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" | |
0b1b213f | 37 | #include "xfs_trace.h" |
2a82b8be DC |
38 | |
39 | #ifdef XFS_FILESTREAMS_TRACE | |
40 | ||
41 | ktrace_t *xfs_filestreams_trace_buf; | |
42 | ||
43 | STATIC void | |
44 | xfs_filestreams_trace( | |
45 | xfs_mount_t *mp, /* mount point */ | |
46 | int type, /* type of trace */ | |
47 | const char *func, /* source function */ | |
48 | int line, /* source line number */ | |
49 | __psunsigned_t arg0, | |
50 | __psunsigned_t arg1, | |
51 | __psunsigned_t arg2, | |
52 | __psunsigned_t arg3, | |
53 | __psunsigned_t arg4, | |
54 | __psunsigned_t arg5) | |
55 | { | |
56 | ktrace_enter(xfs_filestreams_trace_buf, | |
57 | (void *)(__psint_t)(type | (line << 16)), | |
58 | (void *)func, | |
59 | (void *)(__psunsigned_t)current_pid(), | |
60 | (void *)mp, | |
61 | (void *)(__psunsigned_t)arg0, | |
62 | (void *)(__psunsigned_t)arg1, | |
63 | (void *)(__psunsigned_t)arg2, | |
64 | (void *)(__psunsigned_t)arg3, | |
65 | (void *)(__psunsigned_t)arg4, | |
66 | (void *)(__psunsigned_t)arg5, | |
67 | NULL, NULL, NULL, NULL, NULL, NULL); | |
68 | } | |
69 | ||
70 | #define TRACE0(mp,t) TRACE6(mp,t,0,0,0,0,0,0) | |
71 | #define TRACE1(mp,t,a0) TRACE6(mp,t,a0,0,0,0,0,0) | |
72 | #define TRACE2(mp,t,a0,a1) TRACE6(mp,t,a0,a1,0,0,0,0) | |
73 | #define TRACE3(mp,t,a0,a1,a2) TRACE6(mp,t,a0,a1,a2,0,0,0) | |
74 | #define TRACE4(mp,t,a0,a1,a2,a3) TRACE6(mp,t,a0,a1,a2,a3,0,0) | |
75 | #define TRACE5(mp,t,a0,a1,a2,a3,a4) TRACE6(mp,t,a0,a1,a2,a3,a4,0) | |
76 | #define TRACE6(mp,t,a0,a1,a2,a3,a4,a5) \ | |
34a622b2 | 77 | xfs_filestreams_trace(mp, t, __func__, __LINE__, \ |
2a82b8be DC |
78 | (__psunsigned_t)a0, (__psunsigned_t)a1, \ |
79 | (__psunsigned_t)a2, (__psunsigned_t)a3, \ | |
80 | (__psunsigned_t)a4, (__psunsigned_t)a5) | |
81 | ||
82 | #define TRACE_AG_SCAN(mp, ag, ag2) \ | |
83 | TRACE2(mp, XFS_FSTRM_KTRACE_AGSCAN, ag, ag2); | |
84 | #define TRACE_AG_PICK1(mp, max_ag, maxfree) \ | |
85 | TRACE2(mp, XFS_FSTRM_KTRACE_AGPICK1, max_ag, maxfree); | |
86 | #define TRACE_AG_PICK2(mp, ag, ag2, cnt, free, scan, flag) \ | |
87 | TRACE6(mp, XFS_FSTRM_KTRACE_AGPICK2, ag, ag2, \ | |
88 | cnt, free, scan, flag) | |
89 | #define TRACE_UPDATE(mp, ip, ag, cnt, ag2, cnt2) \ | |
90 | TRACE5(mp, XFS_FSTRM_KTRACE_UPDATE, ip, ag, cnt, ag2, cnt2) | |
91 | #define TRACE_FREE(mp, ip, pip, ag, cnt) \ | |
92 | TRACE4(mp, XFS_FSTRM_KTRACE_FREE, ip, pip, ag, cnt) | |
93 | #define TRACE_LOOKUP(mp, ip, pip, ag, cnt) \ | |
94 | TRACE4(mp, XFS_FSTRM_KTRACE_ITEM_LOOKUP, ip, pip, ag, cnt) | |
95 | #define TRACE_ASSOCIATE(mp, ip, pip, ag, cnt) \ | |
96 | TRACE4(mp, XFS_FSTRM_KTRACE_ASSOCIATE, ip, pip, ag, cnt) | |
97 | #define TRACE_MOVEAG(mp, ip, pip, oag, ocnt, nag, ncnt) \ | |
98 | TRACE6(mp, XFS_FSTRM_KTRACE_MOVEAG, ip, pip, oag, ocnt, nag, ncnt) | |
99 | #define TRACE_ORPHAN(mp, ip, ag) \ | |
100 | TRACE2(mp, XFS_FSTRM_KTRACE_ORPHAN, ip, ag); | |
101 | ||
102 | ||
103 | #else | |
104 | #define TRACE_AG_SCAN(mp, ag, ag2) | |
105 | #define TRACE_AG_PICK1(mp, max_ag, maxfree) | |
106 | #define TRACE_AG_PICK2(mp, ag, ag2, cnt, free, scan, flag) | |
107 | #define TRACE_UPDATE(mp, ip, ag, cnt, ag2, cnt2) | |
108 | #define TRACE_FREE(mp, ip, pip, ag, cnt) | |
109 | #define TRACE_LOOKUP(mp, ip, pip, ag, cnt) | |
110 | #define TRACE_ASSOCIATE(mp, ip, pip, ag, cnt) | |
111 | #define TRACE_MOVEAG(mp, ip, pip, oag, ocnt, nag, ncnt) | |
112 | #define TRACE_ORPHAN(mp, ip, ag) | |
113 | #endif | |
114 | ||
115 | static kmem_zone_t *item_zone; | |
116 | ||
117 | /* | |
118 | * Structure for associating a file or a directory with an allocation group. | |
119 | * The parent directory pointer is only needed for files, but since there will | |
120 | * generally be vastly more files than directories in the cache, using the same | |
121 | * data structure simplifies the code with very little memory overhead. | |
122 | */ | |
123 | typedef struct fstrm_item | |
124 | { | |
125 | xfs_agnumber_t ag; /* AG currently in use for the file/directory. */ | |
126 | xfs_inode_t *ip; /* inode self-pointer. */ | |
127 | xfs_inode_t *pip; /* Parent directory inode pointer. */ | |
128 | } fstrm_item_t; | |
129 | ||
130 | ||
131 | /* | |
132 | * Scan the AGs starting at startag looking for an AG that isn't in use and has | |
133 | * at least minlen blocks free. | |
134 | */ | |
135 | static int | |
136 | _xfs_filestream_pick_ag( | |
137 | xfs_mount_t *mp, | |
138 | xfs_agnumber_t startag, | |
139 | xfs_agnumber_t *agp, | |
140 | int flags, | |
141 | xfs_extlen_t minlen) | |
142 | { | |
143 | int err, trylock, nscan; | |
6cc87645 | 144 | xfs_extlen_t longest, free, minfree, maxfree = 0; |
2a82b8be DC |
145 | xfs_agnumber_t ag, max_ag = NULLAGNUMBER; |
146 | struct xfs_perag *pag; | |
147 | ||
148 | /* 2% of an AG's blocks must be free for it to be chosen. */ | |
149 | minfree = mp->m_sb.sb_agblocks / 50; | |
150 | ||
151 | ag = startag; | |
152 | *agp = NULLAGNUMBER; | |
153 | ||
154 | /* For the first pass, don't sleep trying to init the per-AG. */ | |
155 | trylock = XFS_ALLOC_FLAG_TRYLOCK; | |
156 | ||
157 | for (nscan = 0; 1; nscan++) { | |
158 | ||
159 | TRACE_AG_SCAN(mp, ag, xfs_filestream_peek_ag(mp, ag)); | |
160 | ||
161 | pag = mp->m_perag + ag; | |
162 | ||
163 | if (!pag->pagf_init) { | |
164 | err = xfs_alloc_pagf_init(mp, NULL, ag, trylock); | |
165 | if (err && !trylock) | |
166 | return err; | |
167 | } | |
168 | ||
169 | /* Might fail sometimes during the 1st pass with trylock set. */ | |
170 | if (!pag->pagf_init) | |
171 | goto next_ag; | |
172 | ||
173 | /* Keep track of the AG with the most free blocks. */ | |
174 | if (pag->pagf_freeblks > maxfree) { | |
175 | maxfree = pag->pagf_freeblks; | |
176 | max_ag = ag; | |
177 | } | |
178 | ||
179 | /* | |
180 | * The AG reference count does two things: it enforces mutual | |
181 | * exclusion when examining the suitability of an AG in this | |
182 | * loop, and it guards against two filestreams being established | |
183 | * in the same AG as each other. | |
184 | */ | |
185 | if (xfs_filestream_get_ag(mp, ag) > 1) { | |
186 | xfs_filestream_put_ag(mp, ag); | |
187 | goto next_ag; | |
188 | } | |
189 | ||
6cc87645 | 190 | longest = xfs_alloc_longest_free_extent(mp, pag); |
2a82b8be DC |
191 | if (((minlen && longest >= minlen) || |
192 | (!minlen && pag->pagf_freeblks >= minfree)) && | |
193 | (!pag->pagf_metadata || !(flags & XFS_PICK_USERDATA) || | |
194 | (flags & XFS_PICK_LOWSPACE))) { | |
195 | ||
196 | /* Break out, retaining the reference on the AG. */ | |
197 | free = pag->pagf_freeblks; | |
198 | *agp = ag; | |
199 | break; | |
200 | } | |
201 | ||
202 | /* Drop the reference on this AG, it's not usable. */ | |
203 | xfs_filestream_put_ag(mp, ag); | |
204 | next_ag: | |
205 | /* Move to the next AG, wrapping to AG 0 if necessary. */ | |
206 | if (++ag >= mp->m_sb.sb_agcount) | |
207 | ag = 0; | |
208 | ||
209 | /* If a full pass of the AGs hasn't been done yet, continue. */ | |
210 | if (ag != startag) | |
211 | continue; | |
212 | ||
213 | /* Allow sleeping in xfs_alloc_pagf_init() on the 2nd pass. */ | |
214 | if (trylock != 0) { | |
215 | trylock = 0; | |
216 | continue; | |
217 | } | |
218 | ||
219 | /* Finally, if lowspace wasn't set, set it for the 3rd pass. */ | |
220 | if (!(flags & XFS_PICK_LOWSPACE)) { | |
221 | flags |= XFS_PICK_LOWSPACE; | |
222 | continue; | |
223 | } | |
224 | ||
225 | /* | |
226 | * Take the AG with the most free space, regardless of whether | |
227 | * it's already in use by another filestream. | |
228 | */ | |
229 | if (max_ag != NULLAGNUMBER) { | |
230 | xfs_filestream_get_ag(mp, max_ag); | |
231 | TRACE_AG_PICK1(mp, max_ag, maxfree); | |
232 | free = maxfree; | |
233 | *agp = max_ag; | |
234 | break; | |
235 | } | |
236 | ||
237 | /* take AG 0 if none matched */ | |
238 | TRACE_AG_PICK1(mp, max_ag, maxfree); | |
239 | *agp = 0; | |
240 | return 0; | |
241 | } | |
242 | ||
243 | TRACE_AG_PICK2(mp, startag, *agp, xfs_filestream_peek_ag(mp, *agp), | |
244 | free, nscan, flags); | |
245 | ||
246 | return 0; | |
247 | } | |
248 | ||
249 | /* | |
250 | * Set the allocation group number for a file or a directory, updating inode | |
251 | * references and per-AG references as appropriate. Must be called with the | |
252 | * m_peraglock held in read mode. | |
253 | */ | |
254 | static int | |
255 | _xfs_filestream_update_ag( | |
256 | xfs_inode_t *ip, | |
257 | xfs_inode_t *pip, | |
258 | xfs_agnumber_t ag) | |
259 | { | |
260 | int err = 0; | |
261 | xfs_mount_t *mp; | |
262 | xfs_mru_cache_t *cache; | |
263 | fstrm_item_t *item; | |
264 | xfs_agnumber_t old_ag; | |
265 | xfs_inode_t *old_pip; | |
266 | ||
267 | /* | |
268 | * Either ip is a regular file and pip is a directory, or ip is a | |
269 | * directory and pip is NULL. | |
270 | */ | |
271 | ASSERT(ip && (((ip->i_d.di_mode & S_IFREG) && pip && | |
272 | (pip->i_d.di_mode & S_IFDIR)) || | |
273 | ((ip->i_d.di_mode & S_IFDIR) && !pip))); | |
274 | ||
275 | mp = ip->i_mount; | |
276 | cache = mp->m_filestream; | |
277 | ||
278 | item = xfs_mru_cache_lookup(cache, ip->i_ino); | |
279 | if (item) { | |
280 | ASSERT(item->ip == ip); | |
281 | old_ag = item->ag; | |
282 | item->ag = ag; | |
283 | old_pip = item->pip; | |
284 | item->pip = pip; | |
285 | xfs_mru_cache_done(cache); | |
286 | ||
287 | /* | |
288 | * If the AG has changed, drop the old ref and take a new one, | |
289 | * effectively transferring the reference from old to new AG. | |
290 | */ | |
291 | if (ag != old_ag) { | |
292 | xfs_filestream_put_ag(mp, old_ag); | |
293 | xfs_filestream_get_ag(mp, ag); | |
294 | } | |
295 | ||
296 | /* | |
297 | * If ip is a file and its pip has changed, drop the old ref and | |
298 | * take a new one. | |
299 | */ | |
300 | if (pip && pip != old_pip) { | |
301 | IRELE(old_pip); | |
302 | IHOLD(pip); | |
303 | } | |
304 | ||
305 | TRACE_UPDATE(mp, ip, old_ag, xfs_filestream_peek_ag(mp, old_ag), | |
306 | ag, xfs_filestream_peek_ag(mp, ag)); | |
307 | return 0; | |
308 | } | |
309 | ||
310 | item = kmem_zone_zalloc(item_zone, KM_MAYFAIL); | |
311 | if (!item) | |
312 | return ENOMEM; | |
313 | ||
314 | item->ag = ag; | |
315 | item->ip = ip; | |
316 | item->pip = pip; | |
317 | ||
318 | err = xfs_mru_cache_insert(cache, ip->i_ino, item); | |
319 | if (err) { | |
320 | kmem_zone_free(item_zone, item); | |
321 | return err; | |
322 | } | |
323 | ||
324 | /* Take a reference on the AG. */ | |
325 | xfs_filestream_get_ag(mp, ag); | |
326 | ||
327 | /* | |
328 | * Take a reference on the inode itself regardless of whether it's a | |
329 | * regular file or a directory. | |
330 | */ | |
331 | IHOLD(ip); | |
332 | ||
333 | /* | |
334 | * In the case of a regular file, take a reference on the parent inode | |
335 | * as well to ensure it remains in-core. | |
336 | */ | |
337 | if (pip) | |
338 | IHOLD(pip); | |
339 | ||
340 | TRACE_UPDATE(mp, ip, ag, xfs_filestream_peek_ag(mp, ag), | |
341 | ag, xfs_filestream_peek_ag(mp, ag)); | |
342 | ||
343 | return 0; | |
344 | } | |
345 | ||
346 | /* xfs_fstrm_free_func(): callback for freeing cached stream items. */ | |
a8272ce0 | 347 | STATIC void |
2a82b8be | 348 | xfs_fstrm_free_func( |
bcc7b445 ES |
349 | unsigned long ino, |
350 | void *data) | |
2a82b8be | 351 | { |
bcc7b445 | 352 | fstrm_item_t *item = (fstrm_item_t *)data; |
2a82b8be DC |
353 | xfs_inode_t *ip = item->ip; |
354 | int ref; | |
355 | ||
356 | ASSERT(ip->i_ino == ino); | |
357 | ||
358 | xfs_iflags_clear(ip, XFS_IFILESTREAM); | |
359 | ||
360 | /* Drop the reference taken on the AG when the item was added. */ | |
361 | ref = xfs_filestream_put_ag(ip->i_mount, item->ag); | |
362 | ||
363 | ASSERT(ref >= 0); | |
364 | TRACE_FREE(ip->i_mount, ip, item->pip, item->ag, | |
365 | xfs_filestream_peek_ag(ip->i_mount, item->ag)); | |
366 | ||
367 | /* | |
368 | * _xfs_filestream_update_ag() always takes a reference on the inode | |
369 | * itself, whether it's a file or a directory. Release it here. | |
370 | * This can result in the inode being freed and so we must | |
371 | * not hold any inode locks when freeing filesstreams objects | |
372 | * otherwise we can deadlock here. | |
373 | */ | |
374 | IRELE(ip); | |
375 | ||
376 | /* | |
377 | * In the case of a regular file, _xfs_filestream_update_ag() also | |
378 | * takes a ref on the parent inode to keep it in-core. Release that | |
379 | * too. | |
380 | */ | |
381 | if (item->pip) | |
382 | IRELE(item->pip); | |
383 | ||
384 | /* Finally, free the memory allocated for the item. */ | |
385 | kmem_zone_free(item_zone, item); | |
386 | } | |
387 | ||
388 | /* | |
389 | * xfs_filestream_init() is called at xfs initialisation time to set up the | |
390 | * memory zone that will be used for filestream data structure allocation. | |
391 | */ | |
392 | int | |
393 | xfs_filestream_init(void) | |
394 | { | |
395 | item_zone = kmem_zone_init(sizeof(fstrm_item_t), "fstrm_item"); | |
9f8868ff CH |
396 | if (!item_zone) |
397 | return -ENOMEM; | |
0b1b213f | 398 | |
9f8868ff | 399 | return 0; |
2a82b8be DC |
400 | } |
401 | ||
402 | /* | |
403 | * xfs_filestream_uninit() is called at xfs termination time to destroy the | |
404 | * memory zone that was used for filestream data structure allocation. | |
405 | */ | |
406 | void | |
407 | xfs_filestream_uninit(void) | |
408 | { | |
2a82b8be DC |
409 | kmem_zone_destroy(item_zone); |
410 | } | |
411 | ||
412 | /* | |
413 | * xfs_filestream_mount() is called when a file system is mounted with the | |
414 | * filestream option. It is responsible for allocating the data structures | |
415 | * needed to track the new file system's file streams. | |
416 | */ | |
417 | int | |
418 | xfs_filestream_mount( | |
419 | xfs_mount_t *mp) | |
420 | { | |
421 | int err; | |
422 | unsigned int lifetime, grp_count; | |
423 | ||
424 | /* | |
425 | * The filestream timer tunable is currently fixed within the range of | |
426 | * one second to four minutes, with five seconds being the default. The | |
427 | * group count is somewhat arbitrary, but it'd be nice to adhere to the | |
428 | * timer tunable to within about 10 percent. This requires at least 10 | |
429 | * groups. | |
430 | */ | |
431 | lifetime = xfs_fstrm_centisecs * 10; | |
432 | grp_count = 10; | |
433 | ||
434 | err = xfs_mru_cache_create(&mp->m_filestream, lifetime, grp_count, | |
bcc7b445 | 435 | xfs_fstrm_free_func); |
2a82b8be DC |
436 | |
437 | return err; | |
438 | } | |
439 | ||
440 | /* | |
441 | * xfs_filestream_unmount() is called when a file system that was mounted with | |
442 | * the filestream option is unmounted. It drains the data structures created | |
443 | * to track the file system's file streams and frees all the memory that was | |
444 | * allocated. | |
445 | */ | |
446 | void | |
447 | xfs_filestream_unmount( | |
448 | xfs_mount_t *mp) | |
449 | { | |
450 | xfs_mru_cache_destroy(mp->m_filestream); | |
451 | } | |
452 | ||
453 | /* | |
454 | * If the mount point's m_perag array is going to be reallocated, all | |
455 | * outstanding cache entries must be flushed to avoid accessing reference count | |
456 | * addresses that have been freed. The call to xfs_filestream_flush() must be | |
457 | * made inside the block that holds the m_peraglock in write mode to do the | |
458 | * reallocation. | |
459 | */ | |
460 | void | |
461 | xfs_filestream_flush( | |
462 | xfs_mount_t *mp) | |
463 | { | |
65de5567 | 464 | xfs_mru_cache_flush(mp->m_filestream); |
2a82b8be DC |
465 | } |
466 | ||
467 | /* | |
468 | * Return the AG of the filestream the file or directory belongs to, or | |
469 | * NULLAGNUMBER otherwise. | |
470 | */ | |
471 | xfs_agnumber_t | |
472 | xfs_filestream_lookup_ag( | |
473 | xfs_inode_t *ip) | |
474 | { | |
475 | xfs_mru_cache_t *cache; | |
476 | fstrm_item_t *item; | |
477 | xfs_agnumber_t ag; | |
478 | int ref; | |
479 | ||
480 | if (!(ip->i_d.di_mode & (S_IFREG | S_IFDIR))) { | |
481 | ASSERT(0); | |
482 | return NULLAGNUMBER; | |
483 | } | |
484 | ||
485 | cache = ip->i_mount->m_filestream; | |
486 | item = xfs_mru_cache_lookup(cache, ip->i_ino); | |
487 | if (!item) { | |
488 | TRACE_LOOKUP(ip->i_mount, ip, NULL, NULLAGNUMBER, 0); | |
489 | return NULLAGNUMBER; | |
490 | } | |
491 | ||
492 | ASSERT(ip == item->ip); | |
493 | ag = item->ag; | |
494 | ref = xfs_filestream_peek_ag(ip->i_mount, ag); | |
495 | xfs_mru_cache_done(cache); | |
496 | ||
497 | TRACE_LOOKUP(ip->i_mount, ip, item->pip, ag, ref); | |
498 | return ag; | |
499 | } | |
500 | ||
501 | /* | |
502 | * xfs_filestream_associate() should only be called to associate a regular file | |
503 | * with its parent directory. Calling it with a child directory isn't | |
504 | * appropriate because filestreams don't apply to entire directory hierarchies. | |
505 | * Creating a file in a child directory of an existing filestream directory | |
506 | * starts a new filestream with its own allocation group association. | |
507 | * | |
508 | * Returns < 0 on error, 0 if successful association occurred, > 0 if | |
509 | * we failed to get an association because of locking issues. | |
510 | */ | |
511 | int | |
512 | xfs_filestream_associate( | |
513 | xfs_inode_t *pip, | |
514 | xfs_inode_t *ip) | |
515 | { | |
516 | xfs_mount_t *mp; | |
517 | xfs_mru_cache_t *cache; | |
518 | fstrm_item_t *item; | |
519 | xfs_agnumber_t ag, rotorstep, startag; | |
520 | int err = 0; | |
521 | ||
522 | ASSERT(pip->i_d.di_mode & S_IFDIR); | |
523 | ASSERT(ip->i_d.di_mode & S_IFREG); | |
524 | if (!(pip->i_d.di_mode & S_IFDIR) || !(ip->i_d.di_mode & S_IFREG)) | |
525 | return -EINVAL; | |
526 | ||
527 | mp = pip->i_mount; | |
528 | cache = mp->m_filestream; | |
529 | down_read(&mp->m_peraglock); | |
530 | ||
531 | /* | |
532 | * We have a problem, Houston. | |
533 | * | |
534 | * Taking the iolock here violates inode locking order - we already | |
535 | * hold the ilock. Hence if we block getting this lock we may never | |
536 | * wake. Unfortunately, that means if we can't get the lock, we're | |
537 | * screwed in terms of getting a stream association - we can't spin | |
538 | * waiting for the lock because someone else is waiting on the lock we | |
539 | * hold and we cannot drop that as we are in a transaction here. | |
540 | * | |
075fe102 CH |
541 | * Lucky for us, this inversion is not a problem because it's a |
542 | * directory inode that we are trying to lock here. | |
2a82b8be DC |
543 | * |
544 | * So, if we can't get the iolock without sleeping then just give up | |
545 | */ | |
546 | if (!xfs_ilock_nowait(pip, XFS_IOLOCK_EXCL)) { | |
547 | up_read(&mp->m_peraglock); | |
548 | return 1; | |
549 | } | |
550 | ||
551 | /* If the parent directory is already in the cache, use its AG. */ | |
552 | item = xfs_mru_cache_lookup(cache, pip->i_ino); | |
553 | if (item) { | |
554 | ASSERT(item->ip == pip); | |
555 | ag = item->ag; | |
556 | xfs_mru_cache_done(cache); | |
557 | ||
558 | TRACE_LOOKUP(mp, pip, pip, ag, xfs_filestream_peek_ag(mp, ag)); | |
559 | err = _xfs_filestream_update_ag(ip, pip, ag); | |
560 | ||
561 | goto exit; | |
562 | } | |
563 | ||
564 | /* | |
565 | * Set the starting AG using the rotor for inode32, otherwise | |
566 | * use the directory inode's AG. | |
567 | */ | |
568 | if (mp->m_flags & XFS_MOUNT_32BITINODES) { | |
569 | rotorstep = xfs_rotorstep; | |
570 | startag = (mp->m_agfrotor / rotorstep) % mp->m_sb.sb_agcount; | |
571 | mp->m_agfrotor = (mp->m_agfrotor + 1) % | |
572 | (mp->m_sb.sb_agcount * rotorstep); | |
573 | } else | |
574 | startag = XFS_INO_TO_AGNO(mp, pip->i_ino); | |
575 | ||
576 | /* Pick a new AG for the parent inode starting at startag. */ | |
577 | err = _xfs_filestream_pick_ag(mp, startag, &ag, 0, 0); | |
578 | if (err || ag == NULLAGNUMBER) | |
579 | goto exit_did_pick; | |
580 | ||
581 | /* Associate the parent inode with the AG. */ | |
582 | err = _xfs_filestream_update_ag(pip, NULL, ag); | |
583 | if (err) | |
584 | goto exit_did_pick; | |
585 | ||
586 | /* Associate the file inode with the AG. */ | |
587 | err = _xfs_filestream_update_ag(ip, pip, ag); | |
588 | if (err) | |
589 | goto exit_did_pick; | |
590 | ||
591 | TRACE_ASSOCIATE(mp, ip, pip, ag, xfs_filestream_peek_ag(mp, ag)); | |
592 | ||
593 | exit_did_pick: | |
594 | /* | |
595 | * If _xfs_filestream_pick_ag() returned a valid AG, remove the | |
596 | * reference it took on it, since the file and directory will have taken | |
597 | * their own now if they were successfully cached. | |
598 | */ | |
599 | if (ag != NULLAGNUMBER) | |
600 | xfs_filestream_put_ag(mp, ag); | |
601 | ||
602 | exit: | |
603 | xfs_iunlock(pip, XFS_IOLOCK_EXCL); | |
604 | up_read(&mp->m_peraglock); | |
605 | return -err; | |
606 | } | |
607 | ||
608 | /* | |
609 | * Pick a new allocation group for the current file and its file stream. This | |
610 | * function is called by xfs_bmap_filestreams() with the mount point's per-ag | |
611 | * lock held. | |
612 | */ | |
613 | int | |
614 | xfs_filestream_new_ag( | |
615 | xfs_bmalloca_t *ap, | |
616 | xfs_agnumber_t *agp) | |
617 | { | |
618 | int flags, err; | |
619 | xfs_inode_t *ip, *pip = NULL; | |
620 | xfs_mount_t *mp; | |
621 | xfs_mru_cache_t *cache; | |
622 | xfs_extlen_t minlen; | |
623 | fstrm_item_t *dir, *file; | |
624 | xfs_agnumber_t ag = NULLAGNUMBER; | |
625 | ||
626 | ip = ap->ip; | |
627 | mp = ip->i_mount; | |
628 | cache = mp->m_filestream; | |
629 | minlen = ap->alen; | |
630 | *agp = NULLAGNUMBER; | |
631 | ||
632 | /* | |
633 | * Look for the file in the cache, removing it if it's found. Doing | |
634 | * this allows it to be held across the dir lookup that follows. | |
635 | */ | |
636 | file = xfs_mru_cache_remove(cache, ip->i_ino); | |
637 | if (file) { | |
638 | ASSERT(ip == file->ip); | |
639 | ||
640 | /* Save the file's parent inode and old AG number for later. */ | |
641 | pip = file->pip; | |
642 | ag = file->ag; | |
643 | ||
644 | /* Look for the file's directory in the cache. */ | |
645 | dir = xfs_mru_cache_lookup(cache, pip->i_ino); | |
646 | if (dir) { | |
647 | ASSERT(pip == dir->ip); | |
648 | ||
649 | /* | |
650 | * If the directory has already moved on to a new AG, | |
651 | * use that AG as the new AG for the file. Don't | |
652 | * forget to twiddle the AG refcounts to match the | |
653 | * movement. | |
654 | */ | |
655 | if (dir->ag != file->ag) { | |
656 | xfs_filestream_put_ag(mp, file->ag); | |
657 | xfs_filestream_get_ag(mp, dir->ag); | |
658 | *agp = file->ag = dir->ag; | |
659 | } | |
660 | ||
661 | xfs_mru_cache_done(cache); | |
662 | } | |
663 | ||
664 | /* | |
665 | * Put the file back in the cache. If this fails, the free | |
666 | * function needs to be called to tidy up in the same way as if | |
667 | * the item had simply expired from the cache. | |
668 | */ | |
669 | err = xfs_mru_cache_insert(cache, ip->i_ino, file); | |
670 | if (err) { | |
671 | xfs_fstrm_free_func(ip->i_ino, file); | |
672 | return err; | |
673 | } | |
674 | ||
675 | /* | |
676 | * If the file's AG was moved to the directory's new AG, there's | |
677 | * nothing more to be done. | |
678 | */ | |
679 | if (*agp != NULLAGNUMBER) { | |
680 | TRACE_MOVEAG(mp, ip, pip, | |
681 | ag, xfs_filestream_peek_ag(mp, ag), | |
682 | *agp, xfs_filestream_peek_ag(mp, *agp)); | |
683 | return 0; | |
684 | } | |
685 | } | |
686 | ||
687 | /* | |
688 | * If the file's parent directory is known, take its iolock in exclusive | |
689 | * mode to prevent two sibling files from racing each other to migrate | |
690 | * themselves and their parent to different AGs. | |
691 | */ | |
692 | if (pip) | |
693 | xfs_ilock(pip, XFS_IOLOCK_EXCL); | |
694 | ||
695 | /* | |
696 | * A new AG needs to be found for the file. If the file's parent | |
697 | * directory is also known, it will be moved to the new AG as well to | |
698 | * ensure that files created inside it in future use the new AG. | |
699 | */ | |
700 | ag = (ag == NULLAGNUMBER) ? 0 : (ag + 1) % mp->m_sb.sb_agcount; | |
701 | flags = (ap->userdata ? XFS_PICK_USERDATA : 0) | | |
702 | (ap->low ? XFS_PICK_LOWSPACE : 0); | |
703 | ||
704 | err = _xfs_filestream_pick_ag(mp, ag, agp, flags, minlen); | |
705 | if (err || *agp == NULLAGNUMBER) | |
706 | goto exit; | |
707 | ||
708 | /* | |
709 | * If the file wasn't found in the file cache, then its parent directory | |
710 | * inode isn't known. For this to have happened, the file must either | |
711 | * be pre-existing, or it was created long enough ago that its cache | |
712 | * entry has expired. This isn't the sort of usage that the filestreams | |
713 | * allocator is trying to optimise, so there's no point trying to track | |
714 | * its new AG somehow in the filestream data structures. | |
715 | */ | |
716 | if (!pip) { | |
717 | TRACE_ORPHAN(mp, ip, *agp); | |
718 | goto exit; | |
719 | } | |
720 | ||
721 | /* Associate the parent inode with the AG. */ | |
722 | err = _xfs_filestream_update_ag(pip, NULL, *agp); | |
723 | if (err) | |
724 | goto exit; | |
725 | ||
726 | /* Associate the file inode with the AG. */ | |
727 | err = _xfs_filestream_update_ag(ip, pip, *agp); | |
728 | if (err) | |
729 | goto exit; | |
730 | ||
731 | TRACE_MOVEAG(mp, ip, pip, NULLAGNUMBER, 0, | |
732 | *agp, xfs_filestream_peek_ag(mp, *agp)); | |
733 | ||
734 | exit: | |
735 | /* | |
736 | * If _xfs_filestream_pick_ag() returned a valid AG, remove the | |
737 | * reference it took on it, since the file and directory will have taken | |
738 | * their own now if they were successfully cached. | |
739 | */ | |
740 | if (*agp != NULLAGNUMBER) | |
741 | xfs_filestream_put_ag(mp, *agp); | |
742 | else | |
743 | *agp = 0; | |
744 | ||
745 | if (pip) | |
746 | xfs_iunlock(pip, XFS_IOLOCK_EXCL); | |
747 | ||
748 | return err; | |
749 | } | |
750 | ||
751 | /* | |
752 | * Remove an association between an inode and a filestream object. | |
753 | * Typically this is done on last close of an unlinked file. | |
754 | */ | |
755 | void | |
756 | xfs_filestream_deassociate( | |
757 | xfs_inode_t *ip) | |
758 | { | |
759 | xfs_mru_cache_t *cache = ip->i_mount->m_filestream; | |
760 | ||
761 | xfs_mru_cache_delete(cache, ip->i_ino); | |
762 | } |