[XFS] Fix speculative allocation beyond eof
[linux-2.6.git] / fs / xfs / xfs_iomap.c
1 /*
2  * Copyright (c) 2000-2006 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_ialloc.h"
39 #include "xfs_btree.h"
40 #include "xfs_bmap.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_buf_item.h"
48 #include "xfs_trans_space.h"
49 #include "xfs_utils.h"
50 #include "xfs_iomap.h"
51
52 #if defined(XFS_RW_TRACE)
53 void
54 xfs_iomap_enter_trace(
55         int             tag,
56         xfs_inode_t     *ip,
57         xfs_off_t       offset,
58         ssize_t         count)
59 {
60         if (!ip->i_rwtrace)
61                 return;
62
63         ktrace_enter(ip->i_rwtrace,
64                 (void *)((unsigned long)tag),
65                 (void *)ip,
66                 (void *)((unsigned long)((ip->i_d.di_size >> 32) & 0xffffffff)),
67                 (void *)((unsigned long)(ip->i_d.di_size & 0xffffffff)),
68                 (void *)((unsigned long)((offset >> 32) & 0xffffffff)),
69                 (void *)((unsigned long)(offset & 0xffffffff)),
70                 (void *)((unsigned long)count),
71                 (void *)((unsigned long)((ip->i_new_size >> 32) & 0xffffffff)),
72                 (void *)((unsigned long)(ip->i_new_size & 0xffffffff)),
73                 (void *)((unsigned long)current_pid()),
74                 (void *)NULL,
75                 (void *)NULL,
76                 (void *)NULL,
77                 (void *)NULL,
78                 (void *)NULL,
79                 (void *)NULL);
80 }
81
82 void
83 xfs_iomap_map_trace(
84         int             tag,
85         xfs_inode_t     *ip,
86         xfs_off_t       offset,
87         ssize_t         count,
88         xfs_iomap_t     *iomapp,
89         xfs_bmbt_irec_t *imapp,
90         int             flags)
91 {
92         if (!ip->i_rwtrace)
93                 return;
94
95         ktrace_enter(ip->i_rwtrace,
96                 (void *)((unsigned long)tag),
97                 (void *)ip,
98                 (void *)((unsigned long)((ip->i_d.di_size >> 32) & 0xffffffff)),
99                 (void *)((unsigned long)(ip->i_d.di_size & 0xffffffff)),
100                 (void *)((unsigned long)((offset >> 32) & 0xffffffff)),
101                 (void *)((unsigned long)(offset & 0xffffffff)),
102                 (void *)((unsigned long)count),
103                 (void *)((unsigned long)flags),
104                 (void *)((unsigned long)((iomapp->iomap_offset >> 32) & 0xffffffff)),
105                 (void *)((unsigned long)(iomapp->iomap_offset & 0xffffffff)),
106                 (void *)((unsigned long)(iomapp->iomap_delta)),
107                 (void *)((unsigned long)(iomapp->iomap_bsize)),
108                 (void *)((unsigned long)(iomapp->iomap_bn)),
109                 (void *)(__psint_t)(imapp->br_startoff),
110                 (void *)((unsigned long)(imapp->br_blockcount)),
111                 (void *)(__psint_t)(imapp->br_startblock));
112 }
113 #else
114 #define xfs_iomap_enter_trace(tag, io, offset, count)
115 #define xfs_iomap_map_trace(tag, io, offset, count, iomapp, imapp, flags)
116 #endif
117
118 #define XFS_WRITEIO_ALIGN(mp,off)       (((off) >> mp->m_writeio_log) \
119                                                 << mp->m_writeio_log)
120 #define XFS_STRAT_WRITE_IMAPS   2
121 #define XFS_WRITE_IMAPS         XFS_BMAP_MAX_NMAP
122
123 STATIC int
124 xfs_imap_to_bmap(
125         xfs_inode_t     *ip,
126         xfs_off_t       offset,
127         xfs_bmbt_irec_t *imap,
128         xfs_iomap_t     *iomapp,
129         int             imaps,                  /* Number of imap entries */
130         int             iomaps,                 /* Number of iomap entries */
131         int             flags)
132 {
133         xfs_mount_t     *mp = ip->i_mount;
134         int             pbm;
135         xfs_fsblock_t   start_block;
136
137
138         for (pbm = 0; imaps && pbm < iomaps; imaps--, iomapp++, imap++, pbm++) {
139                 iomapp->iomap_offset = XFS_FSB_TO_B(mp, imap->br_startoff);
140                 iomapp->iomap_delta = offset - iomapp->iomap_offset;
141                 iomapp->iomap_bsize = XFS_FSB_TO_B(mp, imap->br_blockcount);
142                 iomapp->iomap_flags = flags;
143
144                 if (XFS_IS_REALTIME_INODE(ip)) {
145                         iomapp->iomap_flags |= IOMAP_REALTIME;
146                         iomapp->iomap_target = mp->m_rtdev_targp;
147                 } else {
148                         iomapp->iomap_target = mp->m_ddev_targp;
149                 }
150                 start_block = imap->br_startblock;
151                 if (start_block == HOLESTARTBLOCK) {
152                         iomapp->iomap_bn = IOMAP_DADDR_NULL;
153                         iomapp->iomap_flags |= IOMAP_HOLE;
154                 } else if (start_block == DELAYSTARTBLOCK) {
155                         iomapp->iomap_bn = IOMAP_DADDR_NULL;
156                         iomapp->iomap_flags |= IOMAP_DELAY;
157                 } else {
158                         iomapp->iomap_bn = XFS_FSB_TO_DB(ip, start_block);
159                         if (ISUNWRITTEN(imap))
160                                 iomapp->iomap_flags |= IOMAP_UNWRITTEN;
161                 }
162
163                 offset += iomapp->iomap_bsize - iomapp->iomap_delta;
164         }
165         return pbm;     /* Return the number filled */
166 }
167
168 int
169 xfs_iomap(
170         xfs_inode_t     *ip,
171         xfs_off_t       offset,
172         ssize_t         count,
173         int             flags,
174         xfs_iomap_t     *iomapp,
175         int             *niomaps)
176 {
177         xfs_mount_t     *mp = ip->i_mount;
178         xfs_fileoff_t   offset_fsb, end_fsb;
179         int             error = 0;
180         int             lockmode = 0;
181         xfs_bmbt_irec_t imap;
182         int             nimaps = 1;
183         int             bmapi_flags = 0;
184         int             iomap_flags = 0;
185
186         ASSERT((ip->i_d.di_mode & S_IFMT) == S_IFREG);
187
188         if (XFS_FORCED_SHUTDOWN(mp))
189                 return XFS_ERROR(EIO);
190
191         switch (flags & (BMAPI_READ | BMAPI_WRITE | BMAPI_ALLOCATE)) {
192         case BMAPI_READ:
193                 xfs_iomap_enter_trace(XFS_IOMAP_READ_ENTER, ip, offset, count);
194                 lockmode = xfs_ilock_map_shared(ip);
195                 bmapi_flags = XFS_BMAPI_ENTIRE;
196                 break;
197         case BMAPI_WRITE:
198                 xfs_iomap_enter_trace(XFS_IOMAP_WRITE_ENTER, ip, offset, count);
199                 lockmode = XFS_ILOCK_EXCL;
200                 if (flags & BMAPI_IGNSTATE)
201                         bmapi_flags |= XFS_BMAPI_IGSTATE|XFS_BMAPI_ENTIRE;
202                 xfs_ilock(ip, lockmode);
203                 break;
204         case BMAPI_ALLOCATE:
205                 xfs_iomap_enter_trace(XFS_IOMAP_ALLOC_ENTER, ip, offset, count);
206                 lockmode = XFS_ILOCK_SHARED;
207                 bmapi_flags = XFS_BMAPI_ENTIRE;
208
209                 /* Attempt non-blocking lock */
210                 if (flags & BMAPI_TRYLOCK) {
211                         if (!xfs_ilock_nowait(ip, lockmode))
212                                 return XFS_ERROR(EAGAIN);
213                 } else {
214                         xfs_ilock(ip, lockmode);
215                 }
216                 break;
217         default:
218                 BUG();
219         }
220
221         ASSERT(offset <= mp->m_maxioffset);
222         if ((xfs_fsize_t)offset + count > mp->m_maxioffset)
223                 count = mp->m_maxioffset - offset;
224         end_fsb = XFS_B_TO_FSB(mp, (xfs_ufsize_t)offset + count);
225         offset_fsb = XFS_B_TO_FSBT(mp, offset);
226
227         error = xfs_bmapi(NULL, ip, offset_fsb,
228                         (xfs_filblks_t)(end_fsb - offset_fsb),
229                         bmapi_flags,  NULL, 0, &imap,
230                         &nimaps, NULL, NULL);
231
232         if (error)
233                 goto out;
234
235         switch (flags & (BMAPI_WRITE|BMAPI_ALLOCATE)) {
236         case BMAPI_WRITE:
237                 /* If we found an extent, return it */
238                 if (nimaps &&
239                     (imap.br_startblock != HOLESTARTBLOCK) &&
240                     (imap.br_startblock != DELAYSTARTBLOCK)) {
241                         xfs_iomap_map_trace(XFS_IOMAP_WRITE_MAP, ip,
242                                         offset, count, iomapp, &imap, flags);
243                         break;
244                 }
245
246                 if (flags & (BMAPI_DIRECT|BMAPI_MMAP)) {
247                         error = xfs_iomap_write_direct(ip, offset, count, flags,
248                                                        &imap, &nimaps, nimaps);
249                 } else {
250                         error = xfs_iomap_write_delay(ip, offset, count, flags,
251                                                       &imap, &nimaps);
252                 }
253                 if (!error) {
254                         xfs_iomap_map_trace(XFS_IOMAP_ALLOC_MAP, ip,
255                                         offset, count, iomapp, &imap, flags);
256                 }
257                 iomap_flags = IOMAP_NEW;
258                 break;
259         case BMAPI_ALLOCATE:
260                 /* If we found an extent, return it */
261                 xfs_iunlock(ip, lockmode);
262                 lockmode = 0;
263
264                 if (nimaps && !ISNULLSTARTBLOCK(imap.br_startblock)) {
265                         xfs_iomap_map_trace(XFS_IOMAP_WRITE_MAP, ip,
266                                         offset, count, iomapp, &imap, flags);
267                         break;
268                 }
269
270                 error = xfs_iomap_write_allocate(ip, offset, count,
271                                                  &imap, &nimaps);
272                 break;
273         }
274
275         if (nimaps) {
276                 *niomaps = xfs_imap_to_bmap(ip, offset, &imap,
277                                             iomapp, nimaps, *niomaps, iomap_flags);
278         } else if (niomaps) {
279                 *niomaps = 0;
280         }
281
282 out:
283         if (lockmode)
284                 xfs_iunlock(ip, lockmode);
285         return XFS_ERROR(error);
286 }
287
288
289 STATIC int
290 xfs_iomap_eof_align_last_fsb(
291         xfs_mount_t     *mp,
292         xfs_inode_t     *ip,
293         xfs_extlen_t    extsize,
294         xfs_fileoff_t   *last_fsb)
295 {
296         xfs_fileoff_t   new_last_fsb = 0;
297         xfs_extlen_t    align;
298         int             eof, error;
299
300         if (XFS_IS_REALTIME_INODE(ip))
301                 ;
302         /*
303          * If mounted with the "-o swalloc" option, roundup the allocation
304          * request to a stripe width boundary if the file size is >=
305          * stripe width and we are allocating past the allocation eof.
306          */
307         else if (mp->m_swidth && (mp->m_flags & XFS_MOUNT_SWALLOC) &&
308                 (ip->i_size >= XFS_FSB_TO_B(mp, mp->m_swidth)))
309                 new_last_fsb = roundup_64(*last_fsb, mp->m_swidth);
310         /*
311          * Roundup the allocation request to a stripe unit (m_dalign) boundary
312          * if the file size is >= stripe unit size, and we are allocating past
313          * the allocation eof.
314          */
315         else if (mp->m_dalign && (ip->i_size >= XFS_FSB_TO_B(mp, mp->m_dalign)))
316                 new_last_fsb = roundup_64(*last_fsb, mp->m_dalign);
317
318         /*
319          * Always round up the allocation request to an extent boundary
320          * (when file on a real-time subvolume or has di_extsize hint).
321          */
322         if (extsize) {
323                 if (new_last_fsb)
324                         align = roundup_64(new_last_fsb, extsize);
325                 else
326                         align = extsize;
327                 new_last_fsb = roundup_64(*last_fsb, align);
328         }
329
330         if (new_last_fsb) {
331                 error = xfs_bmap_eof(ip, new_last_fsb, XFS_DATA_FORK, &eof);
332                 if (error)
333                         return error;
334                 if (eof)
335                         *last_fsb = new_last_fsb;
336         }
337         return 0;
338 }
339
340 STATIC int
341 xfs_flush_space(
342         xfs_inode_t     *ip,
343         int             *fsynced,
344         int             *ioflags)
345 {
346         switch (*fsynced) {
347         case 0:
348                 if (ip->i_delayed_blks) {
349                         xfs_iunlock(ip, XFS_ILOCK_EXCL);
350                         xfs_flush_inode(ip);
351                         xfs_ilock(ip, XFS_ILOCK_EXCL);
352                         *fsynced = 1;
353                 } else {
354                         *ioflags |= BMAPI_SYNC;
355                         *fsynced = 2;
356                 }
357                 return 0;
358         case 1:
359                 *fsynced = 2;
360                 *ioflags |= BMAPI_SYNC;
361                 return 0;
362         case 2:
363                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
364                 xfs_flush_device(ip);
365                 xfs_ilock(ip, XFS_ILOCK_EXCL);
366                 *fsynced = 3;
367                 return 0;
368         }
369         return 1;
370 }
371
372 STATIC int
373 xfs_cmn_err_fsblock_zero(
374         xfs_inode_t     *ip,
375         xfs_bmbt_irec_t *imap)
376 {
377         xfs_cmn_err(XFS_PTAG_FSBLOCK_ZERO, CE_ALERT, ip->i_mount,
378                         "Access to block zero in inode %llu "
379                         "start_block: %llx start_off: %llx "
380                         "blkcnt: %llx extent-state: %x\n",
381                 (unsigned long long)ip->i_ino,
382                 (unsigned long long)imap->br_startblock,
383                 (unsigned long long)imap->br_startoff,
384                 (unsigned long long)imap->br_blockcount,
385                 imap->br_state);
386         return EFSCORRUPTED;
387 }
388
389 int
390 xfs_iomap_write_direct(
391         xfs_inode_t     *ip,
392         xfs_off_t       offset,
393         size_t          count,
394         int             flags,
395         xfs_bmbt_irec_t *ret_imap,
396         int             *nmaps,
397         int             found)
398 {
399         xfs_mount_t     *mp = ip->i_mount;
400         xfs_fileoff_t   offset_fsb;
401         xfs_fileoff_t   last_fsb;
402         xfs_filblks_t   count_fsb, resaligned;
403         xfs_fsblock_t   firstfsb;
404         xfs_extlen_t    extsz, temp;
405         int             nimaps;
406         int             bmapi_flag;
407         int             quota_flag;
408         int             rt;
409         xfs_trans_t     *tp;
410         xfs_bmbt_irec_t imap;
411         xfs_bmap_free_t free_list;
412         uint            qblocks, resblks, resrtextents;
413         int             committed;
414         int             error;
415
416         /*
417          * Make sure that the dquots are there. This doesn't hold
418          * the ilock across a disk read.
419          */
420         error = XFS_QM_DQATTACH(ip->i_mount, ip, XFS_QMOPT_ILOCKED);
421         if (error)
422                 return XFS_ERROR(error);
423
424         rt = XFS_IS_REALTIME_INODE(ip);
425         extsz = xfs_get_extsz_hint(ip);
426
427         offset_fsb = XFS_B_TO_FSBT(mp, offset);
428         last_fsb = XFS_B_TO_FSB(mp, ((xfs_ufsize_t)(offset + count)));
429         if ((offset + count) > ip->i_size) {
430                 error = xfs_iomap_eof_align_last_fsb(mp, ip, extsz, &last_fsb);
431                 if (error)
432                         goto error_out;
433         } else {
434                 if (found && (ret_imap->br_startblock == HOLESTARTBLOCK))
435                         last_fsb = MIN(last_fsb, (xfs_fileoff_t)
436                                         ret_imap->br_blockcount +
437                                         ret_imap->br_startoff);
438         }
439         count_fsb = last_fsb - offset_fsb;
440         ASSERT(count_fsb > 0);
441
442         resaligned = count_fsb;
443         if (unlikely(extsz)) {
444                 if ((temp = do_mod(offset_fsb, extsz)))
445                         resaligned += temp;
446                 if ((temp = do_mod(resaligned, extsz)))
447                         resaligned += extsz - temp;
448         }
449
450         if (unlikely(rt)) {
451                 resrtextents = qblocks = resaligned;
452                 resrtextents /= mp->m_sb.sb_rextsize;
453                 resblks = XFS_DIOSTRAT_SPACE_RES(mp, 0);
454                 quota_flag = XFS_QMOPT_RES_RTBLKS;
455         } else {
456                 resrtextents = 0;
457                 resblks = qblocks = XFS_DIOSTRAT_SPACE_RES(mp, resaligned);
458                 quota_flag = XFS_QMOPT_RES_REGBLKS;
459         }
460
461         /*
462          * Allocate and setup the transaction
463          */
464         xfs_iunlock(ip, XFS_ILOCK_EXCL);
465         tp = xfs_trans_alloc(mp, XFS_TRANS_DIOSTRAT);
466         error = xfs_trans_reserve(tp, resblks,
467                         XFS_WRITE_LOG_RES(mp), resrtextents,
468                         XFS_TRANS_PERM_LOG_RES,
469                         XFS_WRITE_LOG_COUNT);
470         /*
471          * Check for running out of space, note: need lock to return
472          */
473         if (error)
474                 xfs_trans_cancel(tp, 0);
475         xfs_ilock(ip, XFS_ILOCK_EXCL);
476         if (error)
477                 goto error_out;
478
479         error = XFS_TRANS_RESERVE_QUOTA_NBLKS(mp, tp, ip,
480                                               qblocks, 0, quota_flag);
481         if (error)
482                 goto error1;
483
484         xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
485         xfs_trans_ihold(tp, ip);
486
487         bmapi_flag = XFS_BMAPI_WRITE;
488         if ((flags & BMAPI_DIRECT) && (offset < ip->i_size || extsz))
489                 bmapi_flag |= XFS_BMAPI_PREALLOC;
490
491         /*
492          * Issue the xfs_bmapi() call to allocate the blocks
493          */
494         XFS_BMAP_INIT(&free_list, &firstfsb);
495         nimaps = 1;
496         error = xfs_bmapi(tp, ip, offset_fsb, count_fsb, bmapi_flag,
497                 &firstfsb, 0, &imap, &nimaps, &free_list, NULL);
498         if (error)
499                 goto error0;
500
501         /*
502          * Complete the transaction
503          */
504         error = xfs_bmap_finish(&tp, &free_list, &committed);
505         if (error)
506                 goto error0;
507         error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
508         if (error)
509                 goto error_out;
510
511         /*
512          * Copy any maps to caller's array and return any error.
513          */
514         if (nimaps == 0) {
515                 error = ENOSPC;
516                 goto error_out;
517         }
518
519         if (!(imap.br_startblock || XFS_IS_REALTIME_INODE(ip))) {
520                 error = xfs_cmn_err_fsblock_zero(ip, &imap);
521                 goto error_out;
522         }
523
524         *ret_imap = imap;
525         *nmaps = 1;
526         return 0;
527
528 error0: /* Cancel bmap, unlock inode, unreserve quota blocks, cancel trans */
529         xfs_bmap_cancel(&free_list);
530         XFS_TRANS_UNRESERVE_QUOTA_NBLKS(mp, tp, ip, qblocks, 0, quota_flag);
531
532 error1: /* Just cancel transaction */
533         xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES | XFS_TRANS_ABORT);
534         *nmaps = 0;     /* nothing set-up here */
535
536 error_out:
537         return XFS_ERROR(error);
538 }
539
540 /*
541  * If the caller is doing a write at the end of the file,
542  * then extend the allocation out to the file system's write
543  * iosize.  We clean up any extra space left over when the
544  * file is closed in xfs_inactive().
545  *
546  * For sync writes, we are flushing delayed allocate space to
547  * try to make additional space available for allocation near
548  * the filesystem full boundary - preallocation hurts in that
549  * situation, of course.
550  */
551 STATIC int
552 xfs_iomap_eof_want_preallocate(
553         xfs_mount_t     *mp,
554         xfs_inode_t     *ip,
555         xfs_off_t       offset,
556         size_t          count,
557         int             ioflag,
558         xfs_bmbt_irec_t *imap,
559         int             nimaps,
560         int             *prealloc)
561 {
562         xfs_fileoff_t   start_fsb;
563         xfs_filblks_t   count_fsb;
564         xfs_fsblock_t   firstblock;
565         int             n, error, imaps;
566
567         *prealloc = 0;
568         if ((ioflag & BMAPI_SYNC) || (offset + count) <= ip->i_size)
569                 return 0;
570
571         /*
572          * If there are any real blocks past eof, then don't
573          * do any speculative allocation.
574          */
575         start_fsb = XFS_B_TO_FSBT(mp, ((xfs_ufsize_t)(offset + count - 1)));
576         count_fsb = XFS_B_TO_FSB(mp, (xfs_ufsize_t)XFS_MAXIOFFSET(mp));
577         while (count_fsb > 0) {
578                 imaps = nimaps;
579                 firstblock = NULLFSBLOCK;
580                 error = xfs_bmapi(NULL, ip, start_fsb, count_fsb, 0,
581                                   &firstblock, 0, imap, &imaps, NULL, NULL);
582                 if (error)
583                         return error;
584                 for (n = 0; n < imaps; n++) {
585                         if ((imap[n].br_startblock != HOLESTARTBLOCK) &&
586                             (imap[n].br_startblock != DELAYSTARTBLOCK))
587                                 return 0;
588                         start_fsb += imap[n].br_blockcount;
589                         count_fsb -= imap[n].br_blockcount;
590                 }
591         }
592         *prealloc = 1;
593         return 0;
594 }
595
596 int
597 xfs_iomap_write_delay(
598         xfs_inode_t     *ip,
599         xfs_off_t       offset,
600         size_t          count,
601         int             ioflag,
602         xfs_bmbt_irec_t *ret_imap,
603         int             *nmaps)
604 {
605         xfs_mount_t     *mp = ip->i_mount;
606         xfs_fileoff_t   offset_fsb;
607         xfs_fileoff_t   last_fsb;
608         xfs_off_t       aligned_offset;
609         xfs_fileoff_t   ioalign;
610         xfs_fsblock_t   firstblock;
611         xfs_extlen_t    extsz;
612         int             nimaps;
613         xfs_bmbt_irec_t imap[XFS_WRITE_IMAPS];
614         int             prealloc, fsynced = 0;
615         int             error;
616
617         ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
618
619         /*
620          * Make sure that the dquots are there. This doesn't hold
621          * the ilock across a disk read.
622          */
623         error = XFS_QM_DQATTACH(mp, ip, XFS_QMOPT_ILOCKED);
624         if (error)
625                 return XFS_ERROR(error);
626
627         extsz = xfs_get_extsz_hint(ip);
628         offset_fsb = XFS_B_TO_FSBT(mp, offset);
629
630 retry:
631         error = xfs_iomap_eof_want_preallocate(mp, ip, offset, count,
632                                 ioflag, imap, XFS_WRITE_IMAPS, &prealloc);
633         if (error)
634                 return error;
635
636         if (prealloc) {
637                 aligned_offset = XFS_WRITEIO_ALIGN(mp, (offset + count - 1));
638                 ioalign = XFS_B_TO_FSBT(mp, aligned_offset);
639                 last_fsb = ioalign + mp->m_writeio_blocks;
640         } else {
641                 last_fsb = XFS_B_TO_FSB(mp, ((xfs_ufsize_t)(offset + count)));
642         }
643
644         if (prealloc || extsz) {
645                 error = xfs_iomap_eof_align_last_fsb(mp, ip, extsz, &last_fsb);
646                 if (error)
647                         return error;
648         }
649
650         nimaps = XFS_WRITE_IMAPS;
651         firstblock = NULLFSBLOCK;
652         error = xfs_bmapi(NULL, ip, offset_fsb,
653                           (xfs_filblks_t)(last_fsb - offset_fsb),
654                           XFS_BMAPI_DELAY | XFS_BMAPI_WRITE |
655                           XFS_BMAPI_ENTIRE, &firstblock, 1, imap,
656                           &nimaps, NULL, NULL);
657         if (error && (error != ENOSPC))
658                 return XFS_ERROR(error);
659
660         /*
661          * If bmapi returned us nothing, and if we didn't get back EDQUOT,
662          * then we must have run out of space - flush delalloc, and retry..
663          */
664         if (nimaps == 0) {
665                 xfs_iomap_enter_trace(XFS_IOMAP_WRITE_NOSPACE,
666                                         ip, offset, count);
667                 if (xfs_flush_space(ip, &fsynced, &ioflag))
668                         return XFS_ERROR(ENOSPC);
669
670                 error = 0;
671                 goto retry;
672         }
673
674         if (!(imap[0].br_startblock || XFS_IS_REALTIME_INODE(ip)))
675                 return xfs_cmn_err_fsblock_zero(ip, &imap[0]);
676
677         *ret_imap = imap[0];
678         *nmaps = 1;
679
680         return 0;
681 }
682
683 /*
684  * Pass in a delayed allocate extent, convert it to real extents;
685  * return to the caller the extent we create which maps on top of
686  * the originating callers request.
687  *
688  * Called without a lock on the inode.
689  *
690  * We no longer bother to look at the incoming map - all we have to
691  * guarantee is that whatever we allocate fills the required range.
692  */
693 int
694 xfs_iomap_write_allocate(
695         xfs_inode_t     *ip,
696         xfs_off_t       offset,
697         size_t          count,
698         xfs_bmbt_irec_t *map,
699         int             *retmap)
700 {
701         xfs_mount_t     *mp = ip->i_mount;
702         xfs_fileoff_t   offset_fsb, last_block;
703         xfs_fileoff_t   end_fsb, map_start_fsb;
704         xfs_fsblock_t   first_block;
705         xfs_bmap_free_t free_list;
706         xfs_filblks_t   count_fsb;
707         xfs_bmbt_irec_t imap;
708         xfs_trans_t     *tp;
709         int             nimaps, committed;
710         int             error = 0;
711         int             nres;
712
713         *retmap = 0;
714
715         /*
716          * Make sure that the dquots are there.
717          */
718         if ((error = XFS_QM_DQATTACH(mp, ip, 0)))
719                 return XFS_ERROR(error);
720
721         offset_fsb = XFS_B_TO_FSBT(mp, offset);
722         count_fsb = map->br_blockcount;
723         map_start_fsb = map->br_startoff;
724
725         XFS_STATS_ADD(xs_xstrat_bytes, XFS_FSB_TO_B(mp, count_fsb));
726
727         while (count_fsb != 0) {
728                 /*
729                  * Set up a transaction with which to allocate the
730                  * backing store for the file.  Do allocations in a
731                  * loop until we get some space in the range we are
732                  * interested in.  The other space that might be allocated
733                  * is in the delayed allocation extent on which we sit
734                  * but before our buffer starts.
735                  */
736
737                 nimaps = 0;
738                 while (nimaps == 0) {
739                         tp = xfs_trans_alloc(mp, XFS_TRANS_STRAT_WRITE);
740                         tp->t_flags |= XFS_TRANS_RESERVE;
741                         nres = XFS_EXTENTADD_SPACE_RES(mp, XFS_DATA_FORK);
742                         error = xfs_trans_reserve(tp, nres,
743                                         XFS_WRITE_LOG_RES(mp),
744                                         0, XFS_TRANS_PERM_LOG_RES,
745                                         XFS_WRITE_LOG_COUNT);
746                         if (error) {
747                                 xfs_trans_cancel(tp, 0);
748                                 return XFS_ERROR(error);
749                         }
750                         xfs_ilock(ip, XFS_ILOCK_EXCL);
751                         xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
752                         xfs_trans_ihold(tp, ip);
753
754                         XFS_BMAP_INIT(&free_list, &first_block);
755
756                         /*
757                          * it is possible that the extents have changed since
758                          * we did the read call as we dropped the ilock for a
759                          * while. We have to be careful about truncates or hole
760                          * punchs here - we are not allowed to allocate
761                          * non-delalloc blocks here.
762                          *
763                          * The only protection against truncation is the pages
764                          * for the range we are being asked to convert are
765                          * locked and hence a truncate will block on them
766                          * first.
767                          *
768                          * As a result, if we go beyond the range we really
769                          * need and hit an delalloc extent boundary followed by
770                          * a hole while we have excess blocks in the map, we
771                          * will fill the hole incorrectly and overrun the
772                          * transaction reservation.
773                          *
774                          * Using a single map prevents this as we are forced to
775                          * check each map we look for overlap with the desired
776                          * range and abort as soon as we find it. Also, given
777                          * that we only return a single map, having one beyond
778                          * what we can return is probably a bit silly.
779                          *
780                          * We also need to check that we don't go beyond EOF;
781                          * this is a truncate optimisation as a truncate sets
782                          * the new file size before block on the pages we
783                          * currently have locked under writeback. Because they
784                          * are about to be tossed, we don't need to write them
785                          * back....
786                          */
787                         nimaps = 1;
788                         end_fsb = XFS_B_TO_FSB(mp, ip->i_size);
789                         error = xfs_bmap_last_offset(NULL, ip, &last_block,
790                                                         XFS_DATA_FORK);
791                         if (error)
792                                 goto trans_cancel;
793
794                         last_block = XFS_FILEOFF_MAX(last_block, end_fsb);
795                         if ((map_start_fsb + count_fsb) > last_block) {
796                                 count_fsb = last_block - map_start_fsb;
797                                 if (count_fsb == 0) {
798                                         error = EAGAIN;
799                                         goto trans_cancel;
800                                 }
801                         }
802
803                         /* Go get the actual blocks */
804                         error = xfs_bmapi(tp, ip, map_start_fsb, count_fsb,
805                                         XFS_BMAPI_WRITE, &first_block, 1,
806                                         &imap, &nimaps, &free_list, NULL);
807                         if (error)
808                                 goto trans_cancel;
809
810                         error = xfs_bmap_finish(&tp, &free_list, &committed);
811                         if (error)
812                                 goto trans_cancel;
813
814                         error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
815                         if (error)
816                                 goto error0;
817
818                         xfs_iunlock(ip, XFS_ILOCK_EXCL);
819                 }
820
821                 /*
822                  * See if we were able to allocate an extent that
823                  * covers at least part of the callers request
824                  */
825                 if (!(imap.br_startblock || XFS_IS_REALTIME_INODE(ip)))
826                         return xfs_cmn_err_fsblock_zero(ip, &imap);
827
828                 if ((offset_fsb >= imap.br_startoff) &&
829                     (offset_fsb < (imap.br_startoff +
830                                    imap.br_blockcount))) {
831                         *map = imap;
832                         *retmap = 1;
833                         XFS_STATS_INC(xs_xstrat_quick);
834                         return 0;
835                 }
836
837                 /*
838                  * So far we have not mapped the requested part of the
839                  * file, just surrounding data, try again.
840                  */
841                 count_fsb -= imap.br_blockcount;
842                 map_start_fsb = imap.br_startoff + imap.br_blockcount;
843         }
844
845 trans_cancel:
846         xfs_bmap_cancel(&free_list);
847         xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES | XFS_TRANS_ABORT);
848 error0:
849         xfs_iunlock(ip, XFS_ILOCK_EXCL);
850         return XFS_ERROR(error);
851 }
852
853 int
854 xfs_iomap_write_unwritten(
855         xfs_inode_t     *ip,
856         xfs_off_t       offset,
857         size_t          count)
858 {
859         xfs_mount_t     *mp = ip->i_mount;
860         xfs_fileoff_t   offset_fsb;
861         xfs_filblks_t   count_fsb;
862         xfs_filblks_t   numblks_fsb;
863         xfs_fsblock_t   firstfsb;
864         int             nimaps;
865         xfs_trans_t     *tp;
866         xfs_bmbt_irec_t imap;
867         xfs_bmap_free_t free_list;
868         uint            resblks;
869         int             committed;
870         int             error;
871
872         xfs_iomap_enter_trace(XFS_IOMAP_UNWRITTEN, ip, offset, count);
873
874         offset_fsb = XFS_B_TO_FSBT(mp, offset);
875         count_fsb = XFS_B_TO_FSB(mp, (xfs_ufsize_t)offset + count);
876         count_fsb = (xfs_filblks_t)(count_fsb - offset_fsb);
877
878         /*
879          * Reserve enough blocks in this transaction for two complete extent
880          * btree splits.  We may be converting the middle part of an unwritten
881          * extent and in this case we will insert two new extents in the btree
882          * each of which could cause a full split.
883          *
884          * This reservation amount will be used in the first call to
885          * xfs_bmbt_split() to select an AG with enough space to satisfy the
886          * rest of the operation.
887          */
888         resblks = XFS_DIOSTRAT_SPACE_RES(mp, 0) << 1;
889
890         do {
891                 /*
892                  * set up a transaction to convert the range of extents
893                  * from unwritten to real. Do allocations in a loop until
894                  * we have covered the range passed in.
895                  */
896                 tp = xfs_trans_alloc(mp, XFS_TRANS_STRAT_WRITE);
897                 tp->t_flags |= XFS_TRANS_RESERVE;
898                 error = xfs_trans_reserve(tp, resblks,
899                                 XFS_WRITE_LOG_RES(mp), 0,
900                                 XFS_TRANS_PERM_LOG_RES,
901                                 XFS_WRITE_LOG_COUNT);
902                 if (error) {
903                         xfs_trans_cancel(tp, 0);
904                         return XFS_ERROR(error);
905                 }
906
907                 xfs_ilock(ip, XFS_ILOCK_EXCL);
908                 xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
909                 xfs_trans_ihold(tp, ip);
910
911                 /*
912                  * Modify the unwritten extent state of the buffer.
913                  */
914                 XFS_BMAP_INIT(&free_list, &firstfsb);
915                 nimaps = 1;
916                 error = xfs_bmapi(tp, ip, offset_fsb, count_fsb,
917                                   XFS_BMAPI_WRITE|XFS_BMAPI_CONVERT, &firstfsb,
918                                   1, &imap, &nimaps, &free_list, NULL);
919                 if (error)
920                         goto error_on_bmapi_transaction;
921
922                 error = xfs_bmap_finish(&(tp), &(free_list), &committed);
923                 if (error)
924                         goto error_on_bmapi_transaction;
925
926                 error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
927                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
928                 if (error)
929                         return XFS_ERROR(error);
930
931                 if (!(imap.br_startblock || XFS_IS_REALTIME_INODE(ip)))
932                         return xfs_cmn_err_fsblock_zero(ip, &imap);
933
934                 if ((numblks_fsb = imap.br_blockcount) == 0) {
935                         /*
936                          * The numblks_fsb value should always get
937                          * smaller, otherwise the loop is stuck.
938                          */
939                         ASSERT(imap.br_blockcount);
940                         break;
941                 }
942                 offset_fsb += numblks_fsb;
943                 count_fsb -= numblks_fsb;
944         } while (count_fsb > 0);
945
946         return 0;
947
948 error_on_bmapi_transaction:
949         xfs_bmap_cancel(&free_list);
950         xfs_trans_cancel(tp, (XFS_TRANS_RELEASE_LOG_RES | XFS_TRANS_ABORT));
951         xfs_iunlock(ip, XFS_ILOCK_EXCL);
952         return XFS_ERROR(error);
953 }