xfs: add size update tracepoint to IO completion
[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_alloc.h"
27 #include "xfs_quota.h"
28 #include "xfs_mount.h"
29 #include "xfs_bmap_btree.h"
30 #include "xfs_alloc_btree.h"
31 #include "xfs_ialloc_btree.h"
32 #include "xfs_dinode.h"
33 #include "xfs_inode.h"
34 #include "xfs_btree.h"
35 #include "xfs_bmap.h"
36 #include "xfs_rtalloc.h"
37 #include "xfs_error.h"
38 #include "xfs_itable.h"
39 #include "xfs_rw.h"
40 #include "xfs_attr.h"
41 #include "xfs_buf_item.h"
42 #include "xfs_trans_space.h"
43 #include "xfs_utils.h"
44 #include "xfs_iomap.h"
45 #include "xfs_trace.h"
46
47
48 #define XFS_WRITEIO_ALIGN(mp,off)       (((off) >> mp->m_writeio_log) \
49                                                 << mp->m_writeio_log)
50 #define XFS_WRITE_IMAPS         XFS_BMAP_MAX_NMAP
51
52 STATIC int
53 xfs_iomap_eof_align_last_fsb(
54         xfs_mount_t     *mp,
55         xfs_inode_t     *ip,
56         xfs_extlen_t    extsize,
57         xfs_fileoff_t   *last_fsb)
58 {
59         xfs_fileoff_t   new_last_fsb = 0;
60         xfs_extlen_t    align;
61         int             eof, error;
62
63         if (XFS_IS_REALTIME_INODE(ip))
64                 ;
65         /*
66          * If mounted with the "-o swalloc" option, roundup the allocation
67          * request to a stripe width boundary if the file size is >=
68          * stripe width and we are allocating past the allocation eof.
69          */
70         else if (mp->m_swidth && (mp->m_flags & XFS_MOUNT_SWALLOC) &&
71                 (ip->i_size >= XFS_FSB_TO_B(mp, mp->m_swidth)))
72                 new_last_fsb = roundup_64(*last_fsb, mp->m_swidth);
73         /*
74          * Roundup the allocation request to a stripe unit (m_dalign) boundary
75          * if the file size is >= stripe unit size, and we are allocating past
76          * the allocation eof.
77          */
78         else if (mp->m_dalign && (ip->i_size >= XFS_FSB_TO_B(mp, mp->m_dalign)))
79                 new_last_fsb = roundup_64(*last_fsb, mp->m_dalign);
80
81         /*
82          * Always round up the allocation request to an extent boundary
83          * (when file on a real-time subvolume or has di_extsize hint).
84          */
85         if (extsize) {
86                 if (new_last_fsb)
87                         align = roundup_64(new_last_fsb, extsize);
88                 else
89                         align = extsize;
90                 new_last_fsb = roundup_64(*last_fsb, align);
91         }
92
93         if (new_last_fsb) {
94                 error = xfs_bmap_eof(ip, new_last_fsb, XFS_DATA_FORK, &eof);
95                 if (error)
96                         return error;
97                 if (eof)
98                         *last_fsb = new_last_fsb;
99         }
100         return 0;
101 }
102
103 STATIC int
104 xfs_alert_fsblock_zero(
105         xfs_inode_t     *ip,
106         xfs_bmbt_irec_t *imap)
107 {
108         xfs_alert_tag(ip->i_mount, XFS_PTAG_FSBLOCK_ZERO,
109                         "Access to block zero in inode %llu "
110                         "start_block: %llx start_off: %llx "
111                         "blkcnt: %llx extent-state: %x\n",
112                 (unsigned long long)ip->i_ino,
113                 (unsigned long long)imap->br_startblock,
114                 (unsigned long long)imap->br_startoff,
115                 (unsigned long long)imap->br_blockcount,
116                 imap->br_state);
117         return EFSCORRUPTED;
118 }
119
120 int
121 xfs_iomap_write_direct(
122         xfs_inode_t     *ip,
123         xfs_off_t       offset,
124         size_t          count,
125         xfs_bmbt_irec_t *imap,
126         int             nmaps)
127 {
128         xfs_mount_t     *mp = ip->i_mount;
129         xfs_fileoff_t   offset_fsb;
130         xfs_fileoff_t   last_fsb;
131         xfs_filblks_t   count_fsb, resaligned;
132         xfs_fsblock_t   firstfsb;
133         xfs_extlen_t    extsz, temp;
134         int             nimaps;
135         int             bmapi_flag;
136         int             quota_flag;
137         int             rt;
138         xfs_trans_t     *tp;
139         xfs_bmap_free_t free_list;
140         uint            qblocks, resblks, resrtextents;
141         int             committed;
142         int             error;
143
144         /*
145          * Make sure that the dquots are there. This doesn't hold
146          * the ilock across a disk read.
147          */
148         error = xfs_qm_dqattach_locked(ip, 0);
149         if (error)
150                 return XFS_ERROR(error);
151
152         rt = XFS_IS_REALTIME_INODE(ip);
153         extsz = xfs_get_extsz_hint(ip);
154
155         offset_fsb = XFS_B_TO_FSBT(mp, offset);
156         last_fsb = XFS_B_TO_FSB(mp, ((xfs_ufsize_t)(offset + count)));
157         if ((offset + count) > ip->i_size) {
158                 error = xfs_iomap_eof_align_last_fsb(mp, ip, extsz, &last_fsb);
159                 if (error)
160                         goto error_out;
161         } else {
162                 if (nmaps && (imap->br_startblock == HOLESTARTBLOCK))
163                         last_fsb = MIN(last_fsb, (xfs_fileoff_t)
164                                         imap->br_blockcount +
165                                         imap->br_startoff);
166         }
167         count_fsb = last_fsb - offset_fsb;
168         ASSERT(count_fsb > 0);
169
170         resaligned = count_fsb;
171         if (unlikely(extsz)) {
172                 if ((temp = do_mod(offset_fsb, extsz)))
173                         resaligned += temp;
174                 if ((temp = do_mod(resaligned, extsz)))
175                         resaligned += extsz - temp;
176         }
177
178         if (unlikely(rt)) {
179                 resrtextents = qblocks = resaligned;
180                 resrtextents /= mp->m_sb.sb_rextsize;
181                 resblks = XFS_DIOSTRAT_SPACE_RES(mp, 0);
182                 quota_flag = XFS_QMOPT_RES_RTBLKS;
183         } else {
184                 resrtextents = 0;
185                 resblks = qblocks = XFS_DIOSTRAT_SPACE_RES(mp, resaligned);
186                 quota_flag = XFS_QMOPT_RES_REGBLKS;
187         }
188
189         /*
190          * Allocate and setup the transaction
191          */
192         xfs_iunlock(ip, XFS_ILOCK_EXCL);
193         tp = xfs_trans_alloc(mp, XFS_TRANS_DIOSTRAT);
194         error = xfs_trans_reserve(tp, resblks,
195                         XFS_WRITE_LOG_RES(mp), resrtextents,
196                         XFS_TRANS_PERM_LOG_RES,
197                         XFS_WRITE_LOG_COUNT);
198         /*
199          * Check for running out of space, note: need lock to return
200          */
201         if (error)
202                 xfs_trans_cancel(tp, 0);
203         xfs_ilock(ip, XFS_ILOCK_EXCL);
204         if (error)
205                 goto error_out;
206
207         error = xfs_trans_reserve_quota_nblks(tp, ip, qblocks, 0, quota_flag);
208         if (error)
209                 goto error1;
210
211         xfs_trans_ijoin(tp, ip);
212
213         bmapi_flag = XFS_BMAPI_WRITE;
214         if (offset < ip->i_size || extsz)
215                 bmapi_flag |= XFS_BMAPI_PREALLOC;
216
217         /*
218          * Issue the xfs_bmapi() call to allocate the blocks.
219          *
220          * From this point onwards we overwrite the imap pointer that the
221          * caller gave to us.
222          */
223         xfs_bmap_init(&free_list, &firstfsb);
224         nimaps = 1;
225         error = xfs_bmapi(tp, ip, offset_fsb, count_fsb, bmapi_flag,
226                 &firstfsb, 0, imap, &nimaps, &free_list);
227         if (error)
228                 goto error0;
229
230         /*
231          * Complete the transaction
232          */
233         error = xfs_bmap_finish(&tp, &free_list, &committed);
234         if (error)
235                 goto error0;
236         error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
237         if (error)
238                 goto error_out;
239
240         /*
241          * Copy any maps to caller's array and return any error.
242          */
243         if (nimaps == 0) {
244                 error = ENOSPC;
245                 goto error_out;
246         }
247
248         if (!(imap->br_startblock || XFS_IS_REALTIME_INODE(ip))) {
249                 error = xfs_alert_fsblock_zero(ip, imap);
250                 goto error_out;
251         }
252
253         return 0;
254
255 error0: /* Cancel bmap, unlock inode, unreserve quota blocks, cancel trans */
256         xfs_bmap_cancel(&free_list);
257         xfs_trans_unreserve_quota_nblks(tp, ip, qblocks, 0, quota_flag);
258
259 error1: /* Just cancel transaction */
260         xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES | XFS_TRANS_ABORT);
261
262 error_out:
263         return XFS_ERROR(error);
264 }
265
266 /*
267  * If the caller is doing a write at the end of the file, then extend the
268  * allocation out to the file system's write iosize.  We clean up any extra
269  * space left over when the file is closed in xfs_inactive().
270  *
271  * If we find we already have delalloc preallocation beyond EOF, don't do more
272  * preallocation as it it not needed.
273  */
274 STATIC int
275 xfs_iomap_eof_want_preallocate(
276         xfs_mount_t     *mp,
277         xfs_inode_t     *ip,
278         xfs_off_t       offset,
279         size_t          count,
280         xfs_bmbt_irec_t *imap,
281         int             nimaps,
282         int             *prealloc)
283 {
284         xfs_fileoff_t   start_fsb;
285         xfs_filblks_t   count_fsb;
286         xfs_fsblock_t   firstblock;
287         int             n, error, imaps;
288         int             found_delalloc = 0;
289
290         *prealloc = 0;
291         if ((offset + count) <= ip->i_size)
292                 return 0;
293
294         /*
295          * If there are any real blocks past eof, then don't
296          * do any speculative allocation.
297          */
298         start_fsb = XFS_B_TO_FSBT(mp, ((xfs_ufsize_t)(offset + count - 1)));
299         count_fsb = XFS_B_TO_FSB(mp, (xfs_ufsize_t)XFS_MAXIOFFSET(mp));
300         while (count_fsb > 0) {
301                 imaps = nimaps;
302                 firstblock = NULLFSBLOCK;
303                 error = xfs_bmapi(NULL, ip, start_fsb, count_fsb, 0,
304                                   &firstblock, 0, imap, &imaps, NULL);
305                 if (error)
306                         return error;
307                 for (n = 0; n < imaps; n++) {
308                         if ((imap[n].br_startblock != HOLESTARTBLOCK) &&
309                             (imap[n].br_startblock != DELAYSTARTBLOCK))
310                                 return 0;
311                         start_fsb += imap[n].br_blockcount;
312                         count_fsb -= imap[n].br_blockcount;
313
314                         if (imap[n].br_startblock == DELAYSTARTBLOCK)
315                                 found_delalloc = 1;
316                 }
317         }
318         if (!found_delalloc)
319                 *prealloc = 1;
320         return 0;
321 }
322
323 /*
324  * If we don't have a user specified preallocation size, dynamically increase
325  * the preallocation size as the size of the file grows. Cap the maximum size
326  * at a single extent or less if the filesystem is near full. The closer the
327  * filesystem is to full, the smaller the maximum prealocation.
328  */
329 STATIC xfs_fsblock_t
330 xfs_iomap_prealloc_size(
331         struct xfs_mount        *mp,
332         struct xfs_inode        *ip)
333 {
334         xfs_fsblock_t           alloc_blocks = 0;
335
336         if (!(mp->m_flags & XFS_MOUNT_DFLT_IOSIZE)) {
337                 int shift = 0;
338                 int64_t freesp;
339
340                 /*
341                  * rounddown_pow_of_two() returns an undefined result
342                  * if we pass in alloc_blocks = 0. Hence the "+ 1" to
343                  * ensure we always pass in a non-zero value.
344                  */
345                 alloc_blocks = XFS_B_TO_FSB(mp, ip->i_size) + 1;
346                 alloc_blocks = XFS_FILEOFF_MIN(MAXEXTLEN,
347                                         rounddown_pow_of_two(alloc_blocks));
348
349                 xfs_icsb_sync_counters(mp, XFS_ICSB_LAZY_COUNT);
350                 freesp = mp->m_sb.sb_fdblocks;
351                 if (freesp < mp->m_low_space[XFS_LOWSP_5_PCNT]) {
352                         shift = 2;
353                         if (freesp < mp->m_low_space[XFS_LOWSP_4_PCNT])
354                                 shift++;
355                         if (freesp < mp->m_low_space[XFS_LOWSP_3_PCNT])
356                                 shift++;
357                         if (freesp < mp->m_low_space[XFS_LOWSP_2_PCNT])
358                                 shift++;
359                         if (freesp < mp->m_low_space[XFS_LOWSP_1_PCNT])
360                                 shift++;
361                 }
362                 if (shift)
363                         alloc_blocks >>= shift;
364         }
365
366         if (alloc_blocks < mp->m_writeio_blocks)
367                 alloc_blocks = mp->m_writeio_blocks;
368
369         return alloc_blocks;
370 }
371
372 int
373 xfs_iomap_write_delay(
374         xfs_inode_t     *ip,
375         xfs_off_t       offset,
376         size_t          count,
377         xfs_bmbt_irec_t *ret_imap)
378 {
379         xfs_mount_t     *mp = ip->i_mount;
380         xfs_fileoff_t   offset_fsb;
381         xfs_fileoff_t   last_fsb;
382         xfs_off_t       aligned_offset;
383         xfs_fileoff_t   ioalign;
384         xfs_fsblock_t   firstblock;
385         xfs_extlen_t    extsz;
386         int             nimaps;
387         xfs_bmbt_irec_t imap[XFS_WRITE_IMAPS];
388         int             prealloc, flushed = 0;
389         int             error;
390
391         ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
392
393         /*
394          * Make sure that the dquots are there. This doesn't hold
395          * the ilock across a disk read.
396          */
397         error = xfs_qm_dqattach_locked(ip, 0);
398         if (error)
399                 return XFS_ERROR(error);
400
401         extsz = xfs_get_extsz_hint(ip);
402         offset_fsb = XFS_B_TO_FSBT(mp, offset);
403
404
405         error = xfs_iomap_eof_want_preallocate(mp, ip, offset, count,
406                                 imap, XFS_WRITE_IMAPS, &prealloc);
407         if (error)
408                 return error;
409
410 retry:
411         if (prealloc) {
412                 xfs_fsblock_t   alloc_blocks = xfs_iomap_prealloc_size(mp, ip);
413
414                 aligned_offset = XFS_WRITEIO_ALIGN(mp, (offset + count - 1));
415                 ioalign = XFS_B_TO_FSBT(mp, aligned_offset);
416                 last_fsb = ioalign + alloc_blocks;
417         } else {
418                 last_fsb = XFS_B_TO_FSB(mp, ((xfs_ufsize_t)(offset + count)));
419         }
420
421         if (prealloc || extsz) {
422                 error = xfs_iomap_eof_align_last_fsb(mp, ip, extsz, &last_fsb);
423                 if (error)
424                         return error;
425         }
426
427         nimaps = XFS_WRITE_IMAPS;
428         firstblock = NULLFSBLOCK;
429         error = xfs_bmapi(NULL, ip, offset_fsb,
430                           (xfs_filblks_t)(last_fsb - offset_fsb),
431                           XFS_BMAPI_DELAY | XFS_BMAPI_WRITE |
432                           XFS_BMAPI_ENTIRE, &firstblock, 1, imap,
433                           &nimaps, NULL);
434         switch (error) {
435         case 0:
436         case ENOSPC:
437         case EDQUOT:
438                 break;
439         default:
440                 return XFS_ERROR(error);
441         }
442
443         /*
444          * If bmapi returned us nothing, we got either ENOSPC or EDQUOT.  For
445          * ENOSPC, * flush all other inodes with delalloc blocks to free up
446          * some of the excess reserved metadata space. For both cases, retry
447          * without EOF preallocation.
448          */
449         if (nimaps == 0) {
450                 trace_xfs_delalloc_enospc(ip, offset, count);
451                 if (flushed)
452                         return XFS_ERROR(error ? error : ENOSPC);
453
454                 if (error == ENOSPC) {
455                         xfs_iunlock(ip, XFS_ILOCK_EXCL);
456                         xfs_flush_inodes(ip);
457                         xfs_ilock(ip, XFS_ILOCK_EXCL);
458                 }
459
460                 flushed = 1;
461                 error = 0;
462                 prealloc = 0;
463                 goto retry;
464         }
465
466         if (!(imap[0].br_startblock || XFS_IS_REALTIME_INODE(ip)))
467                 return xfs_alert_fsblock_zero(ip, &imap[0]);
468
469         *ret_imap = imap[0];
470         return 0;
471 }
472
473 /*
474  * Pass in a delayed allocate extent, convert it to real extents;
475  * return to the caller the extent we create which maps on top of
476  * the originating callers request.
477  *
478  * Called without a lock on the inode.
479  *
480  * We no longer bother to look at the incoming map - all we have to
481  * guarantee is that whatever we allocate fills the required range.
482  */
483 int
484 xfs_iomap_write_allocate(
485         xfs_inode_t     *ip,
486         xfs_off_t       offset,
487         size_t          count,
488         xfs_bmbt_irec_t *imap)
489 {
490         xfs_mount_t     *mp = ip->i_mount;
491         xfs_fileoff_t   offset_fsb, last_block;
492         xfs_fileoff_t   end_fsb, map_start_fsb;
493         xfs_fsblock_t   first_block;
494         xfs_bmap_free_t free_list;
495         xfs_filblks_t   count_fsb;
496         xfs_trans_t     *tp;
497         int             nimaps, committed;
498         int             error = 0;
499         int             nres;
500
501         /*
502          * Make sure that the dquots are there.
503          */
504         error = xfs_qm_dqattach(ip, 0);
505         if (error)
506                 return XFS_ERROR(error);
507
508         offset_fsb = XFS_B_TO_FSBT(mp, offset);
509         count_fsb = imap->br_blockcount;
510         map_start_fsb = imap->br_startoff;
511
512         XFS_STATS_ADD(xs_xstrat_bytes, XFS_FSB_TO_B(mp, count_fsb));
513
514         while (count_fsb != 0) {
515                 /*
516                  * Set up a transaction with which to allocate the
517                  * backing store for the file.  Do allocations in a
518                  * loop until we get some space in the range we are
519                  * interested in.  The other space that might be allocated
520                  * is in the delayed allocation extent on which we sit
521                  * but before our buffer starts.
522                  */
523
524                 nimaps = 0;
525                 while (nimaps == 0) {
526                         tp = xfs_trans_alloc(mp, XFS_TRANS_STRAT_WRITE);
527                         tp->t_flags |= XFS_TRANS_RESERVE;
528                         nres = XFS_EXTENTADD_SPACE_RES(mp, XFS_DATA_FORK);
529                         error = xfs_trans_reserve(tp, nres,
530                                         XFS_WRITE_LOG_RES(mp),
531                                         0, XFS_TRANS_PERM_LOG_RES,
532                                         XFS_WRITE_LOG_COUNT);
533                         if (error) {
534                                 xfs_trans_cancel(tp, 0);
535                                 return XFS_ERROR(error);
536                         }
537                         xfs_ilock(ip, XFS_ILOCK_EXCL);
538                         xfs_trans_ijoin(tp, ip);
539
540                         xfs_bmap_init(&free_list, &first_block);
541
542                         /*
543                          * it is possible that the extents have changed since
544                          * we did the read call as we dropped the ilock for a
545                          * while. We have to be careful about truncates or hole
546                          * punchs here - we are not allowed to allocate
547                          * non-delalloc blocks here.
548                          *
549                          * The only protection against truncation is the pages
550                          * for the range we are being asked to convert are
551                          * locked and hence a truncate will block on them
552                          * first.
553                          *
554                          * As a result, if we go beyond the range we really
555                          * need and hit an delalloc extent boundary followed by
556                          * a hole while we have excess blocks in the map, we
557                          * will fill the hole incorrectly and overrun the
558                          * transaction reservation.
559                          *
560                          * Using a single map prevents this as we are forced to
561                          * check each map we look for overlap with the desired
562                          * range and abort as soon as we find it. Also, given
563                          * that we only return a single map, having one beyond
564                          * what we can return is probably a bit silly.
565                          *
566                          * We also need to check that we don't go beyond EOF;
567                          * this is a truncate optimisation as a truncate sets
568                          * the new file size before block on the pages we
569                          * currently have locked under writeback. Because they
570                          * are about to be tossed, we don't need to write them
571                          * back....
572                          */
573                         nimaps = 1;
574                         end_fsb = XFS_B_TO_FSB(mp, ip->i_size);
575                         error = xfs_bmap_last_offset(NULL, ip, &last_block,
576                                                         XFS_DATA_FORK);
577                         if (error)
578                                 goto trans_cancel;
579
580                         last_block = XFS_FILEOFF_MAX(last_block, end_fsb);
581                         if ((map_start_fsb + count_fsb) > last_block) {
582                                 count_fsb = last_block - map_start_fsb;
583                                 if (count_fsb == 0) {
584                                         error = EAGAIN;
585                                         goto trans_cancel;
586                                 }
587                         }
588
589                         /*
590                          * Go get the actual blocks.
591                          *
592                          * From this point onwards we overwrite the imap
593                          * pointer that the caller gave to us.
594                          */
595                         error = xfs_bmapi(tp, ip, map_start_fsb, count_fsb,
596                                         XFS_BMAPI_WRITE, &first_block, 1,
597                                         imap, &nimaps, &free_list);
598                         if (error)
599                                 goto trans_cancel;
600
601                         error = xfs_bmap_finish(&tp, &free_list, &committed);
602                         if (error)
603                                 goto trans_cancel;
604
605                         error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
606                         if (error)
607                                 goto error0;
608
609                         xfs_iunlock(ip, XFS_ILOCK_EXCL);
610                 }
611
612                 /*
613                  * See if we were able to allocate an extent that
614                  * covers at least part of the callers request
615                  */
616                 if (!(imap->br_startblock || XFS_IS_REALTIME_INODE(ip)))
617                         return xfs_alert_fsblock_zero(ip, imap);
618
619                 if ((offset_fsb >= imap->br_startoff) &&
620                     (offset_fsb < (imap->br_startoff +
621                                    imap->br_blockcount))) {
622                         XFS_STATS_INC(xs_xstrat_quick);
623                         return 0;
624                 }
625
626                 /*
627                  * So far we have not mapped the requested part of the
628                  * file, just surrounding data, try again.
629                  */
630                 count_fsb -= imap->br_blockcount;
631                 map_start_fsb = imap->br_startoff + imap->br_blockcount;
632         }
633
634 trans_cancel:
635         xfs_bmap_cancel(&free_list);
636         xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES | XFS_TRANS_ABORT);
637 error0:
638         xfs_iunlock(ip, XFS_ILOCK_EXCL);
639         return XFS_ERROR(error);
640 }
641
642 int
643 xfs_iomap_write_unwritten(
644         xfs_inode_t     *ip,
645         xfs_off_t       offset,
646         size_t          count)
647 {
648         xfs_mount_t     *mp = ip->i_mount;
649         xfs_fileoff_t   offset_fsb;
650         xfs_filblks_t   count_fsb;
651         xfs_filblks_t   numblks_fsb;
652         xfs_fsblock_t   firstfsb;
653         int             nimaps;
654         xfs_trans_t     *tp;
655         xfs_bmbt_irec_t imap;
656         xfs_bmap_free_t free_list;
657         uint            resblks;
658         int             committed;
659         int             error;
660
661         trace_xfs_unwritten_convert(ip, offset, count);
662
663         offset_fsb = XFS_B_TO_FSBT(mp, offset);
664         count_fsb = XFS_B_TO_FSB(mp, (xfs_ufsize_t)offset + count);
665         count_fsb = (xfs_filblks_t)(count_fsb - offset_fsb);
666
667         /*
668          * Reserve enough blocks in this transaction for two complete extent
669          * btree splits.  We may be converting the middle part of an unwritten
670          * extent and in this case we will insert two new extents in the btree
671          * each of which could cause a full split.
672          *
673          * This reservation amount will be used in the first call to
674          * xfs_bmbt_split() to select an AG with enough space to satisfy the
675          * rest of the operation.
676          */
677         resblks = XFS_DIOSTRAT_SPACE_RES(mp, 0) << 1;
678
679         do {
680                 /*
681                  * set up a transaction to convert the range of extents
682                  * from unwritten to real. Do allocations in a loop until
683                  * we have covered the range passed in.
684                  *
685                  * Note that we open code the transaction allocation here
686                  * to pass KM_NOFS--we can't risk to recursing back into
687                  * the filesystem here as we might be asked to write out
688                  * the same inode that we complete here and might deadlock
689                  * on the iolock.
690                  */
691                 xfs_wait_for_freeze(mp, SB_FREEZE_TRANS);
692                 tp = _xfs_trans_alloc(mp, XFS_TRANS_STRAT_WRITE, KM_NOFS);
693                 tp->t_flags |= XFS_TRANS_RESERVE;
694                 error = xfs_trans_reserve(tp, resblks,
695                                 XFS_WRITE_LOG_RES(mp), 0,
696                                 XFS_TRANS_PERM_LOG_RES,
697                                 XFS_WRITE_LOG_COUNT);
698                 if (error) {
699                         xfs_trans_cancel(tp, 0);
700                         return XFS_ERROR(error);
701                 }
702
703                 xfs_ilock(ip, XFS_ILOCK_EXCL);
704                 xfs_trans_ijoin(tp, ip);
705
706                 /*
707                  * Modify the unwritten extent state of the buffer.
708                  */
709                 xfs_bmap_init(&free_list, &firstfsb);
710                 nimaps = 1;
711                 error = xfs_bmapi(tp, ip, offset_fsb, count_fsb,
712                                   XFS_BMAPI_WRITE|XFS_BMAPI_CONVERT, &firstfsb,
713                                   1, &imap, &nimaps, &free_list);
714                 if (error)
715                         goto error_on_bmapi_transaction;
716
717                 error = xfs_bmap_finish(&(tp), &(free_list), &committed);
718                 if (error)
719                         goto error_on_bmapi_transaction;
720
721                 error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
722                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
723                 if (error)
724                         return XFS_ERROR(error);
725
726                 if (!(imap.br_startblock || XFS_IS_REALTIME_INODE(ip)))
727                         return xfs_alert_fsblock_zero(ip, &imap);
728
729                 if ((numblks_fsb = imap.br_blockcount) == 0) {
730                         /*
731                          * The numblks_fsb value should always get
732                          * smaller, otherwise the loop is stuck.
733                          */
734                         ASSERT(imap.br_blockcount);
735                         break;
736                 }
737                 offset_fsb += numblks_fsb;
738                 count_fsb -= numblks_fsb;
739         } while (count_fsb > 0);
740
741         return 0;
742
743 error_on_bmapi_transaction:
744         xfs_bmap_cancel(&free_list);
745         xfs_trans_cancel(tp, (XFS_TRANS_RELEASE_LOG_RES | XFS_TRANS_ABORT));
746         xfs_iunlock(ip, XFS_ILOCK_EXCL);
747         return XFS_ERROR(error);
748 }