Linux-2.6.12-rc2
[linux-2.6.git] / fs / xfs / linux-2.6 / xfs_lrw.c
1 /*
2  * Copyright (c) 2000-2003 Silicon Graphics, Inc.  All Rights Reserved.
3  *
4  * This program is free software; you can redistribute it and/or modify it
5  * under the terms of version 2 of the GNU General Public License as
6  * published by the Free Software Foundation.
7  *
8  * This program is distributed in the hope that it would be useful, but
9  * WITHOUT ANY WARRANTY; without even the implied warranty of
10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
11  *
12  * Further, this software is distributed without any warranty that it is
13  * free of the rightful claim of any third person regarding infringement
14  * or the like.  Any license provided herein, whether implied or
15  * otherwise, applies only to this software file.  Patent licenses, if
16  * any, provided herein do not apply to combinations of this program with
17  * other software, or any other product whatsoever.
18  *
19  * You should have received a copy of the GNU General Public License along
20  * with this program; if not, write the Free Software Foundation, Inc., 59
21  * Temple Place - Suite 330, Boston MA 02111-1307, USA.
22  *
23  * Contact information: Silicon Graphics, Inc., 1600 Amphitheatre Pkwy,
24  * Mountain View, CA  94043, or:
25  *
26  * http://www.sgi.com
27  *
28  * For further information regarding this notice, see:
29  *
30  * http://oss.sgi.com/projects/GenInfo/SGIGPLNoticeExplan/
31  */
32 /*
33  *  fs/xfs/linux/xfs_lrw.c (Linux Read Write stuff)
34  *
35  */
36
37 #include "xfs.h"
38
39 #include "xfs_fs.h"
40 #include "xfs_inum.h"
41 #include "xfs_log.h"
42 #include "xfs_trans.h"
43 #include "xfs_sb.h"
44 #include "xfs_ag.h"
45 #include "xfs_dir.h"
46 #include "xfs_dir2.h"
47 #include "xfs_alloc.h"
48 #include "xfs_dmapi.h"
49 #include "xfs_quota.h"
50 #include "xfs_mount.h"
51 #include "xfs_alloc_btree.h"
52 #include "xfs_bmap_btree.h"
53 #include "xfs_ialloc_btree.h"
54 #include "xfs_btree.h"
55 #include "xfs_ialloc.h"
56 #include "xfs_attr_sf.h"
57 #include "xfs_dir_sf.h"
58 #include "xfs_dir2_sf.h"
59 #include "xfs_dinode.h"
60 #include "xfs_inode.h"
61 #include "xfs_bmap.h"
62 #include "xfs_bit.h"
63 #include "xfs_rtalloc.h"
64 #include "xfs_error.h"
65 #include "xfs_itable.h"
66 #include "xfs_rw.h"
67 #include "xfs_acl.h"
68 #include "xfs_cap.h"
69 #include "xfs_mac.h"
70 #include "xfs_attr.h"
71 #include "xfs_inode_item.h"
72 #include "xfs_buf_item.h"
73 #include "xfs_utils.h"
74 #include "xfs_iomap.h"
75
76 #include <linux/capability.h>
77 #include <linux/writeback.h>
78
79
80 #if defined(XFS_RW_TRACE)
81 void
82 xfs_rw_enter_trace(
83         int                     tag,
84         xfs_iocore_t            *io,
85         void                    *data,
86         size_t                  segs,
87         loff_t                  offset,
88         int                     ioflags)
89 {
90         xfs_inode_t     *ip = XFS_IO_INODE(io);
91
92         if (ip->i_rwtrace == NULL)
93                 return;
94         ktrace_enter(ip->i_rwtrace,
95                 (void *)(unsigned long)tag,
96                 (void *)ip,
97                 (void *)((unsigned long)((ip->i_d.di_size >> 32) & 0xffffffff)),
98                 (void *)((unsigned long)(ip->i_d.di_size & 0xffffffff)),
99                 (void *)data,
100                 (void *)((unsigned long)segs),
101                 (void *)((unsigned long)((offset >> 32) & 0xffffffff)),
102                 (void *)((unsigned long)(offset & 0xffffffff)),
103                 (void *)((unsigned long)ioflags),
104                 (void *)((unsigned long)((io->io_new_size >> 32) & 0xffffffff)),
105                 (void *)((unsigned long)(io->io_new_size & 0xffffffff)),
106                 (void *)NULL,
107                 (void *)NULL,
108                 (void *)NULL,
109                 (void *)NULL,
110                 (void *)NULL);
111 }
112
113 void
114 xfs_inval_cached_trace(
115         xfs_iocore_t    *io,
116         xfs_off_t       offset,
117         xfs_off_t       len,
118         xfs_off_t       first,
119         xfs_off_t       last)
120 {
121         xfs_inode_t     *ip = XFS_IO_INODE(io);
122
123         if (ip->i_rwtrace == NULL)
124                 return;
125         ktrace_enter(ip->i_rwtrace,
126                 (void *)(__psint_t)XFS_INVAL_CACHED,
127                 (void *)ip,
128                 (void *)((unsigned long)((offset >> 32) & 0xffffffff)),
129                 (void *)((unsigned long)(offset & 0xffffffff)),
130                 (void *)((unsigned long)((len >> 32) & 0xffffffff)),
131                 (void *)((unsigned long)(len & 0xffffffff)),
132                 (void *)((unsigned long)((first >> 32) & 0xffffffff)),
133                 (void *)((unsigned long)(first & 0xffffffff)),
134                 (void *)((unsigned long)((last >> 32) & 0xffffffff)),
135                 (void *)((unsigned long)(last & 0xffffffff)),
136                 (void *)NULL,
137                 (void *)NULL,
138                 (void *)NULL,
139                 (void *)NULL,
140                 (void *)NULL,
141                 (void *)NULL);
142 }
143 #endif
144
145 /*
146  *      xfs_iozero
147  *
148  *      xfs_iozero clears the specified range of buffer supplied,
149  *      and marks all the affected blocks as valid and modified.  If
150  *      an affected block is not allocated, it will be allocated.  If
151  *      an affected block is not completely overwritten, and is not
152  *      valid before the operation, it will be read from disk before
153  *      being partially zeroed.
154  */
155 STATIC int
156 xfs_iozero(
157         struct inode            *ip,    /* inode                        */
158         loff_t                  pos,    /* offset in file               */
159         size_t                  count,  /* size of data to zero         */
160         loff_t                  end_size)       /* max file size to set */
161 {
162         unsigned                bytes;
163         struct page             *page;
164         struct address_space    *mapping;
165         char                    *kaddr;
166         int                     status;
167
168         mapping = ip->i_mapping;
169         do {
170                 unsigned long index, offset;
171
172                 offset = (pos & (PAGE_CACHE_SIZE -1)); /* Within page */
173                 index = pos >> PAGE_CACHE_SHIFT;
174                 bytes = PAGE_CACHE_SIZE - offset;
175                 if (bytes > count)
176                         bytes = count;
177
178                 status = -ENOMEM;
179                 page = grab_cache_page(mapping, index);
180                 if (!page)
181                         break;
182
183                 kaddr = kmap(page);
184                 status = mapping->a_ops->prepare_write(NULL, page, offset,
185                                                         offset + bytes);
186                 if (status) {
187                         goto unlock;
188                 }
189
190                 memset((void *) (kaddr + offset), 0, bytes);
191                 flush_dcache_page(page);
192                 status = mapping->a_ops->commit_write(NULL, page, offset,
193                                                         offset + bytes);
194                 if (!status) {
195                         pos += bytes;
196                         count -= bytes;
197                         if (pos > i_size_read(ip))
198                                 i_size_write(ip, pos < end_size ? pos : end_size);
199                 }
200
201 unlock:
202                 kunmap(page);
203                 unlock_page(page);
204                 page_cache_release(page);
205                 if (status)
206                         break;
207         } while (count);
208
209         return (-status);
210 }
211
212 /*
213  * xfs_inval_cached_pages
214  * 
215  * This routine is responsible for keeping direct I/O and buffered I/O
216  * somewhat coherent.  From here we make sure that we're at least
217  * temporarily holding the inode I/O lock exclusively and then call
218  * the page cache to flush and invalidate any cached pages.  If there
219  * are no cached pages this routine will be very quick.
220  */
221 void
222 xfs_inval_cached_pages(
223         vnode_t         *vp,
224         xfs_iocore_t    *io,
225         xfs_off_t       offset,
226         int             write,
227         int             relock)
228 {
229         if (VN_CACHED(vp)) {
230                 xfs_inval_cached_trace(io, offset, -1, ctooff(offtoct(offset)), -1);
231                 VOP_FLUSHINVAL_PAGES(vp, ctooff(offtoct(offset)), -1, FI_REMAPF_LOCKED);
232         }
233
234 }
235
236 ssize_t                 /* bytes read, or (-)  error */
237 xfs_read(
238         bhv_desc_t              *bdp,
239         struct kiocb            *iocb,
240         const struct iovec      *iovp,
241         unsigned int            segs,
242         loff_t                  *offset,
243         int                     ioflags,
244         cred_t                  *credp)
245 {
246         struct file             *file = iocb->ki_filp;
247         struct inode            *inode = file->f_mapping->host;
248         size_t                  size = 0;
249         ssize_t                 ret;
250         xfs_fsize_t             n;
251         xfs_inode_t             *ip;
252         xfs_mount_t             *mp;
253         vnode_t                 *vp;
254         unsigned long           seg;
255
256         ip = XFS_BHVTOI(bdp);
257         vp = BHV_TO_VNODE(bdp);
258         mp = ip->i_mount;
259
260         XFS_STATS_INC(xs_read_calls);
261
262         /* START copy & waste from filemap.c */
263         for (seg = 0; seg < segs; seg++) {
264                 const struct iovec *iv = &iovp[seg];
265
266                 /*
267                  * If any segment has a negative length, or the cumulative
268                  * length ever wraps negative then return -EINVAL.
269                  */
270                 size += iv->iov_len;
271                 if (unlikely((ssize_t)(size|iv->iov_len) < 0))
272                         return XFS_ERROR(-EINVAL);
273         }
274         /* END copy & waste from filemap.c */
275
276         if (unlikely(ioflags & IO_ISDIRECT)) {
277                 xfs_buftarg_t   *target =
278                         (ip->i_d.di_flags & XFS_DIFLAG_REALTIME) ?
279                                 mp->m_rtdev_targp : mp->m_ddev_targp;
280                 if ((*offset & target->pbr_smask) ||
281                     (size & target->pbr_smask)) {
282                         if (*offset == ip->i_d.di_size) {
283                                 return (0);
284                         }
285                         return -XFS_ERROR(EINVAL);
286                 }
287         }
288
289         n = XFS_MAXIOFFSET(mp) - *offset;
290         if ((n <= 0) || (size == 0))
291                 return 0;
292
293         if (n < size)
294                 size = n;
295
296         if (XFS_FORCED_SHUTDOWN(mp)) {
297                 return -EIO;
298         }
299
300         if (unlikely(ioflags & IO_ISDIRECT))
301                 down(&inode->i_sem);
302         xfs_ilock(ip, XFS_IOLOCK_SHARED);
303
304         if (DM_EVENT_ENABLED(vp->v_vfsp, ip, DM_EVENT_READ) &&
305             !(ioflags & IO_INVIS)) {
306                 vrwlock_t locktype = VRWLOCK_READ;
307
308                 ret = -XFS_SEND_DATA(mp, DM_EVENT_READ,
309                                         BHV_TO_VNODE(bdp), *offset, size,
310                                         FILP_DELAY_FLAG(file), &locktype);
311                 if (ret) {
312                         xfs_iunlock(ip, XFS_IOLOCK_SHARED);
313                         goto unlock_isem;
314                 }
315         }
316
317         xfs_rw_enter_trace(XFS_READ_ENTER, &ip->i_iocore,
318                                 (void *)iovp, segs, *offset, ioflags);
319         ret = __generic_file_aio_read(iocb, iovp, segs, offset);
320         if (ret == -EIOCBQUEUED && !(ioflags & IO_ISAIO))
321                 ret = wait_on_sync_kiocb(iocb);
322         if (ret > 0)
323                 XFS_STATS_ADD(xs_read_bytes, ret);
324
325         xfs_iunlock(ip, XFS_IOLOCK_SHARED);
326
327         if (likely(!(ioflags & IO_INVIS)))
328                 xfs_ichgtime(ip, XFS_ICHGTIME_ACC);
329
330 unlock_isem:
331         if (unlikely(ioflags & IO_ISDIRECT))
332                 up(&inode->i_sem);
333         return ret;
334 }
335
336 ssize_t
337 xfs_sendfile(
338         bhv_desc_t              *bdp,
339         struct file             *filp,
340         loff_t                  *offset,
341         int                     ioflags,
342         size_t                  count,
343         read_actor_t            actor,
344         void                    *target,
345         cred_t                  *credp)
346 {
347         ssize_t                 ret;
348         xfs_fsize_t             n;
349         xfs_inode_t             *ip;
350         xfs_mount_t             *mp;
351         vnode_t                 *vp;
352
353         ip = XFS_BHVTOI(bdp);
354         vp = BHV_TO_VNODE(bdp);
355         mp = ip->i_mount;
356
357         XFS_STATS_INC(xs_read_calls);
358
359         n = XFS_MAXIOFFSET(mp) - *offset;
360         if ((n <= 0) || (count == 0))
361                 return 0;
362
363         if (n < count)
364                 count = n;
365
366         if (XFS_FORCED_SHUTDOWN(ip->i_mount))
367                 return -EIO;
368
369         xfs_ilock(ip, XFS_IOLOCK_SHARED);
370
371         if (DM_EVENT_ENABLED(vp->v_vfsp, ip, DM_EVENT_READ) &&
372             (!(ioflags & IO_INVIS))) {
373                 vrwlock_t locktype = VRWLOCK_READ;
374                 int error;
375
376                 error = XFS_SEND_DATA(mp, DM_EVENT_READ, BHV_TO_VNODE(bdp), *offset, count,
377                                       FILP_DELAY_FLAG(filp), &locktype);
378                 if (error) {
379                         xfs_iunlock(ip, XFS_IOLOCK_SHARED);
380                         return -error;
381                 }
382         }
383         xfs_rw_enter_trace(XFS_SENDFILE_ENTER, &ip->i_iocore,
384                    (void *)(unsigned long)target, count, *offset, ioflags);
385         ret = generic_file_sendfile(filp, offset, count, actor, target);
386
387         xfs_iunlock(ip, XFS_IOLOCK_SHARED);
388
389         if (ret > 0)
390                 XFS_STATS_ADD(xs_read_bytes, ret);
391
392         if (likely(!(ioflags & IO_INVIS)))
393                 xfs_ichgtime(ip, XFS_ICHGTIME_ACC);
394
395         return ret;
396 }
397
398 /*
399  * This routine is called to handle zeroing any space in the last
400  * block of the file that is beyond the EOF.  We do this since the
401  * size is being increased without writing anything to that block
402  * and we don't want anyone to read the garbage on the disk.
403  */
404 STATIC int                              /* error (positive) */
405 xfs_zero_last_block(
406         struct inode    *ip,
407         xfs_iocore_t    *io,
408         xfs_off_t       offset,
409         xfs_fsize_t     isize,
410         xfs_fsize_t     end_size)
411 {
412         xfs_fileoff_t   last_fsb;
413         xfs_mount_t     *mp;
414         int             nimaps;
415         int             zero_offset;
416         int             zero_len;
417         int             isize_fsb_offset;
418         int             error = 0;
419         xfs_bmbt_irec_t imap;
420         loff_t          loff;
421         size_t          lsize;
422
423         ASSERT(ismrlocked(io->io_lock, MR_UPDATE) != 0);
424         ASSERT(offset > isize);
425
426         mp = io->io_mount;
427
428         isize_fsb_offset = XFS_B_FSB_OFFSET(mp, isize);
429         if (isize_fsb_offset == 0) {
430                 /*
431                  * There are no extra bytes in the last block on disk to
432                  * zero, so return.
433                  */
434                 return 0;
435         }
436
437         last_fsb = XFS_B_TO_FSBT(mp, isize);
438         nimaps = 1;
439         error = XFS_BMAPI(mp, NULL, io, last_fsb, 1, 0, NULL, 0, &imap,
440                           &nimaps, NULL);
441         if (error) {
442                 return error;
443         }
444         ASSERT(nimaps > 0);
445         /*
446          * If the block underlying isize is just a hole, then there
447          * is nothing to zero.
448          */
449         if (imap.br_startblock == HOLESTARTBLOCK) {
450                 return 0;
451         }
452         /*
453          * Zero the part of the last block beyond the EOF, and write it
454          * out sync.  We need to drop the ilock while we do this so we
455          * don't deadlock when the buffer cache calls back to us.
456          */
457         XFS_IUNLOCK(mp, io, XFS_ILOCK_EXCL| XFS_EXTSIZE_RD);
458         loff = XFS_FSB_TO_B(mp, last_fsb);
459         lsize = XFS_FSB_TO_B(mp, 1);
460
461         zero_offset = isize_fsb_offset;
462         zero_len = mp->m_sb.sb_blocksize - isize_fsb_offset;
463
464         error = xfs_iozero(ip, loff + zero_offset, zero_len, end_size);
465
466         XFS_ILOCK(mp, io, XFS_ILOCK_EXCL|XFS_EXTSIZE_RD);
467         ASSERT(error >= 0);
468         return error;
469 }
470
471 /*
472  * Zero any on disk space between the current EOF and the new,
473  * larger EOF.  This handles the normal case of zeroing the remainder
474  * of the last block in the file and the unusual case of zeroing blocks
475  * out beyond the size of the file.  This second case only happens
476  * with fixed size extents and when the system crashes before the inode
477  * size was updated but after blocks were allocated.  If fill is set,
478  * then any holes in the range are filled and zeroed.  If not, the holes
479  * are left alone as holes.
480  */
481
482 int                                     /* error (positive) */
483 xfs_zero_eof(
484         vnode_t         *vp,
485         xfs_iocore_t    *io,
486         xfs_off_t       offset,         /* starting I/O offset */
487         xfs_fsize_t     isize,          /* current inode size */
488         xfs_fsize_t     end_size)       /* terminal inode size */
489 {
490         struct inode    *ip = LINVFS_GET_IP(vp);
491         xfs_fileoff_t   start_zero_fsb;
492         xfs_fileoff_t   end_zero_fsb;
493         xfs_fileoff_t   prev_zero_fsb;
494         xfs_fileoff_t   zero_count_fsb;
495         xfs_fileoff_t   last_fsb;
496         xfs_extlen_t    buf_len_fsb;
497         xfs_extlen_t    prev_zero_count;
498         xfs_mount_t     *mp;
499         int             nimaps;
500         int             error = 0;
501         xfs_bmbt_irec_t imap;
502         loff_t          loff;
503         size_t          lsize;
504
505         ASSERT(ismrlocked(io->io_lock, MR_UPDATE));
506         ASSERT(ismrlocked(io->io_iolock, MR_UPDATE));
507
508         mp = io->io_mount;
509
510         /*
511          * First handle zeroing the block on which isize resides.
512          * We only zero a part of that block so it is handled specially.
513          */
514         error = xfs_zero_last_block(ip, io, offset, isize, end_size);
515         if (error) {
516                 ASSERT(ismrlocked(io->io_lock, MR_UPDATE));
517                 ASSERT(ismrlocked(io->io_iolock, MR_UPDATE));
518                 return error;
519         }
520
521         /*
522          * Calculate the range between the new size and the old
523          * where blocks needing to be zeroed may exist.  To get the
524          * block where the last byte in the file currently resides,
525          * we need to subtract one from the size and truncate back
526          * to a block boundary.  We subtract 1 in case the size is
527          * exactly on a block boundary.
528          */
529         last_fsb = isize ? XFS_B_TO_FSBT(mp, isize - 1) : (xfs_fileoff_t)-1;
530         start_zero_fsb = XFS_B_TO_FSB(mp, (xfs_ufsize_t)isize);
531         end_zero_fsb = XFS_B_TO_FSBT(mp, offset - 1);
532         ASSERT((xfs_sfiloff_t)last_fsb < (xfs_sfiloff_t)start_zero_fsb);
533         if (last_fsb == end_zero_fsb) {
534                 /*
535                  * The size was only incremented on its last block.
536                  * We took care of that above, so just return.
537                  */
538                 return 0;
539         }
540
541         ASSERT(start_zero_fsb <= end_zero_fsb);
542         prev_zero_fsb = NULLFILEOFF;
543         prev_zero_count = 0;
544         while (start_zero_fsb <= end_zero_fsb) {
545                 nimaps = 1;
546                 zero_count_fsb = end_zero_fsb - start_zero_fsb + 1;
547                 error = XFS_BMAPI(mp, NULL, io, start_zero_fsb, zero_count_fsb,
548                                   0, NULL, 0, &imap, &nimaps, NULL);
549                 if (error) {
550                         ASSERT(ismrlocked(io->io_lock, MR_UPDATE));
551                         ASSERT(ismrlocked(io->io_iolock, MR_UPDATE));
552                         return error;
553                 }
554                 ASSERT(nimaps > 0);
555
556                 if (imap.br_state == XFS_EXT_UNWRITTEN ||
557                     imap.br_startblock == HOLESTARTBLOCK) {
558                         /*
559                          * This loop handles initializing pages that were
560                          * partially initialized by the code below this
561                          * loop. It basically zeroes the part of the page
562                          * that sits on a hole and sets the page as P_HOLE
563                          * and calls remapf if it is a mapped file.
564                          */
565                         prev_zero_fsb = NULLFILEOFF;
566                         prev_zero_count = 0;
567                         start_zero_fsb = imap.br_startoff +
568                                          imap.br_blockcount;
569                         ASSERT(start_zero_fsb <= (end_zero_fsb + 1));
570                         continue;
571                 }
572
573                 /*
574                  * There are blocks in the range requested.
575                  * Zero them a single write at a time.  We actually
576                  * don't zero the entire range returned if it is
577                  * too big and simply loop around to get the rest.
578                  * That is not the most efficient thing to do, but it
579                  * is simple and this path should not be exercised often.
580                  */
581                 buf_len_fsb = XFS_FILBLKS_MIN(imap.br_blockcount,
582                                               mp->m_writeio_blocks << 8);
583                 /*
584                  * Drop the inode lock while we're doing the I/O.
585                  * We'll still have the iolock to protect us.
586                  */
587                 XFS_IUNLOCK(mp, io, XFS_ILOCK_EXCL|XFS_EXTSIZE_RD);
588
589                 loff = XFS_FSB_TO_B(mp, start_zero_fsb);
590                 lsize = XFS_FSB_TO_B(mp, buf_len_fsb);
591
592                 error = xfs_iozero(ip, loff, lsize, end_size);
593
594                 if (error) {
595                         goto out_lock;
596                 }
597
598                 prev_zero_fsb = start_zero_fsb;
599                 prev_zero_count = buf_len_fsb;
600                 start_zero_fsb = imap.br_startoff + buf_len_fsb;
601                 ASSERT(start_zero_fsb <= (end_zero_fsb + 1));
602
603                 XFS_ILOCK(mp, io, XFS_ILOCK_EXCL|XFS_EXTSIZE_RD);
604         }
605
606         return 0;
607
608 out_lock:
609
610         XFS_ILOCK(mp, io, XFS_ILOCK_EXCL|XFS_EXTSIZE_RD);
611         ASSERT(error >= 0);
612         return error;
613 }
614
615 ssize_t                         /* bytes written, or (-) error */
616 xfs_write(
617         bhv_desc_t              *bdp,
618         struct kiocb            *iocb,
619         const struct iovec      *iovp,
620         unsigned int            nsegs,
621         loff_t                  *offset,
622         int                     ioflags,
623         cred_t                  *credp)
624 {
625         struct file             *file = iocb->ki_filp;
626         struct address_space    *mapping = file->f_mapping;
627         struct inode            *inode = mapping->host;
628         unsigned long           segs = nsegs;
629         xfs_inode_t             *xip;
630         xfs_mount_t             *mp;
631         ssize_t                 ret = 0, error = 0;
632         xfs_fsize_t             isize, new_size;
633         xfs_iocore_t            *io;
634         vnode_t                 *vp;
635         unsigned long           seg;
636         int                     iolock;
637         int                     eventsent = 0;
638         vrwlock_t               locktype;
639         size_t                  ocount = 0, count;
640         loff_t                  pos;
641         int                     need_isem = 1, need_flush = 0;
642
643         XFS_STATS_INC(xs_write_calls);
644
645         vp = BHV_TO_VNODE(bdp);
646         xip = XFS_BHVTOI(bdp);
647
648         for (seg = 0; seg < segs; seg++) {
649                 const struct iovec *iv = &iovp[seg];
650
651                 /*
652                  * If any segment has a negative length, or the cumulative
653                  * length ever wraps negative then return -EINVAL.
654                  */
655                 ocount += iv->iov_len;
656                 if (unlikely((ssize_t)(ocount|iv->iov_len) < 0))
657                         return -EINVAL;
658                 if (access_ok(VERIFY_READ, iv->iov_base, iv->iov_len))
659                         continue;
660                 if (seg == 0)
661                         return -EFAULT;
662                 segs = seg;
663                 ocount -= iv->iov_len;  /* This segment is no good */
664                 break;
665         }
666
667         count = ocount;
668         pos = *offset;
669
670         if (count == 0)
671                 return 0;
672
673         io = &xip->i_iocore;
674         mp = io->io_mount;
675
676         if (XFS_FORCED_SHUTDOWN(mp))
677                 return -EIO;
678
679         fs_check_frozen(vp->v_vfsp, SB_FREEZE_WRITE);
680
681         if (ioflags & IO_ISDIRECT) {
682                 xfs_buftarg_t   *target =
683                         (xip->i_d.di_flags & XFS_DIFLAG_REALTIME) ?
684                                 mp->m_rtdev_targp : mp->m_ddev_targp;
685
686                 if ((pos & target->pbr_smask) || (count & target->pbr_smask))
687                         return XFS_ERROR(-EINVAL);
688
689                 if (!VN_CACHED(vp) && pos < i_size_read(inode))
690                         need_isem = 0;
691
692                 if (VN_CACHED(vp))
693                         need_flush = 1;
694         }
695
696 relock:
697         if (need_isem) {
698                 iolock = XFS_IOLOCK_EXCL;
699                 locktype = VRWLOCK_WRITE;
700
701                 down(&inode->i_sem);
702         } else {
703                 iolock = XFS_IOLOCK_SHARED;
704                 locktype = VRWLOCK_WRITE_DIRECT;
705         }
706
707         xfs_ilock(xip, XFS_ILOCK_EXCL|iolock);
708
709         isize = i_size_read(inode);
710
711         if (file->f_flags & O_APPEND)
712                 *offset = isize;
713
714 start:
715         error = -generic_write_checks(file, &pos, &count,
716                                         S_ISBLK(inode->i_mode));
717         if (error) {
718                 xfs_iunlock(xip, XFS_ILOCK_EXCL|iolock);
719                 goto out_unlock_isem;
720         }
721
722         new_size = pos + count;
723         if (new_size > isize)
724                 io->io_new_size = new_size;
725
726         if ((DM_EVENT_ENABLED(vp->v_vfsp, xip, DM_EVENT_WRITE) &&
727             !(ioflags & IO_INVIS) && !eventsent)) {
728                 loff_t          savedsize = pos;
729                 int             dmflags = FILP_DELAY_FLAG(file);
730
731                 if (need_isem)
732                         dmflags |= DM_FLAGS_ISEM;
733
734                 xfs_iunlock(xip, XFS_ILOCK_EXCL);
735                 error = XFS_SEND_DATA(xip->i_mount, DM_EVENT_WRITE, vp,
736                                       pos, count,
737                                       dmflags, &locktype);
738                 if (error) {
739                         xfs_iunlock(xip, iolock);
740                         goto out_unlock_isem;
741                 }
742                 xfs_ilock(xip, XFS_ILOCK_EXCL);
743                 eventsent = 1;
744
745                 /*
746                  * The iolock was dropped and reaquired in XFS_SEND_DATA
747                  * so we have to recheck the size when appending.
748                  * We will only "goto start;" once, since having sent the
749                  * event prevents another call to XFS_SEND_DATA, which is
750                  * what allows the size to change in the first place.
751                  */
752                 if ((file->f_flags & O_APPEND) && savedsize != isize) {
753                         pos = isize = xip->i_d.di_size;
754                         goto start;
755                 }
756         }
757
758         /*
759          * On Linux, generic_file_write updates the times even if
760          * no data is copied in so long as the write had a size.
761          *
762          * We must update xfs' times since revalidate will overcopy xfs.
763          */
764         if (!(ioflags & IO_INVIS)) {
765                 xfs_ichgtime(xip, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
766                 inode_update_time(inode, 1);
767         }
768
769         /*
770          * If the offset is beyond the size of the file, we have a couple
771          * of things to do. First, if there is already space allocated
772          * we need to either create holes or zero the disk or ...
773          *
774          * If there is a page where the previous size lands, we need
775          * to zero it out up to the new size.
776          */
777
778         if (pos > isize) {
779                 error = xfs_zero_eof(BHV_TO_VNODE(bdp), io, pos,
780                                         isize, pos + count);
781                 if (error) {
782                         xfs_iunlock(xip, XFS_ILOCK_EXCL|iolock);
783                         goto out_unlock_isem;
784                 }
785         }
786         xfs_iunlock(xip, XFS_ILOCK_EXCL);
787
788         /*
789          * If we're writing the file then make sure to clear the
790          * setuid and setgid bits if the process is not being run
791          * by root.  This keeps people from modifying setuid and
792          * setgid binaries.
793          */
794
795         if (((xip->i_d.di_mode & S_ISUID) ||
796             ((xip->i_d.di_mode & (S_ISGID | S_IXGRP)) ==
797                 (S_ISGID | S_IXGRP))) &&
798              !capable(CAP_FSETID)) {
799                 error = xfs_write_clear_setuid(xip);
800                 if (likely(!error))
801                         error = -remove_suid(file->f_dentry);
802                 if (unlikely(error)) {
803                         xfs_iunlock(xip, iolock);
804                         goto out_unlock_isem;
805                 }
806         }
807
808 retry:
809         /* We can write back this queue in page reclaim */
810         current->backing_dev_info = mapping->backing_dev_info;
811
812         if ((ioflags & IO_ISDIRECT)) {
813                 if (need_flush) {
814                         xfs_inval_cached_trace(io, pos, -1,
815                                         ctooff(offtoct(pos)), -1);
816                         VOP_FLUSHINVAL_PAGES(vp, ctooff(offtoct(pos)),
817                                         -1, FI_REMAPF_LOCKED);
818                 }
819
820                 if (need_isem) {
821                         /* demote the lock now the cached pages are gone */
822                         XFS_ILOCK_DEMOTE(mp, io, XFS_IOLOCK_EXCL);
823                         up(&inode->i_sem);
824
825                         iolock = XFS_IOLOCK_SHARED;
826                         locktype = VRWLOCK_WRITE_DIRECT;
827                         need_isem = 0;
828                 }
829
830                 xfs_rw_enter_trace(XFS_DIOWR_ENTER, io, (void *)iovp, segs,
831                                 *offset, ioflags);
832                 ret = generic_file_direct_write(iocb, iovp,
833                                 &segs, pos, offset, count, ocount);
834
835                 /*
836                  * direct-io write to a hole: fall through to buffered I/O
837                  * for completing the rest of the request.
838                  */
839                 if (ret >= 0 && ret != count) {
840                         XFS_STATS_ADD(xs_write_bytes, ret);
841
842                         pos += ret;
843                         count -= ret;
844
845                         need_isem = 1;
846                         ioflags &= ~IO_ISDIRECT;
847                         xfs_iunlock(xip, iolock);
848                         goto relock;
849                 }
850         } else {
851                 xfs_rw_enter_trace(XFS_WRITE_ENTER, io, (void *)iovp, segs,
852                                 *offset, ioflags);
853                 ret = generic_file_buffered_write(iocb, iovp, segs,
854                                 pos, offset, count, ret);
855         }
856
857         current->backing_dev_info = NULL;
858
859         if (ret == -EIOCBQUEUED && !(ioflags & IO_ISAIO))
860                 ret = wait_on_sync_kiocb(iocb);
861
862         if ((ret == -ENOSPC) &&
863             DM_EVENT_ENABLED(vp->v_vfsp, xip, DM_EVENT_NOSPACE) &&
864             !(ioflags & IO_INVIS)) {
865
866                 xfs_rwunlock(bdp, locktype);
867                 error = XFS_SEND_NAMESP(xip->i_mount, DM_EVENT_NOSPACE, vp,
868                                 DM_RIGHT_NULL, vp, DM_RIGHT_NULL, NULL, NULL,
869                                 0, 0, 0); /* Delay flag intentionally  unused */
870                 if (error)
871                         goto out_unlock_isem;
872                 xfs_rwlock(bdp, locktype);
873                 pos = xip->i_d.di_size;
874                 ret = 0;
875                 goto retry;
876         }
877
878         if (*offset > xip->i_d.di_size) {
879                 xfs_ilock(xip, XFS_ILOCK_EXCL);
880                 if (*offset > xip->i_d.di_size) {
881                         xip->i_d.di_size = *offset;
882                         i_size_write(inode, *offset);
883                         xip->i_update_core = 1;
884                         xip->i_update_size = 1;
885                 }
886                 xfs_iunlock(xip, XFS_ILOCK_EXCL);
887         }
888
889         error = -ret;
890         if (ret <= 0)
891                 goto out_unlock_internal;
892
893         XFS_STATS_ADD(xs_write_bytes, ret);
894
895         /* Handle various SYNC-type writes */
896         if ((file->f_flags & O_SYNC) || IS_SYNC(inode)) {
897                 /*
898                  * If we're treating this as O_DSYNC and we have not updated the
899                  * size, force the log.
900                  */
901                 if (!(mp->m_flags & XFS_MOUNT_OSYNCISOSYNC) &&
902                     !(xip->i_update_size)) {
903                         xfs_inode_log_item_t    *iip = xip->i_itemp;
904
905                         /*
906                          * If an allocation transaction occurred
907                          * without extending the size, then we have to force
908                          * the log up the proper point to ensure that the
909                          * allocation is permanent.  We can't count on
910                          * the fact that buffered writes lock out direct I/O
911                          * writes - the direct I/O write could have extended
912                          * the size nontransactionally, then finished before
913                          * we started.  xfs_write_file will think that the file
914                          * didn't grow but the update isn't safe unless the
915                          * size change is logged.
916                          *
917                          * Force the log if we've committed a transaction
918                          * against the inode or if someone else has and
919                          * the commit record hasn't gone to disk (e.g.
920                          * the inode is pinned).  This guarantees that
921                          * all changes affecting the inode are permanent
922                          * when we return.
923                          */
924                         if (iip && iip->ili_last_lsn) {
925                                 xfs_log_force(mp, iip->ili_last_lsn,
926                                                 XFS_LOG_FORCE | XFS_LOG_SYNC);
927                         } else if (xfs_ipincount(xip) > 0) {
928                                 xfs_log_force(mp, (xfs_lsn_t)0,
929                                                 XFS_LOG_FORCE | XFS_LOG_SYNC);
930                         }
931
932                 } else {
933                         xfs_trans_t     *tp;
934
935                         /*
936                          * O_SYNC or O_DSYNC _with_ a size update are handled
937                          * the same way.
938                          *
939                          * If the write was synchronous then we need to make
940                          * sure that the inode modification time is permanent.
941                          * We'll have updated the timestamp above, so here
942                          * we use a synchronous transaction to log the inode.
943                          * It's not fast, but it's necessary.
944                          *
945                          * If this a dsync write and the size got changed
946                          * non-transactionally, then we need to ensure that
947                          * the size change gets logged in a synchronous
948                          * transaction.
949                          */
950
951                         tp = xfs_trans_alloc(mp, XFS_TRANS_WRITE_SYNC);
952                         if ((error = xfs_trans_reserve(tp, 0,
953                                                       XFS_SWRITE_LOG_RES(mp),
954                                                       0, 0, 0))) {
955                                 /* Transaction reserve failed */
956                                 xfs_trans_cancel(tp, 0);
957                         } else {
958                                 /* Transaction reserve successful */
959                                 xfs_ilock(xip, XFS_ILOCK_EXCL);
960                                 xfs_trans_ijoin(tp, xip, XFS_ILOCK_EXCL);
961                                 xfs_trans_ihold(tp, xip);
962                                 xfs_trans_log_inode(tp, xip, XFS_ILOG_CORE);
963                                 xfs_trans_set_sync(tp);
964                                 error = xfs_trans_commit(tp, 0, NULL);
965                                 xfs_iunlock(xip, XFS_ILOCK_EXCL);
966                         }
967                         if (error)
968                                 goto out_unlock_internal;
969                 }
970         
971                 xfs_rwunlock(bdp, locktype);
972                 if (need_isem)
973                         up(&inode->i_sem);
974
975                 error = sync_page_range(inode, mapping, pos, ret);
976                 if (!error)
977                         error = ret;
978                 return error;
979         }
980
981  out_unlock_internal:
982         xfs_rwunlock(bdp, locktype);
983  out_unlock_isem:
984         if (need_isem)
985                 up(&inode->i_sem);
986         return -error;
987 }
988
989 /*
990  * All xfs metadata buffers except log state machine buffers
991  * get this attached as their b_bdstrat callback function.
992  * This is so that we can catch a buffer
993  * after prematurely unpinning it to forcibly shutdown the filesystem.
994  */
995 int
996 xfs_bdstrat_cb(struct xfs_buf *bp)
997 {
998         xfs_mount_t     *mp;
999
1000         mp = XFS_BUF_FSPRIVATE3(bp, xfs_mount_t *);
1001         if (!XFS_FORCED_SHUTDOWN(mp)) {
1002                 pagebuf_iorequest(bp);
1003                 return 0;
1004         } else {
1005                 xfs_buftrace("XFS__BDSTRAT IOERROR", bp);
1006                 /*
1007                  * Metadata write that didn't get logged but
1008                  * written delayed anyway. These aren't associated
1009                  * with a transaction, and can be ignored.
1010                  */
1011                 if (XFS_BUF_IODONE_FUNC(bp) == NULL &&
1012                     (XFS_BUF_ISREAD(bp)) == 0)
1013                         return (xfs_bioerror_relse(bp));
1014                 else
1015                         return (xfs_bioerror(bp));
1016         }
1017 }
1018
1019
1020 int
1021 xfs_bmap(bhv_desc_t     *bdp,
1022         xfs_off_t       offset,
1023         ssize_t         count,
1024         int             flags,
1025         xfs_iomap_t     *iomapp,
1026         int             *niomaps)
1027 {
1028         xfs_inode_t     *ip = XFS_BHVTOI(bdp);
1029         xfs_iocore_t    *io = &ip->i_iocore;
1030
1031         ASSERT((ip->i_d.di_mode & S_IFMT) == S_IFREG);
1032         ASSERT(((ip->i_d.di_flags & XFS_DIFLAG_REALTIME) != 0) ==
1033                ((ip->i_iocore.io_flags & XFS_IOCORE_RT) != 0));
1034
1035         return xfs_iomap(io, offset, count, flags, iomapp, niomaps);
1036 }
1037
1038 /*
1039  * Wrapper around bdstrat so that we can stop data
1040  * from going to disk in case we are shutting down the filesystem.
1041  * Typically user data goes thru this path; one of the exceptions
1042  * is the superblock.
1043  */
1044 int
1045 xfsbdstrat(
1046         struct xfs_mount        *mp,
1047         struct xfs_buf          *bp)
1048 {
1049         ASSERT(mp);
1050         if (!XFS_FORCED_SHUTDOWN(mp)) {
1051                 /* Grio redirection would go here
1052                  * if (XFS_BUF_IS_GRIO(bp)) {
1053                  */
1054
1055                 pagebuf_iorequest(bp);
1056                 return 0;
1057         }
1058
1059         xfs_buftrace("XFSBDSTRAT IOERROR", bp);
1060         return (xfs_bioerror_relse(bp));
1061 }
1062
1063 /*
1064  * If the underlying (data/log/rt) device is readonly, there are some
1065  * operations that cannot proceed.
1066  */
1067 int
1068 xfs_dev_is_read_only(
1069         xfs_mount_t             *mp,
1070         char                    *message)
1071 {
1072         if (xfs_readonly_buftarg(mp->m_ddev_targp) ||
1073             xfs_readonly_buftarg(mp->m_logdev_targp) ||
1074             (mp->m_rtdev_targp && xfs_readonly_buftarg(mp->m_rtdev_targp))) {
1075                 cmn_err(CE_NOTE,
1076                         "XFS: %s required on read-only device.", message);
1077                 cmn_err(CE_NOTE,
1078                         "XFS: write access unavailable, cannot proceed.");
1079                 return EROFS;
1080         }
1081         return 0;
1082 }