]> git.karo-electronics.de Git - mv-sheeva.git/blob - fs/xfs/xfs_mount.c
[XFS] Fix up noattr2 so that it will properly update the versionnum and
[mv-sheeva.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 #include "xfs_utils.h"
47
48 STATIC int      xfs_mount_log_sb(xfs_mount_t *, __int64_t);
49 STATIC int      xfs_uuid_mount(xfs_mount_t *);
50 STATIC void     xfs_uuid_unmount(xfs_mount_t *mp);
51 STATIC void     xfs_unmountfs_wait(xfs_mount_t *);
52
53
54 #ifdef HAVE_PERCPU_SB
55 STATIC void     xfs_icsb_destroy_counters(xfs_mount_t *);
56 STATIC void     xfs_icsb_balance_counter(xfs_mount_t *, xfs_sb_field_t,
57                                                 int);
58 STATIC void     xfs_icsb_balance_counter_locked(xfs_mount_t *, xfs_sb_field_t,
59                                                 int);
60 STATIC int      xfs_icsb_modify_counters(xfs_mount_t *, xfs_sb_field_t,
61                                                 int64_t, int);
62 STATIC void     xfs_icsb_disable_counter(xfs_mount_t *, xfs_sb_field_t);
63
64 #else
65
66 #define xfs_icsb_destroy_counters(mp)                   do { } while (0)
67 #define xfs_icsb_balance_counter(mp, a, b)              do { } while (0)
68 #define xfs_icsb_balance_counter_locked(mp, a, b)       do { } while (0)
69 #define xfs_icsb_modify_counters(mp, a, b, c)           do { } while (0)
70
71 #endif
72
73 static const struct {
74         short offset;
75         short type;     /* 0 = integer
76                          * 1 = binary / string (no translation)
77                          */
78 } xfs_sb_info[] = {
79     { offsetof(xfs_sb_t, sb_magicnum),   0 },
80     { offsetof(xfs_sb_t, sb_blocksize),  0 },
81     { offsetof(xfs_sb_t, sb_dblocks),    0 },
82     { offsetof(xfs_sb_t, sb_rblocks),    0 },
83     { offsetof(xfs_sb_t, sb_rextents),   0 },
84     { offsetof(xfs_sb_t, sb_uuid),       1 },
85     { offsetof(xfs_sb_t, sb_logstart),   0 },
86     { offsetof(xfs_sb_t, sb_rootino),    0 },
87     { offsetof(xfs_sb_t, sb_rbmino),     0 },
88     { offsetof(xfs_sb_t, sb_rsumino),    0 },
89     { offsetof(xfs_sb_t, sb_rextsize),   0 },
90     { offsetof(xfs_sb_t, sb_agblocks),   0 },
91     { offsetof(xfs_sb_t, sb_agcount),    0 },
92     { offsetof(xfs_sb_t, sb_rbmblocks),  0 },
93     { offsetof(xfs_sb_t, sb_logblocks),  0 },
94     { offsetof(xfs_sb_t, sb_versionnum), 0 },
95     { offsetof(xfs_sb_t, sb_sectsize),   0 },
96     { offsetof(xfs_sb_t, sb_inodesize),  0 },
97     { offsetof(xfs_sb_t, sb_inopblock),  0 },
98     { offsetof(xfs_sb_t, sb_fname[0]),   1 },
99     { offsetof(xfs_sb_t, sb_blocklog),   0 },
100     { offsetof(xfs_sb_t, sb_sectlog),    0 },
101     { offsetof(xfs_sb_t, sb_inodelog),   0 },
102     { offsetof(xfs_sb_t, sb_inopblog),   0 },
103     { offsetof(xfs_sb_t, sb_agblklog),   0 },
104     { offsetof(xfs_sb_t, sb_rextslog),   0 },
105     { offsetof(xfs_sb_t, sb_inprogress), 0 },
106     { offsetof(xfs_sb_t, sb_imax_pct),   0 },
107     { offsetof(xfs_sb_t, sb_icount),     0 },
108     { offsetof(xfs_sb_t, sb_ifree),      0 },
109     { offsetof(xfs_sb_t, sb_fdblocks),   0 },
110     { offsetof(xfs_sb_t, sb_frextents),  0 },
111     { offsetof(xfs_sb_t, sb_uquotino),   0 },
112     { offsetof(xfs_sb_t, sb_gquotino),   0 },
113     { offsetof(xfs_sb_t, sb_qflags),     0 },
114     { offsetof(xfs_sb_t, sb_flags),      0 },
115     { offsetof(xfs_sb_t, sb_shared_vn),  0 },
116     { offsetof(xfs_sb_t, sb_inoalignmt), 0 },
117     { offsetof(xfs_sb_t, sb_unit),       0 },
118     { offsetof(xfs_sb_t, sb_width),      0 },
119     { offsetof(xfs_sb_t, sb_dirblklog),  0 },
120     { offsetof(xfs_sb_t, sb_logsectlog), 0 },
121     { offsetof(xfs_sb_t, sb_logsectsize),0 },
122     { offsetof(xfs_sb_t, sb_logsunit),   0 },
123     { offsetof(xfs_sb_t, sb_features2),  0 },
124     { offsetof(xfs_sb_t, sb_bad_features2), 0 },
125     { sizeof(xfs_sb_t),                  0 }
126 };
127
128 /*
129  * Return a pointer to an initialized xfs_mount structure.
130  */
131 xfs_mount_t *
132 xfs_mount_init(void)
133 {
134         xfs_mount_t *mp;
135
136         mp = kmem_zalloc(sizeof(xfs_mount_t), KM_SLEEP);
137
138         if (xfs_icsb_init_counters(mp)) {
139                 mp->m_flags |= XFS_MOUNT_NO_PERCPU_SB;
140         }
141
142         spin_lock_init(&mp->m_sb_lock);
143         mutex_init(&mp->m_ilock);
144         mutex_init(&mp->m_growlock);
145         atomic_set(&mp->m_active_trans, 0);
146
147         return mp;
148 }
149
150 /*
151  * Free up the resources associated with a mount structure.  Assume that
152  * the structure was initially zeroed, so we can tell which fields got
153  * initialized.
154  */
155 void
156 xfs_mount_free(
157         xfs_mount_t     *mp)
158 {
159         if (mp->m_perag) {
160                 int     agno;
161
162                 for (agno = 0; agno < mp->m_maxagi; agno++)
163                         if (mp->m_perag[agno].pagb_list)
164                                 kmem_free(mp->m_perag[agno].pagb_list,
165                                                 sizeof(xfs_perag_busy_t) *
166                                                         XFS_PAGB_NUM_SLOTS);
167                 kmem_free(mp->m_perag,
168                           sizeof(xfs_perag_t) * mp->m_sb.sb_agcount);
169         }
170
171         spinlock_destroy(&mp->m_ail_lock);
172         spinlock_destroy(&mp->m_sb_lock);
173         mutex_destroy(&mp->m_ilock);
174         mutex_destroy(&mp->m_growlock);
175         if (mp->m_quotainfo)
176                 XFS_QM_DONE(mp);
177
178         if (mp->m_fsname != NULL)
179                 kmem_free(mp->m_fsname, mp->m_fsname_len);
180         if (mp->m_rtname != NULL)
181                 kmem_free(mp->m_rtname, strlen(mp->m_rtname) + 1);
182         if (mp->m_logname != NULL)
183                 kmem_free(mp->m_logname, strlen(mp->m_logname) + 1);
184
185         xfs_icsb_destroy_counters(mp);
186 }
187
188 /*
189  * Check size of device based on the (data/realtime) block count.
190  * Note: this check is used by the growfs code as well as mount.
191  */
192 int
193 xfs_sb_validate_fsb_count(
194         xfs_sb_t        *sbp,
195         __uint64_t      nblocks)
196 {
197         ASSERT(PAGE_SHIFT >= sbp->sb_blocklog);
198         ASSERT(sbp->sb_blocklog >= BBSHIFT);
199
200 #if XFS_BIG_BLKNOS     /* Limited by ULONG_MAX of page cache index */
201         if (nblocks >> (PAGE_CACHE_SHIFT - sbp->sb_blocklog) > ULONG_MAX)
202                 return E2BIG;
203 #else                  /* Limited by UINT_MAX of sectors */
204         if (nblocks << (sbp->sb_blocklog - BBSHIFT) > UINT_MAX)
205                 return E2BIG;
206 #endif
207         return 0;
208 }
209
210 /*
211  * Check the validity of the SB found.
212  */
213 STATIC int
214 xfs_mount_validate_sb(
215         xfs_mount_t     *mp,
216         xfs_sb_t        *sbp,
217         int             flags)
218 {
219         /*
220          * If the log device and data device have the
221          * same device number, the log is internal.
222          * Consequently, the sb_logstart should be non-zero.  If
223          * we have a zero sb_logstart in this case, we may be trying to mount
224          * a volume filesystem in a non-volume manner.
225          */
226         if (sbp->sb_magicnum != XFS_SB_MAGIC) {
227                 xfs_fs_mount_cmn_err(flags, "bad magic number");
228                 return XFS_ERROR(EWRONGFS);
229         }
230
231         if (!xfs_sb_good_version(sbp)) {
232                 xfs_fs_mount_cmn_err(flags, "bad version");
233                 return XFS_ERROR(EWRONGFS);
234         }
235
236         if (unlikely(
237             sbp->sb_logstart == 0 && mp->m_logdev_targp == mp->m_ddev_targp)) {
238                 xfs_fs_mount_cmn_err(flags,
239                         "filesystem is marked as having an external log; "
240                         "specify logdev on the\nmount command line.");
241                 return XFS_ERROR(EINVAL);
242         }
243
244         if (unlikely(
245             sbp->sb_logstart != 0 && mp->m_logdev_targp != mp->m_ddev_targp)) {
246                 xfs_fs_mount_cmn_err(flags,
247                         "filesystem is marked as having an internal log; "
248                         "do not specify logdev on\nthe mount command line.");
249                 return XFS_ERROR(EINVAL);
250         }
251
252         /*
253          * More sanity checking. These were stolen directly from
254          * xfs_repair.
255          */
256         if (unlikely(
257             sbp->sb_agcount <= 0                                        ||
258             sbp->sb_sectsize < XFS_MIN_SECTORSIZE                       ||
259             sbp->sb_sectsize > XFS_MAX_SECTORSIZE                       ||
260             sbp->sb_sectlog < XFS_MIN_SECTORSIZE_LOG                    ||
261             sbp->sb_sectlog > XFS_MAX_SECTORSIZE_LOG                    ||
262             sbp->sb_blocksize < XFS_MIN_BLOCKSIZE                       ||
263             sbp->sb_blocksize > XFS_MAX_BLOCKSIZE                       ||
264             sbp->sb_blocklog < XFS_MIN_BLOCKSIZE_LOG                    ||
265             sbp->sb_blocklog > XFS_MAX_BLOCKSIZE_LOG                    ||
266             sbp->sb_inodesize < XFS_DINODE_MIN_SIZE                     ||
267             sbp->sb_inodesize > XFS_DINODE_MAX_SIZE                     ||
268             sbp->sb_inodelog < XFS_DINODE_MIN_LOG                       ||
269             sbp->sb_inodelog > XFS_DINODE_MAX_LOG                       ||
270             (sbp->sb_blocklog - sbp->sb_inodelog != sbp->sb_inopblog)   ||
271             (sbp->sb_rextsize * sbp->sb_blocksize > XFS_MAX_RTEXTSIZE)  ||
272             (sbp->sb_rextsize * sbp->sb_blocksize < XFS_MIN_RTEXTSIZE)  ||
273             (sbp->sb_imax_pct > 100 /* zero sb_imax_pct is valid */))) {
274                 xfs_fs_mount_cmn_err(flags, "SB sanity check 1 failed");
275                 return XFS_ERROR(EFSCORRUPTED);
276         }
277
278         /*
279          * Sanity check AG count, size fields against data size field
280          */
281         if (unlikely(
282             sbp->sb_dblocks == 0 ||
283             sbp->sb_dblocks >
284              (xfs_drfsbno_t)sbp->sb_agcount * sbp->sb_agblocks ||
285             sbp->sb_dblocks < (xfs_drfsbno_t)(sbp->sb_agcount - 1) *
286                               sbp->sb_agblocks + XFS_MIN_AG_BLOCKS)) {
287                 xfs_fs_mount_cmn_err(flags, "SB sanity check 2 failed");
288                 return XFS_ERROR(EFSCORRUPTED);
289         }
290
291         if (xfs_sb_validate_fsb_count(sbp, sbp->sb_dblocks) ||
292             xfs_sb_validate_fsb_count(sbp, sbp->sb_rblocks)) {
293                 xfs_fs_mount_cmn_err(flags,
294                         "file system too large to be mounted on this system.");
295                 return XFS_ERROR(E2BIG);
296         }
297
298         if (unlikely(sbp->sb_inprogress)) {
299                 xfs_fs_mount_cmn_err(flags, "file system busy");
300                 return XFS_ERROR(EFSCORRUPTED);
301         }
302
303         /*
304          * Version 1 directory format has never worked on Linux.
305          */
306         if (unlikely(!xfs_sb_version_hasdirv2(sbp))) {
307                 xfs_fs_mount_cmn_err(flags,
308                         "file system using version 1 directory format");
309                 return XFS_ERROR(ENOSYS);
310         }
311
312         /*
313          * Until this is fixed only page-sized or smaller data blocks work.
314          */
315         if (unlikely(sbp->sb_blocksize > PAGE_SIZE)) {
316                 xfs_fs_mount_cmn_err(flags,
317                         "file system with blocksize %d bytes",
318                         sbp->sb_blocksize);
319                 xfs_fs_mount_cmn_err(flags,
320                         "only pagesize (%ld) or less will currently work.",
321                         PAGE_SIZE);
322                 return XFS_ERROR(ENOSYS);
323         }
324
325         return 0;
326 }
327
328 STATIC void
329 xfs_initialize_perag_icache(
330         xfs_perag_t     *pag)
331 {
332         if (!pag->pag_ici_init) {
333                 rwlock_init(&pag->pag_ici_lock);
334                 INIT_RADIX_TREE(&pag->pag_ici_root, GFP_ATOMIC);
335                 pag->pag_ici_init = 1;
336         }
337 }
338
339 xfs_agnumber_t
340 xfs_initialize_perag(
341         xfs_mount_t     *mp,
342         xfs_agnumber_t  agcount)
343 {
344         xfs_agnumber_t  index, max_metadata;
345         xfs_perag_t     *pag;
346         xfs_agino_t     agino;
347         xfs_ino_t       ino;
348         xfs_sb_t        *sbp = &mp->m_sb;
349         xfs_ino_t       max_inum = XFS_MAXINUMBER_32;
350
351         /* Check to see if the filesystem can overflow 32 bit inodes */
352         agino = XFS_OFFBNO_TO_AGINO(mp, sbp->sb_agblocks - 1, 0);
353         ino = XFS_AGINO_TO_INO(mp, agcount - 1, agino);
354
355         /* Clear the mount flag if no inode can overflow 32 bits
356          * on this filesystem, or if specifically requested..
357          */
358         if ((mp->m_flags & XFS_MOUNT_SMALL_INUMS) && ino > max_inum) {
359                 mp->m_flags |= XFS_MOUNT_32BITINODES;
360         } else {
361                 mp->m_flags &= ~XFS_MOUNT_32BITINODES;
362         }
363
364         /* If we can overflow then setup the ag headers accordingly */
365         if (mp->m_flags & XFS_MOUNT_32BITINODES) {
366                 /* Calculate how much should be reserved for inodes to
367                  * meet the max inode percentage.
368                  */
369                 if (mp->m_maxicount) {
370                         __uint64_t      icount;
371
372                         icount = sbp->sb_dblocks * sbp->sb_imax_pct;
373                         do_div(icount, 100);
374                         icount += sbp->sb_agblocks - 1;
375                         do_div(icount, sbp->sb_agblocks);
376                         max_metadata = icount;
377                 } else {
378                         max_metadata = agcount;
379                 }
380                 for (index = 0; index < agcount; index++) {
381                         ino = XFS_AGINO_TO_INO(mp, index, agino);
382                         if (ino > max_inum) {
383                                 index++;
384                                 break;
385                         }
386
387                         /* This ag is preferred for inodes */
388                         pag = &mp->m_perag[index];
389                         pag->pagi_inodeok = 1;
390                         if (index < max_metadata)
391                                 pag->pagf_metadata = 1;
392                         xfs_initialize_perag_icache(pag);
393                 }
394         } else {
395                 /* Setup default behavior for smaller filesystems */
396                 for (index = 0; index < agcount; index++) {
397                         pag = &mp->m_perag[index];
398                         pag->pagi_inodeok = 1;
399                         xfs_initialize_perag_icache(pag);
400                 }
401         }
402         return index;
403 }
404
405 void
406 xfs_sb_from_disk(
407         xfs_sb_t        *to,
408         xfs_dsb_t       *from)
409 {
410         to->sb_magicnum = be32_to_cpu(from->sb_magicnum);
411         to->sb_blocksize = be32_to_cpu(from->sb_blocksize);
412         to->sb_dblocks = be64_to_cpu(from->sb_dblocks);
413         to->sb_rblocks = be64_to_cpu(from->sb_rblocks);
414         to->sb_rextents = be64_to_cpu(from->sb_rextents);
415         memcpy(&to->sb_uuid, &from->sb_uuid, sizeof(to->sb_uuid));
416         to->sb_logstart = be64_to_cpu(from->sb_logstart);
417         to->sb_rootino = be64_to_cpu(from->sb_rootino);
418         to->sb_rbmino = be64_to_cpu(from->sb_rbmino);
419         to->sb_rsumino = be64_to_cpu(from->sb_rsumino);
420         to->sb_rextsize = be32_to_cpu(from->sb_rextsize);
421         to->sb_agblocks = be32_to_cpu(from->sb_agblocks);
422         to->sb_agcount = be32_to_cpu(from->sb_agcount);
423         to->sb_rbmblocks = be32_to_cpu(from->sb_rbmblocks);
424         to->sb_logblocks = be32_to_cpu(from->sb_logblocks);
425         to->sb_versionnum = be16_to_cpu(from->sb_versionnum);
426         to->sb_sectsize = be16_to_cpu(from->sb_sectsize);
427         to->sb_inodesize = be16_to_cpu(from->sb_inodesize);
428         to->sb_inopblock = be16_to_cpu(from->sb_inopblock);
429         memcpy(&to->sb_fname, &from->sb_fname, sizeof(to->sb_fname));
430         to->sb_blocklog = from->sb_blocklog;
431         to->sb_sectlog = from->sb_sectlog;
432         to->sb_inodelog = from->sb_inodelog;
433         to->sb_inopblog = from->sb_inopblog;
434         to->sb_agblklog = from->sb_agblklog;
435         to->sb_rextslog = from->sb_rextslog;
436         to->sb_inprogress = from->sb_inprogress;
437         to->sb_imax_pct = from->sb_imax_pct;
438         to->sb_icount = be64_to_cpu(from->sb_icount);
439         to->sb_ifree = be64_to_cpu(from->sb_ifree);
440         to->sb_fdblocks = be64_to_cpu(from->sb_fdblocks);
441         to->sb_frextents = be64_to_cpu(from->sb_frextents);
442         to->sb_uquotino = be64_to_cpu(from->sb_uquotino);
443         to->sb_gquotino = be64_to_cpu(from->sb_gquotino);
444         to->sb_qflags = be16_to_cpu(from->sb_qflags);
445         to->sb_flags = from->sb_flags;
446         to->sb_shared_vn = from->sb_shared_vn;
447         to->sb_inoalignmt = be32_to_cpu(from->sb_inoalignmt);
448         to->sb_unit = be32_to_cpu(from->sb_unit);
449         to->sb_width = be32_to_cpu(from->sb_width);
450         to->sb_dirblklog = from->sb_dirblklog;
451         to->sb_logsectlog = from->sb_logsectlog;
452         to->sb_logsectsize = be16_to_cpu(from->sb_logsectsize);
453         to->sb_logsunit = be32_to_cpu(from->sb_logsunit);
454         to->sb_features2 = be32_to_cpu(from->sb_features2);
455         to->sb_bad_features2 = be32_to_cpu(from->sb_bad_features2);
456 }
457
458 /*
459  * Copy in core superblock to ondisk one.
460  *
461  * The fields argument is mask of superblock fields to copy.
462  */
463 void
464 xfs_sb_to_disk(
465         xfs_dsb_t       *to,
466         xfs_sb_t        *from,
467         __int64_t       fields)
468 {
469         xfs_caddr_t     to_ptr = (xfs_caddr_t)to;
470         xfs_caddr_t     from_ptr = (xfs_caddr_t)from;
471         xfs_sb_field_t  f;
472         int             first;
473         int             size;
474
475         ASSERT(fields);
476         if (!fields)
477                 return;
478
479         while (fields) {
480                 f = (xfs_sb_field_t)xfs_lowbit64((__uint64_t)fields);
481                 first = xfs_sb_info[f].offset;
482                 size = xfs_sb_info[f + 1].offset - first;
483
484                 ASSERT(xfs_sb_info[f].type == 0 || xfs_sb_info[f].type == 1);
485
486                 if (size == 1 || xfs_sb_info[f].type == 1) {
487                         memcpy(to_ptr + first, from_ptr + first, size);
488                 } else {
489                         switch (size) {
490                         case 2:
491                                 *(__be16 *)(to_ptr + first) =
492                                         cpu_to_be16(*(__u16 *)(from_ptr + first));
493                                 break;
494                         case 4:
495                                 *(__be32 *)(to_ptr + first) =
496                                         cpu_to_be32(*(__u32 *)(from_ptr + first));
497                                 break;
498                         case 8:
499                                 *(__be64 *)(to_ptr + first) =
500                                         cpu_to_be64(*(__u64 *)(from_ptr + first));
501                                 break;
502                         default:
503                                 ASSERT(0);
504                         }
505                 }
506
507                 fields &= ~(1LL << f);
508         }
509 }
510
511 /*
512  * xfs_readsb
513  *
514  * Does the initial read of the superblock.
515  */
516 int
517 xfs_readsb(xfs_mount_t *mp, int flags)
518 {
519         unsigned int    sector_size;
520         unsigned int    extra_flags;
521         xfs_buf_t       *bp;
522         int             error;
523
524         ASSERT(mp->m_sb_bp == NULL);
525         ASSERT(mp->m_ddev_targp != NULL);
526
527         /*
528          * Allocate a (locked) buffer to hold the superblock.
529          * This will be kept around at all times to optimize
530          * access to the superblock.
531          */
532         sector_size = xfs_getsize_buftarg(mp->m_ddev_targp);
533         extra_flags = XFS_BUF_LOCK | XFS_BUF_MANAGE | XFS_BUF_MAPPED;
534
535         bp = xfs_buf_read_flags(mp->m_ddev_targp, XFS_SB_DADDR,
536                                 BTOBB(sector_size), extra_flags);
537         if (!bp || XFS_BUF_ISERROR(bp)) {
538                 xfs_fs_mount_cmn_err(flags, "SB read failed");
539                 error = bp ? XFS_BUF_GETERROR(bp) : ENOMEM;
540                 goto fail;
541         }
542         ASSERT(XFS_BUF_ISBUSY(bp));
543         ASSERT(XFS_BUF_VALUSEMA(bp) <= 0);
544
545         /*
546          * Initialize the mount structure from the superblock.
547          * But first do some basic consistency checking.
548          */
549         xfs_sb_from_disk(&mp->m_sb, XFS_BUF_TO_SBP(bp));
550
551         error = xfs_mount_validate_sb(mp, &(mp->m_sb), flags);
552         if (error) {
553                 xfs_fs_mount_cmn_err(flags, "SB validate failed");
554                 goto fail;
555         }
556
557         /*
558          * We must be able to do sector-sized and sector-aligned IO.
559          */
560         if (sector_size > mp->m_sb.sb_sectsize) {
561                 xfs_fs_mount_cmn_err(flags,
562                         "device supports only %u byte sectors (not %u)",
563                         sector_size, mp->m_sb.sb_sectsize);
564                 error = ENOSYS;
565                 goto fail;
566         }
567
568         /*
569          * If device sector size is smaller than the superblock size,
570          * re-read the superblock so the buffer is correctly sized.
571          */
572         if (sector_size < mp->m_sb.sb_sectsize) {
573                 XFS_BUF_UNMANAGE(bp);
574                 xfs_buf_relse(bp);
575                 sector_size = mp->m_sb.sb_sectsize;
576                 bp = xfs_buf_read_flags(mp->m_ddev_targp, XFS_SB_DADDR,
577                                         BTOBB(sector_size), extra_flags);
578                 if (!bp || XFS_BUF_ISERROR(bp)) {
579                         xfs_fs_mount_cmn_err(flags, "SB re-read failed");
580                         error = bp ? XFS_BUF_GETERROR(bp) : ENOMEM;
581                         goto fail;
582                 }
583                 ASSERT(XFS_BUF_ISBUSY(bp));
584                 ASSERT(XFS_BUF_VALUSEMA(bp) <= 0);
585         }
586
587         /* Initialize per-cpu counters */
588         xfs_icsb_reinit_counters(mp);
589
590         mp->m_sb_bp = bp;
591         xfs_buf_relse(bp);
592         ASSERT(XFS_BUF_VALUSEMA(bp) > 0);
593         return 0;
594
595  fail:
596         if (bp) {
597                 XFS_BUF_UNMANAGE(bp);
598                 xfs_buf_relse(bp);
599         }
600         return error;
601 }
602
603
604 /*
605  * xfs_mount_common
606  *
607  * Mount initialization code establishing various mount
608  * fields from the superblock associated with the given
609  * mount structure
610  */
611 STATIC void
612 xfs_mount_common(xfs_mount_t *mp, xfs_sb_t *sbp)
613 {
614         int     i;
615
616         mp->m_agfrotor = mp->m_agirotor = 0;
617         spin_lock_init(&mp->m_agirotor_lock);
618         mp->m_maxagi = mp->m_sb.sb_agcount;
619         mp->m_blkbit_log = sbp->sb_blocklog + XFS_NBBYLOG;
620         mp->m_blkbb_log = sbp->sb_blocklog - BBSHIFT;
621         mp->m_sectbb_log = sbp->sb_sectlog - BBSHIFT;
622         mp->m_agno_log = xfs_highbit32(sbp->sb_agcount - 1) + 1;
623         mp->m_agino_log = sbp->sb_inopblog + sbp->sb_agblklog;
624         mp->m_litino = sbp->sb_inodesize -
625                 ((uint)sizeof(xfs_dinode_core_t) + (uint)sizeof(xfs_agino_t));
626         mp->m_blockmask = sbp->sb_blocksize - 1;
627         mp->m_blockwsize = sbp->sb_blocksize >> XFS_WORDLOG;
628         mp->m_blockwmask = mp->m_blockwsize - 1;
629         INIT_LIST_HEAD(&mp->m_del_inodes);
630
631         /*
632          * Setup for attributes, in case they get created.
633          * This value is for inodes getting attributes for the first time,
634          * the per-inode value is for old attribute values.
635          */
636         ASSERT(sbp->sb_inodesize >= 256 && sbp->sb_inodesize <= 2048);
637         switch (sbp->sb_inodesize) {
638         case 256:
639                 mp->m_attroffset = XFS_LITINO(mp) -
640                                    XFS_BMDR_SPACE_CALC(MINABTPTRS);
641                 break;
642         case 512:
643         case 1024:
644         case 2048:
645                 mp->m_attroffset = XFS_BMDR_SPACE_CALC(6 * MINABTPTRS);
646                 break;
647         default:
648                 ASSERT(0);
649         }
650         ASSERT(mp->m_attroffset < XFS_LITINO(mp));
651
652         for (i = 0; i < 2; i++) {
653                 mp->m_alloc_mxr[i] = XFS_BTREE_BLOCK_MAXRECS(sbp->sb_blocksize,
654                         xfs_alloc, i == 0);
655                 mp->m_alloc_mnr[i] = XFS_BTREE_BLOCK_MINRECS(sbp->sb_blocksize,
656                         xfs_alloc, i == 0);
657         }
658         for (i = 0; i < 2; i++) {
659                 mp->m_bmap_dmxr[i] = XFS_BTREE_BLOCK_MAXRECS(sbp->sb_blocksize,
660                         xfs_bmbt, i == 0);
661                 mp->m_bmap_dmnr[i] = XFS_BTREE_BLOCK_MINRECS(sbp->sb_blocksize,
662                         xfs_bmbt, i == 0);
663         }
664         for (i = 0; i < 2; i++) {
665                 mp->m_inobt_mxr[i] = XFS_BTREE_BLOCK_MAXRECS(sbp->sb_blocksize,
666                         xfs_inobt, i == 0);
667                 mp->m_inobt_mnr[i] = XFS_BTREE_BLOCK_MINRECS(sbp->sb_blocksize,
668                         xfs_inobt, i == 0);
669         }
670
671         mp->m_bsize = XFS_FSB_TO_BB(mp, 1);
672         mp->m_ialloc_inos = (int)MAX((__uint16_t)XFS_INODES_PER_CHUNK,
673                                         sbp->sb_inopblock);
674         mp->m_ialloc_blks = mp->m_ialloc_inos >> sbp->sb_inopblog;
675 }
676
677 /*
678  * xfs_initialize_perag_data
679  *
680  * Read in each per-ag structure so we can count up the number of
681  * allocated inodes, free inodes and used filesystem blocks as this
682  * information is no longer persistent in the superblock. Once we have
683  * this information, write it into the in-core superblock structure.
684  */
685 STATIC int
686 xfs_initialize_perag_data(xfs_mount_t *mp, xfs_agnumber_t agcount)
687 {
688         xfs_agnumber_t  index;
689         xfs_perag_t     *pag;
690         xfs_sb_t        *sbp = &mp->m_sb;
691         uint64_t        ifree = 0;
692         uint64_t        ialloc = 0;
693         uint64_t        bfree = 0;
694         uint64_t        bfreelst = 0;
695         uint64_t        btree = 0;
696         int             error;
697
698         for (index = 0; index < agcount; index++) {
699                 /*
700                  * read the agf, then the agi. This gets us
701                  * all the inforamtion we need and populates the
702                  * per-ag structures for us.
703                  */
704                 error = xfs_alloc_pagf_init(mp, NULL, index, 0);
705                 if (error)
706                         return error;
707
708                 error = xfs_ialloc_pagi_init(mp, NULL, index);
709                 if (error)
710                         return error;
711                 pag = &mp->m_perag[index];
712                 ifree += pag->pagi_freecount;
713                 ialloc += pag->pagi_count;
714                 bfree += pag->pagf_freeblks;
715                 bfreelst += pag->pagf_flcount;
716                 btree += pag->pagf_btreeblks;
717         }
718         /*
719          * Overwrite incore superblock counters with just-read data
720          */
721         spin_lock(&mp->m_sb_lock);
722         sbp->sb_ifree = ifree;
723         sbp->sb_icount = ialloc;
724         sbp->sb_fdblocks = bfree + bfreelst + btree;
725         spin_unlock(&mp->m_sb_lock);
726
727         /* Fixup the per-cpu counters as well. */
728         xfs_icsb_reinit_counters(mp);
729
730         return 0;
731 }
732
733 /*
734  * Update alignment values based on mount options and sb values
735  */
736 STATIC int
737 xfs_update_alignment(xfs_mount_t *mp, int mfsi_flags, __uint64_t *update_flags)
738 {
739         xfs_sb_t        *sbp = &(mp->m_sb);
740
741         if (mp->m_dalign && !(mfsi_flags & XFS_MFSI_SECOND)) {
742                 /*
743                  * If stripe unit and stripe width are not multiples
744                  * of the fs blocksize turn off alignment.
745                  */
746                 if ((BBTOB(mp->m_dalign) & mp->m_blockmask) ||
747                     (BBTOB(mp->m_swidth) & mp->m_blockmask)) {
748                         if (mp->m_flags & XFS_MOUNT_RETERR) {
749                                 cmn_err(CE_WARN,
750                                         "XFS: alignment check 1 failed");
751                                 return XFS_ERROR(EINVAL);
752                         }
753                         mp->m_dalign = mp->m_swidth = 0;
754                 } else {
755                         /*
756                          * Convert the stripe unit and width to FSBs.
757                          */
758                         mp->m_dalign = XFS_BB_TO_FSBT(mp, mp->m_dalign);
759                         if (mp->m_dalign && (sbp->sb_agblocks % mp->m_dalign)) {
760                                 if (mp->m_flags & XFS_MOUNT_RETERR) {
761                                         return XFS_ERROR(EINVAL);
762                                 }
763                                 xfs_fs_cmn_err(CE_WARN, mp,
764 "stripe alignment turned off: sunit(%d)/swidth(%d) incompatible with agsize(%d)",
765                                         mp->m_dalign, mp->m_swidth,
766                                         sbp->sb_agblocks);
767
768                                 mp->m_dalign = 0;
769                                 mp->m_swidth = 0;
770                         } else if (mp->m_dalign) {
771                                 mp->m_swidth = XFS_BB_TO_FSBT(mp, mp->m_swidth);
772                         } else {
773                                 if (mp->m_flags & XFS_MOUNT_RETERR) {
774                                         xfs_fs_cmn_err(CE_WARN, mp,
775 "stripe alignment turned off: sunit(%d) less than bsize(%d)",
776                                                 mp->m_dalign,
777                                                 mp->m_blockmask +1);
778                                         return XFS_ERROR(EINVAL);
779                                 }
780                                 mp->m_swidth = 0;
781                         }
782                 }
783
784                 /*
785                  * Update superblock with new values
786                  * and log changes
787                  */
788                 if (xfs_sb_version_hasdalign(sbp)) {
789                         if (sbp->sb_unit != mp->m_dalign) {
790                                 sbp->sb_unit = mp->m_dalign;
791                                 *update_flags |= XFS_SB_UNIT;
792                         }
793                         if (sbp->sb_width != mp->m_swidth) {
794                                 sbp->sb_width = mp->m_swidth;
795                                 *update_flags |= XFS_SB_WIDTH;
796                         }
797                 }
798         } else if ((mp->m_flags & XFS_MOUNT_NOALIGN) != XFS_MOUNT_NOALIGN &&
799                     xfs_sb_version_hasdalign(&mp->m_sb)) {
800                         mp->m_dalign = sbp->sb_unit;
801                         mp->m_swidth = sbp->sb_width;
802         }
803
804         return 0;
805 }
806
807 /*
808  * Set the maximum inode count for this filesystem
809  */
810 STATIC void
811 xfs_set_maxicount(xfs_mount_t *mp)
812 {
813         xfs_sb_t        *sbp = &(mp->m_sb);
814         __uint64_t      icount;
815
816         if (sbp->sb_imax_pct) {
817                 /*
818                  * Make sure the maximum inode count is a multiple
819                  * of the units we allocate inodes in.
820                  */
821                 icount = sbp->sb_dblocks * sbp->sb_imax_pct;
822                 do_div(icount, 100);
823                 do_div(icount, mp->m_ialloc_blks);
824                 mp->m_maxicount = (icount * mp->m_ialloc_blks)  <<
825                                    sbp->sb_inopblog;
826         } else {
827                 mp->m_maxicount = 0;
828         }
829 }
830
831 /*
832  * Set the default minimum read and write sizes unless
833  * already specified in a mount option.
834  * We use smaller I/O sizes when the file system
835  * is being used for NFS service (wsync mount option).
836  */
837 STATIC void
838 xfs_set_rw_sizes(xfs_mount_t *mp)
839 {
840         xfs_sb_t        *sbp = &(mp->m_sb);
841         int             readio_log, writeio_log;
842
843         if (!(mp->m_flags & XFS_MOUNT_DFLT_IOSIZE)) {
844                 if (mp->m_flags & XFS_MOUNT_WSYNC) {
845                         readio_log = XFS_WSYNC_READIO_LOG;
846                         writeio_log = XFS_WSYNC_WRITEIO_LOG;
847                 } else {
848                         readio_log = XFS_READIO_LOG_LARGE;
849                         writeio_log = XFS_WRITEIO_LOG_LARGE;
850                 }
851         } else {
852                 readio_log = mp->m_readio_log;
853                 writeio_log = mp->m_writeio_log;
854         }
855
856         if (sbp->sb_blocklog > readio_log) {
857                 mp->m_readio_log = sbp->sb_blocklog;
858         } else {
859                 mp->m_readio_log = readio_log;
860         }
861         mp->m_readio_blocks = 1 << (mp->m_readio_log - sbp->sb_blocklog);
862         if (sbp->sb_blocklog > writeio_log) {
863                 mp->m_writeio_log = sbp->sb_blocklog;
864         } else {
865                 mp->m_writeio_log = writeio_log;
866         }
867         mp->m_writeio_blocks = 1 << (mp->m_writeio_log - sbp->sb_blocklog);
868 }
869
870 /*
871  * Set whether we're using inode alignment.
872  */
873 STATIC void
874 xfs_set_inoalignment(xfs_mount_t *mp)
875 {
876         if (xfs_sb_version_hasalign(&mp->m_sb) &&
877             mp->m_sb.sb_inoalignmt >=
878             XFS_B_TO_FSBT(mp, mp->m_inode_cluster_size))
879                 mp->m_inoalign_mask = mp->m_sb.sb_inoalignmt - 1;
880         else
881                 mp->m_inoalign_mask = 0;
882         /*
883          * If we are using stripe alignment, check whether
884          * the stripe unit is a multiple of the inode alignment
885          */
886         if (mp->m_dalign && mp->m_inoalign_mask &&
887             !(mp->m_dalign & mp->m_inoalign_mask))
888                 mp->m_sinoalign = mp->m_dalign;
889         else
890                 mp->m_sinoalign = 0;
891 }
892
893 /*
894  * Check that the data (and log if separate) are an ok size.
895  */
896 STATIC int
897 xfs_check_sizes(xfs_mount_t *mp, int mfsi_flags)
898 {
899         xfs_buf_t       *bp;
900         xfs_daddr_t     d;
901         int             error;
902
903         d = (xfs_daddr_t)XFS_FSB_TO_BB(mp, mp->m_sb.sb_dblocks);
904         if (XFS_BB_TO_FSB(mp, d) != mp->m_sb.sb_dblocks) {
905                 cmn_err(CE_WARN, "XFS: size check 1 failed");
906                 return XFS_ERROR(E2BIG);
907         }
908         error = xfs_read_buf(mp, mp->m_ddev_targp,
909                              d - XFS_FSS_TO_BB(mp, 1),
910                              XFS_FSS_TO_BB(mp, 1), 0, &bp);
911         if (!error) {
912                 xfs_buf_relse(bp);
913         } else {
914                 cmn_err(CE_WARN, "XFS: size check 2 failed");
915                 if (error == ENOSPC)
916                         error = XFS_ERROR(E2BIG);
917                 return error;
918         }
919
920         if (((mfsi_flags & XFS_MFSI_CLIENT) == 0) &&
921             mp->m_logdev_targp != mp->m_ddev_targp) {
922                 d = (xfs_daddr_t)XFS_FSB_TO_BB(mp, mp->m_sb.sb_logblocks);
923                 if (XFS_BB_TO_FSB(mp, d) != mp->m_sb.sb_logblocks) {
924                         cmn_err(CE_WARN, "XFS: size check 3 failed");
925                         return XFS_ERROR(E2BIG);
926                 }
927                 error = xfs_read_buf(mp, mp->m_logdev_targp,
928                                      d - XFS_FSB_TO_BB(mp, 1),
929                                      XFS_FSB_TO_BB(mp, 1), 0, &bp);
930                 if (!error) {
931                         xfs_buf_relse(bp);
932                 } else {
933                         cmn_err(CE_WARN, "XFS: size check 3 failed");
934                         if (error == ENOSPC)
935                                 error = XFS_ERROR(E2BIG);
936                         return error;
937                 }
938         }
939         return 0;
940 }
941
942 /*
943  * xfs_mountfs
944  *
945  * This function does the following on an initial mount of a file system:
946  *      - reads the superblock from disk and init the mount struct
947  *      - if we're a 32-bit kernel, do a size check on the superblock
948  *              so we don't mount terabyte filesystems
949  *      - init mount struct realtime fields
950  *      - allocate inode hash table for fs
951  *      - init directory manager
952  *      - perform recovery and init the log manager
953  */
954 int
955 xfs_mountfs(
956         xfs_mount_t     *mp,
957         int             mfsi_flags)
958 {
959         xfs_sb_t        *sbp = &(mp->m_sb);
960         xfs_inode_t     *rip;
961         __uint64_t      resblks;
962         __int64_t       update_flags = 0LL;
963         uint            quotamount, quotaflags;
964         int             agno;
965         int             uuid_mounted = 0;
966         int             error = 0;
967
968         xfs_mount_common(mp, sbp);
969
970         /*
971          * Check for a mismatched features2 values.  Older kernels
972          * read & wrote into the wrong sb offset for sb_features2
973          * on some platforms due to xfs_sb_t not being 64bit size aligned
974          * when sb_features2 was added, which made older superblock
975          * reading/writing routines swap it as a 64-bit value.
976          *
977          * For backwards compatibility, we make both slots equal.
978          *
979          * If we detect a mismatched field, we OR the set bits into the
980          * existing features2 field in case it has already been modified; we
981          * don't want to lose any features.  We then update the bad location
982          * with the ORed value so that older kernels will see any features2
983          * flags, and mark the two fields as needing updates once the
984          * transaction subsystem is online.
985          */
986         if (xfs_sb_has_mismatched_features2(sbp)) {
987                 cmn_err(CE_WARN,
988                         "XFS: correcting sb_features alignment problem");
989                 sbp->sb_features2 |= sbp->sb_bad_features2;
990                 sbp->sb_bad_features2 = sbp->sb_features2;
991                 update_flags |= XFS_SB_FEATURES2 | XFS_SB_BAD_FEATURES2;
992
993                 /*
994                  * Re-check for ATTR2 in case it was found in bad_features2
995                  * slot.
996                  */
997                 if (xfs_sb_version_hasattr2(&mp->m_sb) &&
998                    !(mp->m_flags & XFS_MOUNT_NOATTR2))
999                         mp->m_flags |= XFS_MOUNT_ATTR2;
1000         }
1001
1002         if (xfs_sb_version_hasattr2(&mp->m_sb) &&
1003            (mp->m_flags & XFS_MOUNT_NOATTR2)) {
1004                 xfs_sb_version_removeattr2(&mp->m_sb);
1005                 update_flags |= XFS_SB_FEATURES2;
1006
1007                 /* update sb_versionnum for the clearing of the morebits */
1008                 if (!sbp->sb_features2)
1009                         update_flags |= XFS_SB_VERSIONNUM;
1010         }
1011
1012         /*
1013          * Check if sb_agblocks is aligned at stripe boundary
1014          * If sb_agblocks is NOT aligned turn off m_dalign since
1015          * allocator alignment is within an ag, therefore ag has
1016          * to be aligned at stripe boundary.
1017          */
1018         error = xfs_update_alignment(mp, mfsi_flags, &update_flags);
1019         if (error)
1020                 goto error1;
1021
1022         xfs_alloc_compute_maxlevels(mp);
1023         xfs_bmap_compute_maxlevels(mp, XFS_DATA_FORK);
1024         xfs_bmap_compute_maxlevels(mp, XFS_ATTR_FORK);
1025         xfs_ialloc_compute_maxlevels(mp);
1026
1027         xfs_set_maxicount(mp);
1028
1029         mp->m_maxioffset = xfs_max_file_offset(sbp->sb_blocklog);
1030
1031         /*
1032          * XFS uses the uuid from the superblock as the unique
1033          * identifier for fsid.  We can not use the uuid from the volume
1034          * since a single partition filesystem is identical to a single
1035          * partition volume/filesystem.
1036          */
1037         if ((mfsi_flags & XFS_MFSI_SECOND) == 0 &&
1038             (mp->m_flags & XFS_MOUNT_NOUUID) == 0) {
1039                 if (xfs_uuid_mount(mp)) {
1040                         error = XFS_ERROR(EINVAL);
1041                         goto error1;
1042                 }
1043                 uuid_mounted=1;
1044         }
1045
1046         /*
1047          * Set the minimum read and write sizes
1048          */
1049         xfs_set_rw_sizes(mp);
1050
1051         /*
1052          * Set the inode cluster size.
1053          * This may still be overridden by the file system
1054          * block size if it is larger than the chosen cluster size.
1055          */
1056         mp->m_inode_cluster_size = XFS_INODE_BIG_CLUSTER_SIZE;
1057
1058         /*
1059          * Set inode alignment fields
1060          */
1061         xfs_set_inoalignment(mp);
1062
1063         /*
1064          * Check that the data (and log if separate) are an ok size.
1065          */
1066         error = xfs_check_sizes(mp, mfsi_flags);
1067         if (error)
1068                 goto error1;
1069
1070         /*
1071          * Initialize realtime fields in the mount structure
1072          */
1073         error = xfs_rtmount_init(mp);
1074         if (error) {
1075                 cmn_err(CE_WARN, "XFS: RT mount failed");
1076                 goto error1;
1077         }
1078
1079         /*
1080          * For client case we are done now
1081          */
1082         if (mfsi_flags & XFS_MFSI_CLIENT) {
1083                 return 0;
1084         }
1085
1086         /*
1087          *  Copies the low order bits of the timestamp and the randomly
1088          *  set "sequence" number out of a UUID.
1089          */
1090         uuid_getnodeuniq(&sbp->sb_uuid, mp->m_fixedfsid);
1091
1092         mp->m_dmevmask = 0;     /* not persistent; set after each mount */
1093
1094         xfs_dir_mount(mp);
1095
1096         /*
1097          * Initialize the attribute manager's entries.
1098          */
1099         mp->m_attr_magicpct = (mp->m_sb.sb_blocksize * 37) / 100;
1100
1101         /*
1102          * Initialize the precomputed transaction reservations values.
1103          */
1104         xfs_trans_init(mp);
1105
1106         /*
1107          * Allocate and initialize the per-ag data.
1108          */
1109         init_rwsem(&mp->m_peraglock);
1110         mp->m_perag =
1111                 kmem_zalloc(sbp->sb_agcount * sizeof(xfs_perag_t), KM_SLEEP);
1112
1113         mp->m_maxagi = xfs_initialize_perag(mp, sbp->sb_agcount);
1114
1115         /*
1116          * log's mount-time initialization. Perform 1st part recovery if needed
1117          */
1118         if (likely(sbp->sb_logblocks > 0)) {    /* check for volume case */
1119                 error = xfs_log_mount(mp, mp->m_logdev_targp,
1120                                       XFS_FSB_TO_DADDR(mp, sbp->sb_logstart),
1121                                       XFS_FSB_TO_BB(mp, sbp->sb_logblocks));
1122                 if (error) {
1123                         cmn_err(CE_WARN, "XFS: log mount failed");
1124                         goto error2;
1125                 }
1126         } else {        /* No log has been defined */
1127                 cmn_err(CE_WARN, "XFS: no log defined");
1128                 XFS_ERROR_REPORT("xfs_mountfs_int(1)", XFS_ERRLEVEL_LOW, mp);
1129                 error = XFS_ERROR(EFSCORRUPTED);
1130                 goto error2;
1131         }
1132
1133         /*
1134          * Now the log is mounted, we know if it was an unclean shutdown or
1135          * not. If it was, with the first phase of recovery has completed, we
1136          * have consistent AG blocks on disk. We have not recovered EFIs yet,
1137          * but they are recovered transactionally in the second recovery phase
1138          * later.
1139          *
1140          * Hence we can safely re-initialise incore superblock counters from
1141          * the per-ag data. These may not be correct if the filesystem was not
1142          * cleanly unmounted, so we need to wait for recovery to finish before
1143          * doing this.
1144          *
1145          * If the filesystem was cleanly unmounted, then we can trust the
1146          * values in the superblock to be correct and we don't need to do
1147          * anything here.
1148          *
1149          * If we are currently making the filesystem, the initialisation will
1150          * fail as the perag data is in an undefined state.
1151          */
1152
1153         if (xfs_sb_version_haslazysbcount(&mp->m_sb) &&
1154             !XFS_LAST_UNMOUNT_WAS_CLEAN(mp) &&
1155              !mp->m_sb.sb_inprogress) {
1156                 error = xfs_initialize_perag_data(mp, sbp->sb_agcount);
1157                 if (error) {
1158                         goto error2;
1159                 }
1160         }
1161         /*
1162          * Get and sanity-check the root inode.
1163          * Save the pointer to it in the mount structure.
1164          */
1165         error = xfs_iget(mp, NULL, sbp->sb_rootino, 0, XFS_ILOCK_EXCL, &rip, 0);
1166         if (error) {
1167                 cmn_err(CE_WARN, "XFS: failed to read root inode");
1168                 goto error3;
1169         }
1170
1171         ASSERT(rip != NULL);
1172
1173         if (unlikely((rip->i_d.di_mode & S_IFMT) != S_IFDIR)) {
1174                 cmn_err(CE_WARN, "XFS: corrupted root inode");
1175                 cmn_err(CE_WARN, "Device %s - root %llu is not a directory",
1176                         XFS_BUFTARG_NAME(mp->m_ddev_targp),
1177                         (unsigned long long)rip->i_ino);
1178                 xfs_iunlock(rip, XFS_ILOCK_EXCL);
1179                 XFS_ERROR_REPORT("xfs_mountfs_int(2)", XFS_ERRLEVEL_LOW,
1180                                  mp);
1181                 error = XFS_ERROR(EFSCORRUPTED);
1182                 goto error4;
1183         }
1184         mp->m_rootip = rip;     /* save it */
1185
1186         xfs_iunlock(rip, XFS_ILOCK_EXCL);
1187
1188         /*
1189          * Initialize realtime inode pointers in the mount structure
1190          */
1191         error = xfs_rtmount_inodes(mp);
1192         if (error) {
1193                 /*
1194                  * Free up the root inode.
1195                  */
1196                 cmn_err(CE_WARN, "XFS: failed to read RT inodes");
1197                 goto error4;
1198         }
1199
1200         /*
1201          * If fs is not mounted readonly, then update the superblock changes.
1202          */
1203         if (update_flags && !(mp->m_flags & XFS_MOUNT_RDONLY)) {
1204                 error = xfs_mount_log_sb(mp, update_flags);
1205                 if (error) {
1206                         cmn_err(CE_WARN, "XFS: failed to write sb changes");
1207                         goto error4;
1208                 }
1209         }
1210
1211         /*
1212          * Initialise the XFS quota management subsystem for this mount
1213          */
1214         error = XFS_QM_INIT(mp, &quotamount, &quotaflags);
1215         if (error)
1216                 goto error4;
1217
1218         /*
1219          * Finish recovering the file system.  This part needed to be
1220          * delayed until after the root and real-time bitmap inodes
1221          * were consistently read in.
1222          */
1223         error = xfs_log_mount_finish(mp, mfsi_flags);
1224         if (error) {
1225                 cmn_err(CE_WARN, "XFS: log mount finish failed");
1226                 goto error4;
1227         }
1228
1229         /*
1230          * Complete the quota initialisation, post-log-replay component.
1231          */
1232         error = XFS_QM_MOUNT(mp, quotamount, quotaflags, mfsi_flags);
1233         if (error)
1234                 goto error4;
1235
1236         /*
1237          * Now we are mounted, reserve a small amount of unused space for
1238          * privileged transactions. This is needed so that transaction
1239          * space required for critical operations can dip into this pool
1240          * when at ENOSPC. This is needed for operations like create with
1241          * attr, unwritten extent conversion at ENOSPC, etc. Data allocations
1242          * are not allowed to use this reserved space.
1243          *
1244          * We default to 5% or 1024 fsbs of space reserved, whichever is smaller.
1245          * This may drive us straight to ENOSPC on mount, but that implies
1246          * we were already there on the last unmount. Warn if this occurs.
1247          */
1248         resblks = mp->m_sb.sb_dblocks;
1249         do_div(resblks, 20);
1250         resblks = min_t(__uint64_t, resblks, 1024);
1251         error = xfs_reserve_blocks(mp, &resblks, NULL);
1252         if (error)
1253                 cmn_err(CE_WARN, "XFS: Unable to allocate reserve blocks. "
1254                                 "Continuing without a reserve pool.");
1255
1256         return 0;
1257
1258  error4:
1259         /*
1260          * Free up the root inode.
1261          */
1262         IRELE(rip);
1263  error3:
1264         xfs_log_unmount_dealloc(mp);
1265  error2:
1266         for (agno = 0; agno < sbp->sb_agcount; agno++)
1267                 if (mp->m_perag[agno].pagb_list)
1268                         kmem_free(mp->m_perag[agno].pagb_list,
1269                           sizeof(xfs_perag_busy_t) * XFS_PAGB_NUM_SLOTS);
1270         kmem_free(mp->m_perag, sbp->sb_agcount * sizeof(xfs_perag_t));
1271         mp->m_perag = NULL;
1272         /* FALLTHROUGH */
1273  error1:
1274         if (uuid_mounted)
1275                 xfs_uuid_unmount(mp);
1276         xfs_freesb(mp);
1277         return error;
1278 }
1279
1280 /*
1281  * xfs_unmountfs
1282  *
1283  * This flushes out the inodes,dquots and the superblock, unmounts the
1284  * log and makes sure that incore structures are freed.
1285  */
1286 int
1287 xfs_unmountfs(xfs_mount_t *mp, struct cred *cr)
1288 {
1289         __uint64_t      resblks;
1290         int             error = 0;
1291
1292         /*
1293          * We can potentially deadlock here if we have an inode cluster
1294          * that has been freed has it's buffer still pinned in memory because
1295          * the transaction is still sitting in a iclog. The stale inodes
1296          * on that buffer will have their flush locks held until the
1297          * transaction hits the disk and the callbacks run. the inode
1298          * flush takes the flush lock unconditionally and with nothing to
1299          * push out the iclog we will never get that unlocked. hence we
1300          * need to force the log first.
1301          */
1302         xfs_log_force(mp, (xfs_lsn_t)0, XFS_LOG_FORCE | XFS_LOG_SYNC);
1303         xfs_iflush_all(mp);
1304
1305         XFS_QM_DQPURGEALL(mp, XFS_QMOPT_QUOTALL | XFS_QMOPT_UMOUNTING);
1306
1307         /*
1308          * Flush out the log synchronously so that we know for sure
1309          * that nothing is pinned.  This is important because bflush()
1310          * will skip pinned buffers.
1311          */
1312         xfs_log_force(mp, (xfs_lsn_t)0, XFS_LOG_FORCE | XFS_LOG_SYNC);
1313
1314         xfs_binval(mp->m_ddev_targp);
1315         if (mp->m_rtdev_targp) {
1316                 xfs_binval(mp->m_rtdev_targp);
1317         }
1318
1319         /*
1320          * Unreserve any blocks we have so that when we unmount we don't account
1321          * the reserved free space as used. This is really only necessary for
1322          * lazy superblock counting because it trusts the incore superblock
1323          * counters to be aboslutely correct on clean unmount.
1324          *
1325          * We don't bother correcting this elsewhere for lazy superblock
1326          * counting because on mount of an unclean filesystem we reconstruct the
1327          * correct counter value and this is irrelevant.
1328          *
1329          * For non-lazy counter filesystems, this doesn't matter at all because
1330          * we only every apply deltas to the superblock and hence the incore
1331          * value does not matter....
1332          */
1333         resblks = 0;
1334         error = xfs_reserve_blocks(mp, &resblks, NULL);
1335         if (error)
1336                 cmn_err(CE_WARN, "XFS: Unable to free reserved block pool. "
1337                                 "Freespace may not be correct on next mount.");
1338
1339         error = xfs_log_sbcount(mp, 1);
1340         if (error)
1341                 cmn_err(CE_WARN, "XFS: Unable to update superblock counters. "
1342                                 "Freespace may not be correct on next mount.");
1343         xfs_unmountfs_writesb(mp);
1344         xfs_unmountfs_wait(mp);                 /* wait for async bufs */
1345         xfs_log_unmount(mp);                    /* Done! No more fs ops. */
1346
1347         xfs_freesb(mp);
1348
1349         /*
1350          * All inodes from this mount point should be freed.
1351          */
1352         ASSERT(mp->m_inodes == NULL);
1353
1354         xfs_unmountfs_close(mp, cr);
1355         if ((mp->m_flags & XFS_MOUNT_NOUUID) == 0)
1356                 xfs_uuid_unmount(mp);
1357
1358 #if defined(DEBUG) || defined(INDUCE_IO_ERROR)
1359         xfs_errortag_clearall(mp, 0);
1360 #endif
1361         xfs_mount_free(mp);
1362         return 0;
1363 }
1364
1365 void
1366 xfs_unmountfs_close(xfs_mount_t *mp, struct cred *cr)
1367 {
1368         if (mp->m_logdev_targp && mp->m_logdev_targp != mp->m_ddev_targp)
1369                 xfs_free_buftarg(mp->m_logdev_targp, 1);
1370         if (mp->m_rtdev_targp)
1371                 xfs_free_buftarg(mp->m_rtdev_targp, 1);
1372         xfs_free_buftarg(mp->m_ddev_targp, 0);
1373 }
1374
1375 STATIC void
1376 xfs_unmountfs_wait(xfs_mount_t *mp)
1377 {
1378         if (mp->m_logdev_targp != mp->m_ddev_targp)
1379                 xfs_wait_buftarg(mp->m_logdev_targp);
1380         if (mp->m_rtdev_targp)
1381                 xfs_wait_buftarg(mp->m_rtdev_targp);
1382         xfs_wait_buftarg(mp->m_ddev_targp);
1383 }
1384
1385 int
1386 xfs_fs_writable(xfs_mount_t *mp)
1387 {
1388         return !(xfs_test_for_freeze(mp) || XFS_FORCED_SHUTDOWN(mp) ||
1389                 (mp->m_flags & XFS_MOUNT_RDONLY));
1390 }
1391
1392 /*
1393  * xfs_log_sbcount
1394  *
1395  * Called either periodically to keep the on disk superblock values
1396  * roughly up to date or from unmount to make sure the values are
1397  * correct on a clean unmount.
1398  *
1399  * Note this code can be called during the process of freezing, so
1400  * we may need to use the transaction allocator which does not not
1401  * block when the transaction subsystem is in its frozen state.
1402  */
1403 int
1404 xfs_log_sbcount(
1405         xfs_mount_t     *mp,
1406         uint            sync)
1407 {
1408         xfs_trans_t     *tp;
1409         int             error;
1410
1411         if (!xfs_fs_writable(mp))
1412                 return 0;
1413
1414         xfs_icsb_sync_counters(mp, 0);
1415
1416         /*
1417          * we don't need to do this if we are updating the superblock
1418          * counters on every modification.
1419          */
1420         if (!xfs_sb_version_haslazysbcount(&mp->m_sb))
1421                 return 0;
1422
1423         tp = _xfs_trans_alloc(mp, XFS_TRANS_SB_COUNT);
1424         error = xfs_trans_reserve(tp, 0, mp->m_sb.sb_sectsize + 128, 0, 0,
1425                                         XFS_DEFAULT_LOG_COUNT);
1426         if (error) {
1427                 xfs_trans_cancel(tp, 0);
1428                 return error;
1429         }
1430
1431         xfs_mod_sb(tp, XFS_SB_IFREE | XFS_SB_ICOUNT | XFS_SB_FDBLOCKS);
1432         if (sync)
1433                 xfs_trans_set_sync(tp);
1434         error = xfs_trans_commit(tp, 0);
1435         return error;
1436 }
1437
1438 STATIC void
1439 xfs_mark_shared_ro(
1440         xfs_mount_t     *mp,
1441         xfs_buf_t       *bp)
1442 {
1443         xfs_dsb_t       *sb = XFS_BUF_TO_SBP(bp);
1444         __uint16_t      version;
1445
1446         if (!(sb->sb_flags & XFS_SBF_READONLY))
1447                 sb->sb_flags |= XFS_SBF_READONLY;
1448
1449         version = be16_to_cpu(sb->sb_versionnum);
1450         if ((version & XFS_SB_VERSION_NUMBITS) != XFS_SB_VERSION_4 ||
1451             !(version & XFS_SB_VERSION_SHAREDBIT))
1452                 version |= XFS_SB_VERSION_SHAREDBIT;
1453         sb->sb_versionnum = cpu_to_be16(version);
1454 }
1455
1456 int
1457 xfs_unmountfs_writesb(xfs_mount_t *mp)
1458 {
1459         xfs_buf_t       *sbp;
1460         int             error = 0;
1461
1462         /*
1463          * skip superblock write if fs is read-only, or
1464          * if we are doing a forced umount.
1465          */
1466         if (!((mp->m_flags & XFS_MOUNT_RDONLY) ||
1467                 XFS_FORCED_SHUTDOWN(mp))) {
1468
1469                 sbp = xfs_getsb(mp, 0);
1470
1471                 /*
1472                  * mark shared-readonly if desired
1473                  */
1474                 if (mp->m_mk_sharedro)
1475                         xfs_mark_shared_ro(mp, sbp);
1476
1477                 XFS_BUF_UNDONE(sbp);
1478                 XFS_BUF_UNREAD(sbp);
1479                 XFS_BUF_UNDELAYWRITE(sbp);
1480                 XFS_BUF_WRITE(sbp);
1481                 XFS_BUF_UNASYNC(sbp);
1482                 ASSERT(XFS_BUF_TARGET(sbp) == mp->m_ddev_targp);
1483                 xfsbdstrat(mp, sbp);
1484                 error = xfs_iowait(sbp);
1485                 if (error)
1486                         xfs_ioerror_alert("xfs_unmountfs_writesb",
1487                                           mp, sbp, XFS_BUF_ADDR(sbp));
1488                 if (error && mp->m_mk_sharedro)
1489                         xfs_fs_cmn_err(CE_ALERT, mp, "Superblock write error detected while unmounting.  Filesystem may not be marked shared readonly");
1490                 xfs_buf_relse(sbp);
1491         }
1492         return error;
1493 }
1494
1495 /*
1496  * xfs_mod_sb() can be used to copy arbitrary changes to the
1497  * in-core superblock into the superblock buffer to be logged.
1498  * It does not provide the higher level of locking that is
1499  * needed to protect the in-core superblock from concurrent
1500  * access.
1501  */
1502 void
1503 xfs_mod_sb(xfs_trans_t *tp, __int64_t fields)
1504 {
1505         xfs_buf_t       *bp;
1506         int             first;
1507         int             last;
1508         xfs_mount_t     *mp;
1509         xfs_sb_field_t  f;
1510
1511         ASSERT(fields);
1512         if (!fields)
1513                 return;
1514         mp = tp->t_mountp;
1515         bp = xfs_trans_getsb(tp, mp, 0);
1516         first = sizeof(xfs_sb_t);
1517         last = 0;
1518
1519         /* translate/copy */
1520
1521         xfs_sb_to_disk(XFS_BUF_TO_SBP(bp), &mp->m_sb, fields);
1522
1523         /* find modified range */
1524
1525         f = (xfs_sb_field_t)xfs_lowbit64((__uint64_t)fields);
1526         ASSERT((1LL << f) & XFS_SB_MOD_BITS);
1527         first = xfs_sb_info[f].offset;
1528
1529         f = (xfs_sb_field_t)xfs_highbit64((__uint64_t)fields);
1530         ASSERT((1LL << f) & XFS_SB_MOD_BITS);
1531         last = xfs_sb_info[f + 1].offset - 1;
1532
1533         xfs_trans_log_buf(tp, bp, first, last);
1534 }
1535
1536
1537 /*
1538  * xfs_mod_incore_sb_unlocked() is a utility routine common used to apply
1539  * a delta to a specified field in the in-core superblock.  Simply
1540  * switch on the field indicated and apply the delta to that field.
1541  * Fields are not allowed to dip below zero, so if the delta would
1542  * do this do not apply it and return EINVAL.
1543  *
1544  * The m_sb_lock must be held when this routine is called.
1545  */
1546 int
1547 xfs_mod_incore_sb_unlocked(
1548         xfs_mount_t     *mp,
1549         xfs_sb_field_t  field,
1550         int64_t         delta,
1551         int             rsvd)
1552 {
1553         int             scounter;       /* short counter for 32 bit fields */
1554         long long       lcounter;       /* long counter for 64 bit fields */
1555         long long       res_used, rem;
1556
1557         /*
1558          * With the in-core superblock spin lock held, switch
1559          * on the indicated field.  Apply the delta to the
1560          * proper field.  If the fields value would dip below
1561          * 0, then do not apply the delta and return EINVAL.
1562          */
1563         switch (field) {
1564         case XFS_SBS_ICOUNT:
1565                 lcounter = (long long)mp->m_sb.sb_icount;
1566                 lcounter += delta;
1567                 if (lcounter < 0) {
1568                         ASSERT(0);
1569                         return XFS_ERROR(EINVAL);
1570                 }
1571                 mp->m_sb.sb_icount = lcounter;
1572                 return 0;
1573         case XFS_SBS_IFREE:
1574                 lcounter = (long long)mp->m_sb.sb_ifree;
1575                 lcounter += delta;
1576                 if (lcounter < 0) {
1577                         ASSERT(0);
1578                         return XFS_ERROR(EINVAL);
1579                 }
1580                 mp->m_sb.sb_ifree = lcounter;
1581                 return 0;
1582         case XFS_SBS_FDBLOCKS:
1583                 lcounter = (long long)
1584                         mp->m_sb.sb_fdblocks - XFS_ALLOC_SET_ASIDE(mp);
1585                 res_used = (long long)(mp->m_resblks - mp->m_resblks_avail);
1586
1587                 if (delta > 0) {                /* Putting blocks back */
1588                         if (res_used > delta) {
1589                                 mp->m_resblks_avail += delta;
1590                         } else {
1591                                 rem = delta - res_used;
1592                                 mp->m_resblks_avail = mp->m_resblks;
1593                                 lcounter += rem;
1594                         }
1595                 } else {                                /* Taking blocks away */
1596
1597                         lcounter += delta;
1598
1599                 /*
1600                  * If were out of blocks, use any available reserved blocks if
1601                  * were allowed to.
1602                  */
1603
1604                         if (lcounter < 0) {
1605                                 if (rsvd) {
1606                                         lcounter = (long long)mp->m_resblks_avail + delta;
1607                                         if (lcounter < 0) {
1608                                                 return XFS_ERROR(ENOSPC);
1609                                         }
1610                                         mp->m_resblks_avail = lcounter;
1611                                         return 0;
1612                                 } else {        /* not reserved */
1613                                         return XFS_ERROR(ENOSPC);
1614                                 }
1615                         }
1616                 }
1617
1618                 mp->m_sb.sb_fdblocks = lcounter + XFS_ALLOC_SET_ASIDE(mp);
1619                 return 0;
1620         case XFS_SBS_FREXTENTS:
1621                 lcounter = (long long)mp->m_sb.sb_frextents;
1622                 lcounter += delta;
1623                 if (lcounter < 0) {
1624                         return XFS_ERROR(ENOSPC);
1625                 }
1626                 mp->m_sb.sb_frextents = lcounter;
1627                 return 0;
1628         case XFS_SBS_DBLOCKS:
1629                 lcounter = (long long)mp->m_sb.sb_dblocks;
1630                 lcounter += delta;
1631                 if (lcounter < 0) {
1632                         ASSERT(0);
1633                         return XFS_ERROR(EINVAL);
1634                 }
1635                 mp->m_sb.sb_dblocks = lcounter;
1636                 return 0;
1637         case XFS_SBS_AGCOUNT:
1638                 scounter = mp->m_sb.sb_agcount;
1639                 scounter += delta;
1640                 if (scounter < 0) {
1641                         ASSERT(0);
1642                         return XFS_ERROR(EINVAL);
1643                 }
1644                 mp->m_sb.sb_agcount = scounter;
1645                 return 0;
1646         case XFS_SBS_IMAX_PCT:
1647                 scounter = mp->m_sb.sb_imax_pct;
1648                 scounter += delta;
1649                 if (scounter < 0) {
1650                         ASSERT(0);
1651                         return XFS_ERROR(EINVAL);
1652                 }
1653                 mp->m_sb.sb_imax_pct = scounter;
1654                 return 0;
1655         case XFS_SBS_REXTSIZE:
1656                 scounter = mp->m_sb.sb_rextsize;
1657                 scounter += delta;
1658                 if (scounter < 0) {
1659                         ASSERT(0);
1660                         return XFS_ERROR(EINVAL);
1661                 }
1662                 mp->m_sb.sb_rextsize = scounter;
1663                 return 0;
1664         case XFS_SBS_RBMBLOCKS:
1665                 scounter = mp->m_sb.sb_rbmblocks;
1666                 scounter += delta;
1667                 if (scounter < 0) {
1668                         ASSERT(0);
1669                         return XFS_ERROR(EINVAL);
1670                 }
1671                 mp->m_sb.sb_rbmblocks = scounter;
1672                 return 0;
1673         case XFS_SBS_RBLOCKS:
1674                 lcounter = (long long)mp->m_sb.sb_rblocks;
1675                 lcounter += delta;
1676                 if (lcounter < 0) {
1677                         ASSERT(0);
1678                         return XFS_ERROR(EINVAL);
1679                 }
1680                 mp->m_sb.sb_rblocks = lcounter;
1681                 return 0;
1682         case XFS_SBS_REXTENTS:
1683                 lcounter = (long long)mp->m_sb.sb_rextents;
1684                 lcounter += delta;
1685                 if (lcounter < 0) {
1686                         ASSERT(0);
1687                         return XFS_ERROR(EINVAL);
1688                 }
1689                 mp->m_sb.sb_rextents = lcounter;
1690                 return 0;
1691         case XFS_SBS_REXTSLOG:
1692                 scounter = mp->m_sb.sb_rextslog;
1693                 scounter += delta;
1694                 if (scounter < 0) {
1695                         ASSERT(0);
1696                         return XFS_ERROR(EINVAL);
1697                 }
1698                 mp->m_sb.sb_rextslog = scounter;
1699                 return 0;
1700         default:
1701                 ASSERT(0);
1702                 return XFS_ERROR(EINVAL);
1703         }
1704 }
1705
1706 /*
1707  * xfs_mod_incore_sb() is used to change a field in the in-core
1708  * superblock structure by the specified delta.  This modification
1709  * is protected by the m_sb_lock.  Just use the xfs_mod_incore_sb_unlocked()
1710  * routine to do the work.
1711  */
1712 int
1713 xfs_mod_incore_sb(
1714         xfs_mount_t     *mp,
1715         xfs_sb_field_t  field,
1716         int64_t         delta,
1717         int             rsvd)
1718 {
1719         int     status;
1720
1721         /* check for per-cpu counters */
1722         switch (field) {
1723 #ifdef HAVE_PERCPU_SB
1724         case XFS_SBS_ICOUNT:
1725         case XFS_SBS_IFREE:
1726         case XFS_SBS_FDBLOCKS:
1727                 if (!(mp->m_flags & XFS_MOUNT_NO_PERCPU_SB)) {
1728                         status = xfs_icsb_modify_counters(mp, field,
1729                                                         delta, rsvd);
1730                         break;
1731                 }
1732                 /* FALLTHROUGH */
1733 #endif
1734         default:
1735                 spin_lock(&mp->m_sb_lock);
1736                 status = xfs_mod_incore_sb_unlocked(mp, field, delta, rsvd);
1737                 spin_unlock(&mp->m_sb_lock);
1738                 break;
1739         }
1740
1741         return status;
1742 }
1743
1744 /*
1745  * xfs_mod_incore_sb_batch() is used to change more than one field
1746  * in the in-core superblock structure at a time.  This modification
1747  * is protected by a lock internal to this module.  The fields and
1748  * changes to those fields are specified in the array of xfs_mod_sb
1749  * structures passed in.
1750  *
1751  * Either all of the specified deltas will be applied or none of
1752  * them will.  If any modified field dips below 0, then all modifications
1753  * will be backed out and EINVAL will be returned.
1754  */
1755 int
1756 xfs_mod_incore_sb_batch(xfs_mount_t *mp, xfs_mod_sb_t *msb, uint nmsb, int rsvd)
1757 {
1758         int             status=0;
1759         xfs_mod_sb_t    *msbp;
1760
1761         /*
1762          * Loop through the array of mod structures and apply each
1763          * individually.  If any fail, then back out all those
1764          * which have already been applied.  Do all of this within
1765          * the scope of the m_sb_lock so that all of the changes will
1766          * be atomic.
1767          */
1768         spin_lock(&mp->m_sb_lock);
1769         msbp = &msb[0];
1770         for (msbp = &msbp[0]; msbp < (msb + nmsb); msbp++) {
1771                 /*
1772                  * Apply the delta at index n.  If it fails, break
1773                  * from the loop so we'll fall into the undo loop
1774                  * below.
1775                  */
1776                 switch (msbp->msb_field) {
1777 #ifdef HAVE_PERCPU_SB
1778                 case XFS_SBS_ICOUNT:
1779                 case XFS_SBS_IFREE:
1780                 case XFS_SBS_FDBLOCKS:
1781                         if (!(mp->m_flags & XFS_MOUNT_NO_PERCPU_SB)) {
1782                                 spin_unlock(&mp->m_sb_lock);
1783                                 status = xfs_icsb_modify_counters(mp,
1784                                                         msbp->msb_field,
1785                                                         msbp->msb_delta, rsvd);
1786                                 spin_lock(&mp->m_sb_lock);
1787                                 break;
1788                         }
1789                         /* FALLTHROUGH */
1790 #endif
1791                 default:
1792                         status = xfs_mod_incore_sb_unlocked(mp,
1793                                                 msbp->msb_field,
1794                                                 msbp->msb_delta, rsvd);
1795                         break;
1796                 }
1797
1798                 if (status != 0) {
1799                         break;
1800                 }
1801         }
1802
1803         /*
1804          * If we didn't complete the loop above, then back out
1805          * any changes made to the superblock.  If you add code
1806          * between the loop above and here, make sure that you
1807          * preserve the value of status. Loop back until
1808          * we step below the beginning of the array.  Make sure
1809          * we don't touch anything back there.
1810          */
1811         if (status != 0) {
1812                 msbp--;
1813                 while (msbp >= msb) {
1814                         switch (msbp->msb_field) {
1815 #ifdef HAVE_PERCPU_SB
1816                         case XFS_SBS_ICOUNT:
1817                         case XFS_SBS_IFREE:
1818                         case XFS_SBS_FDBLOCKS:
1819                                 if (!(mp->m_flags & XFS_MOUNT_NO_PERCPU_SB)) {
1820                                         spin_unlock(&mp->m_sb_lock);
1821                                         status = xfs_icsb_modify_counters(mp,
1822                                                         msbp->msb_field,
1823                                                         -(msbp->msb_delta),
1824                                                         rsvd);
1825                                         spin_lock(&mp->m_sb_lock);
1826                                         break;
1827                                 }
1828                                 /* FALLTHROUGH */
1829 #endif
1830                         default:
1831                                 status = xfs_mod_incore_sb_unlocked(mp,
1832                                                         msbp->msb_field,
1833                                                         -(msbp->msb_delta),
1834                                                         rsvd);
1835                                 break;
1836                         }
1837                         ASSERT(status == 0);
1838                         msbp--;
1839                 }
1840         }
1841         spin_unlock(&mp->m_sb_lock);
1842         return status;
1843 }
1844
1845 /*
1846  * xfs_getsb() is called to obtain the buffer for the superblock.
1847  * The buffer is returned locked and read in from disk.
1848  * The buffer should be released with a call to xfs_brelse().
1849  *
1850  * If the flags parameter is BUF_TRYLOCK, then we'll only return
1851  * the superblock buffer if it can be locked without sleeping.
1852  * If it can't then we'll return NULL.
1853  */
1854 xfs_buf_t *
1855 xfs_getsb(
1856         xfs_mount_t     *mp,
1857         int             flags)
1858 {
1859         xfs_buf_t       *bp;
1860
1861         ASSERT(mp->m_sb_bp != NULL);
1862         bp = mp->m_sb_bp;
1863         if (flags & XFS_BUF_TRYLOCK) {
1864                 if (!XFS_BUF_CPSEMA(bp)) {
1865                         return NULL;
1866                 }
1867         } else {
1868                 XFS_BUF_PSEMA(bp, PRIBIO);
1869         }
1870         XFS_BUF_HOLD(bp);
1871         ASSERT(XFS_BUF_ISDONE(bp));
1872         return bp;
1873 }
1874
1875 /*
1876  * Used to free the superblock along various error paths.
1877  */
1878 void
1879 xfs_freesb(
1880         xfs_mount_t     *mp)
1881 {
1882         xfs_buf_t       *bp;
1883
1884         /*
1885          * Use xfs_getsb() so that the buffer will be locked
1886          * when we call xfs_buf_relse().
1887          */
1888         bp = xfs_getsb(mp, 0);
1889         XFS_BUF_UNMANAGE(bp);
1890         xfs_buf_relse(bp);
1891         mp->m_sb_bp = NULL;
1892 }
1893
1894 /*
1895  * See if the UUID is unique among mounted XFS filesystems.
1896  * Mount fails if UUID is nil or a FS with the same UUID is already mounted.
1897  */
1898 STATIC int
1899 xfs_uuid_mount(
1900         xfs_mount_t     *mp)
1901 {
1902         if (uuid_is_nil(&mp->m_sb.sb_uuid)) {
1903                 cmn_err(CE_WARN,
1904                         "XFS: Filesystem %s has nil UUID - can't mount",
1905                         mp->m_fsname);
1906                 return -1;
1907         }
1908         if (!uuid_table_insert(&mp->m_sb.sb_uuid)) {
1909                 cmn_err(CE_WARN,
1910                         "XFS: Filesystem %s has duplicate UUID - can't mount",
1911                         mp->m_fsname);
1912                 return -1;
1913         }
1914         return 0;
1915 }
1916
1917 /*
1918  * Remove filesystem from the UUID table.
1919  */
1920 STATIC void
1921 xfs_uuid_unmount(
1922         xfs_mount_t     *mp)
1923 {
1924         uuid_table_remove(&mp->m_sb.sb_uuid);
1925 }
1926
1927 /*
1928  * Used to log changes to the superblock unit and width fields which could
1929  * be altered by the mount options, as well as any potential sb_features2
1930  * fixup. Only the first superblock is updated.
1931  */
1932 STATIC int
1933 xfs_mount_log_sb(
1934         xfs_mount_t     *mp,
1935         __int64_t       fields)
1936 {
1937         xfs_trans_t     *tp;
1938         int             error;
1939
1940         ASSERT(fields & (XFS_SB_UNIT | XFS_SB_WIDTH | XFS_SB_UUID |
1941                          XFS_SB_FEATURES2 | XFS_SB_BAD_FEATURES2));
1942
1943         tp = xfs_trans_alloc(mp, XFS_TRANS_SB_UNIT);
1944         error = xfs_trans_reserve(tp, 0, mp->m_sb.sb_sectsize + 128, 0, 0,
1945                                 XFS_DEFAULT_LOG_COUNT);
1946         if (error) {
1947                 xfs_trans_cancel(tp, 0);
1948                 return error;
1949         }
1950         xfs_mod_sb(tp, fields);
1951         error = xfs_trans_commit(tp, 0);
1952         return error;
1953 }
1954
1955
1956 #ifdef HAVE_PERCPU_SB
1957 /*
1958  * Per-cpu incore superblock counters
1959  *
1960  * Simple concept, difficult implementation
1961  *
1962  * Basically, replace the incore superblock counters with a distributed per cpu
1963  * counter for contended fields (e.g.  free block count).
1964  *
1965  * Difficulties arise in that the incore sb is used for ENOSPC checking, and
1966  * hence needs to be accurately read when we are running low on space. Hence
1967  * there is a method to enable and disable the per-cpu counters based on how
1968  * much "stuff" is available in them.
1969  *
1970  * Basically, a counter is enabled if there is enough free resource to justify
1971  * running a per-cpu fast-path. If the per-cpu counter runs out (i.e. a local
1972  * ENOSPC), then we disable the counters to synchronise all callers and
1973  * re-distribute the available resources.
1974  *
1975  * If, once we redistributed the available resources, we still get a failure,
1976  * we disable the per-cpu counter and go through the slow path.
1977  *
1978  * The slow path is the current xfs_mod_incore_sb() function.  This means that
1979  * when we disable a per-cpu counter, we need to drain it's resources back to
1980  * the global superblock. We do this after disabling the counter to prevent
1981  * more threads from queueing up on the counter.
1982  *
1983  * Essentially, this means that we still need a lock in the fast path to enable
1984  * synchronisation between the global counters and the per-cpu counters. This
1985  * is not a problem because the lock will be local to a CPU almost all the time
1986  * and have little contention except when we get to ENOSPC conditions.
1987  *
1988  * Basically, this lock becomes a barrier that enables us to lock out the fast
1989  * path while we do things like enabling and disabling counters and
1990  * synchronising the counters.
1991  *
1992  * Locking rules:
1993  *
1994  *      1. m_sb_lock before picking up per-cpu locks
1995  *      2. per-cpu locks always picked up via for_each_online_cpu() order
1996  *      3. accurate counter sync requires m_sb_lock + per cpu locks
1997  *      4. modifying per-cpu counters requires holding per-cpu lock
1998  *      5. modifying global counters requires holding m_sb_lock
1999  *      6. enabling or disabling a counter requires holding the m_sb_lock 
2000  *         and _none_ of the per-cpu locks.
2001  *
2002  * Disabled counters are only ever re-enabled by a balance operation
2003  * that results in more free resources per CPU than a given threshold.
2004  * To ensure counters don't remain disabled, they are rebalanced when
2005  * the global resource goes above a higher threshold (i.e. some hysteresis
2006  * is present to prevent thrashing).
2007  */
2008
2009 #ifdef CONFIG_HOTPLUG_CPU
2010 /*
2011  * hot-plug CPU notifier support.
2012  *
2013  * We need a notifier per filesystem as we need to be able to identify
2014  * the filesystem to balance the counters out. This is achieved by
2015  * having a notifier block embedded in the xfs_mount_t and doing pointer
2016  * magic to get the mount pointer from the notifier block address.
2017  */
2018 STATIC int
2019 xfs_icsb_cpu_notify(
2020         struct notifier_block *nfb,
2021         unsigned long action,
2022         void *hcpu)
2023 {
2024         xfs_icsb_cnts_t *cntp;
2025         xfs_mount_t     *mp;
2026
2027         mp = (xfs_mount_t *)container_of(nfb, xfs_mount_t, m_icsb_notifier);
2028         cntp = (xfs_icsb_cnts_t *)
2029                         per_cpu_ptr(mp->m_sb_cnts, (unsigned long)hcpu);
2030         switch (action) {
2031         case CPU_UP_PREPARE:
2032         case CPU_UP_PREPARE_FROZEN:
2033                 /* Easy Case - initialize the area and locks, and
2034                  * then rebalance when online does everything else for us. */
2035                 memset(cntp, 0, sizeof(xfs_icsb_cnts_t));
2036                 break;
2037         case CPU_ONLINE:
2038         case CPU_ONLINE_FROZEN:
2039                 xfs_icsb_lock(mp);
2040                 xfs_icsb_balance_counter(mp, XFS_SBS_ICOUNT, 0);
2041                 xfs_icsb_balance_counter(mp, XFS_SBS_IFREE, 0);
2042                 xfs_icsb_balance_counter(mp, XFS_SBS_FDBLOCKS, 0);
2043                 xfs_icsb_unlock(mp);
2044                 break;
2045         case CPU_DEAD:
2046         case CPU_DEAD_FROZEN:
2047                 /* Disable all the counters, then fold the dead cpu's
2048                  * count into the total on the global superblock and
2049                  * re-enable the counters. */
2050                 xfs_icsb_lock(mp);
2051                 spin_lock(&mp->m_sb_lock);
2052                 xfs_icsb_disable_counter(mp, XFS_SBS_ICOUNT);
2053                 xfs_icsb_disable_counter(mp, XFS_SBS_IFREE);
2054                 xfs_icsb_disable_counter(mp, XFS_SBS_FDBLOCKS);
2055
2056                 mp->m_sb.sb_icount += cntp->icsb_icount;
2057                 mp->m_sb.sb_ifree += cntp->icsb_ifree;
2058                 mp->m_sb.sb_fdblocks += cntp->icsb_fdblocks;
2059
2060                 memset(cntp, 0, sizeof(xfs_icsb_cnts_t));
2061
2062                 xfs_icsb_balance_counter_locked(mp, XFS_SBS_ICOUNT, 0);
2063                 xfs_icsb_balance_counter_locked(mp, XFS_SBS_IFREE, 0);
2064                 xfs_icsb_balance_counter_locked(mp, XFS_SBS_FDBLOCKS, 0);
2065                 spin_unlock(&mp->m_sb_lock);
2066                 xfs_icsb_unlock(mp);
2067                 break;
2068         }
2069
2070         return NOTIFY_OK;
2071 }
2072 #endif /* CONFIG_HOTPLUG_CPU */
2073
2074 int
2075 xfs_icsb_init_counters(
2076         xfs_mount_t     *mp)
2077 {
2078         xfs_icsb_cnts_t *cntp;
2079         int             i;
2080
2081         mp->m_sb_cnts = alloc_percpu(xfs_icsb_cnts_t);
2082         if (mp->m_sb_cnts == NULL)
2083                 return -ENOMEM;
2084
2085 #ifdef CONFIG_HOTPLUG_CPU
2086         mp->m_icsb_notifier.notifier_call = xfs_icsb_cpu_notify;
2087         mp->m_icsb_notifier.priority = 0;
2088         register_hotcpu_notifier(&mp->m_icsb_notifier);
2089 #endif /* CONFIG_HOTPLUG_CPU */
2090
2091         for_each_online_cpu(i) {
2092                 cntp = (xfs_icsb_cnts_t *)per_cpu_ptr(mp->m_sb_cnts, i);
2093                 memset(cntp, 0, sizeof(xfs_icsb_cnts_t));
2094         }
2095
2096         mutex_init(&mp->m_icsb_mutex);
2097
2098         /*
2099          * start with all counters disabled so that the
2100          * initial balance kicks us off correctly
2101          */
2102         mp->m_icsb_counters = -1;
2103         return 0;
2104 }
2105
2106 void
2107 xfs_icsb_reinit_counters(
2108         xfs_mount_t     *mp)
2109 {
2110         xfs_icsb_lock(mp);
2111         /*
2112          * start with all counters disabled so that the
2113          * initial balance kicks us off correctly
2114          */
2115         mp->m_icsb_counters = -1;
2116         xfs_icsb_balance_counter(mp, XFS_SBS_ICOUNT, 0);
2117         xfs_icsb_balance_counter(mp, XFS_SBS_IFREE, 0);
2118         xfs_icsb_balance_counter(mp, XFS_SBS_FDBLOCKS, 0);
2119         xfs_icsb_unlock(mp);
2120 }
2121
2122 STATIC void
2123 xfs_icsb_destroy_counters(
2124         xfs_mount_t     *mp)
2125 {
2126         if (mp->m_sb_cnts) {
2127                 unregister_hotcpu_notifier(&mp->m_icsb_notifier);
2128                 free_percpu(mp->m_sb_cnts);
2129         }
2130         mutex_destroy(&mp->m_icsb_mutex);
2131 }
2132
2133 STATIC_INLINE void
2134 xfs_icsb_lock_cntr(
2135         xfs_icsb_cnts_t *icsbp)
2136 {
2137         while (test_and_set_bit(XFS_ICSB_FLAG_LOCK, &icsbp->icsb_flags)) {
2138                 ndelay(1000);
2139         }
2140 }
2141
2142 STATIC_INLINE void
2143 xfs_icsb_unlock_cntr(
2144         xfs_icsb_cnts_t *icsbp)
2145 {
2146         clear_bit(XFS_ICSB_FLAG_LOCK, &icsbp->icsb_flags);
2147 }
2148
2149
2150 STATIC_INLINE void
2151 xfs_icsb_lock_all_counters(
2152         xfs_mount_t     *mp)
2153 {
2154         xfs_icsb_cnts_t *cntp;
2155         int             i;
2156
2157         for_each_online_cpu(i) {
2158                 cntp = (xfs_icsb_cnts_t *)per_cpu_ptr(mp->m_sb_cnts, i);
2159                 xfs_icsb_lock_cntr(cntp);
2160         }
2161 }
2162
2163 STATIC_INLINE void
2164 xfs_icsb_unlock_all_counters(
2165         xfs_mount_t     *mp)
2166 {
2167         xfs_icsb_cnts_t *cntp;
2168         int             i;
2169
2170         for_each_online_cpu(i) {
2171                 cntp = (xfs_icsb_cnts_t *)per_cpu_ptr(mp->m_sb_cnts, i);
2172                 xfs_icsb_unlock_cntr(cntp);
2173         }
2174 }
2175
2176 STATIC void
2177 xfs_icsb_count(
2178         xfs_mount_t     *mp,
2179         xfs_icsb_cnts_t *cnt,
2180         int             flags)
2181 {
2182         xfs_icsb_cnts_t *cntp;
2183         int             i;
2184
2185         memset(cnt, 0, sizeof(xfs_icsb_cnts_t));
2186
2187         if (!(flags & XFS_ICSB_LAZY_COUNT))
2188                 xfs_icsb_lock_all_counters(mp);
2189
2190         for_each_online_cpu(i) {
2191                 cntp = (xfs_icsb_cnts_t *)per_cpu_ptr(mp->m_sb_cnts, i);
2192                 cnt->icsb_icount += cntp->icsb_icount;
2193                 cnt->icsb_ifree += cntp->icsb_ifree;
2194                 cnt->icsb_fdblocks += cntp->icsb_fdblocks;
2195         }
2196
2197         if (!(flags & XFS_ICSB_LAZY_COUNT))
2198                 xfs_icsb_unlock_all_counters(mp);
2199 }
2200
2201 STATIC int
2202 xfs_icsb_counter_disabled(
2203         xfs_mount_t     *mp,
2204         xfs_sb_field_t  field)
2205 {
2206         ASSERT((field >= XFS_SBS_ICOUNT) && (field <= XFS_SBS_FDBLOCKS));
2207         return test_bit(field, &mp->m_icsb_counters);
2208 }
2209
2210 STATIC void
2211 xfs_icsb_disable_counter(
2212         xfs_mount_t     *mp,
2213         xfs_sb_field_t  field)
2214 {
2215         xfs_icsb_cnts_t cnt;
2216
2217         ASSERT((field >= XFS_SBS_ICOUNT) && (field <= XFS_SBS_FDBLOCKS));
2218
2219         /*
2220          * If we are already disabled, then there is nothing to do
2221          * here. We check before locking all the counters to avoid
2222          * the expensive lock operation when being called in the
2223          * slow path and the counter is already disabled. This is
2224          * safe because the only time we set or clear this state is under
2225          * the m_icsb_mutex.
2226          */
2227         if (xfs_icsb_counter_disabled(mp, field))
2228                 return;
2229
2230         xfs_icsb_lock_all_counters(mp);
2231         if (!test_and_set_bit(field, &mp->m_icsb_counters)) {
2232                 /* drain back to superblock */
2233
2234                 xfs_icsb_count(mp, &cnt, XFS_ICSB_LAZY_COUNT);
2235                 switch(field) {
2236                 case XFS_SBS_ICOUNT:
2237                         mp->m_sb.sb_icount = cnt.icsb_icount;
2238                         break;
2239                 case XFS_SBS_IFREE:
2240                         mp->m_sb.sb_ifree = cnt.icsb_ifree;
2241                         break;
2242                 case XFS_SBS_FDBLOCKS:
2243                         mp->m_sb.sb_fdblocks = cnt.icsb_fdblocks;
2244                         break;
2245                 default:
2246                         BUG();
2247                 }
2248         }
2249
2250         xfs_icsb_unlock_all_counters(mp);
2251 }
2252
2253 STATIC void
2254 xfs_icsb_enable_counter(
2255         xfs_mount_t     *mp,
2256         xfs_sb_field_t  field,
2257         uint64_t        count,
2258         uint64_t        resid)
2259 {
2260         xfs_icsb_cnts_t *cntp;
2261         int             i;
2262
2263         ASSERT((field >= XFS_SBS_ICOUNT) && (field <= XFS_SBS_FDBLOCKS));
2264
2265         xfs_icsb_lock_all_counters(mp);
2266         for_each_online_cpu(i) {
2267                 cntp = per_cpu_ptr(mp->m_sb_cnts, i);
2268                 switch (field) {
2269                 case XFS_SBS_ICOUNT:
2270                         cntp->icsb_icount = count + resid;
2271                         break;
2272                 case XFS_SBS_IFREE:
2273                         cntp->icsb_ifree = count + resid;
2274                         break;
2275                 case XFS_SBS_FDBLOCKS:
2276                         cntp->icsb_fdblocks = count + resid;
2277                         break;
2278                 default:
2279                         BUG();
2280                         break;
2281                 }
2282                 resid = 0;
2283         }
2284         clear_bit(field, &mp->m_icsb_counters);
2285         xfs_icsb_unlock_all_counters(mp);
2286 }
2287
2288 void
2289 xfs_icsb_sync_counters_locked(
2290         xfs_mount_t     *mp,
2291         int             flags)
2292 {
2293         xfs_icsb_cnts_t cnt;
2294
2295         xfs_icsb_count(mp, &cnt, flags);
2296
2297         if (!xfs_icsb_counter_disabled(mp, XFS_SBS_ICOUNT))
2298                 mp->m_sb.sb_icount = cnt.icsb_icount;
2299         if (!xfs_icsb_counter_disabled(mp, XFS_SBS_IFREE))
2300                 mp->m_sb.sb_ifree = cnt.icsb_ifree;
2301         if (!xfs_icsb_counter_disabled(mp, XFS_SBS_FDBLOCKS))
2302                 mp->m_sb.sb_fdblocks = cnt.icsb_fdblocks;
2303 }
2304
2305 /*
2306  * Accurate update of per-cpu counters to incore superblock
2307  */
2308 void
2309 xfs_icsb_sync_counters(
2310         xfs_mount_t     *mp,
2311         int             flags)
2312 {
2313         spin_lock(&mp->m_sb_lock);
2314         xfs_icsb_sync_counters_locked(mp, flags);
2315         spin_unlock(&mp->m_sb_lock);
2316 }
2317
2318 /*
2319  * Balance and enable/disable counters as necessary.
2320  *
2321  * Thresholds for re-enabling counters are somewhat magic.  inode counts are
2322  * chosen to be the same number as single on disk allocation chunk per CPU, and
2323  * free blocks is something far enough zero that we aren't going thrash when we
2324  * get near ENOSPC. We also need to supply a minimum we require per cpu to
2325  * prevent looping endlessly when xfs_alloc_space asks for more than will
2326  * be distributed to a single CPU but each CPU has enough blocks to be
2327  * reenabled.
2328  *
2329  * Note that we can be called when counters are already disabled.
2330  * xfs_icsb_disable_counter() optimises the counter locking in this case to
2331  * prevent locking every per-cpu counter needlessly.
2332  */
2333
2334 #define XFS_ICSB_INO_CNTR_REENABLE      (uint64_t)64
2335 #define XFS_ICSB_FDBLK_CNTR_REENABLE(mp) \
2336                 (uint64_t)(512 + XFS_ALLOC_SET_ASIDE(mp))
2337 STATIC void
2338 xfs_icsb_balance_counter_locked(
2339         xfs_mount_t     *mp,
2340         xfs_sb_field_t  field,
2341         int             min_per_cpu)
2342 {
2343         uint64_t        count, resid;
2344         int             weight = num_online_cpus();
2345         uint64_t        min = (uint64_t)min_per_cpu;
2346
2347         /* disable counter and sync counter */
2348         xfs_icsb_disable_counter(mp, field);
2349
2350         /* update counters  - first CPU gets residual*/
2351         switch (field) {
2352         case XFS_SBS_ICOUNT:
2353                 count = mp->m_sb.sb_icount;
2354                 resid = do_div(count, weight);
2355                 if (count < max(min, XFS_ICSB_INO_CNTR_REENABLE))
2356                         return;
2357                 break;
2358         case XFS_SBS_IFREE:
2359                 count = mp->m_sb.sb_ifree;
2360                 resid = do_div(count, weight);
2361                 if (count < max(min, XFS_ICSB_INO_CNTR_REENABLE))
2362                         return;
2363                 break;
2364         case XFS_SBS_FDBLOCKS:
2365                 count = mp->m_sb.sb_fdblocks;
2366                 resid = do_div(count, weight);
2367                 if (count < max(min, XFS_ICSB_FDBLK_CNTR_REENABLE(mp)))
2368                         return;
2369                 break;
2370         default:
2371                 BUG();
2372                 count = resid = 0;      /* quiet, gcc */
2373                 break;
2374         }
2375
2376         xfs_icsb_enable_counter(mp, field, count, resid);
2377 }
2378
2379 STATIC void
2380 xfs_icsb_balance_counter(
2381         xfs_mount_t     *mp,
2382         xfs_sb_field_t  fields,
2383         int             min_per_cpu)
2384 {
2385         spin_lock(&mp->m_sb_lock);
2386         xfs_icsb_balance_counter_locked(mp, fields, min_per_cpu);
2387         spin_unlock(&mp->m_sb_lock);
2388 }
2389
2390 STATIC int
2391 xfs_icsb_modify_counters(
2392         xfs_mount_t     *mp,
2393         xfs_sb_field_t  field,
2394         int64_t         delta,
2395         int             rsvd)
2396 {
2397         xfs_icsb_cnts_t *icsbp;
2398         long long       lcounter;       /* long counter for 64 bit fields */
2399         int             cpu, ret = 0;
2400
2401         might_sleep();
2402 again:
2403         cpu = get_cpu();
2404         icsbp = (xfs_icsb_cnts_t *)per_cpu_ptr(mp->m_sb_cnts, cpu);
2405
2406         /*
2407          * if the counter is disabled, go to slow path
2408          */
2409         if (unlikely(xfs_icsb_counter_disabled(mp, field)))
2410                 goto slow_path;
2411         xfs_icsb_lock_cntr(icsbp);
2412         if (unlikely(xfs_icsb_counter_disabled(mp, field))) {
2413                 xfs_icsb_unlock_cntr(icsbp);
2414                 goto slow_path;
2415         }
2416
2417         switch (field) {
2418         case XFS_SBS_ICOUNT:
2419                 lcounter = icsbp->icsb_icount;
2420                 lcounter += delta;
2421                 if (unlikely(lcounter < 0))
2422                         goto balance_counter;
2423                 icsbp->icsb_icount = lcounter;
2424                 break;
2425
2426         case XFS_SBS_IFREE:
2427                 lcounter = icsbp->icsb_ifree;
2428                 lcounter += delta;
2429                 if (unlikely(lcounter < 0))
2430                         goto balance_counter;
2431                 icsbp->icsb_ifree = lcounter;
2432                 break;
2433
2434         case XFS_SBS_FDBLOCKS:
2435                 BUG_ON((mp->m_resblks - mp->m_resblks_avail) != 0);
2436
2437                 lcounter = icsbp->icsb_fdblocks - XFS_ALLOC_SET_ASIDE(mp);
2438                 lcounter += delta;
2439                 if (unlikely(lcounter < 0))
2440                         goto balance_counter;
2441                 icsbp->icsb_fdblocks = lcounter + XFS_ALLOC_SET_ASIDE(mp);
2442                 break;
2443         default:
2444                 BUG();
2445                 break;
2446         }
2447         xfs_icsb_unlock_cntr(icsbp);
2448         put_cpu();
2449         return 0;
2450
2451 slow_path:
2452         put_cpu();
2453
2454         /*
2455          * serialise with a mutex so we don't burn lots of cpu on
2456          * the superblock lock. We still need to hold the superblock
2457          * lock, however, when we modify the global structures.
2458          */
2459         xfs_icsb_lock(mp);
2460
2461         /*
2462          * Now running atomically.
2463          *
2464          * If the counter is enabled, someone has beaten us to rebalancing.
2465          * Drop the lock and try again in the fast path....
2466          */
2467         if (!(xfs_icsb_counter_disabled(mp, field))) {
2468                 xfs_icsb_unlock(mp);
2469                 goto again;
2470         }
2471
2472         /*
2473          * The counter is currently disabled. Because we are
2474          * running atomically here, we know a rebalance cannot
2475          * be in progress. Hence we can go straight to operating
2476          * on the global superblock. We do not call xfs_mod_incore_sb()
2477          * here even though we need to get the m_sb_lock. Doing so
2478          * will cause us to re-enter this function and deadlock.
2479          * Hence we get the m_sb_lock ourselves and then call
2480          * xfs_mod_incore_sb_unlocked() as the unlocked path operates
2481          * directly on the global counters.
2482          */
2483         spin_lock(&mp->m_sb_lock);
2484         ret = xfs_mod_incore_sb_unlocked(mp, field, delta, rsvd);
2485         spin_unlock(&mp->m_sb_lock);
2486
2487         /*
2488          * Now that we've modified the global superblock, we
2489          * may be able to re-enable the distributed counters
2490          * (e.g. lots of space just got freed). After that
2491          * we are done.
2492          */
2493         if (ret != ENOSPC)
2494                 xfs_icsb_balance_counter(mp, field, 0);
2495         xfs_icsb_unlock(mp);
2496         return ret;
2497
2498 balance_counter:
2499         xfs_icsb_unlock_cntr(icsbp);
2500         put_cpu();
2501
2502         /*
2503          * We may have multiple threads here if multiple per-cpu
2504          * counters run dry at the same time. This will mean we can
2505          * do more balances than strictly necessary but it is not
2506          * the common slowpath case.
2507          */
2508         xfs_icsb_lock(mp);
2509
2510         /*
2511          * running atomically.
2512          *
2513          * This will leave the counter in the correct state for future
2514          * accesses. After the rebalance, we simply try again and our retry
2515          * will either succeed through the fast path or slow path without
2516          * another balance operation being required.
2517          */
2518         xfs_icsb_balance_counter(mp, field, delta);
2519         xfs_icsb_unlock(mp);
2520         goto again;
2521 }
2522
2523 #endif