip->ili_last_lsn,
882                                                 XFS_LOG_FORCE | XFS_LOG_SYNC);
883                         } else if (xfs_ipincount(xip) > 0) {
884                                 xfs_log_force(mp, (xfs_lsn_t)0,
885                                                 XFS_LOG_FORCE | XFS_LOG_SYNC);
886                         }
887
888                 } else {
889                         xfs_trans_t     *tp;
890
891                         /*
892                          * O_SYNC or O_DSYNC _with_ a size update are handled
893                          * the same way.
894                          *
895                          * If the write was synchronous then we need to make
896                          * sure that the inode modification time is permanent.
897                          * We'll have updated the timestamp above, so here
898                          * we use a synchronous transaction to log the inode.
899                          * It's not fast, but it's necessary.
900                          *
901                          * If this a dsync write and the size got changed
902                          * non-transactionally, then we need to ensure that
903                          * the size change gets logged in a synchronous
904                          * transaction.
905                          */
906
907                         tp = xfs_trans_alloc(mp, XFS_TRANS_WRITE_SYNC);
908                         if ((error = xfs_trans_reserve(tp, 0,
909                                                       XFS_SWRITE_LOG_RES(mp),
910                                                       0, 0, 0))) {
911                                 /* Transaction reserve failed */
912                                 xfs_trans_cancel(tp, 0);
913                         } else {
914                                 /* Transaction reserve successful */
915                                 xfs_ilock(xip, XFS_ILOCK_EXCL);
916                                 xfs_trans_ijoin(tp, xip, XFS_ILOCK_EXCL);
917                                 xfs_trans_ihold(tp, xip);
918                                 xfs_trans_log_inode(tp, xip, XFS_ILOG_CORE);
919                                 xfs_trans_set_sync(tp);
920                                 error = xfs_trans_commit(tp, 0, NULL);
921                                 xfs_iunlock(xip, XFS_ILOCK_EXCL);
922                         }
923                         if (error)
924                                 goto out_unlock_internal;
925                 }
926         
927                 xfs_rwunlock(bdp, locktype);
928                 if (need_isem)
929                         up(&inode->i_sem);
930
931                 error = sync_page_range(inode, mapping, pos, ret);
932                 if (!error)
933                         error = ret;
934                 return error;
935         }
936
937  out_unlock_internal:
938         xfs_rwunlock(bdp, locktype);
939  out_unlock_isem:
940         if (need_isem)
941                 up(&inode->i_sem);
942  out_nounlocks:
943         return -error;
944 }
945
946 /*
947  * All xfs metadata buffers except log state machine buffers
948  * get this attached as their b_bdstrat callback function.
949  * This is so that we can catch a buffer
950  * after prematurely unpinning it to forcibly shutdown the filesystem.
951  */
952 int
953 xfs_bdstrat_cb(struct xfs_buf *bp)
954 {
955         xfs_mount_t     *mp;
956
957         mp = XFS_BUF_FSPRIVATE3(bp, xfs_mount_t *);
958         if (!XFS_FORCED_SHUTDOWN(mp)) {
959                 pagebuf_iorequest(bp);
960                 return 0;
961         } else {
962                 xfs_buftrace("XFS__BDSTRAT IOERROR", bp);
963                 /*
964                  * Metadata write that didn't get logged but
965                  * written delayed anyway. These aren't associated
966                  * with a transaction, and can be ignored.
967                  */
968                 if (XFS_BUF_IODONE_FUNC(bp) == NULL &&
969                     (XFS_BUF_ISREAD(bp)) == 0)
970                         return (xfs_bioerror_relse(bp));
971                 else
972                         return (xfs_bioerror(bp));
973         }
974 }
975
976
977 int
978 xfs_bmap(bhv_desc_t     *bdp,
979         xfs_off_t       offset,
980         ssize_t         count,
981         int             flags,
982         xfs_iomap_t     *iomapp,
983         int             *niomaps)
984 {
985         xfs_inode_t     *ip = XFS_BHVTOI(bdp);
986         xfs_iocore_t    *io = &ip->i_iocore;
987
988         ASSERT((ip->i_d.di_mode & S_IFMT) == S_IFREG);
989         ASSERT(((ip->i_d.di_flags & XFS_DIFLAG_REALTIME) != 0) ==
990                ((ip->i_iocore.io_flags & XFS_IOCORE_RT) != 0));
991
992         return xfs_iomap(io, offset, count, flags, iomapp, niomaps);
993 }
994
995 /*
996  * Wrapper around bdstrat so that we can stop data
997  * from going to disk in case we are shutting down the filesystem.
998  * Typically user data goes thru this path; one of the exceptions
999  * is the superblock.
1000  */
1001 int
1002 xfsbdstrat(
1003         struct xfs_mount        *mp,
1004         struct xfs_buf          *bp)
1005 {
1006         ASSERT(mp);
1007         if (!XFS_FORCED_SHUTDOWN(mp)) {
1008                 /* Grio redirection would go here
1009                  * if (XFS_BUF_IS_GRIO(bp)) {
1010                  */
1011
1012                 pagebuf_iorequest(bp);
1013                 return 0;
1014         }
1015
1016         xfs_buftrace("XFSBDSTRAT IOERROR", bp);
1017         return (xfs_bioerror_relse(bp));
1018 }
1019
1020 /*
1021  * If the underlying (data/log/rt) device is readonly, there are some
1022  * operations that cannot proceed.
1023  */
1024 int
1025 xfs_dev_is_read_only(
1026         xfs_mount_t             *mp,
1027         char                    *message)
1028 {
1029         if (xfs_readonly_buftarg(mp->m_ddev_targp) ||
1030             xfs_readonly_buftarg(mp->m_logdev_targp) ||
1031             (mp->m_rtdev_targp && xfs_readonly_buftarg(mp->m_rtdev_targp))) {
1032                 cmn_err(CE_NOTE,
1033                         "XFS: %s required on read-only device.", message);
1034                 cmn_err(CE_NOTE,
1035                         "XFS: write access unavailable, cannot proceed.");
1036                 return EROFS;
1037         }
1038         return 0;
1039 }