Merge branch 'linus' into x86/tracehook
[linux-2.6] / fs / xfs / xfs_mount.c
1 /*
2  * Copyright (c) 2000-2005 Silicon Graphics, Inc.
3  * All Rights Reserved.
4  *
5  * This program is free software; you can redistribute it and/or
6  * modify it under the terms of the GNU General Public License as
7  * published by the Free Software Foundation.
8  *
9  * This program is distributed in the hope that it would be useful,
10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
12  * GNU General Public License for more details.
13  *
14  * You should have received a copy of the GNU General Public License
15  * along with this program; if not, write the Free Software Foundation,
16  * Inc.,  51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
17  */
18 #include "xfs.h"
19 #include "xfs_fs.h"
20 #include "xfs_types.h"
21 #include "xfs_bit.h"
22 #include "xfs_log.h"
23 #include "xfs_inum.h"
24 #include "xfs_trans.h"
25 #include "xfs_sb.h"
26 #include "xfs_ag.h"
27 #include "xfs_dir2.h"
28 #include "xfs_dmapi.h"
29 #include "xfs_mount.h"
30 #include "xfs_bmap_btree.h"
31 #include "xfs_alloc_btree.h"
32 #include "xfs_ialloc_btree.h"
33 #include "xfs_dir2_sf.h"
34 #include "xfs_attr_sf.h"
35 #include "xfs_dinode.h"
36 #include "xfs_inode.h"
37 #include "xfs_btree.h"
38 #include "xfs_ialloc.h"
39 #include "xfs_alloc.h"
40 #include "xfs_rtalloc.h"
41 #include "xfs_bmap.h"
42 #include "xfs_error.h"
43 #include "xfs_rw.h"
44 #include "xfs_quota.h"
45 #include "xfs_fsops.h"
46 #include "xfs_utils.h"
47
48 STATIC int      xfs_mount_log_sb(xfs_mount_t *, __int64_t);
49 STATIC int      xfs_uuid_mount(xfs_mount_t *);
50 STATIC void     xfs_unmountfs_wait(xfs_mount_t *);
51
52
53 #ifdef HAVE_PERCPU_SB
54 STATIC void     xfs_icsb_balance_counter(xfs_mount_t *, xfs_sb_field_t,
55                                                 int);
56 STATIC void     xfs_icsb_balance_counter_locked(xfs_mount_t *, xfs_sb_field_t,
57                                                 int);
58 STATIC int      xfs_icsb_modify_counters(xfs_mount_t *, xfs_sb_field_t,
59                                                 int64_t, int);
60 STATIC void     xfs_icsb_disable_counter(xfs_mount_t *, xfs_sb_field_t);
61
62 #else
63
64 #define xfs_icsb_balance_counter(mp, a, b)              do { } while (0)
65 #define xfs_icsb_balance_counter_locked(mp, a, b)       do { } while (0)
66 #define xfs_icsb_modify_counters(mp, a, b, c)           do { } while (0)
67
68 #endif
69
70 static const struct {
71         short offset;
72         short type;     /* 0 = integer
73                          * 1 = binary / string (no translation)
74                          */
75 } xfs_sb_info[] = {
76     { offsetof(xfs_sb_t, sb_magicnum),   0 },
77     { offsetof(xfs_sb_t, sb_blocksize),  0 },
78     { offsetof(xfs_sb_t, sb_dblocks),    0 },
79     { offsetof(xfs_sb_t, sb_rblocks),    0 },
80     { offsetof(xfs_sb_t, sb_rextents),   0 },
81     { offsetof(xfs_sb_t, sb_uuid),       1 },
82     { offsetof(xfs_sb_t, sb_logstart),   0 },
83     { offsetof(xfs_sb_t, sb_rootino),    0 },
84     { offsetof(xfs_sb_t, sb_rbmino),     0 },
85     { offsetof(xfs_sb_t, sb_rsumino),    0 },
86     { offsetof(xfs_sb_t, sb_rextsize),   0 },
87     { offsetof(xfs_sb_t, sb_agblocks),   0 },
88     { offsetof(xfs_sb_t, sb_agcount),    0 },
89     { offsetof(xfs_sb_t, sb_rbmblocks),  0 },
90     { offsetof(xfs_sb_t, sb_logblocks),  0 },
91     { offsetof(xfs_sb_t, sb_versionnum), 0 },
92     { offsetof(xfs_sb_t, sb_sectsize),   0 },
93     { offsetof(xfs_sb_t, sb_inodesize),  0 },
94     { offsetof(xfs_sb_t, sb_inopblock),  0 },
95     { offsetof(xfs_sb_t, sb_fname[0]),   1 },
96     { offsetof(xfs_sb_t, sb_blocklog),   0 },
97     { offsetof(xfs_sb_t, sb_sectlog),    0 },
98     { offsetof(xfs_sb_t, sb_inodelog),   0 },
99     { offsetof(xfs_sb_t, sb_inopblog),   0 },
100     { offsetof(xfs_sb_t, sb_agblklog),   0 },
101     { offsetof(xfs_sb_t, sb_rextslog),   0 },
102     { offsetof(xfs_sb_t, sb_inprogress), 0 },
103     { offsetof(xfs_sb_t, sb_imax_pct),   0 },
104     { offsetof(xfs_sb_t, sb_icount),     0 },
105     { offsetof(xfs_sb_t, sb_ifree),      0 },
106     { offsetof(xfs_sb_t, sb_fdblocks),   0 },
107     { offsetof(xfs_sb_t, sb_frextents),  0 },
108     { offsetof(xfs_sb_t, sb_uquotino),   0 },
109     { offsetof(xfs_sb_t, sb_gquotino),   0 },
110     { offsetof(xfs_sb_t, sb_qflags),     0 },
111     { offsetof(xfs_sb_t, sb_flags),      0 },
112     { offsetof(xfs_sb_t, sb_shared_vn),  0 },
113     { offsetof(xfs_sb_t, sb_inoalignmt), 0 },
114     { offsetof(xfs_sb_t, sb_unit),       0 },
115     { offsetof(xfs_sb_t, sb_width),      0 },
116     { offsetof(xfs_sb_t, sb_dirblklog),  0 },
117     { offsetof(xfs_sb_t, sb_logsectlog), 0 },
118     { offsetof(xfs_sb_t, sb_logsectsize),0 },
119     { offsetof(xfs_sb_t, sb_logsunit),   0 },
120     { offsetof(xfs_sb_t, sb_features2),  0 },
121     { offsetof(xfs_sb_t, sb_bad_features2), 0 },
122     { sizeof(xfs_sb_t),                  0 }
123 };
124
125 /*
126  * Free up the resources associated with a mount structure.  Assume that
127  * the structure was initially zeroed, so we can tell which fields got
128  * initialized.
129  */
130 STATIC void
131 xfs_free_perag(
132         xfs_mount_t     *mp)
133 {
134         if (mp->m_perag) {
135                 int     agno;
136
137                 for (agno = 0; agno < mp->m_maxagi; agno++)
138                         if (mp->m_perag[agno].pagb_list)
139                                 kmem_free(mp->m_perag[agno].pagb_list);
140                 kmem_free(mp->m_perag);
141         }
142 }
143
144 /*
145  * Check size of device based on the (data/realtime) block count.
146  * Note: this check is used by the growfs code as well as mount.
147  */
148 int
149 xfs_sb_validate_fsb_count(
150         xfs_sb_t        *sbp,
151         __uint64_t      nblocks)
152 {
153         ASSERT(PAGE_SHIFT >= sbp->sb_blocklog);
154         ASSERT(sbp->sb_blocklog >= BBSHIFT);
155
156 #if XFS_BIG_BLKNOS     /* Limited by ULONG_MAX of page cache index */
157         if (nblocks >> (PAGE_CACHE_SHIFT - sbp->sb_blocklog) > ULONG_MAX)
158                 return E2BIG;
159 #else                  /* Limited by UINT_MAX of sectors */
160         if (nblocks << (sbp->sb_blocklog - BBSHIFT) > UINT_MAX)
161                 return E2BIG;
162 #endif
163         return 0;
164 }
165
166 /*
167  * Check the validity of the SB found.
168  */
169 STATIC int
170 xfs_mount_validate_sb(
171         xfs_mount_t     *mp,
172         xfs_sb_t        *sbp,
173         int             flags)
174 {
175         /*
176          * If the log device and data device have the
177          * same device number, the log is internal.
178          * Consequently, the sb_logstart should be non-zero.  If
179          * we have a zero sb_logstart in this case, we may be trying to mount
180          * a volume filesystem in a non-volume manner.
181          */
182         if (sbp->sb_magicnum != XFS_SB_MAGIC) {
183                 xfs_fs_mount_cmn_err(flags, "bad magic number");
184                 return XFS_ERROR(EWRONGFS);
185         }
186
187         if (!xfs_sb_good_version(sbp)) {
188                 xfs_fs_mount_cmn_err(flags, "bad version");
189                 return XFS_ERROR(EWRONGFS);
190         }
191
192         if (unlikely(
193             sbp->sb_logstart == 0 && mp->m_logdev_targp == mp->m_ddev_targp)) {
194                 xfs_fs_mount_cmn_err(flags,
195                         "filesystem is marked as having an external log; "
196                         "specify logdev on the\nmount command line.");
197                 return XFS_ERROR(EINVAL);
198         }
199
200         if (unlikely(
201             sbp->sb_logstart != 0 && mp->m_logdev_targp != mp->m_ddev_targp)) {
202                 xfs_fs_mount_cmn_err(flags,
203                         "filesystem is marked as having an internal log; "
204                         "do not specify logdev on\nthe mount command line.");
205                 return XFS_ERROR(EINVAL);
206         }
207
208         /*
209          * More sanity checking. These were stolen directly from
210          * xfs_repair.
211          */
212         if (unlikely(
213             sbp->sb_agcount <= 0                                        ||
214             sbp->sb_sectsize < XFS_MIN_SECTORSIZE                       ||
215             sbp->sb_sectsize > XFS_MAX_SECTORSIZE                       ||
216             sbp->sb_sectlog < XFS_MIN_SECTORSIZE_LOG                    ||
217             sbp->sb_sectlog > XFS_MAX_SECTORSIZE_LOG                    ||
218             sbp->sb_blocksize < XFS_MIN_BLOCKSIZE                       ||
219             sbp->sb_blocksize > XFS_MAX_BLOCKSIZE                       ||
220             sbp->sb_blocklog < XFS_MIN_BLOCKSIZE_LOG                    ||
221             sbp->sb_blocklog > XFS_MAX_BLOCKSIZE_LOG                    ||
222             sbp->sb_inodesize < XFS_DINODE_MIN_SIZE                     ||
223             sbp->sb_inodesize > XFS_DINODE_MAX_SIZE                     ||
224             sbp->sb_inodelog < XFS_DINODE_MIN_LOG                       ||
225             sbp->sb_inodelog > XFS_DINODE_MAX_LOG                       ||
226             (sbp->sb_blocklog - sbp->sb_inodelog != sbp->sb_inopblog)   ||
227             (sbp->sb_rextsize * sbp->sb_blocksize > XFS_MAX_RTEXTSIZE)  ||
228             (sbp->sb_rextsize * sbp->sb_blocksize < XFS_MIN_RTEXTSIZE)  ||
229             (sbp->sb_imax_pct > 100 /* zero sb_imax_pct is valid */))) {
230                 xfs_fs_mount_cmn_err(flags, "SB sanity check 1 failed");
231                 return XFS_ERROR(EFSCORRUPTED);
232         }
233
234         /*
235          * Sanity check AG count, size fields against data size field
236          */
237         if (unlikely(
238             sbp->sb_dblocks == 0 ||
239             sbp->sb_dblocks >
240              (xfs_drfsbno_t)sbp->sb_agcount * sbp->sb_agblocks ||
241             sbp->sb_dblocks < (xfs_drfsbno_t)(sbp->sb_agcount - 1) *
242                               sbp->sb_agblocks + XFS_MIN_AG_BLOCKS)) {
243                 xfs_fs_mount_cmn_err(flags, "SB sanity check 2 failed");
244                 return XFS_ERROR(EFSCORRUPTED);
245         }
246
247         /*
248          * Until this is fixed only page-sized or smaller data blocks work.
249          */
250         if (unlikely(sbp->sb_blocksize > PAGE_SIZE)) {
251                 xfs_fs_mount_cmn_err(flags,
252                         "file system with blocksize %d bytes",
253                         sbp->sb_blocksize);
254                 xfs_fs_mount_cmn_err(flags,
255                         "only pagesize (%ld) or less will currently work.",
256                         PAGE_SIZE);
257                 return XFS_ERROR(ENOSYS);
258         }
259
260         if (xfs_sb_validate_fsb_count(sbp, sbp->sb_dblocks) ||
261             xfs_sb_validate_fsb_count(sbp, sbp->sb_rblocks)) {
262                 xfs_fs_mount_cmn_err(flags,
263                         "file system too large to be mounted on this system.");
264                 return XFS_ERROR(E2BIG);
265         }
266
267         if (unlikely(sbp->sb_inprogress)) {
268                 xfs_fs_mount_cmn_err(flags, "file system busy");
269                 return XFS_ERROR(EFSCORRUPTED);
270         }
271
272         /*
273          * Version 1 directory format has never worked on Linux.
274          */
275         if (unlikely(!xfs_sb_version_hasdirv2(sbp))) {
276                 xfs_fs_mount_cmn_err(flags,
277                         "file system using version 1 directory format");
278                 return XFS_ERROR(ENOSYS);
279         }
280
281         return 0;
282 }
283
284 STATIC void
285 xfs_initialize_perag_icache(
286         xfs_perag_t     *pag)
287 {
288         if (!pag->pag_ici_init) {
289                 rwlock_init(&pag->pag_ici_lock);
290                 INIT_RADIX_TREE(&pag->pag_ici_root, GFP_ATOMIC);
291                 pag->pag_ici_init = 1;
292         }
293 }
294
295 xfs_agnumber_t
296 xfs_initialize_perag(
297         xfs_mount_t     *mp,
298         xfs_agnumber_t  agcount)
299 {
300         xfs_agnumber_t  index, max_metadata;
301         xfs_perag_t     *pag;
302         xfs_agino_t     agino;
303         xfs_ino_t       ino;
304         xfs_sb_t        *sbp = &mp->m_sb;
305         xfs_ino_t       max_inum = XFS_MAXINUMBER_32;
306
307         /* Check to see if the filesystem can overflow 32 bit inodes */
308         agino = XFS_OFFBNO_TO_AGINO(mp, sbp->sb_agblocks - 1, 0);
309         ino = XFS_AGINO_TO_INO(mp, agcount - 1, agino);
310
311         /* Clear the mount flag if no inode can overflow 32 bits
312          * on this filesystem, or if specifically requested..
313          */
314         if ((mp->m_flags & XFS_MOUNT_SMALL_INUMS) && ino > max_inum) {
315                 mp->m_flags |= XFS_MOUNT_32BITINODES;
316         } else {
317                 mp->m_flags &= ~XFS_MOUNT_32BITINODES;
318         }
319
320         /* If we can overflow then setup the ag headers accordingly */
321         if (mp->m_flags & XFS_MOUNT_32BITINODES) {
322                 /* Calculate how much should be reserved for inodes to
323                  * meet the max inode percentage.
324                  */
325                 if (mp->m_maxicount) {
326                         __uint64_t      icount;
327
328                         icount = sbp->sb_dblocks * sbp->sb_imax_pct;
329                         do_div(icount, 100);
330                         icount += sbp->sb_agblocks - 1;
331                         do_div(icount, sbp->sb_agblocks);
332                         max_metadata = icount;
333                 } else {
334                         max_metadata = agcount;
335                 }
336                 for (index = 0; index < agcount; index++) {
337                         ino = XFS_AGINO_TO_INO(mp, index, agino);
338                         if (ino > max_inum) {
339                                 index++;
340                                 break;
341                         }
342
343                         /* This ag is preferred for inodes */
344                         pag = &mp->m_perag[index];
345                         pag->pagi_inodeok = 1;
346                         if (index < max_metadata)
347                                 pag->pagf_metadata = 1;
348                         xfs_initialize_perag_icache(pag);
349                 }
350         } else {
351                 /* Setup default behavior for smaller filesystems */
352                 for (index = 0; index < agcount; index++) {
353                         pag = &mp->m_perag[index];
354                         pag->pagi_inodeok = 1;
355                         xfs_initialize_perag_icache(pag);
356                 }
357         }
358         return index;
359 }
360
361 void
362 xfs_sb_from_disk(
363         xfs_sb_t        *to,
364         xfs_dsb_t       *from)
365 {
366         to->sb_magicnum = be32_to_cpu(from->sb_magicnum);
367         to->sb_blocksize = be32_to_cpu(from->sb_blocksize);
368         to->sb_dblocks = be64_to_cpu(from->sb_dblocks);
369         to->sb_rblocks = be64_to_cpu(from->sb_rblocks);
370         to->sb_rextents = be64_to_cpu(from->sb_rextents);
371         memcpy(&to->sb_uuid, &from->sb_uuid, sizeof(to->sb_uuid));
372         to->sb_logstart = be64_to_cpu(from->sb_logstart);
373         to->sb_rootino = be64_to_cpu(from->sb_rootino);
374         to->sb_rbmino = be64_to_cpu(from->sb_rbmino);
375         to->sb_rsumino = be64_to_cpu(from->sb_rsumino);
376         to->sb_rextsize = be32_to_cpu(from->sb_rextsize);
377         to->sb_agblocks = be32_to_cpu(from->sb_agblocks);
378         to->sb_agcount = be32_to_cpu(from->sb_agcount);
379         to->sb_rbmblocks = be32_to_cpu(from->sb_rbmblocks);
380         to->sb_logblocks = be32_to_cpu(from->sb_logblocks);
381         to->sb_versionnum = be16_to_cpu(from->sb_versionnum);
382         to->sb_sectsize = be16_to_cpu(from->sb_sectsize);
383         to->sb_inodesize = be16_to_cpu(from->sb_inodesize);
384         to->sb_inopblock = be16_to_cpu(from->sb_inopblock);
385         memcpy(&to->sb_fname, &from->sb_fname, sizeof(to->sb_fname));
386         to->sb_blocklog = from->sb_blocklog;
387         to->sb_sectlog = from->sb_sectlog;
388         to->sb_inodelog = from->sb_inodelog;
389         to->sb_inopblog = from->sb_inopblog;
390         to->sb_agblklog = from->sb_agblklog;
391         to->sb_rextslog = from->sb_rextslog;
392         to->sb_inprogress = from->sb_inprogress;
393         to->sb_imax_pct = from->sb_imax_pct;
394         to->sb_icount = be64_to_cpu(from->sb_icount);
395         to->sb_ifree = be64_to_cpu(from->sb_ifree);
396         to->sb_fdblocks = be64_to_cpu(from->sb_fdblocks);
397         to->sb_frextents = be64_to_cpu(from->sb_frextents);
398         to->sb_uquotino = be64_to_cpu(from->sb_uquotino);
399         to->sb_gquotino = be64_to_cpu(from->sb_gquotino);
400         to->sb_qflags = be16_to_cpu(from->sb_qflags);
401         to->sb_flags = from->sb_flags;
402         to->sb_shared_vn = from->sb_shared_vn;
403         to->sb_inoalignmt = be32_to_cpu(from->sb_inoalignmt);
404         to->sb_unit = be32_to_cpu(from->sb_unit);
405         to->sb_width = be32_to_cpu(from->sb_width);
406         to->sb_dirblklog = from->sb_dirblklog;
407         to->sb_logsectlog = from->sb_logsectlog;
408         to->sb_logsectsize = be16_to_cpu(from->sb_logsectsize);
409         to->sb_logsunit = be32_to_cpu(from->sb_logsunit);
410         to->sb_features2 = be32_to_cpu(from->sb_features2);
411         to->sb_bad_features2 = be32_to_cpu(from->sb_bad_features2);
412 }
413
414 /*
415  * Copy in core superblock to ondisk one.
416  *
417  * The fields argument is mask of superblock fields to copy.
418  */
419 void
420 xfs_sb_to_disk(
421         xfs_dsb_t       *to,
422         xfs_sb_t        *from,
423         __int64_t       fields)
424 {
425         xfs_caddr_t     to_ptr = (xfs_caddr_t)to;
426         xfs_caddr_t     from_ptr = (xfs_caddr_t)from;
427         xfs_sb_field_t  f;
428         int             first;
429         int             size;
430
431         ASSERT(fields);
432         if (!fields)
433                 return;
434
435         while (fields) {
436                 f = (xfs_sb_field_t)xfs_lowbit64((__uint64_t)fields);
437                 first = xfs_sb_info[f].offset;
438                 size = xfs_sb_info[f + 1].offset - first;
439
440                 ASSERT(xfs_sb_info[f].type == 0 || xfs_sb_info[f].type == 1);
441
442                 if (size == 1 || xfs_sb_info[f].type == 1) {
443                         memcpy(to_ptr + first, from_ptr + first, size);
444                 } else {
445                         switch (size) {
446                         case 2:
447                                 *(__be16 *)(to_ptr + first) =
448                                         cpu_to_be16(*(__u16 *)(from_ptr + first));
449                                 break;
450                         case 4:
451                                 *(__be32 *)(to_ptr + first) =
452                                         cpu_to_be32(*(__u32 *)(from_ptr + first));
453                                 break;
454                         case 8:
455                                 *(__be64 *)(to_ptr + first) =
456                                         cpu_to_be64(*(__u64 *)(from_ptr + first));
457                                 break;
458                         default:
459                                 ASSERT(0);
460                         }
461                 }
462
463                 fields &= ~(1LL << f);
464         }
465 }
466
467 /*
468  * xfs_readsb
469  *
470  * Does the initial read of the superblock.
471  */
472 int
473 xfs_readsb(xfs_mount_t *mp, int flags)
474 {
475         unsigned int    sector_size;
476         unsigned int    extra_flags;
477         xfs_buf_t       *bp;
478         int             error;
479
480         ASSERT(mp->m_sb_bp == NULL);
481         ASSERT(mp->m_ddev_targp != NULL);
482
483         /*
484          * Allocate a (locked) buffer to hold the superblock.
485          * This will be kept around at all times to optimize
486          * access to the superblock.
487          */
488         sector_size = xfs_getsize_buftarg(mp->m_ddev_targp);
489         extra_flags = XFS_BUF_LOCK | XFS_BUF_MANAGE | XFS_BUF_MAPPED;
490
491         bp = xfs_buf_read_flags(mp->m_ddev_targp, XFS_SB_DADDR,
492                                 BTOBB(sector_size), extra_flags);
493         if (!bp || XFS_BUF_ISERROR(bp)) {
494                 xfs_fs_mount_cmn_err(flags, "SB read failed");
495                 error = bp ? XFS_BUF_GETERROR(bp) : ENOMEM;
496                 goto fail;
497         }
498         ASSERT(XFS_BUF_ISBUSY(bp));
499         ASSERT(XFS_BUF_VALUSEMA(bp) <= 0);
500
501         /*
502          * Initialize the mount structure from the superblock.
503          * But first do some basic consistency checking.
504          */
505         xfs_sb_from_disk(&mp->m_sb, XFS_BUF_TO_SBP(bp));
506
507         error = xfs_mount_validate_sb(mp, &(mp->m_sb), flags);
508         if (error) {
509                 xfs_fs_mount_cmn_err(flags, "SB validate failed");
510                 goto fail;
511         }
512
513         /*
514          * We must be able to do sector-sized and sector-aligned IO.
515          */
516         if (sector_size > mp->m_sb.sb_sectsize) {
517                 xfs_fs_mount_cmn_err(flags,
518                         "device supports only %u byte sectors (not %u)",
519                         sector_size, mp->m_sb.sb_sectsize);
520                 error = ENOSYS;
521                 goto fail;
522         }
523
524         /*
525          * If device sector size is smaller than the superblock size,
526          * re-read the superblock so the buffer is correctly sized.
527          */
528         if (sector_size < mp->m_sb.sb_sectsize) {
529                 XFS_BUF_UNMANAGE(bp);
530                 xfs_buf_relse(bp);
531                 sector_size = mp->m_sb.sb_sectsize;
532                 bp = xfs_buf_read_flags(mp->m_ddev_targp, XFS_SB_DADDR,
533                                         BTOBB(sector_size), extra_flags);
534                 if (!bp || XFS_BUF_ISERROR(bp)) {
535                         xfs_fs_mount_cmn_err(flags, "SB re-read failed");
536                         error = bp ? XFS_BUF_GETERROR(bp) : ENOMEM;
537                         goto fail;
538                 }
539                 ASSERT(XFS_BUF_ISBUSY(bp));
540                 ASSERT(XFS_BUF_VALUSEMA(bp) <= 0);
541         }
542
543         /* Initialize per-cpu counters */
544         xfs_icsb_reinit_counters(mp);
545
546         mp->m_sb_bp = bp;
547         xfs_buf_relse(bp);
548         ASSERT(XFS_BUF_VALUSEMA(bp) > 0);
549         return 0;
550
551  fail:
552         if (bp) {
553                 XFS_BUF_UNMANAGE(bp);
554                 xfs_buf_relse(bp);
555         }
556         return error;
557 }
558
559
560 /*
561  * xfs_mount_common
562  *
563  * Mount initialization code establishing various mount
564  * fields from the superblock associated with the given
565  * mount structure
566  */
567 STATIC void
568 xfs_mount_common(xfs_mount_t *mp, xfs_sb_t *sbp)
569 {
570         int     i;
571
572         mp->m_agfrotor = mp->m_agirotor = 0;
573         spin_lock_init(&mp->m_agirotor_lock);
574         mp->m_maxagi = mp->m_sb.sb_agcount;
575         mp->m_blkbit_log = sbp->sb_blocklog + XFS_NBBYLOG;
576         mp->m_blkbb_log = sbp->sb_blocklog - BBSHIFT;
577         mp->m_sectbb_log = sbp->sb_sectlog - BBSHIFT;
578         mp->m_agno_log = xfs_highbit32(sbp->sb_agcount - 1) + 1;
579         mp->m_agino_log = sbp->sb_inopblog + sbp->sb_agblklog;
580         mp->m_litino = sbp->sb_inodesize -
581                 ((uint)sizeof(xfs_dinode_core_t) + (uint)sizeof(xfs_agino_t));
582         mp->m_blockmask = sbp->sb_blocksize - 1;
583         mp->m_blockwsize = sbp->sb_blocksize >> XFS_WORDLOG;
584         mp->m_blockwmask = mp->m_blockwsize - 1;
585         INIT_LIST_HEAD(&mp->m_del_inodes);
586
587         /*
588          * Setup for attributes, in case they get created.
589          * This value is for inodes getting attributes for the first time,
590          * the per-inode value is for old attribute values.
591          */
592         ASSERT(sbp->sb_inodesize >= 256 && sbp->sb_inodesize <= 2048);
593         switch (sbp->sb_inodesize) {
594         case 256:
595                 mp->m_attroffset = XFS_LITINO(mp) -
596                                    XFS_BMDR_SPACE_CALC(MINABTPTRS);
597                 break;
598         case 512:
599         case 1024:
600         case 2048:
601                 mp->m_attroffset = XFS_BMDR_SPACE_CALC(6 * MINABTPTRS);
602                 break;
603         default:
604                 ASSERT(0);
605         }
606         ASSERT(mp->m_attroffset < XFS_LITINO(mp));
607
608         for (i = 0; i < 2; i++) {
609                 mp->m_alloc_mxr[i] = XFS_BTREE_BLOCK_MAXRECS(sbp->sb_blocksize,
610                         xfs_alloc, i == 0);
611                 mp->m_alloc_mnr[i] = XFS_BTREE_BLOCK_MINRECS(sbp->sb_blocksize,
612                         xfs_alloc, i == 0);
613         }
614         for (i = 0; i < 2; i++) {
615                 mp->m_bmap_dmxr[i] = XFS_BTREE_BLOCK_MAXRECS(sbp->sb_blocksize,
616                         xfs_bmbt, i == 0);
617                 mp->m_bmap_dmnr[i] = XFS_BTREE_BLOCK_MINRECS(sbp->sb_blocksize,
618                         xfs_bmbt, i == 0);
619         }
620         for (i = 0; i < 2; i++) {
621                 mp->m_inobt_mxr[i] = XFS_BTREE_BLOCK_MAXRECS(sbp->sb_blocksize,
622                         xfs_inobt, i == 0);
623                 mp->m_inobt_mnr[i] = XFS_BTREE_BLOCK_MINRECS(sbp->sb_blocksize,
624                         xfs_inobt, i == 0);
625         }
626
627         mp->m_bsize = XFS_FSB_TO_BB(mp, 1);
628         mp->m_ialloc_inos = (int)MAX((__uint16_t)XFS_INODES_PER_CHUNK,
629                                         sbp->sb_inopblock);
630         mp->m_ialloc_blks = mp->m_ialloc_inos >> sbp->sb_inopblog;
631 }
632
633 /*
634  * xfs_initialize_perag_data
635  *
636  * Read in each per-ag structure so we can count up the number of
637  * allocated inodes, free inodes and used filesystem blocks as this
638  * information is no longer persistent in the superblock. Once we have
639  * this information, write it into the in-core superblock structure.
640  */
641 STATIC int
642 xfs_initialize_perag_data(xfs_mount_t *mp, xfs_agnumber_t agcount)
643 {
644         xfs_agnumber_t  index;
645         xfs_perag_t     *pag;
646         xfs_sb_t        *sbp = &mp->m_sb;
647         uint64_t        ifree = 0;
648         uint64_t        ialloc = 0;
649         uint64_t        bfree = 0;
650         uint64_t        bfreelst = 0;
651         uint64_t        btree = 0;
652         int             error;
653
654         for (index = 0; index < agcount; index++) {
655                 /*
656                  * read the agf, then the agi. This gets us
657                  * all the inforamtion we need and populates the
658                  * per-ag structures for us.
659                  */
660                 error = xfs_alloc_pagf_init(mp, NULL, index, 0);
661                 if (error)
662                         return error;
663
664                 error = xfs_ialloc_pagi_init(mp, NULL, index);
665                 if (error)
666                         return error;
667                 pag = &mp->m_perag[index];
668                 ifree += pag->pagi_freecount;
669                 ialloc += pag->pagi_count;
670                 bfree += pag->pagf_freeblks;
671                 bfreelst += pag->pagf_flcount;
672                 btree += pag->pagf_btreeblks;
673         }
674         /*
675          * Overwrite incore superblock counters with just-read data
676          */
677         spin_lock(&mp->m_sb_lock);
678         sbp->sb_ifree = ifree;
679         sbp->sb_icount = ialloc;
680         sbp->sb_fdblocks = bfree + bfreelst + btree;
681         spin_unlock(&mp->m_sb_lock);
682
683         /* Fixup the per-cpu counters as well. */
684         xfs_icsb_reinit_counters(mp);
685
686         return 0;
687 }
688
689 /*
690  * Update alignment values based on mount options and sb values
691  */
692 STATIC int
693 xfs_update_alignment(xfs_mount_t *mp, __uint64_t *update_flags)
694 {
695         xfs_sb_t        *sbp = &(mp->m_sb);
696
697         if (mp->m_dalign) {
698                 /*
699                  * If stripe unit and stripe width are not multiples
700                  * of the fs blocksize turn off alignment.
701                  */
702                 if ((BBTOB(mp->m_dalign) & mp->m_blockmask) ||
703                     (BBTOB(mp->m_swidth) & mp->m_blockmask)) {
704                         if (mp->m_flags & XFS_MOUNT_RETERR) {
705                                 cmn_err(CE_WARN,
706                                         "XFS: alignment check 1 failed");
707                                 return XFS_ERROR(EINVAL);
708                         }
709                         mp->m_dalign = mp->m_swidth = 0;
710                 } else {
711                         /*
712                          * Convert the stripe unit and width to FSBs.
713                          */
714                         mp->m_dalign = XFS_BB_TO_FSBT(mp, mp->m_dalign);
715                         if (mp->m_dalign && (sbp->sb_agblocks % mp->m_dalign)) {
716                                 if (mp->m_flags & XFS_MOUNT_RETERR) {
717                                         return XFS_ERROR(EINVAL);
718                                 }
719                                 xfs_fs_cmn_err(CE_WARN, mp,
720 "stripe alignment turned off: sunit(%d)/swidth(%d) incompatible with agsize(%d)",
721                                         mp->m_dalign, mp->m_swidth,
722                                         sbp->sb_agblocks);
723
724                                 mp->m_dalign = 0;
725                                 mp->m_swidth = 0;
726                         } else if (mp->m_dalign) {
727                                 mp->m_swidth = XFS_BB_TO_FSBT(mp, mp->m_swidth);
728                         } else {
729                                 if (mp->m_flags & XFS_MOUNT_RETERR) {
730                                         xfs_fs_cmn_err(CE_WARN, mp,
731 "stripe alignment turned off: sunit(%d) less than bsize(%d)",
732                                                 mp->m_dalign,
733                                                 mp->m_blockmask +1);
734                                         return XFS_ERROR(EINVAL);
735                                 }
736                                 mp->m_swidth = 0;
737                         }
738                 }
739
740                 /*
741                  * Update superblock with new values
742                  * and log changes
743                  */
744                 if (xfs_sb_version_hasdalign(sbp)) {
745                         if (sbp->sb_unit != mp->m_dalign) {
746                                 sbp->sb_unit = mp->m_dalign;
747                                 *update_flags |= XFS_SB_UNIT;
748                         }
749                         if (sbp->sb_width != mp->m_swidth) {
750                                 sbp->sb_width = mp->m_swidth;
751                                 *update_flags |= XFS_SB_WIDTH;
752                         }
753                 }
754         } else if ((mp->m_flags & XFS_MOUNT_NOALIGN) != XFS_MOUNT_NOALIGN &&
755                     xfs_sb_version_hasdalign(&mp->m_sb)) {
756                         mp->m_dalign = sbp->sb_unit;
757                         mp->m_swidth = sbp->sb_width;
758         }
759
760         return 0;
761 }
762
763 /*
764  * Set the maximum inode count for this filesystem
765  */
766 STATIC void
767 xfs_set_maxicount(xfs_mount_t *mp)
768 {
769         xfs_sb_t        *sbp = &(mp->m_sb);
770         __uint64_t      icount;
771
772         if (sbp->sb_imax_pct) {
773                 /*
774                  * Make sure the maximum inode count is a multiple
775                  * of the units we allocate inodes in.
776                  */
777                 icount = sbp->sb_dblocks * sbp->sb_imax_pct;
778                 do_div(icount, 100);
779                 do_div(icount, mp->m_ialloc_blks);
780                 mp->m_maxicount = (icount * mp->m_ialloc_blks)  <<
781                                    sbp->sb_inopblog;
782         } else {
783                 mp->m_maxicount = 0;
784         }
785 }
786
787 /*
788  * Set the default minimum read and write sizes unless
789  * already specified in a mount option.
790  * We use smaller I/O sizes when the file system
791  * is being used for NFS service (wsync mount option).
792  */
793 STATIC void
794 xfs_set_rw_sizes(xfs_mount_t *mp)
795 {
796         xfs_sb_t        *sbp = &(mp->m_sb);
797         int             readio_log, writeio_log;
798
799         if (!(mp->m_flags & XFS_MOUNT_DFLT_IOSIZE)) {
800                 if (mp->m_flags & XFS_MOUNT_WSYNC) {
801                         readio_log = XFS_WSYNC_READIO_LOG;
802                         writeio_log = XFS_WSYNC_WRITEIO_LOG;
803                 } else {
804                         readio_log = XFS_READIO_LOG_LARGE;
805                         writeio_log = XFS_WRITEIO_LOG_LARGE;
806                 }
807         } else {
808                 readio_log = mp->m_readio_log;
809                 writeio_log = mp->m_writeio_log;
810         }
811
812         if (sbp->sb_blocklog > readio_log) {
813                 mp->m_readio_log = sbp->sb_blocklog;
814         } else {
815                 mp->m_readio_log = readio_log;
816         }
817         mp->m_readio_blocks = 1 << (mp->m_readio_log - sbp->sb_blocklog);
818         if (sbp->sb_blocklog > writeio_log) {
819                 mp->m_writeio_log = sbp->sb_blocklog;
820         } else {
821                 mp->m_writeio_log = writeio_log;
822         }
823         mp->m_writeio_blocks = 1 << (mp->m_writeio_log - sbp->sb_blocklog);
824 }
825
826 /*
827  * Set whether we're using inode alignment.
828  */
829 STATIC void
830 xfs_set_inoalignment(xfs_mount_t *mp)
831 {
832         if (xfs_sb_version_hasalign(&mp->m_sb) &&
833             mp->m_sb.sb_inoalignmt >=
834             XFS_B_TO_FSBT(mp, mp->m_inode_cluster_size))
835                 mp->m_inoalign_mask = mp->m_sb.sb_inoalignmt - 1;
836         else
837                 mp->m_inoalign_mask = 0;
838         /*
839          * If we are using stripe alignment, check whether
840          * the stripe unit is a multiple of the inode alignment
841          */
842         if (mp->m_dalign && mp->m_inoalign_mask &&
843             !(mp->m_dalign & mp->m_inoalign_mask))
844                 mp->m_sinoalign = mp->m_dalign;
845         else
846                 mp->m_sinoalign = 0;
847 }
848
849 /*
850  * Check that the data (and log if separate) are an ok size.
851  */
852 STATIC int
853 xfs_check_sizes(xfs_mount_t *mp)
854 {
855         xfs_buf_t       *bp;
856         xfs_daddr_t     d;
857         int             error;
858
859         d = (xfs_daddr_t)XFS_FSB_TO_BB(mp, mp->m_sb.sb_dblocks);
860         if (XFS_BB_TO_FSB(mp, d) != mp->m_sb.sb_dblocks) {
861                 cmn_err(CE_WARN, "XFS: size check 1 failed");
862                 return XFS_ERROR(E2BIG);
863         }
864         error = xfs_read_buf(mp, mp->m_ddev_targp,
865                              d - XFS_FSS_TO_BB(mp, 1),
866                              XFS_FSS_TO_BB(mp, 1), 0, &bp);
867         if (!error) {
868                 xfs_buf_relse(bp);
869         } else {
870                 cmn_err(CE_WARN, "XFS: size check 2 failed");
871                 if (error == ENOSPC)
872                         error = XFS_ERROR(E2BIG);
873                 return error;
874         }
875
876         if (mp->m_logdev_targp != mp->m_ddev_targp) {
877                 d = (xfs_daddr_t)XFS_FSB_TO_BB(mp, mp->m_sb.sb_logblocks);
878                 if (XFS_BB_TO_FSB(mp, d) != mp->m_sb.sb_logblocks) {
879                         cmn_err(CE_WARN, "XFS: size check 3 failed");
880                         return XFS_ERROR(E2BIG);
881                 }
882                 error = xfs_read_buf(mp, mp->m_logdev_targp,
883                                      d - XFS_FSB_TO_BB(mp, 1),
884                                      XFS_FSB_TO_BB(mp, 1), 0, &bp);
885                 if (!error) {
886                         xfs_buf_relse(bp);
887                 } else {
888                         cmn_err(CE_WARN, "XFS: size check 3 failed");
889                         if (error == ENOSPC)
890                                 error = XFS_ERROR(E2BIG);
891                         return error;
892                 }
893         }
894         return 0;
895 }
896
897 /*
898  * xfs_mountfs
899  *
900  * This function does the following on an initial mount of a file system:
901  *      - reads the superblock from disk and init the mount struct
902  *      - if we're a 32-bit kernel, do a size check on the superblock
903  *              so we don't mount terabyte filesystems
904  *      - init mount struct realtime fields
905  *      - allocate inode hash table for fs
906  *      - init directory manager
907  *      - perform recovery and init the log manager
908  */
909 int
910 xfs_mountfs(
911         xfs_mount_t     *mp)
912 {
913         xfs_sb_t        *sbp = &(mp->m_sb);
914         xfs_inode_t     *rip;
915         __uint64_t      resblks;
916         __int64_t       update_flags = 0LL;
917         uint            quotamount, quotaflags;
918         int             uuid_mounted = 0;
919         int             error = 0;
920
921         xfs_mount_common(mp, sbp);
922
923         /*
924          * Check for a mismatched features2 values.  Older kernels
925          * read & wrote into the wrong sb offset for sb_features2
926          * on some platforms due to xfs_sb_t not being 64bit size aligned
927          * when sb_features2 was added, which made older superblock
928          * reading/writing routines swap it as a 64-bit value.
929          *
930          * For backwards compatibility, we make both slots equal.
931          *
932          * If we detect a mismatched field, we OR the set bits into the
933          * existing features2 field in case it has already been modified; we
934          * don't want to lose any features.  We then update the bad location
935          * with the ORed value so that older kernels will see any features2
936          * flags, and mark the two fields as needing updates once the
937          * transaction subsystem is online.
938          */
939         if (xfs_sb_has_mismatched_features2(sbp)) {
940                 cmn_err(CE_WARN,
941                         "XFS: correcting sb_features alignment problem");
942                 sbp->sb_features2 |= sbp->sb_bad_features2;
943                 sbp->sb_bad_features2 = sbp->sb_features2;
944                 update_flags |= XFS_SB_FEATURES2 | XFS_SB_BAD_FEATURES2;
945
946                 /*
947                  * Re-check for ATTR2 in case it was found in bad_features2
948                  * slot.
949                  */
950                 if (xfs_sb_version_hasattr2(&mp->m_sb) &&
951                    !(mp->m_flags & XFS_MOUNT_NOATTR2))
952                         mp->m_flags |= XFS_MOUNT_ATTR2;
953         }
954
955         if (xfs_sb_version_hasattr2(&mp->m_sb) &&
956            (mp->m_flags & XFS_MOUNT_NOATTR2)) {
957                 xfs_sb_version_removeattr2(&mp->m_sb);
958                 update_flags |= XFS_SB_FEATURES2;
959
960                 /* update sb_versionnum for the clearing of the morebits */
961                 if (!sbp->sb_features2)
962                         update_flags |= XFS_SB_VERSIONNUM;
963         }
964
965         /*
966          * Check if sb_agblocks is aligned at stripe boundary
967          * If sb_agblocks is NOT aligned turn off m_dalign since
968          * allocator alignment is within an ag, therefore ag has
969          * to be aligned at stripe boundary.
970          */
971         error = xfs_update_alignment(mp, &update_flags);
972         if (error)
973                 goto error1;
974
975         xfs_alloc_compute_maxlevels(mp);
976         xfs_bmap_compute_maxlevels(mp, XFS_DATA_FORK);
977         xfs_bmap_compute_maxlevels(mp, XFS_ATTR_FORK);
978         xfs_ialloc_compute_maxlevels(mp);
979
980         xfs_set_maxicount(mp);
981
982         mp->m_maxioffset = xfs_max_file_offset(sbp->sb_blocklog);
983
984         /*
985          * XFS uses the uuid from the superblock as the unique
986          * identifier for fsid.  We can not use the uuid from the volume
987          * since a single partition filesystem is identical to a single
988          * partition volume/filesystem.
989          */
990         if ((mp->m_flags & XFS_MOUNT_NOUUID) == 0) {
991                 if (xfs_uuid_mount(mp)) {
992                         error = XFS_ERROR(EINVAL);
993                         goto error1;
994                 }
995                 uuid_mounted=1;
996         }
997
998         /*
999          * Set the minimum read and write sizes
1000          */
1001         xfs_set_rw_sizes(mp);
1002
1003         /*
1004          * Set the inode cluster size.
1005          * This may still be overridden by the file system
1006          * block size if it is larger than the chosen cluster size.
1007          */
1008         mp->m_inode_cluster_size = XFS_INODE_BIG_CLUSTER_SIZE;
1009
1010         /*
1011          * Set inode alignment fields
1012          */
1013         xfs_set_inoalignment(mp);
1014
1015         /*
1016          * Check that the data (and log if separate) are an ok size.
1017          */
1018         error = xfs_check_sizes(mp);
1019         if (error)
1020                 goto error1;
1021
1022         /*
1023          * Initialize realtime fields in the mount structure
1024          */
1025         error = xfs_rtmount_init(mp);
1026         if (error) {
1027                 cmn_err(CE_WARN, "XFS: RT mount failed");
1028                 goto error1;
1029         }
1030
1031         /*
1032          *  Copies the low order bits of the timestamp and the randomly
1033          *  set "sequence" number out of a UUID.
1034          */
1035         uuid_getnodeuniq(&sbp->sb_uuid, mp->m_fixedfsid);
1036
1037         mp->m_dmevmask = 0;     /* not persistent; set after each mount */
1038
1039         xfs_dir_mount(mp);
1040
1041         /*
1042          * Initialize the attribute manager's entries.
1043          */
1044         mp->m_attr_magicpct = (mp->m_sb.sb_blocksize * 37) / 100;
1045
1046         /*
1047          * Initialize the precomputed transaction reservations values.
1048          */
1049         xfs_trans_init(mp);
1050
1051         /*
1052          * Allocate and initialize the per-ag data.
1053          */
1054         init_rwsem(&mp->m_peraglock);
1055         mp->m_perag = kmem_zalloc(sbp->sb_agcount * sizeof(xfs_perag_t),
1056                                   KM_MAYFAIL);
1057         if (!mp->m_perag)
1058                 goto error1;
1059
1060         mp->m_maxagi = xfs_initialize_perag(mp, sbp->sb_agcount);
1061
1062         /*
1063          * log's mount-time initialization. Perform 1st part recovery if needed
1064          */
1065         if (likely(sbp->sb_logblocks > 0)) {    /* check for volume case */
1066                 error = xfs_log_mount(mp, mp->m_logdev_targp,
1067                                       XFS_FSB_TO_DADDR(mp, sbp->sb_logstart),
1068                                       XFS_FSB_TO_BB(mp, sbp->sb_logblocks));
1069                 if (error) {
1070                         cmn_err(CE_WARN, "XFS: log mount failed");
1071                         goto error2;
1072                 }
1073         } else {        /* No log has been defined */
1074                 cmn_err(CE_WARN, "XFS: no log defined");
1075                 XFS_ERROR_REPORT("xfs_mountfs_int(1)", XFS_ERRLEVEL_LOW, mp);
1076                 error = XFS_ERROR(EFSCORRUPTED);
1077                 goto error2;
1078         }
1079
1080         /*
1081          * Now the log is mounted, we know if it was an unclean shutdown or
1082          * not. If it was, with the first phase of recovery has completed, we
1083          * have consistent AG blocks on disk. We have not recovered EFIs yet,
1084          * but they are recovered transactionally in the second recovery phase
1085          * later.
1086          *
1087          * Hence we can safely re-initialise incore superblock counters from
1088          * the per-ag data. These may not be correct if the filesystem was not
1089          * cleanly unmounted, so we need to wait for recovery to finish before
1090          * doing this.
1091          *
1092          * If the filesystem was cleanly unmounted, then we can trust the
1093          * values in the superblock to be correct and we don't need to do
1094          * anything here.
1095          *
1096          * If we are currently making the filesystem, the initialisation will
1097          * fail as the perag data is in an undefined state.
1098          */
1099
1100         if (xfs_sb_version_haslazysbcount(&mp->m_sb) &&
1101             !XFS_LAST_UNMOUNT_WAS_CLEAN(mp) &&
1102              !mp->m_sb.sb_inprogress) {
1103                 error = xfs_initialize_perag_data(mp, sbp->sb_agcount);
1104                 if (error) {
1105                         goto error2;
1106                 }
1107         }
1108         /*
1109          * Get and sanity-check the root inode.
1110          * Save the pointer to it in the mount structure.
1111          */
1112         error = xfs_iget(mp, NULL, sbp->sb_rootino, 0, XFS_ILOCK_EXCL, &rip, 0);
1113         if (error) {
1114                 cmn_err(CE_WARN, "XFS: failed to read root inode");
1115                 goto error3;
1116         }
1117
1118         ASSERT(rip != NULL);
1119
1120         if (unlikely((rip->i_d.di_mode & S_IFMT) != S_IFDIR)) {
1121                 cmn_err(CE_WARN, "XFS: corrupted root inode");
1122                 cmn_err(CE_WARN, "Device %s - root %llu is not a directory",
1123                         XFS_BUFTARG_NAME(mp->m_ddev_targp),
1124                         (unsigned long long)rip->i_ino);
1125                 xfs_iunlock(rip, XFS_ILOCK_EXCL);
1126                 XFS_ERROR_REPORT("xfs_mountfs_int(2)", XFS_ERRLEVEL_LOW,
1127                                  mp);
1128                 error = XFS_ERROR(EFSCORRUPTED);
1129                 goto error4;
1130         }
1131         mp->m_rootip = rip;     /* save it */
1132
1133         xfs_iunlock(rip, XFS_ILOCK_EXCL);
1134
1135         /*
1136          * Initialize realtime inode pointers in the mount structure
1137          */
1138         error = xfs_rtmount_inodes(mp);
1139         if (error) {
1140                 /*
1141                  * Free up the root inode.
1142                  */
1143                 cmn_err(CE_WARN, "XFS: failed to read RT inodes");
1144                 goto error4;
1145         }
1146
1147         /*
1148          * If fs is not mounted readonly, then update the superblock changes.
1149          */
1150         if (update_flags && !(mp->m_flags & XFS_MOUNT_RDONLY)) {
1151                 error = xfs_mount_log_sb(mp, update_flags);
1152                 if (error) {
1153                         cmn_err(CE_WARN, "XFS: failed to write sb changes");
1154                         goto error4;
1155                 }
1156         }
1157
1158         /*
1159          * Initialise the XFS quota management subsystem for this mount
1160          */
1161         error = XFS_QM_INIT(mp, &quotamount, &quotaflags);
1162         if (error)
1163                 goto error4;
1164
1165         /*
1166          * Finish recovering the file system.  This part needed to be
1167          * delayed until after the root and real-time bitmap inodes
1168          * were consistently read in.
1169          */
1170         error = xfs_log_mount_finish(mp);
1171         if (error) {
1172                 cmn_err(CE_WARN, "XFS: log mount finish failed");
1173                 goto error4;
1174         }
1175
1176         /*
1177          * Complete the quota initialisation, post-log-replay component.
1178          */
1179         error = XFS_QM_MOUNT(mp, quotamount, quotaflags);
1180         if (error)
1181                 goto error4;
1182
1183         /*
1184          * Now we are mounted, reserve a small amount of unused space for
1185          * privileged transactions. This is needed so that transaction
1186          * space required for critical operations can dip into this pool
1187          * when at ENOSPC. This is needed for operations like create with
1188          * attr, unwritten extent conversion at ENOSPC, etc. Data allocations
1189          * are not allowed to use this reserved space.
1190          *
1191          * We default to 5% or 1024 fsbs of space reserved, whichever is smaller.
1192          * This may drive us straight to ENOSPC on mount, but that implies
1193          * we were already there on the last unmount. Warn if this occurs.
1194          */
1195         resblks = mp->m_sb.sb_dblocks;
1196         do_div(resblks, 20);
1197         resblks = min_t(__uint64_t, resblks, 1024);
1198         error = xfs_reserve_blocks(mp, &resblks, NULL);
1199         if (error)
1200                 cmn_err(CE_WARN, "XFS: Unable to allocate reserve blocks. "
1201                                 "Continuing without a reserve pool.");
1202
1203         return 0;
1204
1205  error4:
1206         /*
1207          * Free up the root inode.
1208          */
1209         IRELE(rip);
1210  error3:
1211         xfs_log_unmount_dealloc(mp);
1212  error2:
1213         xfs_free_perag(mp);
1214  error1:
1215         if (uuid_mounted)
1216                 uuid_table_remove(&mp->m_sb.sb_uuid);
1217         return error;
1218 }
1219
1220 /*
1221  * This flushes out the inodes,dquots and the superblock, unmounts the
1222  * log and makes sure that incore structures are freed.
1223  */
1224 void
1225 xfs_unmountfs(
1226         struct xfs_mount        *mp)
1227 {
1228         __uint64_t              resblks;
1229         int                     error;
1230
1231         IRELE(mp->m_rootip);
1232
1233         /*
1234          * We can potentially deadlock here if we have an inode cluster
1235          * that has been freed has it's buffer still pinned in memory because
1236          * the transaction is still sitting in a iclog. The stale inodes
1237          * on that buffer will have their flush locks held until the
1238          * transaction hits the disk and the callbacks run. the inode
1239          * flush takes the flush lock unconditionally and with nothing to
1240          * push out the iclog we will never get that unlocked. hence we
1241          * need to force the log first.
1242          */
1243         xfs_log_force(mp, (xfs_lsn_t)0, XFS_LOG_FORCE | XFS_LOG_SYNC);
1244         xfs_iflush_all(mp);
1245
1246         XFS_QM_DQPURGEALL(mp, XFS_QMOPT_QUOTALL | XFS_QMOPT_UMOUNTING);
1247
1248         /*
1249          * Flush out the log synchronously so that we know for sure
1250          * that nothing is pinned.  This is important because bflush()
1251          * will skip pinned buffers.
1252          */
1253         xfs_log_force(mp, (xfs_lsn_t)0, XFS_LOG_FORCE | XFS_LOG_SYNC);
1254
1255         xfs_binval(mp->m_ddev_targp);
1256         if (mp->m_rtdev_targp) {
1257                 xfs_binval(mp->m_rtdev_targp);
1258         }
1259
1260         /*
1261          * Unreserve any blocks we have so that when we unmount we don't account
1262          * the reserved free space as used. This is really only necessary for
1263          * lazy superblock counting because it trusts the incore superblock
1264          * counters to be aboslutely correct on clean unmount.
1265          *
1266          * We don't bother correcting this elsewhere for lazy superblock
1267          * counting because on mount of an unclean filesystem we reconstruct the
1268          * correct counter value and this is irrelevant.
1269          *
1270          * For non-lazy counter filesystems, this doesn't matter at all because
1271          * we only every apply deltas to the superblock and hence the incore
1272          * value does not matter....
1273          */
1274         resblks = 0;
1275         error = xfs_reserve_blocks(mp, &resblks, NULL);
1276         if (error)
1277                 cmn_err(CE_WARN, "XFS: Unable to free reserved block pool. "
1278                                 "Freespace may not be correct on next mount.");
1279
1280         error = xfs_log_sbcount(mp, 1);
1281         if (error)
1282                 cmn_err(CE_WARN, "XFS: Unable to update superblock counters. "
1283                                 "Freespace may not be correct on next mount.");
1284         xfs_unmountfs_writesb(mp);
1285         xfs_unmountfs_wait(mp);                 /* wait for async bufs */
1286         xfs_log_unmount(mp);                    /* Done! No more fs ops. */
1287
1288         /*
1289          * All inodes from this mount point should be freed.
1290          */
1291         ASSERT(mp->m_inodes == NULL);
1292
1293         if ((mp->m_flags & XFS_MOUNT_NOUUID) == 0)
1294                 uuid_table_remove(&mp->m_sb.sb_uuid);
1295
1296 #if defined(DEBUG)
1297         xfs_errortag_clearall(mp, 0);
1298 #endif
1299         xfs_free_perag(mp);
1300         if (mp->m_quotainfo)
1301                 XFS_QM_DONE(mp);
1302 }
1303
1304 STATIC void
1305 xfs_unmountfs_wait(xfs_mount_t *mp)
1306 {
1307         if (mp->m_logdev_targp != mp->m_ddev_targp)
1308                 xfs_wait_buftarg(mp->m_logdev_targp);
1309         if (mp->m_rtdev_targp)
1310                 xfs_wait_buftarg(mp->m_rtdev_targp);
1311         xfs_wait_buftarg(mp->m_ddev_targp);
1312 }
1313
1314 int
1315 xfs_fs_writable(xfs_mount_t *mp)
1316 {
1317         return !(xfs_test_for_freeze(mp) || XFS_FORCED_SHUTDOWN(mp) ||
1318                 (mp->m_flags & XFS_MOUNT_RDONLY));
1319 }
1320
1321 /*
1322  * xfs_log_sbcount
1323  *
1324  * Called either periodically to keep the on disk superblock values
1325  * roughly up to date or from unmount to make sure the values are
1326  * correct on a clean unmount.
1327  *
1328  * Note this code can be called during the process of freezing, so
1329  * we may need to use the transaction allocator which does not not
1330  * block when the transaction subsystem is in its frozen state.
1331  */
1332 int
1333 xfs_log_sbcount(
1334         xfs_mount_t     *mp,
1335         uint            sync)
1336 {
1337         xfs_trans_t     *tp;
1338         int             error;
1339
1340         if (!xfs_fs_writable(mp))
1341                 return 0;
1342
1343         xfs_icsb_sync_counters(mp, 0);
1344
1345         /*
1346          * we don't need to do this if we are updating the superblock
1347          * counters on every modification.
1348          */
1349         if (!xfs_sb_version_haslazysbcount(&mp->m_sb))
1350                 return 0;
1351
1352         tp = _xfs_trans_alloc(mp, XFS_TRANS_SB_COUNT);
1353         error = xfs_trans_reserve(tp, 0, mp->m_sb.sb_sectsize + 128, 0, 0,
1354                                         XFS_DEFAULT_LOG_COUNT);
1355         if (error) {
1356                 xfs_trans_cancel(tp, 0);
1357                 return error;
1358         }
1359
1360         xfs_mod_sb(tp, XFS_SB_IFREE | XFS_SB_ICOUNT | XFS_SB_FDBLOCKS);
1361         if (sync)
1362                 xfs_trans_set_sync(tp);
1363         error = xfs_trans_commit(tp, 0);
1364         return error;
1365 }
1366
1367 STATIC void
1368 xfs_mark_shared_ro(
1369         xfs_mount_t     *mp,
1370         xfs_buf_t       *bp)
1371 {
1372         xfs_dsb_t       *sb = XFS_BUF_TO_SBP(bp);
1373         __uint16_t      version;
1374
1375         if (!(sb->sb_flags & XFS_SBF_READONLY))
1376                 sb->sb_flags |= XFS_SBF_READONLY;
1377
1378         version = be16_to_cpu(sb->sb_versionnum);
1379         if ((version & XFS_SB_VERSION_NUMBITS) != XFS_SB_VERSION_4 ||
1380             !(version & XFS_SB_VERSION_SHAREDBIT))
1381                 version |= XFS_SB_VERSION_SHAREDBIT;
1382         sb->sb_versionnum = cpu_to_be16(version);
1383 }
1384
1385 int
1386 xfs_unmountfs_writesb(xfs_mount_t *mp)
1387 {
1388         xfs_buf_t       *sbp;
1389         int             error = 0;
1390
1391         /*
1392          * skip superblock write if fs is read-only, or
1393          * if we are doing a forced umount.
1394          */
1395         if (!((mp->m_flags & XFS_MOUNT_RDONLY) ||
1396                 XFS_FORCED_SHUTDOWN(mp))) {
1397
1398                 sbp = xfs_getsb(mp, 0);
1399
1400                 /*
1401                  * mark shared-readonly if desired
1402                  */
1403                 if (mp->m_mk_sharedro)
1404                         xfs_mark_shared_ro(mp, sbp);
1405
1406                 XFS_BUF_UNDONE(sbp);
1407                 XFS_BUF_UNREAD(sbp);
1408                 XFS_BUF_UNDELAYWRITE(sbp);
1409                 XFS_BUF_WRITE(sbp);
1410                 XFS_BUF_UNASYNC(sbp);
1411                 ASSERT(XFS_BUF_TARGET(sbp) == mp->m_ddev_targp);
1412                 xfsbdstrat(mp, sbp);
1413                 error = xfs_iowait(sbp);
1414                 if (error)
1415                         xfs_ioerror_alert("xfs_unmountfs_writesb",
1416                                           mp, sbp, XFS_BUF_ADDR(sbp));
1417                 if (error && mp->m_mk_sharedro)
1418                         xfs_fs_cmn_err(CE_ALERT, mp, "Superblock write error detected while unmounting.  Filesystem may not be marked shared readonly");
1419                 xfs_buf_relse(sbp);
1420         }
1421         return error;
1422 }
1423
1424 /*
1425  * xfs_mod_sb() can be used to copy arbitrary changes to the
1426  * in-core superblock into the superblock buffer to be logged.
1427  * It does not provide the higher level of locking that is
1428  * needed to protect the in-core superblock from concurrent
1429  * access.
1430  */
1431 void
1432 xfs_mod_sb(xfs_trans_t *tp, __int64_t fields)
1433 {
1434         xfs_buf_t       *bp;
1435         int             first;
1436         int             last;
1437         xfs_mount_t     *mp;
1438         xfs_sb_field_t  f;
1439
1440         ASSERT(fields);
1441         if (!fields)
1442                 return;
1443         mp = tp->t_mountp;
1444         bp = xfs_trans_getsb(tp, mp, 0);
1445         first = sizeof(xfs_sb_t);
1446         last = 0;
1447
1448         /* translate/copy */
1449
1450         xfs_sb_to_disk(XFS_BUF_TO_SBP(bp), &mp->m_sb, fields);
1451
1452         /* find modified range */
1453
1454         f = (xfs_sb_field_t)xfs_lowbit64((__uint64_t)fields);
1455         ASSERT((1LL << f) & XFS_SB_MOD_BITS);
1456         first = xfs_sb_info[f].offset;
1457
1458         f = (xfs_sb_field_t)xfs_highbit64((__uint64_t)fields);
1459         ASSERT((1LL << f) & XFS_SB_MOD_BITS);
1460         last = xfs_sb_info[f + 1].offset - 1;
1461
1462         xfs_trans_log_buf(tp, bp, first, last);
1463 }
1464
1465
1466 /*
1467  * xfs_mod_incore_sb_unlocked() is a utility routine common used to apply
1468  * a delta to a specified field in the in-core superblock.  Simply
1469  * switch on the field indicated and apply the delta to that field.
1470  * Fields are not allowed to dip below zero, so if the delta would
1471  * do this do not apply it and return EINVAL.
1472  *
1473  * The m_sb_lock must be held when this routine is called.
1474  */
1475 int
1476 xfs_mod_incore_sb_unlocked(
1477         xfs_mount_t     *mp,
1478         xfs_sb_field_t  field,
1479         int64_t         delta,
1480         int             rsvd)
1481 {
1482         int             scounter;       /* short counter for 32 bit fields */
1483         long long       lcounter;       /* long counter for 64 bit fields */
1484         long long       res_used, rem;
1485
1486         /*
1487          * With the in-core superblock spin lock held, switch
1488          * on the indicated field.  Apply the delta to the
1489          * proper field.  If the fields value would dip below
1490          * 0, then do not apply the delta and return EINVAL.
1491          */
1492         switch (field) {
1493         case XFS_SBS_ICOUNT:
1494                 lcounter = (long long)mp->m_sb.sb_icount;
1495                 lcounter += delta;
1496                 if (lcounter < 0) {
1497                         ASSERT(0);
1498                         return XFS_ERROR(EINVAL);
1499                 }
1500                 mp->m_sb.sb_icount = lcounter;
1501                 return 0;
1502         case XFS_SBS_IFREE:
1503                 lcounter = (long long)mp->m_sb.sb_ifree;
1504                 lcounter += delta;
1505                 if (lcounter < 0) {
1506                         ASSERT(0);
1507                         return XFS_ERROR(EINVAL);
1508                 }
1509                 mp->m_sb.sb_ifree = lcounter;
1510                 return 0;
1511         case XFS_SBS_FDBLOCKS:
1512                 lcounter = (long long)
1513                         mp->m_sb.sb_fdblocks - XFS_ALLOC_SET_ASIDE(mp);
1514                 res_used = (long long)(mp->m_resblks - mp->m_resblks_avail);
1515
1516                 if (delta > 0) {                /* Putting blocks back */
1517                         if (res_used > delta) {
1518                                 mp->m_resblks_avail += delta;
1519                         } else {
1520                                 rem = delta - res_used;
1521                                 mp->m_resblks_avail = mp->m_resblks;
1522                                 lcounter += rem;
1523                         }
1524                 } else {                                /* Taking blocks away */
1525
1526                         lcounter += delta;
1527
1528                 /*
1529                  * If were out of blocks, use any available reserved blocks if
1530                  * were allowed to.
1531                  */
1532
1533                         if (lcounter < 0) {
1534                                 if (rsvd) {
1535                                         lcounter = (long long)mp->m_resblks_avail + delta;
1536                                         if (lcounter < 0) {
1537                                                 return XFS_ERROR(ENOSPC);
1538                                         }
1539                                         mp->m_resblks_avail = lcounter;
1540                                         return 0;
1541                                 } else {        /* not reserved */
1542                                         return XFS_ERROR(ENOSPC);
1543                                 }
1544                         }
1545                 }
1546
1547                 mp->m_sb.sb_fdblocks = lcounter + XFS_ALLOC_SET_ASIDE(mp);
1548                 return 0;
1549         case XFS_SBS_FREXTENTS:
1550                 lcounter = (long long)mp->m_sb.sb_frextents;
1551                 lcounter += delta;
1552                 if (lcounter < 0) {
1553                         return XFS_ERROR(ENOSPC);
1554                 }
1555                 mp->m_sb.sb_frextents = lcounter;
1556                 return 0;
1557         case XFS_SBS_DBLOCKS:
1558                 lcounter = (long long)mp->m_sb.sb_dblocks;
1559                 lcounter += delta;
1560                 if (lcounter < 0) {
1561                         ASSERT(0);
1562                         return XFS_ERROR(EINVAL);
1563                 }
1564                 mp->m_sb.sb_dblocks = lcounter;
1565                 return 0;
1566         case XFS_SBS_AGCOUNT:
1567                 scounter = mp->m_sb.sb_agcount;
1568                 scounter += delta;
1569                 if (scounter < 0) {
1570                         ASSERT(0);
1571                         return XFS_ERROR(EINVAL);
1572                 }
1573                 mp->m_sb.sb_agcount = scounter;
1574                 return 0;
1575         case XFS_SBS_IMAX_PCT:
1576                 scounter = mp->m_sb.sb_imax_pct;
1577                 scounter += delta;
1578                 if (scounter < 0) {
1579                         ASSERT(0);
1580                         return XFS_ERROR(EINVAL);
1581                 }
1582                 mp->m_sb.sb_imax_pct = scounter;
1583                 return 0;
1584         case XFS_SBS_REXTSIZE:
1585                 scounter = mp->m_sb.sb_rextsize;
1586                 scounter += delta;
1587                 if (scounter < 0) {
1588                         ASSERT(0);
1589                         return XFS_ERROR(EINVAL);
1590                 }
1591                 mp->m_sb.sb_rextsize = scounter;
1592                 return 0;
1593         case XFS_SBS_RBMBLOCKS:
1594                 scounter = mp->m_sb.sb_rbmblocks;
1595                 scounter += delta;
1596                 if (scounter < 0) {
1597                         ASSERT(0);
1598                         return XFS_ERROR(EINVAL);
1599                 }
1600                 mp->m_sb.sb_rbmblocks = scounter;
1601                 return 0;
1602         case XFS_SBS_RBLOCKS:
1603                 lcounter = (long long)mp->m_sb.sb_rblocks;
1604                 lcounter += delta;
1605                 if (lcounter < 0) {
1606                         ASSERT(0);
1607                         return XFS_ERROR(EINVAL);
1608                 }
1609                 mp->m_sb.sb_rblocks = lcounter;
1610                 return 0;
1611         case XFS_SBS_REXTENTS:
1612                 lcounter = (long long)mp->m_sb.sb_rextents;
1613                 lcounter += delta;
1614                 if (lcounter < 0) {
1615                         ASSERT(0);
1616                         return XFS_ERROR(EINVAL);
1617                 }
1618                 mp->m_sb.sb_rextents = lcounter;
1619                 return 0;
1620         case XFS_SBS_REXTSLOG:
1621                 scounter = mp->m_sb.sb_rextslog;
1622                 scounter += delta;
1623                 if (scounter < 0) {
1624                         ASSERT(0);
1625                         return XFS_ERROR(EINVAL);
1626                 }
1627                 mp->m_sb.sb_rextslog = scounter;
1628                 return 0;
1629         default:
1630                 ASSERT(0);
1631                 return XFS_ERROR(EINVAL);
1632         }
1633 }
1634
1635 /*
1636  * xfs_mod_incore_sb() is used to change a field in the in-core
1637  * superblock structure by the specified delta.  This modification
1638  * is protected by the m_sb_lock.  Just use the xfs_mod_incore_sb_unlocked()
1639  * routine to do the work.
1640  */
1641 int
1642 xfs_mod_incore_sb(
1643         xfs_mount_t     *mp,
1644         xfs_sb_field_t  field,
1645         int64_t         delta,
1646         int             rsvd)
1647 {
1648         int     status;
1649
1650         /* check for per-cpu counters */
1651         switch (field) {
1652 #ifdef HAVE_PERCPU_SB
1653         case XFS_SBS_ICOUNT:
1654         case XFS_SBS_IFREE:
1655         case XFS_SBS_FDBLOCKS:
1656                 if (!(mp->m_flags & XFS_MOUNT_NO_PERCPU_SB)) {
1657                         status = xfs_icsb_modify_counters(mp, field,
1658                                                         delta, rsvd);
1659                         break;
1660                 }
1661                 /* FALLTHROUGH */
1662 #endif
1663         default:
1664                 spin_lock(&mp->m_sb_lock);
1665                 status = xfs_mod_incore_sb_unlocked(mp, field, delta, rsvd);
1666                 spin_unlock(&mp->m_sb_lock);
1667                 break;
1668         }
1669
1670         return status;
1671 }
1672
1673 /*
1674  * xfs_mod_incore_sb_batch() is used to change more than one field
1675  * in the in-core superblock structure at a time.  This modification
1676  * is protected by a lock internal to this module.  The fields and
1677  * changes to those fields are specified in the array of xfs_mod_sb
1678  * structures passed in.
1679  *
1680  * Either all of the specified deltas will be applied or none of
1681  * them will.  If any modified field dips below 0, then all modifications
1682  * will be backed out and EINVAL will be returned.
1683  */
1684 int
1685 xfs_mod_incore_sb_batch(xfs_mount_t *mp, xfs_mod_sb_t *msb, uint nmsb, int rsvd)
1686 {
1687         int             status=0;
1688         xfs_mod_sb_t    *msbp;
1689
1690         /*
1691          * Loop through the array of mod structures and apply each
1692          * individually.  If any fail, then back out all those
1693          * which have already been applied.  Do all of this within
1694          * the scope of the m_sb_lock so that all of the changes will
1695          * be atomic.
1696          */
1697         spin_lock(&mp->m_sb_lock);
1698         msbp = &msb[0];
1699         for (msbp = &msbp[0]; msbp < (msb + nmsb); msbp++) {
1700                 /*
1701                  * Apply the delta at index n.  If it fails, break
1702                  * from the loop so we'll fall into the undo loop
1703                  * below.
1704                  */
1705                 switch (msbp->msb_field) {
1706 #ifdef HAVE_PERCPU_SB
1707                 case XFS_SBS_ICOUNT:
1708                 case XFS_SBS_IFREE:
1709                 case XFS_SBS_FDBLOCKS:
1710                         if (!(mp->m_flags & XFS_MOUNT_NO_PERCPU_SB)) {
1711                                 spin_unlock(&mp->m_sb_lock);
1712                                 status = xfs_icsb_modify_counters(mp,
1713                                                         msbp->msb_field,
1714                                                         msbp->msb_delta, rsvd);
1715                                 spin_lock(&mp->m_sb_lock);
1716                                 break;
1717                         }
1718                         /* FALLTHROUGH */
1719 #endif
1720                 default:
1721                         status = xfs_mod_incore_sb_unlocked(mp,
1722                                                 msbp->msb_field,
1723                                                 msbp->msb_delta, rsvd);
1724                         break;
1725                 }
1726
1727                 if (status != 0) {
1728                         break;
1729                 }
1730         }
1731
1732         /*
1733          * If we didn't complete the loop above, then back out
1734          * any changes made to the superblock.  If you add code
1735          * between the loop above and here, make sure that you
1736          * preserve the value of status. Loop back until
1737          * we step below the beginning of the array.  Make sure
1738          * we don't touch anything back there.
1739          */
1740         if (status != 0) {
1741                 msbp--;
1742                 while (msbp >= msb) {
1743                         switch (msbp->msb_field) {
1744 #ifdef HAVE_PERCPU_SB
1745                         case XFS_SBS_ICOUNT:
1746                         case XFS_SBS_IFREE:
1747                         case XFS_SBS_FDBLOCKS:
1748                                 if (!(mp->m_flags & XFS_MOUNT_NO_PERCPU_SB)) {
1749                                         spin_unlock(&mp->m_sb_lock);
1750                                         status = xfs_icsb_modify_counters(mp,
1751                                                         msbp->msb_field,
1752                                                         -(msbp->msb_delta),
1753                                                         rsvd);
1754                                         spin_lock(&mp->m_sb_lock);
1755                                         break;
1756                                 }
1757                                 /* FALLTHROUGH */
1758 #endif
1759                         default:
1760                                 status = xfs_mod_incore_sb_unlocked(mp,
1761                                                         msbp->msb_field,
1762                                                         -(msbp->msb_delta),
1763                                                         rsvd);
1764                                 break;
1765                         }
1766                         ASSERT(status == 0);
1767                         msbp--;
1768                 }
1769         }
1770         spin_unlock(&mp->m_sb_lock);
1771         return status;
1772 }
1773
1774 /*
1775  * xfs_getsb() is called to obtain the buffer for the superblock.
1776  * The buffer is returned locked and read in from disk.
1777  * The buffer should be released with a call to xfs_brelse().
1778  *
1779  * If the flags parameter is BUF_TRYLOCK, then we'll only return
1780  * the superblock buffer if it can be locked without sleeping.
1781  * If it can't then we'll return NULL.
1782  */
1783 xfs_buf_t *
1784 xfs_getsb(
1785         xfs_mount_t     *mp,
1786         int             flags)
1787 {
1788         xfs_buf_t       *bp;
1789
1790         ASSERT(mp->m_sb_bp != NULL);
1791         bp = mp->m_sb_bp;
1792         if (flags & XFS_BUF_TRYLOCK) {
1793                 if (!XFS_BUF_CPSEMA(bp)) {
1794                         return NULL;
1795                 }
1796         } else {
1797                 XFS_BUF_PSEMA(bp, PRIBIO);
1798         }
1799         XFS_BUF_HOLD(bp);
1800         ASSERT(XFS_BUF_ISDONE(bp));
1801         return bp;
1802 }
1803
1804 /*
1805  * Used to free the superblock along various error paths.
1806  */
1807 void
1808 xfs_freesb(
1809         xfs_mount_t     *mp)
1810 {
1811         xfs_buf_t       *bp;
1812
1813         /*
1814          * Use xfs_getsb() so that the buffer will be locked
1815          * when we call xfs_buf_relse().
1816          */
1817         bp = xfs_getsb(mp, 0);
1818         XFS_BUF_UNMANAGE(bp);
1819         xfs_buf_relse(bp);
1820         mp->m_sb_bp = NULL;
1821 }
1822
1823 /*
1824  * See if the UUID is unique among mounted XFS filesystems.
1825  * Mount fails if UUID is nil or a FS with the same UUID is already mounted.
1826  */
1827 STATIC int
1828 xfs_uuid_mount(
1829         xfs_mount_t     *mp)
1830 {
1831         if (uuid_is_nil(&mp->m_sb.sb_uuid)) {
1832                 cmn_err(CE_WARN,
1833                         "XFS: Filesystem %s has nil UUID - can't mount",
1834                         mp->m_fsname);
1835                 return -1;
1836         }
1837         if (!uuid_table_insert(&mp->m_sb.sb_uuid)) {
1838                 cmn_err(CE_WARN,
1839                         "XFS: Filesystem %s has duplicate UUID - can't mount",
1840                         mp->m_fsname);
1841                 return -1;
1842         }
1843         return 0;
1844 }
1845
1846 /*
1847  * Used to log changes to the superblock unit and width fields which could
1848  * be altered by the mount options, as well as any potential sb_features2
1849  * fixup. Only the first superblock is updated.
1850  */
1851 STATIC int
1852 xfs_mount_log_sb(
1853         xfs_mount_t     *mp,
1854         __int64_t       fields)
1855 {
1856         xfs_trans_t     *tp;
1857         int             error;
1858
1859         ASSERT(fields & (XFS_SB_UNIT | XFS_SB_WIDTH | XFS_SB_UUID |
1860                          XFS_SB_FEATURES2 | XFS_SB_BAD_FEATURES2 |
1861                          XFS_SB_VERSIONNUM));
1862
1863         tp = xfs_trans_alloc(mp, XFS_TRANS_SB_UNIT);
1864         error = xfs_trans_reserve(tp, 0, mp->m_sb.sb_sectsize + 128, 0, 0,
1865                                 XFS_DEFAULT_LOG_COUNT);
1866         if (error) {
1867                 xfs_trans_cancel(tp, 0);
1868                 return error;
1869         }
1870         xfs_mod_sb(tp, fields);
1871         error = xfs_trans_commit(tp, 0);
1872         return error;
1873 }
1874
1875
1876 #ifdef HAVE_PERCPU_SB
1877 /*
1878  * Per-cpu incore superblock counters
1879  *
1880  * Simple concept, difficult implementation
1881  *
1882  * Basically, replace the incore superblock counters with a distributed per cpu
1883  * counter for contended fields (e.g.  free block count).
1884  *
1885  * Difficulties arise in that the incore sb is used for ENOSPC checking, and
1886  * hence needs to be accurately read when we are running low on space. Hence
1887  * there is a method to enable and disable the per-cpu counters based on how
1888  * much "stuff" is available in them.
1889  *
1890  * Basically, a counter is enabled if there is enough free resource to justify
1891  * running a per-cpu fast-path. If the per-cpu counter runs out (i.e. a local
1892  * ENOSPC), then we disable the counters to synchronise all callers and
1893  * re-distribute the available resources.
1894  *
1895  * If, once we redistributed the available resources, we still get a failure,
1896  * we disable the per-cpu counter and go through the slow path.
1897  *
1898  * The slow path is the current xfs_mod_incore_sb() function.  This means that
1899  * when we disable a per-cpu counter, we need to drain it's resources back to
1900  * the global superblock. We do this after disabling the counter to prevent
1901  * more threads from queueing up on the counter.
1902  *
1903  * Essentially, this means that we still need a lock in the fast path to enable
1904  * synchronisation between the global counters and the per-cpu counters. This
1905  * is not a problem because the lock will be local to a CPU almost all the time
1906  * and have little contention except when we get to ENOSPC conditions.
1907  *
1908  * Basically, this lock becomes a barrier that enables us to lock out the fast
1909  * path while we do things like enabling and disabling counters and
1910  * synchronising the counters.
1911  *
1912  * Locking rules:
1913  *
1914  *      1. m_sb_lock before picking up per-cpu locks
1915  *      2. per-cpu locks always picked up via for_each_online_cpu() order
1916  *      3. accurate counter sync requires m_sb_lock + per cpu locks
1917  *      4. modifying per-cpu counters requires holding per-cpu lock
1918  *      5. modifying global counters requires holding m_sb_lock
1919  *      6. enabling or disabling a counter requires holding the m_sb_lock 
1920  *         and _none_ of the per-cpu locks.
1921  *
1922  * Disabled counters are only ever re-enabled by a balance operation
1923  * that results in more free resources per CPU than a given threshold.
1924  * To ensure counters don't remain disabled, they are rebalanced when
1925  * the global resource goes above a higher threshold (i.e. some hysteresis
1926  * is present to prevent thrashing).
1927  */
1928
1929 #ifdef CONFIG_HOTPLUG_CPU
1930 /*
1931  * hot-plug CPU notifier support.
1932  *
1933  * We need a notifier per filesystem as we need to be able to identify
1934  * the filesystem to balance the counters out. This is achieved by
1935  * having a notifier block embedded in the xfs_mount_t and doing pointer
1936  * magic to get the mount pointer from the notifier block address.
1937  */
1938 STATIC int
1939 xfs_icsb_cpu_notify(
1940         struct notifier_block *nfb,
1941         unsigned long action,
1942         void *hcpu)
1943 {
1944         xfs_icsb_cnts_t *cntp;
1945         xfs_mount_t     *mp;
1946
1947         mp = (xfs_mount_t *)container_of(nfb, xfs_mount_t, m_icsb_notifier);
1948         cntp = (xfs_icsb_cnts_t *)
1949                         per_cpu_ptr(mp->m_sb_cnts, (unsigned long)hcpu);
1950         switch (action) {
1951         case CPU_UP_PREPARE:
1952         case CPU_UP_PREPARE_FROZEN:
1953                 /* Easy Case - initialize the area and locks, and
1954                  * then rebalance when online does everything else for us. */
1955                 memset(cntp, 0, sizeof(xfs_icsb_cnts_t));
1956                 break;
1957         case CPU_ONLINE:
1958         case CPU_ONLINE_FROZEN:
1959                 xfs_icsb_lock(mp);
1960                 xfs_icsb_balance_counter(mp, XFS_SBS_ICOUNT, 0);
1961                 xfs_icsb_balance_counter(mp, XFS_SBS_IFREE, 0);
1962                 xfs_icsb_balance_counter(mp, XFS_SBS_FDBLOCKS, 0);
1963                 xfs_icsb_unlock(mp);
1964                 break;
1965         case CPU_DEAD:
1966         case CPU_DEAD_FROZEN:
1967                 /* Disable all the counters, then fold the dead cpu's
1968                  * count into the total on the global superblock and
1969                  * re-enable the counters. */
1970                 xfs_icsb_lock(mp);
1971                 spin_lock(&mp->m_sb_lock);
1972                 xfs_icsb_disable_counter(mp, XFS_SBS_ICOUNT);
1973                 xfs_icsb_disable_counter(mp, XFS_SBS_IFREE);
1974                 xfs_icsb_disable_counter(mp, XFS_SBS_FDBLOCKS);
1975
1976                 mp->m_sb.sb_icount += cntp->icsb_icount;
1977                 mp->m_sb.sb_ifree += cntp->icsb_ifree;
1978                 mp->m_sb.sb_fdblocks += cntp->icsb_fdblocks;
1979
1980                 memset(cntp, 0, sizeof(xfs_icsb_cnts_t));
1981
1982                 xfs_icsb_balance_counter_locked(mp, XFS_SBS_ICOUNT, 0);
1983                 xfs_icsb_balance_counter_locked(mp, XFS_SBS_IFREE, 0);
1984                 xfs_icsb_balance_counter_locked(mp, XFS_SBS_FDBLOCKS, 0);
1985                 spin_unlock(&mp->m_sb_lock);
1986                 xfs_icsb_unlock(mp);
1987                 break;
1988         }
1989
1990         return NOTIFY_OK;
1991 }
1992 #endif /* CONFIG_HOTPLUG_CPU */
1993
1994 int
1995 xfs_icsb_init_counters(
1996         xfs_mount_t     *mp)
1997 {
1998         xfs_icsb_cnts_t *cntp;
1999         int             i;
2000
2001         mp->m_sb_cnts = alloc_percpu(xfs_icsb_cnts_t);
2002         if (mp->m_sb_cnts == NULL)
2003                 return -ENOMEM;
2004
2005 #ifdef CONFIG_HOTPLUG_CPU
2006         mp->m_icsb_notifier.notifier_call = xfs_icsb_cpu_notify;
2007         mp->m_icsb_notifier.priority = 0;
2008         register_hotcpu_notifier(&mp->m_icsb_notifier);
2009 #endif /* CONFIG_HOTPLUG_CPU */
2010
2011         for_each_online_cpu(i) {
2012                 cntp = (xfs_icsb_cnts_t *)per_cpu_ptr(mp->m_sb_cnts, i);
2013                 memset(cntp, 0, sizeof(xfs_icsb_cnts_t));
2014         }
2015
2016         mutex_init(&mp->m_icsb_mutex);
2017
2018         /*
2019          * start with all counters disabled so that the
2020          * initial balance kicks us off correctly
2021          */
2022         mp->m_icsb_counters = -1;
2023         return 0;
2024 }
2025
2026 void
2027 xfs_icsb_reinit_counters(
2028         xfs_mount_t     *mp)
2029 {
2030         xfs_icsb_lock(mp);
2031         /*
2032          * start with all counters disabled so that the
2033          * initial balance kicks us off correctly
2034          */
2035         mp->m_icsb_counters = -1;
2036         xfs_icsb_balance_counter(mp, XFS_SBS_ICOUNT, 0);
2037         xfs_icsb_balance_counter(mp, XFS_SBS_IFREE, 0);
2038         xfs_icsb_balance_counter(mp, XFS_SBS_FDBLOCKS, 0);
2039         xfs_icsb_unlock(mp);
2040 }
2041
2042 void
2043 xfs_icsb_destroy_counters(
2044         xfs_mount_t     *mp)
2045 {
2046         if (mp->m_sb_cnts) {
2047                 unregister_hotcpu_notifier(&mp->m_icsb_notifier);
2048                 free_percpu(mp->m_sb_cnts);
2049         }
2050         mutex_destroy(&mp->m_icsb_mutex);
2051 }
2052
2053 STATIC_INLINE void
2054 xfs_icsb_lock_cntr(
2055         xfs_icsb_cnts_t *icsbp)
2056 {
2057         while (test_and_set_bit(XFS_ICSB_FLAG_LOCK, &icsbp->icsb_flags)) {
2058                 ndelay(1000);
2059         }
2060 }
2061
2062 STATIC_INLINE void
2063 xfs_icsb_unlock_cntr(
2064         xfs_icsb_cnts_t *icsbp)
2065 {
2066         clear_bit(XFS_ICSB_FLAG_LOCK, &icsbp->icsb_flags);
2067 }
2068
2069
2070 STATIC_INLINE void
2071 xfs_icsb_lock_all_counters(
2072         xfs_mount_t     *mp)
2073 {
2074         xfs_icsb_cnts_t *cntp;
2075         int             i;
2076
2077         for_each_online_cpu(i) {
2078                 cntp = (xfs_icsb_cnts_t *)per_cpu_ptr(mp->m_sb_cnts, i);
2079                 xfs_icsb_lock_cntr(cntp);
2080         }
2081 }
2082
2083 STATIC_INLINE void
2084 xfs_icsb_unlock_all_counters(
2085         xfs_mount_t     *mp)
2086 {
2087         xfs_icsb_cnts_t *cntp;
2088         int             i;
2089
2090         for_each_online_cpu(i) {
2091                 cntp = (xfs_icsb_cnts_t *)per_cpu_ptr(mp->m_sb_cnts, i);
2092                 xfs_icsb_unlock_cntr(cntp);
2093         }
2094 }
2095
2096 STATIC void
2097 xfs_icsb_count(
2098         xfs_mount_t     *mp,
2099         xfs_icsb_cnts_t *cnt,
2100         int             flags)
2101 {
2102         xfs_icsb_cnts_t *cntp;
2103         int             i;
2104
2105         memset(cnt, 0, sizeof(xfs_icsb_cnts_t));
2106
2107         if (!(flags & XFS_ICSB_LAZY_COUNT))
2108                 xfs_icsb_lock_all_counters(mp);
2109
2110         for_each_online_cpu(i) {
2111                 cntp = (xfs_icsb_cnts_t *)per_cpu_ptr(mp->m_sb_cnts, i);
2112                 cnt->icsb_icount += cntp->icsb_icount;
2113                 cnt->icsb_ifree += cntp->icsb_ifree;
2114                 cnt->icsb_fdblocks += cntp->icsb_fdblocks;
2115         }
2116
2117         if (!(flags & XFS_ICSB_LAZY_COUNT))
2118                 xfs_icsb_unlock_all_counters(mp);
2119 }
2120
2121 STATIC int
2122 xfs_icsb_counter_disabled(
2123         xfs_mount_t     *mp,
2124         xfs_sb_field_t  field)
2125 {
2126         ASSERT((field >= XFS_SBS_ICOUNT) && (field <= XFS_SBS_FDBLOCKS));
2127         return test_bit(field, &mp->m_icsb_counters);
2128 }
2129
2130 STATIC void
2131 xfs_icsb_disable_counter(
2132         xfs_mount_t     *mp,
2133         xfs_sb_field_t  field)
2134 {
2135         xfs_icsb_cnts_t cnt;
2136
2137         ASSERT((field >= XFS_SBS_ICOUNT) && (field <= XFS_SBS_FDBLOCKS));
2138
2139         /*
2140          * If we are already disabled, then there is nothing to do
2141          * here. We check before locking all the counters to avoid
2142          * the expensive lock operation when being called in the
2143          * slow path and the counter is already disabled. This is
2144          * safe because the only time we set or clear this state is under
2145          * the m_icsb_mutex.
2146          */
2147         if (xfs_icsb_counter_disabled(mp, field))
2148                 return;
2149
2150         xfs_icsb_lock_all_counters(mp);
2151         if (!test_and_set_bit(field, &mp->m_icsb_counters)) {
2152                 /* drain back to superblock */
2153
2154                 xfs_icsb_count(mp, &cnt, XFS_ICSB_LAZY_COUNT);
2155                 switch(field) {
2156                 case XFS_SBS_ICOUNT:
2157                         mp->m_sb.sb_icount = cnt.icsb_icount;
2158                         break;
2159                 case XFS_SBS_IFREE:
2160                         mp->m_sb.sb_ifree = cnt.icsb_ifree;
2161                         break;
2162                 case XFS_SBS_FDBLOCKS:
2163                         mp->m_sb.sb_fdblocks = cnt.icsb_fdblocks;
2164                         break;
2165                 default:
2166                         BUG();
2167                 }
2168         }
2169
2170         xfs_icsb_unlock_all_counters(mp);
2171 }
2172
2173 STATIC void
2174 xfs_icsb_enable_counter(
2175         xfs_mount_t     *mp,
2176         xfs_sb_field_t  field,
2177         uint64_t        count,
2178         uint64_t        resid)
2179 {
2180         xfs_icsb_cnts_t *cntp;
2181         int             i;
2182
2183         ASSERT((field >= XFS_SBS_ICOUNT) && (field <= XFS_SBS_FDBLOCKS));
2184
2185         xfs_icsb_lock_all_counters(mp);
2186         for_each_online_cpu(i) {
2187                 cntp = per_cpu_ptr(mp->m_sb_cnts, i);
2188                 switch (field) {
2189                 case XFS_SBS_ICOUNT:
2190                         cntp->icsb_icount = count + resid;
2191                         break;
2192                 case XFS_SBS_IFREE:
2193                         cntp->icsb_ifree = count + resid;
2194                         break;
2195                 case XFS_SBS_FDBLOCKS:
2196                         cntp->icsb_fdblocks = count + resid;
2197                         break;
2198                 default:
2199                         BUG();
2200                         break;
2201                 }
2202                 resid = 0;
2203         }
2204         clear_bit(field, &mp->m_icsb_counters);
2205         xfs_icsb_unlock_all_counters(mp);
2206 }
2207
2208 void
2209 xfs_icsb_sync_counters_locked(
2210         xfs_mount_t     *mp,
2211         int             flags)
2212 {
2213         xfs_icsb_cnts_t cnt;
2214
2215         xfs_icsb_count(mp, &cnt, flags);
2216
2217         if (!xfs_icsb_counter_disabled(mp, XFS_SBS_ICOUNT))
2218                 mp->m_sb.sb_icount = cnt.icsb_icount;
2219         if (!xfs_icsb_counter_disabled(mp, XFS_SBS_IFREE))
2220                 mp->m_sb.sb_ifree = cnt.icsb_ifree;
2221         if (!xfs_icsb_counter_disabled(mp, XFS_SBS_FDBLOCKS))
2222                 mp->m_sb.sb_fdblocks = cnt.icsb_fdblocks;
2223 }
2224
2225 /*
2226  * Accurate update of per-cpu counters to incore superblock
2227  */
2228 void
2229 xfs_icsb_sync_counters(
2230         xfs_mount_t     *mp,
2231         int             flags)
2232 {
2233         spin_lock(&mp->m_sb_lock);
2234         xfs_icsb_sync_counters_locked(mp, flags);
2235         spin_unlock(&mp->m_sb_lock);
2236 }
2237
2238 /*
2239  * Balance and enable/disable counters as necessary.
2240  *
2241  * Thresholds for re-enabling counters are somewhat magic.  inode counts are
2242  * chosen to be the same number as single on disk allocation chunk per CPU, and
2243  * free blocks is something far enough zero that we aren't going thrash when we
2244  * get near ENOSPC. We also need to supply a minimum we require per cpu to
2245  * prevent looping endlessly when xfs_alloc_space asks for more than will
2246  * be distributed to a single CPU but each CPU has enough blocks to be
2247  * reenabled.
2248  *
2249  * Note that we can be called when counters are already disabled.
2250  * xfs_icsb_disable_counter() optimises the counter locking in this case to
2251  * prevent locking every per-cpu counter needlessly.
2252  */
2253
2254 #define XFS_ICSB_INO_CNTR_REENABLE      (uint64_t)64
2255 #define XFS_ICSB_FDBLK_CNTR_REENABLE(mp) \
2256                 (uint64_t)(512 + XFS_ALLOC_SET_ASIDE(mp))
2257 STATIC void
2258 xfs_icsb_balance_counter_locked(
2259         xfs_mount_t     *mp,
2260         xfs_sb_field_t  field,
2261         int             min_per_cpu)
2262 {
2263         uint64_t        count, resid;
2264         int             weight = num_online_cpus();
2265         uint64_t        min = (uint64_t)min_per_cpu;
2266
2267         /* disable counter and sync counter */
2268         xfs_icsb_disable_counter(mp, field);
2269
2270         /* update counters  - first CPU gets residual*/
2271         switch (field) {
2272         case XFS_SBS_ICOUNT:
2273                 count = mp->m_sb.sb_icount;
2274                 resid = do_div(count, weight);
2275                 if (count < max(min, XFS_ICSB_INO_CNTR_REENABLE))
2276                         return;
2277                 break;
2278         case XFS_SBS_IFREE:
2279                 count = mp->m_sb.sb_ifree;
2280                 resid = do_div(count, weight);
2281                 if (count < max(min, XFS_ICSB_INO_CNTR_REENABLE))
2282                         return;
2283                 break;
2284         case XFS_SBS_FDBLOCKS:
2285                 count = mp->m_sb.sb_fdblocks;
2286                 resid = do_div(count, weight);
2287                 if (count < max(min, XFS_ICSB_FDBLK_CNTR_REENABLE(mp)))
2288                         return;
2289                 break;
2290         default:
2291                 BUG();
2292                 count = resid = 0;      /* quiet, gcc */
2293                 break;
2294         }
2295
2296         xfs_icsb_enable_counter(mp, field, count, resid);
2297 }
2298
2299 STATIC void
2300 xfs_icsb_balance_counter(
2301         xfs_mount_t     *mp,
2302         xfs_sb_field_t  fields,
2303         int             min_per_cpu)
2304 {
2305         spin_lock(&mp->m_sb_lock);
2306         xfs_icsb_balance_counter_locked(mp, fields, min_per_cpu);
2307         spin_unlock(&mp->m_sb_lock);
2308 }
2309
2310 STATIC int
2311 xfs_icsb_modify_counters(
2312         xfs_mount_t     *mp,
2313         xfs_sb_field_t  field,
2314         int64_t         delta,
2315         int             rsvd)
2316 {
2317         xfs_icsb_cnts_t *icsbp;
2318         long long       lcounter;       /* long counter for 64 bit fields */
2319         int             cpu, ret = 0;
2320
2321         might_sleep();
2322 again:
2323         cpu = get_cpu();
2324         icsbp = (xfs_icsb_cnts_t *)per_cpu_ptr(mp->m_sb_cnts, cpu);
2325
2326         /*
2327          * if the counter is disabled, go to slow path
2328          */
2329         if (unlikely(xfs_icsb_counter_disabled(mp, field)))
2330                 goto slow_path;
2331         xfs_icsb_lock_cntr(icsbp);
2332         if (unlikely(xfs_icsb_counter_disabled(mp, field))) {
2333                 xfs_icsb_unlock_cntr(icsbp);
2334                 goto slow_path;
2335         }
2336
2337         switch (field) {
2338         case XFS_SBS_ICOUNT:
2339                 lcounter = icsbp->icsb_icount;
2340                 lcounter += delta;
2341                 if (unlikely(lcounter < 0))
2342                         goto balance_counter;
2343                 icsbp->icsb_icount = lcounter;
2344                 break;
2345
2346         case XFS_SBS_IFREE:
2347                 lcounter = icsbp->icsb_ifree;
2348                 lcounter += delta;
2349                 if (unlikely(lcounter < 0))
2350                         goto balance_counter;
2351                 icsbp->icsb_ifree = lcounter;
2352                 break;
2353
2354         case XFS_SBS_FDBLOCKS:
2355                 BUG_ON((mp->m_resblks - mp->m_resblks_avail) != 0);
2356
2357                 lcounter = icsbp->icsb_fdblocks - XFS_ALLOC_SET_ASIDE(mp);
2358                 lcounter += delta;
2359                 if (unlikely(lcounter < 0))
2360                         goto balance_counter;
2361                 icsbp->icsb_fdblocks = lcounter + XFS_ALLOC_SET_ASIDE(mp);
2362                 break;
2363         default:
2364                 BUG();
2365                 break;
2366         }
2367         xfs_icsb_unlock_cntr(icsbp);
2368         put_cpu();
2369         return 0;
2370
2371 slow_path:
2372         put_cpu();
2373
2374         /*
2375          * serialise with a mutex so we don't burn lots of cpu on
2376          * the superblock lock. We still need to hold the superblock
2377          * lock, however, when we modify the global structures.
2378          */
2379         xfs_icsb_lock(mp);
2380
2381         /*
2382          * Now running atomically.
2383          *
2384          * If the counter is enabled, someone has beaten us to rebalancing.
2385          * Drop the lock and try again in the fast path....
2386          */
2387         if (!(xfs_icsb_counter_disabled(mp, field))) {
2388                 xfs_icsb_unlock(mp);
2389                 goto again;
2390         }
2391
2392         /*
2393          * The counter is currently disabled. Because we are
2394          * running atomically here, we know a rebalance cannot
2395          * be in progress. Hence we can go straight to operating
2396          * on the global superblock. We do not call xfs_mod_incore_sb()
2397          * here even though we need to get the m_sb_lock. Doing so
2398          * will cause us to re-enter this function and deadlock.
2399          * Hence we get the m_sb_lock ourselves and then call
2400          * xfs_mod_incore_sb_unlocked() as the unlocked path operates
2401          * directly on the global counters.
2402          */
2403         spin_lock(&mp->m_sb_lock);
2404         ret = xfs_mod_incore_sb_unlocked(mp, field, delta, rsvd);
2405         spin_unlock(&mp->m_sb_lock);
2406
2407         /*
2408          * Now that we've modified the global superblock, we
2409          * may be able to re-enable the distributed counters
2410          * (e.g. lots of space just got freed). After that
2411          * we are done.
2412          */
2413         if (ret != ENOSPC)
2414                 xfs_icsb_balance_counter(mp, field, 0);
2415         xfs_icsb_unlock(mp);
2416         return ret;
2417
2418 balance_counter:
2419         xfs_icsb_unlock_cntr(icsbp);
2420         put_cpu();
2421
2422         /*
2423          * We may have multiple threads here if multiple per-cpu
2424          * counters run dry at the same time. This will mean we can
2425          * do more balances than strictly necessary but it is not
2426          * the common slowpath case.
2427          */
2428         xfs_icsb_lock(mp);
2429
2430         /*
2431          * running atomically.
2432          *
2433          * This will leave the counter in the correct state for future
2434          * accesses. After the rebalance, we simply try again and our retry
2435          * will either succeed through the fast path or slow path without
2436          * another balance operation being required.
2437          */
2438         xfs_icsb_balance_counter(mp, field, delta);
2439         xfs_icsb_unlock(mp);
2440         goto again;
2441 }
2442
2443 #endif