99bab1e372b1ce03e7932781a446577c00c8ce04
[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_sbunit(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     { sizeof(xfs_sb_t),                  0 }
123 };
124
125 /*
126  * Return a pointer to an initialized xfs_mount structure.
127  */
128 xfs_mount_t *
129 xfs_mount_init(void)
130 {
131         xfs_mount_t *mp;
132
133         mp = kmem_zalloc(sizeof(xfs_mount_t), KM_SLEEP);
134
135         if (xfs_icsb_init_counters(mp)) {
136                 mp->m_flags |= XFS_MOUNT_NO_PERCPU_SB;
137         }
138
139         spin_lock_init(&mp->m_sb_lock);
140         mutex_init(&mp->m_ilock);
141         mutex_init(&mp->m_growlock);
142         atomic_set(&mp->m_active_trans, 0);
143
144         return mp;
145 }
146
147 /*
148  * Free up the resources associated with a mount structure.  Assume that
149  * the structure was initially zeroed, so we can tell which fields got
150  * initialized.
151  */
152 void
153 xfs_mount_free(
154         xfs_mount_t     *mp)
155 {
156         if (mp->m_perag) {
157                 int     agno;
158
159                 for (agno = 0; agno < mp->m_maxagi; agno++)
160                         if (mp->m_perag[agno].pagb_list)
161                                 kmem_free(mp->m_perag[agno].pagb_list,
162                                                 sizeof(xfs_perag_busy_t) *
163                                                         XFS_PAGB_NUM_SLOTS);
164                 kmem_free(mp->m_perag,
165                           sizeof(xfs_perag_t) * mp->m_sb.sb_agcount);
166         }
167
168         spinlock_destroy(&mp->m_ail_lock);
169         spinlock_destroy(&mp->m_sb_lock);
170         mutex_destroy(&mp->m_ilock);
171         mutex_destroy(&mp->m_growlock);
172         if (mp->m_quotainfo)
173                 XFS_QM_DONE(mp);
174
175         if (mp->m_fsname != NULL)
176                 kmem_free(mp->m_fsname, mp->m_fsname_len);
177         if (mp->m_rtname != NULL)
178                 kmem_free(mp->m_rtname, strlen(mp->m_rtname) + 1);
179         if (mp->m_logname != NULL)
180                 kmem_free(mp->m_logname, strlen(mp->m_logname) + 1);
181
182         xfs_icsb_destroy_counters(mp);
183 }
184
185 /*
186  * Check size of device based on the (data/realtime) block count.
187  * Note: this check is used by the growfs code as well as mount.
188  */
189 int
190 xfs_sb_validate_fsb_count(
191         xfs_sb_t        *sbp,
192         __uint64_t      nblocks)
193 {
194         ASSERT(PAGE_SHIFT >= sbp->sb_blocklog);
195         ASSERT(sbp->sb_blocklog >= BBSHIFT);
196
197 #if XFS_BIG_BLKNOS     /* Limited by ULONG_MAX of page cache index */
198         if (nblocks >> (PAGE_CACHE_SHIFT - sbp->sb_blocklog) > ULONG_MAX)
199                 return E2BIG;
200 #else                  /* Limited by UINT_MAX of sectors */
201         if (nblocks << (sbp->sb_blocklog - BBSHIFT) > UINT_MAX)
202                 return E2BIG;
203 #endif
204         return 0;
205 }
206
207 /*
208  * Check the validity of the SB found.
209  */
210 STATIC int
211 xfs_mount_validate_sb(
212         xfs_mount_t     *mp,
213         xfs_sb_t        *sbp,
214         int             flags)
215 {
216         /*
217          * If the log device and data device have the
218          * same device number, the log is internal.
219          * Consequently, the sb_logstart should be non-zero.  If
220          * we have a zero sb_logstart in this case, we may be trying to mount
221          * a volume filesystem in a non-volume manner.
222          */
223         if (sbp->sb_magicnum != XFS_SB_MAGIC) {
224                 xfs_fs_mount_cmn_err(flags, "bad magic number");
225                 return XFS_ERROR(EWRONGFS);
226         }
227
228         if (!xfs_sb_good_version(sbp)) {
229                 xfs_fs_mount_cmn_err(flags, "bad version");
230                 return XFS_ERROR(EWRONGFS);
231         }
232
233         if (unlikely(
234             sbp->sb_logstart == 0 && mp->m_logdev_targp == mp->m_ddev_targp)) {
235                 xfs_fs_mount_cmn_err(flags,
236                         "filesystem is marked as having an external log; "
237                         "specify logdev on the\nmount command line.");
238                 return XFS_ERROR(EINVAL);
239         }
240
241         if (unlikely(
242             sbp->sb_logstart != 0 && mp->m_logdev_targp != mp->m_ddev_targp)) {
243                 xfs_fs_mount_cmn_err(flags,
244                         "filesystem is marked as having an internal log; "
245                         "do not specify logdev on\nthe mount command line.");
246                 return XFS_ERROR(EINVAL);
247         }
248
249         /*
250          * More sanity checking. These were stolen directly from
251          * xfs_repair.
252          */
253         if (unlikely(
254             sbp->sb_agcount <= 0                                        ||
255             sbp->sb_sectsize < XFS_MIN_SECTORSIZE                       ||
256             sbp->sb_sectsize > XFS_MAX_SECTORSIZE                       ||
257             sbp->sb_sectlog < XFS_MIN_SECTORSIZE_LOG                    ||
258             sbp->sb_sectlog > XFS_MAX_SECTORSIZE_LOG                    ||
259             sbp->sb_blocksize < XFS_MIN_BLOCKSIZE                       ||
260             sbp->sb_blocksize > XFS_MAX_BLOCKSIZE                       ||
261             sbp->sb_blocklog < XFS_MIN_BLOCKSIZE_LOG                    ||
262             sbp->sb_blocklog > XFS_MAX_BLOCKSIZE_LOG                    ||
263             sbp->sb_inodesize < XFS_DINODE_MIN_SIZE                     ||
264             sbp->sb_inodesize > XFS_DINODE_MAX_SIZE                     ||
265             sbp->sb_inodelog < XFS_DINODE_MIN_LOG                       ||
266             sbp->sb_inodelog > XFS_DINODE_MAX_LOG                       ||
267             (sbp->sb_blocklog - sbp->sb_inodelog != sbp->sb_inopblog)   ||
268             (sbp->sb_rextsize * sbp->sb_blocksize > XFS_MAX_RTEXTSIZE)  ||
269             (sbp->sb_rextsize * sbp->sb_blocksize < XFS_MIN_RTEXTSIZE)  ||
270             (sbp->sb_imax_pct > 100 /* zero sb_imax_pct is valid */))) {
271                 xfs_fs_mount_cmn_err(flags, "SB sanity check 1 failed");
272                 return XFS_ERROR(EFSCORRUPTED);
273         }
274
275         /*
276          * Sanity check AG count, size fields against data size field
277          */
278         if (unlikely(
279             sbp->sb_dblocks == 0 ||
280             sbp->sb_dblocks >
281              (xfs_drfsbno_t)sbp->sb_agcount * sbp->sb_agblocks ||
282             sbp->sb_dblocks < (xfs_drfsbno_t)(sbp->sb_agcount - 1) *
283                               sbp->sb_agblocks + XFS_MIN_AG_BLOCKS)) {
284                 xfs_fs_mount_cmn_err(flags, "SB sanity check 2 failed");
285                 return XFS_ERROR(EFSCORRUPTED);
286         }
287
288         if (xfs_sb_validate_fsb_count(sbp, sbp->sb_dblocks) ||
289             xfs_sb_validate_fsb_count(sbp, sbp->sb_rblocks)) {
290                 xfs_fs_mount_cmn_err(flags,
291                         "file system too large to be mounted on this system.");
292                 return XFS_ERROR(E2BIG);
293         }
294
295         if (unlikely(sbp->sb_inprogress)) {
296                 xfs_fs_mount_cmn_err(flags, "file system busy");
297                 return XFS_ERROR(EFSCORRUPTED);
298         }
299
300         /*
301          * Version 1 directory format has never worked on Linux.
302          */
303         if (unlikely(!xfs_sb_version_hasdirv2(sbp))) {
304                 xfs_fs_mount_cmn_err(flags,
305                         "file system using version 1 directory format");
306                 return XFS_ERROR(ENOSYS);
307         }
308
309         /*
310          * Until this is fixed only page-sized or smaller data blocks work.
311          */
312         if (unlikely(sbp->sb_blocksize > PAGE_SIZE)) {
313                 xfs_fs_mount_cmn_err(flags,
314                         "file system with blocksize %d bytes",
315                         sbp->sb_blocksize);
316                 xfs_fs_mount_cmn_err(flags,
317                         "only pagesize (%ld) or less will currently work.",
318                         PAGE_SIZE);
319                 return XFS_ERROR(ENOSYS);
320         }
321
322         return 0;
323 }
324
325 STATIC void
326 xfs_initialize_perag_icache(
327         xfs_perag_t     *pag)
328 {
329         if (!pag->pag_ici_init) {
330                 rwlock_init(&pag->pag_ici_lock);
331                 INIT_RADIX_TREE(&pag->pag_ici_root, GFP_ATOMIC);
332                 pag->pag_ici_init = 1;
333         }
334 }
335
336 xfs_agnumber_t
337 xfs_initialize_perag(
338         xfs_mount_t     *mp,
339         xfs_agnumber_t  agcount)
340 {
341         xfs_agnumber_t  index, max_metadata;
342         xfs_perag_t     *pag;
343         xfs_agino_t     agino;
344         xfs_ino_t       ino;
345         xfs_sb_t        *sbp = &mp->m_sb;
346         xfs_ino_t       max_inum = XFS_MAXINUMBER_32;
347
348         /* Check to see if the filesystem can overflow 32 bit inodes */
349         agino = XFS_OFFBNO_TO_AGINO(mp, sbp->sb_agblocks - 1, 0);
350         ino = XFS_AGINO_TO_INO(mp, agcount - 1, agino);
351
352         /* Clear the mount flag if no inode can overflow 32 bits
353          * on this filesystem, or if specifically requested..
354          */
355         if ((mp->m_flags & XFS_MOUNT_SMALL_INUMS) && ino > max_inum) {
356                 mp->m_flags |= XFS_MOUNT_32BITINODES;
357         } else {
358                 mp->m_flags &= ~XFS_MOUNT_32BITINODES;
359         }
360
361         /* If we can overflow then setup the ag headers accordingly */
362         if (mp->m_flags & XFS_MOUNT_32BITINODES) {
363                 /* Calculate how much should be reserved for inodes to
364                  * meet the max inode percentage.
365                  */
366                 if (mp->m_maxicount) {
367                         __uint64_t      icount;
368
369                         icount = sbp->sb_dblocks * sbp->sb_imax_pct;
370                         do_div(icount, 100);
371                         icount += sbp->sb_agblocks - 1;
372                         do_div(icount, sbp->sb_agblocks);
373                         max_metadata = icount;
374                 } else {
375                         max_metadata = agcount;
376                 }
377                 for (index = 0; index < agcount; index++) {
378                         ino = XFS_AGINO_TO_INO(mp, index, agino);
379                         if (ino > max_inum) {
380                                 index++;
381                                 break;
382                         }
383
384                         /* This ag is preferred for inodes */
385                         pag = &mp->m_perag[index];
386                         pag->pagi_inodeok = 1;
387                         if (index < max_metadata)
388                                 pag->pagf_metadata = 1;
389                         xfs_initialize_perag_icache(pag);
390                 }
391         } else {
392                 /* Setup default behavior for smaller filesystems */
393                 for (index = 0; index < agcount; index++) {
394                         pag = &mp->m_perag[index];
395                         pag->pagi_inodeok = 1;
396                         xfs_initialize_perag_icache(pag);
397                 }
398         }
399         return index;
400 }
401
402 void
403 xfs_sb_from_disk(
404         xfs_sb_t        *to,
405         xfs_dsb_t       *from)
406 {
407         to->sb_magicnum = be32_to_cpu(from->sb_magicnum);
408         to->sb_blocksize = be32_to_cpu(from->sb_blocksize);
409         to->sb_dblocks = be64_to_cpu(from->sb_dblocks);
410         to->sb_rblocks = be64_to_cpu(from->sb_rblocks);
411         to->sb_rextents = be64_to_cpu(from->sb_rextents);
412         memcpy(&to->sb_uuid, &from->sb_uuid, sizeof(to->sb_uuid));
413         to->sb_logstart = be64_to_cpu(from->sb_logstart);
414         to->sb_rootino = be64_to_cpu(from->sb_rootino);
415         to->sb_rbmino = be64_to_cpu(from->sb_rbmino);
416         to->sb_rsumino = be64_to_cpu(from->sb_rsumino);
417         to->sb_rextsize = be32_to_cpu(from->sb_rextsize);
418         to->sb_agblocks = be32_to_cpu(from->sb_agblocks);
419         to->sb_agcount = be32_to_cpu(from->sb_agcount);
420         to->sb_rbmblocks = be32_to_cpu(from->sb_rbmblocks);
421         to->sb_logblocks = be32_to_cpu(from->sb_logblocks);
422         to->sb_versionnum = be16_to_cpu(from->sb_versionnum);
423         to->sb_sectsize = be16_to_cpu(from->sb_sectsize);
424         to->sb_inodesize = be16_to_cpu(from->sb_inodesize);
425         to->sb_inopblock = be16_to_cpu(from->sb_inopblock);
426         memcpy(&to->sb_fname, &from->sb_fname, sizeof(to->sb_fname));
427         to->sb_blocklog = from->sb_blocklog;
428         to->sb_sectlog = from->sb_sectlog;
429         to->sb_inodelog = from->sb_inodelog;
430         to->sb_inopblog = from->sb_inopblog;
431         to->sb_agblklog = from->sb_agblklog;
432         to->sb_rextslog = from->sb_rextslog;
433         to->sb_inprogress = from->sb_inprogress;
434         to->sb_imax_pct = from->sb_imax_pct;
435         to->sb_icount = be64_to_cpu(from->sb_icount);
436         to->sb_ifree = be64_to_cpu(from->sb_ifree);
437         to->sb_fdblocks = be64_to_cpu(from->sb_fdblocks);
438         to->sb_frextents = be64_to_cpu(from->sb_frextents);
439         to->sb_uquotino = be64_to_cpu(from->sb_uquotino);
440         to->sb_gquotino = be64_to_cpu(from->sb_gquotino);
441         to->sb_qflags = be16_to_cpu(from->sb_qflags);
442         to->sb_flags = from->sb_flags;
443         to->sb_shared_vn = from->sb_shared_vn;
444         to->sb_inoalignmt = be32_to_cpu(from->sb_inoalignmt);
445         to->sb_unit = be32_to_cpu(from->sb_unit);
446         to->sb_width = be32_to_cpu(from->sb_width);
447         to->sb_dirblklog = from->sb_dirblklog;
448         to->sb_logsectlog = from->sb_logsectlog;
449         to->sb_logsectsize = be16_to_cpu(from->sb_logsectsize);
450         to->sb_logsunit = be32_to_cpu(from->sb_logsunit);
451         to->sb_features2 = be32_to_cpu(from->sb_features2);
452 }
453
454 /*
455  * Copy in core superblock to ondisk one.
456  *
457  * The fields argument is mask of superblock fields to copy.
458  */
459 void
460 xfs_sb_to_disk(
461         xfs_dsb_t       *to,
462         xfs_sb_t        *from,
463         __int64_t       fields)
464 {
465         xfs_caddr_t     to_ptr = (xfs_caddr_t)to;
466         xfs_caddr_t     from_ptr = (xfs_caddr_t)from;
467         xfs_sb_field_t  f;
468         int             first;
469         int             size;
470
471         ASSERT(fields);
472         if (!fields)
473                 return;
474
475         while (fields) {
476                 f = (xfs_sb_field_t)xfs_lowbit64((__uint64_t)fields);
477                 first = xfs_sb_info[f].offset;
478                 size = xfs_sb_info[f + 1].offset - first;
479
480                 ASSERT(xfs_sb_info[f].type == 0 || xfs_sb_info[f].type == 1);
481
482                 if (size == 1 || xfs_sb_info[f].type == 1) {
483                         memcpy(to_ptr + first, from_ptr + first, size);
484                 } else {
485                         switch (size) {
486                         case 2:
487                                 *(__be16 *)(to_ptr + first) =
488                                         cpu_to_be16(*(__u16 *)(from_ptr + first));
489                                 break;
490                         case 4:
491                                 *(__be32 *)(to_ptr + first) =
492                                         cpu_to_be32(*(__u32 *)(from_ptr + first));
493                                 break;
494                         case 8:
495                                 *(__be64 *)(to_ptr + first) =
496                                         cpu_to_be64(*(__u64 *)(from_ptr + first));
497                                 break;
498                         default:
499                                 ASSERT(0);
500                         }
501                 }
502
503                 fields &= ~(1LL << f);
504         }
505 }
506
507 /*
508  * xfs_readsb
509  *
510  * Does the initial read of the superblock.
511  */
512 int
513 xfs_readsb(xfs_mount_t *mp, int flags)
514 {
515         unsigned int    sector_size;
516         unsigned int    extra_flags;
517         xfs_buf_t       *bp;
518         int             error;
519
520         ASSERT(mp->m_sb_bp == NULL);
521         ASSERT(mp->m_ddev_targp != NULL);
522
523         /*
524          * Allocate a (locked) buffer to hold the superblock.
525          * This will be kept around at all times to optimize
526          * access to the superblock.
527          */
528         sector_size = xfs_getsize_buftarg(mp->m_ddev_targp);
529         extra_flags = XFS_BUF_LOCK | XFS_BUF_MANAGE | XFS_BUF_MAPPED;
530
531         bp = xfs_buf_read_flags(mp->m_ddev_targp, XFS_SB_DADDR,
532                                 BTOBB(sector_size), extra_flags);
533         if (!bp || XFS_BUF_ISERROR(bp)) {
534                 xfs_fs_mount_cmn_err(flags, "SB read failed");
535                 error = bp ? XFS_BUF_GETERROR(bp) : ENOMEM;
536                 goto fail;
537         }
538         ASSERT(XFS_BUF_ISBUSY(bp));
539         ASSERT(XFS_BUF_VALUSEMA(bp) <= 0);
540
541         /*
542          * Initialize the mount structure from the superblock.
543          * But first do some basic consistency checking.
544          */
545         xfs_sb_from_disk(&mp->m_sb, XFS_BUF_TO_SBP(bp));
546
547         error = xfs_mount_validate_sb(mp, &(mp->m_sb), flags);
548         if (error) {
549                 xfs_fs_mount_cmn_err(flags, "SB validate failed");
550                 goto fail;
551         }
552
553         /*
554          * We must be able to do sector-sized and sector-aligned IO.
555          */
556         if (sector_size > mp->m_sb.sb_sectsize) {
557                 xfs_fs_mount_cmn_err(flags,
558                         "device supports only %u byte sectors (not %u)",
559                         sector_size, mp->m_sb.sb_sectsize);
560                 error = ENOSYS;
561                 goto fail;
562         }
563
564         /*
565          * If device sector size is smaller than the superblock size,
566          * re-read the superblock so the buffer is correctly sized.
567          */
568         if (sector_size < mp->m_sb.sb_sectsize) {
569                 XFS_BUF_UNMANAGE(bp);
570                 xfs_buf_relse(bp);
571                 sector_size = mp->m_sb.sb_sectsize;
572                 bp = xfs_buf_read_flags(mp->m_ddev_targp, XFS_SB_DADDR,
573                                         BTOBB(sector_size), extra_flags);
574                 if (!bp || XFS_BUF_ISERROR(bp)) {
575                         xfs_fs_mount_cmn_err(flags, "SB re-read failed");
576                         error = bp ? XFS_BUF_GETERROR(bp) : ENOMEM;
577                         goto fail;
578                 }
579                 ASSERT(XFS_BUF_ISBUSY(bp));
580                 ASSERT(XFS_BUF_VALUSEMA(bp) <= 0);
581         }
582
583         /* Initialize per-cpu counters */
584         xfs_icsb_reinit_counters(mp);
585
586         mp->m_sb_bp = bp;
587         xfs_buf_relse(bp);
588         ASSERT(XFS_BUF_VALUSEMA(bp) > 0);
589         return 0;
590
591  fail:
592         if (bp) {
593                 XFS_BUF_UNMANAGE(bp);
594                 xfs_buf_relse(bp);
595         }
596         return error;
597 }
598
599
600 /*
601  * xfs_mount_common
602  *
603  * Mount initialization code establishing various mount
604  * fields from the superblock associated with the given
605  * mount structure
606  */
607 STATIC void
608 xfs_mount_common(xfs_mount_t *mp, xfs_sb_t *sbp)
609 {
610         int     i;
611
612         mp->m_agfrotor = mp->m_agirotor = 0;
613         spin_lock_init(&mp->m_agirotor_lock);
614         mp->m_maxagi = mp->m_sb.sb_agcount;
615         mp->m_blkbit_log = sbp->sb_blocklog + XFS_NBBYLOG;
616         mp->m_blkbb_log = sbp->sb_blocklog - BBSHIFT;
617         mp->m_sectbb_log = sbp->sb_sectlog - BBSHIFT;
618         mp->m_agno_log = xfs_highbit32(sbp->sb_agcount - 1) + 1;
619         mp->m_agino_log = sbp->sb_inopblog + sbp->sb_agblklog;
620         mp->m_litino = sbp->sb_inodesize -
621                 ((uint)sizeof(xfs_dinode_core_t) + (uint)sizeof(xfs_agino_t));
622         mp->m_blockmask = sbp->sb_blocksize - 1;
623         mp->m_blockwsize = sbp->sb_blocksize >> XFS_WORDLOG;
624         mp->m_blockwmask = mp->m_blockwsize - 1;
625         INIT_LIST_HEAD(&mp->m_del_inodes);
626
627         /*
628          * Setup for attributes, in case they get created.
629          * This value is for inodes getting attributes for the first time,
630          * the per-inode value is for old attribute values.
631          */
632         ASSERT(sbp->sb_inodesize >= 256 && sbp->sb_inodesize <= 2048);
633         switch (sbp->sb_inodesize) {
634         case 256:
635                 mp->m_attroffset = XFS_LITINO(mp) -
636                                    XFS_BMDR_SPACE_CALC(MINABTPTRS);
637                 break;
638         case 512:
639         case 1024:
640         case 2048:
641                 mp->m_attroffset = XFS_BMDR_SPACE_CALC(6 * MINABTPTRS);
642                 break;
643         default:
644                 ASSERT(0);
645         }
646         ASSERT(mp->m_attroffset < XFS_LITINO(mp));
647
648         for (i = 0; i < 2; i++) {
649                 mp->m_alloc_mxr[i] = XFS_BTREE_BLOCK_MAXRECS(sbp->sb_blocksize,
650                         xfs_alloc, i == 0);
651                 mp->m_alloc_mnr[i] = XFS_BTREE_BLOCK_MINRECS(sbp->sb_blocksize,
652                         xfs_alloc, i == 0);
653         }
654         for (i = 0; i < 2; i++) {
655                 mp->m_bmap_dmxr[i] = XFS_BTREE_BLOCK_MAXRECS(sbp->sb_blocksize,
656                         xfs_bmbt, i == 0);
657                 mp->m_bmap_dmnr[i] = XFS_BTREE_BLOCK_MINRECS(sbp->sb_blocksize,
658                         xfs_bmbt, i == 0);
659         }
660         for (i = 0; i < 2; i++) {
661                 mp->m_inobt_mxr[i] = XFS_BTREE_BLOCK_MAXRECS(sbp->sb_blocksize,
662                         xfs_inobt, i == 0);
663                 mp->m_inobt_mnr[i] = XFS_BTREE_BLOCK_MINRECS(sbp->sb_blocksize,
664                         xfs_inobt, i == 0);
665         }
666
667         mp->m_bsize = XFS_FSB_TO_BB(mp, 1);
668         mp->m_ialloc_inos = (int)MAX((__uint16_t)XFS_INODES_PER_CHUNK,
669                                         sbp->sb_inopblock);
670         mp->m_ialloc_blks = mp->m_ialloc_inos >> sbp->sb_inopblog;
671 }
672
673 /*
674  * xfs_initialize_perag_data
675  *
676  * Read in each per-ag structure so we can count up the number of
677  * allocated inodes, free inodes and used filesystem blocks as this
678  * information is no longer persistent in the superblock. Once we have
679  * this information, write it into the in-core superblock structure.
680  */
681 STATIC int
682 xfs_initialize_perag_data(xfs_mount_t *mp, xfs_agnumber_t agcount)
683 {
684         xfs_agnumber_t  index;
685         xfs_perag_t     *pag;
686         xfs_sb_t        *sbp = &mp->m_sb;
687         uint64_t        ifree = 0;
688         uint64_t        ialloc = 0;
689         uint64_t        bfree = 0;
690         uint64_t        bfreelst = 0;
691         uint64_t        btree = 0;
692         int             error;
693
694         for (index = 0; index < agcount; index++) {
695                 /*
696                  * read the agf, then the agi. This gets us
697                  * all the inforamtion we need and populates the
698                  * per-ag structures for us.
699                  */
700                 error = xfs_alloc_pagf_init(mp, NULL, index, 0);
701                 if (error)
702                         return error;
703
704                 error = xfs_ialloc_pagi_init(mp, NULL, index);
705                 if (error)
706                         return error;
707                 pag = &mp->m_perag[index];
708                 ifree += pag->pagi_freecount;
709                 ialloc += pag->pagi_count;
710                 bfree += pag->pagf_freeblks;
711                 bfreelst += pag->pagf_flcount;
712                 btree += pag->pagf_btreeblks;
713         }
714         /*
715          * Overwrite incore superblock counters with just-read data
716          */
717         spin_lock(&mp->m_sb_lock);
718         sbp->sb_ifree = ifree;
719         sbp->sb_icount = ialloc;
720         sbp->sb_fdblocks = bfree + bfreelst + btree;
721         spin_unlock(&mp->m_sb_lock);
722
723         /* Fixup the per-cpu counters as well. */
724         xfs_icsb_reinit_counters(mp);
725
726         return 0;
727 }
728
729 /*
730  * Update alignment values based on mount options and sb values
731  */
732 STATIC int
733 xfs_update_alignment(xfs_mount_t *mp, int mfsi_flags, __uint64_t *update_flags)
734 {
735         xfs_sb_t        *sbp = &(mp->m_sb);
736
737         if (mp->m_dalign && !(mfsi_flags & XFS_MFSI_SECOND)) {
738                 /*
739                  * If stripe unit and stripe width are not multiples
740                  * of the fs blocksize turn off alignment.
741                  */
742                 if ((BBTOB(mp->m_dalign) & mp->m_blockmask) ||
743                     (BBTOB(mp->m_swidth) & mp->m_blockmask)) {
744                         if (mp->m_flags & XFS_MOUNT_RETERR) {
745                                 cmn_err(CE_WARN,
746                                         "XFS: alignment check 1 failed");
747                                 return XFS_ERROR(EINVAL);
748                         }
749                         mp->m_dalign = mp->m_swidth = 0;
750                 } else {
751                         /*
752                          * Convert the stripe unit and width to FSBs.
753                          */
754                         mp->m_dalign = XFS_BB_TO_FSBT(mp, mp->m_dalign);
755                         if (mp->m_dalign && (sbp->sb_agblocks % mp->m_dalign)) {
756                                 if (mp->m_flags & XFS_MOUNT_RETERR) {
757                                         return XFS_ERROR(EINVAL);
758                                 }
759                                 xfs_fs_cmn_err(CE_WARN, mp,
760 "stripe alignment turned off: sunit(%d)/swidth(%d) incompatible with agsize(%d)",
761                                         mp->m_dalign, mp->m_swidth,
762                                         sbp->sb_agblocks);
763
764                                 mp->m_dalign = 0;
765                                 mp->m_swidth = 0;
766                         } else if (mp->m_dalign) {
767                                 mp->m_swidth = XFS_BB_TO_FSBT(mp, mp->m_swidth);
768                         } else {
769                                 if (mp->m_flags & XFS_MOUNT_RETERR) {
770                                         xfs_fs_cmn_err(CE_WARN, mp,
771 "stripe alignment turned off: sunit(%d) less than bsize(%d)",
772                                                 mp->m_dalign,
773                                                 mp->m_blockmask +1);
774                                         return XFS_ERROR(EINVAL);
775                                 }
776                                 mp->m_swidth = 0;
777                         }
778                 }
779
780                 /*
781                  * Update superblock with new values
782                  * and log changes
783                  */
784                 if (xfs_sb_version_hasdalign(sbp)) {
785                         if (sbp->sb_unit != mp->m_dalign) {
786                                 sbp->sb_unit = mp->m_dalign;
787                                 *update_flags |= XFS_SB_UNIT;
788                         }
789                         if (sbp->sb_width != mp->m_swidth) {
790                                 sbp->sb_width = mp->m_swidth;
791                                 *update_flags |= XFS_SB_WIDTH;
792                         }
793                 }
794         } else if ((mp->m_flags & XFS_MOUNT_NOALIGN) != XFS_MOUNT_NOALIGN &&
795                     xfs_sb_version_hasdalign(&mp->m_sb)) {
796                         mp->m_dalign = sbp->sb_unit;
797                         mp->m_swidth = sbp->sb_width;
798         }
799
800         return 0;
801 }
802
803 /*
804  * Set the maximum inode count for this filesystem
805  */
806 STATIC void
807 xfs_set_maxicount(xfs_mount_t *mp)
808 {
809         xfs_sb_t        *sbp = &(mp->m_sb);
810         __uint64_t      icount;
811
812         if (sbp->sb_imax_pct) {
813                 /*
814                  * Make sure the maximum inode count is a multiple
815                  * of the units we allocate inodes in.
816                  */
817                 icount = sbp->sb_dblocks * sbp->sb_imax_pct;
818                 do_div(icount, 100);
819                 do_div(icount, mp->m_ialloc_blks);
820                 mp->m_maxicount = (icount * mp->m_ialloc_blks)  <<
821                                    sbp->sb_inopblog;
822         } else {
823                 mp->m_maxicount = 0;
824         }
825 }
826
827 /*
828  * Set the default minimum read and write sizes unless
829  * already specified in a mount option.
830  * We use smaller I/O sizes when the file system
831  * is being used for NFS service (wsync mount option).
832  */
833 STATIC void
834 xfs_set_rw_sizes(xfs_mount_t *mp)
835 {
836         xfs_sb_t        *sbp = &(mp->m_sb);
837         int             readio_log, writeio_log;
838
839         if (!(mp->m_flags & XFS_MOUNT_DFLT_IOSIZE)) {
840                 if (mp->m_flags & XFS_MOUNT_WSYNC) {
841                         readio_log = XFS_WSYNC_READIO_LOG;
842                         writeio_log = XFS_WSYNC_WRITEIO_LOG;
843                 } else {
844                         readio_log = XFS_READIO_LOG_LARGE;
845                         writeio_log = XFS_WRITEIO_LOG_LARGE;
846                 }
847         } else {
848                 readio_log = mp->m_readio_log;
849                 writeio_log = mp->m_writeio_log;
850         }
851
852         if (sbp->sb_blocklog > readio_log) {
853                 mp->m_readio_log = sbp->sb_blocklog;
854         } else {
855                 mp->m_readio_log = readio_log;
856         }
857         mp->m_readio_blocks = 1 << (mp->m_readio_log - sbp->sb_blocklog);
858         if (sbp->sb_blocklog > writeio_log) {
859                 mp->m_writeio_log = sbp->sb_blocklog;
860         } else {
861                 mp->m_writeio_log = writeio_log;
862         }
863         mp->m_writeio_blocks = 1 << (mp->m_writeio_log - sbp->sb_blocklog);
864 }
865
866 /*
867  * Set whether we're using inode alignment.
868  */
869 STATIC void
870 xfs_set_inoalignment(xfs_mount_t *mp)
871 {
872         if (xfs_sb_version_hasalign(&mp->m_sb) &&
873             mp->m_sb.sb_inoalignmt >=
874             XFS_B_TO_FSBT(mp, mp->m_inode_cluster_size))
875                 mp->m_inoalign_mask = mp->m_sb.sb_inoalignmt - 1;
876         else
877                 mp->m_inoalign_mask = 0;
878         /*
879          * If we are using stripe alignment, check whether
880          * the stripe unit is a multiple of the inode alignment
881          */
882         if (mp->m_dalign && mp->m_inoalign_mask &&
883             !(mp->m_dalign & mp->m_inoalign_mask))
884                 mp->m_sinoalign = mp->m_dalign;
885         else
886                 mp->m_sinoalign = 0;
887 }
888
889 /*
890  * Check that the data (and log if separate) are an ok size.
891  */
892 STATIC int
893 xfs_check_sizes(xfs_mount_t *mp, int mfsi_flags)
894 {
895         xfs_buf_t       *bp;
896         xfs_daddr_t     d;
897         int             error;
898
899         d = (xfs_daddr_t)XFS_FSB_TO_BB(mp, mp->m_sb.sb_dblocks);
900         if (XFS_BB_TO_FSB(mp, d) != mp->m_sb.sb_dblocks) {
901                 cmn_err(CE_WARN, "XFS: size check 1 failed");
902                 return XFS_ERROR(E2BIG);
903         }
904         error = xfs_read_buf(mp, mp->m_ddev_targp,
905                              d - XFS_FSS_TO_BB(mp, 1),
906                              XFS_FSS_TO_BB(mp, 1), 0, &bp);
907         if (!error) {
908                 xfs_buf_relse(bp);
909         } else {
910                 cmn_err(CE_WARN, "XFS: size check 2 failed");
911                 if (error == ENOSPC)
912                         error = XFS_ERROR(E2BIG);
913                 return error;
914         }
915
916         if (((mfsi_flags & XFS_MFSI_CLIENT) == 0) &&
917             mp->m_logdev_targp != mp->m_ddev_targp) {
918                 d = (xfs_daddr_t)XFS_FSB_TO_BB(mp, mp->m_sb.sb_logblocks);
919                 if (XFS_BB_TO_FSB(mp, d) != mp->m_sb.sb_logblocks) {
920                         cmn_err(CE_WARN, "XFS: size check 3 failed");
921                         return XFS_ERROR(E2BIG);
922                 }
923                 error = xfs_read_buf(mp, mp->m_logdev_targp,
924                                      d - XFS_FSB_TO_BB(mp, 1),
925                                      XFS_FSB_TO_BB(mp, 1), 0, &bp);
926                 if (!error) {
927                         xfs_buf_relse(bp);
928                 } else {
929                         cmn_err(CE_WARN, "XFS: size check 3 failed");
930                         if (error == ENOSPC)
931                                 error = XFS_ERROR(E2BIG);
932                         return error;
933                 }
934         }
935         return 0;
936 }
937
938 /*
939  * xfs_mountfs
940  *
941  * This function does the following on an initial mount of a file system:
942  *      - reads the superblock from disk and init the mount struct
943  *      - if we're a 32-bit kernel, do a size check on the superblock
944  *              so we don't mount terabyte filesystems
945  *      - init mount struct realtime fields
946  *      - allocate inode hash table for fs
947  *      - init directory manager
948  *      - perform recovery and init the log manager
949  */
950 int
951 xfs_mountfs(
952         xfs_mount_t     *mp,
953         int             mfsi_flags)
954 {
955         xfs_sb_t        *sbp = &(mp->m_sb);
956         xfs_inode_t     *rip;
957         bhv_vnode_t     *rvp = NULL;
958         __uint64_t      resblks;
959         __int64_t       update_flags = 0LL;
960         uint            quotamount, quotaflags;
961         int             agno;
962         int             uuid_mounted = 0;
963         int             error = 0;
964
965         if (mp->m_sb_bp == NULL) {
966                 error = xfs_readsb(mp, mfsi_flags);
967                 if (error)
968                         return error;
969         }
970         xfs_mount_common(mp, sbp);
971
972         /*
973          * Check if sb_agblocks is aligned at stripe boundary
974          * If sb_agblocks is NOT aligned turn off m_dalign since
975          * allocator alignment is within an ag, therefore ag has
976          * to be aligned at stripe boundary.
977          */
978         error = xfs_update_alignment(mp, mfsi_flags, &update_flags);
979         if (error)
980                 goto error1;
981
982         xfs_alloc_compute_maxlevels(mp);
983         xfs_bmap_compute_maxlevels(mp, XFS_DATA_FORK);
984         xfs_bmap_compute_maxlevels(mp, XFS_ATTR_FORK);
985         xfs_ialloc_compute_maxlevels(mp);
986
987         xfs_set_maxicount(mp);
988
989         mp->m_maxioffset = xfs_max_file_offset(sbp->sb_blocklog);
990
991         /*
992          * XFS uses the uuid from the superblock as the unique
993          * identifier for fsid.  We can not use the uuid from the volume
994          * since a single partition filesystem is identical to a single
995          * partition volume/filesystem.
996          */
997         if ((mfsi_flags & XFS_MFSI_SECOND) == 0 &&
998             (mp->m_flags & XFS_MOUNT_NOUUID) == 0) {
999                 if (xfs_uuid_mount(mp)) {
1000                         error = XFS_ERROR(EINVAL);
1001                         goto error1;
1002                 }
1003                 uuid_mounted=1;
1004         }
1005
1006         /*
1007          * Set the minimum read and write sizes
1008          */
1009         xfs_set_rw_sizes(mp);
1010
1011         /*
1012          * Set the inode cluster size.
1013          * This may still be overridden by the file system
1014          * block size if it is larger than the chosen cluster size.
1015          */
1016         mp->m_inode_cluster_size = XFS_INODE_BIG_CLUSTER_SIZE;
1017
1018         /*
1019          * Set inode alignment fields
1020          */
1021         xfs_set_inoalignment(mp);
1022
1023         /*
1024          * Check that the data (and log if separate) are an ok size.
1025          */
1026         error = xfs_check_sizes(mp, mfsi_flags);
1027         if (error)
1028                 goto error1;
1029
1030         /*
1031          * Initialize realtime fields in the mount structure
1032          */
1033         error = xfs_rtmount_init(mp);
1034         if (error) {
1035                 cmn_err(CE_WARN, "XFS: RT mount failed");
1036                 goto error1;
1037         }
1038
1039         /*
1040          * For client case we are done now
1041          */
1042         if (mfsi_flags & XFS_MFSI_CLIENT) {
1043                 return 0;
1044         }
1045
1046         /*
1047          *  Copies the low order bits of the timestamp and the randomly
1048          *  set "sequence" number out of a UUID.
1049          */
1050         uuid_getnodeuniq(&sbp->sb_uuid, mp->m_fixedfsid);
1051
1052         mp->m_dmevmask = 0;     /* not persistent; set after each mount */
1053
1054         xfs_dir_mount(mp);
1055
1056         /*
1057          * Initialize the attribute manager's entries.
1058          */
1059         mp->m_attr_magicpct = (mp->m_sb.sb_blocksize * 37) / 100;
1060
1061         /*
1062          * Initialize the precomputed transaction reservations values.
1063          */
1064         xfs_trans_init(mp);
1065
1066         /*
1067          * Allocate and initialize the per-ag data.
1068          */
1069         init_rwsem(&mp->m_peraglock);
1070         mp->m_perag =
1071                 kmem_zalloc(sbp->sb_agcount * sizeof(xfs_perag_t), KM_SLEEP);
1072
1073         mp->m_maxagi = xfs_initialize_perag(mp, sbp->sb_agcount);
1074
1075         /*
1076          * log's mount-time initialization. Perform 1st part recovery if needed
1077          */
1078         if (likely(sbp->sb_logblocks > 0)) {    /* check for volume case */
1079                 error = xfs_log_mount(mp, mp->m_logdev_targp,
1080                                       XFS_FSB_TO_DADDR(mp, sbp->sb_logstart),
1081                                       XFS_FSB_TO_BB(mp, sbp->sb_logblocks));
1082                 if (error) {
1083                         cmn_err(CE_WARN, "XFS: log mount failed");
1084                         goto error2;
1085                 }
1086         } else {        /* No log has been defined */
1087                 cmn_err(CE_WARN, "XFS: no log defined");
1088                 XFS_ERROR_REPORT("xfs_mountfs_int(1)", XFS_ERRLEVEL_LOW, mp);
1089                 error = XFS_ERROR(EFSCORRUPTED);
1090                 goto error2;
1091         }
1092
1093         /*
1094          * Now the log is mounted, we know if it was an unclean shutdown or
1095          * not. If it was, with the first phase of recovery has completed, we
1096          * have consistent AG blocks on disk. We have not recovered EFIs yet,
1097          * but they are recovered transactionally in the second recovery phase
1098          * later.
1099          *
1100          * Hence we can safely re-initialise incore superblock counters from
1101          * the per-ag data. These may not be correct if the filesystem was not
1102          * cleanly unmounted, so we need to wait for recovery to finish before
1103          * doing this.
1104          *
1105          * If the filesystem was cleanly unmounted, then we can trust the
1106          * values in the superblock to be correct and we don't need to do
1107          * anything here.
1108          *
1109          * If we are currently making the filesystem, the initialisation will
1110          * fail as the perag data is in an undefined state.
1111          */
1112
1113         if (xfs_sb_version_haslazysbcount(&mp->m_sb) &&
1114             !XFS_LAST_UNMOUNT_WAS_CLEAN(mp) &&
1115              !mp->m_sb.sb_inprogress) {
1116                 error = xfs_initialize_perag_data(mp, sbp->sb_agcount);
1117                 if (error) {
1118                         goto error2;
1119                 }
1120         }
1121         /*
1122          * Get and sanity-check the root inode.
1123          * Save the pointer to it in the mount structure.
1124          */
1125         error = xfs_iget(mp, NULL, sbp->sb_rootino, 0, XFS_ILOCK_EXCL, &rip, 0);
1126         if (error) {
1127                 cmn_err(CE_WARN, "XFS: failed to read root inode");
1128                 goto error3;
1129         }
1130
1131         ASSERT(rip != NULL);
1132         rvp = XFS_ITOV(rip);
1133
1134         if (unlikely((rip->i_d.di_mode & S_IFMT) != S_IFDIR)) {
1135                 cmn_err(CE_WARN, "XFS: corrupted root inode");
1136                 cmn_err(CE_WARN, "Device %s - root %llu is not a directory",
1137                         XFS_BUFTARG_NAME(mp->m_ddev_targp),
1138                         (unsigned long long)rip->i_ino);
1139                 xfs_iunlock(rip, XFS_ILOCK_EXCL);
1140                 XFS_ERROR_REPORT("xfs_mountfs_int(2)", XFS_ERRLEVEL_LOW,
1141                                  mp);
1142                 error = XFS_ERROR(EFSCORRUPTED);
1143                 goto error4;
1144         }
1145         mp->m_rootip = rip;     /* save it */
1146
1147         xfs_iunlock(rip, XFS_ILOCK_EXCL);
1148
1149         /*
1150          * Initialize realtime inode pointers in the mount structure
1151          */
1152         error = xfs_rtmount_inodes(mp);
1153         if (error) {
1154                 /*
1155                  * Free up the root inode.
1156                  */
1157                 cmn_err(CE_WARN, "XFS: failed to read RT inodes");
1158                 goto error4;
1159         }
1160
1161         /*
1162          * If fs is not mounted readonly, then update the superblock
1163          * unit and width changes.
1164          */
1165         if (update_flags && !(mp->m_flags & XFS_MOUNT_RDONLY))
1166                 xfs_mount_log_sbunit(mp, update_flags);
1167
1168         /*
1169          * Initialise the XFS quota management subsystem for this mount
1170          */
1171         error = XFS_QM_INIT(mp, &quotamount, &quotaflags);
1172         if (error)
1173                 goto error4;
1174
1175         /*
1176          * Finish recovering the file system.  This part needed to be
1177          * delayed until after the root and real-time bitmap inodes
1178          * were consistently read in.
1179          */
1180         error = xfs_log_mount_finish(mp, mfsi_flags);
1181         if (error) {
1182                 cmn_err(CE_WARN, "XFS: log mount finish failed");
1183                 goto error4;
1184         }
1185
1186         /*
1187          * Complete the quota initialisation, post-log-replay component.
1188          */
1189         error = XFS_QM_MOUNT(mp, quotamount, quotaflags, mfsi_flags);
1190         if (error)
1191                 goto error4;
1192
1193         /*
1194          * Now we are mounted, reserve a small amount of unused space for
1195          * privileged transactions. This is needed so that transaction
1196          * space required for critical operations can dip into this pool
1197          * when at ENOSPC. This is needed for operations like create with
1198          * attr, unwritten extent conversion at ENOSPC, etc. Data allocations
1199          * are not allowed to use this reserved space.
1200          *
1201          * We default to 5% or 1024 fsbs of space reserved, whichever is smaller.
1202          * This may drive us straight to ENOSPC on mount, but that implies
1203          * we were already there on the last unmount.
1204          */
1205         resblks = mp->m_sb.sb_dblocks;
1206         do_div(resblks, 20);
1207         resblks = min_t(__uint64_t, resblks, 1024);
1208         xfs_reserve_blocks(mp, &resblks, NULL);
1209
1210         return 0;
1211
1212  error4:
1213         /*
1214          * Free up the root inode.
1215          */
1216         VN_RELE(rvp);
1217  error3:
1218         xfs_log_unmount_dealloc(mp);
1219  error2:
1220         for (agno = 0; agno < sbp->sb_agcount; agno++)
1221                 if (mp->m_perag[agno].pagb_list)
1222                         kmem_free(mp->m_perag[agno].pagb_list,
1223                           sizeof(xfs_perag_busy_t) * XFS_PAGB_NUM_SLOTS);
1224         kmem_free(mp->m_perag, sbp->sb_agcount * sizeof(xfs_perag_t));
1225         mp->m_perag = NULL;
1226         /* FALLTHROUGH */
1227  error1:
1228         if (uuid_mounted)
1229                 xfs_uuid_unmount(mp);
1230         xfs_freesb(mp);
1231         return error;
1232 }
1233
1234 /*
1235  * xfs_unmountfs
1236  *
1237  * This flushes out the inodes,dquots and the superblock, unmounts the
1238  * log and makes sure that incore structures are freed.
1239  */
1240 int
1241 xfs_unmountfs(xfs_mount_t *mp, struct cred *cr)
1242 {
1243         __uint64_t      resblks;
1244
1245         /*
1246          * We can potentially deadlock here if we have an inode cluster
1247          * that has been freed has it's buffer still pinned in memory because
1248          * the transaction is still sitting in a iclog. The stale inodes
1249          * on that buffer will have their flush locks held until the
1250          * transaction hits the disk and the callbacks run. the inode
1251          * flush takes the flush lock unconditionally and with nothing to
1252          * push out the iclog we will never get that unlocked. hence we
1253          * need to force the log first.
1254          */
1255         xfs_log_force(mp, (xfs_lsn_t)0, XFS_LOG_FORCE | XFS_LOG_SYNC);
1256         xfs_iflush_all(mp);
1257
1258         XFS_QM_DQPURGEALL(mp, XFS_QMOPT_QUOTALL | XFS_QMOPT_UMOUNTING);
1259
1260         /*
1261          * Flush out the log synchronously so that we know for sure
1262          * that nothing is pinned.  This is important because bflush()
1263          * will skip pinned buffers.
1264          */
1265         xfs_log_force(mp, (xfs_lsn_t)0, XFS_LOG_FORCE | XFS_LOG_SYNC);
1266
1267         xfs_binval(mp->m_ddev_targp);
1268         if (mp->m_rtdev_targp) {
1269                 xfs_binval(mp->m_rtdev_targp);
1270         }
1271
1272         /*
1273          * Unreserve any blocks we have so that when we unmount we don't account
1274          * the reserved free space as used. This is really only necessary for
1275          * lazy superblock counting because it trusts the incore superblock
1276          * counters to be aboslutely correct on clean unmount.
1277          *
1278          * We don't bother correcting this elsewhere for lazy superblock
1279          * counting because on mount of an unclean filesystem we reconstruct the
1280          * correct counter value and this is irrelevant.
1281          *
1282          * For non-lazy counter filesystems, this doesn't matter at all because
1283          * we only every apply deltas to the superblock and hence the incore
1284          * value does not matter....
1285          */
1286         resblks = 0;
1287         xfs_reserve_blocks(mp, &resblks, NULL);
1288
1289         xfs_log_sbcount(mp, 1);
1290         xfs_unmountfs_writesb(mp);
1291         xfs_unmountfs_wait(mp);                 /* wait for async bufs */
1292         xfs_log_unmount(mp);                    /* Done! No more fs ops. */
1293
1294         xfs_freesb(mp);
1295
1296         /*
1297          * All inodes from this mount point should be freed.
1298          */
1299         ASSERT(mp->m_inodes == NULL);
1300
1301         xfs_unmountfs_close(mp, cr);
1302         if ((mp->m_flags & XFS_MOUNT_NOUUID) == 0)
1303                 xfs_uuid_unmount(mp);
1304
1305 #if defined(DEBUG) || defined(INDUCE_IO_ERROR)
1306         xfs_errortag_clearall(mp, 0);
1307 #endif
1308         xfs_mount_free(mp);
1309         return 0;
1310 }
1311
1312 void
1313 xfs_unmountfs_close(xfs_mount_t *mp, struct cred *cr)
1314 {
1315         if (mp->m_logdev_targp && mp->m_logdev_targp != mp->m_ddev_targp)
1316                 xfs_free_buftarg(mp->m_logdev_targp, 1);
1317         if (mp->m_rtdev_targp)
1318                 xfs_free_buftarg(mp->m_rtdev_targp, 1);
1319         xfs_free_buftarg(mp->m_ddev_targp, 0);
1320 }
1321
1322 STATIC void
1323 xfs_unmountfs_wait(xfs_mount_t *mp)
1324 {
1325         if (mp->m_logdev_targp != mp->m_ddev_targp)
1326                 xfs_wait_buftarg(mp->m_logdev_targp);
1327         if (mp->m_rtdev_targp)
1328                 xfs_wait_buftarg(mp->m_rtdev_targp);
1329         xfs_wait_buftarg(mp->m_ddev_targp);
1330 }
1331
1332 int
1333 xfs_fs_writable(xfs_mount_t *mp)
1334 {
1335         return !(xfs_test_for_freeze(mp) || XFS_FORCED_SHUTDOWN(mp) ||
1336                 (mp->m_flags & XFS_MOUNT_RDONLY));
1337 }
1338
1339 /*
1340  * xfs_log_sbcount
1341  *
1342  * Called either periodically to keep the on disk superblock values
1343  * roughly up to date or from unmount to make sure the values are
1344  * correct on a clean unmount.
1345  *
1346  * Note this code can be called during the process of freezing, so
1347  * we may need to use the transaction allocator which does not not
1348  * block when the transaction subsystem is in its frozen state.
1349  */
1350 int
1351 xfs_log_sbcount(
1352         xfs_mount_t     *mp,
1353         uint            sync)
1354 {
1355         xfs_trans_t     *tp;
1356         int             error;
1357
1358         if (!xfs_fs_writable(mp))
1359                 return 0;
1360
1361         xfs_icsb_sync_counters(mp);
1362
1363         /*
1364          * we don't need to do this if we are updating the superblock
1365          * counters on every modification.
1366          */
1367         if (!xfs_sb_version_haslazysbcount(&mp->m_sb))
1368                 return 0;
1369
1370         tp = _xfs_trans_alloc(mp, XFS_TRANS_SB_COUNT);
1371         error = xfs_trans_reserve(tp, 0, mp->m_sb.sb_sectsize + 128, 0, 0,
1372                                         XFS_DEFAULT_LOG_COUNT);
1373         if (error) {
1374                 xfs_trans_cancel(tp, 0);
1375                 return error;
1376         }
1377
1378         xfs_mod_sb(tp, XFS_SB_IFREE | XFS_SB_ICOUNT | XFS_SB_FDBLOCKS);
1379         if (sync)
1380                 xfs_trans_set_sync(tp);
1381         xfs_trans_commit(tp, 0);
1382
1383         return 0;
1384 }
1385
1386 STATIC void
1387 xfs_mark_shared_ro(
1388         xfs_mount_t     *mp,
1389         xfs_buf_t       *bp)
1390 {
1391         xfs_dsb_t       *sb = XFS_BUF_TO_SBP(bp);
1392         __uint16_t      version;
1393
1394         if (!(sb->sb_flags & XFS_SBF_READONLY))
1395                 sb->sb_flags |= XFS_SBF_READONLY;
1396
1397         version = be16_to_cpu(sb->sb_versionnum);
1398         if ((version & XFS_SB_VERSION_NUMBITS) != XFS_SB_VERSION_4 ||
1399             !(version & XFS_SB_VERSION_SHAREDBIT))
1400                 version |= XFS_SB_VERSION_SHAREDBIT;
1401         sb->sb_versionnum = cpu_to_be16(version);
1402 }
1403
1404 int
1405 xfs_unmountfs_writesb(xfs_mount_t *mp)
1406 {
1407         xfs_buf_t       *sbp;
1408         int             error = 0;
1409
1410         /*
1411          * skip superblock write if fs is read-only, or
1412          * if we are doing a forced umount.
1413          */
1414         if (!((mp->m_flags & XFS_MOUNT_RDONLY) ||
1415                 XFS_FORCED_SHUTDOWN(mp))) {
1416
1417                 sbp = xfs_getsb(mp, 0);
1418
1419                 /*
1420                  * mark shared-readonly if desired
1421                  */
1422                 if (mp->m_mk_sharedro)
1423                         xfs_mark_shared_ro(mp, sbp);
1424
1425                 XFS_BUF_UNDONE(sbp);
1426                 XFS_BUF_UNREAD(sbp);
1427                 XFS_BUF_UNDELAYWRITE(sbp);
1428                 XFS_BUF_WRITE(sbp);
1429                 XFS_BUF_UNASYNC(sbp);
1430                 ASSERT(XFS_BUF_TARGET(sbp) == mp->m_ddev_targp);
1431                 xfsbdstrat(mp, sbp);
1432                 /* Nevermind errors we might get here. */
1433                 error = xfs_iowait(sbp);
1434                 if (error)
1435                         xfs_ioerror_alert("xfs_unmountfs_writesb",
1436                                           mp, sbp, XFS_BUF_ADDR(sbp));
1437                 if (error && mp->m_mk_sharedro)
1438                         xfs_fs_cmn_err(CE_ALERT, mp, "Superblock write error detected while unmounting.  Filesystem may not be marked shared readonly");
1439                 xfs_buf_relse(sbp);
1440         }
1441         return error;
1442 }
1443
1444 /*
1445  * xfs_mod_sb() can be used to copy arbitrary changes to the
1446  * in-core superblock into the superblock buffer to be logged.
1447  * It does not provide the higher level of locking that is
1448  * needed to protect the in-core superblock from concurrent
1449  * access.
1450  */
1451 void
1452 xfs_mod_sb(xfs_trans_t *tp, __int64_t fields)
1453 {
1454         xfs_buf_t       *bp;
1455         int             first;
1456         int             last;
1457         xfs_mount_t     *mp;
1458         xfs_sb_field_t  f;
1459
1460         ASSERT(fields);
1461         if (!fields)
1462                 return;
1463         mp = tp->t_mountp;
1464         bp = xfs_trans_getsb(tp, mp, 0);
1465         first = sizeof(xfs_sb_t);
1466         last = 0;
1467
1468         /* translate/copy */
1469
1470         xfs_sb_to_disk(XFS_BUF_TO_SBP(bp), &mp->m_sb, fields);
1471
1472         /* find modified range */
1473
1474         f = (xfs_sb_field_t)xfs_lowbit64((__uint64_t)fields);
1475         ASSERT((1LL << f) & XFS_SB_MOD_BITS);
1476         first = xfs_sb_info[f].offset;
1477
1478         f = (xfs_sb_field_t)xfs_highbit64((__uint64_t)fields);
1479         ASSERT((1LL << f) & XFS_SB_MOD_BITS);
1480         last = xfs_sb_info[f + 1].offset - 1;
1481
1482         xfs_trans_log_buf(tp, bp, first, last);
1483 }
1484
1485
1486 /*
1487  * xfs_mod_incore_sb_unlocked() is a utility routine common used to apply
1488  * a delta to a specified field in the in-core superblock.  Simply
1489  * switch on the field indicated and apply the delta to that field.
1490  * Fields are not allowed to dip below zero, so if the delta would
1491  * do this do not apply it and return EINVAL.
1492  *
1493  * The m_sb_lock must be held when this routine is called.
1494  */
1495 int
1496 xfs_mod_incore_sb_unlocked(
1497         xfs_mount_t     *mp,
1498         xfs_sb_field_t  field,
1499         int64_t         delta,
1500         int             rsvd)
1501 {
1502         int             scounter;       /* short counter for 32 bit fields */
1503         long long       lcounter;       /* long counter for 64 bit fields */
1504         long long       res_used, rem;
1505
1506         /*
1507          * With the in-core superblock spin lock held, switch
1508          * on the indicated field.  Apply the delta to the
1509          * proper field.  If the fields value would dip below
1510          * 0, then do not apply the delta and return EINVAL.
1511          */
1512         switch (field) {
1513         case XFS_SBS_ICOUNT:
1514                 lcounter = (long long)mp->m_sb.sb_icount;
1515                 lcounter += delta;
1516                 if (lcounter < 0) {
1517                         ASSERT(0);
1518                         return XFS_ERROR(EINVAL);
1519                 }
1520                 mp->m_sb.sb_icount = lcounter;
1521                 return 0;
1522         case XFS_SBS_IFREE:
1523                 lcounter = (long long)mp->m_sb.sb_ifree;
1524                 lcounter += delta;
1525                 if (lcounter < 0) {
1526                         ASSERT(0);
1527                         return XFS_ERROR(EINVAL);
1528                 }
1529                 mp->m_sb.sb_ifree = lcounter;
1530                 return 0;
1531         case XFS_SBS_FDBLOCKS:
1532                 lcounter = (long long)
1533                         mp->m_sb.sb_fdblocks - XFS_ALLOC_SET_ASIDE(mp);
1534                 res_used = (long long)(mp->m_resblks - mp->m_resblks_avail);
1535
1536                 if (delta > 0) {                /* Putting blocks back */
1537                         if (res_used > delta) {
1538                                 mp->m_resblks_avail += delta;
1539                         } else {
1540                                 rem = delta - res_used;
1541                                 mp->m_resblks_avail = mp->m_resblks;
1542                                 lcounter += rem;
1543                         }
1544                 } else {                                /* Taking blocks away */
1545
1546                         lcounter += delta;
1547
1548                 /*
1549                  * If were out of blocks, use any available reserved blocks if
1550                  * were allowed to.
1551                  */
1552
1553                         if (lcounter < 0) {
1554                                 if (rsvd) {
1555                                         lcounter = (long long)mp->m_resblks_avail + delta;
1556                                         if (lcounter < 0) {
1557                                                 return XFS_ERROR(ENOSPC);
1558                                         }
1559                                         mp->m_resblks_avail = lcounter;
1560                                         return 0;
1561                                 } else {        /* not reserved */
1562                                         return XFS_ERROR(ENOSPC);
1563                                 }
1564                         }
1565                 }
1566
1567                 mp->m_sb.sb_fdblocks = lcounter + XFS_ALLOC_SET_ASIDE(mp);
1568                 return 0;
1569         case XFS_SBS_FREXTENTS:
1570                 lcounter = (long long)mp->m_sb.sb_frextents;
1571                 lcounter += delta;
1572                 if (lcounter < 0) {
1573                         return XFS_ERROR(ENOSPC);
1574                 }
1575                 mp->m_sb.sb_frextents = lcounter;
1576                 return 0;
1577         case XFS_SBS_DBLOCKS:
1578                 lcounter = (long long)mp->m_sb.sb_dblocks;
1579                 lcounter += delta;
1580                 if (lcounter < 0) {
1581                         ASSERT(0);
1582                         return XFS_ERROR(EINVAL);
1583                 }
1584                 mp->m_sb.sb_dblocks = lcounter;
1585                 return 0;
1586         case XFS_SBS_AGCOUNT:
1587                 scounter = mp->m_sb.sb_agcount;
1588                 scounter += delta;
1589                 if (scounter < 0) {
1590                         ASSERT(0);
1591                         return XFS_ERROR(EINVAL);
1592                 }
1593                 mp->m_sb.sb_agcount = scounter;
1594                 return 0;
1595         case XFS_SBS_IMAX_PCT:
1596                 scounter = mp->m_sb.sb_imax_pct;
1597                 scounter += delta;
1598                 if (scounter < 0) {
1599                         ASSERT(0);
1600                         return XFS_ERROR(EINVAL);
1601                 }
1602                 mp->m_sb.sb_imax_pct = scounter;
1603                 return 0;
1604         case XFS_SBS_REXTSIZE:
1605                 scounter = mp->m_sb.sb_rextsize;
1606                 scounter += delta;
1607                 if (scounter < 0) {
1608                         ASSERT(0);
1609                         return XFS_ERROR(EINVAL);
1610                 }
1611                 mp->m_sb.sb_rextsize = scounter;
1612                 return 0;
1613         case XFS_SBS_RBMBLOCKS:
1614                 scounter = mp->m_sb.sb_rbmblocks;
1615                 scounter += delta;
1616                 if (scounter < 0) {
1617                         ASSERT(0);
1618                         return XFS_ERROR(EINVAL);
1619                 }
1620                 mp->m_sb.sb_rbmblocks = scounter;
1621                 return 0;
1622         case XFS_SBS_RBLOCKS:
1623                 lcounter = (long long)mp->m_sb.sb_rblocks;
1624                 lcounter += delta;
1625                 if (lcounter < 0) {
1626                         ASSERT(0);
1627                         return XFS_ERROR(EINVAL);
1628                 }
1629                 mp->m_sb.sb_rblocks = lcounter;
1630                 return 0;
1631         case XFS_SBS_REXTENTS:
1632                 lcounter = (long long)mp->m_sb.sb_rextents;
1633                 lcounter += delta;
1634                 if (lcounter < 0) {
1635                         ASSERT(0);
1636                         return XFS_ERROR(EINVAL);
1637                 }
1638                 mp->m_sb.sb_rextents = lcounter;
1639                 return 0;
1640         case XFS_SBS_REXTSLOG:
1641                 scounter = mp->m_sb.sb_rextslog;
1642                 scounter += delta;
1643                 if (scounter < 0) {
1644                         ASSERT(0);
1645                         return XFS_ERROR(EINVAL);
1646                 }
1647                 mp->m_sb.sb_rextslog = scounter;
1648                 return 0;
1649         default:
1650                 ASSERT(0);
1651                 return XFS_ERROR(EINVAL);
1652         }
1653 }
1654
1655 /*
1656  * xfs_mod_incore_sb() is used to change a field in the in-core
1657  * superblock structure by the specified delta.  This modification
1658  * is protected by the m_sb_lock.  Just use the xfs_mod_incore_sb_unlocked()
1659  * routine to do the work.
1660  */
1661 int
1662 xfs_mod_incore_sb(
1663         xfs_mount_t     *mp,
1664         xfs_sb_field_t  field,
1665         int64_t         delta,
1666         int             rsvd)
1667 {
1668         int     status;
1669
1670         /* check for per-cpu counters */
1671         switch (field) {
1672 #ifdef HAVE_PERCPU_SB
1673         case XFS_SBS_ICOUNT:
1674         case XFS_SBS_IFREE:
1675         case XFS_SBS_FDBLOCKS:
1676                 if (!(mp->m_flags & XFS_MOUNT_NO_PERCPU_SB)) {
1677                         status = xfs_icsb_modify_counters(mp, field,
1678                                                         delta, rsvd);
1679                         break;
1680                 }
1681                 /* FALLTHROUGH */
1682 #endif
1683         default:
1684                 spin_lock(&mp->m_sb_lock);
1685                 status = xfs_mod_incore_sb_unlocked(mp, field, delta, rsvd);
1686                 spin_unlock(&mp->m_sb_lock);
1687                 break;
1688         }
1689
1690         return status;
1691 }
1692
1693 /*
1694  * xfs_mod_incore_sb_batch() is used to change more than one field
1695  * in the in-core superblock structure at a time.  This modification
1696  * is protected by a lock internal to this module.  The fields and
1697  * changes to those fields are specified in the array of xfs_mod_sb
1698  * structures passed in.
1699  *
1700  * Either all of the specified deltas will be applied or none of
1701  * them will.  If any modified field dips below 0, then all modifications
1702  * will be backed out and EINVAL will be returned.
1703  */
1704 int
1705 xfs_mod_incore_sb_batch(xfs_mount_t *mp, xfs_mod_sb_t *msb, uint nmsb, int rsvd)
1706 {
1707         int             status=0;
1708         xfs_mod_sb_t    *msbp;
1709
1710         /*
1711          * Loop through the array of mod structures and apply each
1712          * individually.  If any fail, then back out all those
1713          * which have already been applied.  Do all of this within
1714          * the scope of the m_sb_lock so that all of the changes will
1715          * be atomic.
1716          */
1717         spin_lock(&mp->m_sb_lock);
1718         msbp = &msb[0];
1719         for (msbp = &msbp[0]; msbp < (msb + nmsb); msbp++) {
1720                 /*
1721                  * Apply the delta at index n.  If it fails, break
1722                  * from the loop so we'll fall into the undo loop
1723                  * below.
1724                  */
1725                 switch (msbp->msb_field) {
1726 #ifdef HAVE_PERCPU_SB
1727                 case XFS_SBS_ICOUNT:
1728                 case XFS_SBS_IFREE:
1729                 case XFS_SBS_FDBLOCKS:
1730                         if (!(mp->m_flags & XFS_MOUNT_NO_PERCPU_SB)) {
1731                                 spin_unlock(&mp->m_sb_lock);
1732                                 status = xfs_icsb_modify_counters(mp,
1733                                                         msbp->msb_field,
1734                                                         msbp->msb_delta, rsvd);
1735                                 spin_lock(&mp->m_sb_lock);
1736                                 break;
1737                         }
1738                         /* FALLTHROUGH */
1739 #endif
1740                 default:
1741                         status = xfs_mod_incore_sb_unlocked(mp,
1742                                                 msbp->msb_field,
1743                                                 msbp->msb_delta, rsvd);
1744                         break;
1745                 }
1746
1747                 if (status != 0) {
1748                         break;
1749                 }
1750         }
1751
1752         /*
1753          * If we didn't complete the loop above, then back out
1754          * any changes made to the superblock.  If you add code
1755          * between the loop above and here, make sure that you
1756          * preserve the value of status. Loop back until
1757          * we step below the beginning of the array.  Make sure
1758          * we don't touch anything back there.
1759          */
1760         if (status != 0) {
1761                 msbp--;
1762                 while (msbp >= msb) {
1763                         switch (msbp->msb_field) {
1764 #ifdef HAVE_PERCPU_SB
1765                         case XFS_SBS_ICOUNT:
1766                         case XFS_SBS_IFREE:
1767                         case XFS_SBS_FDBLOCKS:
1768                                 if (!(mp->m_flags & XFS_MOUNT_NO_PERCPU_SB)) {
1769                                         spin_unlock(&mp->m_sb_lock);
1770                                         status = xfs_icsb_modify_counters(mp,
1771                                                         msbp->msb_field,
1772                                                         -(msbp->msb_delta),
1773                                                         rsvd);
1774                                         spin_lock(&mp->m_sb_lock);
1775                                         break;
1776                                 }
1777                                 /* FALLTHROUGH */
1778 #endif
1779                         default:
1780                                 status = xfs_mod_incore_sb_unlocked(mp,
1781                                                         msbp->msb_field,
1782                                                         -(msbp->msb_delta),
1783                                                         rsvd);
1784                                 break;
1785                         }
1786                         ASSERT(status == 0);
1787                         msbp--;
1788                 }
1789         }
1790         spin_unlock(&mp->m_sb_lock);
1791         return status;
1792 }
1793
1794 /*
1795  * xfs_getsb() is called to obtain the buffer for the superblock.
1796  * The buffer is returned locked and read in from disk.
1797  * The buffer should be released with a call to xfs_brelse().
1798  *
1799  * If the flags parameter is BUF_TRYLOCK, then we'll only return
1800  * the superblock buffer if it can be locked without sleeping.
1801  * If it can't then we'll return NULL.
1802  */
1803 xfs_buf_t *
1804 xfs_getsb(
1805         xfs_mount_t     *mp,
1806         int             flags)
1807 {
1808         xfs_buf_t       *bp;
1809
1810         ASSERT(mp->m_sb_bp != NULL);
1811         bp = mp->m_sb_bp;
1812         if (flags & XFS_BUF_TRYLOCK) {
1813                 if (!XFS_BUF_CPSEMA(bp)) {
1814                         return NULL;
1815                 }
1816         } else {
1817                 XFS_BUF_PSEMA(bp, PRIBIO);
1818         }
1819         XFS_BUF_HOLD(bp);
1820         ASSERT(XFS_BUF_ISDONE(bp));
1821         return bp;
1822 }
1823
1824 /*
1825  * Used to free the superblock along various error paths.
1826  */
1827 void
1828 xfs_freesb(
1829         xfs_mount_t     *mp)
1830 {
1831         xfs_buf_t       *bp;
1832
1833         /*
1834          * Use xfs_getsb() so that the buffer will be locked
1835          * when we call xfs_buf_relse().
1836          */
1837         bp = xfs_getsb(mp, 0);
1838         XFS_BUF_UNMANAGE(bp);
1839         xfs_buf_relse(bp);
1840         mp->m_sb_bp = NULL;
1841 }
1842
1843 /*
1844  * See if the UUID is unique among mounted XFS filesystems.
1845  * Mount fails if UUID is nil or a FS with the same UUID is already mounted.
1846  */
1847 STATIC int
1848 xfs_uuid_mount(
1849         xfs_mount_t     *mp)
1850 {
1851         if (uuid_is_nil(&mp->m_sb.sb_uuid)) {
1852                 cmn_err(CE_WARN,
1853                         "XFS: Filesystem %s has nil UUID - can't mount",
1854                         mp->m_fsname);
1855                 return -1;
1856         }
1857         if (!uuid_table_insert(&mp->m_sb.sb_uuid)) {
1858                 cmn_err(CE_WARN,
1859                         "XFS: Filesystem %s has duplicate UUID - can't mount",
1860                         mp->m_fsname);
1861                 return -1;
1862         }
1863         return 0;
1864 }
1865
1866 /*
1867  * Remove filesystem from the UUID table.
1868  */
1869 STATIC void
1870 xfs_uuid_unmount(
1871         xfs_mount_t     *mp)
1872 {
1873         uuid_table_remove(&mp->m_sb.sb_uuid);
1874 }
1875
1876 /*
1877  * Used to log changes to the superblock unit and width fields which could
1878  * be altered by the mount options. Only the first superblock is updated.
1879  */
1880 STATIC void
1881 xfs_mount_log_sbunit(
1882         xfs_mount_t     *mp,
1883         __int64_t       fields)
1884 {
1885         xfs_trans_t     *tp;
1886
1887         ASSERT(fields & (XFS_SB_UNIT|XFS_SB_WIDTH|XFS_SB_UUID));
1888
1889         tp = xfs_trans_alloc(mp, XFS_TRANS_SB_UNIT);
1890         if (xfs_trans_reserve(tp, 0, mp->m_sb.sb_sectsize + 128, 0, 0,
1891                                 XFS_DEFAULT_LOG_COUNT)) {
1892                 xfs_trans_cancel(tp, 0);
1893                 return;
1894         }
1895         xfs_mod_sb(tp, fields);
1896         xfs_trans_commit(tp, 0);
1897 }
1898
1899
1900 #ifdef HAVE_PERCPU_SB
1901 /*
1902  * Per-cpu incore superblock counters
1903  *
1904  * Simple concept, difficult implementation
1905  *
1906  * Basically, replace the incore superblock counters with a distributed per cpu
1907  * counter for contended fields (e.g.  free block count).
1908  *
1909  * Difficulties arise in that the incore sb is used for ENOSPC checking, and
1910  * hence needs to be accurately read when we are running low on space. Hence
1911  * there is a method to enable and disable the per-cpu counters based on how
1912  * much "stuff" is available in them.
1913  *
1914  * Basically, a counter is enabled if there is enough free resource to justify
1915  * running a per-cpu fast-path. If the per-cpu counter runs out (i.e. a local
1916  * ENOSPC), then we disable the counters to synchronise all callers and
1917  * re-distribute the available resources.
1918  *
1919  * If, once we redistributed the available resources, we still get a failure,
1920  * we disable the per-cpu counter and go through the slow path.
1921  *
1922  * The slow path is the current xfs_mod_incore_sb() function.  This means that
1923  * when we disable a per-cpu counter, we need to drain it's resources back to
1924  * the global superblock. We do this after disabling the counter to prevent
1925  * more threads from queueing up on the counter.
1926  *
1927  * Essentially, this means that we still need a lock in the fast path to enable
1928  * synchronisation between the global counters and the per-cpu counters. This
1929  * is not a problem because the lock will be local to a CPU almost all the time
1930  * and have little contention except when we get to ENOSPC conditions.
1931  *
1932  * Basically, this lock becomes a barrier that enables us to lock out the fast
1933  * path while we do things like enabling and disabling counters and
1934  * synchronising the counters.
1935  *
1936  * Locking rules:
1937  *
1938  *      1. m_sb_lock before picking up per-cpu locks
1939  *      2. per-cpu locks always picked up via for_each_online_cpu() order
1940  *      3. accurate counter sync requires m_sb_lock + per cpu locks
1941  *      4. modifying per-cpu counters requires holding per-cpu lock
1942  *      5. modifying global counters requires holding m_sb_lock
1943  *      6. enabling or disabling a counter requires holding the m_sb_lock 
1944  *         and _none_ of the per-cpu locks.
1945  *
1946  * Disabled counters are only ever re-enabled by a balance operation
1947  * that results in more free resources per CPU than a given threshold.
1948  * To ensure counters don't remain disabled, they are rebalanced when
1949  * the global resource goes above a higher threshold (i.e. some hysteresis
1950  * is present to prevent thrashing).
1951  */
1952
1953 #ifdef CONFIG_HOTPLUG_CPU
1954 /*
1955  * hot-plug CPU notifier support.
1956  *
1957  * We need a notifier per filesystem as we need to be able to identify
1958  * the filesystem to balance the counters out. This is achieved by
1959  * having a notifier block embedded in the xfs_mount_t and doing pointer
1960  * magic to get the mount pointer from the notifier block address.
1961  */
1962 STATIC int
1963 xfs_icsb_cpu_notify(
1964         struct notifier_block *nfb,
1965         unsigned long action,
1966         void *hcpu)
1967 {
1968         xfs_icsb_cnts_t *cntp;
1969         xfs_mount_t     *mp;
1970
1971         mp = (xfs_mount_t *)container_of(nfb, xfs_mount_t, m_icsb_notifier);
1972         cntp = (xfs_icsb_cnts_t *)
1973                         per_cpu_ptr(mp->m_sb_cnts, (unsigned long)hcpu);
1974         switch (action) {
1975         case CPU_UP_PREPARE:
1976         case CPU_UP_PREPARE_FROZEN:
1977                 /* Easy Case - initialize the area and locks, and
1978                  * then rebalance when online does everything else for us. */
1979                 memset(cntp, 0, sizeof(xfs_icsb_cnts_t));
1980                 break;
1981         case CPU_ONLINE:
1982         case CPU_ONLINE_FROZEN:
1983                 xfs_icsb_lock(mp);
1984                 xfs_icsb_balance_counter(mp, XFS_SBS_ICOUNT, 0, 0);
1985                 xfs_icsb_balance_counter(mp, XFS_SBS_IFREE, 0, 0);
1986                 xfs_icsb_balance_counter(mp, XFS_SBS_FDBLOCKS, 0, 0);
1987                 xfs_icsb_unlock(mp);
1988                 break;
1989         case CPU_DEAD:
1990         case CPU_DEAD_FROZEN:
1991                 /* Disable all the counters, then fold the dead cpu's
1992                  * count into the total on the global superblock and
1993                  * re-enable the counters. */
1994                 xfs_icsb_lock(mp);
1995                 spin_lock(&mp->m_sb_lock);
1996                 xfs_icsb_disable_counter(mp, XFS_SBS_ICOUNT);
1997                 xfs_icsb_disable_counter(mp, XFS_SBS_IFREE);
1998                 xfs_icsb_disable_counter(mp, XFS_SBS_FDBLOCKS);
1999
2000                 mp->m_sb.sb_icount += cntp->icsb_icount;
2001                 mp->m_sb.sb_ifree += cntp->icsb_ifree;
2002                 mp->m_sb.sb_fdblocks += cntp->icsb_fdblocks;
2003
2004                 memset(cntp, 0, sizeof(xfs_icsb_cnts_t));
2005
2006                 xfs_icsb_balance_counter(mp, XFS_SBS_ICOUNT,
2007                                          XFS_ICSB_SB_LOCKED, 0);
2008                 xfs_icsb_balance_counter(mp, XFS_SBS_IFREE,
2009                                          XFS_ICSB_SB_LOCKED, 0);
2010                 xfs_icsb_balance_counter(mp, XFS_SBS_FDBLOCKS,
2011                                          XFS_ICSB_SB_LOCKED, 0);
2012                 spin_unlock(&mp->m_sb_lock);
2013                 xfs_icsb_unlock(mp);
2014                 break;
2015         }
2016
2017         return NOTIFY_OK;
2018 }
2019 #endif /* CONFIG_HOTPLUG_CPU */
2020
2021 int
2022 xfs_icsb_init_counters(
2023         xfs_mount_t     *mp)
2024 {
2025         xfs_icsb_cnts_t *cntp;
2026         int             i;
2027
2028         mp->m_sb_cnts = alloc_percpu(xfs_icsb_cnts_t);
2029         if (mp->m_sb_cnts == NULL)
2030                 return -ENOMEM;
2031
2032 #ifdef CONFIG_HOTPLUG_CPU
2033         mp->m_icsb_notifier.notifier_call = xfs_icsb_cpu_notify;
2034         mp->m_icsb_notifier.priority = 0;
2035         register_hotcpu_notifier(&mp->m_icsb_notifier);
2036 #endif /* CONFIG_HOTPLUG_CPU */
2037
2038         for_each_online_cpu(i) {
2039                 cntp = (xfs_icsb_cnts_t *)per_cpu_ptr(mp->m_sb_cnts, i);
2040                 memset(cntp, 0, sizeof(xfs_icsb_cnts_t));
2041         }
2042
2043         mutex_init(&mp->m_icsb_mutex);
2044
2045         /*
2046          * start with all counters disabled so that the
2047          * initial balance kicks us off correctly
2048          */
2049         mp->m_icsb_counters = -1;
2050         return 0;
2051 }
2052
2053 void
2054 xfs_icsb_reinit_counters(
2055         xfs_mount_t     *mp)
2056 {
2057         xfs_icsb_lock(mp);
2058         /*
2059          * start with all counters disabled so that the
2060          * initial balance kicks us off correctly
2061          */
2062         mp->m_icsb_counters = -1;
2063         xfs_icsb_balance_counter(mp, XFS_SBS_ICOUNT, 0, 0);
2064         xfs_icsb_balance_counter(mp, XFS_SBS_IFREE, 0, 0);
2065         xfs_icsb_balance_counter(mp, XFS_SBS_FDBLOCKS, 0, 0);
2066         xfs_icsb_unlock(mp);
2067 }
2068
2069 STATIC void
2070 xfs_icsb_destroy_counters(
2071         xfs_mount_t     *mp)
2072 {
2073         if (mp->m_sb_cnts) {
2074                 unregister_hotcpu_notifier(&mp->m_icsb_notifier);
2075                 free_percpu(mp->m_sb_cnts);
2076         }
2077         mutex_destroy(&mp->m_icsb_mutex);
2078 }
2079
2080 STATIC_INLINE void
2081 xfs_icsb_lock_cntr(
2082         xfs_icsb_cnts_t *icsbp)
2083 {
2084         while (test_and_set_bit(XFS_ICSB_FLAG_LOCK, &icsbp->icsb_flags)) {
2085                 ndelay(1000);
2086         }
2087 }
2088
2089 STATIC_INLINE void
2090 xfs_icsb_unlock_cntr(
2091         xfs_icsb_cnts_t *icsbp)
2092 {
2093         clear_bit(XFS_ICSB_FLAG_LOCK, &icsbp->icsb_flags);
2094 }
2095
2096
2097 STATIC_INLINE void
2098 xfs_icsb_lock_all_counters(
2099         xfs_mount_t     *mp)
2100 {
2101         xfs_icsb_cnts_t *cntp;
2102         int             i;
2103
2104         for_each_online_cpu(i) {
2105                 cntp = (xfs_icsb_cnts_t *)per_cpu_ptr(mp->m_sb_cnts, i);
2106                 xfs_icsb_lock_cntr(cntp);
2107         }
2108 }
2109
2110 STATIC_INLINE void
2111 xfs_icsb_unlock_all_counters(
2112         xfs_mount_t     *mp)
2113 {
2114         xfs_icsb_cnts_t *cntp;
2115         int             i;
2116
2117         for_each_online_cpu(i) {
2118                 cntp = (xfs_icsb_cnts_t *)per_cpu_ptr(mp->m_sb_cnts, i);
2119                 xfs_icsb_unlock_cntr(cntp);
2120         }
2121 }
2122
2123 STATIC void
2124 xfs_icsb_count(
2125         xfs_mount_t     *mp,
2126         xfs_icsb_cnts_t *cnt,
2127         int             flags)
2128 {
2129         xfs_icsb_cnts_t *cntp;
2130         int             i;
2131
2132         memset(cnt, 0, sizeof(xfs_icsb_cnts_t));
2133
2134         if (!(flags & XFS_ICSB_LAZY_COUNT))
2135                 xfs_icsb_lock_all_counters(mp);
2136
2137         for_each_online_cpu(i) {
2138                 cntp = (xfs_icsb_cnts_t *)per_cpu_ptr(mp->m_sb_cnts, i);
2139                 cnt->icsb_icount += cntp->icsb_icount;
2140                 cnt->icsb_ifree += cntp->icsb_ifree;
2141                 cnt->icsb_fdblocks += cntp->icsb_fdblocks;
2142         }
2143
2144         if (!(flags & XFS_ICSB_LAZY_COUNT))
2145                 xfs_icsb_unlock_all_counters(mp);
2146 }
2147
2148 STATIC int
2149 xfs_icsb_counter_disabled(
2150         xfs_mount_t     *mp,
2151         xfs_sb_field_t  field)
2152 {
2153         ASSERT((field >= XFS_SBS_ICOUNT) && (field <= XFS_SBS_FDBLOCKS));
2154         return test_bit(field, &mp->m_icsb_counters);
2155 }
2156
2157 STATIC int
2158 xfs_icsb_disable_counter(
2159         xfs_mount_t     *mp,
2160         xfs_sb_field_t  field)
2161 {
2162         xfs_icsb_cnts_t cnt;
2163
2164         ASSERT((field >= XFS_SBS_ICOUNT) && (field <= XFS_SBS_FDBLOCKS));
2165
2166         /*
2167          * If we are already disabled, then there is nothing to do
2168          * here. We check before locking all the counters to avoid
2169          * the expensive lock operation when being called in the
2170          * slow path and the counter is already disabled. This is
2171          * safe because the only time we set or clear this state is under
2172          * the m_icsb_mutex.
2173          */
2174         if (xfs_icsb_counter_disabled(mp, field))
2175                 return 0;
2176
2177         xfs_icsb_lock_all_counters(mp);
2178         if (!test_and_set_bit(field, &mp->m_icsb_counters)) {
2179                 /* drain back to superblock */
2180
2181                 xfs_icsb_count(mp, &cnt, XFS_ICSB_SB_LOCKED|XFS_ICSB_LAZY_COUNT);
2182                 switch(field) {
2183                 case XFS_SBS_ICOUNT:
2184                         mp->m_sb.sb_icount = cnt.icsb_icount;
2185                         break;
2186                 case XFS_SBS_IFREE:
2187                         mp->m_sb.sb_ifree = cnt.icsb_ifree;
2188                         break;
2189                 case XFS_SBS_FDBLOCKS:
2190                         mp->m_sb.sb_fdblocks = cnt.icsb_fdblocks;
2191                         break;
2192                 default:
2193                         BUG();
2194                 }
2195         }
2196
2197         xfs_icsb_unlock_all_counters(mp);
2198
2199         return 0;
2200 }
2201
2202 STATIC void
2203 xfs_icsb_enable_counter(
2204         xfs_mount_t     *mp,
2205         xfs_sb_field_t  field,
2206         uint64_t        count,
2207         uint64_t        resid)
2208 {
2209         xfs_icsb_cnts_t *cntp;
2210         int             i;
2211
2212         ASSERT((field >= XFS_SBS_ICOUNT) && (field <= XFS_SBS_FDBLOCKS));
2213
2214         xfs_icsb_lock_all_counters(mp);
2215         for_each_online_cpu(i) {
2216                 cntp = per_cpu_ptr(mp->m_sb_cnts, i);
2217                 switch (field) {
2218                 case XFS_SBS_ICOUNT:
2219                         cntp->icsb_icount = count + resid;
2220                         break;
2221                 case XFS_SBS_IFREE:
2222                         cntp->icsb_ifree = count + resid;
2223                         break;
2224                 case XFS_SBS_FDBLOCKS:
2225                         cntp->icsb_fdblocks = count + resid;
2226                         break;
2227                 default:
2228                         BUG();
2229                         break;
2230                 }
2231                 resid = 0;
2232         }
2233         clear_bit(field, &mp->m_icsb_counters);
2234         xfs_icsb_unlock_all_counters(mp);
2235 }
2236
2237 void
2238 xfs_icsb_sync_counters_flags(
2239         xfs_mount_t     *mp,
2240         int             flags)
2241 {
2242         xfs_icsb_cnts_t cnt;
2243
2244         /* Pass 1: lock all counters */
2245         if ((flags & XFS_ICSB_SB_LOCKED) == 0)
2246                 spin_lock(&mp->m_sb_lock);
2247
2248         xfs_icsb_count(mp, &cnt, flags);
2249
2250         /* Step 3: update mp->m_sb fields */
2251         if (!xfs_icsb_counter_disabled(mp, XFS_SBS_ICOUNT))
2252                 mp->m_sb.sb_icount = cnt.icsb_icount;
2253         if (!xfs_icsb_counter_disabled(mp, XFS_SBS_IFREE))
2254                 mp->m_sb.sb_ifree = cnt.icsb_ifree;
2255         if (!xfs_icsb_counter_disabled(mp, XFS_SBS_FDBLOCKS))
2256                 mp->m_sb.sb_fdblocks = cnt.icsb_fdblocks;
2257
2258         if ((flags & XFS_ICSB_SB_LOCKED) == 0)
2259                 spin_unlock(&mp->m_sb_lock);
2260 }
2261
2262 /*
2263  * Accurate update of per-cpu counters to incore superblock
2264  */
2265 STATIC void
2266 xfs_icsb_sync_counters(
2267         xfs_mount_t     *mp)
2268 {
2269         xfs_icsb_sync_counters_flags(mp, 0);
2270 }
2271
2272 /*
2273  * Balance and enable/disable counters as necessary.
2274  *
2275  * Thresholds for re-enabling counters are somewhat magic.  inode counts are
2276  * chosen to be the same number as single on disk allocation chunk per CPU, and
2277  * free blocks is something far enough zero that we aren't going thrash when we
2278  * get near ENOSPC. We also need to supply a minimum we require per cpu to
2279  * prevent looping endlessly when xfs_alloc_space asks for more than will
2280  * be distributed to a single CPU but each CPU has enough blocks to be
2281  * reenabled.
2282  *
2283  * Note that we can be called when counters are already disabled.
2284  * xfs_icsb_disable_counter() optimises the counter locking in this case to
2285  * prevent locking every per-cpu counter needlessly.
2286  */
2287
2288 #define XFS_ICSB_INO_CNTR_REENABLE      (uint64_t)64
2289 #define XFS_ICSB_FDBLK_CNTR_REENABLE(mp) \
2290                 (uint64_t)(512 + XFS_ALLOC_SET_ASIDE(mp))
2291 STATIC void
2292 xfs_icsb_balance_counter(
2293         xfs_mount_t     *mp,
2294         xfs_sb_field_t  field,
2295         int             flags,
2296         int             min_per_cpu)
2297 {
2298         uint64_t        count, resid;
2299         int             weight = num_online_cpus();
2300         uint64_t        min = (uint64_t)min_per_cpu;
2301
2302         if (!(flags & XFS_ICSB_SB_LOCKED))
2303                 spin_lock(&mp->m_sb_lock);
2304
2305         /* disable counter and sync counter */
2306         xfs_icsb_disable_counter(mp, field);
2307
2308         /* update counters  - first CPU gets residual*/
2309         switch (field) {
2310         case XFS_SBS_ICOUNT:
2311                 count = mp->m_sb.sb_icount;
2312                 resid = do_div(count, weight);
2313                 if (count < max(min, XFS_ICSB_INO_CNTR_REENABLE))
2314                         goto out;
2315                 break;
2316         case XFS_SBS_IFREE:
2317                 count = mp->m_sb.sb_ifree;
2318                 resid = do_div(count, weight);
2319                 if (count < max(min, XFS_ICSB_INO_CNTR_REENABLE))
2320                         goto out;
2321                 break;
2322         case XFS_SBS_FDBLOCKS:
2323                 count = mp->m_sb.sb_fdblocks;
2324                 resid = do_div(count, weight);
2325                 if (count < max(min, XFS_ICSB_FDBLK_CNTR_REENABLE(mp)))
2326                         goto out;
2327                 break;
2328         default:
2329                 BUG();
2330                 count = resid = 0;      /* quiet, gcc */
2331                 break;
2332         }
2333
2334         xfs_icsb_enable_counter(mp, field, count, resid);
2335 out:
2336         if (!(flags & XFS_ICSB_SB_LOCKED))
2337                 spin_unlock(&mp->m_sb_lock);
2338 }
2339
2340 STATIC int
2341 xfs_icsb_modify_counters(
2342         xfs_mount_t     *mp,
2343         xfs_sb_field_t  field,
2344         int64_t         delta,
2345         int             rsvd)
2346 {
2347         xfs_icsb_cnts_t *icsbp;
2348         long long       lcounter;       /* long counter for 64 bit fields */
2349         int             cpu, ret = 0;
2350
2351         might_sleep();
2352 again:
2353         cpu = get_cpu();
2354         icsbp = (xfs_icsb_cnts_t *)per_cpu_ptr(mp->m_sb_cnts, cpu);
2355
2356         /*
2357          * if the counter is disabled, go to slow path
2358          */
2359         if (unlikely(xfs_icsb_counter_disabled(mp, field)))
2360                 goto slow_path;
2361         xfs_icsb_lock_cntr(icsbp);
2362         if (unlikely(xfs_icsb_counter_disabled(mp, field))) {
2363                 xfs_icsb_unlock_cntr(icsbp);
2364                 goto slow_path;
2365         }
2366
2367         switch (field) {
2368         case XFS_SBS_ICOUNT:
2369                 lcounter = icsbp->icsb_icount;
2370                 lcounter += delta;
2371                 if (unlikely(lcounter < 0))
2372                         goto balance_counter;
2373                 icsbp->icsb_icount = lcounter;
2374                 break;
2375
2376         case XFS_SBS_IFREE:
2377                 lcounter = icsbp->icsb_ifree;
2378                 lcounter += delta;
2379                 if (unlikely(lcounter < 0))
2380                         goto balance_counter;
2381                 icsbp->icsb_ifree = lcounter;
2382                 break;
2383
2384         case XFS_SBS_FDBLOCKS:
2385                 BUG_ON((mp->m_resblks - mp->m_resblks_avail) != 0);
2386
2387                 lcounter = icsbp->icsb_fdblocks - XFS_ALLOC_SET_ASIDE(mp);
2388                 lcounter += delta;
2389                 if (unlikely(lcounter < 0))
2390                         goto balance_counter;
2391                 icsbp->icsb_fdblocks = lcounter + XFS_ALLOC_SET_ASIDE(mp);
2392                 break;
2393         default:
2394                 BUG();
2395                 break;
2396         }
2397         xfs_icsb_unlock_cntr(icsbp);
2398         put_cpu();
2399         return 0;
2400
2401 slow_path:
2402         put_cpu();
2403
2404         /*
2405          * serialise with a mutex so we don't burn lots of cpu on
2406          * the superblock lock. We still need to hold the superblock
2407          * lock, however, when we modify the global structures.
2408          */
2409         xfs_icsb_lock(mp);
2410
2411         /*
2412          * Now running atomically.
2413          *
2414          * If the counter is enabled, someone has beaten us to rebalancing.
2415          * Drop the lock and try again in the fast path....
2416          */
2417         if (!(xfs_icsb_counter_disabled(mp, field))) {
2418                 xfs_icsb_unlock(mp);
2419                 goto again;
2420         }
2421
2422         /*
2423          * The counter is currently disabled. Because we are
2424          * running atomically here, we know a rebalance cannot
2425          * be in progress. Hence we can go straight to operating
2426          * on the global superblock. We do not call xfs_mod_incore_sb()
2427          * here even though we need to get the m_sb_lock. Doing so
2428          * will cause us to re-enter this function and deadlock.
2429          * Hence we get the m_sb_lock ourselves and then call
2430          * xfs_mod_incore_sb_unlocked() as the unlocked path operates
2431          * directly on the global counters.
2432          */
2433         spin_lock(&mp->m_sb_lock);
2434         ret = xfs_mod_incore_sb_unlocked(mp, field, delta, rsvd);
2435         spin_unlock(&mp->m_sb_lock);
2436
2437         /*
2438          * Now that we've modified the global superblock, we
2439          * may be able to re-enable the distributed counters
2440          * (e.g. lots of space just got freed). After that
2441          * we are done.
2442          */
2443         if (ret != ENOSPC)
2444                 xfs_icsb_balance_counter(mp, field, 0, 0);
2445         xfs_icsb_unlock(mp);
2446         return ret;
2447
2448 balance_counter:
2449         xfs_icsb_unlock_cntr(icsbp);
2450         put_cpu();
2451
2452         /*
2453          * We may have multiple threads here if multiple per-cpu
2454          * counters run dry at the same time. This will mean we can
2455          * do more balances than strictly necessary but it is not
2456          * the common slowpath case.
2457          */
2458         xfs_icsb_lock(mp);
2459
2460         /*
2461          * running atomically.
2462          *
2463          * This will leave the counter in the correct state for future
2464          * accesses. After the rebalance, we simply try again and our retry
2465          * will either succeed through the fast path or slow path without
2466          * another balance operation being required.
2467          */
2468         xfs_icsb_balance_counter(mp, field, 0, delta);
2469         xfs_icsb_unlock(mp);
2470         goto again;
2471 }
2472
2473 #endif