Commit | Line | Data |
---|---|---|
2a82b8be DC |
1 | /* |
2 | * Copyright (c) 2006-2007 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_bmap_btree.h" | |
20 | #include "xfs_inum.h" | |
21 | #include "xfs_dir2.h" | |
22 | #include "xfs_dir2_sf.h" | |
23 | #include "xfs_attr_sf.h" | |
24 | #include "xfs_dinode.h" | |
25 | #include "xfs_inode.h" | |
26 | #include "xfs_ag.h" | |
27 | #include "xfs_dmapi.h" | |
28 | #include "xfs_log.h" | |
29 | #include "xfs_trans.h" | |
30 | #include "xfs_sb.h" | |
31 | #include "xfs_mount.h" | |
32 | #include "xfs_bmap.h" | |
33 | #include "xfs_alloc.h" | |
34 | #include "xfs_utils.h" | |
35 | #include "xfs_mru_cache.h" | |
36 | #include "xfs_filestream.h" | |
37 | ||
38 | #ifdef XFS_FILESTREAMS_TRACE | |
39 | ||
40 | ktrace_t *xfs_filestreams_trace_buf; | |
41 | ||
42 | STATIC void | |
43 | xfs_filestreams_trace( | |
44 | xfs_mount_t *mp, /* mount point */ | |
45 | int type, /* type of trace */ | |
46 | const char *func, /* source function */ | |
47 | int line, /* source line number */ | |
48 | __psunsigned_t arg0, | |
49 | __psunsigned_t arg1, | |
50 | __psunsigned_t arg2, | |
51 | __psunsigned_t arg3, | |
52 | __psunsigned_t arg4, | |
53 | __psunsigned_t arg5) | |
54 | { | |
55 | ktrace_enter(xfs_filestreams_trace_buf, | |
56 | (void *)(__psint_t)(type | (line << 16)), | |
57 | (void *)func, | |
58 | (void *)(__psunsigned_t)current_pid(), | |
59 | (void *)mp, | |
60 | (void *)(__psunsigned_t)arg0, | |
61 | (void *)(__psunsigned_t)arg1, | |
62 | (void *)(__psunsigned_t)arg2, | |
63 | (void *)(__psunsigned_t)arg3, | |
64 | (void *)(__psunsigned_t)arg4, | |
65 | (void *)(__psunsigned_t)arg5, | |
66 | NULL, NULL, NULL, NULL, NULL, NULL); | |
67 | } | |
68 | ||
69 | #define TRACE0(mp,t) TRACE6(mp,t,0,0,0,0,0,0) | |
70 | #define TRACE1(mp,t,a0) TRACE6(mp,t,a0,0,0,0,0,0) | |
71 | #define TRACE2(mp,t,a0,a1) TRACE6(mp,t,a0,a1,0,0,0,0) | |
72 | #define TRACE3(mp,t,a0,a1,a2) TRACE6(mp,t,a0,a1,a2,0,0,0) | |
73 | #define TRACE4(mp,t,a0,a1,a2,a3) TRACE6(mp,t,a0,a1,a2,a3,0,0) | |
74 | #define TRACE5(mp,t,a0,a1,a2,a3,a4) TRACE6(mp,t,a0,a1,a2,a3,a4,0) | |
75 | #define TRACE6(mp,t,a0,a1,a2,a3,a4,a5) \ | |
34a622b2 | 76 | xfs_filestreams_trace(mp, t, __func__, __LINE__, \ |
2a82b8be DC |
77 | (__psunsigned_t)a0, (__psunsigned_t)a1, \ |
78 | (__psunsigned_t)a2, (__psunsigned_t)a3, \ | |
79 | (__psunsigned_t)a4, (__psunsigned_t)a5) | |
80 | ||
81 | #define TRACE_AG_SCAN(mp, ag, ag2) \ | |
82 | TRACE2(mp, XFS_FSTRM_KTRACE_AGSCAN, ag, ag2); | |
83 | #define TRACE_AG_PICK1(mp, max_ag, maxfree) \ | |
84 | TRACE2(mp, XFS_FSTRM_KTRACE_AGPICK1, max_ag, maxfree); | |
85 | #define TRACE_AG_PICK2(mp, ag, ag2, cnt, free, scan, flag) \ | |
86 | TRACE6(mp, XFS_FSTRM_KTRACE_AGPICK2, ag, ag2, \ | |
87 | cnt, free, scan, flag) | |
88 | #define TRACE_UPDATE(mp, ip, ag, cnt, ag2, cnt2) \ | |
89 | TRACE5(mp, XFS_FSTRM_KTRACE_UPDATE, ip, ag, cnt, ag2, cnt2) | |
90 | #define TRACE_FREE(mp, ip, pip, ag, cnt) \ | |
91 | TRACE4(mp, XFS_FSTRM_KTRACE_FREE, ip, pip, ag, cnt) | |
92 | #define TRACE_LOOKUP(mp, ip, pip, ag, cnt) \ | |
93 | TRACE4(mp, XFS_FSTRM_KTRACE_ITEM_LOOKUP, ip, pip, ag, cnt) | |
94 | #define TRACE_ASSOCIATE(mp, ip, pip, ag, cnt) \ | |
95 | TRACE4(mp, XFS_FSTRM_KTRACE_ASSOCIATE, ip, pip, ag, cnt) | |
96 | #define TRACE_MOVEAG(mp, ip, pip, oag, ocnt, nag, ncnt) \ | |
97 | TRACE6(mp, XFS_FSTRM_KTRACE_MOVEAG, ip, pip, oag, ocnt, nag, ncnt) | |
98 | #define TRACE_ORPHAN(mp, ip, ag) \ | |
99 | TRACE2(mp, XFS_FSTRM_KTRACE_ORPHAN, ip, ag); | |
100 | ||
101 | ||
102 | #else | |
103 | #define TRACE_AG_SCAN(mp, ag, ag2) | |
104 | #define TRACE_AG_PICK1(mp, max_ag, maxfree) | |
105 | #define TRACE_AG_PICK2(mp, ag, ag2, cnt, free, scan, flag) | |
106 | #define TRACE_UPDATE(mp, ip, ag, cnt, ag2, cnt2) | |
107 | #define TRACE_FREE(mp, ip, pip, ag, cnt) | |
108 | #define TRACE_LOOKUP(mp, ip, pip, ag, cnt) | |
109 | #define TRACE_ASSOCIATE(mp, ip, pip, ag, cnt) | |
110 | #define TRACE_MOVEAG(mp, ip, pip, oag, ocnt, nag, ncnt) | |
111 | #define TRACE_ORPHAN(mp, ip, ag) | |
112 | #endif | |
113 | ||
114 | static kmem_zone_t *item_zone; | |
115 | ||
116 | /* | |
117 | * Structure for associating a file or a directory with an allocation group. | |
118 | * The parent directory pointer is only needed for files, but since there will | |
119 | * generally be vastly more files than directories in the cache, using the same | |
120 | * data structure simplifies the code with very little memory overhead. | |
121 | */ | |
122 | typedef struct fstrm_item | |
123 | { | |
124 | xfs_agnumber_t ag; /* AG currently in use for the file/directory. */ | |
125 | xfs_inode_t *ip; /* inode self-pointer. */ | |
126 | xfs_inode_t *pip; /* Parent directory inode pointer. */ | |
127 | } fstrm_item_t; | |
128 | ||
129 | ||
130 | /* | |
131 | * Scan the AGs starting at startag looking for an AG that isn't in use and has | |
132 | * at least minlen blocks free. | |
133 | */ | |
134 | static int | |
135 | _xfs_filestream_pick_ag( | |
136 | xfs_mount_t *mp, | |
137 | xfs_agnumber_t startag, | |
138 | xfs_agnumber_t *agp, | |
139 | int flags, | |
140 | xfs_extlen_t minlen) | |
141 | { | |
142 | int err, trylock, nscan; | |
6cc87645 | 143 | xfs_extlen_t longest, free, minfree, maxfree = 0; |
2a82b8be DC |
144 | xfs_agnumber_t ag, max_ag = NULLAGNUMBER; |
145 | struct xfs_perag *pag; | |
146 | ||
147 | /* 2% of an AG's blocks must be free for it to be chosen. */ | |
148 | minfree = mp->m_sb.sb_agblocks / 50; | |
149 | ||
150 | ag = startag; | |
151 | *agp = NULLAGNUMBER; | |
152 | ||
153 | /* For the first pass, don't sleep trying to init the per-AG. */ | |
154 | trylock = XFS_ALLOC_FLAG_TRYLOCK; | |
155 | ||
156 | for (nscan = 0; 1; nscan++) { | |
157 | ||
158 | TRACE_AG_SCAN(mp, ag, xfs_filestream_peek_ag(mp, ag)); | |
159 | ||
160 | pag = mp->m_perag + ag; | |
161 | ||
162 | if (!pag->pagf_init) { | |
163 | err = xfs_alloc_pagf_init(mp, NULL, ag, trylock); | |
164 | if (err && !trylock) | |
165 | return err; | |
166 | } | |
167 | ||
168 | /* Might fail sometimes during the 1st pass with trylock set. */ | |
169 | if (!pag->pagf_init) | |
170 | goto next_ag; | |
171 | ||
172 | /* Keep track of the AG with the most free blocks. */ | |
173 | if (pag->pagf_freeblks > maxfree) { | |
174 | maxfree = pag->pagf_freeblks; | |
175 | max_ag = ag; | |
176 | } | |
177 | ||
178 | /* | |
179 | * The AG reference count does two things: it enforces mutual | |
180 | * exclusion when examining the suitability of an AG in this | |
181 | * loop, and it guards against two filestreams being established | |
182 | * in the same AG as each other. | |
183 | */ | |
184 | if (xfs_filestream_get_ag(mp, ag) > 1) { | |
185 | xfs_filestream_put_ag(mp, ag); | |
186 | goto next_ag; | |
187 | } | |
188 | ||
6cc87645 | 189 | longest = xfs_alloc_longest_free_extent(mp, pag); |
2a82b8be DC |
190 | if (((minlen && longest >= minlen) || |
191 | (!minlen && pag->pagf_freeblks >= minfree)) && | |
192 | (!pag->pagf_metadata || !(flags & XFS_PICK_USERDATA) || | |
193 | (flags & XFS_PICK_LOWSPACE))) { | |
194 | ||
195 | /* Break out, retaining the reference on the AG. */ | |
196 | free = pag->pagf_freeblks; | |
197 | *agp = ag; | |
198 | break; | |
199 | } | |
200 | ||
201 | /* Drop the reference on this AG, it's not usable. */ | |
202 | xfs_filestream_put_ag(mp, ag); | |
203 | next_ag: | |
204 | /* Move to the next AG, wrapping to AG 0 if necessary. */ | |
205 | if (++ag >= mp->m_sb.sb_agcount) | |
206 | ag = 0; | |
207 | ||
208 | /* If a full pass of the AGs hasn't been done yet, continue. */ | |
209 | if (ag != startag) | |
210 | continue; | |
211 | ||
212 | /* Allow sleeping in xfs_alloc_pagf_init() on the 2nd pass. */ | |
213 | if (trylock != 0) { | |
214 | trylock = 0; | |
215 | continue; | |
216 | } | |
217 | ||
218 | /* Finally, if lowspace wasn't set, set it for the 3rd pass. */ | |
219 | if (!(flags & XFS_PICK_LOWSPACE)) { | |
220 | flags |= XFS_PICK_LOWSPACE; | |
221 | continue; | |
222 | } | |
223 | ||
224 | /* | |
225 | * Take the AG with the most free space, regardless of whether | |
226 | * it's already in use by another filestream. | |
227 | */ | |
228 | if (max_ag != NULLAGNUMBER) { | |
229 | xfs_filestream_get_ag(mp, max_ag); | |
230 | TRACE_AG_PICK1(mp, max_ag, maxfree); | |
231 | free = maxfree; | |
232 | *agp = max_ag; | |
233 | break; | |
234 | } | |
235 | ||
236 | /* take AG 0 if none matched */ | |
237 | TRACE_AG_PICK1(mp, max_ag, maxfree); | |
238 | *agp = 0; | |
239 | return 0; | |
240 | } | |
241 | ||
242 | TRACE_AG_PICK2(mp, startag, *agp, xfs_filestream_peek_ag(mp, *agp), | |
243 | free, nscan, flags); | |
244 | ||
245 | return 0; | |
246 | } | |
247 | ||
248 | /* | |
249 | * Set the allocation group number for a file or a directory, updating inode | |
250 | * references and per-AG references as appropriate. Must be called with the | |
251 | * m_peraglock held in read mode. | |
252 | */ | |
253 | static int | |
254 | _xfs_filestream_update_ag( | |
255 | xfs_inode_t *ip, | |
256 | xfs_inode_t *pip, | |
257 | xfs_agnumber_t ag) | |
258 | { | |
259 | int err = 0; | |
260 | xfs_mount_t *mp; | |
261 | xfs_mru_cache_t *cache; | |
262 | fstrm_item_t *item; | |
263 | xfs_agnumber_t old_ag; | |
264 | xfs_inode_t *old_pip; | |
265 | ||
266 | /* | |
267 | * Either ip is a regular file and pip is a directory, or ip is a | |
268 | * directory and pip is NULL. | |
269 | */ | |
270 | ASSERT(ip && (((ip->i_d.di_mode & S_IFREG) && pip && | |
271 | (pip->i_d.di_mode & S_IFDIR)) || | |
272 | ((ip->i_d.di_mode & S_IFDIR) && !pip))); | |
273 | ||
274 | mp = ip->i_mount; | |
275 | cache = mp->m_filestream; | |
276 | ||
277 | item = xfs_mru_cache_lookup(cache, ip->i_ino); | |
278 | if (item) { | |
279 | ASSERT(item->ip == ip); | |
280 | old_ag = item->ag; | |
281 | item->ag = ag; | |
282 | old_pip = item->pip; | |
283 | item->pip = pip; | |
284 | xfs_mru_cache_done(cache); | |
285 | ||
286 | /* | |
287 | * If the AG has changed, drop the old ref and take a new one, | |
288 | * effectively transferring the reference from old to new AG. | |
289 | */ | |
290 | if (ag != old_ag) { | |
291 | xfs_filestream_put_ag(mp, old_ag); | |
292 | xfs_filestream_get_ag(mp, ag); | |
293 | } | |
294 | ||
295 | /* | |
296 | * If ip is a file and its pip has changed, drop the old ref and | |
297 | * take a new one. | |
298 | */ | |
299 | if (pip && pip != old_pip) { | |
300 | IRELE(old_pip); | |
301 | IHOLD(pip); | |
302 | } | |
303 | ||
304 | TRACE_UPDATE(mp, ip, old_ag, xfs_filestream_peek_ag(mp, old_ag), | |
305 | ag, xfs_filestream_peek_ag(mp, ag)); | |
306 | return 0; | |
307 | } | |
308 | ||
309 | item = kmem_zone_zalloc(item_zone, KM_MAYFAIL); | |
310 | if (!item) | |
311 | return ENOMEM; | |
312 | ||
313 | item->ag = ag; | |
314 | item->ip = ip; | |
315 | item->pip = pip; | |
316 | ||
317 | err = xfs_mru_cache_insert(cache, ip->i_ino, item); | |
318 | if (err) { | |
319 | kmem_zone_free(item_zone, item); | |
320 | return err; | |
321 | } | |
322 | ||
323 | /* Take a reference on the AG. */ | |
324 | xfs_filestream_get_ag(mp, ag); | |
325 | ||
326 | /* | |
327 | * Take a reference on the inode itself regardless of whether it's a | |
328 | * regular file or a directory. | |
329 | */ | |
330 | IHOLD(ip); | |
331 | ||
332 | /* | |
333 | * In the case of a regular file, take a reference on the parent inode | |
334 | * as well to ensure it remains in-core. | |
335 | */ | |
336 | if (pip) | |
337 | IHOLD(pip); | |
338 | ||
339 | TRACE_UPDATE(mp, ip, ag, xfs_filestream_peek_ag(mp, ag), | |
340 | ag, xfs_filestream_peek_ag(mp, ag)); | |
341 | ||
342 | return 0; | |
343 | } | |
344 | ||
345 | /* xfs_fstrm_free_func(): callback for freeing cached stream items. */ | |
a8272ce0 | 346 | STATIC void |
2a82b8be | 347 | xfs_fstrm_free_func( |
bcc7b445 ES |
348 | unsigned long ino, |
349 | void *data) | |
2a82b8be | 350 | { |
bcc7b445 | 351 | fstrm_item_t *item = (fstrm_item_t *)data; |
2a82b8be DC |
352 | xfs_inode_t *ip = item->ip; |
353 | int ref; | |
354 | ||
355 | ASSERT(ip->i_ino == ino); | |
356 | ||
357 | xfs_iflags_clear(ip, XFS_IFILESTREAM); | |
358 | ||
359 | /* Drop the reference taken on the AG when the item was added. */ | |
360 | ref = xfs_filestream_put_ag(ip->i_mount, item->ag); | |
361 | ||
362 | ASSERT(ref >= 0); | |
363 | TRACE_FREE(ip->i_mount, ip, item->pip, item->ag, | |
364 | xfs_filestream_peek_ag(ip->i_mount, item->ag)); | |
365 | ||
366 | /* | |
367 | * _xfs_filestream_update_ag() always takes a reference on the inode | |
368 | * itself, whether it's a file or a directory. Release it here. | |
369 | * This can result in the inode being freed and so we must | |
370 | * not hold any inode locks when freeing filesstreams objects | |
371 | * otherwise we can deadlock here. | |
372 | */ | |
373 | IRELE(ip); | |
374 | ||
375 | /* | |
376 | * In the case of a regular file, _xfs_filestream_update_ag() also | |
377 | * takes a ref on the parent inode to keep it in-core. Release that | |
378 | * too. | |
379 | */ | |
380 | if (item->pip) | |
381 | IRELE(item->pip); | |
382 | ||
383 | /* Finally, free the memory allocated for the item. */ | |
384 | kmem_zone_free(item_zone, item); | |
385 | } | |
386 | ||
387 | /* | |
388 | * xfs_filestream_init() is called at xfs initialisation time to set up the | |
389 | * memory zone that will be used for filestream data structure allocation. | |
390 | */ | |
391 | int | |
392 | xfs_filestream_init(void) | |
393 | { | |
394 | item_zone = kmem_zone_init(sizeof(fstrm_item_t), "fstrm_item"); | |
9f8868ff CH |
395 | if (!item_zone) |
396 | return -ENOMEM; | |
2a82b8be | 397 | #ifdef XFS_FILESTREAMS_TRACE |
5695ef46 | 398 | xfs_filestreams_trace_buf = ktrace_alloc(XFS_FSTRM_KTRACE_SIZE, KM_NOFS); |
2a82b8be | 399 | #endif |
9f8868ff | 400 | return 0; |
2a82b8be DC |
401 | } |
402 | ||
403 | /* | |
404 | * xfs_filestream_uninit() is called at xfs termination time to destroy the | |
405 | * memory zone that was used for filestream data structure allocation. | |
406 | */ | |
407 | void | |
408 | xfs_filestream_uninit(void) | |
409 | { | |
410 | #ifdef XFS_FILESTREAMS_TRACE | |
411 | ktrace_free(xfs_filestreams_trace_buf); | |
412 | #endif | |
413 | kmem_zone_destroy(item_zone); | |
414 | } | |
415 | ||
416 | /* | |
417 | * xfs_filestream_mount() is called when a file system is mounted with the | |
418 | * filestream option. It is responsible for allocating the data structures | |
419 | * needed to track the new file system's file streams. | |
420 | */ | |
421 | int | |
422 | xfs_filestream_mount( | |
423 | xfs_mount_t *mp) | |
424 | { | |
425 | int err; | |
426 | unsigned int lifetime, grp_count; | |
427 | ||
428 | /* | |
429 | * The filestream timer tunable is currently fixed within the range of | |
430 | * one second to four minutes, with five seconds being the default. The | |
431 | * group count is somewhat arbitrary, but it'd be nice to adhere to the | |
432 | * timer tunable to within about 10 percent. This requires at least 10 | |
433 | * groups. | |
434 | */ | |
435 | lifetime = xfs_fstrm_centisecs * 10; | |
436 | grp_count = 10; | |
437 | ||
438 | err = xfs_mru_cache_create(&mp->m_filestream, lifetime, grp_count, | |
bcc7b445 | 439 | xfs_fstrm_free_func); |
2a82b8be DC |
440 | |
441 | return err; | |
442 | } | |
443 | ||
444 | /* | |
445 | * xfs_filestream_unmount() is called when a file system that was mounted with | |
446 | * the filestream option is unmounted. It drains the data structures created | |
447 | * to track the file system's file streams and frees all the memory that was | |
448 | * allocated. | |
449 | */ | |
450 | void | |
451 | xfs_filestream_unmount( | |
452 | xfs_mount_t *mp) | |
453 | { | |
454 | xfs_mru_cache_destroy(mp->m_filestream); | |
455 | } | |
456 | ||
457 | /* | |
458 | * If the mount point's m_perag array is going to be reallocated, all | |
459 | * outstanding cache entries must be flushed to avoid accessing reference count | |
460 | * addresses that have been freed. The call to xfs_filestream_flush() must be | |
461 | * made inside the block that holds the m_peraglock in write mode to do the | |
462 | * reallocation. | |
463 | */ | |
464 | void | |
465 | xfs_filestream_flush( | |
466 | xfs_mount_t *mp) | |
467 | { | |
65de5567 | 468 | xfs_mru_cache_flush(mp->m_filestream); |
2a82b8be DC |
469 | } |
470 | ||
471 | /* | |
472 | * Return the AG of the filestream the file or directory belongs to, or | |
473 | * NULLAGNUMBER otherwise. | |
474 | */ | |
475 | xfs_agnumber_t | |
476 | xfs_filestream_lookup_ag( | |
477 | xfs_inode_t *ip) | |
478 | { | |
479 | xfs_mru_cache_t *cache; | |
480 | fstrm_item_t *item; | |
481 | xfs_agnumber_t ag; | |
482 | int ref; | |
483 | ||
484 | if (!(ip->i_d.di_mode & (S_IFREG | S_IFDIR))) { | |
485 | ASSERT(0); | |
486 | return NULLAGNUMBER; | |
487 | } | |
488 | ||
489 | cache = ip->i_mount->m_filestream; | |
490 | item = xfs_mru_cache_lookup(cache, ip->i_ino); | |
491 | if (!item) { | |
492 | TRACE_LOOKUP(ip->i_mount, ip, NULL, NULLAGNUMBER, 0); | |
493 | return NULLAGNUMBER; | |
494 | } | |
495 | ||
496 | ASSERT(ip == item->ip); | |
497 | ag = item->ag; | |
498 | ref = xfs_filestream_peek_ag(ip->i_mount, ag); | |
499 | xfs_mru_cache_done(cache); | |
500 | ||
501 | TRACE_LOOKUP(ip->i_mount, ip, item->pip, ag, ref); | |
502 | return ag; | |
503 | } | |
504 | ||
505 | /* | |
506 | * xfs_filestream_associate() should only be called to associate a regular file | |
507 | * with its parent directory. Calling it with a child directory isn't | |
508 | * appropriate because filestreams don't apply to entire directory hierarchies. | |
509 | * Creating a file in a child directory of an existing filestream directory | |
510 | * starts a new filestream with its own allocation group association. | |
511 | * | |
512 | * Returns < 0 on error, 0 if successful association occurred, > 0 if | |
513 | * we failed to get an association because of locking issues. | |
514 | */ | |
515 | int | |
516 | xfs_filestream_associate( | |
517 | xfs_inode_t *pip, | |
518 | xfs_inode_t *ip) | |
519 | { | |
520 | xfs_mount_t *mp; | |
521 | xfs_mru_cache_t *cache; | |
522 | fstrm_item_t *item; | |
523 | xfs_agnumber_t ag, rotorstep, startag; | |
524 | int err = 0; | |
525 | ||
526 | ASSERT(pip->i_d.di_mode & S_IFDIR); | |
527 | ASSERT(ip->i_d.di_mode & S_IFREG); | |
528 | if (!(pip->i_d.di_mode & S_IFDIR) || !(ip->i_d.di_mode & S_IFREG)) | |
529 | return -EINVAL; | |
530 | ||
531 | mp = pip->i_mount; | |
532 | cache = mp->m_filestream; | |
533 | down_read(&mp->m_peraglock); | |
534 | ||
535 | /* | |
536 | * We have a problem, Houston. | |
537 | * | |
538 | * Taking the iolock here violates inode locking order - we already | |
539 | * hold the ilock. Hence if we block getting this lock we may never | |
540 | * wake. Unfortunately, that means if we can't get the lock, we're | |
541 | * screwed in terms of getting a stream association - we can't spin | |
542 | * waiting for the lock because someone else is waiting on the lock we | |
543 | * hold and we cannot drop that as we are in a transaction here. | |
544 | * | |
075fe102 CH |
545 | * Lucky for us, this inversion is not a problem because it's a |
546 | * directory inode that we are trying to lock here. | |
2a82b8be DC |
547 | * |
548 | * So, if we can't get the iolock without sleeping then just give up | |
549 | */ | |
550 | if (!xfs_ilock_nowait(pip, XFS_IOLOCK_EXCL)) { | |
551 | up_read(&mp->m_peraglock); | |
552 | return 1; | |
553 | } | |
554 | ||
555 | /* If the parent directory is already in the cache, use its AG. */ | |
556 | item = xfs_mru_cache_lookup(cache, pip->i_ino); | |
557 | if (item) { | |
558 | ASSERT(item->ip == pip); | |
559 | ag = item->ag; | |
560 | xfs_mru_cache_done(cache); | |
561 | ||
562 | TRACE_LOOKUP(mp, pip, pip, ag, xfs_filestream_peek_ag(mp, ag)); | |
563 | err = _xfs_filestream_update_ag(ip, pip, ag); | |
564 | ||
565 | goto exit; | |
566 | } | |
567 | ||
568 | /* | |
569 | * Set the starting AG using the rotor for inode32, otherwise | |
570 | * use the directory inode's AG. | |
571 | */ | |
572 | if (mp->m_flags & XFS_MOUNT_32BITINODES) { | |
573 | rotorstep = xfs_rotorstep; | |
574 | startag = (mp->m_agfrotor / rotorstep) % mp->m_sb.sb_agcount; | |
575 | mp->m_agfrotor = (mp->m_agfrotor + 1) % | |
576 | (mp->m_sb.sb_agcount * rotorstep); | |
577 | } else | |
578 | startag = XFS_INO_TO_AGNO(mp, pip->i_ino); | |
579 | ||
580 | /* Pick a new AG for the parent inode starting at startag. */ | |
581 | err = _xfs_filestream_pick_ag(mp, startag, &ag, 0, 0); | |
582 | if (err || ag == NULLAGNUMBER) | |
583 | goto exit_did_pick; | |
584 | ||
585 | /* Associate the parent inode with the AG. */ | |
586 | err = _xfs_filestream_update_ag(pip, NULL, ag); | |
587 | if (err) | |
588 | goto exit_did_pick; | |
589 | ||
590 | /* Associate the file inode with the AG. */ | |
591 | err = _xfs_filestream_update_ag(ip, pip, ag); | |
592 | if (err) | |
593 | goto exit_did_pick; | |
594 | ||
595 | TRACE_ASSOCIATE(mp, ip, pip, ag, xfs_filestream_peek_ag(mp, ag)); | |
596 | ||
597 | exit_did_pick: | |
598 | /* | |
599 | * If _xfs_filestream_pick_ag() returned a valid AG, remove the | |
600 | * reference it took on it, since the file and directory will have taken | |
601 | * their own now if they were successfully cached. | |
602 | */ | |
603 | if (ag != NULLAGNUMBER) | |
604 | xfs_filestream_put_ag(mp, ag); | |
605 | ||
606 | exit: | |
607 | xfs_iunlock(pip, XFS_IOLOCK_EXCL); | |
608 | up_read(&mp->m_peraglock); | |
609 | return -err; | |
610 | } | |
611 | ||
612 | /* | |
613 | * Pick a new allocation group for the current file and its file stream. This | |
614 | * function is called by xfs_bmap_filestreams() with the mount point's per-ag | |
615 | * lock held. | |
616 | */ | |
617 | int | |
618 | xfs_filestream_new_ag( | |
619 | xfs_bmalloca_t *ap, | |
620 | xfs_agnumber_t *agp) | |
621 | { | |
622 | int flags, err; | |
623 | xfs_inode_t *ip, *pip = NULL; | |
624 | xfs_mount_t *mp; | |
625 | xfs_mru_cache_t *cache; | |
626 | xfs_extlen_t minlen; | |
627 | fstrm_item_t *dir, *file; | |
628 | xfs_agnumber_t ag = NULLAGNUMBER; | |
629 | ||
630 | ip = ap->ip; | |
631 | mp = ip->i_mount; | |
632 | cache = mp->m_filestream; | |
633 | minlen = ap->alen; | |
634 | *agp = NULLAGNUMBER; | |
635 | ||
636 | /* | |
637 | * Look for the file in the cache, removing it if it's found. Doing | |
638 | * this allows it to be held across the dir lookup that follows. | |
639 | */ | |
640 | file = xfs_mru_cache_remove(cache, ip->i_ino); | |
641 | if (file) { | |
642 | ASSERT(ip == file->ip); | |
643 | ||
644 | /* Save the file's parent inode and old AG number for later. */ | |
645 | pip = file->pip; | |
646 | ag = file->ag; | |
647 | ||
648 | /* Look for the file's directory in the cache. */ | |
649 | dir = xfs_mru_cache_lookup(cache, pip->i_ino); | |
650 | if (dir) { | |
651 | ASSERT(pip == dir->ip); | |
652 | ||
653 | /* | |
654 | * If the directory has already moved on to a new AG, | |
655 | * use that AG as the new AG for the file. Don't | |
656 | * forget to twiddle the AG refcounts to match the | |
657 | * movement. | |
658 | */ | |
659 | if (dir->ag != file->ag) { | |
660 | xfs_filestream_put_ag(mp, file->ag); | |
661 | xfs_filestream_get_ag(mp, dir->ag); | |
662 | *agp = file->ag = dir->ag; | |
663 | } | |
664 | ||
665 | xfs_mru_cache_done(cache); | |
666 | } | |
667 | ||
668 | /* | |
669 | * Put the file back in the cache. If this fails, the free | |
670 | * function needs to be called to tidy up in the same way as if | |
671 | * the item had simply expired from the cache. | |
672 | */ | |
673 | err = xfs_mru_cache_insert(cache, ip->i_ino, file); | |
674 | if (err) { | |
675 | xfs_fstrm_free_func(ip->i_ino, file); | |
676 | return err; | |
677 | } | |
678 | ||
679 | /* | |
680 | * If the file's AG was moved to the directory's new AG, there's | |
681 | * nothing more to be done. | |
682 | */ | |
683 | if (*agp != NULLAGNUMBER) { | |
684 | TRACE_MOVEAG(mp, ip, pip, | |
685 | ag, xfs_filestream_peek_ag(mp, ag), | |
686 | *agp, xfs_filestream_peek_ag(mp, *agp)); | |
687 | return 0; | |
688 | } | |
689 | } | |
690 | ||
691 | /* | |
692 | * If the file's parent directory is known, take its iolock in exclusive | |
693 | * mode to prevent two sibling files from racing each other to migrate | |
694 | * themselves and their parent to different AGs. | |
695 | */ | |
696 | if (pip) | |
697 | xfs_ilock(pip, XFS_IOLOCK_EXCL); | |
698 | ||
699 | /* | |
700 | * A new AG needs to be found for the file. If the file's parent | |
701 | * directory is also known, it will be moved to the new AG as well to | |
702 | * ensure that files created inside it in future use the new AG. | |
703 | */ | |
704 | ag = (ag == NULLAGNUMBER) ? 0 : (ag + 1) % mp->m_sb.sb_agcount; | |
705 | flags = (ap->userdata ? XFS_PICK_USERDATA : 0) | | |
706 | (ap->low ? XFS_PICK_LOWSPACE : 0); | |
707 | ||
708 | err = _xfs_filestream_pick_ag(mp, ag, agp, flags, minlen); | |
709 | if (err || *agp == NULLAGNUMBER) | |
710 | goto exit; | |
711 | ||
712 | /* | |
713 | * If the file wasn't found in the file cache, then its parent directory | |
714 | * inode isn't known. For this to have happened, the file must either | |
715 | * be pre-existing, or it was created long enough ago that its cache | |
716 | * entry has expired. This isn't the sort of usage that the filestreams | |
717 | * allocator is trying to optimise, so there's no point trying to track | |
718 | * its new AG somehow in the filestream data structures. | |
719 | */ | |
720 | if (!pip) { | |
721 | TRACE_ORPHAN(mp, ip, *agp); | |
722 | goto exit; | |
723 | } | |
724 | ||
725 | /* Associate the parent inode with the AG. */ | |
726 | err = _xfs_filestream_update_ag(pip, NULL, *agp); | |
727 | if (err) | |
728 | goto exit; | |
729 | ||
730 | /* Associate the file inode with the AG. */ | |
731 | err = _xfs_filestream_update_ag(ip, pip, *agp); | |
732 | if (err) | |
733 | goto exit; | |
734 | ||
735 | TRACE_MOVEAG(mp, ip, pip, NULLAGNUMBER, 0, | |
736 | *agp, xfs_filestream_peek_ag(mp, *agp)); | |
737 | ||
738 | exit: | |
739 | /* | |
740 | * If _xfs_filestream_pick_ag() returned a valid AG, remove the | |
741 | * reference it took on it, since the file and directory will have taken | |
742 | * their own now if they were successfully cached. | |
743 | */ | |
744 | if (*agp != NULLAGNUMBER) | |
745 | xfs_filestream_put_ag(mp, *agp); | |
746 | else | |
747 | *agp = 0; | |
748 | ||
749 | if (pip) | |
750 | xfs_iunlock(pip, XFS_IOLOCK_EXCL); | |
751 | ||
752 | return err; | |
753 | } | |
754 | ||
755 | /* | |
756 | * Remove an association between an inode and a filestream object. | |
757 | * Typically this is done on last close of an unlinked file. | |
758 | */ | |
759 | void | |
760 | xfs_filestream_deassociate( | |
761 | xfs_inode_t *ip) | |
762 | { | |
763 | xfs_mru_cache_t *cache = ip->i_mount->m_filestream; | |
764 | ||
765 | xfs_mru_cache_delete(cache, ip->i_ino); | |
766 | } |