[XFS] Fix superblock features2 field alignment problem
[sfrench/cifs-2.6.git] / 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
47 STATIC void     xfs_mount_log_sb(xfs_mount_t *, __int64_t);
48 STATIC int      xfs_uuid_mount(xfs_mount_t *);
49 STATIC void     xfs_uuid_unmount(xfs_mount_t *mp);
50 STATIC void     xfs_unmountfs_wait(xfs_mount_t *);
51
52
53 #ifdef HAVE_PERCPU_SB
54 STATIC void     xfs_icsb_destroy_counters(xfs_mount_t *);
55 STATIC void     xfs_icsb_balance_counter(xfs_mount_t *, xfs_sb_field_t,
56                                                 int, int);
57 STATIC void     xfs_icsb_sync_counters(xfs_mount_t *);
58 STATIC int      xfs_icsb_modify_counters(xfs_mount_t *, xfs_sb_field_t,
59                                                 int64_t, int);
60 STATIC int      xfs_icsb_disable_counter(xfs_mount_t *, xfs_sb_field_t);
61
62 #else
63
64 #define xfs_icsb_destroy_counters(mp)                   do { } while (0)
65 #define xfs_icsb_balance_counter(mp, a, b, c)           do { } while (0)
66 #define xfs_icsb_sync_counters(mp)                      do { } while (0)
67 #define xfs_icsb_modify_counters(mp, a, b, c)           do { } while (0)
68
69 #endif
70
71 static const struct {
72         short offset;
73         short type;     /* 0 = integer
74                          * 1 = binary / string (no translation)
75                          */
76 } xfs_sb_info[] = {
77     { offsetof(xfs_sb_t, sb_magicnum),   0 },
78     { offsetof(xfs_sb_t, sb_blocksize),  0 },
79     { offsetof(xfs_sb_t, sb_dblocks),    0 },
80     { offsetof(xfs_sb_t, sb_rblocks),    0 },
81     { offsetof(xfs_sb_t, sb_rextents),   0 },
82     { offsetof(xfs_sb_t, sb_uuid),       1 },
83     { offsetof(xfs_sb_t, sb_logstart),   0 },
84     { offsetof(xfs_sb_t, sb_rootino),    0 },
85     { offsetof(xfs_sb_t, sb_rbmino),     0 },
86     { offsetof(xfs_sb_t, sb_rsumino),    0 },
87     { offsetof(xfs_sb_t, sb_rextsize),   0 },
88     { offsetof(xfs_sb_t, sb_agblocks),   0 },
89     { offsetof(xfs_sb_t, sb_agcount),    0 },
90     { offsetof(xfs_sb_t, sb_rbmblocks),  0 },
91     { offsetof(xfs_sb_t, sb_logblocks),  0 },
92     { offsetof(xfs_sb_t, sb_versionnum), 0 },
93     { offsetof(xfs_sb_t, sb_sectsize),   0 },
94     { offsetof(xfs_sb_t, sb_inodesize),  0 },
95     { offsetof(xfs_sb_t, sb_inopblock),  0 },
96     { offsetof(xfs_sb_t, sb_fname[0]),   1 },
97     { offsetof(xfs_sb_t, sb_blocklog),   0 },
98     { offsetof(xfs_sb_t, sb_sectlog),    0 },
99     { offsetof(xfs_sb_t, sb_inodelog),   0 },
100     { offsetof(xfs_sb_t, sb_inopblog),   0 },
101     { offsetof(xfs_sb_t, sb_agblklog),   0 },
102     { offsetof(xfs_sb_t, sb_rextslog),   0 },
103     { offsetof(xfs_sb_t, sb_inprogress), 0 },
104     { offsetof(xfs_sb_t, sb_imax_pct),   0 },
105     { offsetof(xfs_sb_t, sb_icount),     0 },
106     { offsetof(xfs_sb_t, sb_ifree),      0 },
107     { offsetof(xfs_sb_t, sb_fdblocks),   0 },
108     { offsetof(xfs_sb_t, sb_frextents),  0 },
109     { offsetof(xfs_sb_t, sb_uquotino),   0 },
110     { offsetof(xfs_sb_t, sb_gquotino),   0 },
111     { offsetof(xfs_sb_t, sb_qflags),     0 },
112     { offsetof(xfs_sb_t, sb_flags),      0 },
113     { offsetof(xfs_sb_t, sb_shared_vn),  0 },
114     { offsetof(xfs_sb_t, sb_inoalignmt), 0 },
115     { offsetof(xfs_sb_t, sb_unit),       0 },
116     { offsetof(xfs_sb_t, sb_width),      0 },
117     { offsetof(xfs_sb_t, sb_dirblklog),  0 },
118     { offsetof(xfs_sb_t, sb_logsectlog), 0 },
119     { offsetof(xfs_sb_t, sb_logsectsize),0 },
120     { offsetof(xfs_sb_t, sb_logsunit),   0 },
121     { offsetof(xfs_sb_t, sb_features2),  0 },
122     { offsetof(xfs_sb_t, sb_bad_features2), 0 },
123     { sizeof(xfs_sb_t),                  0 }
124 };
125
126 /*
127  * Return a pointer to an initialized xfs_mount structure.
128  */
129 xfs_mount_t *
130 xfs_mount_init(void)
131 {
132         xfs_mount_t *mp;
133
134         mp = kmem_zalloc(sizeof(xfs_mount_t), KM_SLEEP);
135
136         if (xfs_icsb_init_counters(mp)) {
137                 mp->m_flags |= XFS_MOUNT_NO_PERCPU_SB;
138         }
139
140         spin_lock_init(&mp->m_sb_lock);
141         mutex_init(&mp->m_ilock);
142         mutex_init(&mp->m_growlock);
143         atomic_set(&mp->m_active_trans, 0);
144
145         return mp;
146 }
147
148 /*
149  * Free up the resources associated with a mount structure.  Assume that
150  * the structure was initially zeroed, so we can tell which fields got
151  * initialized.
152  */
153 void
154 xfs_mount_free(
155         xfs_mount_t     *mp)
156 {
157         if (mp->m_perag) {
158                 int     agno;
159
160                 for (agno = 0; agno < mp->m_maxagi; agno++)
161                         if (mp->m_perag[agno].pagb_list)
162                                 kmem_free(mp->m_perag[agno].pagb_list,
163                                                 sizeof(xfs_perag_busy_t) *
164                                                         XFS_PAGB_NUM_SLOTS);
165                 kmem_free(mp->m_perag,
166                           sizeof(xfs_perag_t) * mp->m_sb.sb_agcount);
167         }
168
169         spinlock_destroy(&mp->m_ail_lock);
170         spinlock_destroy(&mp->m_sb_lock);
171         mutex_destroy(&mp->m_ilock);
172         mutex_destroy(&mp->m_growlock);
173         if (mp->m_quotainfo)
174                 XFS_QM_DONE(mp);
175
176         if (mp->m_fsname != NULL)
177                 kmem_free(mp->m_fsname, mp->m_fsname_len);
178         if (mp->m_rtname != NULL)
179                 kmem_free(mp->m_rtname, strlen(mp->m_rtname) + 1);
180         if (mp->m_logname != NULL)
181                 kmem_free(mp->m_logname, strlen(mp->m_logname) + 1);
182
183         xfs_icsb_destroy_counters(mp);
184 }
185
186 /*
187  * Check size of device based on the (data/realtime) block count.
188  * Note: this check is used by the growfs code as well as mount.
189  */
190 int
191 xfs_sb_validate_fsb_count(
192         xfs_sb_t        *sbp,
193         __uint64_t      nblocks)
194 {
195         ASSERT(PAGE_SHIFT >= sbp->sb_blocklog);
196         ASSERT(sbp->sb_blocklog >= BBSHIFT);
197
198 #if XFS_BIG_BLKNOS     /* Limited by ULONG_MAX of page cache index */
199         if (nblocks >> (PAGE_CACHE_SHIFT - sbp->sb_blocklog) > ULONG_MAX)
200                 return E2BIG;
201 #else                  /* Limited by UINT_MAX of sectors */
202         if (nblocks << (sbp->sb_blocklog - BBSHIFT) > UINT_MAX)
203                 return E2BIG;
204 #endif
205         return 0;
206 }
207
208 /*
209  * Check the validity of the SB found.
210  */
211 STATIC int
212 xfs_mount_validate_sb(
213         xfs_mount_t     *mp,
214         xfs_sb_t        *sbp,
215         int             flags)
216 {
217         /*
218          * If the log device and data device have the
219          * same device number, the log is internal.
220          * Consequently, the sb_logstart should be non-zero.  If
221          * we have a zero sb_logstart in this case, we may be trying to mount
222          * a volume filesystem in a non-volume manner.
223          */
224         if (sbp->sb_magicnum != XFS_SB_MAGIC) {
225                 xfs_fs_mount_cmn_err(flags, "bad magic number");
226                 return XFS_ERROR(EWRONGFS);
227         }
228
229         if (!xfs_sb_good_version(sbp)) {
230                 xfs_fs_mount_cmn_err(flags, "bad version");
231                 return XFS_ERROR(EWRONGFS);
232         }
233
234         if (unlikely(
235             sbp->sb_logstart == 0 && mp->m_logdev_targp == mp->m_ddev_targp)) {
236                 xfs_fs_mount_cmn_err(flags,
237                         "filesystem is marked as having an external log; "
238                         "specify logdev on the\nmount command line.");
239                 return XFS_ERROR(EINVAL);
240         }
241
242         if (unlikely(
243             sbp->sb_logstart != 0 && mp->m_logdev_targp != mp->m_ddev_targp)) {
244                 xfs_fs_mount_cmn_err(flags,
245                         "filesystem is marked as having an internal log; "
246                         "do not specify logdev on\nthe mount command line.");
247                 return XFS_ERROR(EINVAL);
248         }
249
250         /*
251          * More sanity checking. These were stolen directly from
252          * xfs_repair.
253          */
254         if (unlikely(
255             sbp->sb_agcount <= 0                                        ||
256             sbp->sb_sectsize < XFS_MIN_SECTORSIZE                       ||
257             sbp->sb_sectsize > XFS_MAX_SECTORSIZE                       ||
258             sbp->sb_sectlog < XFS_MIN_SECTORSIZE_LOG                    ||
259             sbp->sb_sectlog > XFS_MAX_SECTORSIZE_LOG                    ||
260             sbp->sb_blocksize < XFS_MIN_BLOCKSIZE                       ||
261             sbp->sb_blocksize > XFS_MAX_BLOCKSIZE                       ||
262             sbp->sb_blocklog < XFS_MIN_BLOCKSIZE_LOG                    ||
263             sbp->sb_blocklog > XFS_MAX_BLOCKSIZE_LOG                    ||
264             sbp->sb_inodesize < XFS_DINODE_MIN_SIZE                     ||
265             sbp->sb_inodesize > XFS_DINODE_MAX_SIZE                     ||
266             sbp->sb_inodelog < XFS_DINODE_MIN_LOG                       ||
267             sbp->sb_inodelog > XFS_DINODE_MAX_LOG                       ||
268             (sbp->sb_blocklog - sbp->sb_inodelog != sbp->sb_inopblog)   ||
269             (sbp->sb_rextsize * sbp->sb_blocksize > XFS_MAX_RTEXTSIZE)  ||
270             (sbp->sb_rextsize * sbp->sb_blocksize < XFS_MIN_RTEXTSIZE)  ||
271             (sbp->sb_imax_pct > 100 /* zero sb_imax_pct is valid */))) {
272                 xfs_fs_mount_cmn_err(flags, "SB sanity check 1 failed");
273                 return XFS_ERROR(EFSCORRUPTED);
274         }
275
276         /*
277          * Sanity check AG count, size fields against data size field
278          */
279         if (unlikely(
280             sbp->sb_dblocks == 0 ||
281             sbp->sb_dblocks >
282              (xfs_drfsbno_t)sbp->sb_agcount * sbp->sb_agblocks ||
283             sbp->sb_dblocks < (xfs_drfsbno_t)(sbp->sb_agcount - 1) *
284                               sbp->sb_agblocks + XFS_MIN_AG_BLOCKS)) {
285                 xfs_fs_mount_cmn_err(flags, "SB sanity check 2 failed");
286                 return XFS_ERROR(EFSCORRUPTED);
287         }
288
289         if (xfs_sb_validate_fsb_count(sbp, sbp->sb_dblocks) ||
290             xfs_sb_validate_fsb_count(sbp, sbp->sb_rblocks)) {
291                 xfs_fs_mount_cmn_err(flags,
292                         "file system too large to be mounted on this system.");
293                 return XFS_ERROR(E2BIG);
294         }
295
296         if (unlikely(sbp->sb_inprogress)) {
297                 xfs_fs_mount_cmn_err(flags, "file system busy");
298                 return XFS_ERROR(EFSCORRUPTED);
299         }
300
301         /*
302          * Version 1 directory format has never worked on Linux.
303          */
304         if (unlikely(!xfs_sb_version_hasdirv2(sbp))) {
305                 xfs_fs_mount_cmn_err(flags,
306                         "file system using version 1 directory format");
307                 return XFS_ERROR(ENOSYS);
308         }
309
310         /*
311          * Until this is fixed only page-sized or smaller data blocks work.
312          */
313         if (unlikely(sbp->sb_blocksize > PAGE_SIZE)) {
314                 xfs_fs_mount_cmn_err(flags,
315                         "file system with blocksize %d bytes",
316                         sbp->sb_blocksize);
317                 xfs_fs_mount_cmn_err(flags,
318                         "only pagesize (%ld) or less will currently work.",
319                         PAGE_SIZE);
320                 return XFS_ERROR(ENOSYS);
321         }
322
323         return 0;
324 }
325
326 STATIC void
327 xfs_initialize_perag_icache(
328         xfs_perag_t     *pag)
329 {
330         if (!pag->pag_ici_init) {
331                 rwlock_init(&pag->pag_ici_lock);
332                 INIT_RADIX_TREE(&pag->pag_ici_root, GFP_ATOMIC);
333                 pag->pag_ici_init = 1;
334         }
335 }
336
337 xfs_agnumber_t
338 xfs_initialize_perag(
339         xfs_mount_t     *mp,
340         xfs_agnumber_t  agcount)
341 {
342         xfs_agnumber_t  index, max_metadata;
343         xfs_perag_t     *pag;
344         xfs_agino_t     agino;
345         xfs_ino_t       ino;
346         xfs_sb_t        *sbp = &mp->m_sb;
347         xfs_ino_t       max_inum = XFS_MAXINUMBER_32;
348
349         /* Check to see if the filesystem can overflow 32 bit inodes */
350         agino = XFS_OFFBNO_TO_AGINO(mp, sbp->sb_agblocks - 1, 0);
351         ino = XFS_AGINO_TO_INO(mp, agcount - 1, agino);
352
353         /* Clear the mount flag if no inode can overflow 32 bits
354          * on this filesystem, or if specifically requested..
355          */
356         if ((mp->m_flags & XFS_MOUNT_SMALL_INUMS) && ino > max_inum) {
357                 mp->m_flags |= XFS_MOUNT_32BITINODES;
358         } else {
359                 mp->m_flags &= ~XFS_MOUNT_32BITINODES;
360         }
361
362         /* If we can overflow then setup the ag headers accordingly */
363         if (mp->m_flags & XFS_MOUNT_32BITINODES) {
364                 /* Calculate how much should be reserved for inodes to
365                  * meet the max inode percentage.
366                  */
367                 if (mp->m_maxicount) {
368                         __uint64_t      icount;
369
370                         icount = sbp->sb_dblocks * sbp->sb_imax_pct;
371                         do_div(icount, 100);
372                         icount += sbp->sb_agblocks - 1;
373                         do_div(icount, sbp->sb_agblocks);
374                         max_metadata = icount;
375                 } else {
376                         max_metadata = agcount;
377                 }
378                 for (index = 0; index < agcount; index++) {
379                         ino = XFS_AGINO_TO_INO(mp, index, agino);
380                         if (ino > max_inum) {
381                                 index++;
382                                 break;
383                         }
384
385                         /* This ag is preferred for inodes */
386                         pag = &mp->m_perag[index];
387                         pag->pagi_inodeok = 1;
388                         if (index < max_metadata)
389                                 pag->pagf_metadata = 1;
390                         xfs_initialize_perag_icache(pag);
391                 }
392         } else {
393                 /* Setup default behavior for smaller filesystems */
394                 for (index = 0; index < agcount; index++) {
395                         pag = &mp->m_perag[index];
396                         pag->pagi_inodeok = 1;
397                         xfs_initialize_perag_icache(pag);
398                 }
399         }
400         return index;
401 }
402
403 void
404 xfs_sb_from_disk(
405         xfs_sb_t        *to,
406         xfs_dsb_t       *from)
407 {
408         to->sb_magicnum = be32_to_cpu(from->sb_magicnum);
409         to->sb_blocksize = be32_to_cpu(from->sb_blocksize);
410         to->sb_dblocks = be64_to_cpu(from->sb_dblocks);
411         to->sb_rblocks = be64_to_cpu(from->sb_rblocks);
412         to->sb_rextents = be64_to_cpu(from->sb_rextents);
413         memcpy(&to->sb_uuid, &from->sb_uuid, sizeof(to->sb_uuid));
414         to->sb_logstart = be64_to_cpu(from->sb_logstart);
415         to->sb_rootino = be64_to_cpu(from->sb_rootino);
416         to->sb_rbmino = be64_to_cpu(from->sb_rbmino);
417         to->sb_rsumino = be64_to_cpu(from->sb_rsumino);
418         to->sb_rextsize = be32_to_cpu(from->sb_rextsize);
419         to->sb_agblocks = be32_to_cpu(from->sb_agblocks);
420         to->sb_agcount = be32_to_cpu(from->sb_agcount);
421         to->sb_rbmblocks = be32_to_cpu(from->sb_rbmblocks);
422         to->sb_logblocks = be32_to_cpu(from->sb_logblocks);
423         to->sb_versionnum = be16_to_cpu(from->sb_versionnum);
424         to->sb_sectsize = be16_to_cpu(from->sb_sectsize);
425         to->sb_inodesize = be16_to_cpu(from->sb_inodesize);
426         to->sb_inopblock = be16_to_cpu(from->sb_inopblock);
427         memcpy(&to->sb_fname, &from->sb_fname, sizeof(to->sb_fname));
428         to->sb_blocklog = from->sb_blocklog;
429         to->sb_sectlog = from->sb_sectlog;
430         to->sb_inodelog = from->sb_inodelog;
431         to->sb_inopblog = from->sb_inopblog;
432         to->sb_agblklog = from->sb_agblklog;
433         to->sb_rextslog = from->sb_rextslog;
434         to->sb_inprogress = from->sb_inprogress;
435         to->sb_imax_pct = from->sb_imax_pct;
436         to->sb_icount = be64_to_cpu(from->sb_icount);
437         to->sb_ifree = be64_to_cpu(from->sb_ifree);
438         to->sb_fdblocks = be64_to_cpu(from->sb_fdblocks);
439         to->sb_frextents = be64_to_cpu(from->sb_frextents);
440         to->sb_uquotino = be64_to_cpu(from->sb_uquotino);
441         to->sb_gquotino = be64_to_cpu(from->sb_gquotino);
442         to->sb_qflags = be16_to_cpu(from->sb_qflags);
443         to->sb_flags = from->sb_flags;
444         to->sb_shared_vn = from->sb_shared_vn;
445         to->sb_inoalignmt = be32_to_cpu(from->sb_inoalignmt);
446         to->sb_unit = be32_to_cpu(from->sb_unit);
447         to->sb_width = be32_to_cpu(from->sb_width);
448         to->sb_dirblklog = from->sb_dirblklog;
449         to->sb_logsectlog = from->sb_logsectlog;
450         to->sb_logsectsize = be16_to_cpu(from->sb_logsectsize);
451         to->sb_logsunit = be32_to_cpu(from->sb_logsunit);
452         to->sb_features2 = be32_to_cpu(from->sb_features2);
453         to->sb_bad_features2 = be32_to_cpu(from->sb_bad_features2);
454 }
455
456 /*
457  * Copy in core superblock to ondisk one.
458  *
459  * The fields argument is mask of superblock fields to copy.
460  */
461 void
462 xfs_sb_to_disk(
463         xfs_dsb_t       *to,
464         xfs_sb_t        *from,
465         __int64_t       fields)
466 {
467         xfs_caddr_t     to_ptr = (xfs_caddr_t)to;
468         xfs_caddr_t     from_ptr = (xfs_caddr_t)from;
469         xfs_sb_field_t  f;
470         int             first;
471         int             size;
472
473         ASSERT(fields);
474         if (!fields)
475                 return;
476
477         while (fields) {
478                 f = (xfs_sb_field_t)xfs_lowbit64((__uint64_t)fields);
479                 first = xfs_sb_info[f].offset;
480                 size = xfs_sb_info[f + 1].offset - first;
481
482                 ASSERT(xfs_sb_info[f].type == 0 || xfs_sb_info[f].type == 1);
483
484                 if (size == 1 || xfs_sb_info[f].type == 1) {
485                         memcpy(to_ptr + first, from_ptr + first, size);
486                 } else {
487                         switch (size) {
488                         case 2:
489                                 *(__be16 *)(to_ptr + first) =
490                                         cpu_to_be16(*(__u16 *)(from_ptr + first));
491                                 break;
492                         case 4:
493                                 *(__be32 *)(to_ptr + first) =
494                                         cpu_to_be32(*(__u32 *)(from_ptr + first));
495                                 break;
496                         case 8:
497                                 *(__be64 *)(to_ptr + first) =
498                                         cpu_to_be64(*(__u64 *)(from_ptr + first));
499                                 break;
500                         default:
501                                 ASSERT(0);
502                         }
503                 }
504
505                 fields &= ~(1LL << f);
506         }
507 }
508
509 /*
510  * xfs_readsb
511  *
512  * Does the initial read of the superblock.
513  */
514 int
515 xfs_readsb(xfs_mount_t *mp, int flags)
516 {
517         unsigned int    sector_size;
518         unsigned int    extra_flags;
519         xfs_buf_t       *bp;
520         int             error;
521
522         ASSERT(mp->m_sb_bp == NULL);
523         ASSERT(mp->m_ddev_targp != NULL);
524
525         /*
526          * Allocate a (locked) buffer to hold the superblock.
527          * This will be kept around at all times to optimize
528          * access to the superblock.
529          */
530         sector_size = xfs_getsize_buftarg(mp->m_ddev_targp);
531         extra_flags = XFS_BUF_LOCK | XFS_BUF_MANAGE | XFS_BUF_MAPPED;
532
533         bp = xfs_buf_read_flags(mp->m_ddev_targp, XFS_SB_DADDR,
534                                 BTOBB(sector_size), extra_flags);
535         if (!bp || XFS_BUF_ISERROR(bp)) {
536                 xfs_fs_mount_cmn_err(flags, "SB read failed");
537                 error = bp ? XFS_BUF_GETERROR(bp) : ENOMEM;
538                 goto fail;
539         }
540         ASSERT(XFS_BUF_ISBUSY(bp));
541         ASSERT(XFS_BUF_VALUSEMA(bp) <= 0);
542
543         /*
544          * Initialize the mount structure from the superblock.
545          * But first do some basic consistency checking.
546          */
547         xfs_sb_from_disk(&mp->m_sb, XFS_BUF_TO_SBP(bp));
548
549         error = xfs_mount_validate_sb(mp, &(mp->m_sb), flags);
550         if (error) {
551                 xfs_fs_mount_cmn_err(flags, "SB validate failed");
552                 goto fail;
553         }
554
555         /*
556          * We must be able to do sector-sized and sector-aligned IO.
557          */
558         if (sector_size > mp->m_sb.sb_sectsize) {
559                 xfs_fs_mount_cmn_err(flags,
560                         "device supports only %u byte sectors (not %u)",
561                         sector_size, mp->m_sb.sb_sectsize);
562                 error = ENOSYS;
563                 goto fail;
564         }
565
566         /*
567          * If device sector size is smaller than the superblock size,
568          * re-read the superblock so the buffer is correctly sized.
569          */
570         if (sector_size < mp->m_sb.sb_sectsize) {
571                 XFS_BUF_UNMANAGE(bp);
572                 xfs_buf_relse(bp);
573                 sector_size = mp->m_sb.sb_sectsize;
574                 bp = xfs_buf_read_flags(mp->m_ddev_targp, XFS_SB_DADDR,
575                                         BTOBB(sector_size), extra_flags);
576                 if (!bp || XFS_BUF_ISERROR(bp)) {
577                         xfs_fs_mount_cmn_err(flags, "SB re-read failed");
578                         error = bp ? XFS_BUF_GETERROR(bp) : ENOMEM;
579                         goto fail;
580                 }
581                 ASSERT(XFS_BUF_ISBUSY(bp));
582                 ASSERT(XFS_BUF_VALUSEMA(bp) <= 0);
583         }
584
585         /* Initialize per-cpu counters */
586         xfs_icsb_reinit_counters(mp);
587
588         mp->m_sb_bp = bp;
589         xfs_buf_relse(bp);
590         ASSERT(XFS_BUF_VALUSEMA(bp) > 0);
591         return 0;
592
593  fail:
594         if (bp) {
595                 XFS_BUF_UNMANAGE(bp);
596                 xfs_buf_relse(bp);
597         }
598         return error;
599 }
600
601
602 /*
603  * xfs_mount_common
604  *
605  * Mount initialization code establishing various mount
606  * fields from the superblock associated with the given
607  * mount structure
608  */
609 STATIC void
610 xfs_mount_common(xfs_mount_t *mp, xfs_sb_t *sbp)
611 {
612         int     i;
613
614         mp->m_agfrotor = mp->m_agirotor = 0;
615         spin_lock_init(&mp->m_agirotor_lock);
616         mp->m_maxagi = mp->m_sb.sb_agcount;
617         mp->m_blkbit_log = sbp->sb_blocklog + XFS_NBBYLOG;
618         mp->m_blkbb_log = sbp->sb_blocklog - BBSHIFT;
619         mp->m_sectbb_log = sbp->sb_sectlog - BBSHIFT;
620         mp->m_agno_log = xfs_highbit32(sbp->sb_agcount - 1) + 1;
621         mp->m_agino_log = sbp->sb_inopblog + sbp->sb_agblklog;
622         mp->m_litino = sbp->sb_inodesize -
623                 ((uint)sizeof(xfs_dinode_core_t) + (uint)sizeof(xfs_agino_t));
624         mp->m_blockmask = sbp->sb_blocksize - 1;
625         mp->m_blockwsize = sbp->sb_blocksize >> XFS_WORDLOG;
626         mp->m_blockwmask = mp->m_blockwsize - 1;
627         INIT_LIST_HEAD(&mp->m_del_inodes);
628
629         /*
630          * Setup for attributes, in case they get created.
631          * This value is for inodes getting attributes for the first time,
632          * the per-inode value is for old attribute values.
633          */
634         ASSERT(sbp->sb_inodesize >= 256 && sbp->sb_inodesize <= 2048);
635         switch (sbp->sb_inodesize) {
636         case 256:
637                 mp->m_attroffset = XFS_LITINO(mp) -
638                                    XFS_BMDR_SPACE_CALC(MINABTPTRS);
639                 break;
640         case 512:
641         case 1024:
642         case 2048:
643                 mp->m_attroffset = XFS_BMDR_SPACE_CALC(6 * MINABTPTRS);
644                 break;
645         default:
646                 ASSERT(0);
647         }
648         ASSERT(mp->m_attroffset < XFS_LITINO(mp));
649
650         for (i = 0; i < 2; i++) {
651                 mp->m_alloc_mxr[i] = XFS_BTREE_BLOCK_MAXRECS(sbp->sb_blocksize,
652                         xfs_alloc, i == 0);
653                 mp->m_alloc_mnr[i] = XFS_BTREE_BLOCK_MINRECS(sbp->sb_blocksize,
654                         xfs_alloc, i == 0);
655         }
656         for (i = 0; i < 2; i++) {
657                 mp->m_bmap_dmxr[i] = XFS_BTREE_BLOCK_MAXRECS(sbp->sb_blocksize,
658                         xfs_bmbt, i == 0);
659                 mp->m_bmap_dmnr[i] = XFS_BTREE_BLOCK_MINRECS(sbp->sb_blocksize,
660                         xfs_bmbt, i == 0);
661         }
662         for (i = 0; i < 2; i++) {
663                 mp->m_inobt_mxr[i] = XFS_BTREE_BLOCK_MAXRECS(sbp->sb_blocksize,
664                         xfs_inobt, i == 0);
665                 mp->m_inobt_mnr[i] = XFS_BTREE_BLOCK_MINRECS(sbp->sb_blocksize,
666                         xfs_inobt, i == 0);
667         }
668
669         mp->m_bsize = XFS_FSB_TO_BB(mp, 1);
670         mp->m_ialloc_inos = (int)MAX((__uint16_t)XFS_INODES_PER_CHUNK,
671                                         sbp->sb_inopblock);
672         mp->m_ialloc_blks = mp->m_ialloc_inos >> sbp->sb_inopblog;
673 }
674
675 /*
676  * xfs_initialize_perag_data
677  *
678  * Read in each per-ag structure so we can count up the number of
679  * allocated inodes, free inodes and used filesystem blocks as this
680  * information is no longer persistent in the superblock. Once we have
681  * this information, write it into the in-core superblock structure.
682  */
683 STATIC int
684 xfs_initialize_perag_data(xfs_mount_t *mp, xfs_agnumber_t agcount)
685 {
686         xfs_agnumber_t  index;
687         xfs_perag_t     *pag;
688         xfs_sb_t        *sbp = &mp->m_sb;
689         uint64_t        ifree = 0;
690         uint64_t        ialloc = 0;
691         uint64_t        bfree = 0;
692         uint64_t        bfreelst = 0;
693         uint64_t        btree = 0;
694         int             error;
695
696         for (index = 0; index < agcount; index++) {
697                 /*
698                  * read the agf, then the agi. This gets us
699                  * all the inforamtion we need and populates the
700                  * per-ag structures for us.
701                  */
702                 error = xfs_alloc_pagf_init(mp, NULL, index, 0);
703                 if (error)
704                         return error;
705
706                 error = xfs_ialloc_pagi_init(mp, NULL, index);
707                 if (error)
708                         return error;
709                 pag = &mp->m_perag[index];
710                 ifree += pag->pagi_freecount;
711                 ialloc += pag->pagi_count;
712                 bfree += pag->pagf_freeblks;
713                 bfreelst += pag->pagf_flcount;
714                 btree += pag->pagf_btreeblks;
715         }
716         /*
717          * Overwrite incore superblock counters with just-read data
718          */
719         spin_lock(&mp->m_sb_lock);
720         sbp->sb_ifree = ifree;
721         sbp->sb_icount = ialloc;
722         sbp->sb_fdblocks = bfree + bfreelst + btree;
723         spin_unlock(&mp->m_sb_lock);
724
725         /* Fixup the per-cpu counters as well. */
726         xfs_icsb_reinit_counters(mp);
727
728         return 0;
729 }
730
731 /*
732  * Update alignment values based on mount options and sb values
733  */
734 STATIC int
735 xfs_update_alignment(xfs_mount_t *mp, int mfsi_flags, __uint64_t *update_flags)
736 {
737         xfs_sb_t        *sbp = &(mp->m_sb);
738
739         if (mp->m_dalign && !(mfsi_flags & XFS_MFSI_SECOND)) {
740                 /*
741                  * If stripe unit and stripe width are not multiples
742                  * of the fs blocksize turn off alignment.
743                  */
744                 if ((BBTOB(mp->m_dalign) & mp->m_blockmask) ||
745                     (BBTOB(mp->m_swidth) & mp->m_blockmask)) {
746                         if (mp->m_flags & XFS_MOUNT_RETERR) {
747                                 cmn_err(CE_WARN,
748                                         "XFS: alignment check 1 failed");
749                                 return XFS_ERROR(EINVAL);
750                         }
751                         mp->m_dalign = mp->m_swidth = 0;
752                 } else {
753                         /*
754                          * Convert the stripe unit and width to FSBs.
755                          */
756                         mp->m_dalign = XFS_BB_TO_FSBT(mp, mp->m_dalign);
757                         if (mp->m_dalign && (sbp->sb_agblocks % mp->m_dalign)) {
758                                 if (mp->m_flags & XFS_MOUNT_RETERR) {
759                                         return XFS_ERROR(EINVAL);
760                                 }
761                                 xfs_fs_cmn_err(CE_WARN, mp,
762 "stripe alignment turned off: sunit(%d)/swidth(%d) incompatible with agsize(%d)",
763                                         mp->m_dalign, mp->m_swidth,
764                                         sbp->sb_agblocks);
765
766                                 mp->m_dalign = 0;
767                                 mp->m_swidth = 0;
768                         } else if (mp->m_dalign) {
769                                 mp->m_swidth = XFS_BB_TO_FSBT(mp, mp->m_swidth);
770                         } else {
771                                 if (mp->m_flags & XFS_MOUNT_RETERR) {
772                                         xfs_fs_cmn_err(CE_WARN, mp,
773 "stripe alignment turned off: sunit(%d) less than bsize(%d)",
774                                                 mp->m_dalign,
775                                                 mp->m_blockmask +1);
776                                         return XFS_ERROR(EINVAL);
777                                 }
778                                 mp->m_swidth = 0;
779                         }
780                 }
781
782                 /*
783                  * Update superblock with new values
784                  * and log changes
785                  */
786                 if (xfs_sb_version_hasdalign(sbp)) {
787                         if (sbp->sb_unit != mp->m_dalign) {
788                                 sbp->sb_unit = mp->m_dalign;
789                                 *update_flags |= XFS_SB_UNIT;
790                         }
791                         if (sbp->sb_width != mp->m_swidth) {
792                                 sbp->sb_width = mp->m_swidth;
793                                 *update_flags |= XFS_SB_WIDTH;
794                         }
795                 }
796         } else if ((mp->m_flags & XFS_MOUNT_NOALIGN) != XFS_MOUNT_NOALIGN &&
797                     xfs_sb_version_hasdalign(&mp->m_sb)) {
798                         mp->m_dalign = sbp->sb_unit;
799                         mp->m_swidth = sbp->sb_width;
800         }
801
802         return 0;
803 }
804
805 /*
806  * Set the maximum inode count for this filesystem
807  */
808 STATIC void
809 xfs_set_maxicount(xfs_mount_t *mp)
810 {
811         xfs_sb_t        *sbp = &(mp->m_sb);
812         __uint64_t      icount;
813
814         if (sbp->sb_imax_pct) {
815                 /*
816                  * Make sure the maximum inode count is a multiple
817                  * of the units we allocate inodes in.
818                  */
819                 icount = sbp->sb_dblocks * sbp->sb_imax_pct;
820                 do_div(icount, 100);
821                 do_div(icount, mp->m_ialloc_blks);
822                 mp->m_maxicount = (icount * mp->m_ialloc_blks)  <<
823                                    sbp->sb_inopblog;
824         } else {
825                 mp->m_maxicount = 0;
826         }
827 }
828
829 /*
830  * Set the default minimum read and write sizes unless
831  * already specified in a mount option.
832  * We use smaller I/O sizes when the file system
833  * is being used for NFS service (wsync mount option).
834  */
835 STATIC void
836 xfs_set_rw_sizes(xfs_mount_t *mp)
837 {
838         xfs_sb_t        *sbp = &(mp->m_sb);
839         int             readio_log, writeio_log;
840
841         if (!(mp->m_flags & XFS_MOUNT_DFLT_IOSIZE)) {
842                 if (mp->m_flags & XFS_MOUNT_WSYNC) {
843                         readio_log = XFS_WSYNC_READIO_LOG;
844                         writeio_log = XFS_WSYNC_WRITEIO_LOG;
845                 } else {
846                         readio_log = XFS_READIO_LOG_LARGE;
847                         writeio_log = XFS_WRITEIO_LOG_LARGE;
848                 }
849         } else {
850                 readio_log = mp->m_readio_log;
851                 writeio_log = mp->m_writeio_log;
852         }
853
854         if (sbp->sb_blocklog > readio_log) {
855                 mp->m_readio_log = sbp->sb_blocklog;
856         } else {
857                 mp->m_readio_log = readio_log;
858         }
859         mp->m_readio_blocks = 1 << (mp->m_readio_log - sbp->sb_blocklog);
860         if (sbp->sb_blocklog > writeio_log) {
861                 mp->m_writeio_log = sbp->sb_blocklog;
862         } else {
863                 mp->m_writeio_log = writeio_log;
864         }
865         mp->m_writeio_blocks = 1 << (mp->m_writeio_log - sbp->sb_blocklog);
866 }
867
868 /*
869  * Set whether we're using inode alignment.
870  */
871 STATIC void
872 xfs_set_inoalignment(xfs_mount_t *mp)
873 {
874         if (xfs_sb_version_hasalign(&mp->m_sb) &&
875             mp->m_sb.sb_inoalignmt >=
876             XFS_B_TO_FSBT(mp, mp->m_inode_cluster_size))
877                 mp->m_inoalign_mask = mp->m_sb.sb_inoalignmt - 1;
878         else
879                 mp->m_inoalign_mask = 0;
880         /*
881          * If we are using stripe alignment, check whether
882          * the stripe unit is a multiple of the inode alignment
883          */
884         if (mp->m_dalign && mp->m_inoalign_mask &&
885             !(mp->m_dalign & mp->m_inoalign_mask))
886                 mp->m_sinoalign = mp->m_dalign;
887         else
888                 mp->m_sinoalign = 0;
889 }
890
891 /*
892  * Check that the data (and log if separate) are an ok size.
893  */
894 STATIC int
895 xfs_check_sizes(xfs_mount_t *mp, int mfsi_flags)
896 {
897         xfs_buf_t       *bp;
898         xfs_daddr_t     d;
899         int             error;
900
901         d = (xfs_daddr_t)XFS_FSB_TO_BB(mp, mp->m_sb.sb_dblocks);
902         if (XFS_BB_TO_FSB(mp, d) != mp->m_sb.sb_dblocks) {
903                 cmn_err(CE_WARN, "XFS: size check 1 failed");
904                 return XFS_ERROR(E2BIG);
905         }
906         error = xfs_read_buf(mp, mp->m_ddev_targp,
907                              d - XFS_FSS_TO_BB(mp, 1),
908                              XFS_FSS_TO_BB(mp, 1), 0, &bp);
909         if (!error) {
910                 xfs_buf_relse(bp);
911         } else {
912                 cmn_err(CE_WARN, "XFS: size check 2 failed");
913                 if (error == ENOSPC)
914                         error = XFS_ERROR(E2BIG);
915                 return error;
916         }
917
918         if (((mfsi_flags & XFS_MFSI_CLIENT) == 0) &&
919             mp->m_logdev_targp != mp->m_ddev_targp) {
920                 d = (xfs_daddr_t)XFS_FSB_TO_BB(mp, mp->m_sb.sb_logblocks);
921                 if (XFS_BB_TO_FSB(mp, d) != mp->m_sb.sb_logblocks) {
922                         cmn_err(CE_WARN, "XFS: size check 3 failed");
923                         return XFS_ERROR(E2BIG);
924                 }
925                 error = xfs_read_buf(mp, mp->m_logdev_targp,
926                                      d - XFS_FSB_TO_BB(mp, 1),
927                                      XFS_FSB_TO_BB(mp, 1), 0, &bp);
928                 if (!error) {
929                         xfs_buf_relse(bp);
930                 } else {
931                         cmn_err(CE_WARN, "XFS: size check 3 failed");
932                         if (error == ENOSPC)
933                                 error = XFS_ERROR(E2BIG);
934                         return error;
935                 }
936         }
937         return 0;
938 }
939
940 /*
941  * xfs_mountfs
942  *
943  * This function does the following on an initial mount of a file system:
944  *      - reads the superblock from disk and init the mount struct
945  *      - if we're a 32-bit kernel, do a size check on the superblock
946  *              so we don't mount terabyte filesystems
947  *      - init mount struct realtime fields
948  *      - allocate inode hash table for fs
949  *      - init directory manager
950  *      - perform recovery and init the log manager
951  */
952 int
953 xfs_mountfs(
954         xfs_mount_t     *mp,
955         int             mfsi_flags)
956 {
957         xfs_sb_t        *sbp = &(mp->m_sb);
958         xfs_inode_t     *rip;
959         bhv_vnode_t     *rvp = NULL;
960         __uint64_t      resblks;
961         __int64_t       update_flags = 0LL;
962         uint            quotamount, quotaflags;
963         int             agno;
964         int             uuid_mounted = 0;
965         int             error = 0;
966
967         if (mp->m_sb_bp == NULL) {
968                 error = xfs_readsb(mp, mfsi_flags);
969                 if (error)
970                         return error;
971         }
972         xfs_mount_common(mp, sbp);
973
974         /*
975          * Check for a bad features2 field alignment. This happened on
976          * some platforms due to xfs_sb_t not being 64bit size aligned
977          * when sb_features was added and hence the compiler put it in
978          * the wrong place.
979          *
980          * If we detect a bad field, we or the set bits into the existing
981          * features2 field in case it has already been modified and we
982          * don't want to lose any features. Zero the bad one and mark
983          * the two fields as needing updates once the transaction subsystem
984          * is online.
985          */
986         if (xfs_sb_has_bad_features2(sbp)) {
987                 cmn_err(CE_WARN,
988                         "XFS: correcting sb_features alignment problem");
989                 sbp->sb_features2 |= sbp->sb_bad_features2;
990                 sbp->sb_bad_features2 = 0;
991                 update_flags |= XFS_SB_FEATURES2 | XFS_SB_BAD_FEATURES2;
992         }
993
994         /*
995          * Check if sb_agblocks is aligned at stripe boundary
996          * If sb_agblocks is NOT aligned turn off m_dalign since
997          * allocator alignment is within an ag, therefore ag has
998          * to be aligned at stripe boundary.
999          */
1000         error = xfs_update_alignment(mp, mfsi_flags, &update_flags);
1001         if (error)
1002                 goto error1;
1003
1004         xfs_alloc_compute_maxlevels(mp);
1005         xfs_bmap_compute_maxlevels(mp, XFS_DATA_FORK);
1006         xfs_bmap_compute_maxlevels(mp, XFS_ATTR_FORK);
1007         xfs_ialloc_compute_maxlevels(mp);
1008
1009         xfs_set_maxicount(mp);
1010
1011         mp->m_maxioffset = xfs_max_file_offset(sbp->sb_blocklog);
1012
1013         /*
1014          * XFS uses the uuid from the superblock as the unique
1015          * identifier for fsid.  We can not use the uuid from the volume
1016          * since a single partition filesystem is identical to a single
1017          * partition volume/filesystem.
1018          */
1019         if ((mfsi_flags & XFS_MFSI_SECOND) == 0 &&
1020             (mp->m_flags & XFS_MOUNT_NOUUID) == 0) {
1021                 if (xfs_uuid_mount(mp)) {
1022                         error = XFS_ERROR(EINVAL);
1023                         goto error1;
1024                 }
1025                 uuid_mounted=1;
1026         }
1027
1028         /*
1029          * Set the minimum read and write sizes
1030          */
1031         xfs_set_rw_sizes(mp);
1032
1033         /*
1034          * Set the inode cluster size.
1035          * This may still be overridden by the file system
1036          * block size if it is larger than the chosen cluster size.
1037          */
1038         mp->m_inode_cluster_size = XFS_INODE_BIG_CLUSTER_SIZE;
1039
1040         /*
1041          * Set inode alignment fields
1042          */
1043         xfs_set_inoalignment(mp);
1044
1045         /*
1046          * Check that the data (and log if separate) are an ok size.
1047          */
1048         error = xfs_check_sizes(mp, mfsi_flags);
1049         if (error)
1050                 goto error1;
1051
1052         /*
1053          * Initialize realtime fields in the mount structure
1054          */
1055         error = xfs_rtmount_init(mp);
1056         if (error) {
1057                 cmn_err(CE_WARN, "XFS: RT mount failed");
1058                 goto error1;
1059         }
1060
1061         /*
1062          * For client case we are done now
1063          */
1064         if (mfsi_flags & XFS_MFSI_CLIENT) {
1065                 return 0;
1066         }
1067
1068         /*
1069          *  Copies the low order bits of the timestamp and the randomly
1070          *  set "sequence" number out of a UUID.
1071          */
1072         uuid_getnodeuniq(&sbp->sb_uuid, mp->m_fixedfsid);
1073
1074         mp->m_dmevmask = 0;     /* not persistent; set after each mount */
1075
1076         xfs_dir_mount(mp);
1077
1078         /*
1079          * Initialize the attribute manager's entries.
1080          */
1081         mp->m_attr_magicpct = (mp->m_sb.sb_blocksize * 37) / 100;
1082
1083         /*
1084          * Initialize the precomputed transaction reservations values.
1085          */
1086         xfs_trans_init(mp);
1087
1088         /*
1089          * Allocate and initialize the per-ag data.
1090          */
1091         init_rwsem(&mp->m_peraglock);
1092         mp->m_perag =
1093                 kmem_zalloc(sbp->sb_agcount * sizeof(xfs_perag_t), KM_SLEEP);
1094
1095         mp->m_maxagi = xfs_initialize_perag(mp, sbp->sb_agcount);
1096
1097         /*
1098          * log's mount-time initialization. Perform 1st part recovery if needed
1099          */
1100         if (likely(sbp->sb_logblocks > 0)) {    /* check for volume case */
1101                 error = xfs_log_mount(mp, mp->m_logdev_targp,
1102                                       XFS_FSB_TO_DADDR(mp, sbp->sb_logstart),
1103                                       XFS_FSB_TO_BB(mp, sbp->sb_logblocks));
1104                 if (error) {
1105                         cmn_err(CE_WARN, "XFS: log mount failed");
1106                         goto error2;
1107                 }
1108         } else {        /* No log has been defined */
1109                 cmn_err(CE_WARN, "XFS: no log defined");
1110                 XFS_ERROR_REPORT("xfs_mountfs_int(1)", XFS_ERRLEVEL_LOW, mp);
1111                 error = XFS_ERROR(EFSCORRUPTED);
1112                 goto error2;
1113         }
1114
1115         /*
1116          * Now the log is mounted, we know if it was an unclean shutdown or
1117          * not. If it was, with the first phase of recovery has completed, we
1118          * have consistent AG blocks on disk. We have not recovered EFIs yet,
1119          * but they are recovered transactionally in the second recovery phase
1120          * later.
1121          *
1122          * Hence we can safely re-initialise incore superblock counters from
1123          * the per-ag data. These may not be correct if the filesystem was not
1124          * cleanly unmounted, so we need to wait for recovery to finish before
1125          * doing this.
1126          *
1127          * If the filesystem was cleanly unmounted, then we can trust the
1128          * values in the superblock to be correct and we don't need to do
1129          * anything here.
1130          *
1131          * If we are currently making the filesystem, the initialisation will
1132          * fail as the perag data is in an undefined state.
1133          */
1134
1135         if (xfs_sb_version_haslazysbcount(&mp->m_sb) &&
1136             !XFS_LAST_UNMOUNT_WAS_CLEAN(mp) &&
1137              !mp->m_sb.sb_inprogress) {
1138                 error = xfs_initialize_perag_data(mp, sbp->sb_agcount);
1139                 if (error) {
1140                         goto error2;
1141                 }
1142         }
1143         /*
1144          * Get and sanity-check the root inode.
1145          * Save the pointer to it in the mount structure.
1146          */
1147         error = xfs_iget(mp, NULL, sbp->sb_rootino, 0, XFS_ILOCK_EXCL, &rip, 0);
1148         if (error) {
1149                 cmn_err(CE_WARN, "XFS: failed to read root inode");
1150                 goto error3;
1151         }
1152
1153         ASSERT(rip != NULL);
1154         rvp = XFS_ITOV(rip);
1155
1156         if (unlikely((rip->i_d.di_mode & S_IFMT) != S_IFDIR)) {
1157                 cmn_err(CE_WARN, "XFS: corrupted root inode");
1158                 cmn_err(CE_WARN, "Device %s - root %llu is not a directory",
1159                         XFS_BUFTARG_NAME(mp->m_ddev_targp),
1160                         (unsigned long long)rip->i_ino);
1161                 xfs_iunlock(rip, XFS_ILOCK_EXCL);
1162                 XFS_ERROR_REPORT("xfs_mountfs_int(2)", XFS_ERRLEVEL_LOW,
1163                                  mp);
1164                 error = XFS_ERROR(EFSCORRUPTED);
1165                 goto error4;
1166         }
1167         mp->m_rootip = rip;     /* save it */
1168
1169         xfs_iunlock(rip, XFS_ILOCK_EXCL);
1170
1171         /*
1172          * Initialize realtime inode pointers in the mount structure
1173          */
1174         error = xfs_rtmount_inodes(mp);
1175         if (error) {
1176                 /*
1177                  * Free up the root inode.
1178                  */
1179                 cmn_err(CE_WARN, "XFS: failed to read RT inodes");
1180                 goto error4;
1181         }
1182
1183         /*
1184          * If fs is not mounted readonly, then update the superblock changes.
1185          */
1186         if (update_flags && !(mp->m_flags & XFS_MOUNT_RDONLY))
1187                 xfs_mount_log_sb(mp, update_flags);
1188
1189         /*
1190          * Initialise the XFS quota management subsystem for this mount
1191          */
1192         error = XFS_QM_INIT(mp, &quotamount, &quotaflags);
1193         if (error)
1194                 goto error4;
1195
1196         /*
1197          * Finish recovering the file system.  This part needed to be
1198          * delayed until after the root and real-time bitmap inodes
1199          * were consistently read in.
1200          */
1201         error = xfs_log_mount_finish(mp, mfsi_flags);
1202         if (error) {
1203                 cmn_err(CE_WARN, "XFS: log mount finish failed");
1204                 goto error4;
1205         }
1206
1207         /*
1208          * Complete the quota initialisation, post-log-replay component.
1209          */
1210         error = XFS_QM_MOUNT(mp, quotamount, quotaflags, mfsi_flags);
1211         if (error)
1212                 goto error4;
1213
1214         /*
1215          * Now we are mounted, reserve a small amount of unused space for
1216          * privileged transactions. This is needed so that transaction
1217          * space required for critical operations can dip into this pool
1218          * when at ENOSPC. This is needed for operations like create with
1219          * attr, unwritten extent conversion at ENOSPC, etc. Data allocations
1220          * are not allowed to use this reserved space.
1221          *
1222          * We default to 5% or 1024 fsbs of space reserved, whichever is smaller.
1223          * This may drive us straight to ENOSPC on mount, but that implies
1224          * we were already there on the last unmount.
1225          */
1226         resblks = mp->m_sb.sb_dblocks;
1227         do_div(resblks, 20);
1228         resblks = min_t(__uint64_t, resblks, 1024);
1229         xfs_reserve_blocks(mp, &resblks, NULL);
1230
1231         return 0;
1232
1233  error4:
1234         /*
1235          * Free up the root inode.
1236          */
1237         VN_RELE(rvp);
1238  error3:
1239         xfs_log_unmount_dealloc(mp);
1240  error2:
1241         for (agno = 0; agno < sbp->sb_agcount; agno++)
1242                 if (mp->m_perag[agno].pagb_list)
1243                         kmem_free(mp->m_perag[agno].pagb_list,
1244                           sizeof(xfs_perag_busy_t) * XFS_PAGB_NUM_SLOTS);
1245         kmem_free(mp->m_perag, sbp->sb_agcount * sizeof(xfs_perag_t));
1246         mp->m_perag = NULL;
1247         /* FALLTHROUGH */
1248  error1:
1249         if (uuid_mounted)
1250                 xfs_uuid_unmount(mp);
1251         xfs_freesb(mp);
1252         return error;
1253 }
1254
1255 /*
1256  * xfs_unmountfs
1257  *
1258  * This flushes out the inodes,dquots and the superblock, unmounts the
1259  * log and makes sure that incore structures are freed.
1260  */
1261 int
1262 xfs_unmountfs(xfs_mount_t *mp, struct cred *cr)
1263 {
1264         __uint64_t      resblks;
1265
1266         /*
1267          * We can potentially deadlock here if we have an inode cluster
1268          * that has been freed has it's buffer still pinned in memory because
1269          * the transaction is still sitting in a iclog. The stale inodes
1270          * on that buffer will have their flush locks held until the
1271          * transaction hits the disk and the callbacks run. the inode
1272          * flush takes the flush lock unconditionally and with nothing to
1273          * push out the iclog we will never get that unlocked. hence we
1274          * need to force the log first.
1275          */
1276         xfs_log_force(mp, (xfs_lsn_t)0, XFS_LOG_FORCE | XFS_LOG_SYNC);
1277         xfs_iflush_all(mp);
1278
1279         XFS_QM_DQPURGEALL(mp, XFS_QMOPT_QUOTALL | XFS_QMOPT_UMOUNTING);
1280
1281         /*
1282          * Flush out the log synchronously so that we know for sure
1283          * that nothing is pinned.  This is important because bflush()
1284          * will skip pinned buffers.
1285          */
1286         xfs_log_force(mp, (xfs_lsn_t)0, XFS_LOG_FORCE | XFS_LOG_SYNC);
1287
1288         xfs_binval(mp->m_ddev_targp);
1289         if (mp->m_rtdev_targp) {
1290                 xfs_binval(mp->m_rtdev_targp);
1291         }
1292
1293         /*
1294          * Unreserve any blocks we have so that when we unmount we don't account
1295          * the reserved free space as used. This is really only necessary for
1296          * lazy superblock counting because it trusts the incore superblock
1297          * counters to be aboslutely correct on clean unmount.
1298          *
1299          * We don't bother correcting this elsewhere for lazy superblock
1300          * counting because on mount of an unclean filesystem we reconstruct the
1301          * correct counter value and this is irrelevant.
1302          *
1303          * For non-lazy counter filesystems, this doesn't matter at all because
1304          * we only every apply deltas to the superblock and hence the incore
1305          * value does not matter....
1306          */
1307         resblks = 0;
1308         xfs_reserve_blocks(mp, &resblks, NULL);
1309
1310         xfs_log_sbcount(mp, 1);
1311         xfs_unmountfs_writesb(mp);
1312         xfs_unmountfs_wait(mp);                 /* wait for async bufs */
1313         xfs_log_unmount(mp);                    /* Done! No more fs ops. */
1314
1315         xfs_freesb(mp);
1316
1317         /*
1318          * All inodes from this mount point should be freed.
1319          */
1320         ASSERT(mp->m_inodes == NULL);
1321
1322         xfs_unmountfs_close(mp, cr);
1323         if ((mp->m_flags & XFS_MOUNT_NOUUID) == 0)
1324                 xfs_uuid_unmount(mp);
1325
1326 #if defined(DEBUG) || defined(INDUCE_IO_ERROR)
1327         xfs_errortag_clearall(mp, 0);
1328 #endif
1329         xfs_mount_free(mp);
1330         return 0;
1331 }
1332
1333 void
1334 xfs_unmountfs_close(xfs_mount_t *mp, struct cred *cr)
1335 {
1336         if (mp->m_logdev_targp && mp->m_logdev_targp != mp->m_ddev_targp)
1337                 xfs_free_buftarg(mp->m_logdev_targp, 1);
1338         if (mp->m_rtdev_targp)
1339                 xfs_free_buftarg(mp->m_rtdev_targp, 1);
1340         xfs_free_buftarg(mp->m_ddev_targp, 0);
1341 }
1342
1343 STATIC void
1344 xfs_unmountfs_wait(xfs_mount_t *mp)
1345 {
1346         if (mp->m_logdev_targp != mp->m_ddev_targp)
1347                 xfs_wait_buftarg(mp->m_logdev_targp);
1348         if (mp->m_rtdev_targp)
1349                 xfs_wait_buftarg(mp->m_rtdev_targp);
1350         xfs_wait_buftarg(mp->m_ddev_targp);
1351 }
1352
1353 int
1354 xfs_fs_writable(xfs_mount_t *mp)
1355 {
1356         return !(xfs_test_for_freeze(mp) || XFS_FORCED_SHUTDOWN(mp) ||
1357                 (mp->m_flags & XFS_MOUNT_RDONLY));
1358 }
1359
1360 /*
1361  * xfs_log_sbcount
1362  *
1363  * Called either periodically to keep the on disk superblock values
1364  * roughly up to date or from unmount to make sure the values are
1365  * correct on a clean unmount.
1366  *
1367  * Note this code can be called during the process of freezing, so
1368  * we may need to use the transaction allocator which does not not
1369  * block when the transaction subsystem is in its frozen state.
1370  */
1371 int
1372 xfs_log_sbcount(
1373         xfs_mount_t     *mp,
1374         uint            sync)
1375 {
1376         xfs_trans_t     *tp;
1377         int             error;
1378
1379         if (!xfs_fs_writable(mp))
1380                 return 0;
1381
1382         xfs_icsb_sync_counters(mp);
1383
1384         /*
1385          * we don't need to do this if we are updating the superblock
1386          * counters on every modification.
1387          */
1388         if (!xfs_sb_version_haslazysbcount(&mp->m_sb))
1389                 return 0;
1390
1391         tp = _xfs_trans_alloc(mp, XFS_TRANS_SB_COUNT);
1392         error = xfs_trans_reserve(tp, 0, mp->m_sb.sb_sectsize + 128, 0, 0,
1393                                         XFS_DEFAULT_LOG_COUNT);
1394         if (error) {
1395                 xfs_trans_cancel(tp, 0);
1396                 return error;
1397         }
1398
1399         xfs_mod_sb(tp, XFS_SB_IFREE | XFS_SB_ICOUNT | XFS_SB_FDBLOCKS);
1400         if (sync)
1401                 xfs_trans_set_sync(tp);
1402         xfs_trans_commit(tp, 0);
1403
1404         return 0;
1405 }
1406
1407 STATIC void
1408 xfs_mark_shared_ro(
1409         xfs_mount_t     *mp,
1410         xfs_buf_t       *bp)
1411 {
1412         xfs_dsb_t       *sb = XFS_BUF_TO_SBP(bp);
1413         __uint16_t      version;
1414
1415         if (!(sb->sb_flags & XFS_SBF_READONLY))
1416                 sb->sb_flags |= XFS_SBF_READONLY;
1417
1418         version = be16_to_cpu(sb->sb_versionnum);
1419         if ((version & XFS_SB_VERSION_NUMBITS) != XFS_SB_VERSION_4 ||
1420             !(version & XFS_SB_VERSION_SHAREDBIT))
1421                 version |= XFS_SB_VERSION_SHAREDBIT;
1422         sb->sb_versionnum = cpu_to_be16(version);
1423 }
1424
1425 int
1426 xfs_unmountfs_writesb(xfs_mount_t *mp)
1427 {
1428         xfs_buf_t       *sbp;
1429         int             error = 0;
1430
1431         /*
1432          * skip superblock write if fs is read-only, or
1433          * if we are doing a forced umount.
1434          */
1435         if (!((mp->m_flags & XFS_MOUNT_RDONLY) ||
1436                 XFS_FORCED_SHUTDOWN(mp))) {
1437
1438                 sbp = xfs_getsb(mp, 0);
1439
1440                 /*
1441                  * mark shared-readonly if desired
1442                  */
1443                 if (mp->m_mk_sharedro)
1444                         xfs_mark_shared_ro(mp, sbp);
1445
1446                 XFS_BUF_UNDONE(sbp);
1447                 XFS_BUF_UNREAD(sbp);
1448                 XFS_BUF_UNDELAYWRITE(sbp);
1449                 XFS_BUF_WRITE(sbp);
1450                 XFS_BUF_UNASYNC(sbp);
1451                 ASSERT(XFS_BUF_TARGET(sbp) == mp->m_ddev_targp);
1452                 xfsbdstrat(mp, sbp);
1453                 /* Nevermind errors we might get here. */
1454                 error = xfs_iowait(sbp);
1455                 if (error)
1456                         xfs_ioerror_alert("xfs_unmountfs_writesb",
1457                                           mp, sbp, XFS_BUF_ADDR(sbp));
1458                 if (error && mp->m_mk_sharedro)
1459                         xfs_fs_cmn_err(CE_ALERT, mp, "Superblock write error detected while unmounting.  Filesystem may not be marked shared readonly");
1460                 xfs_buf_relse(sbp);
1461         }
1462         return error;
1463 }
1464
1465 /*
1466  * xfs_mod_sb() can be used to copy arbitrary changes to the
1467  * in-core superblock into the superblock buffer to be logged.
1468  * It does not provide the higher level of locking that is
1469  * needed to protect the in-core superblock from concurrent
1470  * access.
1471  */
1472 void
1473 xfs_mod_sb(xfs_trans_t *tp, __int64_t fields)
1474 {
1475         xfs_buf_t       *bp;
1476         int             first;
1477         int             last;
1478         xfs_mount_t     *mp;
1479         xfs_sb_field_t  f;
1480
1481         ASSERT(fields);
1482         if (!fields)
1483                 return;
1484         mp = tp->t_mountp;
1485         bp = xfs_trans_getsb(tp, mp, 0);
1486         first = sizeof(xfs_sb_t);
1487         last = 0;
1488
1489         /* translate/copy */
1490
1491         xfs_sb_to_disk(XFS_BUF_TO_SBP(bp), &mp->m_sb, fields);
1492
1493         /* find modified range */
1494
1495         f = (xfs_sb_field_t)xfs_lowbit64((__uint64_t)fields);
1496         ASSERT((1LL << f) & XFS_SB_MOD_BITS);
1497         first = xfs_sb_info[f].offset;
1498
1499         f = (xfs_sb_field_t)xfs_highbit64((__uint64_t)fields);
1500         ASSERT((1LL << f) & XFS_SB_MOD_BITS);
1501         last = xfs_sb_info[f + 1].offset - 1;
1502
1503         xfs_trans_log_buf(tp, bp, first, last);
1504 }
1505
1506
1507 /*
1508  * xfs_mod_incore_sb_unlocked() is a utility routine common used to apply
1509  * a delta to a specified field in the in-core superblock.  Simply
1510  * switch on the field indicated and apply the delta to that field.
1511  * Fields are not allowed to dip below zero, so if the delta would
1512  * do this do not apply it and return EINVAL.
1513  *
1514  * The m_sb_lock must be held when this routine is called.
1515  */
1516 int
1517 xfs_mod_incore_sb_unlocked(
1518         xfs_mount_t     *mp,
1519         xfs_sb_field_t  field,
1520         int64_t         delta,
1521         int             rsvd)
1522 {
1523         int             scounter;       /* short counter for 32 bit fields */
1524         long long       lcounter;       /* long counter for 64 bit fields */
1525         long long       res_used, rem;
1526
1527         /*
1528          * With the in-core superblock spin lock held, switch
1529          * on the indicated field.  Apply the delta to the
1530          * proper field.  If the fields value would dip below
1531          * 0, then do not apply the delta and return EINVAL.
1532          */
1533         switch (field) {
1534         case XFS_SBS_ICOUNT:
1535                 lcounter = (long long)mp->m_sb.sb_icount;
1536                 lcounter += delta;
1537                 if (lcounter < 0) {
1538                         ASSERT(0);
1539                         return XFS_ERROR(EINVAL);
1540                 }
1541                 mp->m_sb.sb_icount = lcounter;
1542                 return 0;
1543         case XFS_SBS_IFREE:
1544                 lcounter = (long long)mp->m_sb.sb_ifree;
1545                 lcounter += delta;
1546                 if (lcounter < 0) {
1547                         ASSERT(0);
1548                         return XFS_ERROR(EINVAL);
1549                 }
1550                 mp->m_sb.sb_ifree = lcounter;
1551                 return 0;
1552         case XFS_SBS_FDBLOCKS:
1553                 lcounter = (long long)
1554                         mp->m_sb.sb_fdblocks - XFS_ALLOC_SET_ASIDE(mp);
1555                 res_used = (long long)(mp->m_resblks - mp->m_resblks_avail);
1556
1557                 if (delta > 0) {                /* Putting blocks back */
1558                         if (res_used > delta) {
1559                                 mp->m_resblks_avail += delta;
1560                         } else {
1561                                 rem = delta - res_used;
1562                                 mp->m_resblks_avail = mp->m_resblks;
1563                                 lcounter += rem;
1564                         }
1565                 } else {                                /* Taking blocks away */
1566
1567                         lcounter += delta;
1568
1569                 /*
1570                  * If were out of blocks, use any available reserved blocks if
1571                  * were allowed to.
1572                  */
1573
1574                         if (lcounter < 0) {
1575                                 if (rsvd) {
1576                                         lcounter = (long long)mp->m_resblks_avail + delta;
1577                                         if (lcounter < 0) {
1578                                                 return XFS_ERROR(ENOSPC);
1579                                         }
1580                                         mp->m_resblks_avail = lcounter;
1581                                         return 0;
1582                                 } else {        /* not reserved */
1583                                         return XFS_ERROR(ENOSPC);
1584                                 }
1585                         }
1586                 }
1587
1588                 mp->m_sb.sb_fdblocks = lcounter + XFS_ALLOC_SET_ASIDE(mp);
1589                 return 0;
1590         case XFS_SBS_FREXTENTS:
1591                 lcounter = (long long)mp->m_sb.sb_frextents;
1592                 lcounter += delta;
1593                 if (lcounter < 0) {
1594                         return XFS_ERROR(ENOSPC);
1595                 }
1596                 mp->m_sb.sb_frextents = lcounter;
1597                 return 0;
1598         case XFS_SBS_DBLOCKS:
1599                 lcounter = (long long)mp->m_sb.sb_dblocks;
1600                 lcounter += delta;
1601                 if (lcounter < 0) {
1602                         ASSERT(0);
1603                         return XFS_ERROR(EINVAL);
1604                 }
1605                 mp->m_sb.sb_dblocks = lcounter;
1606                 return 0;
1607         case XFS_SBS_AGCOUNT:
1608                 scounter = mp->m_sb.sb_agcount;
1609                 scounter += delta;
1610                 if (scounter < 0) {
1611                         ASSERT(0);
1612                         return XFS_ERROR(EINVAL);
1613                 }
1614                 mp->m_sb.sb_agcount = scounter;
1615                 return 0;
1616         case XFS_SBS_IMAX_PCT:
1617                 scounter = mp->m_sb.sb_imax_pct;
1618                 scounter += delta;
1619                 if (scounter < 0) {
1620                         ASSERT(0);
1621                         return XFS_ERROR(EINVAL);
1622                 }
1623                 mp->m_sb.sb_imax_pct = scounter;
1624                 return 0;
1625         case XFS_SBS_REXTSIZE:
1626                 scounter = mp->m_sb.sb_rextsize;
1627                 scounter += delta;
1628                 if (scounter < 0) {
1629                         ASSERT(0);
1630                         return XFS_ERROR(EINVAL);
1631                 }
1632                 mp->m_sb.sb_rextsize = scounter;
1633                 return 0;
1634         case XFS_SBS_RBMBLOCKS:
1635                 scounter = mp->m_sb.sb_rbmblocks;
1636                 scounter += delta;
1637                 if (scounter < 0) {
1638                         ASSERT(0);
1639                         return XFS_ERROR(EINVAL);
1640                 }
1641                 mp->m_sb.sb_rbmblocks = scounter;
1642                 return 0;
1643         case XFS_SBS_RBLOCKS:
1644                 lcounter = (long long)mp->m_sb.sb_rblocks;
1645                 lcounter += delta;
1646                 if (lcounter < 0) {
1647                         ASSERT(0);
1648                         return XFS_ERROR(EINVAL);
1649                 }
1650                 mp->m_sb.sb_rblocks = lcounter;
1651                 return 0;
1652         case XFS_SBS_REXTENTS:
1653                 lcounter = (long long)mp->m_sb.sb_rextents;
1654                 lcounter += delta;
1655                 if (lcounter < 0) {
1656                         ASSERT(0);
1657                         return XFS_ERROR(EINVAL);
1658                 }
1659                 mp->m_sb.sb_rextents = lcounter;
1660                 return 0;
1661         case XFS_SBS_REXTSLOG:
1662                 scounter = mp->m_sb.sb_rextslog;
1663                 scounter += delta;
1664                 if (scounter < 0) {
1665                         ASSERT(0);
1666                         return XFS_ERROR(EINVAL);
1667                 }
1668                 mp->m_sb.sb_rextslog = scounter;
1669                 return 0;
1670         default:
1671                 ASSERT(0);
1672                 return XFS_ERROR(EINVAL);
1673         }
1674 }
1675
1676 /*
1677  * xfs_mod_incore_sb() is used to change a field in the in-core
1678  * superblock structure by the specified delta.  This modification
1679  * is protected by the m_sb_lock.  Just use the xfs_mod_incore_sb_unlocked()
1680  * routine to do the work.
1681  */
1682 int
1683 xfs_mod_incore_sb(
1684         xfs_mount_t     *mp,
1685         xfs_sb_field_t  field,
1686         int64_t         delta,
1687         int             rsvd)
1688 {
1689         int     status;
1690
1691         /* check for per-cpu counters */
1692         switch (field) {
1693 #ifdef HAVE_PERCPU_SB
1694         case XFS_SBS_ICOUNT:
1695         case XFS_SBS_IFREE:
1696         case XFS_SBS_FDBLOCKS:
1697                 if (!(mp->m_flags & XFS_MOUNT_NO_PERCPU_SB)) {
1698                         status = xfs_icsb_modify_counters(mp, field,
1699                                                         delta, rsvd);
1700                         break;
1701                 }
1702                 /* FALLTHROUGH */
1703 #endif
1704         default:
1705                 spin_lock(&mp->m_sb_lock);
1706                 status = xfs_mod_incore_sb_unlocked(mp, field, delta, rsvd);
1707                 spin_unlock(&mp->m_sb_lock);
1708                 break;
1709         }
1710
1711         return status;
1712 }
1713
1714 /*
1715  * xfs_mod_incore_sb_batch() is used to change more than one field
1716  * in the in-core superblock structure at a time.  This modification
1717  * is protected by a lock internal to this module.  The fields and
1718  * changes to those fields are specified in the array of xfs_mod_sb
1719  * structures passed in.
1720  *
1721  * Either all of the specified deltas will be applied or none of
1722  * them will.  If any modified field dips below 0, then all modifications
1723  * will be backed out and EINVAL will be returned.
1724  */
1725 int
1726 xfs_mod_incore_sb_batch(xfs_mount_t *mp, xfs_mod_sb_t *msb, uint nmsb, int rsvd)
1727 {
1728         int             status=0;
1729         xfs_mod_sb_t    *msbp;
1730
1731         /*
1732          * Loop through the array of mod structures and apply each
1733          * individually.  If any fail, then back out all those
1734          * which have already been applied.  Do all of this within
1735          * the scope of the m_sb_lock so that all of the changes will
1736          * be atomic.
1737          */
1738         spin_lock(&mp->m_sb_lock);
1739         msbp = &msb[0];
1740         for (msbp = &msbp[0]; msbp < (msb + nmsb); msbp++) {
1741                 /*
1742                  * Apply the delta at index n.  If it fails, break
1743                  * from the loop so we'll fall into the undo loop
1744                  * below.
1745                  */
1746                 switch (msbp->msb_field) {
1747 #ifdef HAVE_PERCPU_SB
1748                 case XFS_SBS_ICOUNT:
1749                 case XFS_SBS_IFREE:
1750                 case XFS_SBS_FDBLOCKS:
1751                         if (!(mp->m_flags & XFS_MOUNT_NO_PERCPU_SB)) {
1752                                 spin_unlock(&mp->m_sb_lock);
1753                                 status = xfs_icsb_modify_counters(mp,
1754                                                         msbp->msb_field,
1755                                                         msbp->msb_delta, rsvd);
1756                                 spin_lock(&mp->m_sb_lock);
1757                                 break;
1758                         }
1759                         /* FALLTHROUGH */
1760 #endif
1761                 default:
1762                         status = xfs_mod_incore_sb_unlocked(mp,
1763                                                 msbp->msb_field,
1764                                                 msbp->msb_delta, rsvd);
1765                         break;
1766                 }
1767
1768                 if (status != 0) {
1769                         break;
1770                 }
1771         }
1772
1773         /*
1774          * If we didn't complete the loop above, then back out
1775          * any changes made to the superblock.  If you add code
1776          * between the loop above and here, make sure that you
1777          * preserve the value of status. Loop back until
1778          * we step below the beginning of the array.  Make sure
1779          * we don't touch anything back there.
1780          */
1781         if (status != 0) {
1782                 msbp--;
1783                 while (msbp >= msb) {
1784                         switch (msbp->msb_field) {
1785 #ifdef HAVE_PERCPU_SB
1786                         case XFS_SBS_ICOUNT:
1787                         case XFS_SBS_IFREE:
1788                         case XFS_SBS_FDBLOCKS:
1789                                 if (!(mp->m_flags & XFS_MOUNT_NO_PERCPU_SB)) {
1790                                         spin_unlock(&mp->m_sb_lock);
1791                                         status = xfs_icsb_modify_counters(mp,
1792                                                         msbp->msb_field,
1793                                                         -(msbp->msb_delta),
1794                                                         rsvd);
1795                                         spin_lock(&mp->m_sb_lock);
1796                                         break;
1797                                 }
1798                                 /* FALLTHROUGH */
1799 #endif
1800                         default:
1801                                 status = xfs_mod_incore_sb_unlocked(mp,
1802                                                         msbp->msb_field,
1803                                                         -(msbp->msb_delta),
1804                                                         rsvd);
1805                                 break;
1806                         }
1807                         ASSERT(status == 0);
1808                         msbp--;
1809                 }
1810         }
1811         spin_unlock(&mp->m_sb_lock);
1812         return status;
1813 }
1814
1815 /*
1816  * xfs_getsb() is called to obtain the buffer for the superblock.
1817  * The buffer is returned locked and read in from disk.
1818  * The buffer should be released with a call to xfs_brelse().
1819  *
1820  * If the flags parameter is BUF_TRYLOCK, then we'll only return
1821  * the superblock buffer if it can be locked without sleeping.
1822  * If it can't then we'll return NULL.
1823  */
1824 xfs_buf_t *
1825 xfs_getsb(
1826         xfs_mount_t     *mp,
1827         int             flags)
1828 {
1829         xfs_buf_t       *bp;
1830
1831         ASSERT(mp->m_sb_bp != NULL);
1832         bp = mp->m_sb_bp;
1833         if (flags & XFS_BUF_TRYLOCK) {
1834                 if (!XFS_BUF_CPSEMA(bp)) {
1835                         return NULL;
1836                 }
1837         } else {
1838                 XFS_BUF_PSEMA(bp, PRIBIO);
1839         }
1840         XFS_BUF_HOLD(bp);
1841         ASSERT(XFS_BUF_ISDONE(bp));
1842         return bp;
1843 }
1844
1845 /*
1846  * Used to free the superblock along various error paths.
1847  */
1848 void
1849 xfs_freesb(
1850         xfs_mount_t     *mp)
1851 {
1852         xfs_buf_t       *bp;
1853
1854         /*
1855          * Use xfs_getsb() so that the buffer will be locked
1856          * when we call xfs_buf_relse().
1857          */
1858         bp = xfs_getsb(mp, 0);
1859         XFS_BUF_UNMANAGE(bp);
1860         xfs_buf_relse(bp);
1861         mp->m_sb_bp = NULL;
1862 }
1863
1864 /*
1865  * See if the UUID is unique among mounted XFS filesystems.
1866  * Mount fails if UUID is nil or a FS with the same UUID is already mounted.
1867  */
1868 STATIC int
1869 xfs_uuid_mount(
1870         xfs_mount_t     *mp)
1871 {
1872         if (uuid_is_nil(&mp->m_sb.sb_uuid)) {
1873                 cmn_err(CE_WARN,
1874                         "XFS: Filesystem %s has nil UUID - can't mount",
1875                         mp->m_fsname);
1876                 return -1;
1877         }
1878         if (!uuid_table_insert(&mp->m_sb.sb_uuid)) {
1879                 cmn_err(CE_WARN,
1880                         "XFS: Filesystem %s has duplicate UUID - can't mount",
1881                         mp->m_fsname);
1882                 return -1;
1883         }
1884         return 0;
1885 }
1886
1887 /*
1888  * Remove filesystem from the UUID table.
1889  */
1890 STATIC void
1891 xfs_uuid_unmount(
1892         xfs_mount_t     *mp)
1893 {
1894         uuid_table_remove(&mp->m_sb.sb_uuid);
1895 }
1896
1897 /*
1898  * Used to log changes to the superblock unit and width fields which could
1899  * be altered by the mount options. Only the first superblock is updated.
1900  */
1901 STATIC void
1902 xfs_mount_log_sb(
1903         xfs_mount_t     *mp,
1904         __int64_t       fields)
1905 {
1906         xfs_trans_t     *tp;
1907
1908         ASSERT(fields & (XFS_SB_UNIT | XFS_SB_WIDTH | XFS_SB_UUID |
1909                          XFS_SB_FEATURES2 | XFS_SB_BAD_FEATURES2));
1910
1911         tp = xfs_trans_alloc(mp, XFS_TRANS_SB_UNIT);
1912         if (xfs_trans_reserve(tp, 0, mp->m_sb.sb_sectsize + 128, 0, 0,
1913                                 XFS_DEFAULT_LOG_COUNT)) {
1914                 xfs_trans_cancel(tp, 0);
1915                 return;
1916         }
1917         xfs_mod_sb(tp, fields);
1918         xfs_trans_commit(tp, 0);
1919 }
1920
1921
1922 #ifdef HAVE_PERCPU_SB
1923 /*
1924  * Per-cpu incore superblock counters
1925  *
1926  * Simple concept, difficult implementation
1927  *
1928  * Basically, replace the incore superblock counters with a distributed per cpu
1929  * counter for contended fields (e.g.  free block count).
1930  *
1931  * Difficulties arise in that the incore sb is used for ENOSPC checking, and
1932  * hence needs to be accurately read when we are running low on space. Hence
1933  * there is a method to enable and disable the per-cpu counters based on how
1934  * much "stuff" is available in them.
1935  *
1936  * Basically, a counter is enabled if there is enough free resource to justify
1937  * running a per-cpu fast-path. If the per-cpu counter runs out (i.e. a local
1938  * ENOSPC), then we disable the counters to synchronise all callers and
1939  * re-distribute the available resources.
1940  *
1941  * If, once we redistributed the available resources, we still get a failure,
1942  * we disable the per-cpu counter and go through the slow path.
1943  *
1944  * The slow path is the current xfs_mod_incore_sb() function.  This means that
1945  * when we disable a per-cpu counter, we need to drain it's resources back to
1946  * the global superblock. We do this after disabling the counter to prevent
1947  * more threads from queueing up on the counter.
1948  *
1949  * Essentially, this means that we still need a lock in the fast path to enable
1950  * synchronisation between the global counters and the per-cpu counters. This
1951  * is not a problem because the lock will be local to a CPU almost all the time
1952  * and have little contention except when we get to ENOSPC conditions.
1953  *
1954  * Basically, this lock becomes a barrier that enables us to lock out the fast
1955  * path while we do things like enabling and disabling counters and
1956  * synchronising the counters.
1957  *
1958  * Locking rules:
1959  *
1960  *      1. m_sb_lock before picking up per-cpu locks
1961  *      2. per-cpu locks always picked up via for_each_online_cpu() order
1962  *      3. accurate counter sync requires m_sb_lock + per cpu locks
1963  *      4. modifying per-cpu counters requires holding per-cpu lock
1964  *      5. modifying global counters requires holding m_sb_lock
1965  *      6. enabling or disabling a counter requires holding the m_sb_lock 
1966  *         and _none_ of the per-cpu locks.
1967  *
1968  * Disabled counters are only ever re-enabled by a balance operation
1969  * that results in more free resources per CPU than a given threshold.
1970  * To ensure counters don't remain disabled, they are rebalanced when
1971  * the global resource goes above a higher threshold (i.e. some hysteresis
1972  * is present to prevent thrashing).
1973  */
1974
1975 #ifdef CONFIG_HOTPLUG_CPU
1976 /*
1977  * hot-plug CPU notifier support.
1978  *
1979  * We need a notifier per filesystem as we need to be able to identify
1980  * the filesystem to balance the counters out. This is achieved by
1981  * having a notifier block embedded in the xfs_mount_t and doing pointer
1982  * magic to get the mount pointer from the notifier block address.
1983  */
1984 STATIC int
1985 xfs_icsb_cpu_notify(
1986         struct notifier_block *nfb,
1987         unsigned long action,
1988         void *hcpu)
1989 {
1990         xfs_icsb_cnts_t *cntp;
1991         xfs_mount_t     *mp;
1992
1993         mp = (xfs_mount_t *)container_of(nfb, xfs_mount_t, m_icsb_notifier);
1994         cntp = (xfs_icsb_cnts_t *)
1995                         per_cpu_ptr(mp->m_sb_cnts, (unsigned long)hcpu);
1996         switch (action) {
1997         case CPU_UP_PREPARE:
1998         case CPU_UP_PREPARE_FROZEN:
1999                 /* Easy Case - initialize the area and locks, and
2000                  * then rebalance when online does everything else for us. */
2001                 memset(cntp, 0, sizeof(xfs_icsb_cnts_t));
2002                 break;
2003         case CPU_ONLINE:
2004         case CPU_ONLINE_FROZEN:
2005                 xfs_icsb_lock(mp);
2006                 xfs_icsb_balance_counter(mp, XFS_SBS_ICOUNT, 0, 0);
2007                 xfs_icsb_balance_counter(mp, XFS_SBS_IFREE, 0, 0);
2008                 xfs_icsb_balance_counter(mp, XFS_SBS_FDBLOCKS, 0, 0);
2009                 xfs_icsb_unlock(mp);
2010                 break;
2011         case CPU_DEAD:
2012         case CPU_DEAD_FROZEN:
2013                 /* Disable all the counters, then fold the dead cpu's
2014                  * count into the total on the global superblock and
2015                  * re-enable the counters. */
2016                 xfs_icsb_lock(mp);
2017                 spin_lock(&mp->m_sb_lock);
2018                 xfs_icsb_disable_counter(mp, XFS_SBS_ICOUNT);
2019                 xfs_icsb_disable_counter(mp, XFS_SBS_IFREE);
2020                 xfs_icsb_disable_counter(mp, XFS_SBS_FDBLOCKS);
2021
2022                 mp->m_sb.sb_icount += cntp->icsb_icount;
2023                 mp->m_sb.sb_ifree += cntp->icsb_ifree;
2024                 mp->m_sb.sb_fdblocks += cntp->icsb_fdblocks;
2025
2026                 memset(cntp, 0, sizeof(xfs_icsb_cnts_t));
2027
2028                 xfs_icsb_balance_counter(mp, XFS_SBS_ICOUNT,
2029                                          XFS_ICSB_SB_LOCKED, 0);
2030                 xfs_icsb_balance_counter(mp, XFS_SBS_IFREE,
2031                                          XFS_ICSB_SB_LOCKED, 0);
2032                 xfs_icsb_balance_counter(mp, XFS_SBS_FDBLOCKS,
2033                                          XFS_ICSB_SB_LOCKED, 0);
2034                 spin_unlock(&mp->m_sb_lock);
2035                 xfs_icsb_unlock(mp);
2036                 break;
2037         }
2038
2039         return NOTIFY_OK;
2040 }
2041 #endif /* CONFIG_HOTPLUG_CPU */
2042
2043 int
2044 xfs_icsb_init_counters(
2045         xfs_mount_t     *mp)
2046 {
2047         xfs_icsb_cnts_t *cntp;
2048         int             i;
2049
2050         mp->m_sb_cnts = alloc_percpu(xfs_icsb_cnts_t);
2051         if (mp->m_sb_cnts == NULL)
2052                 return -ENOMEM;
2053
2054 #ifdef CONFIG_HOTPLUG_CPU
2055         mp->m_icsb_notifier.notifier_call = xfs_icsb_cpu_notify;
2056         mp->m_icsb_notifier.priority = 0;
2057         register_hotcpu_notifier(&mp->m_icsb_notifier);
2058 #endif /* CONFIG_HOTPLUG_CPU */
2059
2060         for_each_online_cpu(i) {
2061                 cntp = (xfs_icsb_cnts_t *)per_cpu_ptr(mp->m_sb_cnts, i);
2062                 memset(cntp, 0, sizeof(xfs_icsb_cnts_t));
2063         }
2064
2065         mutex_init(&mp->m_icsb_mutex);
2066
2067         /*
2068          * start with all counters disabled so that the
2069          * initial balance kicks us off correctly
2070          */
2071         mp->m_icsb_counters = -1;
2072         return 0;
2073 }
2074
2075 void
2076 xfs_icsb_reinit_counters(
2077         xfs_mount_t     *mp)
2078 {
2079         xfs_icsb_lock(mp);
2080         /*
2081          * start with all counters disabled so that the
2082          * initial balance kicks us off correctly
2083          */
2084         mp->m_icsb_counters = -1;
2085         xfs_icsb_balance_counter(mp, XFS_SBS_ICOUNT, 0, 0);
2086         xfs_icsb_balance_counter(mp, XFS_SBS_IFREE, 0, 0);
2087         xfs_icsb_balance_counter(mp, XFS_SBS_FDBLOCKS, 0, 0);
2088         xfs_icsb_unlock(mp);
2089 }
2090
2091 STATIC void
2092 xfs_icsb_destroy_counters(
2093         xfs_mount_t     *mp)
2094 {
2095         if (mp->m_sb_cnts) {
2096                 unregister_hotcpu_notifier(&mp->m_icsb_notifier);
2097                 free_percpu(mp->m_sb_cnts);
2098         }
2099         mutex_destroy(&mp->m_icsb_mutex);
2100 }
2101
2102 STATIC_INLINE void
2103 xfs_icsb_lock_cntr(
2104         xfs_icsb_cnts_t *icsbp)
2105 {
2106         while (test_and_set_bit(XFS_ICSB_FLAG_LOCK, &icsbp->icsb_flags)) {
2107                 ndelay(1000);
2108         }
2109 }
2110
2111 STATIC_INLINE void
2112 xfs_icsb_unlock_cntr(
2113         xfs_icsb_cnts_t *icsbp)
2114 {
2115         clear_bit(XFS_ICSB_FLAG_LOCK, &icsbp->icsb_flags);
2116 }
2117
2118
2119 STATIC_INLINE void
2120 xfs_icsb_lock_all_counters(
2121         xfs_mount_t     *mp)
2122 {
2123         xfs_icsb_cnts_t *cntp;
2124         int             i;
2125
2126         for_each_online_cpu(i) {
2127                 cntp = (xfs_icsb_cnts_t *)per_cpu_ptr(mp->m_sb_cnts, i);
2128                 xfs_icsb_lock_cntr(cntp);
2129         }
2130 }
2131
2132 STATIC_INLINE void
2133 xfs_icsb_unlock_all_counters(
2134         xfs_mount_t     *mp)
2135 {
2136         xfs_icsb_cnts_t *cntp;
2137         int             i;
2138
2139         for_each_online_cpu(i) {
2140                 cntp = (xfs_icsb_cnts_t *)per_cpu_ptr(mp->m_sb_cnts, i);
2141                 xfs_icsb_unlock_cntr(cntp);
2142         }
2143 }
2144
2145 STATIC void
2146 xfs_icsb_count(
2147         xfs_mount_t     *mp,
2148         xfs_icsb_cnts_t *cnt,
2149         int             flags)
2150 {
2151         xfs_icsb_cnts_t *cntp;
2152         int             i;
2153
2154         memset(cnt, 0, sizeof(xfs_icsb_cnts_t));
2155
2156         if (!(flags & XFS_ICSB_LAZY_COUNT))
2157                 xfs_icsb_lock_all_counters(mp);
2158
2159         for_each_online_cpu(i) {
2160                 cntp = (xfs_icsb_cnts_t *)per_cpu_ptr(mp->m_sb_cnts, i);
2161                 cnt->icsb_icount += cntp->icsb_icount;
2162                 cnt->icsb_ifree += cntp->icsb_ifree;
2163                 cnt->icsb_fdblocks += cntp->icsb_fdblocks;
2164         }
2165
2166         if (!(flags & XFS_ICSB_LAZY_COUNT))
2167                 xfs_icsb_unlock_all_counters(mp);
2168 }
2169
2170 STATIC int
2171 xfs_icsb_counter_disabled(
2172         xfs_mount_t     *mp,
2173         xfs_sb_field_t  field)
2174 {
2175         ASSERT((field >= XFS_SBS_ICOUNT) && (field <= XFS_SBS_FDBLOCKS));
2176         return test_bit(field, &mp->m_icsb_counters);
2177 }
2178
2179 STATIC int
2180 xfs_icsb_disable_counter(
2181         xfs_mount_t     *mp,
2182         xfs_sb_field_t  field)
2183 {
2184         xfs_icsb_cnts_t cnt;
2185
2186         ASSERT((field >= XFS_SBS_ICOUNT) && (field <= XFS_SBS_FDBLOCKS));
2187
2188         /*
2189          * If we are already disabled, then there is nothing to do
2190          * here. We check before locking all the counters to avoid
2191          * the expensive lock operation when being called in the
2192          * slow path and the counter is already disabled. This is
2193          * safe because the only time we set or clear this state is under
2194          * the m_icsb_mutex.
2195          */
2196         if (xfs_icsb_counter_disabled(mp, field))
2197                 return 0;
2198
2199         xfs_icsb_lock_all_counters(mp);
2200         if (!test_and_set_bit(field, &mp->m_icsb_counters)) {
2201                 /* drain back to superblock */
2202
2203                 xfs_icsb_count(mp, &cnt, XFS_ICSB_SB_LOCKED|XFS_ICSB_LAZY_COUNT);
2204                 switch(field) {
2205                 case XFS_SBS_ICOUNT:
2206                         mp->m_sb.sb_icount = cnt.icsb_icount;
2207                         break;
2208                 case XFS_SBS_IFREE:
2209                         mp->m_sb.sb_ifree = cnt.icsb_ifree;
2210                         break;
2211                 case XFS_SBS_FDBLOCKS:
2212                         mp->m_sb.sb_fdblocks = cnt.icsb_fdblocks;
2213                         break;
2214                 default:
2215                         BUG();
2216                 }
2217         }
2218
2219         xfs_icsb_unlock_all_counters(mp);
2220
2221         return 0;
2222 }
2223
2224 STATIC void
2225 xfs_icsb_enable_counter(
2226         xfs_mount_t     *mp,
2227         xfs_sb_field_t  field,
2228         uint64_t        count,
2229         uint64_t        resid)
2230 {
2231         xfs_icsb_cnts_t *cntp;
2232         int             i;
2233
2234         ASSERT((field >= XFS_SBS_ICOUNT) && (field <= XFS_SBS_FDBLOCKS));
2235
2236         xfs_icsb_lock_all_counters(mp);
2237         for_each_online_cpu(i) {
2238                 cntp = per_cpu_ptr(mp->m_sb_cnts, i);
2239                 switch (field) {
2240                 case XFS_SBS_ICOUNT:
2241                         cntp->icsb_icount = count + resid;
2242                         break;
2243                 case XFS_SBS_IFREE:
2244                         cntp->icsb_ifree = count + resid;
2245                         break;
2246                 case XFS_SBS_FDBLOCKS:
2247                         cntp->icsb_fdblocks = count + resid;
2248                         break;
2249                 default:
2250                         BUG();
2251                         break;
2252                 }
2253                 resid = 0;
2254         }
2255         clear_bit(field, &mp->m_icsb_counters);
2256         xfs_icsb_unlock_all_counters(mp);
2257 }
2258
2259 void
2260 xfs_icsb_sync_counters_flags(
2261         xfs_mount_t     *mp,
2262         int             flags)
2263 {
2264         xfs_icsb_cnts_t cnt;
2265
2266         /* Pass 1: lock all counters */
2267         if ((flags & XFS_ICSB_SB_LOCKED) == 0)
2268                 spin_lock(&mp->m_sb_lock);
2269
2270         xfs_icsb_count(mp, &cnt, flags);
2271
2272         /* Step 3: update mp->m_sb fields */
2273         if (!xfs_icsb_counter_disabled(mp, XFS_SBS_ICOUNT))
2274                 mp->m_sb.sb_icount = cnt.icsb_icount;
2275         if (!xfs_icsb_counter_disabled(mp, XFS_SBS_IFREE))
2276                 mp->m_sb.sb_ifree = cnt.icsb_ifree;
2277         if (!xfs_icsb_counter_disabled(mp, XFS_SBS_FDBLOCKS))
2278                 mp->m_sb.sb_fdblocks = cnt.icsb_fdblocks;
2279
2280         if ((flags & XFS_ICSB_SB_LOCKED) == 0)
2281                 spin_unlock(&mp->m_sb_lock);
2282 }
2283
2284 /*
2285  * Accurate update of per-cpu counters to incore superblock
2286  */
2287 STATIC void
2288 xfs_icsb_sync_counters(
2289         xfs_mount_t     *mp)
2290 {
2291         xfs_icsb_sync_counters_flags(mp, 0);
2292 }
2293
2294 /*
2295  * Balance and enable/disable counters as necessary.
2296  *
2297  * Thresholds for re-enabling counters are somewhat magic.  inode counts are
2298  * chosen to be the same number as single on disk allocation chunk per CPU, and
2299  * free blocks is something far enough zero that we aren't going thrash when we
2300  * get near ENOSPC. We also need to supply a minimum we require per cpu to
2301  * prevent looping endlessly when xfs_alloc_space asks for more than will
2302  * be distributed to a single CPU but each CPU has enough blocks to be
2303  * reenabled.
2304  *
2305  * Note that we can be called when counters are already disabled.
2306  * xfs_icsb_disable_counter() optimises the counter locking in this case to
2307  * prevent locking every per-cpu counter needlessly.
2308  */
2309
2310 #define XFS_ICSB_INO_CNTR_REENABLE      (uint64_t)64
2311 #define XFS_ICSB_FDBLK_CNTR_REENABLE(mp) \
2312                 (uint64_t)(512 + XFS_ALLOC_SET_ASIDE(mp))
2313 STATIC void
2314 xfs_icsb_balance_counter(
2315         xfs_mount_t     *mp,
2316         xfs_sb_field_t  field,
2317         int             flags,
2318         int             min_per_cpu)
2319 {
2320         uint64_t        count, resid;
2321         int             weight = num_online_cpus();
2322         uint64_t        min = (uint64_t)min_per_cpu;
2323
2324         if (!(flags & XFS_ICSB_SB_LOCKED))
2325                 spin_lock(&mp->m_sb_lock);
2326
2327         /* disable counter and sync counter */
2328         xfs_icsb_disable_counter(mp, field);
2329
2330         /* update counters  - first CPU gets residual*/
2331         switch (field) {
2332         case XFS_SBS_ICOUNT:
2333                 count = mp->m_sb.sb_icount;
2334                 resid = do_div(count, weight);
2335                 if (count < max(min, XFS_ICSB_INO_CNTR_REENABLE))
2336                         goto out;
2337                 break;
2338         case XFS_SBS_IFREE:
2339                 count = mp->m_sb.sb_ifree;
2340                 resid = do_div(count, weight);
2341                 if (count < max(min, XFS_ICSB_INO_CNTR_REENABLE))
2342                         goto out;
2343                 break;
2344         case XFS_SBS_FDBLOCKS:
2345                 count = mp->m_sb.sb_fdblocks;
2346                 resid = do_div(count, weight);
2347                 if (count < max(min, XFS_ICSB_FDBLK_CNTR_REENABLE(mp)))
2348                         goto out;
2349                 break;
2350         default:
2351                 BUG();
2352                 count = resid = 0;      /* quiet, gcc */
2353                 break;
2354         }
2355
2356         xfs_icsb_enable_counter(mp, field, count, resid);
2357 out:
2358         if (!(flags & XFS_ICSB_SB_LOCKED))
2359                 spin_unlock(&mp->m_sb_lock);
2360 }
2361
2362 STATIC int
2363 xfs_icsb_modify_counters(
2364         xfs_mount_t     *mp,
2365         xfs_sb_field_t  field,
2366         int64_t         delta,
2367         int             rsvd)
2368 {
2369         xfs_icsb_cnts_t *icsbp;
2370         long long       lcounter;       /* long counter for 64 bit fields */
2371         int             cpu, ret = 0;
2372
2373         might_sleep();
2374 again:
2375         cpu = get_cpu();
2376         icsbp = (xfs_icsb_cnts_t *)per_cpu_ptr(mp->m_sb_cnts, cpu);
2377
2378         /*
2379          * if the counter is disabled, go to slow path
2380          */
2381         if (unlikely(xfs_icsb_counter_disabled(mp, field)))
2382                 goto slow_path;
2383         xfs_icsb_lock_cntr(icsbp);
2384         if (unlikely(xfs_icsb_counter_disabled(mp, field))) {
2385                 xfs_icsb_unlock_cntr(icsbp);
2386                 goto slow_path;
2387         }
2388
2389         switch (field) {
2390         case XFS_SBS_ICOUNT:
2391                 lcounter = icsbp->icsb_icount;
2392                 lcounter += delta;
2393                 if (unlikely(lcounter < 0))
2394                         goto balance_counter;
2395                 icsbp->icsb_icount = lcounter;
2396                 break;
2397
2398         case XFS_SBS_IFREE:
2399                 lcounter = icsbp->icsb_ifree;
2400                 lcounter += delta;
2401                 if (unlikely(lcounter < 0))
2402                         goto balance_counter;
2403                 icsbp->icsb_ifree = lcounter;
2404                 break;
2405
2406         case XFS_SBS_FDBLOCKS:
2407                 BUG_ON((mp->m_resblks - mp->m_resblks_avail) != 0);
2408
2409                 lcounter = icsbp->icsb_fdblocks - XFS_ALLOC_SET_ASIDE(mp);
2410                 lcounter += delta;
2411                 if (unlikely(lcounter < 0))
2412                         goto balance_counter;
2413                 icsbp->icsb_fdblocks = lcounter + XFS_ALLOC_SET_ASIDE(mp);
2414                 break;
2415         default:
2416                 BUG();
2417                 break;
2418         }
2419         xfs_icsb_unlock_cntr(icsbp);
2420         put_cpu();
2421         return 0;
2422
2423 slow_path:
2424         put_cpu();
2425
2426         /*
2427          * serialise with a mutex so we don't burn lots of cpu on
2428          * the superblock lock. We still need to hold the superblock
2429          * lock, however, when we modify the global structures.
2430          */
2431         xfs_icsb_lock(mp);
2432
2433         /*
2434          * Now running atomically.
2435          *
2436          * If the counter is enabled, someone has beaten us to rebalancing.
2437          * Drop the lock and try again in the fast path....
2438          */
2439         if (!(xfs_icsb_counter_disabled(mp, field))) {
2440                 xfs_icsb_unlock(mp);
2441                 goto again;
2442         }
2443
2444         /*
2445          * The counter is currently disabled. Because we are
2446          * running atomically here, we know a rebalance cannot
2447          * be in progress. Hence we can go straight to operating
2448          * on the global superblock. We do not call xfs_mod_incore_sb()
2449          * here even though we need to get the m_sb_lock. Doing so
2450          * will cause us to re-enter this function and deadlock.
2451          * Hence we get the m_sb_lock ourselves and then call
2452          * xfs_mod_incore_sb_unlocked() as the unlocked path operates
2453          * directly on the global counters.
2454          */
2455         spin_lock(&mp->m_sb_lock);
2456         ret = xfs_mod_incore_sb_unlocked(mp, field, delta, rsvd);
2457         spin_unlock(&mp->m_sb_lock);
2458
2459         /*
2460          * Now that we've modified the global superblock, we
2461          * may be able to re-enable the distributed counters
2462          * (e.g. lots of space just got freed). After that
2463          * we are done.
2464          */
2465         if (ret != ENOSPC)
2466                 xfs_icsb_balance_counter(mp, field, 0, 0);
2467         xfs_icsb_unlock(mp);
2468         return ret;
2469
2470 balance_counter:
2471         xfs_icsb_unlock_cntr(icsbp);
2472         put_cpu();
2473
2474         /*
2475          * We may have multiple threads here if multiple per-cpu
2476          * counters run dry at the same time. This will mean we can
2477          * do more balances than strictly necessary but it is not
2478          * the common slowpath case.
2479          */
2480         xfs_icsb_lock(mp);
2481
2482         /*
2483          * running atomically.
2484          *
2485          * This will leave the counter in the correct state for future
2486          * accesses. After the rebalance, we simply try again and our retry
2487          * will either succeed through the fast path or slow path without
2488          * another balance operation being required.
2489          */
2490         xfs_icsb_balance_counter(mp, field, 0, delta);
2491         xfs_icsb_unlock(mp);
2492         goto again;
2493 }
2494
2495 #endif