Merge 'acpi-2.6.12' branch into to-akpm
[linux-2.6] / fs / xfs / xfs_inode_item.c
1 /*
2  * Copyright (c) 2000-2002 Silicon Graphics, Inc.  All Rights Reserved.
3  *
4  * This program is free software; you can redistribute it and/or modify it
5  * under the terms of version 2 of the GNU General Public License as
6  * published by the Free Software Foundation.
7  *
8  * This program is distributed in the hope that it would be useful, but
9  * WITHOUT ANY WARRANTY; without even the implied warranty of
10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
11  *
12  * Further, this software is distributed without any warranty that it is
13  * free of the rightful claim of any third person regarding infringement
14  * or the like.  Any license provided herein, whether implied or
15  * otherwise, applies only to this software file.  Patent licenses, if
16  * any, provided herein do not apply to combinations of this program with
17  * other software, or any other product whatsoever.
18  *
19  * You should have received a copy of the GNU General Public License along
20  * with this program; if not, write the Free Software Foundation, Inc., 59
21  * Temple Place - Suite 330, Boston MA 02111-1307, USA.
22  *
23  * Contact information: Silicon Graphics, Inc., 1600 Amphitheatre Pkwy,
24  * Mountain View, CA  94043, or:
25  *
26  * http://www.sgi.com
27  *
28  * For further information regarding this notice, see:
29  *
30  * http://oss.sgi.com/projects/GenInfo/SGIGPLNoticeExplan/
31  */
32
33 /*
34  * This file contains the implementation of the xfs_inode_log_item.
35  * It contains the item operations used to manipulate the inode log
36  * items as well as utility routines used by the inode specific
37  * transaction routines.
38  */
39 #include "xfs.h"
40 #include "xfs_macros.h"
41 #include "xfs_types.h"
42 #include "xfs_inum.h"
43 #include "xfs_log.h"
44 #include "xfs_trans.h"
45 #include "xfs_buf_item.h"
46 #include "xfs_sb.h"
47 #include "xfs_dir.h"
48 #include "xfs_dir2.h"
49 #include "xfs_dmapi.h"
50 #include "xfs_mount.h"
51 #include "xfs_trans_priv.h"
52 #include "xfs_ag.h"
53 #include "xfs_alloc_btree.h"
54 #include "xfs_bmap_btree.h"
55 #include "xfs_ialloc_btree.h"
56 #include "xfs_btree.h"
57 #include "xfs_ialloc.h"
58 #include "xfs_attr_sf.h"
59 #include "xfs_dir_sf.h"
60 #include "xfs_dir2_sf.h"
61 #include "xfs_dinode.h"
62 #include "xfs_inode_item.h"
63 #include "xfs_inode.h"
64 #include "xfs_rw.h"
65
66
67 kmem_zone_t     *xfs_ili_zone;          /* inode log item zone */
68
69 /*
70  * This returns the number of iovecs needed to log the given inode item.
71  *
72  * We need one iovec for the inode log format structure, one for the
73  * inode core, and possibly one for the inode data/extents/b-tree root
74  * and one for the inode attribute data/extents/b-tree root.
75  */
76 STATIC uint
77 xfs_inode_item_size(
78         xfs_inode_log_item_t    *iip)
79 {
80         uint            nvecs;
81         xfs_inode_t     *ip;
82
83         ip = iip->ili_inode;
84         nvecs = 2;
85
86         /*
87          * Only log the data/extents/b-tree root if there is something
88          * left to log.
89          */
90         iip->ili_format.ilf_fields |= XFS_ILOG_CORE;
91
92         switch (ip->i_d.di_format) {
93         case XFS_DINODE_FMT_EXTENTS:
94                 iip->ili_format.ilf_fields &=
95                         ~(XFS_ILOG_DDATA | XFS_ILOG_DBROOT |
96                           XFS_ILOG_DEV | XFS_ILOG_UUID);
97                 if ((iip->ili_format.ilf_fields & XFS_ILOG_DEXT) &&
98                     (ip->i_d.di_nextents > 0) &&
99                     (ip->i_df.if_bytes > 0)) {
100                         ASSERT(ip->i_df.if_u1.if_extents != NULL);
101                         nvecs++;
102                 } else {
103                         iip->ili_format.ilf_fields &= ~XFS_ILOG_DEXT;
104                 }
105                 break;
106
107         case XFS_DINODE_FMT_BTREE:
108                 ASSERT(ip->i_df.if_ext_max ==
109                        XFS_IFORK_DSIZE(ip) / (uint)sizeof(xfs_bmbt_rec_t));
110                 iip->ili_format.ilf_fields &=
111                         ~(XFS_ILOG_DDATA | XFS_ILOG_DEXT |
112                           XFS_ILOG_DEV | XFS_ILOG_UUID);
113                 if ((iip->ili_format.ilf_fields & XFS_ILOG_DBROOT) &&
114                     (ip->i_df.if_broot_bytes > 0)) {
115                         ASSERT(ip->i_df.if_broot != NULL);
116                         nvecs++;
117                 } else {
118                         ASSERT(!(iip->ili_format.ilf_fields &
119                                  XFS_ILOG_DBROOT));
120 #ifdef XFS_TRANS_DEBUG
121                         if (iip->ili_root_size > 0) {
122                                 ASSERT(iip->ili_root_size ==
123                                        ip->i_df.if_broot_bytes);
124                                 ASSERT(memcmp(iip->ili_orig_root,
125                                             ip->i_df.if_broot,
126                                             iip->ili_root_size) == 0);
127                         } else {
128                                 ASSERT(ip->i_df.if_broot_bytes == 0);
129                         }
130 #endif
131                         iip->ili_format.ilf_fields &= ~XFS_ILOG_DBROOT;
132                 }
133                 break;
134
135         case XFS_DINODE_FMT_LOCAL:
136                 iip->ili_format.ilf_fields &=
137                         ~(XFS_ILOG_DEXT | XFS_ILOG_DBROOT |
138                           XFS_ILOG_DEV | XFS_ILOG_UUID);
139                 if ((iip->ili_format.ilf_fields & XFS_ILOG_DDATA) &&
140                     (ip->i_df.if_bytes > 0)) {
141                         ASSERT(ip->i_df.if_u1.if_data != NULL);
142                         ASSERT(ip->i_d.di_size > 0);
143                         nvecs++;
144                 } else {
145                         iip->ili_format.ilf_fields &= ~XFS_ILOG_DDATA;
146                 }
147                 break;
148
149         case XFS_DINODE_FMT_DEV:
150                 iip->ili_format.ilf_fields &=
151                         ~(XFS_ILOG_DDATA | XFS_ILOG_DBROOT |
152                           XFS_ILOG_DEXT | XFS_ILOG_UUID);
153                 break;
154
155         case XFS_DINODE_FMT_UUID:
156                 iip->ili_format.ilf_fields &=
157                         ~(XFS_ILOG_DDATA | XFS_ILOG_DBROOT |
158                           XFS_ILOG_DEXT | XFS_ILOG_DEV);
159                 break;
160
161         default:
162                 ASSERT(0);
163                 break;
164         }
165
166         /*
167          * If there are no attributes associated with this file,
168          * then there cannot be anything more to log.
169          * Clear all attribute-related log flags.
170          */
171         if (!XFS_IFORK_Q(ip)) {
172                 iip->ili_format.ilf_fields &=
173                         ~(XFS_ILOG_ADATA | XFS_ILOG_ABROOT | XFS_ILOG_AEXT);
174                 return nvecs;
175         }
176
177         /*
178          * Log any necessary attribute data.
179          */
180         switch (ip->i_d.di_aformat) {
181         case XFS_DINODE_FMT_EXTENTS:
182                 iip->ili_format.ilf_fields &=
183                         ~(XFS_ILOG_ADATA | XFS_ILOG_ABROOT);
184                 if ((iip->ili_format.ilf_fields & XFS_ILOG_AEXT) &&
185                     (ip->i_d.di_anextents > 0) &&
186                     (ip->i_afp->if_bytes > 0)) {
187                         ASSERT(ip->i_afp->if_u1.if_extents != NULL);
188                         nvecs++;
189                 } else {
190                         iip->ili_format.ilf_fields &= ~XFS_ILOG_AEXT;
191                 }
192                 break;
193
194         case XFS_DINODE_FMT_BTREE:
195                 iip->ili_format.ilf_fields &=
196                         ~(XFS_ILOG_ADATA | XFS_ILOG_AEXT);
197                 if ((iip->ili_format.ilf_fields & XFS_ILOG_ABROOT) &&
198                     (ip->i_afp->if_broot_bytes > 0)) {
199                         ASSERT(ip->i_afp->if_broot != NULL);
200                         nvecs++;
201                 } else {
202                         iip->ili_format.ilf_fields &= ~XFS_ILOG_ABROOT;
203                 }
204                 break;
205
206         case XFS_DINODE_FMT_LOCAL:
207                 iip->ili_format.ilf_fields &=
208                         ~(XFS_ILOG_AEXT | XFS_ILOG_ABROOT);
209                 if ((iip->ili_format.ilf_fields & XFS_ILOG_ADATA) &&
210                     (ip->i_afp->if_bytes > 0)) {
211                         ASSERT(ip->i_afp->if_u1.if_data != NULL);
212                         nvecs++;
213                 } else {
214                         iip->ili_format.ilf_fields &= ~XFS_ILOG_ADATA;
215                 }
216                 break;
217
218         default:
219                 ASSERT(0);
220                 break;
221         }
222
223         return nvecs;
224 }
225
226 /*
227  * This is called to fill in the vector of log iovecs for the
228  * given inode log item.  It fills the first item with an inode
229  * log format structure, the second with the on-disk inode structure,
230  * and a possible third and/or fourth with the inode data/extents/b-tree
231  * root and inode attributes data/extents/b-tree root.
232  */
233 STATIC void
234 xfs_inode_item_format(
235         xfs_inode_log_item_t    *iip,
236         xfs_log_iovec_t         *log_vector)
237 {
238         uint                    nvecs;
239         xfs_log_iovec_t         *vecp;
240         xfs_inode_t             *ip;
241         size_t                  data_bytes;
242         xfs_bmbt_rec_t          *ext_buffer;
243         int                     nrecs;
244         xfs_mount_t             *mp;
245
246         ip = iip->ili_inode;
247         vecp = log_vector;
248
249         vecp->i_addr = (xfs_caddr_t)&iip->ili_format;
250         vecp->i_len  = sizeof(xfs_inode_log_format_t);
251         vecp++;
252         nvecs        = 1;
253
254         /*
255          * Clear i_update_core if the timestamps (or any other
256          * non-transactional modification) need flushing/logging
257          * and we're about to log them with the rest of the core.
258          *
259          * This is the same logic as xfs_iflush() but this code can't
260          * run at the same time as xfs_iflush because we're in commit
261          * processing here and so we have the inode lock held in
262          * exclusive mode.  Although it doesn't really matter
263          * for the timestamps if both routines were to grab the
264          * timestamps or not.  That would be ok.
265          *
266          * We clear i_update_core before copying out the data.
267          * This is for coordination with our timestamp updates
268          * that don't hold the inode lock. They will always
269          * update the timestamps BEFORE setting i_update_core,
270          * so if we clear i_update_core after they set it we
271          * are guaranteed to see their updates to the timestamps
272          * either here.  Likewise, if they set it after we clear it
273          * here, we'll see it either on the next commit of this
274          * inode or the next time the inode gets flushed via
275          * xfs_iflush().  This depends on strongly ordered memory
276          * semantics, but we have that.  We use the SYNCHRONIZE
277          * macro to make sure that the compiler does not reorder
278          * the i_update_core access below the data copy below.
279          */
280         if (ip->i_update_core)  {
281                 ip->i_update_core = 0;
282                 SYNCHRONIZE();
283         }
284
285         /*
286          * We don't have to worry about re-ordering here because
287          * the update_size field is protected by the inode lock
288          * and we have that held in exclusive mode.
289          */
290         if (ip->i_update_size)
291                 ip->i_update_size = 0;
292
293         vecp->i_addr = (xfs_caddr_t)&ip->i_d;
294         vecp->i_len  = sizeof(xfs_dinode_core_t);
295         vecp++;
296         nvecs++;
297         iip->ili_format.ilf_fields |= XFS_ILOG_CORE;
298
299         /*
300          * If this is really an old format inode, then we need to
301          * log it as such.  This means that we have to copy the link
302          * count from the new field to the old.  We don't have to worry
303          * about the new fields, because nothing trusts them as long as
304          * the old inode version number is there.  If the superblock already
305          * has a new version number, then we don't bother converting back.
306          */
307         mp = ip->i_mount;
308         ASSERT(ip->i_d.di_version == XFS_DINODE_VERSION_1 ||
309                XFS_SB_VERSION_HASNLINK(&mp->m_sb));
310         if (ip->i_d.di_version == XFS_DINODE_VERSION_1) {
311                 if (!XFS_SB_VERSION_HASNLINK(&mp->m_sb)) {
312                         /*
313                          * Convert it back.
314                          */
315                         ASSERT(ip->i_d.di_nlink <= XFS_MAXLINK_1);
316                         ip->i_d.di_onlink = ip->i_d.di_nlink;
317                 } else {
318                         /*
319                          * The superblock version has already been bumped,
320                          * so just make the conversion to the new inode
321                          * format permanent.
322                          */
323                         ip->i_d.di_version = XFS_DINODE_VERSION_2;
324                         ip->i_d.di_onlink = 0;
325                         memset(&(ip->i_d.di_pad[0]), 0, sizeof(ip->i_d.di_pad));
326                 }
327         }
328
329         switch (ip->i_d.di_format) {
330         case XFS_DINODE_FMT_EXTENTS:
331                 ASSERT(!(iip->ili_format.ilf_fields &
332                          (XFS_ILOG_DDATA | XFS_ILOG_DBROOT |
333                           XFS_ILOG_DEV | XFS_ILOG_UUID)));
334                 if (iip->ili_format.ilf_fields & XFS_ILOG_DEXT) {
335                         ASSERT(ip->i_df.if_bytes > 0);
336                         ASSERT(ip->i_df.if_u1.if_extents != NULL);
337                         ASSERT(ip->i_d.di_nextents > 0);
338                         ASSERT(iip->ili_extents_buf == NULL);
339                         nrecs = ip->i_df.if_bytes /
340                                 (uint)sizeof(xfs_bmbt_rec_t);
341                         ASSERT(nrecs > 0);
342 #if __BYTE_ORDER == __BIG_ENDIAN
343                         if (nrecs == ip->i_d.di_nextents) {
344                                 /*
345                                  * There are no delayed allocation
346                                  * extents, so just point to the
347                                  * real extents array.
348                                  */
349                                 vecp->i_addr =
350                                         (char *)(ip->i_df.if_u1.if_extents);
351                                 vecp->i_len = ip->i_df.if_bytes;
352                         } else
353 #endif
354                         {
355                                 /*
356                                  * There are delayed allocation extents
357                                  * in the inode, or we need to convert
358                                  * the extents to on disk format.
359                                  * Use xfs_iextents_copy()
360                                  * to copy only the real extents into
361                                  * a separate buffer.  We'll free the
362                                  * buffer in the unlock routine.
363                                  */
364                                 ext_buffer = kmem_alloc(ip->i_df.if_bytes,
365                                         KM_SLEEP);
366                                 iip->ili_extents_buf = ext_buffer;
367                                 vecp->i_addr = (xfs_caddr_t)ext_buffer;
368                                 vecp->i_len = xfs_iextents_copy(ip, ext_buffer,
369                                                 XFS_DATA_FORK);
370                         }
371                         ASSERT(vecp->i_len <= ip->i_df.if_bytes);
372                         iip->ili_format.ilf_dsize = vecp->i_len;
373                         vecp++;
374                         nvecs++;
375                 }
376                 break;
377
378         case XFS_DINODE_FMT_BTREE:
379                 ASSERT(!(iip->ili_format.ilf_fields &
380                          (XFS_ILOG_DDATA | XFS_ILOG_DEXT |
381                           XFS_ILOG_DEV | XFS_ILOG_UUID)));
382                 if (iip->ili_format.ilf_fields & XFS_ILOG_DBROOT) {
383                         ASSERT(ip->i_df.if_broot_bytes > 0);
384                         ASSERT(ip->i_df.if_broot != NULL);
385                         vecp->i_addr = (xfs_caddr_t)ip->i_df.if_broot;
386                         vecp->i_len = ip->i_df.if_broot_bytes;
387                         vecp++;
388                         nvecs++;
389                         iip->ili_format.ilf_dsize = ip->i_df.if_broot_bytes;
390                 }
391                 break;
392
393         case XFS_DINODE_FMT_LOCAL:
394                 ASSERT(!(iip->ili_format.ilf_fields &
395                          (XFS_ILOG_DBROOT | XFS_ILOG_DEXT |
396                           XFS_ILOG_DEV | XFS_ILOG_UUID)));
397                 if (iip->ili_format.ilf_fields & XFS_ILOG_DDATA) {
398                         ASSERT(ip->i_df.if_bytes > 0);
399                         ASSERT(ip->i_df.if_u1.if_data != NULL);
400                         ASSERT(ip->i_d.di_size > 0);
401
402                         vecp->i_addr = (xfs_caddr_t)ip->i_df.if_u1.if_data;
403                         /*
404                          * Round i_bytes up to a word boundary.
405                          * The underlying memory is guaranteed to
406                          * to be there by xfs_idata_realloc().
407                          */
408                         data_bytes = roundup(ip->i_df.if_bytes, 4);
409                         ASSERT((ip->i_df.if_real_bytes == 0) ||
410                                (ip->i_df.if_real_bytes == data_bytes));
411                         vecp->i_len = (int)data_bytes;
412                         vecp++;
413                         nvecs++;
414                         iip->ili_format.ilf_dsize = (unsigned)data_bytes;
415                 }
416                 break;
417
418         case XFS_DINODE_FMT_DEV:
419                 ASSERT(!(iip->ili_format.ilf_fields &
420                          (XFS_ILOG_DBROOT | XFS_ILOG_DEXT |
421                           XFS_ILOG_DDATA | XFS_ILOG_UUID)));
422                 if (iip->ili_format.ilf_fields & XFS_ILOG_DEV) {
423                         iip->ili_format.ilf_u.ilfu_rdev =
424                                 ip->i_df.if_u2.if_rdev;
425                 }
426                 break;
427
428         case XFS_DINODE_FMT_UUID:
429                 ASSERT(!(iip->ili_format.ilf_fields &
430                          (XFS_ILOG_DBROOT | XFS_ILOG_DEXT |
431                           XFS_ILOG_DDATA | XFS_ILOG_DEV)));
432                 if (iip->ili_format.ilf_fields & XFS_ILOG_UUID) {
433                         iip->ili_format.ilf_u.ilfu_uuid =
434                                 ip->i_df.if_u2.if_uuid;
435                 }
436                 break;
437
438         default:
439                 ASSERT(0);
440                 break;
441         }
442
443         /*
444          * If there are no attributes associated with the file,
445          * then we're done.
446          * Assert that no attribute-related log flags are set.
447          */
448         if (!XFS_IFORK_Q(ip)) {
449                 ASSERT(nvecs == iip->ili_item.li_desc->lid_size);
450                 iip->ili_format.ilf_size = nvecs;
451                 ASSERT(!(iip->ili_format.ilf_fields &
452                          (XFS_ILOG_ADATA | XFS_ILOG_ABROOT | XFS_ILOG_AEXT)));
453                 return;
454         }
455
456         switch (ip->i_d.di_aformat) {
457         case XFS_DINODE_FMT_EXTENTS:
458                 ASSERT(!(iip->ili_format.ilf_fields &
459                          (XFS_ILOG_ADATA | XFS_ILOG_ABROOT)));
460                 if (iip->ili_format.ilf_fields & XFS_ILOG_AEXT) {
461                         ASSERT(ip->i_afp->if_bytes > 0);
462                         ASSERT(ip->i_afp->if_u1.if_extents != NULL);
463                         ASSERT(ip->i_d.di_anextents > 0);
464 #ifdef DEBUG
465                         nrecs = ip->i_afp->if_bytes /
466                                 (uint)sizeof(xfs_bmbt_rec_t);
467 #endif
468                         ASSERT(nrecs > 0);
469                         ASSERT(nrecs == ip->i_d.di_anextents);
470 #if __BYTE_ORDER == __BIG_ENDIAN
471                         /*
472                          * There are not delayed allocation extents
473                          * for attributes, so just point at the array.
474                          */
475                         vecp->i_addr = (char *)(ip->i_afp->if_u1.if_extents);
476                         vecp->i_len = ip->i_afp->if_bytes;
477 #else
478                         ASSERT(iip->ili_aextents_buf == NULL);
479                         /*
480                          * Need to endian flip before logging
481                          */
482                         ext_buffer = kmem_alloc(ip->i_afp->if_bytes,
483                                 KM_SLEEP);
484                         iip->ili_aextents_buf = ext_buffer;
485                         vecp->i_addr = (xfs_caddr_t)ext_buffer;
486                         vecp->i_len = xfs_iextents_copy(ip, ext_buffer,
487                                         XFS_ATTR_FORK);
488 #endif
489                         iip->ili_format.ilf_asize = vecp->i_len;
490                         vecp++;
491                         nvecs++;
492                 }
493                 break;
494
495         case XFS_DINODE_FMT_BTREE:
496                 ASSERT(!(iip->ili_format.ilf_fields &
497                          (XFS_ILOG_ADATA | XFS_ILOG_AEXT)));
498                 if (iip->ili_format.ilf_fields & XFS_ILOG_ABROOT) {
499                         ASSERT(ip->i_afp->if_broot_bytes > 0);
500                         ASSERT(ip->i_afp->if_broot != NULL);
501                         vecp->i_addr = (xfs_caddr_t)ip->i_afp->if_broot;
502                         vecp->i_len = ip->i_afp->if_broot_bytes;
503                         vecp++;
504                         nvecs++;
505                         iip->ili_format.ilf_asize = ip->i_afp->if_broot_bytes;
506                 }
507                 break;
508
509         case XFS_DINODE_FMT_LOCAL:
510                 ASSERT(!(iip->ili_format.ilf_fields &
511                          (XFS_ILOG_ABROOT | XFS_ILOG_AEXT)));
512                 if (iip->ili_format.ilf_fields & XFS_ILOG_ADATA) {
513                         ASSERT(ip->i_afp->if_bytes > 0);
514                         ASSERT(ip->i_afp->if_u1.if_data != NULL);
515
516                         vecp->i_addr = (xfs_caddr_t)ip->i_afp->if_u1.if_data;
517                         /*
518                          * Round i_bytes up to a word boundary.
519                          * The underlying memory is guaranteed to
520                          * to be there by xfs_idata_realloc().
521                          */
522                         data_bytes = roundup(ip->i_afp->if_bytes, 4);
523                         ASSERT((ip->i_afp->if_real_bytes == 0) ||
524                                (ip->i_afp->if_real_bytes == data_bytes));
525                         vecp->i_len = (int)data_bytes;
526                         vecp++;
527                         nvecs++;
528                         iip->ili_format.ilf_asize = (unsigned)data_bytes;
529                 }
530                 break;
531
532         default:
533                 ASSERT(0);
534                 break;
535         }
536
537         ASSERT(nvecs == iip->ili_item.li_desc->lid_size);
538         iip->ili_format.ilf_size = nvecs;
539 }
540
541
542 /*
543  * This is called to pin the inode associated with the inode log
544  * item in memory so it cannot be written out.  Do this by calling
545  * xfs_ipin() to bump the pin count in the inode while holding the
546  * inode pin lock.
547  */
548 STATIC void
549 xfs_inode_item_pin(
550         xfs_inode_log_item_t    *iip)
551 {
552         ASSERT(ismrlocked(&(iip->ili_inode->i_lock), MR_UPDATE));
553         xfs_ipin(iip->ili_inode);
554 }
555
556
557 /*
558  * This is called to unpin the inode associated with the inode log
559  * item which was previously pinned with a call to xfs_inode_item_pin().
560  * Just call xfs_iunpin() on the inode to do this.
561  */
562 /* ARGSUSED */
563 STATIC void
564 xfs_inode_item_unpin(
565         xfs_inode_log_item_t    *iip,
566         int                     stale)
567 {
568         xfs_iunpin(iip->ili_inode);
569 }
570
571 /* ARGSUSED */
572 STATIC void
573 xfs_inode_item_unpin_remove(
574         xfs_inode_log_item_t    *iip,
575         xfs_trans_t             *tp)
576 {
577         xfs_iunpin(iip->ili_inode);
578 }
579
580 /*
581  * This is called to attempt to lock the inode associated with this
582  * inode log item, in preparation for the push routine which does the actual
583  * iflush.  Don't sleep on the inode lock or the flush lock.
584  *
585  * If the flush lock is already held, indicating that the inode has
586  * been or is in the process of being flushed, then (ideally) we'd like to
587  * see if the inode's buffer is still incore, and if so give it a nudge.
588  * We delay doing so until the pushbuf routine, though, to avoid holding
589  * the AIL lock across a call to the blackhole which is the buffercache.
590  * Also we don't want to sleep in any device strategy routines, which can happen
591  * if we do the subsequent bawrite in here.
592  */
593 STATIC uint
594 xfs_inode_item_trylock(
595         xfs_inode_log_item_t    *iip)
596 {
597         register xfs_inode_t    *ip;
598
599         ip = iip->ili_inode;
600
601         if (xfs_ipincount(ip) > 0) {
602                 return XFS_ITEM_PINNED;
603         }
604
605         if (!xfs_ilock_nowait(ip, XFS_ILOCK_SHARED)) {
606                 return XFS_ITEM_LOCKED;
607         }
608
609         if (!xfs_iflock_nowait(ip)) {
610                 /*
611                  * If someone else isn't already trying to push the inode
612                  * buffer, we get to do it.
613                  */
614                 if (iip->ili_pushbuf_flag == 0) {
615                         iip->ili_pushbuf_flag = 1;
616 #ifdef DEBUG
617                         iip->ili_push_owner = get_thread_id();
618 #endif
619                         /*
620                          * Inode is left locked in shared mode.
621                          * Pushbuf routine gets to unlock it.
622                          */
623                         return XFS_ITEM_PUSHBUF;
624                 } else {
625                         /*
626                          * We hold the AIL_LOCK, so we must specify the
627                          * NONOTIFY flag so that we won't double trip.
628                          */
629                         xfs_iunlock(ip, XFS_ILOCK_SHARED|XFS_IUNLOCK_NONOTIFY);
630                         return XFS_ITEM_FLUSHING;
631                 }
632                 /* NOTREACHED */
633         }
634
635         /* Stale items should force out the iclog */
636         if (ip->i_flags & XFS_ISTALE) {
637                 xfs_ifunlock(ip);
638                 xfs_iunlock(ip, XFS_ILOCK_SHARED|XFS_IUNLOCK_NONOTIFY);
639                 return XFS_ITEM_PINNED;
640         }
641
642 #ifdef DEBUG
643         if (!XFS_FORCED_SHUTDOWN(ip->i_mount)) {
644                 ASSERT(iip->ili_format.ilf_fields != 0);
645                 ASSERT(iip->ili_logged == 0);
646                 ASSERT(iip->ili_item.li_flags & XFS_LI_IN_AIL);
647         }
648 #endif
649         return XFS_ITEM_SUCCESS;
650 }
651
652 /*
653  * Unlock the inode associated with the inode log item.
654  * Clear the fields of the inode and inode log item that
655  * are specific to the current transaction.  If the
656  * hold flags is set, do not unlock the inode.
657  */
658 STATIC void
659 xfs_inode_item_unlock(
660         xfs_inode_log_item_t    *iip)
661 {
662         uint            hold;
663         uint            iolocked;
664         uint            lock_flags;
665         xfs_inode_t     *ip;
666
667         ASSERT(iip != NULL);
668         ASSERT(iip->ili_inode->i_itemp != NULL);
669         ASSERT(ismrlocked(&(iip->ili_inode->i_lock), MR_UPDATE));
670         ASSERT((!(iip->ili_inode->i_itemp->ili_flags &
671                   XFS_ILI_IOLOCKED_EXCL)) ||
672                ismrlocked(&(iip->ili_inode->i_iolock), MR_UPDATE));
673         ASSERT((!(iip->ili_inode->i_itemp->ili_flags &
674                   XFS_ILI_IOLOCKED_SHARED)) ||
675                ismrlocked(&(iip->ili_inode->i_iolock), MR_ACCESS));
676         /*
677          * Clear the transaction pointer in the inode.
678          */
679         ip = iip->ili_inode;
680         ip->i_transp = NULL;
681
682         /*
683          * If the inode needed a separate buffer with which to log
684          * its extents, then free it now.
685          */
686         if (iip->ili_extents_buf != NULL) {
687                 ASSERT(ip->i_d.di_format == XFS_DINODE_FMT_EXTENTS);
688                 ASSERT(ip->i_d.di_nextents > 0);
689                 ASSERT(iip->ili_format.ilf_fields & XFS_ILOG_DEXT);
690                 ASSERT(ip->i_df.if_bytes > 0);
691                 kmem_free(iip->ili_extents_buf, ip->i_df.if_bytes);
692                 iip->ili_extents_buf = NULL;
693         }
694         if (iip->ili_aextents_buf != NULL) {
695                 ASSERT(ip->i_d.di_aformat == XFS_DINODE_FMT_EXTENTS);
696                 ASSERT(ip->i_d.di_anextents > 0);
697                 ASSERT(iip->ili_format.ilf_fields & XFS_ILOG_AEXT);
698                 ASSERT(ip->i_afp->if_bytes > 0);
699                 kmem_free(iip->ili_aextents_buf, ip->i_afp->if_bytes);
700                 iip->ili_aextents_buf = NULL;
701         }
702
703         /*
704          * Figure out if we should unlock the inode or not.
705          */
706         hold = iip->ili_flags & XFS_ILI_HOLD;
707
708         /*
709          * Before clearing out the flags, remember whether we
710          * are holding the inode's IO lock.
711          */
712         iolocked = iip->ili_flags & XFS_ILI_IOLOCKED_ANY;
713
714         /*
715          * Clear out the fields of the inode log item particular
716          * to the current transaction.
717          */
718         iip->ili_ilock_recur = 0;
719         iip->ili_iolock_recur = 0;
720         iip->ili_flags = 0;
721
722         /*
723          * Unlock the inode if XFS_ILI_HOLD was not set.
724          */
725         if (!hold) {
726                 lock_flags = XFS_ILOCK_EXCL;
727                 if (iolocked & XFS_ILI_IOLOCKED_EXCL) {
728                         lock_flags |= XFS_IOLOCK_EXCL;
729                 } else if (iolocked & XFS_ILI_IOLOCKED_SHARED) {
730                         lock_flags |= XFS_IOLOCK_SHARED;
731                 }
732                 xfs_iput(iip->ili_inode, lock_flags);
733         }
734 }
735
736 /*
737  * This is called to find out where the oldest active copy of the
738  * inode log item in the on disk log resides now that the last log
739  * write of it completed at the given lsn.  Since we always re-log
740  * all dirty data in an inode, the latest copy in the on disk log
741  * is the only one that matters.  Therefore, simply return the
742  * given lsn.
743  */
744 /*ARGSUSED*/
745 STATIC xfs_lsn_t
746 xfs_inode_item_committed(
747         xfs_inode_log_item_t    *iip,
748         xfs_lsn_t               lsn)
749 {
750         return (lsn);
751 }
752
753 /*
754  * The transaction with the inode locked has aborted.  The inode
755  * must not be dirty within the transaction (unless we're forcibly
756  * shutting down).  We simply unlock just as if the transaction
757  * had been cancelled.
758  */
759 STATIC void
760 xfs_inode_item_abort(
761         xfs_inode_log_item_t    *iip)
762 {
763         xfs_inode_item_unlock(iip);
764         return;
765 }
766
767
768 /*
769  * This gets called by xfs_trans_push_ail(), when IOP_TRYLOCK
770  * failed to get the inode flush lock but did get the inode locked SHARED.
771  * Here we're trying to see if the inode buffer is incore, and if so whether it's
772  * marked delayed write. If that's the case, we'll initiate a bawrite on that
773  * buffer to expedite the process.
774  *
775  * We aren't holding the AIL_LOCK (or the flush lock) when this gets called,
776  * so it is inherently race-y.
777  */
778 STATIC void
779 xfs_inode_item_pushbuf(
780         xfs_inode_log_item_t    *iip)
781 {
782         xfs_inode_t     *ip;
783         xfs_mount_t     *mp;
784         xfs_buf_t       *bp;
785         uint            dopush;
786
787         ip = iip->ili_inode;
788
789         ASSERT(ismrlocked(&(ip->i_lock), MR_ACCESS));
790
791         /*
792          * The ili_pushbuf_flag keeps others from
793          * trying to duplicate our effort.
794          */
795         ASSERT(iip->ili_pushbuf_flag != 0);
796         ASSERT(iip->ili_push_owner == get_thread_id());
797
798         /*
799          * If flushlock isn't locked anymore, chances are that the
800          * inode flush completed and the inode was taken off the AIL.
801          * So, just get out.
802          */
803         if ((valusema(&(ip->i_flock)) > 0)  ||
804             ((iip->ili_item.li_flags & XFS_LI_IN_AIL) == 0)) {
805                 iip->ili_pushbuf_flag = 0;
806                 xfs_iunlock(ip, XFS_ILOCK_SHARED);
807                 return;
808         }
809
810         mp = ip->i_mount;
811         bp = xfs_incore(mp->m_ddev_targp, iip->ili_format.ilf_blkno,
812                     iip->ili_format.ilf_len, XFS_INCORE_TRYLOCK);
813
814         if (bp != NULL) {
815                 if (XFS_BUF_ISDELAYWRITE(bp)) {
816                         /*
817                          * We were racing with iflush because we don't hold
818                          * the AIL_LOCK or the flush lock. However, at this point,
819                          * we have the buffer, and we know that it's dirty.
820                          * So, it's possible that iflush raced with us, and
821                          * this item is already taken off the AIL.
822                          * If not, we can flush it async.
823                          */
824                         dopush = ((iip->ili_item.li_flags & XFS_LI_IN_AIL) &&
825                                   (valusema(&(ip->i_flock)) <= 0));
826                         iip->ili_pushbuf_flag = 0;
827                         xfs_iunlock(ip, XFS_ILOCK_SHARED);
828                         xfs_buftrace("INODE ITEM PUSH", bp);
829                         if (XFS_BUF_ISPINNED(bp)) {
830                                 xfs_log_force(mp, (xfs_lsn_t)0,
831                                               XFS_LOG_FORCE);
832                         }
833                         if (dopush) {
834                                 xfs_bawrite(mp, bp);
835                         } else {
836                                 xfs_buf_relse(bp);
837                         }
838                 } else {
839                         iip->ili_pushbuf_flag = 0;
840                         xfs_iunlock(ip, XFS_ILOCK_SHARED);
841                         xfs_buf_relse(bp);
842                 }
843                 return;
844         }
845         /*
846          * We have to be careful about resetting pushbuf flag too early (above).
847          * Even though in theory we can do it as soon as we have the buflock,
848          * we don't want others to be doing work needlessly. They'll come to
849          * this function thinking that pushing the buffer is their
850          * responsibility only to find that the buffer is still locked by
851          * another doing the same thing
852          */
853         iip->ili_pushbuf_flag = 0;
854         xfs_iunlock(ip, XFS_ILOCK_SHARED);
855         return;
856 }
857
858
859 /*
860  * This is called to asynchronously write the inode associated with this
861  * inode log item out to disk. The inode will already have been locked by
862  * a successful call to xfs_inode_item_trylock().
863  */
864 STATIC void
865 xfs_inode_item_push(
866         xfs_inode_log_item_t    *iip)
867 {
868         xfs_inode_t     *ip;
869
870         ip = iip->ili_inode;
871
872         ASSERT(ismrlocked(&(ip->i_lock), MR_ACCESS));
873         ASSERT(valusema(&(ip->i_flock)) <= 0);
874         /*
875          * Since we were able to lock the inode's flush lock and
876          * we found it on the AIL, the inode must be dirty.  This
877          * is because the inode is removed from the AIL while still
878          * holding the flush lock in xfs_iflush_done().  Thus, if
879          * we found it in the AIL and were able to obtain the flush
880          * lock without sleeping, then there must not have been
881          * anyone in the process of flushing the inode.
882          */
883         ASSERT(XFS_FORCED_SHUTDOWN(ip->i_mount) ||
884                iip->ili_format.ilf_fields != 0);
885
886         /*
887          * Write out the inode.  The completion routine ('iflush_done') will
888          * pull it from the AIL, mark it clean, unlock the flush lock.
889          */
890         (void) xfs_iflush(ip, XFS_IFLUSH_ASYNC);
891         xfs_iunlock(ip, XFS_ILOCK_SHARED);
892
893         return;
894 }
895
896 /*
897  * XXX rcc - this one really has to do something.  Probably needs
898  * to stamp in a new field in the incore inode.
899  */
900 /* ARGSUSED */
901 STATIC void
902 xfs_inode_item_committing(
903         xfs_inode_log_item_t    *iip,
904         xfs_lsn_t               lsn)
905 {
906         iip->ili_last_lsn = lsn;
907         return;
908 }
909
910 /*
911  * This is the ops vector shared by all buf log items.
912  */
913 STATIC struct xfs_item_ops xfs_inode_item_ops = {
914         .iop_size       = (uint(*)(xfs_log_item_t*))xfs_inode_item_size,
915         .iop_format     = (void(*)(xfs_log_item_t*, xfs_log_iovec_t*))
916                                         xfs_inode_item_format,
917         .iop_pin        = (void(*)(xfs_log_item_t*))xfs_inode_item_pin,
918         .iop_unpin      = (void(*)(xfs_log_item_t*, int))xfs_inode_item_unpin,
919         .iop_unpin_remove = (void(*)(xfs_log_item_t*, xfs_trans_t*))
920                                         xfs_inode_item_unpin_remove,
921         .iop_trylock    = (uint(*)(xfs_log_item_t*))xfs_inode_item_trylock,
922         .iop_unlock     = (void(*)(xfs_log_item_t*))xfs_inode_item_unlock,
923         .iop_committed  = (xfs_lsn_t(*)(xfs_log_item_t*, xfs_lsn_t))
924                                         xfs_inode_item_committed,
925         .iop_push       = (void(*)(xfs_log_item_t*))xfs_inode_item_push,
926         .iop_abort      = (void(*)(xfs_log_item_t*))xfs_inode_item_abort,
927         .iop_pushbuf    = (void(*)(xfs_log_item_t*))xfs_inode_item_pushbuf,
928         .iop_committing = (void(*)(xfs_log_item_t*, xfs_lsn_t))
929                                         xfs_inode_item_committing
930 };
931
932
933 /*
934  * Initialize the inode log item for a newly allocated (in-core) inode.
935  */
936 void
937 xfs_inode_item_init(
938         xfs_inode_t     *ip,
939         xfs_mount_t     *mp)
940 {
941         xfs_inode_log_item_t    *iip;
942
943         ASSERT(ip->i_itemp == NULL);
944         iip = ip->i_itemp = kmem_zone_zalloc(xfs_ili_zone, KM_SLEEP);
945
946         iip->ili_item.li_type = XFS_LI_INODE;
947         iip->ili_item.li_ops = &xfs_inode_item_ops;
948         iip->ili_item.li_mountp = mp;
949         iip->ili_inode = ip;
950
951         /*
952            We have zeroed memory. No need ...
953            iip->ili_extents_buf = NULL;
954            iip->ili_pushbuf_flag = 0;
955          */
956
957         iip->ili_format.ilf_type = XFS_LI_INODE;
958         iip->ili_format.ilf_ino = ip->i_ino;
959         iip->ili_format.ilf_blkno = ip->i_blkno;
960         iip->ili_format.ilf_len = ip->i_len;
961         iip->ili_format.ilf_boffset = ip->i_boffset;
962 }
963
964 /*
965  * Free the inode log item and any memory hanging off of it.
966  */
967 void
968 xfs_inode_item_destroy(
969         xfs_inode_t     *ip)
970 {
971 #ifdef XFS_TRANS_DEBUG
972         if (ip->i_itemp->ili_root_size != 0) {
973                 kmem_free(ip->i_itemp->ili_orig_root,
974                           ip->i_itemp->ili_root_size);
975         }
976 #endif
977         kmem_zone_free(xfs_ili_zone, ip->i_itemp);
978 }
979
980
981 /*
982  * This is the inode flushing I/O completion routine.  It is called
983  * from interrupt level when the buffer containing the inode is
984  * flushed to disk.  It is responsible for removing the inode item
985  * from the AIL if it has not been re-logged, and unlocking the inode's
986  * flush lock.
987  */
988 /*ARGSUSED*/
989 void
990 xfs_iflush_done(
991         xfs_buf_t               *bp,
992         xfs_inode_log_item_t    *iip)
993 {
994         xfs_inode_t     *ip;
995         SPLDECL(s);
996
997         ip = iip->ili_inode;
998
999         /*
1000          * We only want to pull the item from the AIL if it is
1001          * actually there and its location in the log has not
1002          * changed since we started the flush.  Thus, we only bother
1003          * if the ili_logged flag is set and the inode's lsn has not
1004          * changed.  First we check the lsn outside
1005          * the lock since it's cheaper, and then we recheck while
1006          * holding the lock before removing the inode from the AIL.
1007          */
1008         if (iip->ili_logged &&
1009             (iip->ili_item.li_lsn == iip->ili_flush_lsn)) {
1010                 AIL_LOCK(ip->i_mount, s);
1011                 if (iip->ili_item.li_lsn == iip->ili_flush_lsn) {
1012                         /*
1013                          * xfs_trans_delete_ail() drops the AIL lock.
1014                          */
1015                         xfs_trans_delete_ail(ip->i_mount,
1016                                              (xfs_log_item_t*)iip, s);
1017                 } else {
1018                         AIL_UNLOCK(ip->i_mount, s);
1019                 }
1020         }
1021
1022         iip->ili_logged = 0;
1023
1024         /*
1025          * Clear the ili_last_fields bits now that we know that the
1026          * data corresponding to them is safely on disk.
1027          */
1028         iip->ili_last_fields = 0;
1029
1030         /*
1031          * Release the inode's flush lock since we're done with it.
1032          */
1033         xfs_ifunlock(ip);
1034
1035         return;
1036 }
1037
1038 /*
1039  * This is the inode flushing abort routine.  It is called
1040  * from xfs_iflush when the filesystem is shutting down to clean
1041  * up the inode state.
1042  * It is responsible for removing the inode item
1043  * from the AIL if it has not been re-logged, and unlocking the inode's
1044  * flush lock.
1045  */
1046 void
1047 xfs_iflush_abort(
1048         xfs_inode_t             *ip)
1049 {
1050         xfs_inode_log_item_t    *iip;
1051         xfs_mount_t             *mp;
1052         SPLDECL(s);
1053
1054         iip = ip->i_itemp;
1055         mp = ip->i_mount;
1056         if (iip) {
1057                 if (iip->ili_item.li_flags & XFS_LI_IN_AIL) {
1058                         AIL_LOCK(mp, s);
1059                         if (iip->ili_item.li_flags & XFS_LI_IN_AIL) {
1060                                 /*
1061                                  * xfs_trans_delete_ail() drops the AIL lock.
1062                                  */
1063                                 xfs_trans_delete_ail(mp, (xfs_log_item_t *)iip,
1064                                         s);
1065                         } else
1066                                 AIL_UNLOCK(mp, s);
1067                 }
1068                 iip->ili_logged = 0;
1069                 /*
1070                  * Clear the ili_last_fields bits now that we know that the
1071                  * data corresponding to them is safely on disk.
1072                  */
1073                 iip->ili_last_fields = 0;
1074                 /*
1075                  * Clear the inode logging fields so no more flushes are
1076                  * attempted.
1077                  */
1078                 iip->ili_format.ilf_fields = 0;
1079         }
1080         /*
1081          * Release the inode's flush lock since we're done with it.
1082          */
1083         xfs_ifunlock(ip);
1084 }
1085
1086 void
1087 xfs_istale_done(
1088         xfs_buf_t               *bp,
1089         xfs_inode_log_item_t    *iip)
1090 {
1091         xfs_iflush_abort(iip->ili_inode);
1092 }