Merge branch 'master'
[linux-2.6] / fs / xfs / xfs_mount.c
1 /*
2  * Copyright (c) 2000-2005 Silicon Graphics, Inc.
3  * All Rights Reserved.
4  *
5  * This program is free software; you can redistribute it and/or
6  * modify it under the terms of the GNU General Public License as
7  * published by the Free Software Foundation.
8  *
9  * This program is distributed in the hope that it would be useful,
10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
12  * GNU General Public License for more details.
13  *
14  * You should have received a copy of the GNU General Public License
15  * along with this program; if not, write the Free Software Foundation,
16  * Inc.,  51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
17  */
18 #include "xfs.h"
19 #include "xfs_fs.h"
20 #include "xfs_types.h"
21 #include "xfs_bit.h"
22 #include "xfs_log.h"
23 #include "xfs_inum.h"
24 #include "xfs_trans.h"
25 #include "xfs_sb.h"
26 #include "xfs_ag.h"
27 #include "xfs_dir.h"
28 #include "xfs_dir2.h"
29 #include "xfs_dmapi.h"
30 #include "xfs_mount.h"
31 #include "xfs_bmap_btree.h"
32 #include "xfs_alloc_btree.h"
33 #include "xfs_ialloc_btree.h"
34 #include "xfs_dir_sf.h"
35 #include "xfs_dir2_sf.h"
36 #include "xfs_attr_sf.h"
37 #include "xfs_dinode.h"
38 #include "xfs_inode.h"
39 #include "xfs_btree.h"
40 #include "xfs_ialloc.h"
41 #include "xfs_alloc.h"
42 #include "xfs_rtalloc.h"
43 #include "xfs_bmap.h"
44 #include "xfs_error.h"
45 #include "xfs_rw.h"
46 #include "xfs_quota.h"
47 #include "xfs_fsops.h"
48
49 STATIC void     xfs_mount_log_sbunit(xfs_mount_t *, __int64_t);
50 STATIC int      xfs_uuid_mount(xfs_mount_t *);
51 STATIC void     xfs_uuid_unmount(xfs_mount_t *mp);
52 STATIC void     xfs_unmountfs_wait(xfs_mount_t *);
53
54
55 #ifdef HAVE_PERCPU_SB
56 STATIC void     xfs_icsb_destroy_counters(xfs_mount_t *);
57 STATIC void     xfs_icsb_balance_counter(xfs_mount_t *, xfs_sb_field_t, int);
58 STATIC void     xfs_icsb_sync_counters(xfs_mount_t *);
59 STATIC int      xfs_icsb_modify_counters(xfs_mount_t *, xfs_sb_field_t,
60                                                 int, int);
61 STATIC int      xfs_icsb_modify_counters_locked(xfs_mount_t *, xfs_sb_field_t,
62                                                 int, int);
63 STATIC int      xfs_icsb_disable_counter(xfs_mount_t *, xfs_sb_field_t);
64
65 #else
66
67 #define xfs_icsb_destroy_counters(mp)                   do { } while (0)
68 #define xfs_icsb_balance_counter(mp, a, b)              do { } while (0)
69 #define xfs_icsb_sync_counters(mp)                      do { } while (0)
70 #define xfs_icsb_modify_counters(mp, a, b, c)           do { } while (0)
71 #define xfs_icsb_modify_counters_locked(mp, a, b, c)    do { } while (0)
72
73 #endif
74
75 static const struct {
76         short offset;
77         short type;     /* 0 = integer
78                          * 1 = binary / string (no translation)
79                          */
80 } xfs_sb_info[] = {
81     { offsetof(xfs_sb_t, sb_magicnum),   0 },
82     { offsetof(xfs_sb_t, sb_blocksize),  0 },
83     { offsetof(xfs_sb_t, sb_dblocks),    0 },
84     { offsetof(xfs_sb_t, sb_rblocks),    0 },
85     { offsetof(xfs_sb_t, sb_rextents),   0 },
86     { offsetof(xfs_sb_t, sb_uuid),       1 },
87     { offsetof(xfs_sb_t, sb_logstart),   0 },
88     { offsetof(xfs_sb_t, sb_rootino),    0 },
89     { offsetof(xfs_sb_t, sb_rbmino),     0 },
90     { offsetof(xfs_sb_t, sb_rsumino),    0 },
91     { offsetof(xfs_sb_t, sb_rextsize),   0 },
92     { offsetof(xfs_sb_t, sb_agblocks),   0 },
93     { offsetof(xfs_sb_t, sb_agcount),    0 },
94     { offsetof(xfs_sb_t, sb_rbmblocks),  0 },
95     { offsetof(xfs_sb_t, sb_logblocks),  0 },
96     { offsetof(xfs_sb_t, sb_versionnum), 0 },
97     { offsetof(xfs_sb_t, sb_sectsize),   0 },
98     { offsetof(xfs_sb_t, sb_inodesize),  0 },
99     { offsetof(xfs_sb_t, sb_inopblock),  0 },
100     { offsetof(xfs_sb_t, sb_fname[0]),   1 },
101     { offsetof(xfs_sb_t, sb_blocklog),   0 },
102     { offsetof(xfs_sb_t, sb_sectlog),    0 },
103     { offsetof(xfs_sb_t, sb_inodelog),   0 },
104     { offsetof(xfs_sb_t, sb_inopblog),   0 },
105     { offsetof(xfs_sb_t, sb_agblklog),   0 },
106     { offsetof(xfs_sb_t, sb_rextslog),   0 },
107     { offsetof(xfs_sb_t, sb_inprogress), 0 },
108     { offsetof(xfs_sb_t, sb_imax_pct),   0 },
109     { offsetof(xfs_sb_t, sb_icount),     0 },
110     { offsetof(xfs_sb_t, sb_ifree),      0 },
111     { offsetof(xfs_sb_t, sb_fdblocks),   0 },
112     { offsetof(xfs_sb_t, sb_frextents),  0 },
113     { offsetof(xfs_sb_t, sb_uquotino),   0 },
114     { offsetof(xfs_sb_t, sb_gquotino),   0 },
115     { offsetof(xfs_sb_t, sb_qflags),     0 },
116     { offsetof(xfs_sb_t, sb_flags),      0 },
117     { offsetof(xfs_sb_t, sb_shared_vn),  0 },
118     { offsetof(xfs_sb_t, sb_inoalignmt), 0 },
119     { offsetof(xfs_sb_t, sb_unit),       0 },
120     { offsetof(xfs_sb_t, sb_width),      0 },
121     { offsetof(xfs_sb_t, sb_dirblklog),  0 },
122     { offsetof(xfs_sb_t, sb_logsectlog), 0 },
123     { offsetof(xfs_sb_t, sb_logsectsize),0 },
124     { offsetof(xfs_sb_t, sb_logsunit),   0 },
125     { offsetof(xfs_sb_t, sb_features2),  0 },
126     { sizeof(xfs_sb_t),                  0 }
127 };
128
129 /*
130  * Return a pointer to an initialized xfs_mount structure.
131  */
132 xfs_mount_t *
133 xfs_mount_init(void)
134 {
135         xfs_mount_t *mp;
136
137         mp = kmem_zalloc(sizeof(xfs_mount_t), KM_SLEEP);
138
139         if (xfs_icsb_init_counters(mp)) {
140                 mp->m_flags |= XFS_MOUNT_NO_PERCPU_SB;
141         }
142
143         AIL_LOCKINIT(&mp->m_ail_lock, "xfs_ail");
144         spinlock_init(&mp->m_sb_lock, "xfs_sb");
145         mutex_init(&mp->m_ilock);
146         initnsema(&mp->m_growlock, 1, "xfs_grow");
147         /*
148          * Initialize the AIL.
149          */
150         xfs_trans_ail_init(mp);
151
152         atomic_set(&mp->m_active_trans, 0);
153
154         return mp;
155 }
156
157 /*
158  * Free up the resources associated with a mount structure.  Assume that
159  * the structure was initially zeroed, so we can tell which fields got
160  * initialized.
161  */
162 void
163 xfs_mount_free(
164         xfs_mount_t     *mp,
165         int             remove_bhv)
166 {
167         if (mp->m_ihash)
168                 xfs_ihash_free(mp);
169         if (mp->m_chash)
170                 xfs_chash_free(mp);
171
172         if (mp->m_perag) {
173                 int     agno;
174
175                 for (agno = 0; agno < mp->m_maxagi; agno++)
176                         if (mp->m_perag[agno].pagb_list)
177                                 kmem_free(mp->m_perag[agno].pagb_list,
178                                                 sizeof(xfs_perag_busy_t) *
179                                                         XFS_PAGB_NUM_SLOTS);
180                 kmem_free(mp->m_perag,
181                           sizeof(xfs_perag_t) * mp->m_sb.sb_agcount);
182         }
183
184         AIL_LOCK_DESTROY(&mp->m_ail_lock);
185         spinlock_destroy(&mp->m_sb_lock);
186         mutex_destroy(&mp->m_ilock);
187         freesema(&mp->m_growlock);
188         if (mp->m_quotainfo)
189                 XFS_QM_DONE(mp);
190
191         if (mp->m_fsname != NULL)
192                 kmem_free(mp->m_fsname, mp->m_fsname_len);
193         if (mp->m_rtname != NULL)
194                 kmem_free(mp->m_rtname, strlen(mp->m_rtname) + 1);
195         if (mp->m_logname != NULL)
196                 kmem_free(mp->m_logname, strlen(mp->m_logname) + 1);
197
198         if (remove_bhv) {
199                 struct vfs      *vfsp = XFS_MTOVFS(mp);
200
201                 bhv_remove_all_vfsops(vfsp, 0);
202                 VFS_REMOVEBHV(vfsp, &mp->m_bhv);
203         }
204
205         xfs_icsb_destroy_counters(mp);
206         kmem_free(mp, sizeof(xfs_mount_t));
207 }
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         ASSERT(PAGE_SHIFT >= sbp->sb_blocklog);
292         ASSERT(sbp->sb_blocklog >= BBSHIFT);
293
294 #if XFS_BIG_BLKNOS     /* Limited by ULONG_MAX of page cache index */
295         if (unlikely(
296             (sbp->sb_dblocks >> (PAGE_SHIFT - sbp->sb_blocklog)) > ULONG_MAX ||
297             (sbp->sb_rblocks >> (PAGE_SHIFT - sbp->sb_blocklog)) > ULONG_MAX)) {
298 #else                  /* Limited by UINT_MAX of sectors */
299         if (unlikely(
300             (sbp->sb_dblocks << (sbp->sb_blocklog - BBSHIFT)) > UINT_MAX ||
301             (sbp->sb_rblocks << (sbp->sb_blocklog - BBSHIFT)) > UINT_MAX)) {
302 #endif
303                 xfs_fs_mount_cmn_err(flags,
304                         "file system too large to be mounted on this system.");
305                 return XFS_ERROR(E2BIG);
306         }
307
308         if (unlikely(sbp->sb_inprogress)) {
309                 xfs_fs_mount_cmn_err(flags, "file system busy");
310                 return XFS_ERROR(EFSCORRUPTED);
311         }
312
313         /*
314          * Version 1 directory format has never worked on Linux.
315          */
316         if (unlikely(!XFS_SB_VERSION_HASDIRV2(sbp))) {
317                 xfs_fs_mount_cmn_err(flags,
318                         "file system using version 1 directory format");
319                 return XFS_ERROR(ENOSYS);
320         }
321
322         /*
323          * Until this is fixed only page-sized or smaller data blocks work.
324          */
325         if (unlikely(sbp->sb_blocksize > PAGE_SIZE)) {
326                 xfs_fs_mount_cmn_err(flags,
327                         "file system with blocksize %d bytes",
328                         sbp->sb_blocksize);
329                 xfs_fs_mount_cmn_err(flags,
330                         "only pagesize (%ld) or less will currently work.",
331                         PAGE_SIZE);
332                 return XFS_ERROR(ENOSYS);
333         }
334
335         return 0;
336 }
337
338 xfs_agnumber_t
339 xfs_initialize_perag(
340         struct vfs      *vfs,
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 ((vfs->vfs_flag & VFS_32BITINODES) && 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                 }
393         } else {
394                 /* Setup default behavior for smaller filesystems */
395                 for (index = 0; index < agcount; index++) {
396                         pag = &mp->m_perag[index];
397                         pag->pagi_inodeok = 1;
398                 }
399         }
400         return index;
401 }
402
403 /*
404  * xfs_xlatesb
405  *
406  *     data       - on disk version of sb
407  *     sb         - a superblock
408  *     dir        - conversion direction: <0 - convert sb to buf
409  *                                        >0 - convert buf to sb
410  *     fields     - which fields to copy (bitmask)
411  */
412 void
413 xfs_xlatesb(
414         void            *data,
415         xfs_sb_t        *sb,
416         int             dir,
417         __int64_t       fields)
418 {
419         xfs_caddr_t     buf_ptr;
420         xfs_caddr_t     mem_ptr;
421         xfs_sb_field_t  f;
422         int             first;
423         int             size;
424
425         ASSERT(dir);
426         ASSERT(fields);
427
428         if (!fields)
429                 return;
430
431         buf_ptr = (xfs_caddr_t)data;
432         mem_ptr = (xfs_caddr_t)sb;
433
434         while (fields) {
435                 f = (xfs_sb_field_t)xfs_lowbit64((__uint64_t)fields);
436                 first = xfs_sb_info[f].offset;
437                 size = xfs_sb_info[f + 1].offset - first;
438
439                 ASSERT(xfs_sb_info[f].type == 0 || xfs_sb_info[f].type == 1);
440
441                 if (size == 1 || xfs_sb_info[f].type == 1) {
442                         if (dir > 0) {
443                                 memcpy(mem_ptr + first, buf_ptr + first, size);
444                         } else {
445                                 memcpy(buf_ptr + first, mem_ptr + first, size);
446                         }
447                 } else {
448                         switch (size) {
449                         case 2:
450                                 INT_XLATE(*(__uint16_t*)(buf_ptr+first),
451                                           *(__uint16_t*)(mem_ptr+first),
452                                           dir, ARCH_CONVERT);
453                                 break;
454                         case 4:
455                                 INT_XLATE(*(__uint32_t*)(buf_ptr+first),
456                                           *(__uint32_t*)(mem_ptr+first),
457                                           dir, ARCH_CONVERT);
458                                 break;
459                         case 8:
460                                 INT_XLATE(*(__uint64_t*)(buf_ptr+first),
461                                           *(__uint64_t*)(mem_ptr+first), dir, ARCH_CONVERT);
462                                 break;
463                         default:
464                                 ASSERT(0);
465                         }
466                 }
467
468                 fields &= ~(1LL << f);
469         }
470 }
471
472 /*
473  * xfs_readsb
474  *
475  * Does the initial read of the superblock.
476  */
477 int
478 xfs_readsb(xfs_mount_t *mp, int flags)
479 {
480         unsigned int    sector_size;
481         unsigned int    extra_flags;
482         xfs_buf_t       *bp;
483         xfs_sb_t        *sbp;
484         int             error;
485
486         ASSERT(mp->m_sb_bp == NULL);
487         ASSERT(mp->m_ddev_targp != NULL);
488
489         /*
490          * Allocate a (locked) buffer to hold the superblock.
491          * This will be kept around at all times to optimize
492          * access to the superblock.
493          */
494         sector_size = xfs_getsize_buftarg(mp->m_ddev_targp);
495         extra_flags = XFS_BUF_LOCK | XFS_BUF_MANAGE | XFS_BUF_MAPPED;
496
497         bp = xfs_buf_read_flags(mp->m_ddev_targp, XFS_SB_DADDR,
498                                 BTOBB(sector_size), extra_flags);
499         if (!bp || XFS_BUF_ISERROR(bp)) {
500                 xfs_fs_mount_cmn_err(flags, "SB read failed");
501                 error = bp ? XFS_BUF_GETERROR(bp) : ENOMEM;
502                 goto fail;
503         }
504         ASSERT(XFS_BUF_ISBUSY(bp));
505         ASSERT(XFS_BUF_VALUSEMA(bp) <= 0);
506
507         /*
508          * Initialize the mount structure from the superblock.
509          * But first do some basic consistency checking.
510          */
511         sbp = XFS_BUF_TO_SBP(bp);
512         xfs_xlatesb(XFS_BUF_PTR(bp), &(mp->m_sb), 1, XFS_SB_ALL_BITS);
513
514         error = xfs_mount_validate_sb(mp, &(mp->m_sb), flags);
515         if (error) {
516                 xfs_fs_mount_cmn_err(flags, "SB validate failed");
517                 goto fail;
518         }
519
520         /*
521          * We must be able to do sector-sized and sector-aligned IO.
522          */
523         if (sector_size > mp->m_sb.sb_sectsize) {
524                 xfs_fs_mount_cmn_err(flags,
525                         "device supports only %u byte sectors (not %u)",
526                         sector_size, mp->m_sb.sb_sectsize);
527                 error = ENOSYS;
528                 goto fail;
529         }
530
531         /*
532          * If device sector size is smaller than the superblock size,
533          * re-read the superblock so the buffer is correctly sized.
534          */
535         if (sector_size < mp->m_sb.sb_sectsize) {
536                 XFS_BUF_UNMANAGE(bp);
537                 xfs_buf_relse(bp);
538                 sector_size = mp->m_sb.sb_sectsize;
539                 bp = xfs_buf_read_flags(mp->m_ddev_targp, XFS_SB_DADDR,
540                                         BTOBB(sector_size), extra_flags);
541                 if (!bp || XFS_BUF_ISERROR(bp)) {
542                         xfs_fs_mount_cmn_err(flags, "SB re-read failed");
543                         error = bp ? XFS_BUF_GETERROR(bp) : ENOMEM;
544                         goto fail;
545                 }
546                 ASSERT(XFS_BUF_ISBUSY(bp));
547                 ASSERT(XFS_BUF_VALUSEMA(bp) <= 0);
548         }
549
550         xfs_icsb_balance_counter(mp, XFS_SBS_ICOUNT, 0);
551         xfs_icsb_balance_counter(mp, XFS_SBS_IFREE, 0);
552         xfs_icsb_balance_counter(mp, XFS_SBS_FDBLOCKS, 0);
553
554         mp->m_sb_bp = bp;
555         xfs_buf_relse(bp);
556         ASSERT(XFS_BUF_VALUSEMA(bp) > 0);
557         return 0;
558
559  fail:
560         if (bp) {
561                 XFS_BUF_UNMANAGE(bp);
562                 xfs_buf_relse(bp);
563         }
564         return error;
565 }
566
567
568 /*
569  * xfs_mount_common
570  *
571  * Mount initialization code establishing various mount
572  * fields from the superblock associated with the given
573  * mount structure
574  */
575 STATIC void
576 xfs_mount_common(xfs_mount_t *mp, xfs_sb_t *sbp)
577 {
578         int     i;
579
580         mp->m_agfrotor = mp->m_agirotor = 0;
581         spinlock_init(&mp->m_agirotor_lock, "m_agirotor_lock");
582         mp->m_maxagi = mp->m_sb.sb_agcount;
583         mp->m_blkbit_log = sbp->sb_blocklog + XFS_NBBYLOG;
584         mp->m_blkbb_log = sbp->sb_blocklog - BBSHIFT;
585         mp->m_sectbb_log = sbp->sb_sectlog - BBSHIFT;
586         mp->m_agno_log = xfs_highbit32(sbp->sb_agcount - 1) + 1;
587         mp->m_agino_log = sbp->sb_inopblog + sbp->sb_agblklog;
588         mp->m_litino = sbp->sb_inodesize -
589                 ((uint)sizeof(xfs_dinode_core_t) + (uint)sizeof(xfs_agino_t));
590         mp->m_blockmask = sbp->sb_blocksize - 1;
591         mp->m_blockwsize = sbp->sb_blocksize >> XFS_WORDLOG;
592         mp->m_blockwmask = mp->m_blockwsize - 1;
593         INIT_LIST_HEAD(&mp->m_del_inodes);
594
595         /*
596          * Setup for attributes, in case they get created.
597          * This value is for inodes getting attributes for the first time,
598          * the per-inode value is for old attribute values.
599          */
600         ASSERT(sbp->sb_inodesize >= 256 && sbp->sb_inodesize <= 2048);
601         switch (sbp->sb_inodesize) {
602         case 256:
603                 mp->m_attroffset = XFS_LITINO(mp) -
604                                    XFS_BMDR_SPACE_CALC(MINABTPTRS);
605                 break;
606         case 512:
607         case 1024:
608         case 2048:
609                 mp->m_attroffset = XFS_BMDR_SPACE_CALC(6 * MINABTPTRS);
610                 break;
611         default:
612                 ASSERT(0);
613         }
614         ASSERT(mp->m_attroffset < XFS_LITINO(mp));
615
616         for (i = 0; i < 2; i++) {
617                 mp->m_alloc_mxr[i] = XFS_BTREE_BLOCK_MAXRECS(sbp->sb_blocksize,
618                         xfs_alloc, i == 0);
619                 mp->m_alloc_mnr[i] = XFS_BTREE_BLOCK_MINRECS(sbp->sb_blocksize,
620                         xfs_alloc, i == 0);
621         }
622         for (i = 0; i < 2; i++) {
623                 mp->m_bmap_dmxr[i] = XFS_BTREE_BLOCK_MAXRECS(sbp->sb_blocksize,
624                         xfs_bmbt, i == 0);
625                 mp->m_bmap_dmnr[i] = XFS_BTREE_BLOCK_MINRECS(sbp->sb_blocksize,
626                         xfs_bmbt, i == 0);
627         }
628         for (i = 0; i < 2; i++) {
629                 mp->m_inobt_mxr[i] = XFS_BTREE_BLOCK_MAXRECS(sbp->sb_blocksize,
630                         xfs_inobt, i == 0);
631                 mp->m_inobt_mnr[i] = XFS_BTREE_BLOCK_MINRECS(sbp->sb_blocksize,
632                         xfs_inobt, i == 0);
633         }
634
635         mp->m_bsize = XFS_FSB_TO_BB(mp, 1);
636         mp->m_ialloc_inos = (int)MAX((__uint16_t)XFS_INODES_PER_CHUNK,
637                                         sbp->sb_inopblock);
638         mp->m_ialloc_blks = mp->m_ialloc_inos >> sbp->sb_inopblog;
639 }
640 /*
641  * xfs_mountfs
642  *
643  * This function does the following on an initial mount of a file system:
644  *      - reads the superblock from disk and init the mount struct
645  *      - if we're a 32-bit kernel, do a size check on the superblock
646  *              so we don't mount terabyte filesystems
647  *      - init mount struct realtime fields
648  *      - allocate inode hash table for fs
649  *      - init directory manager
650  *      - perform recovery and init the log manager
651  */
652 int
653 xfs_mountfs(
654         vfs_t           *vfsp,
655         xfs_mount_t     *mp,
656         int             mfsi_flags)
657 {
658         xfs_buf_t       *bp;
659         xfs_sb_t        *sbp = &(mp->m_sb);
660         xfs_inode_t     *rip;
661         vnode_t         *rvp = NULL;
662         int             readio_log, writeio_log;
663         xfs_daddr_t     d;
664         __uint64_t      ret64;
665         __int64_t       update_flags;
666         uint            quotamount, quotaflags;
667         int             agno;
668         int             uuid_mounted = 0;
669         int             error = 0;
670
671         if (mp->m_sb_bp == NULL) {
672                 if ((error = xfs_readsb(mp, mfsi_flags))) {
673                         return error;
674                 }
675         }
676         xfs_mount_common(mp, sbp);
677
678         /*
679          * Check if sb_agblocks is aligned at stripe boundary
680          * If sb_agblocks is NOT aligned turn off m_dalign since
681          * allocator alignment is within an ag, therefore ag has
682          * to be aligned at stripe boundary.
683          */
684         update_flags = 0LL;
685         if (mp->m_dalign && !(mfsi_flags & XFS_MFSI_SECOND)) {
686                 /*
687                  * If stripe unit and stripe width are not multiples
688                  * of the fs blocksize turn off alignment.
689                  */
690                 if ((BBTOB(mp->m_dalign) & mp->m_blockmask) ||
691                     (BBTOB(mp->m_swidth) & mp->m_blockmask)) {
692                         if (mp->m_flags & XFS_MOUNT_RETERR) {
693                                 cmn_err(CE_WARN,
694                                         "XFS: alignment check 1 failed");
695                                 error = XFS_ERROR(EINVAL);
696                                 goto error1;
697                         }
698                         mp->m_dalign = mp->m_swidth = 0;
699                 } else {
700                         /*
701                          * Convert the stripe unit and width to FSBs.
702                          */
703                         mp->m_dalign = XFS_BB_TO_FSBT(mp, mp->m_dalign);
704                         if (mp->m_dalign && (sbp->sb_agblocks % mp->m_dalign)) {
705                                 if (mp->m_flags & XFS_MOUNT_RETERR) {
706                                         error = XFS_ERROR(EINVAL);
707                                         goto error1;
708                                 }
709                                 xfs_fs_cmn_err(CE_WARN, mp,
710 "stripe alignment turned off: sunit(%d)/swidth(%d) incompatible with agsize(%d)",
711                                         mp->m_dalign, mp->m_swidth,
712                                         sbp->sb_agblocks);
713
714                                 mp->m_dalign = 0;
715                                 mp->m_swidth = 0;
716                         } else if (mp->m_dalign) {
717                                 mp->m_swidth = XFS_BB_TO_FSBT(mp, mp->m_swidth);
718                         } else {
719                                 if (mp->m_flags & XFS_MOUNT_RETERR) {
720                                         xfs_fs_cmn_err(CE_WARN, mp,
721 "stripe alignment turned off: sunit(%d) less than bsize(%d)",
722                                                 mp->m_dalign,
723                                                 mp->m_blockmask +1);
724                                         error = XFS_ERROR(EINVAL);
725                                         goto error1;
726                                 }
727                                 mp->m_swidth = 0;
728                         }
729                 }
730
731                 /*
732                  * Update superblock with new values
733                  * and log changes
734                  */
735                 if (XFS_SB_VERSION_HASDALIGN(sbp)) {
736                         if (sbp->sb_unit != mp->m_dalign) {
737                                 sbp->sb_unit = mp->m_dalign;
738                                 update_flags |= XFS_SB_UNIT;
739                         }
740                         if (sbp->sb_width != mp->m_swidth) {
741                                 sbp->sb_width = mp->m_swidth;
742                                 update_flags |= XFS_SB_WIDTH;
743                         }
744                 }
745         } else if ((mp->m_flags & XFS_MOUNT_NOALIGN) != XFS_MOUNT_NOALIGN &&
746                     XFS_SB_VERSION_HASDALIGN(&mp->m_sb)) {
747                         mp->m_dalign = sbp->sb_unit;
748                         mp->m_swidth = sbp->sb_width;
749         }
750
751         xfs_alloc_compute_maxlevels(mp);
752         xfs_bmap_compute_maxlevels(mp, XFS_DATA_FORK);
753         xfs_bmap_compute_maxlevels(mp, XFS_ATTR_FORK);
754         xfs_ialloc_compute_maxlevels(mp);
755
756         if (sbp->sb_imax_pct) {
757                 __uint64_t      icount;
758
759                 /* Make sure the maximum inode count is a multiple of the
760                  * units we allocate inodes in.
761                  */
762
763                 icount = sbp->sb_dblocks * sbp->sb_imax_pct;
764                 do_div(icount, 100);
765                 do_div(icount, mp->m_ialloc_blks);
766                 mp->m_maxicount = (icount * mp->m_ialloc_blks)  <<
767                                    sbp->sb_inopblog;
768         } else
769                 mp->m_maxicount = 0;
770
771         mp->m_maxioffset = xfs_max_file_offset(sbp->sb_blocklog);
772
773         /*
774          * XFS uses the uuid from the superblock as the unique
775          * identifier for fsid.  We can not use the uuid from the volume
776          * since a single partition filesystem is identical to a single
777          * partition volume/filesystem.
778          */
779         if ((mfsi_flags & XFS_MFSI_SECOND) == 0 &&
780             (mp->m_flags & XFS_MOUNT_NOUUID) == 0) {
781                 if (xfs_uuid_mount(mp)) {
782                         error = XFS_ERROR(EINVAL);
783                         goto error1;
784                 }
785                 uuid_mounted=1;
786                 ret64 = uuid_hash64(&sbp->sb_uuid);
787                 memcpy(&vfsp->vfs_fsid, &ret64, sizeof(ret64));
788         }
789
790         /*
791          * Set the default minimum read and write sizes unless
792          * already specified in a mount option.
793          * We use smaller I/O sizes when the file system
794          * is being used for NFS service (wsync mount option).
795          */
796         if (!(mp->m_flags & XFS_MOUNT_DFLT_IOSIZE)) {
797                 if (mp->m_flags & XFS_MOUNT_WSYNC) {
798                         readio_log = XFS_WSYNC_READIO_LOG;
799                         writeio_log = XFS_WSYNC_WRITEIO_LOG;
800                 } else {
801                         readio_log = XFS_READIO_LOG_LARGE;
802                         writeio_log = XFS_WRITEIO_LOG_LARGE;
803                 }
804         } else {
805                 readio_log = mp->m_readio_log;
806                 writeio_log = mp->m_writeio_log;
807         }
808
809         /*
810          * Set the number of readahead buffers to use based on
811          * physical memory size.
812          */
813         if (xfs_physmem <= 4096)                /* <= 16MB */
814                 mp->m_nreadaheads = XFS_RW_NREADAHEAD_16MB;
815         else if (xfs_physmem <= 8192)   /* <= 32MB */
816                 mp->m_nreadaheads = XFS_RW_NREADAHEAD_32MB;
817         else
818                 mp->m_nreadaheads = XFS_RW_NREADAHEAD_K32;
819         if (sbp->sb_blocklog > readio_log) {
820                 mp->m_readio_log = sbp->sb_blocklog;
821         } else {
822                 mp->m_readio_log = readio_log;
823         }
824         mp->m_readio_blocks = 1 << (mp->m_readio_log - sbp->sb_blocklog);
825         if (sbp->sb_blocklog > writeio_log) {
826                 mp->m_writeio_log = sbp->sb_blocklog;
827         } else {
828                 mp->m_writeio_log = writeio_log;
829         }
830         mp->m_writeio_blocks = 1 << (mp->m_writeio_log - sbp->sb_blocklog);
831
832         /*
833          * Set the inode cluster size based on the physical memory
834          * size.  This may still be overridden by the file system
835          * block size if it is larger than the chosen cluster size.
836          */
837         if (xfs_physmem <= btoc(32 * 1024 * 1024)) { /* <= 32 MB */
838                 mp->m_inode_cluster_size = XFS_INODE_SMALL_CLUSTER_SIZE;
839         } else {
840                 mp->m_inode_cluster_size = XFS_INODE_BIG_CLUSTER_SIZE;
841         }
842         /*
843          * Set whether we're using inode alignment.
844          */
845         if (XFS_SB_VERSION_HASALIGN(&mp->m_sb) &&
846             mp->m_sb.sb_inoalignmt >=
847             XFS_B_TO_FSBT(mp, mp->m_inode_cluster_size))
848                 mp->m_inoalign_mask = mp->m_sb.sb_inoalignmt - 1;
849         else
850                 mp->m_inoalign_mask = 0;
851         /*
852          * If we are using stripe alignment, check whether
853          * the stripe unit is a multiple of the inode alignment
854          */
855         if (mp->m_dalign && mp->m_inoalign_mask &&
856             !(mp->m_dalign & mp->m_inoalign_mask))
857                 mp->m_sinoalign = mp->m_dalign;
858         else
859                 mp->m_sinoalign = 0;
860         /*
861          * Check that the data (and log if separate) are an ok size.
862          */
863         d = (xfs_daddr_t)XFS_FSB_TO_BB(mp, mp->m_sb.sb_dblocks);
864         if (XFS_BB_TO_FSB(mp, d) != mp->m_sb.sb_dblocks) {
865                 cmn_err(CE_WARN, "XFS: size check 1 failed");
866                 error = XFS_ERROR(E2BIG);
867                 goto error1;
868         }
869         error = xfs_read_buf(mp, mp->m_ddev_targp,
870                              d - XFS_FSS_TO_BB(mp, 1),
871                              XFS_FSS_TO_BB(mp, 1), 0, &bp);
872         if (!error) {
873                 xfs_buf_relse(bp);
874         } else {
875                 cmn_err(CE_WARN, "XFS: size check 2 failed");
876                 if (error == ENOSPC) {
877                         error = XFS_ERROR(E2BIG);
878                 }
879                 goto error1;
880         }
881
882         if (((mfsi_flags & XFS_MFSI_CLIENT) == 0) &&
883             mp->m_logdev_targp != mp->m_ddev_targp) {
884                 d = (xfs_daddr_t)XFS_FSB_TO_BB(mp, mp->m_sb.sb_logblocks);
885                 if (XFS_BB_TO_FSB(mp, d) != mp->m_sb.sb_logblocks) {
886                         cmn_err(CE_WARN, "XFS: size check 3 failed");
887                         error = XFS_ERROR(E2BIG);
888                         goto error1;
889                 }
890                 error = xfs_read_buf(mp, mp->m_logdev_targp,
891                                      d - XFS_FSB_TO_BB(mp, 1),
892                                      XFS_FSB_TO_BB(mp, 1), 0, &bp);
893                 if (!error) {
894                         xfs_buf_relse(bp);
895                 } else {
896                         cmn_err(CE_WARN, "XFS: size check 3 failed");
897                         if (error == ENOSPC) {
898                                 error = XFS_ERROR(E2BIG);
899                         }
900                         goto error1;
901                 }
902         }
903
904         /*
905          * Initialize realtime fields in the mount structure
906          */
907         if ((error = xfs_rtmount_init(mp))) {
908                 cmn_err(CE_WARN, "XFS: RT mount failed");
909                 goto error1;
910         }
911
912         /*
913          * For client case we are done now
914          */
915         if (mfsi_flags & XFS_MFSI_CLIENT) {
916                 return 0;
917         }
918
919         /*
920          *  Copies the low order bits of the timestamp and the randomly
921          *  set "sequence" number out of a UUID.
922          */
923         uuid_getnodeuniq(&sbp->sb_uuid, mp->m_fixedfsid);
924
925         /*
926          *  The vfs structure needs to have a file system independent
927          *  way of checking for the invariant file system ID.  Since it
928          *  can't look at mount structures it has a pointer to the data
929          *  in the mount structure.
930          *
931          *  File systems that don't support user level file handles (i.e.
932          *  all of them except for XFS) will leave vfs_altfsid as NULL.
933          */
934         vfsp->vfs_altfsid = (xfs_fsid_t *)mp->m_fixedfsid;
935         mp->m_dmevmask = 0;     /* not persistent; set after each mount */
936
937         /*
938          * Select the right directory manager.
939          */
940         mp->m_dirops =
941                 XFS_SB_VERSION_HASDIRV2(&mp->m_sb) ?
942                         xfsv2_dirops :
943                         xfsv1_dirops;
944
945         /*
946          * Initialize directory manager's entries.
947          */
948         XFS_DIR_MOUNT(mp);
949
950         /*
951          * Initialize the attribute manager's entries.
952          */
953         mp->m_attr_magicpct = (mp->m_sb.sb_blocksize * 37) / 100;
954
955         /*
956          * Initialize the precomputed transaction reservations values.
957          */
958         xfs_trans_init(mp);
959
960         /*
961          * Allocate and initialize the inode hash table for this
962          * file system.
963          */
964         xfs_ihash_init(mp);
965         xfs_chash_init(mp);
966
967         /*
968          * Allocate and initialize the per-ag data.
969          */
970         init_rwsem(&mp->m_peraglock);
971         mp->m_perag =
972                 kmem_zalloc(sbp->sb_agcount * sizeof(xfs_perag_t), KM_SLEEP);
973
974         mp->m_maxagi = xfs_initialize_perag(vfsp, mp, sbp->sb_agcount);
975
976         /*
977          * log's mount-time initialization. Perform 1st part recovery if needed
978          */
979         if (likely(sbp->sb_logblocks > 0)) {    /* check for volume case */
980                 error = xfs_log_mount(mp, mp->m_logdev_targp,
981                                       XFS_FSB_TO_DADDR(mp, sbp->sb_logstart),
982                                       XFS_FSB_TO_BB(mp, sbp->sb_logblocks));
983                 if (error) {
984                         cmn_err(CE_WARN, "XFS: log mount failed");
985                         goto error2;
986                 }
987         } else {        /* No log has been defined */
988                 cmn_err(CE_WARN, "XFS: no log defined");
989                 XFS_ERROR_REPORT("xfs_mountfs_int(1)", XFS_ERRLEVEL_LOW, mp);
990                 error = XFS_ERROR(EFSCORRUPTED);
991                 goto error2;
992         }
993
994         /*
995          * Get and sanity-check the root inode.
996          * Save the pointer to it in the mount structure.
997          */
998         error = xfs_iget(mp, NULL, sbp->sb_rootino, 0, XFS_ILOCK_EXCL, &rip, 0);
999         if (error) {
1000                 cmn_err(CE_WARN, "XFS: failed to read root inode");
1001                 goto error3;
1002         }
1003
1004         ASSERT(rip != NULL);
1005         rvp = XFS_ITOV(rip);
1006
1007         if (unlikely((rip->i_d.di_mode & S_IFMT) != S_IFDIR)) {
1008                 cmn_err(CE_WARN, "XFS: corrupted root inode");
1009                 prdev("Root inode %llu is not a directory",
1010                       mp->m_ddev_targp, (unsigned long long)rip->i_ino);
1011                 xfs_iunlock(rip, XFS_ILOCK_EXCL);
1012                 XFS_ERROR_REPORT("xfs_mountfs_int(2)", XFS_ERRLEVEL_LOW,
1013                                  mp);
1014                 error = XFS_ERROR(EFSCORRUPTED);
1015                 goto error4;
1016         }
1017         mp->m_rootip = rip;     /* save it */
1018
1019         xfs_iunlock(rip, XFS_ILOCK_EXCL);
1020
1021         /*
1022          * Initialize realtime inode pointers in the mount structure
1023          */
1024         if ((error = xfs_rtmount_inodes(mp))) {
1025                 /*
1026                  * Free up the root inode.
1027                  */
1028                 cmn_err(CE_WARN, "XFS: failed to read RT inodes");
1029                 goto error4;
1030         }
1031
1032         /*
1033          * If fs is not mounted readonly, then update the superblock
1034          * unit and width changes.
1035          */
1036         if (update_flags && !(vfsp->vfs_flag & VFS_RDONLY))
1037                 xfs_mount_log_sbunit(mp, update_flags);
1038
1039         /*
1040          * Initialise the XFS quota management subsystem for this mount
1041          */
1042         if ((error = XFS_QM_INIT(mp, &quotamount, &quotaflags)))
1043                 goto error4;
1044
1045         /*
1046          * Finish recovering the file system.  This part needed to be
1047          * delayed until after the root and real-time bitmap inodes
1048          * were consistently read in.
1049          */
1050         error = xfs_log_mount_finish(mp, mfsi_flags);
1051         if (error) {
1052                 cmn_err(CE_WARN, "XFS: log mount finish failed");
1053                 goto error4;
1054         }
1055
1056         /*
1057          * Complete the quota initialisation, post-log-replay component.
1058          */
1059         if ((error = XFS_QM_MOUNT(mp, quotamount, quotaflags, mfsi_flags)))
1060                 goto error4;
1061
1062         return 0;
1063
1064  error4:
1065         /*
1066          * Free up the root inode.
1067          */
1068         VN_RELE(rvp);
1069  error3:
1070         xfs_log_unmount_dealloc(mp);
1071  error2:
1072         xfs_ihash_free(mp);
1073         xfs_chash_free(mp);
1074         for (agno = 0; agno < sbp->sb_agcount; agno++)
1075                 if (mp->m_perag[agno].pagb_list)
1076                         kmem_free(mp->m_perag[agno].pagb_list,
1077                           sizeof(xfs_perag_busy_t) * XFS_PAGB_NUM_SLOTS);
1078         kmem_free(mp->m_perag, sbp->sb_agcount * sizeof(xfs_perag_t));
1079         mp->m_perag = NULL;
1080         /* FALLTHROUGH */
1081  error1:
1082         if (uuid_mounted)
1083                 xfs_uuid_unmount(mp);
1084         xfs_freesb(mp);
1085         return error;
1086 }
1087
1088 /*
1089  * xfs_unmountfs
1090  *
1091  * This flushes out the inodes,dquots and the superblock, unmounts the
1092  * log and makes sure that incore structures are freed.
1093  */
1094 int
1095 xfs_unmountfs(xfs_mount_t *mp, struct cred *cr)
1096 {
1097         struct vfs      *vfsp = XFS_MTOVFS(mp);
1098 #if defined(DEBUG) || defined(INDUCE_IO_ERROR)
1099         int64_t         fsid;
1100 #endif
1101
1102         xfs_iflush_all(mp);
1103
1104         XFS_QM_DQPURGEALL(mp, XFS_QMOPT_QUOTALL | XFS_QMOPT_UMOUNTING);
1105
1106         /*
1107          * Flush out the log synchronously so that we know for sure
1108          * that nothing is pinned.  This is important because bflush()
1109          * will skip pinned buffers.
1110          */
1111         xfs_log_force(mp, (xfs_lsn_t)0, XFS_LOG_FORCE | XFS_LOG_SYNC);
1112
1113         xfs_binval(mp->m_ddev_targp);
1114         if (mp->m_rtdev_targp) {
1115                 xfs_binval(mp->m_rtdev_targp);
1116         }
1117
1118         xfs_unmountfs_writesb(mp);
1119
1120         xfs_unmountfs_wait(mp);                 /* wait for async bufs */
1121
1122         xfs_log_unmount(mp);                    /* Done! No more fs ops. */
1123
1124         xfs_freesb(mp);
1125
1126         /*
1127          * All inodes from this mount point should be freed.
1128          */
1129         ASSERT(mp->m_inodes == NULL);
1130
1131         xfs_unmountfs_close(mp, cr);
1132         if ((mp->m_flags & XFS_MOUNT_NOUUID) == 0)
1133                 xfs_uuid_unmount(mp);
1134
1135 #if defined(DEBUG) || defined(INDUCE_IO_ERROR)
1136         /*
1137          * clear all error tags on this filesystem
1138          */
1139         memcpy(&fsid, &vfsp->vfs_fsid, sizeof(int64_t));
1140         xfs_errortag_clearall_umount(fsid, mp->m_fsname, 0);
1141 #endif
1142         XFS_IODONE(vfsp);
1143         xfs_mount_free(mp, 1);
1144         return 0;
1145 }
1146
1147 void
1148 xfs_unmountfs_close(xfs_mount_t *mp, struct cred *cr)
1149 {
1150         if (mp->m_logdev_targp != mp->m_ddev_targp)
1151                 xfs_free_buftarg(mp->m_logdev_targp, 1);
1152         if (mp->m_rtdev_targp)
1153                 xfs_free_buftarg(mp->m_rtdev_targp, 1);
1154         xfs_free_buftarg(mp->m_ddev_targp, 0);
1155 }
1156
1157 STATIC void
1158 xfs_unmountfs_wait(xfs_mount_t *mp)
1159 {
1160         if (mp->m_logdev_targp != mp->m_ddev_targp)
1161                 xfs_wait_buftarg(mp->m_logdev_targp);
1162         if (mp->m_rtdev_targp)
1163                 xfs_wait_buftarg(mp->m_rtdev_targp);
1164         xfs_wait_buftarg(mp->m_ddev_targp);
1165 }
1166
1167 int
1168 xfs_unmountfs_writesb(xfs_mount_t *mp)
1169 {
1170         xfs_buf_t       *sbp;
1171         xfs_sb_t        *sb;
1172         int             error = 0;
1173
1174         /*
1175          * skip superblock write if fs is read-only, or
1176          * if we are doing a forced umount.
1177          */
1178         sbp = xfs_getsb(mp, 0);
1179         if (!(XFS_MTOVFS(mp)->vfs_flag & VFS_RDONLY ||
1180                 XFS_FORCED_SHUTDOWN(mp))) {
1181
1182                 xfs_icsb_sync_counters(mp);
1183
1184                 /*
1185                  * mark shared-readonly if desired
1186                  */
1187                 sb = XFS_BUF_TO_SBP(sbp);
1188                 if (mp->m_mk_sharedro) {
1189                         if (!(sb->sb_flags & XFS_SBF_READONLY))
1190                                 sb->sb_flags |= XFS_SBF_READONLY;
1191                         if (!XFS_SB_VERSION_HASSHARED(sb))
1192                                 XFS_SB_VERSION_ADDSHARED(sb);
1193                         xfs_fs_cmn_err(CE_NOTE, mp,
1194                                 "Unmounting, marking shared read-only");
1195                 }
1196                 XFS_BUF_UNDONE(sbp);
1197                 XFS_BUF_UNREAD(sbp);
1198                 XFS_BUF_UNDELAYWRITE(sbp);
1199                 XFS_BUF_WRITE(sbp);
1200                 XFS_BUF_UNASYNC(sbp);
1201                 ASSERT(XFS_BUF_TARGET(sbp) == mp->m_ddev_targp);
1202                 xfsbdstrat(mp, sbp);
1203                 /* Nevermind errors we might get here. */
1204                 error = xfs_iowait(sbp);
1205                 if (error)
1206                         xfs_ioerror_alert("xfs_unmountfs_writesb",
1207                                           mp, sbp, XFS_BUF_ADDR(sbp));
1208                 if (error && mp->m_mk_sharedro)
1209                         xfs_fs_cmn_err(CE_ALERT, mp, "Superblock write error detected while unmounting.  Filesystem may not be marked shared readonly");
1210         }
1211         xfs_buf_relse(sbp);
1212         return error;
1213 }
1214
1215 /*
1216  * xfs_mod_sb() can be used to copy arbitrary changes to the
1217  * in-core superblock into the superblock buffer to be logged.
1218  * It does not provide the higher level of locking that is
1219  * needed to protect the in-core superblock from concurrent
1220  * access.
1221  */
1222 void
1223 xfs_mod_sb(xfs_trans_t *tp, __int64_t fields)
1224 {
1225         xfs_buf_t       *bp;
1226         int             first;
1227         int             last;
1228         xfs_mount_t     *mp;
1229         xfs_sb_t        *sbp;
1230         xfs_sb_field_t  f;
1231
1232         ASSERT(fields);
1233         if (!fields)
1234                 return;
1235         mp = tp->t_mountp;
1236         bp = xfs_trans_getsb(tp, mp, 0);
1237         sbp = XFS_BUF_TO_SBP(bp);
1238         first = sizeof(xfs_sb_t);
1239         last = 0;
1240
1241         /* translate/copy */
1242
1243         xfs_xlatesb(XFS_BUF_PTR(bp), &(mp->m_sb), -1, fields);
1244
1245         /* find modified range */
1246
1247         f = (xfs_sb_field_t)xfs_lowbit64((__uint64_t)fields);
1248         ASSERT((1LL << f) & XFS_SB_MOD_BITS);
1249         first = xfs_sb_info[f].offset;
1250
1251         f = (xfs_sb_field_t)xfs_highbit64((__uint64_t)fields);
1252         ASSERT((1LL << f) & XFS_SB_MOD_BITS);
1253         last = xfs_sb_info[f + 1].offset - 1;
1254
1255         xfs_trans_log_buf(tp, bp, first, last);
1256 }
1257 /*
1258  * xfs_mod_incore_sb_unlocked() is a utility routine common used to apply
1259  * a delta to a specified field in the in-core superblock.  Simply
1260  * switch on the field indicated and apply the delta to that field.
1261  * Fields are not allowed to dip below zero, so if the delta would
1262  * do this do not apply it and return EINVAL.
1263  *
1264  * The SB_LOCK must be held when this routine is called.
1265  */
1266 int
1267 xfs_mod_incore_sb_unlocked(xfs_mount_t *mp, xfs_sb_field_t field,
1268                         int delta, int rsvd)
1269 {
1270         int             scounter;       /* short counter for 32 bit fields */
1271         long long       lcounter;       /* long counter for 64 bit fields */
1272         long long       res_used, rem;
1273
1274         /*
1275          * With the in-core superblock spin lock held, switch
1276          * on the indicated field.  Apply the delta to the
1277          * proper field.  If the fields value would dip below
1278          * 0, then do not apply the delta and return EINVAL.
1279          */
1280         switch (field) {
1281         case XFS_SBS_ICOUNT:
1282                 lcounter = (long long)mp->m_sb.sb_icount;
1283                 lcounter += delta;
1284                 if (lcounter < 0) {
1285                         ASSERT(0);
1286                         return XFS_ERROR(EINVAL);
1287                 }
1288                 mp->m_sb.sb_icount = lcounter;
1289                 return 0;
1290         case XFS_SBS_IFREE:
1291                 lcounter = (long long)mp->m_sb.sb_ifree;
1292                 lcounter += delta;
1293                 if (lcounter < 0) {
1294                         ASSERT(0);
1295                         return XFS_ERROR(EINVAL);
1296                 }
1297                 mp->m_sb.sb_ifree = lcounter;
1298                 return 0;
1299         case XFS_SBS_FDBLOCKS:
1300
1301                 lcounter = (long long)mp->m_sb.sb_fdblocks;
1302                 res_used = (long long)(mp->m_resblks - mp->m_resblks_avail);
1303
1304                 if (delta > 0) {                /* Putting blocks back */
1305                         if (res_used > delta) {
1306                                 mp->m_resblks_avail += delta;
1307                         } else {
1308                                 rem = delta - res_used;
1309                                 mp->m_resblks_avail = mp->m_resblks;
1310                                 lcounter += rem;
1311                         }
1312                 } else {                                /* Taking blocks away */
1313
1314                         lcounter += delta;
1315
1316                 /*
1317                  * If were out of blocks, use any available reserved blocks if
1318                  * were allowed to.
1319                  */
1320
1321                         if (lcounter < 0) {
1322                                 if (rsvd) {
1323                                         lcounter = (long long)mp->m_resblks_avail + delta;
1324                                         if (lcounter < 0) {
1325                                                 return XFS_ERROR(ENOSPC);
1326                                         }
1327                                         mp->m_resblks_avail = lcounter;
1328                                         return 0;
1329                                 } else {        /* not reserved */
1330                                         return XFS_ERROR(ENOSPC);
1331                                 }
1332                         }
1333                 }
1334
1335                 mp->m_sb.sb_fdblocks = lcounter;
1336                 return 0;
1337         case XFS_SBS_FREXTENTS:
1338                 lcounter = (long long)mp->m_sb.sb_frextents;
1339                 lcounter += delta;
1340                 if (lcounter < 0) {
1341                         return XFS_ERROR(ENOSPC);
1342                 }
1343                 mp->m_sb.sb_frextents = lcounter;
1344                 return 0;
1345         case XFS_SBS_DBLOCKS:
1346                 lcounter = (long long)mp->m_sb.sb_dblocks;
1347                 lcounter += delta;
1348                 if (lcounter < 0) {
1349                         ASSERT(0);
1350                         return XFS_ERROR(EINVAL);
1351                 }
1352                 mp->m_sb.sb_dblocks = lcounter;
1353                 return 0;
1354         case XFS_SBS_AGCOUNT:
1355                 scounter = mp->m_sb.sb_agcount;
1356                 scounter += delta;
1357                 if (scounter < 0) {
1358                         ASSERT(0);
1359                         return XFS_ERROR(EINVAL);
1360                 }
1361                 mp->m_sb.sb_agcount = scounter;
1362                 return 0;
1363         case XFS_SBS_IMAX_PCT:
1364                 scounter = mp->m_sb.sb_imax_pct;
1365                 scounter += delta;
1366                 if (scounter < 0) {
1367                         ASSERT(0);
1368                         return XFS_ERROR(EINVAL);
1369                 }
1370                 mp->m_sb.sb_imax_pct = scounter;
1371                 return 0;
1372         case XFS_SBS_REXTSIZE:
1373                 scounter = mp->m_sb.sb_rextsize;
1374                 scounter += delta;
1375                 if (scounter < 0) {
1376                         ASSERT(0);
1377                         return XFS_ERROR(EINVAL);
1378                 }
1379                 mp->m_sb.sb_rextsize = scounter;
1380                 return 0;
1381         case XFS_SBS_RBMBLOCKS:
1382                 scounter = mp->m_sb.sb_rbmblocks;
1383                 scounter += delta;
1384                 if (scounter < 0) {
1385                         ASSERT(0);
1386                         return XFS_ERROR(EINVAL);
1387                 }
1388                 mp->m_sb.sb_rbmblocks = scounter;
1389                 return 0;
1390         case XFS_SBS_RBLOCKS:
1391                 lcounter = (long long)mp->m_sb.sb_rblocks;
1392                 lcounter += delta;
1393                 if (lcounter < 0) {
1394                         ASSERT(0);
1395                         return XFS_ERROR(EINVAL);
1396                 }
1397                 mp->m_sb.sb_rblocks = lcounter;
1398                 return 0;
1399         case XFS_SBS_REXTENTS:
1400                 lcounter = (long long)mp->m_sb.sb_rextents;
1401                 lcounter += delta;
1402                 if (lcounter < 0) {
1403                         ASSERT(0);
1404                         return XFS_ERROR(EINVAL);
1405                 }
1406                 mp->m_sb.sb_rextents = lcounter;
1407                 return 0;
1408         case XFS_SBS_REXTSLOG:
1409                 scounter = mp->m_sb.sb_rextslog;
1410                 scounter += delta;
1411                 if (scounter < 0) {
1412                         ASSERT(0);
1413                         return XFS_ERROR(EINVAL);
1414                 }
1415                 mp->m_sb.sb_rextslog = scounter;
1416                 return 0;
1417         default:
1418                 ASSERT(0);
1419                 return XFS_ERROR(EINVAL);
1420         }
1421 }
1422
1423 /*
1424  * xfs_mod_incore_sb() is used to change a field in the in-core
1425  * superblock structure by the specified delta.  This modification
1426  * is protected by the SB_LOCK.  Just use the xfs_mod_incore_sb_unlocked()
1427  * routine to do the work.
1428  */
1429 int
1430 xfs_mod_incore_sb(xfs_mount_t *mp, xfs_sb_field_t field, int delta, int rsvd)
1431 {
1432         unsigned long   s;
1433         int     status;
1434
1435         /* check for per-cpu counters */
1436         switch (field) {
1437 #ifdef HAVE_PERCPU_SB
1438         case XFS_SBS_ICOUNT:
1439         case XFS_SBS_IFREE:
1440         case XFS_SBS_FDBLOCKS:
1441                 if (!(mp->m_flags & XFS_MOUNT_NO_PERCPU_SB)) {
1442                         status = xfs_icsb_modify_counters(mp, field,
1443                                                         delta, rsvd);
1444                         break;
1445                 }
1446                 /* FALLTHROUGH */
1447 #endif
1448         default:
1449                 s = XFS_SB_LOCK(mp);
1450                 status = xfs_mod_incore_sb_unlocked(mp, field, delta, rsvd);
1451                 XFS_SB_UNLOCK(mp, s);
1452                 break;
1453         }
1454
1455         return status;
1456 }
1457
1458 /*
1459  * xfs_mod_incore_sb_batch() is used to change more than one field
1460  * in the in-core superblock structure at a time.  This modification
1461  * is protected by a lock internal to this module.  The fields and
1462  * changes to those fields are specified in the array of xfs_mod_sb
1463  * structures passed in.
1464  *
1465  * Either all of the specified deltas will be applied or none of
1466  * them will.  If any modified field dips below 0, then all modifications
1467  * will be backed out and EINVAL will be returned.
1468  */
1469 int
1470 xfs_mod_incore_sb_batch(xfs_mount_t *mp, xfs_mod_sb_t *msb, uint nmsb, int rsvd)
1471 {
1472         unsigned long   s;
1473         int             status=0;
1474         xfs_mod_sb_t    *msbp;
1475
1476         /*
1477          * Loop through the array of mod structures and apply each
1478          * individually.  If any fail, then back out all those
1479          * which have already been applied.  Do all of this within
1480          * the scope of the SB_LOCK so that all of the changes will
1481          * be atomic.
1482          */
1483         s = XFS_SB_LOCK(mp);
1484         msbp = &msb[0];
1485         for (msbp = &msbp[0]; msbp < (msb + nmsb); msbp++) {
1486                 /*
1487                  * Apply the delta at index n.  If it fails, break
1488                  * from the loop so we'll fall into the undo loop
1489                  * below.
1490                  */
1491                 switch (msbp->msb_field) {
1492 #ifdef HAVE_PERCPU_SB
1493                 case XFS_SBS_ICOUNT:
1494                 case XFS_SBS_IFREE:
1495                 case XFS_SBS_FDBLOCKS:
1496                         if (!(mp->m_flags & XFS_MOUNT_NO_PERCPU_SB)) {
1497                                 status = xfs_icsb_modify_counters_locked(mp,
1498                                                         msbp->msb_field,
1499                                                         msbp->msb_delta, rsvd);
1500                                 break;
1501                         }
1502                         /* FALLTHROUGH */
1503 #endif
1504                 default:
1505                         status = xfs_mod_incore_sb_unlocked(mp,
1506                                                 msbp->msb_field,
1507                                                 msbp->msb_delta, rsvd);
1508                         break;
1509                 }
1510
1511                 if (status != 0) {
1512                         break;
1513                 }
1514         }
1515
1516         /*
1517          * If we didn't complete the loop above, then back out
1518          * any changes made to the superblock.  If you add code
1519          * between the loop above and here, make sure that you
1520          * preserve the value of status. Loop back until
1521          * we step below the beginning of the array.  Make sure
1522          * we don't touch anything back there.
1523          */
1524         if (status != 0) {
1525                 msbp--;
1526                 while (msbp >= msb) {
1527                         switch (msbp->msb_field) {
1528 #ifdef HAVE_PERCPU_SB
1529                         case XFS_SBS_ICOUNT:
1530                         case XFS_SBS_IFREE:
1531                         case XFS_SBS_FDBLOCKS:
1532                                 if (!(mp->m_flags & XFS_MOUNT_NO_PERCPU_SB)) {
1533                                         status =
1534                                             xfs_icsb_modify_counters_locked(mp,
1535                                                         msbp->msb_field,
1536                                                         -(msbp->msb_delta),
1537                                                         rsvd);
1538                                         break;
1539                                 }
1540                                 /* FALLTHROUGH */
1541 #endif
1542                         default:
1543                                 status = xfs_mod_incore_sb_unlocked(mp,
1544                                                         msbp->msb_field,
1545                                                         -(msbp->msb_delta),
1546                                                         rsvd);
1547                                 break;
1548                         }
1549                         ASSERT(status == 0);
1550                         msbp--;
1551                 }
1552         }
1553         XFS_SB_UNLOCK(mp, s);
1554         return status;
1555 }
1556
1557 /*
1558  * xfs_getsb() is called to obtain the buffer for the superblock.
1559  * The buffer is returned locked and read in from disk.
1560  * The buffer should be released with a call to xfs_brelse().
1561  *
1562  * If the flags parameter is BUF_TRYLOCK, then we'll only return
1563  * the superblock buffer if it can be locked without sleeping.
1564  * If it can't then we'll return NULL.
1565  */
1566 xfs_buf_t *
1567 xfs_getsb(
1568         xfs_mount_t     *mp,
1569         int             flags)
1570 {
1571         xfs_buf_t       *bp;
1572
1573         ASSERT(mp->m_sb_bp != NULL);
1574         bp = mp->m_sb_bp;
1575         if (flags & XFS_BUF_TRYLOCK) {
1576                 if (!XFS_BUF_CPSEMA(bp)) {
1577                         return NULL;
1578                 }
1579         } else {
1580                 XFS_BUF_PSEMA(bp, PRIBIO);
1581         }
1582         XFS_BUF_HOLD(bp);
1583         ASSERT(XFS_BUF_ISDONE(bp));
1584         return bp;
1585 }
1586
1587 /*
1588  * Used to free the superblock along various error paths.
1589  */
1590 void
1591 xfs_freesb(
1592         xfs_mount_t     *mp)
1593 {
1594         xfs_buf_t       *bp;
1595
1596         /*
1597          * Use xfs_getsb() so that the buffer will be locked
1598          * when we call xfs_buf_relse().
1599          */
1600         bp = xfs_getsb(mp, 0);
1601         XFS_BUF_UNMANAGE(bp);
1602         xfs_buf_relse(bp);
1603         mp->m_sb_bp = NULL;
1604 }
1605
1606 /*
1607  * See if the UUID is unique among mounted XFS filesystems.
1608  * Mount fails if UUID is nil or a FS with the same UUID is already mounted.
1609  */
1610 STATIC int
1611 xfs_uuid_mount(
1612         xfs_mount_t     *mp)
1613 {
1614         if (uuid_is_nil(&mp->m_sb.sb_uuid)) {
1615                 cmn_err(CE_WARN,
1616                         "XFS: Filesystem %s has nil UUID - can't mount",
1617                         mp->m_fsname);
1618                 return -1;
1619         }
1620         if (!uuid_table_insert(&mp->m_sb.sb_uuid)) {
1621                 cmn_err(CE_WARN,
1622                         "XFS: Filesystem %s has duplicate UUID - can't mount",
1623                         mp->m_fsname);
1624                 return -1;
1625         }
1626         return 0;
1627 }
1628
1629 /*
1630  * Remove filesystem from the UUID table.
1631  */
1632 STATIC void
1633 xfs_uuid_unmount(
1634         xfs_mount_t     *mp)
1635 {
1636         uuid_table_remove(&mp->m_sb.sb_uuid);
1637 }
1638
1639 /*
1640  * Used to log changes to the superblock unit and width fields which could
1641  * be altered by the mount options. Only the first superblock is updated.
1642  */
1643 STATIC void
1644 xfs_mount_log_sbunit(
1645         xfs_mount_t     *mp,
1646         __int64_t       fields)
1647 {
1648         xfs_trans_t     *tp;
1649
1650         ASSERT(fields & (XFS_SB_UNIT|XFS_SB_WIDTH|XFS_SB_UUID));
1651
1652         tp = xfs_trans_alloc(mp, XFS_TRANS_SB_UNIT);
1653         if (xfs_trans_reserve(tp, 0, mp->m_sb.sb_sectsize + 128, 0, 0,
1654                                 XFS_DEFAULT_LOG_COUNT)) {
1655                 xfs_trans_cancel(tp, 0);
1656                 return;
1657         }
1658         xfs_mod_sb(tp, fields);
1659         xfs_trans_commit(tp, 0, NULL);
1660 }
1661
1662
1663 #ifdef HAVE_PERCPU_SB
1664 /*
1665  * Per-cpu incore superblock counters
1666  *
1667  * Simple concept, difficult implementation
1668  *
1669  * Basically, replace the incore superblock counters with a distributed per cpu
1670  * counter for contended fields (e.g.  free block count).
1671  *
1672  * Difficulties arise in that the incore sb is used for ENOSPC checking, and
1673  * hence needs to be accurately read when we are running low on space. Hence
1674  * there is a method to enable and disable the per-cpu counters based on how
1675  * much "stuff" is available in them.
1676  *
1677  * Basically, a counter is enabled if there is enough free resource to justify
1678  * running a per-cpu fast-path. If the per-cpu counter runs out (i.e. a local
1679  * ENOSPC), then we disable the counters to synchronise all callers and
1680  * re-distribute the available resources.
1681  *
1682  * If, once we redistributed the available resources, we still get a failure,
1683  * we disable the per-cpu counter and go through the slow path.
1684  *
1685  * The slow path is the current xfs_mod_incore_sb() function.  This means that
1686  * when we disable a per-cpu counter, we need to drain it's resources back to
1687  * the global superblock. We do this after disabling the counter to prevent
1688  * more threads from queueing up on the counter.
1689  *
1690  * Essentially, this means that we still need a lock in the fast path to enable
1691  * synchronisation between the global counters and the per-cpu counters. This
1692  * is not a problem because the lock will be local to a CPU almost all the time
1693  * and have little contention except when we get to ENOSPC conditions.
1694  *
1695  * Basically, this lock becomes a barrier that enables us to lock out the fast
1696  * path while we do things like enabling and disabling counters and
1697  * synchronising the counters.
1698  *
1699  * Locking rules:
1700  *
1701  *      1. XFS_SB_LOCK() before picking up per-cpu locks
1702  *      2. per-cpu locks always picked up via for_each_online_cpu() order
1703  *      3. accurate counter sync requires XFS_SB_LOCK + per cpu locks
1704  *      4. modifying per-cpu counters requires holding per-cpu lock
1705  *      5. modifying global counters requires holding XFS_SB_LOCK
1706  *      6. enabling or disabling a counter requires holding the XFS_SB_LOCK
1707  *         and _none_ of the per-cpu locks.
1708  *
1709  * Disabled counters are only ever re-enabled by a balance operation
1710  * that results in more free resources per CPU than a given threshold.
1711  * To ensure counters don't remain disabled, they are rebalanced when
1712  * the global resource goes above a higher threshold (i.e. some hysteresis
1713  * is present to prevent thrashing).
1714  */
1715
1716 /*
1717  * hot-plug CPU notifier support.
1718  *
1719  * We cannot use the hotcpu_register() function because it does
1720  * not allow notifier instances. We need a notifier per filesystem
1721  * as we need to be able to identify the filesystem to balance
1722  * the counters out. This is achieved by having a notifier block
1723  * embedded in the xfs_mount_t and doing pointer magic to get the
1724  * mount pointer from the notifier block address.
1725  */
1726 STATIC int
1727 xfs_icsb_cpu_notify(
1728         struct notifier_block *nfb,
1729         unsigned long action,
1730         void *hcpu)
1731 {
1732         xfs_icsb_cnts_t *cntp;
1733         xfs_mount_t     *mp;
1734         int             s;
1735
1736         mp = (xfs_mount_t *)container_of(nfb, xfs_mount_t, m_icsb_notifier);
1737         cntp = (xfs_icsb_cnts_t *)
1738                         per_cpu_ptr(mp->m_sb_cnts, (unsigned long)hcpu);
1739         switch (action) {
1740         case CPU_UP_PREPARE:
1741                 /* Easy Case - initialize the area and locks, and
1742                  * then rebalance when online does everything else for us. */
1743                 memset(cntp, 0, sizeof(xfs_icsb_cnts_t));
1744                 break;
1745         case CPU_ONLINE:
1746                 xfs_icsb_balance_counter(mp, XFS_SBS_ICOUNT, 0);
1747                 xfs_icsb_balance_counter(mp, XFS_SBS_IFREE, 0);
1748                 xfs_icsb_balance_counter(mp, XFS_SBS_FDBLOCKS, 0);
1749                 break;
1750         case CPU_DEAD:
1751                 /* Disable all the counters, then fold the dead cpu's
1752                  * count into the total on the global superblock and
1753                  * re-enable the counters. */
1754                 s = XFS_SB_LOCK(mp);
1755                 xfs_icsb_disable_counter(mp, XFS_SBS_ICOUNT);
1756                 xfs_icsb_disable_counter(mp, XFS_SBS_IFREE);
1757                 xfs_icsb_disable_counter(mp, XFS_SBS_FDBLOCKS);
1758
1759                 mp->m_sb.sb_icount += cntp->icsb_icount;
1760                 mp->m_sb.sb_ifree += cntp->icsb_ifree;
1761                 mp->m_sb.sb_fdblocks += cntp->icsb_fdblocks;
1762
1763                 memset(cntp, 0, sizeof(xfs_icsb_cnts_t));
1764
1765                 xfs_icsb_balance_counter(mp, XFS_SBS_ICOUNT, XFS_ICSB_SB_LOCKED);
1766                 xfs_icsb_balance_counter(mp, XFS_SBS_IFREE, XFS_ICSB_SB_LOCKED);
1767                 xfs_icsb_balance_counter(mp, XFS_SBS_FDBLOCKS, XFS_ICSB_SB_LOCKED);
1768                 XFS_SB_UNLOCK(mp, s);
1769                 break;
1770         }
1771
1772         return NOTIFY_OK;
1773 }
1774
1775 int
1776 xfs_icsb_init_counters(
1777         xfs_mount_t     *mp)
1778 {
1779         xfs_icsb_cnts_t *cntp;
1780         int             i;
1781
1782         mp->m_sb_cnts = alloc_percpu(xfs_icsb_cnts_t);
1783         if (mp->m_sb_cnts == NULL)
1784                 return -ENOMEM;
1785
1786         mp->m_icsb_notifier.notifier_call = xfs_icsb_cpu_notify;
1787         mp->m_icsb_notifier.priority = 0;
1788         register_cpu_notifier(&mp->m_icsb_notifier);
1789
1790         for_each_online_cpu(i) {
1791                 cntp = (xfs_icsb_cnts_t *)per_cpu_ptr(mp->m_sb_cnts, i);
1792                 memset(cntp, 0, sizeof(xfs_icsb_cnts_t));
1793         }
1794         /*
1795          * start with all counters disabled so that the
1796          * initial balance kicks us off correctly
1797          */
1798         mp->m_icsb_counters = -1;
1799         return 0;
1800 }
1801
1802 STATIC void
1803 xfs_icsb_destroy_counters(
1804         xfs_mount_t     *mp)
1805 {
1806         if (mp->m_sb_cnts) {
1807                 unregister_cpu_notifier(&mp->m_icsb_notifier);
1808                 free_percpu(mp->m_sb_cnts);
1809         }
1810 }
1811
1812 STATIC inline void
1813 xfs_icsb_lock_cntr(
1814         xfs_icsb_cnts_t *icsbp)
1815 {
1816         while (test_and_set_bit(XFS_ICSB_FLAG_LOCK, &icsbp->icsb_flags)) {
1817                 ndelay(1000);
1818         }
1819 }
1820
1821 STATIC inline void
1822 xfs_icsb_unlock_cntr(
1823         xfs_icsb_cnts_t *icsbp)
1824 {
1825         clear_bit(XFS_ICSB_FLAG_LOCK, &icsbp->icsb_flags);
1826 }
1827
1828
1829 STATIC inline void
1830 xfs_icsb_lock_all_counters(
1831         xfs_mount_t     *mp)
1832 {
1833         xfs_icsb_cnts_t *cntp;
1834         int             i;
1835
1836         for_each_online_cpu(i) {
1837                 cntp = (xfs_icsb_cnts_t *)per_cpu_ptr(mp->m_sb_cnts, i);
1838                 xfs_icsb_lock_cntr(cntp);
1839         }
1840 }
1841
1842 STATIC inline void
1843 xfs_icsb_unlock_all_counters(
1844         xfs_mount_t     *mp)
1845 {
1846         xfs_icsb_cnts_t *cntp;
1847         int             i;
1848
1849         for_each_online_cpu(i) {
1850                 cntp = (xfs_icsb_cnts_t *)per_cpu_ptr(mp->m_sb_cnts, i);
1851                 xfs_icsb_unlock_cntr(cntp);
1852         }
1853 }
1854
1855 STATIC void
1856 xfs_icsb_count(
1857         xfs_mount_t     *mp,
1858         xfs_icsb_cnts_t *cnt,
1859         int             flags)
1860 {
1861         xfs_icsb_cnts_t *cntp;
1862         int             i;
1863
1864         memset(cnt, 0, sizeof(xfs_icsb_cnts_t));
1865
1866         if (!(flags & XFS_ICSB_LAZY_COUNT))
1867                 xfs_icsb_lock_all_counters(mp);
1868
1869         for_each_online_cpu(i) {
1870                 cntp = (xfs_icsb_cnts_t *)per_cpu_ptr(mp->m_sb_cnts, i);
1871                 cnt->icsb_icount += cntp->icsb_icount;
1872                 cnt->icsb_ifree += cntp->icsb_ifree;
1873                 cnt->icsb_fdblocks += cntp->icsb_fdblocks;
1874         }
1875
1876         if (!(flags & XFS_ICSB_LAZY_COUNT))
1877                 xfs_icsb_unlock_all_counters(mp);
1878 }
1879
1880 STATIC int
1881 xfs_icsb_counter_disabled(
1882         xfs_mount_t     *mp,
1883         xfs_sb_field_t  field)
1884 {
1885         ASSERT((field >= XFS_SBS_ICOUNT) && (field <= XFS_SBS_FDBLOCKS));
1886         return test_bit(field, &mp->m_icsb_counters);
1887 }
1888
1889 STATIC int
1890 xfs_icsb_disable_counter(
1891         xfs_mount_t     *mp,
1892         xfs_sb_field_t  field)
1893 {
1894         xfs_icsb_cnts_t cnt;
1895
1896         ASSERT((field >= XFS_SBS_ICOUNT) && (field <= XFS_SBS_FDBLOCKS));
1897
1898         xfs_icsb_lock_all_counters(mp);
1899         if (!test_and_set_bit(field, &mp->m_icsb_counters)) {
1900                 /* drain back to superblock */
1901
1902                 xfs_icsb_count(mp, &cnt, XFS_ICSB_SB_LOCKED|XFS_ICSB_LAZY_COUNT);
1903                 switch(field) {
1904                 case XFS_SBS_ICOUNT:
1905                         mp->m_sb.sb_icount = cnt.icsb_icount;
1906                         break;
1907                 case XFS_SBS_IFREE:
1908                         mp->m_sb.sb_ifree = cnt.icsb_ifree;
1909                         break;
1910                 case XFS_SBS_FDBLOCKS:
1911                         mp->m_sb.sb_fdblocks = cnt.icsb_fdblocks;
1912                         break;
1913                 default:
1914                         BUG();
1915                 }
1916         }
1917
1918         xfs_icsb_unlock_all_counters(mp);
1919
1920         return 0;
1921 }
1922
1923 STATIC void
1924 xfs_icsb_enable_counter(
1925         xfs_mount_t     *mp,
1926         xfs_sb_field_t  field,
1927         uint64_t        count,
1928         uint64_t        resid)
1929 {
1930         xfs_icsb_cnts_t *cntp;
1931         int             i;
1932
1933         ASSERT((field >= XFS_SBS_ICOUNT) && (field <= XFS_SBS_FDBLOCKS));
1934
1935         xfs_icsb_lock_all_counters(mp);
1936         for_each_online_cpu(i) {
1937                 cntp = per_cpu_ptr(mp->m_sb_cnts, i);
1938                 switch (field) {
1939                 case XFS_SBS_ICOUNT:
1940                         cntp->icsb_icount = count + resid;
1941                         break;
1942                 case XFS_SBS_IFREE:
1943                         cntp->icsb_ifree = count + resid;
1944                         break;
1945                 case XFS_SBS_FDBLOCKS:
1946                         cntp->icsb_fdblocks = count + resid;
1947                         break;
1948                 default:
1949                         BUG();
1950                         break;
1951                 }
1952                 resid = 0;
1953         }
1954         clear_bit(field, &mp->m_icsb_counters);
1955         xfs_icsb_unlock_all_counters(mp);
1956 }
1957
1958 STATIC void
1959 xfs_icsb_sync_counters_int(
1960         xfs_mount_t     *mp,
1961         int             flags)
1962 {
1963         xfs_icsb_cnts_t cnt;
1964         int             s;
1965
1966         /* Pass 1: lock all counters */
1967         if ((flags & XFS_ICSB_SB_LOCKED) == 0)
1968                 s = XFS_SB_LOCK(mp);
1969
1970         xfs_icsb_count(mp, &cnt, flags);
1971
1972         /* Step 3: update mp->m_sb fields */
1973         if (!xfs_icsb_counter_disabled(mp, XFS_SBS_ICOUNT))
1974                 mp->m_sb.sb_icount = cnt.icsb_icount;
1975         if (!xfs_icsb_counter_disabled(mp, XFS_SBS_IFREE))
1976                 mp->m_sb.sb_ifree = cnt.icsb_ifree;
1977         if (!xfs_icsb_counter_disabled(mp, XFS_SBS_FDBLOCKS))
1978                 mp->m_sb.sb_fdblocks = cnt.icsb_fdblocks;
1979
1980         if ((flags & XFS_ICSB_SB_LOCKED) == 0)
1981                 XFS_SB_UNLOCK(mp, s);
1982 }
1983
1984 /*
1985  * Accurate update of per-cpu counters to incore superblock
1986  */
1987 STATIC void
1988 xfs_icsb_sync_counters(
1989         xfs_mount_t     *mp)
1990 {
1991         xfs_icsb_sync_counters_int(mp, 0);
1992 }
1993
1994 /*
1995  * lazy addition used for things like df, background sb syncs, etc
1996  */
1997 void
1998 xfs_icsb_sync_counters_lazy(
1999         xfs_mount_t     *mp)
2000 {
2001         xfs_icsb_sync_counters_int(mp, XFS_ICSB_LAZY_COUNT);
2002 }
2003
2004 /*
2005  * Balance and enable/disable counters as necessary.
2006  *
2007  * Thresholds for re-enabling counters are somewhat magic.
2008  * inode counts are chosen to be the same number as single
2009  * on disk allocation chunk per CPU, and free blocks is
2010  * something far enough zero that we aren't going thrash
2011  * when we get near ENOSPC.
2012  */
2013 #define XFS_ICSB_INO_CNTR_REENABLE      64
2014 #define XFS_ICSB_FDBLK_CNTR_REENABLE    512
2015 STATIC void
2016 xfs_icsb_balance_counter(
2017         xfs_mount_t     *mp,
2018         xfs_sb_field_t  field,
2019         int             flags)
2020 {
2021         uint64_t        count, resid = 0;
2022         int             weight = num_online_cpus();
2023         int             s;
2024
2025         if (!(flags & XFS_ICSB_SB_LOCKED))
2026                 s = XFS_SB_LOCK(mp);
2027
2028         /* disable counter and sync counter */
2029         xfs_icsb_disable_counter(mp, field);
2030
2031         /* update counters  - first CPU gets residual*/
2032         switch (field) {
2033         case XFS_SBS_ICOUNT:
2034                 count = mp->m_sb.sb_icount;
2035                 resid = do_div(count, weight);
2036                 if (count < XFS_ICSB_INO_CNTR_REENABLE)
2037                         goto out;
2038                 break;
2039         case XFS_SBS_IFREE:
2040                 count = mp->m_sb.sb_ifree;
2041                 resid = do_div(count, weight);
2042                 if (count < XFS_ICSB_INO_CNTR_REENABLE)
2043                         goto out;
2044                 break;
2045         case XFS_SBS_FDBLOCKS:
2046                 count = mp->m_sb.sb_fdblocks;
2047                 resid = do_div(count, weight);
2048                 if (count < XFS_ICSB_FDBLK_CNTR_REENABLE)
2049                         goto out;
2050                 break;
2051         default:
2052                 BUG();
2053                 break;
2054         }
2055
2056         xfs_icsb_enable_counter(mp, field, count, resid);
2057 out:
2058         if (!(flags & XFS_ICSB_SB_LOCKED))
2059                 XFS_SB_UNLOCK(mp, s);
2060 }
2061
2062 STATIC int
2063 xfs_icsb_modify_counters_int(
2064         xfs_mount_t     *mp,
2065         xfs_sb_field_t  field,
2066         int             delta,
2067         int             rsvd,
2068         int             flags)
2069 {
2070         xfs_icsb_cnts_t *icsbp;
2071         long long       lcounter;       /* long counter for 64 bit fields */
2072         int             cpu, s, locked = 0;
2073         int             ret = 0, balance_done = 0;
2074
2075 again:
2076         cpu = get_cpu();
2077         icsbp = (xfs_icsb_cnts_t *)per_cpu_ptr(mp->m_sb_cnts, cpu),
2078         xfs_icsb_lock_cntr(icsbp);
2079         if (unlikely(xfs_icsb_counter_disabled(mp, field)))
2080                 goto slow_path;
2081
2082         switch (field) {
2083         case XFS_SBS_ICOUNT:
2084                 lcounter = icsbp->icsb_icount;
2085                 lcounter += delta;
2086                 if (unlikely(lcounter < 0))
2087                         goto slow_path;
2088                 icsbp->icsb_icount = lcounter;
2089                 break;
2090
2091         case XFS_SBS_IFREE:
2092                 lcounter = icsbp->icsb_ifree;
2093                 lcounter += delta;
2094                 if (unlikely(lcounter < 0))
2095                         goto slow_path;
2096                 icsbp->icsb_ifree = lcounter;
2097                 break;
2098
2099         case XFS_SBS_FDBLOCKS:
2100                 BUG_ON((mp->m_resblks - mp->m_resblks_avail) != 0);
2101
2102                 lcounter = icsbp->icsb_fdblocks;
2103                 lcounter += delta;
2104                 if (unlikely(lcounter < 0))
2105                         goto slow_path;
2106                 icsbp->icsb_fdblocks = lcounter;
2107                 break;
2108         default:
2109                 BUG();
2110                 break;
2111         }
2112         xfs_icsb_unlock_cntr(icsbp);
2113         put_cpu();
2114         if (locked)
2115                 XFS_SB_UNLOCK(mp, s);
2116         return 0;
2117
2118         /*
2119          * The slow path needs to be run with the SBLOCK
2120          * held so that we prevent other threads from
2121          * attempting to run this path at the same time.
2122          * this provides exclusion for the balancing code,
2123          * and exclusive fallback if the balance does not
2124          * provide enough resources to continue in an unlocked
2125          * manner.
2126          */
2127 slow_path:
2128         xfs_icsb_unlock_cntr(icsbp);
2129         put_cpu();
2130
2131         /* need to hold superblock incase we need
2132          * to disable a counter */
2133         if (!(flags & XFS_ICSB_SB_LOCKED)) {
2134                 s = XFS_SB_LOCK(mp);
2135                 locked = 1;
2136                 flags |= XFS_ICSB_SB_LOCKED;
2137         }
2138         if (!balance_done) {
2139                 xfs_icsb_balance_counter(mp, field, flags);
2140                 balance_done = 1;
2141                 goto again;
2142         } else {
2143                 /*
2144                  * we might not have enough on this local
2145                  * cpu to allocate for a bulk request.
2146                  * We need to drain this field from all CPUs
2147                  * and disable the counter fastpath
2148                  */
2149                 xfs_icsb_disable_counter(mp, field);
2150         }
2151
2152         ret = xfs_mod_incore_sb_unlocked(mp, field, delta, rsvd);
2153
2154         if (locked)
2155                 XFS_SB_UNLOCK(mp, s);
2156         return ret;
2157 }
2158
2159 STATIC int
2160 xfs_icsb_modify_counters(
2161         xfs_mount_t     *mp,
2162         xfs_sb_field_t  field,
2163         int             delta,
2164         int             rsvd)
2165 {
2166         return xfs_icsb_modify_counters_int(mp, field, delta, rsvd, 0);
2167 }
2168
2169 /*
2170  * Called when superblock is already locked
2171  */
2172 STATIC int
2173 xfs_icsb_modify_counters_locked(
2174         xfs_mount_t     *mp,
2175         xfs_sb_field_t  field,
2176         int             delta,
2177         int             rsvd)
2178 {
2179         return xfs_icsb_modify_counters_int(mp, field, delta,
2180                                                 rsvd, XFS_ICSB_SB_LOCKED);
2181 }
2182 #endif