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" | |
2a82b8be DC |
27 | #include "xfs_log.h" |
28 | #include "xfs_trans.h" | |
29 | #include "xfs_sb.h" | |
30 | #include "xfs_mount.h" | |
31 | #include "xfs_bmap.h" | |
32 | #include "xfs_alloc.h" | |
33 | #include "xfs_utils.h" | |
34 | #include "xfs_mru_cache.h" | |
35 | #include "xfs_filestream.h" | |
0b1b213f | 36 | #include "xfs_trace.h" |
2a82b8be DC |
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 | { | |
4196ac08 | 142 | int streams, max_streams; |
2a82b8be | 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++) { | |
4196ac08 DC |
158 | pag = xfs_perag_get(mp, ag); |
159 | TRACE_AG_SCAN(mp, ag, atomic_read(&pag->pagf_fstrms)); | |
2a82b8be DC |
160 | |
161 | if (!pag->pagf_init) { | |
162 | err = xfs_alloc_pagf_init(mp, NULL, ag, trylock); | |
4196ac08 DC |
163 | if (err && !trylock) { |
164 | xfs_perag_put(pag); | |
2a82b8be | 165 | return err; |
4196ac08 | 166 | } |
2a82b8be DC |
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; | |
4196ac08 | 176 | max_streams = atomic_read(&pag->pagf_fstrms); |
2a82b8be DC |
177 | max_ag = ag; |
178 | } | |
179 | ||
180 | /* | |
181 | * The AG reference count does two things: it enforces mutual | |
182 | * exclusion when examining the suitability of an AG in this | |
183 | * loop, and it guards against two filestreams being established | |
184 | * in the same AG as each other. | |
185 | */ | |
186 | if (xfs_filestream_get_ag(mp, ag) > 1) { | |
187 | xfs_filestream_put_ag(mp, ag); | |
188 | goto next_ag; | |
189 | } | |
190 | ||
6cc87645 | 191 | longest = xfs_alloc_longest_free_extent(mp, pag); |
2a82b8be DC |
192 | if (((minlen && longest >= minlen) || |
193 | (!minlen && pag->pagf_freeblks >= minfree)) && | |
194 | (!pag->pagf_metadata || !(flags & XFS_PICK_USERDATA) || | |
195 | (flags & XFS_PICK_LOWSPACE))) { | |
196 | ||
197 | /* Break out, retaining the reference on the AG. */ | |
198 | free = pag->pagf_freeblks; | |
4196ac08 DC |
199 | streams = atomic_read(&pag->pagf_fstrms); |
200 | xfs_perag_put(pag); | |
2a82b8be DC |
201 | *agp = ag; |
202 | break; | |
203 | } | |
204 | ||
205 | /* Drop the reference on this AG, it's not usable. */ | |
206 | xfs_filestream_put_ag(mp, ag); | |
207 | next_ag: | |
4196ac08 | 208 | xfs_perag_put(pag); |
2a82b8be DC |
209 | /* Move to the next AG, wrapping to AG 0 if necessary. */ |
210 | if (++ag >= mp->m_sb.sb_agcount) | |
211 | ag = 0; | |
212 | ||
213 | /* If a full pass of the AGs hasn't been done yet, continue. */ | |
214 | if (ag != startag) | |
215 | continue; | |
216 | ||
217 | /* Allow sleeping in xfs_alloc_pagf_init() on the 2nd pass. */ | |
218 | if (trylock != 0) { | |
219 | trylock = 0; | |
220 | continue; | |
221 | } | |
222 | ||
223 | /* Finally, if lowspace wasn't set, set it for the 3rd pass. */ | |
224 | if (!(flags & XFS_PICK_LOWSPACE)) { | |
225 | flags |= XFS_PICK_LOWSPACE; | |
226 | continue; | |
227 | } | |
228 | ||
229 | /* | |
230 | * Take the AG with the most free space, regardless of whether | |
231 | * it's already in use by another filestream. | |
232 | */ | |
233 | if (max_ag != NULLAGNUMBER) { | |
234 | xfs_filestream_get_ag(mp, max_ag); | |
235 | TRACE_AG_PICK1(mp, max_ag, maxfree); | |
4196ac08 | 236 | streams = max_streams; |
2a82b8be DC |
237 | free = maxfree; |
238 | *agp = max_ag; | |
239 | break; | |
240 | } | |
241 | ||
242 | /* take AG 0 if none matched */ | |
243 | TRACE_AG_PICK1(mp, max_ag, maxfree); | |
244 | *agp = 0; | |
245 | return 0; | |
246 | } | |
247 | ||
4196ac08 | 248 | TRACE_AG_PICK2(mp, startag, *agp, streams, free, nscan, flags); |
2a82b8be DC |
249 | |
250 | return 0; | |
251 | } | |
252 | ||
253 | /* | |
254 | * Set the allocation group number for a file or a directory, updating inode | |
1c1c6ebc | 255 | * references and per-AG references as appropriate. |
2a82b8be DC |
256 | */ |
257 | static int | |
258 | _xfs_filestream_update_ag( | |
259 | xfs_inode_t *ip, | |
260 | xfs_inode_t *pip, | |
261 | xfs_agnumber_t ag) | |
262 | { | |
263 | int err = 0; | |
264 | xfs_mount_t *mp; | |
265 | xfs_mru_cache_t *cache; | |
266 | fstrm_item_t *item; | |
267 | xfs_agnumber_t old_ag; | |
268 | xfs_inode_t *old_pip; | |
269 | ||
270 | /* | |
271 | * Either ip is a regular file and pip is a directory, or ip is a | |
272 | * directory and pip is NULL. | |
273 | */ | |
274 | ASSERT(ip && (((ip->i_d.di_mode & S_IFREG) && pip && | |
275 | (pip->i_d.di_mode & S_IFDIR)) || | |
276 | ((ip->i_d.di_mode & S_IFDIR) && !pip))); | |
277 | ||
278 | mp = ip->i_mount; | |
279 | cache = mp->m_filestream; | |
280 | ||
281 | item = xfs_mru_cache_lookup(cache, ip->i_ino); | |
282 | if (item) { | |
283 | ASSERT(item->ip == ip); | |
284 | old_ag = item->ag; | |
285 | item->ag = ag; | |
286 | old_pip = item->pip; | |
287 | item->pip = pip; | |
288 | xfs_mru_cache_done(cache); | |
289 | ||
290 | /* | |
291 | * If the AG has changed, drop the old ref and take a new one, | |
292 | * effectively transferring the reference from old to new AG. | |
293 | */ | |
294 | if (ag != old_ag) { | |
295 | xfs_filestream_put_ag(mp, old_ag); | |
296 | xfs_filestream_get_ag(mp, ag); | |
297 | } | |
298 | ||
299 | /* | |
300 | * If ip is a file and its pip has changed, drop the old ref and | |
301 | * take a new one. | |
302 | */ | |
303 | if (pip && pip != old_pip) { | |
304 | IRELE(old_pip); | |
305 | IHOLD(pip); | |
306 | } | |
307 | ||
308 | TRACE_UPDATE(mp, ip, old_ag, xfs_filestream_peek_ag(mp, old_ag), | |
309 | ag, xfs_filestream_peek_ag(mp, ag)); | |
310 | return 0; | |
311 | } | |
312 | ||
313 | item = kmem_zone_zalloc(item_zone, KM_MAYFAIL); | |
314 | if (!item) | |
315 | return ENOMEM; | |
316 | ||
317 | item->ag = ag; | |
318 | item->ip = ip; | |
319 | item->pip = pip; | |
320 | ||
321 | err = xfs_mru_cache_insert(cache, ip->i_ino, item); | |
322 | if (err) { | |
323 | kmem_zone_free(item_zone, item); | |
324 | return err; | |
325 | } | |
326 | ||
327 | /* Take a reference on the AG. */ | |
328 | xfs_filestream_get_ag(mp, ag); | |
329 | ||
330 | /* | |
331 | * Take a reference on the inode itself regardless of whether it's a | |
332 | * regular file or a directory. | |
333 | */ | |
334 | IHOLD(ip); | |
335 | ||
336 | /* | |
337 | * In the case of a regular file, take a reference on the parent inode | |
338 | * as well to ensure it remains in-core. | |
339 | */ | |
340 | if (pip) | |
341 | IHOLD(pip); | |
342 | ||
343 | TRACE_UPDATE(mp, ip, ag, xfs_filestream_peek_ag(mp, ag), | |
344 | ag, xfs_filestream_peek_ag(mp, ag)); | |
345 | ||
346 | return 0; | |
347 | } | |
348 | ||
349 | /* xfs_fstrm_free_func(): callback for freeing cached stream items. */ | |
a8272ce0 | 350 | STATIC void |
2a82b8be | 351 | xfs_fstrm_free_func( |
bcc7b445 ES |
352 | unsigned long ino, |
353 | void *data) | |
2a82b8be | 354 | { |
bcc7b445 | 355 | fstrm_item_t *item = (fstrm_item_t *)data; |
2a82b8be DC |
356 | xfs_inode_t *ip = item->ip; |
357 | int ref; | |
358 | ||
359 | ASSERT(ip->i_ino == ino); | |
360 | ||
361 | xfs_iflags_clear(ip, XFS_IFILESTREAM); | |
362 | ||
363 | /* Drop the reference taken on the AG when the item was added. */ | |
364 | ref = xfs_filestream_put_ag(ip->i_mount, item->ag); | |
365 | ||
366 | ASSERT(ref >= 0); | |
367 | TRACE_FREE(ip->i_mount, ip, item->pip, item->ag, | |
368 | xfs_filestream_peek_ag(ip->i_mount, item->ag)); | |
369 | ||
370 | /* | |
371 | * _xfs_filestream_update_ag() always takes a reference on the inode | |
372 | * itself, whether it's a file or a directory. Release it here. | |
373 | * This can result in the inode being freed and so we must | |
374 | * not hold any inode locks when freeing filesstreams objects | |
375 | * otherwise we can deadlock here. | |
376 | */ | |
377 | IRELE(ip); | |
378 | ||
379 | /* | |
380 | * In the case of a regular file, _xfs_filestream_update_ag() also | |
381 | * takes a ref on the parent inode to keep it in-core. Release that | |
382 | * too. | |
383 | */ | |
384 | if (item->pip) | |
385 | IRELE(item->pip); | |
386 | ||
387 | /* Finally, free the memory allocated for the item. */ | |
388 | kmem_zone_free(item_zone, item); | |
389 | } | |
390 | ||
391 | /* | |
392 | * xfs_filestream_init() is called at xfs initialisation time to set up the | |
393 | * memory zone that will be used for filestream data structure allocation. | |
394 | */ | |
395 | int | |
396 | xfs_filestream_init(void) | |
397 | { | |
398 | item_zone = kmem_zone_init(sizeof(fstrm_item_t), "fstrm_item"); | |
9f8868ff CH |
399 | if (!item_zone) |
400 | return -ENOMEM; | |
0b1b213f | 401 | |
9f8868ff | 402 | return 0; |
2a82b8be DC |
403 | } |
404 | ||
405 | /* | |
406 | * xfs_filestream_uninit() is called at xfs termination time to destroy the | |
407 | * memory zone that was used for filestream data structure allocation. | |
408 | */ | |
409 | void | |
410 | xfs_filestream_uninit(void) | |
411 | { | |
2a82b8be DC |
412 | kmem_zone_destroy(item_zone); |
413 | } | |
414 | ||
415 | /* | |
416 | * xfs_filestream_mount() is called when a file system is mounted with the | |
417 | * filestream option. It is responsible for allocating the data structures | |
418 | * needed to track the new file system's file streams. | |
419 | */ | |
420 | int | |
421 | xfs_filestream_mount( | |
422 | xfs_mount_t *mp) | |
423 | { | |
424 | int err; | |
425 | unsigned int lifetime, grp_count; | |
426 | ||
427 | /* | |
428 | * The filestream timer tunable is currently fixed within the range of | |
429 | * one second to four minutes, with five seconds being the default. The | |
430 | * group count is somewhat arbitrary, but it'd be nice to adhere to the | |
431 | * timer tunable to within about 10 percent. This requires at least 10 | |
432 | * groups. | |
433 | */ | |
434 | lifetime = xfs_fstrm_centisecs * 10; | |
435 | grp_count = 10; | |
436 | ||
437 | err = xfs_mru_cache_create(&mp->m_filestream, lifetime, grp_count, | |
bcc7b445 | 438 | xfs_fstrm_free_func); |
2a82b8be DC |
439 | |
440 | return err; | |
441 | } | |
442 | ||
443 | /* | |
444 | * xfs_filestream_unmount() is called when a file system that was mounted with | |
445 | * the filestream option is unmounted. It drains the data structures created | |
446 | * to track the file system's file streams and frees all the memory that was | |
447 | * allocated. | |
448 | */ | |
449 | void | |
450 | xfs_filestream_unmount( | |
451 | xfs_mount_t *mp) | |
452 | { | |
453 | xfs_mru_cache_destroy(mp->m_filestream); | |
454 | } | |
455 | ||
2a82b8be DC |
456 | /* |
457 | * Return the AG of the filestream the file or directory belongs to, or | |
458 | * NULLAGNUMBER otherwise. | |
459 | */ | |
460 | xfs_agnumber_t | |
461 | xfs_filestream_lookup_ag( | |
462 | xfs_inode_t *ip) | |
463 | { | |
464 | xfs_mru_cache_t *cache; | |
465 | fstrm_item_t *item; | |
466 | xfs_agnumber_t ag; | |
467 | int ref; | |
468 | ||
469 | if (!(ip->i_d.di_mode & (S_IFREG | S_IFDIR))) { | |
470 | ASSERT(0); | |
471 | return NULLAGNUMBER; | |
472 | } | |
473 | ||
474 | cache = ip->i_mount->m_filestream; | |
475 | item = xfs_mru_cache_lookup(cache, ip->i_ino); | |
476 | if (!item) { | |
477 | TRACE_LOOKUP(ip->i_mount, ip, NULL, NULLAGNUMBER, 0); | |
478 | return NULLAGNUMBER; | |
479 | } | |
480 | ||
481 | ASSERT(ip == item->ip); | |
482 | ag = item->ag; | |
483 | ref = xfs_filestream_peek_ag(ip->i_mount, ag); | |
484 | xfs_mru_cache_done(cache); | |
485 | ||
486 | TRACE_LOOKUP(ip->i_mount, ip, item->pip, ag, ref); | |
487 | return ag; | |
488 | } | |
489 | ||
490 | /* | |
491 | * xfs_filestream_associate() should only be called to associate a regular file | |
492 | * with its parent directory. Calling it with a child directory isn't | |
493 | * appropriate because filestreams don't apply to entire directory hierarchies. | |
494 | * Creating a file in a child directory of an existing filestream directory | |
495 | * starts a new filestream with its own allocation group association. | |
496 | * | |
497 | * Returns < 0 on error, 0 if successful association occurred, > 0 if | |
498 | * we failed to get an association because of locking issues. | |
499 | */ | |
500 | int | |
501 | xfs_filestream_associate( | |
502 | xfs_inode_t *pip, | |
503 | xfs_inode_t *ip) | |
504 | { | |
505 | xfs_mount_t *mp; | |
506 | xfs_mru_cache_t *cache; | |
507 | fstrm_item_t *item; | |
508 | xfs_agnumber_t ag, rotorstep, startag; | |
509 | int err = 0; | |
510 | ||
511 | ASSERT(pip->i_d.di_mode & S_IFDIR); | |
512 | ASSERT(ip->i_d.di_mode & S_IFREG); | |
513 | if (!(pip->i_d.di_mode & S_IFDIR) || !(ip->i_d.di_mode & S_IFREG)) | |
514 | return -EINVAL; | |
515 | ||
516 | mp = pip->i_mount; | |
517 | cache = mp->m_filestream; | |
2a82b8be DC |
518 | |
519 | /* | |
520 | * We have a problem, Houston. | |
521 | * | |
522 | * Taking the iolock here violates inode locking order - we already | |
523 | * hold the ilock. Hence if we block getting this lock we may never | |
524 | * wake. Unfortunately, that means if we can't get the lock, we're | |
525 | * screwed in terms of getting a stream association - we can't spin | |
526 | * waiting for the lock because someone else is waiting on the lock we | |
527 | * hold and we cannot drop that as we are in a transaction here. | |
528 | * | |
075fe102 CH |
529 | * Lucky for us, this inversion is not a problem because it's a |
530 | * directory inode that we are trying to lock here. | |
2a82b8be DC |
531 | * |
532 | * So, if we can't get the iolock without sleeping then just give up | |
533 | */ | |
1c1c6ebc | 534 | if (!xfs_ilock_nowait(pip, XFS_IOLOCK_EXCL)) |
2a82b8be | 535 | return 1; |
2a82b8be DC |
536 | |
537 | /* If the parent directory is already in the cache, use its AG. */ | |
538 | item = xfs_mru_cache_lookup(cache, pip->i_ino); | |
539 | if (item) { | |
540 | ASSERT(item->ip == pip); | |
541 | ag = item->ag; | |
542 | xfs_mru_cache_done(cache); | |
543 | ||
544 | TRACE_LOOKUP(mp, pip, pip, ag, xfs_filestream_peek_ag(mp, ag)); | |
545 | err = _xfs_filestream_update_ag(ip, pip, ag); | |
546 | ||
547 | goto exit; | |
548 | } | |
549 | ||
550 | /* | |
551 | * Set the starting AG using the rotor for inode32, otherwise | |
552 | * use the directory inode's AG. | |
553 | */ | |
554 | if (mp->m_flags & XFS_MOUNT_32BITINODES) { | |
555 | rotorstep = xfs_rotorstep; | |
556 | startag = (mp->m_agfrotor / rotorstep) % mp->m_sb.sb_agcount; | |
557 | mp->m_agfrotor = (mp->m_agfrotor + 1) % | |
558 | (mp->m_sb.sb_agcount * rotorstep); | |
559 | } else | |
560 | startag = XFS_INO_TO_AGNO(mp, pip->i_ino); | |
561 | ||
562 | /* Pick a new AG for the parent inode starting at startag. */ | |
563 | err = _xfs_filestream_pick_ag(mp, startag, &ag, 0, 0); | |
564 | if (err || ag == NULLAGNUMBER) | |
565 | goto exit_did_pick; | |
566 | ||
567 | /* Associate the parent inode with the AG. */ | |
568 | err = _xfs_filestream_update_ag(pip, NULL, ag); | |
569 | if (err) | |
570 | goto exit_did_pick; | |
571 | ||
572 | /* Associate the file inode with the AG. */ | |
573 | err = _xfs_filestream_update_ag(ip, pip, ag); | |
574 | if (err) | |
575 | goto exit_did_pick; | |
576 | ||
577 | TRACE_ASSOCIATE(mp, ip, pip, ag, xfs_filestream_peek_ag(mp, ag)); | |
578 | ||
579 | exit_did_pick: | |
580 | /* | |
581 | * If _xfs_filestream_pick_ag() returned a valid AG, remove the | |
582 | * reference it took on it, since the file and directory will have taken | |
583 | * their own now if they were successfully cached. | |
584 | */ | |
585 | if (ag != NULLAGNUMBER) | |
586 | xfs_filestream_put_ag(mp, ag); | |
587 | ||
588 | exit: | |
589 | xfs_iunlock(pip, XFS_IOLOCK_EXCL); | |
2a82b8be DC |
590 | return -err; |
591 | } | |
592 | ||
593 | /* | |
594 | * Pick a new allocation group for the current file and its file stream. This | |
595 | * function is called by xfs_bmap_filestreams() with the mount point's per-ag | |
596 | * lock held. | |
597 | */ | |
598 | int | |
599 | xfs_filestream_new_ag( | |
600 | xfs_bmalloca_t *ap, | |
601 | xfs_agnumber_t *agp) | |
602 | { | |
603 | int flags, err; | |
604 | xfs_inode_t *ip, *pip = NULL; | |
605 | xfs_mount_t *mp; | |
606 | xfs_mru_cache_t *cache; | |
607 | xfs_extlen_t minlen; | |
608 | fstrm_item_t *dir, *file; | |
609 | xfs_agnumber_t ag = NULLAGNUMBER; | |
610 | ||
611 | ip = ap->ip; | |
612 | mp = ip->i_mount; | |
613 | cache = mp->m_filestream; | |
614 | minlen = ap->alen; | |
615 | *agp = NULLAGNUMBER; | |
616 | ||
617 | /* | |
618 | * Look for the file in the cache, removing it if it's found. Doing | |
619 | * this allows it to be held across the dir lookup that follows. | |
620 | */ | |
621 | file = xfs_mru_cache_remove(cache, ip->i_ino); | |
622 | if (file) { | |
623 | ASSERT(ip == file->ip); | |
624 | ||
625 | /* Save the file's parent inode and old AG number for later. */ | |
626 | pip = file->pip; | |
627 | ag = file->ag; | |
628 | ||
629 | /* Look for the file's directory in the cache. */ | |
630 | dir = xfs_mru_cache_lookup(cache, pip->i_ino); | |
631 | if (dir) { | |
632 | ASSERT(pip == dir->ip); | |
633 | ||
634 | /* | |
635 | * If the directory has already moved on to a new AG, | |
636 | * use that AG as the new AG for the file. Don't | |
637 | * forget to twiddle the AG refcounts to match the | |
638 | * movement. | |
639 | */ | |
640 | if (dir->ag != file->ag) { | |
641 | xfs_filestream_put_ag(mp, file->ag); | |
642 | xfs_filestream_get_ag(mp, dir->ag); | |
643 | *agp = file->ag = dir->ag; | |
644 | } | |
645 | ||
646 | xfs_mru_cache_done(cache); | |
647 | } | |
648 | ||
649 | /* | |
650 | * Put the file back in the cache. If this fails, the free | |
651 | * function needs to be called to tidy up in the same way as if | |
652 | * the item had simply expired from the cache. | |
653 | */ | |
654 | err = xfs_mru_cache_insert(cache, ip->i_ino, file); | |
655 | if (err) { | |
656 | xfs_fstrm_free_func(ip->i_ino, file); | |
657 | return err; | |
658 | } | |
659 | ||
660 | /* | |
661 | * If the file's AG was moved to the directory's new AG, there's | |
662 | * nothing more to be done. | |
663 | */ | |
664 | if (*agp != NULLAGNUMBER) { | |
665 | TRACE_MOVEAG(mp, ip, pip, | |
666 | ag, xfs_filestream_peek_ag(mp, ag), | |
667 | *agp, xfs_filestream_peek_ag(mp, *agp)); | |
668 | return 0; | |
669 | } | |
670 | } | |
671 | ||
672 | /* | |
673 | * If the file's parent directory is known, take its iolock in exclusive | |
674 | * mode to prevent two sibling files from racing each other to migrate | |
675 | * themselves and their parent to different AGs. | |
676 | */ | |
677 | if (pip) | |
678 | xfs_ilock(pip, XFS_IOLOCK_EXCL); | |
679 | ||
680 | /* | |
681 | * A new AG needs to be found for the file. If the file's parent | |
682 | * directory is also known, it will be moved to the new AG as well to | |
683 | * ensure that files created inside it in future use the new AG. | |
684 | */ | |
685 | ag = (ag == NULLAGNUMBER) ? 0 : (ag + 1) % mp->m_sb.sb_agcount; | |
686 | flags = (ap->userdata ? XFS_PICK_USERDATA : 0) | | |
687 | (ap->low ? XFS_PICK_LOWSPACE : 0); | |
688 | ||
689 | err = _xfs_filestream_pick_ag(mp, ag, agp, flags, minlen); | |
690 | if (err || *agp == NULLAGNUMBER) | |
691 | goto exit; | |
692 | ||
693 | /* | |
694 | * If the file wasn't found in the file cache, then its parent directory | |
695 | * inode isn't known. For this to have happened, the file must either | |
696 | * be pre-existing, or it was created long enough ago that its cache | |
697 | * entry has expired. This isn't the sort of usage that the filestreams | |
698 | * allocator is trying to optimise, so there's no point trying to track | |
699 | * its new AG somehow in the filestream data structures. | |
700 | */ | |
701 | if (!pip) { | |
702 | TRACE_ORPHAN(mp, ip, *agp); | |
703 | goto exit; | |
704 | } | |
705 | ||
706 | /* Associate the parent inode with the AG. */ | |
707 | err = _xfs_filestream_update_ag(pip, NULL, *agp); | |
708 | if (err) | |
709 | goto exit; | |
710 | ||
711 | /* Associate the file inode with the AG. */ | |
712 | err = _xfs_filestream_update_ag(ip, pip, *agp); | |
713 | if (err) | |
714 | goto exit; | |
715 | ||
716 | TRACE_MOVEAG(mp, ip, pip, NULLAGNUMBER, 0, | |
717 | *agp, xfs_filestream_peek_ag(mp, *agp)); | |
718 | ||
719 | exit: | |
720 | /* | |
721 | * If _xfs_filestream_pick_ag() returned a valid AG, remove the | |
722 | * reference it took on it, since the file and directory will have taken | |
723 | * their own now if they were successfully cached. | |
724 | */ | |
725 | if (*agp != NULLAGNUMBER) | |
726 | xfs_filestream_put_ag(mp, *agp); | |
727 | else | |
728 | *agp = 0; | |
729 | ||
730 | if (pip) | |
731 | xfs_iunlock(pip, XFS_IOLOCK_EXCL); | |
732 | ||
733 | return err; | |
734 | } | |
735 | ||
736 | /* | |
737 | * Remove an association between an inode and a filestream object. | |
738 | * Typically this is done on last close of an unlinked file. | |
739 | */ | |
740 | void | |
741 | xfs_filestream_deassociate( | |
742 | xfs_inode_t *ip) | |
743 | { | |
744 | xfs_mru_cache_t *cache = ip->i_mount->m_filestream; | |
745 | ||
746 | xfs_mru_cache_delete(cache, ip->i_ino); | |
747 | } |