[XFS] Fix block reservation mechanism.
[linux-drm-fsl-dcu.git] / fs / xfs / xfs_mount.c
1 /*
2  * Copyright (c) 2000-2005 Silicon Graphics, Inc.
3  * All Rights Reserved.
4  *
5  * This program is free software; you can redistribute it and/or
6  * modify it under the terms of the GNU General Public License as
7  * published by the Free Software Foundation.
8  *
9  * This program is distributed in the hope that it would be useful,
10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
12  * GNU General Public License for more details.
13  *
14  * You should have received a copy of the GNU General Public License
15  * along with this program; if not, write the Free Software Foundation,
16  * Inc.,  51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
17  */
18 #include "xfs.h"
19 #include "xfs_fs.h"
20 #include "xfs_types.h"
21 #include "xfs_bit.h"
22 #include "xfs_log.h"
23 #include "xfs_inum.h"
24 #include "xfs_trans.h"
25 #include "xfs_sb.h"
26 #include "xfs_ag.h"
27 #include "xfs_dir2.h"
28 #include "xfs_dmapi.h"
29 #include "xfs_mount.h"
30 #include "xfs_bmap_btree.h"
31 #include "xfs_alloc_btree.h"
32 #include "xfs_ialloc_btree.h"
33 #include "xfs_dir2_sf.h"
34 #include "xfs_attr_sf.h"
35 #include "xfs_dinode.h"
36 #include "xfs_inode.h"
37 #include "xfs_btree.h"
38 #include "xfs_ialloc.h"
39 #include "xfs_alloc.h"
40 #include "xfs_rtalloc.h"
41 #include "xfs_bmap.h"
42 #include "xfs_error.h"
43 #include "xfs_rw.h"
44 #include "xfs_quota.h"
45 #include "xfs_fsops.h"
46
47 STATIC void     xfs_mount_log_sbunit(xfs_mount_t *, __int64_t);
48 STATIC int      xfs_uuid_mount(xfs_mount_t *);
49 STATIC void     xfs_uuid_unmount(xfs_mount_t *mp);
50 STATIC void     xfs_unmountfs_wait(xfs_mount_t *);
51
52
53 #ifdef HAVE_PERCPU_SB
54 STATIC void     xfs_icsb_destroy_counters(xfs_mount_t *);
55 STATIC void     xfs_icsb_balance_counter(xfs_mount_t *, xfs_sb_field_t,
56                                                 int, int);
57 STATIC void     xfs_icsb_sync_counters(xfs_mount_t *);
58 STATIC int      xfs_icsb_modify_counters(xfs_mount_t *, xfs_sb_field_t,
59                                                 int64_t, int);
60 STATIC int      xfs_icsb_disable_counter(xfs_mount_t *, xfs_sb_field_t);
61
62 #else
63
64 #define xfs_icsb_destroy_counters(mp)                   do { } while (0)
65 #define xfs_icsb_balance_counter(mp, a, b, c)           do { } while (0)
66 #define xfs_icsb_sync_counters(mp)                      do { } while (0)
67 #define xfs_icsb_modify_counters(mp, a, b, c)           do { } while (0)
68
69 #endif
70
71 static const struct {
72         short offset;
73         short type;     /* 0 = integer
74                          * 1 = binary / string (no translation)
75                          */
76 } xfs_sb_info[] = {
77     { offsetof(xfs_sb_t, sb_magicnum),   0 },
78     { offsetof(xfs_sb_t, sb_blocksize),  0 },
79     { offsetof(xfs_sb_t, sb_dblocks),    0 },
80     { offsetof(xfs_sb_t, sb_rblocks),    0 },
81     { offsetof(xfs_sb_t, sb_rextents),   0 },
82     { offsetof(xfs_sb_t, sb_uuid),       1 },
83     { offsetof(xfs_sb_t, sb_logstart),   0 },
84     { offsetof(xfs_sb_t, sb_rootino),    0 },
85     { offsetof(xfs_sb_t, sb_rbmino),     0 },
86     { offsetof(xfs_sb_t, sb_rsumino),    0 },
87     { offsetof(xfs_sb_t, sb_rextsize),   0 },
88     { offsetof(xfs_sb_t, sb_agblocks),   0 },
89     { offsetof(xfs_sb_t, sb_agcount),    0 },
90     { offsetof(xfs_sb_t, sb_rbmblocks),  0 },
91     { offsetof(xfs_sb_t, sb_logblocks),  0 },
92     { offsetof(xfs_sb_t, sb_versionnum), 0 },
93     { offsetof(xfs_sb_t, sb_sectsize),   0 },
94     { offsetof(xfs_sb_t, sb_inodesize),  0 },
95     { offsetof(xfs_sb_t, sb_inopblock),  0 },
96     { offsetof(xfs_sb_t, sb_fname[0]),   1 },
97     { offsetof(xfs_sb_t, sb_blocklog),   0 },
98     { offsetof(xfs_sb_t, sb_sectlog),    0 },
99     { offsetof(xfs_sb_t, sb_inodelog),   0 },
100     { offsetof(xfs_sb_t, sb_inopblog),   0 },
101     { offsetof(xfs_sb_t, sb_agblklog),   0 },
102     { offsetof(xfs_sb_t, sb_rextslog),   0 },
103     { offsetof(xfs_sb_t, sb_inprogress), 0 },
104     { offsetof(xfs_sb_t, sb_imax_pct),   0 },
105     { offsetof(xfs_sb_t, sb_icount),     0 },
106     { offsetof(xfs_sb_t, sb_ifree),      0 },
107     { offsetof(xfs_sb_t, sb_fdblocks),   0 },
108     { offsetof(xfs_sb_t, sb_frextents),  0 },
109     { offsetof(xfs_sb_t, sb_uquotino),   0 },
110     { offsetof(xfs_sb_t, sb_gquotino),   0 },
111     { offsetof(xfs_sb_t, sb_qflags),     0 },
112     { offsetof(xfs_sb_t, sb_flags),      0 },
113     { offsetof(xfs_sb_t, sb_shared_vn),  0 },
114     { offsetof(xfs_sb_t, sb_inoalignmt), 0 },
115     { offsetof(xfs_sb_t, sb_unit),       0 },
116     { offsetof(xfs_sb_t, sb_width),      0 },
117     { offsetof(xfs_sb_t, sb_dirblklog),  0 },
118     { offsetof(xfs_sb_t, sb_logsectlog), 0 },
119     { offsetof(xfs_sb_t, sb_logsectsize),0 },
120     { offsetof(xfs_sb_t, sb_logsunit),   0 },
121     { offsetof(xfs_sb_t, sb_features2),  0 },
122     { sizeof(xfs_sb_t),                  0 }
123 };
124
125 /*
126  * Return a pointer to an initialized xfs_mount structure.
127  */
128 xfs_mount_t *
129 xfs_mount_init(void)
130 {
131         xfs_mount_t *mp;
132
133         mp = kmem_zalloc(sizeof(xfs_mount_t), KM_SLEEP);
134
135         if (xfs_icsb_init_counters(mp)) {
136                 mp->m_flags |= XFS_MOUNT_NO_PERCPU_SB;
137         }
138
139         AIL_LOCKINIT(&mp->m_ail_lock, "xfs_ail");
140         spinlock_init(&mp->m_sb_lock, "xfs_sb");
141         mutex_init(&mp->m_ilock);
142         initnsema(&mp->m_growlock, 1, "xfs_grow");
143         /*
144          * Initialize the AIL.
145          */
146         xfs_trans_ail_init(mp);
147
148         atomic_set(&mp->m_active_trans, 0);
149
150         return mp;
151 }
152
153 /*
154  * Free up the resources associated with a mount structure.  Assume that
155  * the structure was initially zeroed, so we can tell which fields got
156  * initialized.
157  */
158 void
159 xfs_mount_free(
160         xfs_mount_t     *mp,
161         int             remove_bhv)
162 {
163         if (mp->m_ihash)
164                 xfs_ihash_free(mp);
165         if (mp->m_chash)
166                 xfs_chash_free(mp);
167
168         if (mp->m_perag) {
169                 int     agno;
170
171                 for (agno = 0; agno < mp->m_maxagi; agno++)
172                         if (mp->m_perag[agno].pagb_list)
173                                 kmem_free(mp->m_perag[agno].pagb_list,
174                                                 sizeof(xfs_perag_busy_t) *
175                                                         XFS_PAGB_NUM_SLOTS);
176                 kmem_free(mp->m_perag,
177                           sizeof(xfs_perag_t) * mp->m_sb.sb_agcount);
178         }
179
180         AIL_LOCK_DESTROY(&mp->m_ail_lock);
181         spinlock_destroy(&mp->m_sb_lock);
182         mutex_destroy(&mp->m_ilock);
183         freesema(&mp->m_growlock);
184         if (mp->m_quotainfo)
185                 XFS_QM_DONE(mp);
186
187         if (mp->m_fsname != NULL)
188                 kmem_free(mp->m_fsname, mp->m_fsname_len);
189         if (mp->m_rtname != NULL)
190                 kmem_free(mp->m_rtname, strlen(mp->m_rtname) + 1);
191         if (mp->m_logname != NULL)
192                 kmem_free(mp->m_logname, strlen(mp->m_logname) + 1);
193
194         if (remove_bhv) {
195                 struct bhv_vfs  *vfsp = XFS_MTOVFS(mp);
196
197                 bhv_remove_all_vfsops(vfsp, 0);
198                 VFS_REMOVEBHV(vfsp, &mp->m_bhv);
199         }
200
201         xfs_icsb_destroy_counters(mp);
202         kmem_free(mp, sizeof(xfs_mount_t));
203 }
204
205
206 /*
207  * Check the validity of the SB found.
208  */
209 STATIC int
210 xfs_mount_validate_sb(
211         xfs_mount_t     *mp,
212         xfs_sb_t        *sbp,
213         int             flags)
214 {
215         /*
216          * If the log device and data device have the
217          * same device number, the log is internal.
218          * Consequently, the sb_logstart should be non-zero.  If
219          * we have a zero sb_logstart in this case, we may be trying to mount
220          * a volume filesystem in a non-volume manner.
221          */
222         if (sbp->sb_magicnum != XFS_SB_MAGIC) {
223                 xfs_fs_mount_cmn_err(flags, "bad magic number");
224                 return XFS_ERROR(EWRONGFS);
225         }
226
227         if (!XFS_SB_GOOD_VERSION(sbp)) {
228                 xfs_fs_mount_cmn_err(flags, "bad version");
229                 return XFS_ERROR(EWRONGFS);
230         }
231
232         if (unlikely(
233             sbp->sb_logstart == 0 && mp->m_logdev_targp == mp->m_ddev_targp)) {
234                 xfs_fs_mount_cmn_err(flags,
235                         "filesystem is marked as having an external log; "
236                         "specify logdev on the\nmount command line.");
237                 return XFS_ERROR(EINVAL);
238         }
239
240         if (unlikely(
241             sbp->sb_logstart != 0 && mp->m_logdev_targp != mp->m_ddev_targp)) {
242                 xfs_fs_mount_cmn_err(flags,
243                         "filesystem is marked as having an internal log; "
244                         "do not specify logdev on\nthe mount command line.");
245                 return XFS_ERROR(EINVAL);
246         }
247
248         /*
249          * More sanity checking. These were stolen directly from
250          * xfs_repair.
251          */
252         if (unlikely(
253             sbp->sb_agcount <= 0                                        ||
254             sbp->sb_sectsize < XFS_MIN_SECTORSIZE                       ||
255             sbp->sb_sectsize > XFS_MAX_SECTORSIZE                       ||
256             sbp->sb_sectlog < XFS_MIN_SECTORSIZE_LOG                    ||
257             sbp->sb_sectlog > XFS_MAX_SECTORSIZE_LOG                    ||
258             sbp->sb_blocksize < XFS_MIN_BLOCKSIZE                       ||
259             sbp->sb_blocksize > XFS_MAX_BLOCKSIZE                       ||
260             sbp->sb_blocklog < XFS_MIN_BLOCKSIZE_LOG                    ||
261             sbp->sb_blocklog > XFS_MAX_BLOCKSIZE_LOG                    ||
262             sbp->sb_inodesize < XFS_DINODE_MIN_SIZE                     ||
263             sbp->sb_inodesize > XFS_DINODE_MAX_SIZE                     ||
264             sbp->sb_inodelog < XFS_DINODE_MIN_LOG                       ||
265             sbp->sb_inodelog > XFS_DINODE_MAX_LOG                       ||
266             (sbp->sb_blocklog - sbp->sb_inodelog != sbp->sb_inopblog)   ||
267             (sbp->sb_rextsize * sbp->sb_blocksize > XFS_MAX_RTEXTSIZE)  ||
268             (sbp->sb_rextsize * sbp->sb_blocksize < XFS_MIN_RTEXTSIZE)  ||
269             (sbp->sb_imax_pct > 100 /* zero sb_imax_pct is valid */))) {
270                 xfs_fs_mount_cmn_err(flags, "SB sanity check 1 failed");
271                 return XFS_ERROR(EFSCORRUPTED);
272         }
273
274         /*
275          * Sanity check AG count, size fields against data size field
276          */
277         if (unlikely(
278             sbp->sb_dblocks == 0 ||
279             sbp->sb_dblocks >
280              (xfs_drfsbno_t)sbp->sb_agcount * sbp->sb_agblocks ||
281             sbp->sb_dblocks < (xfs_drfsbno_t)(sbp->sb_agcount - 1) *
282                               sbp->sb_agblocks + XFS_MIN_AG_BLOCKS)) {
283                 xfs_fs_mount_cmn_err(flags, "SB sanity check 2 failed");
284                 return XFS_ERROR(EFSCORRUPTED);
285         }
286
287         ASSERT(PAGE_SHIFT >= sbp->sb_blocklog);
288         ASSERT(sbp->sb_blocklog >= BBSHIFT);
289
290 #if XFS_BIG_BLKNOS     /* Limited by ULONG_MAX of page cache index */
291         if (unlikely(
292             (sbp->sb_dblocks >> (PAGE_SHIFT - sbp->sb_blocklog)) > ULONG_MAX ||
293             (sbp->sb_rblocks >> (PAGE_SHIFT - sbp->sb_blocklog)) > ULONG_MAX)) {
294 #else                  /* Limited by UINT_MAX of sectors */
295         if (unlikely(
296             (sbp->sb_dblocks << (sbp->sb_blocklog - BBSHIFT)) > UINT_MAX ||
297             (sbp->sb_rblocks << (sbp->sb_blocklog - BBSHIFT)) > UINT_MAX)) {
298 #endif
299                 xfs_fs_mount_cmn_err(flags,
300                         "file system too large to be mounted on this system.");
301                 return XFS_ERROR(E2BIG);
302         }
303
304         if (unlikely(sbp->sb_inprogress)) {
305                 xfs_fs_mount_cmn_err(flags, "file system busy");
306                 return XFS_ERROR(EFSCORRUPTED);
307         }
308
309         /*
310          * Version 1 directory format has never worked on Linux.
311          */
312         if (unlikely(!XFS_SB_VERSION_HASDIRV2(sbp))) {
313                 xfs_fs_mount_cmn_err(flags,
314                         "file system using version 1 directory format");
315                 return XFS_ERROR(ENOSYS);
316         }
317
318         /*
319          * Until this is fixed only page-sized or smaller data blocks work.
320          */
321         if (unlikely(sbp->sb_blocksize > PAGE_SIZE)) {
322                 xfs_fs_mount_cmn_err(flags,
323                         "file system with blocksize %d bytes",
324                         sbp->sb_blocksize);
325                 xfs_fs_mount_cmn_err(flags,
326                         "only pagesize (%ld) or less will currently work.",
327                         PAGE_SIZE);
328                 return XFS_ERROR(ENOSYS);
329         }
330
331         return 0;
332 }
333
334 xfs_agnumber_t
335 xfs_initialize_perag(
336         bhv_vfs_t       *vfs,
337         xfs_mount_t     *mp,
338         xfs_agnumber_t  agcount)
339 {
340         xfs_agnumber_t  index, max_metadata;
341         xfs_perag_t     *pag;
342         xfs_agino_t     agino;
343         xfs_ino_t       ino;
344         xfs_sb_t        *sbp = &mp->m_sb;
345         xfs_ino_t       max_inum = XFS_MAXINUMBER_32;
346
347         /* Check to see if the filesystem can overflow 32 bit inodes */
348         agino = XFS_OFFBNO_TO_AGINO(mp, sbp->sb_agblocks - 1, 0);
349         ino = XFS_AGINO_TO_INO(mp, agcount - 1, agino);
350
351         /* Clear the mount flag if no inode can overflow 32 bits
352          * on this filesystem, or if specifically requested..
353          */
354         if ((vfs->vfs_flag & VFS_32BITINODES) && ino > max_inum) {
355                 mp->m_flags |= XFS_MOUNT_32BITINODES;
356         } else {
357                 mp->m_flags &= ~XFS_MOUNT_32BITINODES;
358         }
359
360         /* If we can overflow then setup the ag headers accordingly */
361         if (mp->m_flags & XFS_MOUNT_32BITINODES) {
362                 /* Calculate how much should be reserved for inodes to
363                  * meet the max inode percentage.
364                  */
365                 if (mp->m_maxicount) {
366                         __uint64_t      icount;
367
368                         icount = sbp->sb_dblocks * sbp->sb_imax_pct;
369                         do_div(icount, 100);
370                         icount += sbp->sb_agblocks - 1;
371                         do_div(icount, sbp->sb_agblocks);
372                         max_metadata = icount;
373                 } else {
374                         max_metadata = agcount;
375                 }
376                 for (index = 0; index < agcount; index++) {
377                         ino = XFS_AGINO_TO_INO(mp, index, agino);
378                         if (ino > max_inum) {
379                                 index++;
380                                 break;
381                         }
382
383                         /* This ag is preferred for inodes */
384                         pag = &mp->m_perag[index];
385                         pag->pagi_inodeok = 1;
386                         if (index < max_metadata)
387                                 pag->pagf_metadata = 1;
388                 }
389         } else {
390                 /* Setup default behavior for smaller filesystems */
391                 for (index = 0; index < agcount; index++) {
392                         pag = &mp->m_perag[index];
393                         pag->pagi_inodeok = 1;
394                 }
395         }
396         return index;
397 }
398
399 /*
400  * xfs_xlatesb
401  *
402  *     data       - on disk version of sb
403  *     sb         - a superblock
404  *     dir        - conversion direction: <0 - convert sb to buf
405  *                                        >0 - convert buf to sb
406  *     fields     - which fields to copy (bitmask)
407  */
408 void
409 xfs_xlatesb(
410         void            *data,
411         xfs_sb_t        *sb,
412         int             dir,
413         __int64_t       fields)
414 {
415         xfs_caddr_t     buf_ptr;
416         xfs_caddr_t     mem_ptr;
417         xfs_sb_field_t  f;
418         int             first;
419         int             size;
420
421         ASSERT(dir);
422         ASSERT(fields);
423
424         if (!fields)
425                 return;
426
427         buf_ptr = (xfs_caddr_t)data;
428         mem_ptr = (xfs_caddr_t)sb;
429
430         while (fields) {
431                 f = (xfs_sb_field_t)xfs_lowbit64((__uint64_t)fields);
432                 first = xfs_sb_info[f].offset;
433                 size = xfs_sb_info[f + 1].offset - first;
434
435                 ASSERT(xfs_sb_info[f].type == 0 || xfs_sb_info[f].type == 1);
436
437                 if (size == 1 || xfs_sb_info[f].type == 1) {
438                         if (dir > 0) {
439                                 memcpy(mem_ptr + first, buf_ptr + first, size);
440                         } else {
441                                 memcpy(buf_ptr + first, mem_ptr + first, size);
442                         }
443                 } else {
444                         switch (size) {
445                         case 2:
446                                 INT_XLATE(*(__uint16_t*)(buf_ptr+first),
447                                           *(__uint16_t*)(mem_ptr+first),
448                                           dir, ARCH_CONVERT);
449                                 break;
450                         case 4:
451                                 INT_XLATE(*(__uint32_t*)(buf_ptr+first),
452                                           *(__uint32_t*)(mem_ptr+first),
453                                           dir, ARCH_CONVERT);
454                                 break;
455                         case 8:
456                                 INT_XLATE(*(__uint64_t*)(buf_ptr+first),
457                                           *(__uint64_t*)(mem_ptr+first), dir, ARCH_CONVERT);
458                                 break;
459                         default:
460                                 ASSERT(0);
461                         }
462                 }
463
464                 fields &= ~(1LL << f);
465         }
466 }
467
468 /*
469  * xfs_readsb
470  *
471  * Does the initial read of the superblock.
472  */
473 int
474 xfs_readsb(xfs_mount_t *mp, int flags)
475 {
476         unsigned int    sector_size;
477         unsigned int    extra_flags;
478         xfs_buf_t       *bp;
479         xfs_sb_t        *sbp;
480         int             error;
481
482         ASSERT(mp->m_sb_bp == NULL);
483         ASSERT(mp->m_ddev_targp != NULL);
484
485         /*
486          * Allocate a (locked) buffer to hold the superblock.
487          * This will be kept around at all times to optimize
488          * access to the superblock.
489          */
490         sector_size = xfs_getsize_buftarg(mp->m_ddev_targp);
491         extra_flags = XFS_BUF_LOCK | XFS_BUF_MANAGE | XFS_BUF_MAPPED;
492
493         bp = xfs_buf_read_flags(mp->m_ddev_targp, XFS_SB_DADDR,
494                                 BTOBB(sector_size), extra_flags);
495         if (!bp || XFS_BUF_ISERROR(bp)) {
496                 xfs_fs_mount_cmn_err(flags, "SB read failed");
497                 error = bp ? XFS_BUF_GETERROR(bp) : ENOMEM;
498                 goto fail;
499         }
500         ASSERT(XFS_BUF_ISBUSY(bp));
501         ASSERT(XFS_BUF_VALUSEMA(bp) <= 0);
502
503         /*
504          * Initialize the mount structure from the superblock.
505          * But first do some basic consistency checking.
506          */
507         sbp = XFS_BUF_TO_SBP(bp);
508         xfs_xlatesb(XFS_BUF_PTR(bp), &(mp->m_sb), 1, XFS_SB_ALL_BITS);
509
510         error = xfs_mount_validate_sb(mp, &(mp->m_sb), flags);
511         if (error) {
512                 xfs_fs_mount_cmn_err(flags, "SB validate failed");
513                 goto fail;
514         }
515
516         /*
517          * We must be able to do sector-sized and sector-aligned IO.
518          */
519         if (sector_size > mp->m_sb.sb_sectsize) {
520                 xfs_fs_mount_cmn_err(flags,
521                         "device supports only %u byte sectors (not %u)",
522                         sector_size, mp->m_sb.sb_sectsize);
523                 error = ENOSYS;
524                 goto fail;
525         }
526
527         /*
528          * If device sector size is smaller than the superblock size,
529          * re-read the superblock so the buffer is correctly sized.
530          */
531         if (sector_size < mp->m_sb.sb_sectsize) {
532                 XFS_BUF_UNMANAGE(bp);
533                 xfs_buf_relse(bp);
534                 sector_size = mp->m_sb.sb_sectsize;
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 re-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         xfs_icsb_lock(mp);
547         xfs_icsb_balance_counter(mp, XFS_SBS_ICOUNT, 0, 0);
548         xfs_icsb_balance_counter(mp, XFS_SBS_IFREE, 0, 0);
549         xfs_icsb_balance_counter(mp, XFS_SBS_FDBLOCKS, 0, 0);
550         xfs_icsb_unlock(mp);
551
552         mp->m_sb_bp = bp;
553         xfs_buf_relse(bp);
554         ASSERT(XFS_BUF_VALUSEMA(bp) > 0);
555         return 0;
556
557  fail:
558         if (bp) {
559                 XFS_BUF_UNMANAGE(bp);
560                 xfs_buf_relse(bp);
561         }
562         return error;
563 }
564
565
566 /*
567  * xfs_mount_common
568  *
569  * Mount initialization code establishing various mount
570  * fields from the superblock associated with the given
571  * mount structure
572  */
573 STATIC void
574 xfs_mount_common(xfs_mount_t *mp, xfs_sb_t *sbp)
575 {
576         int     i;
577
578         mp->m_agfrotor = mp->m_agirotor = 0;
579         spinlock_init(&mp->m_agirotor_lock, "m_agirotor_lock");
580         mp->m_maxagi = mp->m_sb.sb_agcount;
581         mp->m_blkbit_log = sbp->sb_blocklog + XFS_NBBYLOG;
582         mp->m_blkbb_log = sbp->sb_blocklog - BBSHIFT;
583         mp->m_sectbb_log = sbp->sb_sectlog - BBSHIFT;
584         mp->m_agno_log = xfs_highbit32(sbp->sb_agcount - 1) + 1;
585         mp->m_agino_log = sbp->sb_inopblog + sbp->sb_agblklog;
586         mp->m_litino = sbp->sb_inodesize -
587                 ((uint)sizeof(xfs_dinode_core_t) + (uint)sizeof(xfs_agino_t));
588         mp->m_blockmask = sbp->sb_blocksize - 1;
589         mp->m_blockwsize = sbp->sb_blocksize >> XFS_WORDLOG;
590         mp->m_blockwmask = mp->m_blockwsize - 1;
591         INIT_LIST_HEAD(&mp->m_del_inodes);
592
593         /*
594          * Setup for attributes, in case they get created.
595          * This value is for inodes getting attributes for the first time,
596          * the per-inode value is for old attribute values.
597          */
598         ASSERT(sbp->sb_inodesize >= 256 && sbp->sb_inodesize <= 2048);
599         switch (sbp->sb_inodesize) {
600         case 256:
601                 mp->m_attroffset = XFS_LITINO(mp) -
602                                    XFS_BMDR_SPACE_CALC(MINABTPTRS);
603                 break;
604         case 512:
605         case 1024:
606         case 2048:
607                 mp->m_attroffset = XFS_BMDR_SPACE_CALC(6 * MINABTPTRS);
608                 break;
609         default:
610                 ASSERT(0);
611         }
612         ASSERT(mp->m_attroffset < XFS_LITINO(mp));
613
614         for (i = 0; i < 2; i++) {
615                 mp->m_alloc_mxr[i] = XFS_BTREE_BLOCK_MAXRECS(sbp->sb_blocksize,
616                         xfs_alloc, i == 0);
617                 mp->m_alloc_mnr[i] = XFS_BTREE_BLOCK_MINRECS(sbp->sb_blocksize,
618                         xfs_alloc, i == 0);
619         }
620         for (i = 0; i < 2; i++) {
621                 mp->m_bmap_dmxr[i] = XFS_BTREE_BLOCK_MAXRECS(sbp->sb_blocksize,
622                         xfs_bmbt, i == 0);
623                 mp->m_bmap_dmnr[i] = XFS_BTREE_BLOCK_MINRECS(sbp->sb_blocksize,
624                         xfs_bmbt, i == 0);
625         }
626         for (i = 0; i < 2; i++) {
627                 mp->m_inobt_mxr[i] = XFS_BTREE_BLOCK_MAXRECS(sbp->sb_blocksize,
628                         xfs_inobt, i == 0);
629                 mp->m_inobt_mnr[i] = XFS_BTREE_BLOCK_MINRECS(sbp->sb_blocksize,
630                         xfs_inobt, i == 0);
631         }
632
633         mp->m_bsize = XFS_FSB_TO_BB(mp, 1);
634         mp->m_ialloc_inos = (int)MAX((__uint16_t)XFS_INODES_PER_CHUNK,
635                                         sbp->sb_inopblock);
636         mp->m_ialloc_blks = mp->m_ialloc_inos >> sbp->sb_inopblog;
637 }
638 /*
639  * xfs_mountfs
640  *
641  * This function does the following on an initial mount of a file system:
642  *      - reads the superblock from disk and init the mount struct
643  *      - if we're a 32-bit kernel, do a size check on the superblock
644  *              so we don't mount terabyte filesystems
645  *      - init mount struct realtime fields
646  *      - allocate inode hash table for fs
647  *      - init directory manager
648  *      - perform recovery and init the log manager
649  */
650 int
651 xfs_mountfs(
652         bhv_vfs_t       *vfsp,
653         xfs_mount_t     *mp,
654         int             mfsi_flags)
655 {
656         xfs_buf_t       *bp;
657         xfs_sb_t        *sbp = &(mp->m_sb);
658         xfs_inode_t     *rip;
659         bhv_vnode_t     *rvp = NULL;
660         int             readio_log, writeio_log;
661         xfs_daddr_t     d;
662         __uint64_t      ret64;
663         __int64_t       update_flags;
664         uint            quotamount, quotaflags;
665         int             agno;
666         int             uuid_mounted = 0;
667         int             error = 0;
668
669         if (mp->m_sb_bp == NULL) {
670                 if ((error = xfs_readsb(mp, mfsi_flags))) {
671                         return error;
672                 }
673         }
674         xfs_mount_common(mp, sbp);
675
676         /*
677          * Check if sb_agblocks is aligned at stripe boundary
678          * If sb_agblocks is NOT aligned turn off m_dalign since
679          * allocator alignment is within an ag, therefore ag has
680          * to be aligned at stripe boundary.
681          */
682         update_flags = 0LL;
683         if (mp->m_dalign && !(mfsi_flags & XFS_MFSI_SECOND)) {
684                 /*
685                  * If stripe unit and stripe width are not multiples
686                  * of the fs blocksize turn off alignment.
687                  */
688                 if ((BBTOB(mp->m_dalign) & mp->m_blockmask) ||
689                     (BBTOB(mp->m_swidth) & mp->m_blockmask)) {
690                         if (mp->m_flags & XFS_MOUNT_RETERR) {
691                                 cmn_err(CE_WARN,
692                                         "XFS: alignment check 1 failed");
693                                 error = XFS_ERROR(EINVAL);
694                                 goto error1;
695                         }
696                         mp->m_dalign = mp->m_swidth = 0;
697                 } else {
698                         /*
699                          * Convert the stripe unit and width to FSBs.
700                          */
701                         mp->m_dalign = XFS_BB_TO_FSBT(mp, mp->m_dalign);
702                         if (mp->m_dalign && (sbp->sb_agblocks % mp->m_dalign)) {
703                                 if (mp->m_flags & XFS_MOUNT_RETERR) {
704                                         error = XFS_ERROR(EINVAL);
705                                         goto error1;
706                                 }
707                                 xfs_fs_cmn_err(CE_WARN, mp,
708 "stripe alignment turned off: sunit(%d)/swidth(%d) incompatible with agsize(%d)",
709                                         mp->m_dalign, mp->m_swidth,
710                                         sbp->sb_agblocks);
711
712                                 mp->m_dalign = 0;
713                                 mp->m_swidth = 0;
714                         } else if (mp->m_dalign) {
715                                 mp->m_swidth = XFS_BB_TO_FSBT(mp, mp->m_swidth);
716                         } else {
717                                 if (mp->m_flags & XFS_MOUNT_RETERR) {
718                                         xfs_fs_cmn_err(CE_WARN, mp,
719 "stripe alignment turned off: sunit(%d) less than bsize(%d)",
720                                                 mp->m_dalign,
721                                                 mp->m_blockmask +1);
722                                         error = XFS_ERROR(EINVAL);
723                                         goto error1;
724                                 }
725                                 mp->m_swidth = 0;
726                         }
727                 }
728
729                 /*
730                  * Update superblock with new values
731                  * and log changes
732                  */
733                 if (XFS_SB_VERSION_HASDALIGN(sbp)) {
734                         if (sbp->sb_unit != mp->m_dalign) {
735                                 sbp->sb_unit = mp->m_dalign;
736                                 update_flags |= XFS_SB_UNIT;
737                         }
738                         if (sbp->sb_width != mp->m_swidth) {
739                                 sbp->sb_width = mp->m_swidth;
740                                 update_flags |= XFS_SB_WIDTH;
741                         }
742                 }
743         } else if ((mp->m_flags & XFS_MOUNT_NOALIGN) != XFS_MOUNT_NOALIGN &&
744                     XFS_SB_VERSION_HASDALIGN(&mp->m_sb)) {
745                         mp->m_dalign = sbp->sb_unit;
746                         mp->m_swidth = sbp->sb_width;
747         }
748
749         xfs_alloc_compute_maxlevels(mp);
750         xfs_bmap_compute_maxlevels(mp, XFS_DATA_FORK);
751         xfs_bmap_compute_maxlevels(mp, XFS_ATTR_FORK);
752         xfs_ialloc_compute_maxlevels(mp);
753
754         if (sbp->sb_imax_pct) {
755                 __uint64_t      icount;
756
757                 /* Make sure the maximum inode count is a multiple of the
758                  * units we allocate inodes in.
759                  */
760
761                 icount = sbp->sb_dblocks * sbp->sb_imax_pct;
762                 do_div(icount, 100);
763                 do_div(icount, mp->m_ialloc_blks);
764                 mp->m_maxicount = (icount * mp->m_ialloc_blks)  <<
765                                    sbp->sb_inopblog;
766         } else
767                 mp->m_maxicount = 0;
768
769         mp->m_maxioffset = xfs_max_file_offset(sbp->sb_blocklog);
770
771         /*
772          * XFS uses the uuid from the superblock as the unique
773          * identifier for fsid.  We can not use the uuid from the volume
774          * since a single partition filesystem is identical to a single
775          * partition volume/filesystem.
776          */
777         if ((mfsi_flags & XFS_MFSI_SECOND) == 0 &&
778             (mp->m_flags & XFS_MOUNT_NOUUID) == 0) {
779                 if (xfs_uuid_mount(mp)) {
780                         error = XFS_ERROR(EINVAL);
781                         goto error1;
782                 }
783                 uuid_mounted=1;
784                 ret64 = uuid_hash64(&sbp->sb_uuid);
785                 memcpy(&vfsp->vfs_fsid, &ret64, sizeof(ret64));
786         }
787
788         /*
789          * Set the default minimum read and write sizes unless
790          * already specified in a mount option.
791          * We use smaller I/O sizes when the file system
792          * is being used for NFS service (wsync mount option).
793          */
794         if (!(mp->m_flags & XFS_MOUNT_DFLT_IOSIZE)) {
795                 if (mp->m_flags & XFS_MOUNT_WSYNC) {
796                         readio_log = XFS_WSYNC_READIO_LOG;
797                         writeio_log = XFS_WSYNC_WRITEIO_LOG;
798                 } else {
799                         readio_log = XFS_READIO_LOG_LARGE;
800                         writeio_log = XFS_WRITEIO_LOG_LARGE;
801                 }
802         } else {
803                 readio_log = mp->m_readio_log;
804                 writeio_log = mp->m_writeio_log;
805         }
806
807         /*
808          * Set the number of readahead buffers to use based on
809          * physical memory size.
810          */
811         if (xfs_physmem <= 4096)                /* <= 16MB */
812                 mp->m_nreadaheads = XFS_RW_NREADAHEAD_16MB;
813         else if (xfs_physmem <= 8192)   /* <= 32MB */
814                 mp->m_nreadaheads = XFS_RW_NREADAHEAD_32MB;
815         else
816                 mp->m_nreadaheads = XFS_RW_NREADAHEAD_K32;
817         if (sbp->sb_blocklog > readio_log) {
818                 mp->m_readio_log = sbp->sb_blocklog;
819         } else {
820                 mp->m_readio_log = readio_log;
821         }
822         mp->m_readio_blocks = 1 << (mp->m_readio_log - sbp->sb_blocklog);
823         if (sbp->sb_blocklog > writeio_log) {
824                 mp->m_writeio_log = sbp->sb_blocklog;
825         } else {
826                 mp->m_writeio_log = writeio_log;
827         }
828         mp->m_writeio_blocks = 1 << (mp->m_writeio_log - sbp->sb_blocklog);
829
830         /*
831          * Set the inode cluster size based on the physical memory
832          * size.  This may still be overridden by the file system
833          * block size if it is larger than the chosen cluster size.
834          */
835         if (xfs_physmem <= btoc(32 * 1024 * 1024)) { /* <= 32 MB */
836                 mp->m_inode_cluster_size = XFS_INODE_SMALL_CLUSTER_SIZE;
837         } else {
838                 mp->m_inode_cluster_size = XFS_INODE_BIG_CLUSTER_SIZE;
839         }
840         /*
841          * Set whether we're using inode alignment.
842          */
843         if (XFS_SB_VERSION_HASALIGN(&mp->m_sb) &&
844             mp->m_sb.sb_inoalignmt >=
845             XFS_B_TO_FSBT(mp, mp->m_inode_cluster_size))
846                 mp->m_inoalign_mask = mp->m_sb.sb_inoalignmt - 1;
847         else
848                 mp->m_inoalign_mask = 0;
849         /*
850          * If we are using stripe alignment, check whether
851          * the stripe unit is a multiple of the inode alignment
852          */
853         if (mp->m_dalign && mp->m_inoalign_mask &&
854             !(mp->m_dalign & mp->m_inoalign_mask))
855                 mp->m_sinoalign = mp->m_dalign;
856         else
857                 mp->m_sinoalign = 0;
858         /*
859          * Check that the data (and log if separate) are an ok size.
860          */
861         d = (xfs_daddr_t)XFS_FSB_TO_BB(mp, mp->m_sb.sb_dblocks);
862         if (XFS_BB_TO_FSB(mp, d) != mp->m_sb.sb_dblocks) {
863                 cmn_err(CE_WARN, "XFS: size check 1 failed");
864                 error = XFS_ERROR(E2BIG);
865                 goto error1;
866         }
867         error = xfs_read_buf(mp, mp->m_ddev_targp,
868                              d - XFS_FSS_TO_BB(mp, 1),
869                              XFS_FSS_TO_BB(mp, 1), 0, &bp);
870         if (!error) {
871                 xfs_buf_relse(bp);
872         } else {
873                 cmn_err(CE_WARN, "XFS: size check 2 failed");
874                 if (error == ENOSPC) {
875                         error = XFS_ERROR(E2BIG);
876                 }
877                 goto error1;
878         }
879
880         if (((mfsi_flags & XFS_MFSI_CLIENT) == 0) &&
881             mp->m_logdev_targp != mp->m_ddev_targp) {
882                 d = (xfs_daddr_t)XFS_FSB_TO_BB(mp, mp->m_sb.sb_logblocks);
883                 if (XFS_BB_TO_FSB(mp, d) != mp->m_sb.sb_logblocks) {
884                         cmn_err(CE_WARN, "XFS: size check 3 failed");
885                         error = XFS_ERROR(E2BIG);
886                         goto error1;
887                 }
888                 error = xfs_read_buf(mp, mp->m_logdev_targp,
889                                      d - XFS_FSB_TO_BB(mp, 1),
890                                      XFS_FSB_TO_BB(mp, 1), 0, &bp);
891                 if (!error) {
892                         xfs_buf_relse(bp);
893                 } else {
894                         cmn_err(CE_WARN, "XFS: size check 3 failed");
895                         if (error == ENOSPC) {
896                                 error = XFS_ERROR(E2BIG);
897                         }
898                         goto error1;
899                 }
900         }
901
902         /*
903          * Initialize realtime fields in the mount structure
904          */
905         if ((error = xfs_rtmount_init(mp))) {
906                 cmn_err(CE_WARN, "XFS: RT mount failed");
907                 goto error1;
908         }
909
910         /*
911          * For client case we are done now
912          */
913         if (mfsi_flags & XFS_MFSI_CLIENT) {
914                 return 0;
915         }
916
917         /*
918          *  Copies the low order bits of the timestamp and the randomly
919          *  set "sequence" number out of a UUID.
920          */
921         uuid_getnodeuniq(&sbp->sb_uuid, mp->m_fixedfsid);
922
923         /*
924          *  The vfs structure needs to have a file system independent
925          *  way of checking for the invariant file system ID.  Since it
926          *  can't look at mount structures it has a pointer to the data
927          *  in the mount structure.
928          *
929          *  File systems that don't support user level file handles (i.e.
930          *  all of them except for XFS) will leave vfs_altfsid as NULL.
931          */
932         vfsp->vfs_altfsid = (xfs_fsid_t *)mp->m_fixedfsid;
933         mp->m_dmevmask = 0;     /* not persistent; set after each mount */
934
935         xfs_dir_mount(mp);
936
937         /*
938          * Initialize the attribute manager's entries.
939          */
940         mp->m_attr_magicpct = (mp->m_sb.sb_blocksize * 37) / 100;
941
942         /*
943          * Initialize the precomputed transaction reservations values.
944          */
945         xfs_trans_init(mp);
946
947         /*
948          * Allocate and initialize the inode hash table for this
949          * file system.
950          */
951         xfs_ihash_init(mp);
952         xfs_chash_init(mp);
953
954         /*
955          * Allocate and initialize the per-ag data.
956          */
957         init_rwsem(&mp->m_peraglock);
958         mp->m_perag =
959                 kmem_zalloc(sbp->sb_agcount * sizeof(xfs_perag_t), KM_SLEEP);
960
961         mp->m_maxagi = xfs_initialize_perag(vfsp, mp, sbp->sb_agcount);
962
963         /*
964          * log's mount-time initialization. Perform 1st part recovery if needed
965          */
966         if (likely(sbp->sb_logblocks > 0)) {    /* check for volume case */
967                 error = xfs_log_mount(mp, mp->m_logdev_targp,
968                                       XFS_FSB_TO_DADDR(mp, sbp->sb_logstart),
969                                       XFS_FSB_TO_BB(mp, sbp->sb_logblocks));
970                 if (error) {
971                         cmn_err(CE_WARN, "XFS: log mount failed");
972                         goto error2;
973                 }
974         } else {        /* No log has been defined */
975                 cmn_err(CE_WARN, "XFS: no log defined");
976                 XFS_ERROR_REPORT("xfs_mountfs_int(1)", XFS_ERRLEVEL_LOW, mp);
977                 error = XFS_ERROR(EFSCORRUPTED);
978                 goto error2;
979         }
980
981         /*
982          * Get and sanity-check the root inode.
983          * Save the pointer to it in the mount structure.
984          */
985         error = xfs_iget(mp, NULL, sbp->sb_rootino, 0, XFS_ILOCK_EXCL, &rip, 0);
986         if (error) {
987                 cmn_err(CE_WARN, "XFS: failed to read root inode");
988                 goto error3;
989         }
990
991         ASSERT(rip != NULL);
992         rvp = XFS_ITOV(rip);
993
994         if (unlikely((rip->i_d.di_mode & S_IFMT) != S_IFDIR)) {
995                 cmn_err(CE_WARN, "XFS: corrupted root inode");
996                 cmn_err(CE_WARN, "Device %s - root %llu is not a directory",
997                         XFS_BUFTARG_NAME(mp->m_ddev_targp),
998                         (unsigned long long)rip->i_ino);
999                 xfs_iunlock(rip, XFS_ILOCK_EXCL);
1000                 XFS_ERROR_REPORT("xfs_mountfs_int(2)", XFS_ERRLEVEL_LOW,
1001                                  mp);
1002                 error = XFS_ERROR(EFSCORRUPTED);
1003                 goto error4;
1004         }
1005         mp->m_rootip = rip;     /* save it */
1006
1007         xfs_iunlock(rip, XFS_ILOCK_EXCL);
1008
1009         /*
1010          * Initialize realtime inode pointers in the mount structure
1011          */
1012         if ((error = xfs_rtmount_inodes(mp))) {
1013                 /*
1014                  * Free up the root inode.
1015                  */
1016                 cmn_err(CE_WARN, "XFS: failed to read RT inodes");
1017                 goto error4;
1018         }
1019
1020         /*
1021          * If fs is not mounted readonly, then update the superblock
1022          * unit and width changes.
1023          */
1024         if (update_flags && !(vfsp->vfs_flag & VFS_RDONLY))
1025                 xfs_mount_log_sbunit(mp, update_flags);
1026
1027         /*
1028          * Initialise the XFS quota management subsystem for this mount
1029          */
1030         if ((error = XFS_QM_INIT(mp, &quotamount, &quotaflags)))
1031                 goto error4;
1032
1033         /*
1034          * Finish recovering the file system.  This part needed to be
1035          * delayed until after the root and real-time bitmap inodes
1036          * were consistently read in.
1037          */
1038         error = xfs_log_mount_finish(mp, mfsi_flags);
1039         if (error) {
1040                 cmn_err(CE_WARN, "XFS: log mount finish failed");
1041                 goto error4;
1042         }
1043
1044         /*
1045          * Complete the quota initialisation, post-log-replay component.
1046          */
1047         if ((error = XFS_QM_MOUNT(mp, quotamount, quotaflags, mfsi_flags)))
1048                 goto error4;
1049
1050         return 0;
1051
1052  error4:
1053         /*
1054          * Free up the root inode.
1055          */
1056         VN_RELE(rvp);
1057  error3:
1058         xfs_log_unmount_dealloc(mp);
1059  error2:
1060         xfs_ihash_free(mp);
1061         xfs_chash_free(mp);
1062         for (agno = 0; agno < sbp->sb_agcount; agno++)
1063                 if (mp->m_perag[agno].pagb_list)
1064                         kmem_free(mp->m_perag[agno].pagb_list,
1065                           sizeof(xfs_perag_busy_t) * XFS_PAGB_NUM_SLOTS);
1066         kmem_free(mp->m_perag, sbp->sb_agcount * sizeof(xfs_perag_t));
1067         mp->m_perag = NULL;
1068         /* FALLTHROUGH */
1069  error1:
1070         if (uuid_mounted)
1071                 xfs_uuid_unmount(mp);
1072         xfs_freesb(mp);
1073         return error;
1074 }
1075
1076 /*
1077  * xfs_unmountfs
1078  *
1079  * This flushes out the inodes,dquots and the superblock, unmounts the
1080  * log and makes sure that incore structures are freed.
1081  */
1082 int
1083 xfs_unmountfs(xfs_mount_t *mp, struct cred *cr)
1084 {
1085         struct bhv_vfs  *vfsp = XFS_MTOVFS(mp);
1086 #if defined(DEBUG) || defined(INDUCE_IO_ERROR)
1087         int64_t         fsid;
1088 #endif
1089
1090         xfs_iflush_all(mp);
1091
1092         XFS_QM_DQPURGEALL(mp, XFS_QMOPT_QUOTALL | XFS_QMOPT_UMOUNTING);
1093
1094         /*
1095          * Flush out the log synchronously so that we know for sure
1096          * that nothing is pinned.  This is important because bflush()
1097          * will skip pinned buffers.
1098          */
1099         xfs_log_force(mp, (xfs_lsn_t)0, XFS_LOG_FORCE | XFS_LOG_SYNC);
1100
1101         xfs_binval(mp->m_ddev_targp);
1102         if (mp->m_rtdev_targp) {
1103                 xfs_binval(mp->m_rtdev_targp);
1104         }
1105
1106         xfs_unmountfs_writesb(mp);
1107
1108         xfs_unmountfs_wait(mp);                 /* wait for async bufs */
1109
1110         xfs_log_unmount(mp);                    /* Done! No more fs ops. */
1111
1112         xfs_freesb(mp);
1113
1114         /*
1115          * All inodes from this mount point should be freed.
1116          */
1117         ASSERT(mp->m_inodes == NULL);
1118
1119         xfs_unmountfs_close(mp, cr);
1120         if ((mp->m_flags & XFS_MOUNT_NOUUID) == 0)
1121                 xfs_uuid_unmount(mp);
1122
1123 #if defined(DEBUG) || defined(INDUCE_IO_ERROR)
1124         /*
1125          * clear all error tags on this filesystem
1126          */
1127         memcpy(&fsid, &vfsp->vfs_fsid, sizeof(int64_t));
1128         xfs_errortag_clearall_umount(fsid, mp->m_fsname, 0);
1129 #endif
1130         XFS_IODONE(vfsp);
1131         xfs_mount_free(mp, 1);
1132         return 0;
1133 }
1134
1135 void
1136 xfs_unmountfs_close(xfs_mount_t *mp, struct cred *cr)
1137 {
1138         if (mp->m_logdev_targp != mp->m_ddev_targp)
1139                 xfs_free_buftarg(mp->m_logdev_targp, 1);
1140         if (mp->m_rtdev_targp)
1141                 xfs_free_buftarg(mp->m_rtdev_targp, 1);
1142         xfs_free_buftarg(mp->m_ddev_targp, 0);
1143 }
1144
1145 STATIC void
1146 xfs_unmountfs_wait(xfs_mount_t *mp)
1147 {
1148         if (mp->m_logdev_targp != mp->m_ddev_targp)
1149                 xfs_wait_buftarg(mp->m_logdev_targp);
1150         if (mp->m_rtdev_targp)
1151                 xfs_wait_buftarg(mp->m_rtdev_targp);
1152         xfs_wait_buftarg(mp->m_ddev_targp);
1153 }
1154
1155 int
1156 xfs_unmountfs_writesb(xfs_mount_t *mp)
1157 {
1158         xfs_buf_t       *sbp;
1159         xfs_sb_t        *sb;
1160         int             error = 0;
1161
1162         /*
1163          * skip superblock write if fs is read-only, or
1164          * if we are doing a forced umount.
1165          */
1166         sbp = xfs_getsb(mp, 0);
1167         if (!(XFS_MTOVFS(mp)->vfs_flag & VFS_RDONLY ||
1168                 XFS_FORCED_SHUTDOWN(mp))) {
1169
1170                 xfs_icsb_sync_counters(mp);
1171
1172                 /*
1173                  * mark shared-readonly if desired
1174                  */
1175                 sb = XFS_BUF_TO_SBP(sbp);
1176                 if (mp->m_mk_sharedro) {
1177                         if (!(sb->sb_flags & XFS_SBF_READONLY))
1178                                 sb->sb_flags |= XFS_SBF_READONLY;
1179                         if (!XFS_SB_VERSION_HASSHARED(sb))
1180                                 XFS_SB_VERSION_ADDSHARED(sb);
1181                         xfs_fs_cmn_err(CE_NOTE, mp,
1182                                 "Unmounting, marking shared read-only");
1183                 }
1184                 XFS_BUF_UNDONE(sbp);
1185                 XFS_BUF_UNREAD(sbp);
1186                 XFS_BUF_UNDELAYWRITE(sbp);
1187                 XFS_BUF_WRITE(sbp);
1188                 XFS_BUF_UNASYNC(sbp);
1189                 ASSERT(XFS_BUF_TARGET(sbp) == mp->m_ddev_targp);
1190                 xfsbdstrat(mp, sbp);
1191                 /* Nevermind errors we might get here. */
1192                 error = xfs_iowait(sbp);
1193                 if (error)
1194                         xfs_ioerror_alert("xfs_unmountfs_writesb",
1195                                           mp, sbp, XFS_BUF_ADDR(sbp));
1196                 if (error && mp->m_mk_sharedro)
1197                         xfs_fs_cmn_err(CE_ALERT, mp, "Superblock write error detected while unmounting.  Filesystem may not be marked shared readonly");
1198         }
1199         xfs_buf_relse(sbp);
1200         return error;
1201 }
1202
1203 /*
1204  * xfs_mod_sb() can be used to copy arbitrary changes to the
1205  * in-core superblock into the superblock buffer to be logged.
1206  * It does not provide the higher level of locking that is
1207  * needed to protect the in-core superblock from concurrent
1208  * access.
1209  */
1210 void
1211 xfs_mod_sb(xfs_trans_t *tp, __int64_t fields)
1212 {
1213         xfs_buf_t       *bp;
1214         int             first;
1215         int             last;
1216         xfs_mount_t     *mp;
1217         xfs_sb_t        *sbp;
1218         xfs_sb_field_t  f;
1219
1220         ASSERT(fields);
1221         if (!fields)
1222                 return;
1223         mp = tp->t_mountp;
1224         bp = xfs_trans_getsb(tp, mp, 0);
1225         sbp = XFS_BUF_TO_SBP(bp);
1226         first = sizeof(xfs_sb_t);
1227         last = 0;
1228
1229         /* translate/copy */
1230
1231         xfs_xlatesb(XFS_BUF_PTR(bp), &(mp->m_sb), -1, fields);
1232
1233         /* find modified range */
1234
1235         f = (xfs_sb_field_t)xfs_lowbit64((__uint64_t)fields);
1236         ASSERT((1LL << f) & XFS_SB_MOD_BITS);
1237         first = xfs_sb_info[f].offset;
1238
1239         f = (xfs_sb_field_t)xfs_highbit64((__uint64_t)fields);
1240         ASSERT((1LL << f) & XFS_SB_MOD_BITS);
1241         last = xfs_sb_info[f + 1].offset - 1;
1242
1243         xfs_trans_log_buf(tp, bp, first, last);
1244 }
1245
1246
1247 /*
1248  * xfs_mod_incore_sb_unlocked() is a utility routine common used to apply
1249  * a delta to a specified field in the in-core superblock.  Simply
1250  * switch on the field indicated and apply the delta to that field.
1251  * Fields are not allowed to dip below zero, so if the delta would
1252  * do this do not apply it and return EINVAL.
1253  *
1254  * The SB_LOCK must be held when this routine is called.
1255  */
1256 int
1257 xfs_mod_incore_sb_unlocked(
1258         xfs_mount_t     *mp,
1259         xfs_sb_field_t  field,
1260         int64_t         delta,
1261         int             rsvd)
1262 {
1263         int             scounter;       /* short counter for 32 bit fields */
1264         long long       lcounter;       /* long counter for 64 bit fields */
1265         long long       res_used, rem;
1266
1267         /*
1268          * With the in-core superblock spin lock held, switch
1269          * on the indicated field.  Apply the delta to the
1270          * proper field.  If the fields value would dip below
1271          * 0, then do not apply the delta and return EINVAL.
1272          */
1273         switch (field) {
1274         case XFS_SBS_ICOUNT:
1275                 lcounter = (long long)mp->m_sb.sb_icount;
1276                 lcounter += delta;
1277                 if (lcounter < 0) {
1278                         ASSERT(0);
1279                         return XFS_ERROR(EINVAL);
1280                 }
1281                 mp->m_sb.sb_icount = lcounter;
1282                 return 0;
1283         case XFS_SBS_IFREE:
1284                 lcounter = (long long)mp->m_sb.sb_ifree;
1285                 lcounter += delta;
1286                 if (lcounter < 0) {
1287                         ASSERT(0);
1288                         return XFS_ERROR(EINVAL);
1289                 }
1290                 mp->m_sb.sb_ifree = lcounter;
1291                 return 0;
1292         case XFS_SBS_FDBLOCKS:
1293                 lcounter = (long long)
1294                         mp->m_sb.sb_fdblocks - XFS_ALLOC_SET_ASIDE(mp);
1295                 res_used = (long long)(mp->m_resblks - mp->m_resblks_avail);
1296
1297                 if (delta > 0) {                /* Putting blocks back */
1298                         if (res_used > delta) {
1299                                 mp->m_resblks_avail += delta;
1300                         } else {
1301                                 rem = delta - res_used;
1302                                 mp->m_resblks_avail = mp->m_resblks;
1303                                 lcounter += rem;
1304                         }
1305                 } else {                                /* Taking blocks away */
1306
1307                         lcounter += delta;
1308
1309                 /*
1310                  * If were out of blocks, use any available reserved blocks if
1311                  * were allowed to.
1312                  */
1313
1314                         if (lcounter < 0) {
1315                                 if (rsvd) {
1316                                         lcounter = (long long)mp->m_resblks_avail + delta;
1317                                         if (lcounter < 0) {
1318                                                 return XFS_ERROR(ENOSPC);
1319                                         }
1320                                         mp->m_resblks_avail = lcounter;
1321                                         return 0;
1322                                 } else {        /* not reserved */
1323                                         return XFS_ERROR(ENOSPC);
1324                                 }
1325                         }
1326                 }
1327
1328                 mp->m_sb.sb_fdblocks = lcounter + XFS_ALLOC_SET_ASIDE(mp);
1329                 return 0;
1330         case XFS_SBS_FREXTENTS:
1331                 lcounter = (long long)mp->m_sb.sb_frextents;
1332                 lcounter += delta;
1333                 if (lcounter < 0) {
1334                         return XFS_ERROR(ENOSPC);
1335                 }
1336                 mp->m_sb.sb_frextents = lcounter;
1337                 return 0;
1338         case XFS_SBS_DBLOCKS:
1339                 lcounter = (long long)mp->m_sb.sb_dblocks;
1340                 lcounter += delta;
1341                 if (lcounter < 0) {
1342                         ASSERT(0);
1343                         return XFS_ERROR(EINVAL);
1344                 }
1345                 mp->m_sb.sb_dblocks = lcounter;
1346                 return 0;
1347         case XFS_SBS_AGCOUNT:
1348                 scounter = mp->m_sb.sb_agcount;
1349                 scounter += delta;
1350                 if (scounter < 0) {
1351                         ASSERT(0);
1352                         return XFS_ERROR(EINVAL);
1353                 }
1354                 mp->m_sb.sb_agcount = scounter;
1355                 return 0;
1356         case XFS_SBS_IMAX_PCT:
1357                 scounter = mp->m_sb.sb_imax_pct;
1358                 scounter += delta;
1359                 if (scounter < 0) {
1360                         ASSERT(0);
1361                         return XFS_ERROR(EINVAL);
1362                 }
1363                 mp->m_sb.sb_imax_pct = scounter;
1364                 return 0;
1365         case XFS_SBS_REXTSIZE:
1366                 scounter = mp->m_sb.sb_rextsize;
1367                 scounter += delta;
1368                 if (scounter < 0) {
1369                         ASSERT(0);
1370                         return XFS_ERROR(EINVAL);
1371                 }
1372                 mp->m_sb.sb_rextsize = scounter;
1373                 return 0;
1374         case XFS_SBS_RBMBLOCKS:
1375                 scounter = mp->m_sb.sb_rbmblocks;
1376                 scounter += delta;
1377                 if (scounter < 0) {
1378                         ASSERT(0);
1379                         return XFS_ERROR(EINVAL);
1380                 }
1381                 mp->m_sb.sb_rbmblocks = scounter;
1382                 return 0;
1383         case XFS_SBS_RBLOCKS:
1384                 lcounter = (long long)mp->m_sb.sb_rblocks;
1385                 lcounter += delta;
1386                 if (lcounter < 0) {
1387                         ASSERT(0);
1388                         return XFS_ERROR(EINVAL);
1389                 }
1390                 mp->m_sb.sb_rblocks = lcounter;
1391                 return 0;
1392         case XFS_SBS_REXTENTS:
1393                 lcounter = (long long)mp->m_sb.sb_rextents;
1394                 lcounter += delta;
1395                 if (lcounter < 0) {
1396                         ASSERT(0);
1397                         return XFS_ERROR(EINVAL);
1398                 }
1399                 mp->m_sb.sb_rextents = lcounter;
1400                 return 0;
1401         case XFS_SBS_REXTSLOG:
1402                 scounter = mp->m_sb.sb_rextslog;
1403                 scounter += delta;
1404                 if (scounter < 0) {
1405                         ASSERT(0);
1406                         return XFS_ERROR(EINVAL);
1407                 }
1408                 mp->m_sb.sb_rextslog = scounter;
1409                 return 0;
1410         default:
1411                 ASSERT(0);
1412                 return XFS_ERROR(EINVAL);
1413         }
1414 }
1415
1416 /*
1417  * xfs_mod_incore_sb() is used to change a field in the in-core
1418  * superblock structure by the specified delta.  This modification
1419  * is protected by the SB_LOCK.  Just use the xfs_mod_incore_sb_unlocked()
1420  * routine to do the work.
1421  */
1422 int
1423 xfs_mod_incore_sb(
1424         xfs_mount_t     *mp,
1425         xfs_sb_field_t  field,
1426         int64_t         delta,
1427         int             rsvd)
1428 {
1429         unsigned long   s;
1430         int     status;
1431
1432         /* check for per-cpu counters */
1433         switch (field) {
1434 #ifdef HAVE_PERCPU_SB
1435         case XFS_SBS_ICOUNT:
1436         case XFS_SBS_IFREE:
1437         case XFS_SBS_FDBLOCKS:
1438                 if (!(mp->m_flags & XFS_MOUNT_NO_PERCPU_SB)) {
1439                         status = xfs_icsb_modify_counters(mp, field,
1440                                                         delta, rsvd);
1441                         break;
1442                 }
1443                 /* FALLTHROUGH */
1444 #endif
1445         default:
1446                 s = XFS_SB_LOCK(mp);
1447                 status = xfs_mod_incore_sb_unlocked(mp, field, delta, rsvd);
1448                 XFS_SB_UNLOCK(mp, s);
1449                 break;
1450         }
1451
1452         return status;
1453 }
1454
1455 /*
1456  * xfs_mod_incore_sb_batch() is used to change more than one field
1457  * in the in-core superblock structure at a time.  This modification
1458  * is protected by a lock internal to this module.  The fields and
1459  * changes to those fields are specified in the array of xfs_mod_sb
1460  * structures passed in.
1461  *
1462  * Either all of the specified deltas will be applied or none of
1463  * them will.  If any modified field dips below 0, then all modifications
1464  * will be backed out and EINVAL will be returned.
1465  */
1466 int
1467 xfs_mod_incore_sb_batch(xfs_mount_t *mp, xfs_mod_sb_t *msb, uint nmsb, int rsvd)
1468 {
1469         unsigned long   s;
1470         int             status=0;
1471         xfs_mod_sb_t    *msbp;
1472
1473         /*
1474          * Loop through the array of mod structures and apply each
1475          * individually.  If any fail, then back out all those
1476          * which have already been applied.  Do all of this within
1477          * the scope of the SB_LOCK so that all of the changes will
1478          * be atomic.
1479          */
1480         s = XFS_SB_LOCK(mp);
1481         msbp = &msb[0];
1482         for (msbp = &msbp[0]; msbp < (msb + nmsb); msbp++) {
1483                 /*
1484                  * Apply the delta at index n.  If it fails, break
1485                  * from the loop so we'll fall into the undo loop
1486                  * below.
1487                  */
1488                 switch (msbp->msb_field) {
1489 #ifdef HAVE_PERCPU_SB
1490                 case XFS_SBS_ICOUNT:
1491                 case XFS_SBS_IFREE:
1492                 case XFS_SBS_FDBLOCKS:
1493                         if (!(mp->m_flags & XFS_MOUNT_NO_PERCPU_SB)) {
1494                                 XFS_SB_UNLOCK(mp, s);
1495                                 status = xfs_icsb_modify_counters(mp,
1496                                                         msbp->msb_field,
1497                                                         msbp->msb_delta, rsvd);
1498                                 s = XFS_SB_LOCK(mp);
1499                                 break;
1500                         }
1501                         /* FALLTHROUGH */
1502 #endif
1503                 default:
1504                         status = xfs_mod_incore_sb_unlocked(mp,
1505                                                 msbp->msb_field,
1506                                                 msbp->msb_delta, rsvd);
1507                         break;
1508                 }
1509
1510                 if (status != 0) {
1511                         break;
1512                 }
1513         }
1514
1515         /*
1516          * If we didn't complete the loop above, then back out
1517          * any changes made to the superblock.  If you add code
1518          * between the loop above and here, make sure that you
1519          * preserve the value of status. Loop back until
1520          * we step below the beginning of the array.  Make sure
1521          * we don't touch anything back there.
1522          */
1523         if (status != 0) {
1524                 msbp--;
1525                 while (msbp >= msb) {
1526                         switch (msbp->msb_field) {
1527 #ifdef HAVE_PERCPU_SB
1528                         case XFS_SBS_ICOUNT:
1529                         case XFS_SBS_IFREE:
1530                         case XFS_SBS_FDBLOCKS:
1531                                 if (!(mp->m_flags & XFS_MOUNT_NO_PERCPU_SB)) {
1532                                         XFS_SB_UNLOCK(mp, s);
1533                                         status = xfs_icsb_modify_counters(mp,
1534                                                         msbp->msb_field,
1535                                                         -(msbp->msb_delta),
1536                                                         rsvd);
1537                                         s = XFS_SB_LOCK(mp);
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 #ifdef CONFIG_HOTPLUG_CPU
1717 /*
1718  * hot-plug CPU notifier support.
1719  *
1720  * We need a notifier per filesystem as we need to be able to identify
1721  * the filesystem to balance the counters out. This is achieved by
1722  * having a notifier block embedded in the xfs_mount_t and doing pointer
1723  * magic to get the mount pointer from the notifier block address.
1724  */
1725 STATIC int
1726 xfs_icsb_cpu_notify(
1727         struct notifier_block *nfb,
1728         unsigned long action,
1729         void *hcpu)
1730 {
1731         xfs_icsb_cnts_t *cntp;
1732         xfs_mount_t     *mp;
1733         int             s;
1734
1735         mp = (xfs_mount_t *)container_of(nfb, xfs_mount_t, m_icsb_notifier);
1736         cntp = (xfs_icsb_cnts_t *)
1737                         per_cpu_ptr(mp->m_sb_cnts, (unsigned long)hcpu);
1738         switch (action) {
1739         case CPU_UP_PREPARE:
1740                 /* Easy Case - initialize the area and locks, and
1741                  * then rebalance when online does everything else for us. */
1742                 memset(cntp, 0, sizeof(xfs_icsb_cnts_t));
1743                 break;
1744         case CPU_ONLINE:
1745                 xfs_icsb_lock(mp);
1746                 xfs_icsb_balance_counter(mp, XFS_SBS_ICOUNT, 0, 0);
1747                 xfs_icsb_balance_counter(mp, XFS_SBS_IFREE, 0, 0);
1748                 xfs_icsb_balance_counter(mp, XFS_SBS_FDBLOCKS, 0, 0);
1749                 xfs_icsb_unlock(mp);
1750                 break;
1751         case CPU_DEAD:
1752                 /* Disable all the counters, then fold the dead cpu's
1753                  * count into the total on the global superblock and
1754                  * re-enable the counters. */
1755                 xfs_icsb_lock(mp);
1756                 s = XFS_SB_LOCK(mp);
1757                 xfs_icsb_disable_counter(mp, XFS_SBS_ICOUNT);
1758                 xfs_icsb_disable_counter(mp, XFS_SBS_IFREE);
1759                 xfs_icsb_disable_counter(mp, XFS_SBS_FDBLOCKS);
1760
1761                 mp->m_sb.sb_icount += cntp->icsb_icount;
1762                 mp->m_sb.sb_ifree += cntp->icsb_ifree;
1763                 mp->m_sb.sb_fdblocks += cntp->icsb_fdblocks;
1764
1765                 memset(cntp, 0, sizeof(xfs_icsb_cnts_t));
1766
1767                 xfs_icsb_balance_counter(mp, XFS_SBS_ICOUNT,
1768                                          XFS_ICSB_SB_LOCKED, 0);
1769                 xfs_icsb_balance_counter(mp, XFS_SBS_IFREE,
1770                                          XFS_ICSB_SB_LOCKED, 0);
1771                 xfs_icsb_balance_counter(mp, XFS_SBS_FDBLOCKS,
1772                                          XFS_ICSB_SB_LOCKED, 0);
1773                 XFS_SB_UNLOCK(mp, s);
1774                 xfs_icsb_unlock(mp);
1775                 break;
1776         }
1777
1778         return NOTIFY_OK;
1779 }
1780 #endif /* CONFIG_HOTPLUG_CPU */
1781
1782 int
1783 xfs_icsb_init_counters(
1784         xfs_mount_t     *mp)
1785 {
1786         xfs_icsb_cnts_t *cntp;
1787         int             i;
1788
1789         mp->m_sb_cnts = alloc_percpu(xfs_icsb_cnts_t);
1790         if (mp->m_sb_cnts == NULL)
1791                 return -ENOMEM;
1792
1793 #ifdef CONFIG_HOTPLUG_CPU
1794         mp->m_icsb_notifier.notifier_call = xfs_icsb_cpu_notify;
1795         mp->m_icsb_notifier.priority = 0;
1796         register_hotcpu_notifier(&mp->m_icsb_notifier);
1797 #endif /* CONFIG_HOTPLUG_CPU */
1798
1799         for_each_online_cpu(i) {
1800                 cntp = (xfs_icsb_cnts_t *)per_cpu_ptr(mp->m_sb_cnts, i);
1801                 memset(cntp, 0, sizeof(xfs_icsb_cnts_t));
1802         }
1803
1804         mutex_init(&mp->m_icsb_mutex);
1805
1806         /*
1807          * start with all counters disabled so that the
1808          * initial balance kicks us off correctly
1809          */
1810         mp->m_icsb_counters = -1;
1811         return 0;
1812 }
1813
1814 STATIC void
1815 xfs_icsb_destroy_counters(
1816         xfs_mount_t     *mp)
1817 {
1818         if (mp->m_sb_cnts) {
1819                 unregister_hotcpu_notifier(&mp->m_icsb_notifier);
1820                 free_percpu(mp->m_sb_cnts);
1821         }
1822         mutex_destroy(&mp->m_icsb_mutex);
1823 }
1824
1825 STATIC_INLINE void
1826 xfs_icsb_lock_cntr(
1827         xfs_icsb_cnts_t *icsbp)
1828 {
1829         while (test_and_set_bit(XFS_ICSB_FLAG_LOCK, &icsbp->icsb_flags)) {
1830                 ndelay(1000);
1831         }
1832 }
1833
1834 STATIC_INLINE void
1835 xfs_icsb_unlock_cntr(
1836         xfs_icsb_cnts_t *icsbp)
1837 {
1838         clear_bit(XFS_ICSB_FLAG_LOCK, &icsbp->icsb_flags);
1839 }
1840
1841
1842 STATIC_INLINE void
1843 xfs_icsb_lock_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_lock_cntr(cntp);
1852         }
1853 }
1854
1855 STATIC_INLINE void
1856 xfs_icsb_unlock_all_counters(
1857         xfs_mount_t     *mp)
1858 {
1859         xfs_icsb_cnts_t *cntp;
1860         int             i;
1861
1862         for_each_online_cpu(i) {
1863                 cntp = (xfs_icsb_cnts_t *)per_cpu_ptr(mp->m_sb_cnts, i);
1864                 xfs_icsb_unlock_cntr(cntp);
1865         }
1866 }
1867
1868 STATIC void
1869 xfs_icsb_count(
1870         xfs_mount_t     *mp,
1871         xfs_icsb_cnts_t *cnt,
1872         int             flags)
1873 {
1874         xfs_icsb_cnts_t *cntp;
1875         int             i;
1876
1877         memset(cnt, 0, sizeof(xfs_icsb_cnts_t));
1878
1879         if (!(flags & XFS_ICSB_LAZY_COUNT))
1880                 xfs_icsb_lock_all_counters(mp);
1881
1882         for_each_online_cpu(i) {
1883                 cntp = (xfs_icsb_cnts_t *)per_cpu_ptr(mp->m_sb_cnts, i);
1884                 cnt->icsb_icount += cntp->icsb_icount;
1885                 cnt->icsb_ifree += cntp->icsb_ifree;
1886                 cnt->icsb_fdblocks += cntp->icsb_fdblocks;
1887         }
1888
1889         if (!(flags & XFS_ICSB_LAZY_COUNT))
1890                 xfs_icsb_unlock_all_counters(mp);
1891 }
1892
1893 STATIC int
1894 xfs_icsb_counter_disabled(
1895         xfs_mount_t     *mp,
1896         xfs_sb_field_t  field)
1897 {
1898         ASSERT((field >= XFS_SBS_ICOUNT) && (field <= XFS_SBS_FDBLOCKS));
1899         return test_bit(field, &mp->m_icsb_counters);
1900 }
1901
1902 STATIC int
1903 xfs_icsb_disable_counter(
1904         xfs_mount_t     *mp,
1905         xfs_sb_field_t  field)
1906 {
1907         xfs_icsb_cnts_t cnt;
1908
1909         ASSERT((field >= XFS_SBS_ICOUNT) && (field <= XFS_SBS_FDBLOCKS));
1910
1911         /*
1912          * If we are already disabled, then there is nothing to do
1913          * here. We check before locking all the counters to avoid
1914          * the expensive lock operation when being called in the
1915          * slow path and the counter is already disabled. This is
1916          * safe because the only time we set or clear this state is under
1917          * the m_icsb_mutex.
1918          */
1919         if (xfs_icsb_counter_disabled(mp, field))
1920                 return 0;
1921
1922         xfs_icsb_lock_all_counters(mp);
1923         if (!test_and_set_bit(field, &mp->m_icsb_counters)) {
1924                 /* drain back to superblock */
1925
1926                 xfs_icsb_count(mp, &cnt, XFS_ICSB_SB_LOCKED|XFS_ICSB_LAZY_COUNT);
1927                 switch(field) {
1928                 case XFS_SBS_ICOUNT:
1929                         mp->m_sb.sb_icount = cnt.icsb_icount;
1930                         break;
1931                 case XFS_SBS_IFREE:
1932                         mp->m_sb.sb_ifree = cnt.icsb_ifree;
1933                         break;
1934                 case XFS_SBS_FDBLOCKS:
1935                         mp->m_sb.sb_fdblocks = cnt.icsb_fdblocks;
1936                         break;
1937                 default:
1938                         BUG();
1939                 }
1940         }
1941
1942         xfs_icsb_unlock_all_counters(mp);
1943
1944         return 0;
1945 }
1946
1947 STATIC void
1948 xfs_icsb_enable_counter(
1949         xfs_mount_t     *mp,
1950         xfs_sb_field_t  field,
1951         uint64_t        count,
1952         uint64_t        resid)
1953 {
1954         xfs_icsb_cnts_t *cntp;
1955         int             i;
1956
1957         ASSERT((field >= XFS_SBS_ICOUNT) && (field <= XFS_SBS_FDBLOCKS));
1958
1959         xfs_icsb_lock_all_counters(mp);
1960         for_each_online_cpu(i) {
1961                 cntp = per_cpu_ptr(mp->m_sb_cnts, i);
1962                 switch (field) {
1963                 case XFS_SBS_ICOUNT:
1964                         cntp->icsb_icount = count + resid;
1965                         break;
1966                 case XFS_SBS_IFREE:
1967                         cntp->icsb_ifree = count + resid;
1968                         break;
1969                 case XFS_SBS_FDBLOCKS:
1970                         cntp->icsb_fdblocks = count + resid;
1971                         break;
1972                 default:
1973                         BUG();
1974                         break;
1975                 }
1976                 resid = 0;
1977         }
1978         clear_bit(field, &mp->m_icsb_counters);
1979         xfs_icsb_unlock_all_counters(mp);
1980 }
1981
1982 void
1983 xfs_icsb_sync_counters_flags(
1984         xfs_mount_t     *mp,
1985         int             flags)
1986 {
1987         xfs_icsb_cnts_t cnt;
1988         int             s;
1989
1990         /* Pass 1: lock all counters */
1991         if ((flags & XFS_ICSB_SB_LOCKED) == 0)
1992                 s = XFS_SB_LOCK(mp);
1993
1994         xfs_icsb_count(mp, &cnt, flags);
1995
1996         /* Step 3: update mp->m_sb fields */
1997         if (!xfs_icsb_counter_disabled(mp, XFS_SBS_ICOUNT))
1998                 mp->m_sb.sb_icount = cnt.icsb_icount;
1999         if (!xfs_icsb_counter_disabled(mp, XFS_SBS_IFREE))
2000                 mp->m_sb.sb_ifree = cnt.icsb_ifree;
2001         if (!xfs_icsb_counter_disabled(mp, XFS_SBS_FDBLOCKS))
2002                 mp->m_sb.sb_fdblocks = cnt.icsb_fdblocks;
2003
2004         if ((flags & XFS_ICSB_SB_LOCKED) == 0)
2005                 XFS_SB_UNLOCK(mp, s);
2006 }
2007
2008 /*
2009  * Accurate update of per-cpu counters to incore superblock
2010  */
2011 STATIC void
2012 xfs_icsb_sync_counters(
2013         xfs_mount_t     *mp)
2014 {
2015         xfs_icsb_sync_counters_flags(mp, 0);
2016 }
2017
2018 /*
2019  * Balance and enable/disable counters as necessary.
2020  *
2021  * Thresholds for re-enabling counters are somewhat magic.  inode counts are
2022  * chosen to be the same number as single on disk allocation chunk per CPU, and
2023  * free blocks is something far enough zero that we aren't going thrash when we
2024  * get near ENOSPC. We also need to supply a minimum we require per cpu to
2025  * prevent looping endlessly when xfs_alloc_space asks for more than will
2026  * be distributed to a single CPU but each CPU has enough blocks to be
2027  * reenabled.
2028  *
2029  * Note that we can be called when counters are already disabled.
2030  * xfs_icsb_disable_counter() optimises the counter locking in this case to
2031  * prevent locking every per-cpu counter needlessly.
2032  */
2033
2034 #define XFS_ICSB_INO_CNTR_REENABLE      (uint64_t)64
2035 #define XFS_ICSB_FDBLK_CNTR_REENABLE(mp) \
2036                 (uint64_t)(512 + XFS_ALLOC_SET_ASIDE(mp))
2037 STATIC void
2038 xfs_icsb_balance_counter(
2039         xfs_mount_t     *mp,
2040         xfs_sb_field_t  field,
2041         int             flags,
2042         int             min_per_cpu)
2043 {
2044         uint64_t        count, resid;
2045         int             weight = num_online_cpus();
2046         int             s;
2047         uint64_t        min = (uint64_t)min_per_cpu;
2048
2049         if (!(flags & XFS_ICSB_SB_LOCKED))
2050                 s = XFS_SB_LOCK(mp);
2051
2052         /* disable counter and sync counter */
2053         xfs_icsb_disable_counter(mp, field);
2054
2055         /* update counters  - first CPU gets residual*/
2056         switch (field) {
2057         case XFS_SBS_ICOUNT:
2058                 count = mp->m_sb.sb_icount;
2059                 resid = do_div(count, weight);
2060                 if (count < max(min, XFS_ICSB_INO_CNTR_REENABLE))
2061                         goto out;
2062                 break;
2063         case XFS_SBS_IFREE:
2064                 count = mp->m_sb.sb_ifree;
2065                 resid = do_div(count, weight);
2066                 if (count < max(min, XFS_ICSB_INO_CNTR_REENABLE))
2067                         goto out;
2068                 break;
2069         case XFS_SBS_FDBLOCKS:
2070                 count = mp->m_sb.sb_fdblocks;
2071                 resid = do_div(count, weight);
2072                 if (count < max(min, XFS_ICSB_FDBLK_CNTR_REENABLE(mp)))
2073                         goto out;
2074                 break;
2075         default:
2076                 BUG();
2077                 count = resid = 0;      /* quiet, gcc */
2078                 break;
2079         }
2080
2081         xfs_icsb_enable_counter(mp, field, count, resid);
2082 out:
2083         if (!(flags & XFS_ICSB_SB_LOCKED))
2084                 XFS_SB_UNLOCK(mp, s);
2085 }
2086
2087 int
2088 xfs_icsb_modify_counters(
2089         xfs_mount_t     *mp,
2090         xfs_sb_field_t  field,
2091         int64_t         delta,
2092         int             rsvd)
2093 {
2094         xfs_icsb_cnts_t *icsbp;
2095         long long       lcounter;       /* long counter for 64 bit fields */
2096         int             cpu, ret = 0, s;
2097
2098         might_sleep();
2099 again:
2100         cpu = get_cpu();
2101         icsbp = (xfs_icsb_cnts_t *)per_cpu_ptr(mp->m_sb_cnts, cpu);
2102
2103         /*
2104          * if the counter is disabled, go to slow path
2105          */
2106         if (unlikely(xfs_icsb_counter_disabled(mp, field)))
2107                 goto slow_path;
2108         xfs_icsb_lock_cntr(icsbp);
2109         if (unlikely(xfs_icsb_counter_disabled(mp, field))) {
2110                 xfs_icsb_unlock_cntr(icsbp);
2111                 goto slow_path;
2112         }
2113
2114         switch (field) {
2115         case XFS_SBS_ICOUNT:
2116                 lcounter = icsbp->icsb_icount;
2117                 lcounter += delta;
2118                 if (unlikely(lcounter < 0))
2119                         goto balance_counter;
2120                 icsbp->icsb_icount = lcounter;
2121                 break;
2122
2123         case XFS_SBS_IFREE:
2124                 lcounter = icsbp->icsb_ifree;
2125                 lcounter += delta;
2126                 if (unlikely(lcounter < 0))
2127                         goto balance_counter;
2128                 icsbp->icsb_ifree = lcounter;
2129                 break;
2130
2131         case XFS_SBS_FDBLOCKS:
2132                 BUG_ON((mp->m_resblks - mp->m_resblks_avail) != 0);
2133
2134                 lcounter = icsbp->icsb_fdblocks - XFS_ALLOC_SET_ASIDE(mp);
2135                 lcounter += delta;
2136                 if (unlikely(lcounter < 0))
2137                         goto balance_counter;
2138                 icsbp->icsb_fdblocks = lcounter + XFS_ALLOC_SET_ASIDE(mp);
2139                 break;
2140         default:
2141                 BUG();
2142                 break;
2143         }
2144         xfs_icsb_unlock_cntr(icsbp);
2145         put_cpu();
2146         return 0;
2147
2148 slow_path:
2149         put_cpu();
2150
2151         /*
2152          * serialise with a mutex so we don't burn lots of cpu on
2153          * the superblock lock. We still need to hold the superblock
2154          * lock, however, when we modify the global structures.
2155          */
2156         xfs_icsb_lock(mp);
2157
2158         /*
2159          * Now running atomically.
2160          *
2161          * If the counter is enabled, someone has beaten us to rebalancing.
2162          * Drop the lock and try again in the fast path....
2163          */
2164         if (!(xfs_icsb_counter_disabled(mp, field))) {
2165                 xfs_icsb_unlock(mp);
2166                 goto again;
2167         }
2168
2169         /*
2170          * The counter is currently disabled. Because we are
2171          * running atomically here, we know a rebalance cannot
2172          * be in progress. Hence we can go straight to operating
2173          * on the global superblock. We do not call xfs_mod_incore_sb()
2174          * here even though we need to get the SB_LOCK. Doing so
2175          * will cause us to re-enter this function and deadlock.
2176          * Hence we get the SB_LOCK ourselves and then call
2177          * xfs_mod_incore_sb_unlocked() as the unlocked path operates
2178          * directly on the global counters.
2179          */
2180         s = XFS_SB_LOCK(mp);
2181         ret = xfs_mod_incore_sb_unlocked(mp, field, delta, rsvd);
2182         XFS_SB_UNLOCK(mp, s);
2183
2184         /*
2185          * Now that we've modified the global superblock, we
2186          * may be able to re-enable the distributed counters
2187          * (e.g. lots of space just got freed). After that
2188          * we are done.
2189          */
2190         if (ret != ENOSPC)
2191                 xfs_icsb_balance_counter(mp, field, 0, 0);
2192         xfs_icsb_unlock(mp);
2193         return ret;
2194
2195 balance_counter:
2196         xfs_icsb_unlock_cntr(icsbp);
2197         put_cpu();
2198
2199         /*
2200          * We may have multiple threads here if multiple per-cpu
2201          * counters run dry at the same time. This will mean we can
2202          * do more balances than strictly necessary but it is not
2203          * the common slowpath case.
2204          */
2205         xfs_icsb_lock(mp);
2206
2207         /*
2208          * running atomically.
2209          *
2210          * This will leave the counter in the correct state for future
2211          * accesses. After the rebalance, we simply try again and our retry
2212          * will either succeed through the fast path or slow path without
2213          * another balance operation being required.
2214          */
2215         xfs_icsb_balance_counter(mp, field, 0, delta);
2216         xfs_icsb_unlock(mp);
2217         goto again;
2218 }
2219
2220 #endif