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" | |
a4fbe6ab | 19 | #include "xfs_format.h" |
239880ef DC |
20 | #include "xfs_log_format.h" |
21 | #include "xfs_trans_resv.h" | |
22 | #include "xfs_ag.h" | |
23 | #include "xfs_sb.h" | |
24 | #include "xfs_mount.h" | |
2a82b8be | 25 | #include "xfs_inum.h" |
2a82b8be | 26 | #include "xfs_inode.h" |
2a82b8be | 27 | #include "xfs_bmap.h" |
68988114 | 28 | #include "xfs_bmap_util.h" |
2a82b8be | 29 | #include "xfs_alloc.h" |
2a82b8be | 30 | #include "xfs_mru_cache.h" |
a4fbe6ab | 31 | #include "xfs_dinode.h" |
2a82b8be | 32 | #include "xfs_filestream.h" |
0b1b213f | 33 | #include "xfs_trace.h" |
2a82b8be DC |
34 | |
35 | #ifdef XFS_FILESTREAMS_TRACE | |
36 | ||
37 | ktrace_t *xfs_filestreams_trace_buf; | |
38 | ||
39 | STATIC void | |
40 | xfs_filestreams_trace( | |
41 | xfs_mount_t *mp, /* mount point */ | |
42 | int type, /* type of trace */ | |
43 | const char *func, /* source function */ | |
44 | int line, /* source line number */ | |
45 | __psunsigned_t arg0, | |
46 | __psunsigned_t arg1, | |
47 | __psunsigned_t arg2, | |
48 | __psunsigned_t arg3, | |
49 | __psunsigned_t arg4, | |
50 | __psunsigned_t arg5) | |
51 | { | |
52 | ktrace_enter(xfs_filestreams_trace_buf, | |
53 | (void *)(__psint_t)(type | (line << 16)), | |
54 | (void *)func, | |
55 | (void *)(__psunsigned_t)current_pid(), | |
56 | (void *)mp, | |
57 | (void *)(__psunsigned_t)arg0, | |
58 | (void *)(__psunsigned_t)arg1, | |
59 | (void *)(__psunsigned_t)arg2, | |
60 | (void *)(__psunsigned_t)arg3, | |
61 | (void *)(__psunsigned_t)arg4, | |
62 | (void *)(__psunsigned_t)arg5, | |
63 | NULL, NULL, NULL, NULL, NULL, NULL); | |
64 | } | |
65 | ||
66 | #define TRACE0(mp,t) TRACE6(mp,t,0,0,0,0,0,0) | |
67 | #define TRACE1(mp,t,a0) TRACE6(mp,t,a0,0,0,0,0,0) | |
68 | #define TRACE2(mp,t,a0,a1) TRACE6(mp,t,a0,a1,0,0,0,0) | |
69 | #define TRACE3(mp,t,a0,a1,a2) TRACE6(mp,t,a0,a1,a2,0,0,0) | |
70 | #define TRACE4(mp,t,a0,a1,a2,a3) TRACE6(mp,t,a0,a1,a2,a3,0,0) | |
71 | #define TRACE5(mp,t,a0,a1,a2,a3,a4) TRACE6(mp,t,a0,a1,a2,a3,a4,0) | |
72 | #define TRACE6(mp,t,a0,a1,a2,a3,a4,a5) \ | |
34a622b2 | 73 | xfs_filestreams_trace(mp, t, __func__, __LINE__, \ |
2a82b8be DC |
74 | (__psunsigned_t)a0, (__psunsigned_t)a1, \ |
75 | (__psunsigned_t)a2, (__psunsigned_t)a3, \ | |
76 | (__psunsigned_t)a4, (__psunsigned_t)a5) | |
77 | ||
78 | #define TRACE_AG_SCAN(mp, ag, ag2) \ | |
79 | TRACE2(mp, XFS_FSTRM_KTRACE_AGSCAN, ag, ag2); | |
80 | #define TRACE_AG_PICK1(mp, max_ag, maxfree) \ | |
81 | TRACE2(mp, XFS_FSTRM_KTRACE_AGPICK1, max_ag, maxfree); | |
82 | #define TRACE_AG_PICK2(mp, ag, ag2, cnt, free, scan, flag) \ | |
83 | TRACE6(mp, XFS_FSTRM_KTRACE_AGPICK2, ag, ag2, \ | |
84 | cnt, free, scan, flag) | |
85 | #define TRACE_UPDATE(mp, ip, ag, cnt, ag2, cnt2) \ | |
86 | TRACE5(mp, XFS_FSTRM_KTRACE_UPDATE, ip, ag, cnt, ag2, cnt2) | |
87 | #define TRACE_FREE(mp, ip, pip, ag, cnt) \ | |
88 | TRACE4(mp, XFS_FSTRM_KTRACE_FREE, ip, pip, ag, cnt) | |
89 | #define TRACE_LOOKUP(mp, ip, pip, ag, cnt) \ | |
90 | TRACE4(mp, XFS_FSTRM_KTRACE_ITEM_LOOKUP, ip, pip, ag, cnt) | |
91 | #define TRACE_ASSOCIATE(mp, ip, pip, ag, cnt) \ | |
92 | TRACE4(mp, XFS_FSTRM_KTRACE_ASSOCIATE, ip, pip, ag, cnt) | |
93 | #define TRACE_MOVEAG(mp, ip, pip, oag, ocnt, nag, ncnt) \ | |
94 | TRACE6(mp, XFS_FSTRM_KTRACE_MOVEAG, ip, pip, oag, ocnt, nag, ncnt) | |
95 | #define TRACE_ORPHAN(mp, ip, ag) \ | |
96 | TRACE2(mp, XFS_FSTRM_KTRACE_ORPHAN, ip, ag); | |
97 | ||
98 | ||
99 | #else | |
100 | #define TRACE_AG_SCAN(mp, ag, ag2) | |
101 | #define TRACE_AG_PICK1(mp, max_ag, maxfree) | |
102 | #define TRACE_AG_PICK2(mp, ag, ag2, cnt, free, scan, flag) | |
103 | #define TRACE_UPDATE(mp, ip, ag, cnt, ag2, cnt2) | |
104 | #define TRACE_FREE(mp, ip, pip, ag, cnt) | |
105 | #define TRACE_LOOKUP(mp, ip, pip, ag, cnt) | |
106 | #define TRACE_ASSOCIATE(mp, ip, pip, ag, cnt) | |
107 | #define TRACE_MOVEAG(mp, ip, pip, oag, ocnt, nag, ncnt) | |
108 | #define TRACE_ORPHAN(mp, ip, ag) | |
109 | #endif | |
110 | ||
111 | static kmem_zone_t *item_zone; | |
112 | ||
113 | /* | |
114 | * Structure for associating a file or a directory with an allocation group. | |
115 | * The parent directory pointer is only needed for files, but since there will | |
116 | * generally be vastly more files than directories in the cache, using the same | |
117 | * data structure simplifies the code with very little memory overhead. | |
118 | */ | |
119 | typedef struct fstrm_item | |
120 | { | |
22328d71 | 121 | struct xfs_mru_cache_elem mru; |
2a82b8be DC |
122 | xfs_agnumber_t ag; /* AG currently in use for the file/directory. */ |
123 | xfs_inode_t *ip; /* inode self-pointer. */ | |
124 | xfs_inode_t *pip; /* Parent directory inode pointer. */ | |
125 | } fstrm_item_t; | |
126 | ||
0664ce8d CH |
127 | /* |
128 | * Allocation group filestream associations are tracked with per-ag atomic | |
129 | * counters. These counters allow _xfs_filestream_pick_ag() to tell whether a | |
130 | * particular AG already has active filestreams associated with it. The mount | |
131 | * point's m_peraglock is used to protect these counters from per-ag array | |
132 | * re-allocation during a growfs operation. When xfs_growfs_data_private() is | |
133 | * about to reallocate the array, it calls xfs_filestream_flush() with the | |
134 | * m_peraglock held in write mode. | |
135 | * | |
136 | * Since xfs_mru_cache_flush() guarantees that all the free functions for all | |
137 | * the cache elements have finished executing before it returns, it's safe for | |
138 | * the free functions to use the atomic counters without m_peraglock protection. | |
139 | * This allows the implementation of xfs_fstrm_free_func() to be agnostic about | |
140 | * whether it was called with the m_peraglock held in read mode, write mode or | |
141 | * not held at all. The race condition this addresses is the following: | |
142 | * | |
143 | * - The work queue scheduler fires and pulls a filestream directory cache | |
144 | * element off the LRU end of the cache for deletion, then gets pre-empted. | |
145 | * - A growfs operation grabs the m_peraglock in write mode, flushes all the | |
146 | * remaining items from the cache and reallocates the mount point's per-ag | |
147 | * array, resetting all the counters to zero. | |
148 | * - The work queue thread resumes and calls the free function for the element | |
149 | * it started cleaning up earlier. In the process it decrements the | |
150 | * filestreams counter for an AG that now has no references. | |
151 | * | |
152 | * With a shrinkfs feature, the above scenario could panic the system. | |
153 | * | |
154 | * All other uses of the following macros should be protected by either the | |
155 | * m_peraglock held in read mode, or the cache's internal locking exposed by the | |
156 | * interval between a call to xfs_mru_cache_lookup() and a call to | |
157 | * xfs_mru_cache_done(). In addition, the m_peraglock must be held in read mode | |
158 | * when new elements are added to the cache. | |
159 | * | |
160 | * Combined, these locking rules ensure that no associations will ever exist in | |
161 | * the cache that reference per-ag array elements that have since been | |
162 | * reallocated. | |
163 | */ | |
164 | static int | |
165 | xfs_filestream_peek_ag( | |
166 | xfs_mount_t *mp, | |
167 | xfs_agnumber_t agno) | |
168 | { | |
169 | struct xfs_perag *pag; | |
170 | int ret; | |
171 | ||
172 | pag = xfs_perag_get(mp, agno); | |
173 | ret = atomic_read(&pag->pagf_fstrms); | |
174 | xfs_perag_put(pag); | |
175 | return ret; | |
176 | } | |
177 | ||
178 | static int | |
179 | xfs_filestream_get_ag( | |
180 | xfs_mount_t *mp, | |
181 | xfs_agnumber_t agno) | |
182 | { | |
183 | struct xfs_perag *pag; | |
184 | int ret; | |
185 | ||
186 | pag = xfs_perag_get(mp, agno); | |
187 | ret = atomic_inc_return(&pag->pagf_fstrms); | |
188 | xfs_perag_put(pag); | |
189 | return ret; | |
190 | } | |
191 | ||
192 | static void | |
193 | xfs_filestream_put_ag( | |
194 | xfs_mount_t *mp, | |
195 | xfs_agnumber_t agno) | |
196 | { | |
197 | struct xfs_perag *pag; | |
198 | ||
199 | pag = xfs_perag_get(mp, agno); | |
200 | atomic_dec(&pag->pagf_fstrms); | |
201 | xfs_perag_put(pag); | |
202 | } | |
2a82b8be DC |
203 | |
204 | /* | |
205 | * Scan the AGs starting at startag looking for an AG that isn't in use and has | |
206 | * at least minlen blocks free. | |
207 | */ | |
208 | static int | |
209 | _xfs_filestream_pick_ag( | |
210 | xfs_mount_t *mp, | |
211 | xfs_agnumber_t startag, | |
212 | xfs_agnumber_t *agp, | |
213 | int flags, | |
214 | xfs_extlen_t minlen) | |
215 | { | |
4196ac08 | 216 | int streams, max_streams; |
2a82b8be | 217 | int err, trylock, nscan; |
6cc87645 | 218 | xfs_extlen_t longest, free, minfree, maxfree = 0; |
2a82b8be DC |
219 | xfs_agnumber_t ag, max_ag = NULLAGNUMBER; |
220 | struct xfs_perag *pag; | |
221 | ||
222 | /* 2% of an AG's blocks must be free for it to be chosen. */ | |
223 | minfree = mp->m_sb.sb_agblocks / 50; | |
224 | ||
225 | ag = startag; | |
226 | *agp = NULLAGNUMBER; | |
227 | ||
228 | /* For the first pass, don't sleep trying to init the per-AG. */ | |
229 | trylock = XFS_ALLOC_FLAG_TRYLOCK; | |
230 | ||
231 | for (nscan = 0; 1; nscan++) { | |
4196ac08 DC |
232 | pag = xfs_perag_get(mp, ag); |
233 | TRACE_AG_SCAN(mp, ag, atomic_read(&pag->pagf_fstrms)); | |
2a82b8be DC |
234 | |
235 | if (!pag->pagf_init) { | |
236 | err = xfs_alloc_pagf_init(mp, NULL, ag, trylock); | |
4196ac08 DC |
237 | if (err && !trylock) { |
238 | xfs_perag_put(pag); | |
2a82b8be | 239 | return err; |
4196ac08 | 240 | } |
2a82b8be DC |
241 | } |
242 | ||
243 | /* Might fail sometimes during the 1st pass with trylock set. */ | |
244 | if (!pag->pagf_init) | |
245 | goto next_ag; | |
246 | ||
247 | /* Keep track of the AG with the most free blocks. */ | |
248 | if (pag->pagf_freeblks > maxfree) { | |
249 | maxfree = pag->pagf_freeblks; | |
4196ac08 | 250 | max_streams = atomic_read(&pag->pagf_fstrms); |
2a82b8be DC |
251 | max_ag = ag; |
252 | } | |
253 | ||
254 | /* | |
255 | * The AG reference count does two things: it enforces mutual | |
256 | * exclusion when examining the suitability of an AG in this | |
257 | * loop, and it guards against two filestreams being established | |
258 | * in the same AG as each other. | |
259 | */ | |
260 | if (xfs_filestream_get_ag(mp, ag) > 1) { | |
261 | xfs_filestream_put_ag(mp, ag); | |
262 | goto next_ag; | |
263 | } | |
264 | ||
6cc87645 | 265 | longest = xfs_alloc_longest_free_extent(mp, pag); |
2a82b8be DC |
266 | if (((minlen && longest >= minlen) || |
267 | (!minlen && pag->pagf_freeblks >= minfree)) && | |
268 | (!pag->pagf_metadata || !(flags & XFS_PICK_USERDATA) || | |
269 | (flags & XFS_PICK_LOWSPACE))) { | |
270 | ||
271 | /* Break out, retaining the reference on the AG. */ | |
272 | free = pag->pagf_freeblks; | |
4196ac08 DC |
273 | streams = atomic_read(&pag->pagf_fstrms); |
274 | xfs_perag_put(pag); | |
2a82b8be DC |
275 | *agp = ag; |
276 | break; | |
277 | } | |
278 | ||
279 | /* Drop the reference on this AG, it's not usable. */ | |
280 | xfs_filestream_put_ag(mp, ag); | |
281 | next_ag: | |
4196ac08 | 282 | xfs_perag_put(pag); |
2a82b8be DC |
283 | /* Move to the next AG, wrapping to AG 0 if necessary. */ |
284 | if (++ag >= mp->m_sb.sb_agcount) | |
285 | ag = 0; | |
286 | ||
287 | /* If a full pass of the AGs hasn't been done yet, continue. */ | |
288 | if (ag != startag) | |
289 | continue; | |
290 | ||
291 | /* Allow sleeping in xfs_alloc_pagf_init() on the 2nd pass. */ | |
292 | if (trylock != 0) { | |
293 | trylock = 0; | |
294 | continue; | |
295 | } | |
296 | ||
297 | /* Finally, if lowspace wasn't set, set it for the 3rd pass. */ | |
298 | if (!(flags & XFS_PICK_LOWSPACE)) { | |
299 | flags |= XFS_PICK_LOWSPACE; | |
300 | continue; | |
301 | } | |
302 | ||
303 | /* | |
304 | * Take the AG with the most free space, regardless of whether | |
305 | * it's already in use by another filestream. | |
306 | */ | |
307 | if (max_ag != NULLAGNUMBER) { | |
308 | xfs_filestream_get_ag(mp, max_ag); | |
309 | TRACE_AG_PICK1(mp, max_ag, maxfree); | |
4196ac08 | 310 | streams = max_streams; |
2a82b8be DC |
311 | free = maxfree; |
312 | *agp = max_ag; | |
313 | break; | |
314 | } | |
315 | ||
316 | /* take AG 0 if none matched */ | |
317 | TRACE_AG_PICK1(mp, max_ag, maxfree); | |
318 | *agp = 0; | |
319 | return 0; | |
320 | } | |
321 | ||
4196ac08 | 322 | TRACE_AG_PICK2(mp, startag, *agp, streams, free, nscan, flags); |
2a82b8be DC |
323 | |
324 | return 0; | |
325 | } | |
326 | ||
327 | /* | |
328 | * Set the allocation group number for a file or a directory, updating inode | |
1c1c6ebc | 329 | * references and per-AG references as appropriate. |
2a82b8be DC |
330 | */ |
331 | static int | |
332 | _xfs_filestream_update_ag( | |
333 | xfs_inode_t *ip, | |
334 | xfs_inode_t *pip, | |
335 | xfs_agnumber_t ag) | |
336 | { | |
337 | int err = 0; | |
338 | xfs_mount_t *mp; | |
2a82b8be DC |
339 | fstrm_item_t *item; |
340 | xfs_agnumber_t old_ag; | |
341 | xfs_inode_t *old_pip; | |
22328d71 | 342 | struct xfs_mru_cache_elem *mru; |
2a82b8be DC |
343 | |
344 | /* | |
345 | * Either ip is a regular file and pip is a directory, or ip is a | |
346 | * directory and pip is NULL. | |
347 | */ | |
abbede1b | 348 | ASSERT(ip && ((S_ISREG(ip->i_d.di_mode) && pip && |
03209378 AV |
349 | S_ISDIR(pip->i_d.di_mode)) || |
350 | (S_ISDIR(ip->i_d.di_mode) && !pip))); | |
2a82b8be DC |
351 | |
352 | mp = ip->i_mount; | |
2a82b8be | 353 | |
22328d71 CH |
354 | mru = xfs_mru_cache_lookup(mp->m_filestream, ip->i_ino); |
355 | if (mru) { | |
356 | item = container_of(mru, fstrm_item_t, mru); | |
357 | ||
2a82b8be DC |
358 | ASSERT(item->ip == ip); |
359 | old_ag = item->ag; | |
360 | item->ag = ag; | |
361 | old_pip = item->pip; | |
362 | item->pip = pip; | |
22328d71 | 363 | xfs_mru_cache_done(mp->m_filestream); |
2a82b8be DC |
364 | |
365 | /* | |
366 | * If the AG has changed, drop the old ref and take a new one, | |
367 | * effectively transferring the reference from old to new AG. | |
368 | */ | |
369 | if (ag != old_ag) { | |
370 | xfs_filestream_put_ag(mp, old_ag); | |
371 | xfs_filestream_get_ag(mp, ag); | |
372 | } | |
373 | ||
374 | /* | |
375 | * If ip is a file and its pip has changed, drop the old ref and | |
376 | * take a new one. | |
377 | */ | |
378 | if (pip && pip != old_pip) { | |
379 | IRELE(old_pip); | |
380 | IHOLD(pip); | |
381 | } | |
382 | ||
383 | TRACE_UPDATE(mp, ip, old_ag, xfs_filestream_peek_ag(mp, old_ag), | |
384 | ag, xfs_filestream_peek_ag(mp, ag)); | |
385 | return 0; | |
386 | } | |
387 | ||
388 | item = kmem_zone_zalloc(item_zone, KM_MAYFAIL); | |
389 | if (!item) | |
390 | return ENOMEM; | |
391 | ||
392 | item->ag = ag; | |
393 | item->ip = ip; | |
394 | item->pip = pip; | |
395 | ||
22328d71 | 396 | err = xfs_mru_cache_insert(mp->m_filestream, ip->i_ino, &item->mru); |
2a82b8be DC |
397 | if (err) { |
398 | kmem_zone_free(item_zone, item); | |
399 | return err; | |
400 | } | |
401 | ||
402 | /* Take a reference on the AG. */ | |
403 | xfs_filestream_get_ag(mp, ag); | |
404 | ||
405 | /* | |
406 | * Take a reference on the inode itself regardless of whether it's a | |
407 | * regular file or a directory. | |
408 | */ | |
409 | IHOLD(ip); | |
410 | ||
411 | /* | |
412 | * In the case of a regular file, take a reference on the parent inode | |
413 | * as well to ensure it remains in-core. | |
414 | */ | |
415 | if (pip) | |
416 | IHOLD(pip); | |
417 | ||
418 | TRACE_UPDATE(mp, ip, ag, xfs_filestream_peek_ag(mp, ag), | |
419 | ag, xfs_filestream_peek_ag(mp, ag)); | |
420 | ||
421 | return 0; | |
422 | } | |
423 | ||
424 | /* xfs_fstrm_free_func(): callback for freeing cached stream items. */ | |
a8272ce0 | 425 | STATIC void |
2a82b8be | 426 | xfs_fstrm_free_func( |
22328d71 | 427 | struct xfs_mru_cache_elem *mru) |
2a82b8be | 428 | { |
22328d71 CH |
429 | fstrm_item_t *item = |
430 | container_of(mru, fstrm_item_t, mru); | |
2a82b8be | 431 | xfs_inode_t *ip = item->ip; |
2a82b8be | 432 | |
2a82b8be | 433 | /* Drop the reference taken on the AG when the item was added. */ |
0664ce8d | 434 | xfs_filestream_put_ag(ip->i_mount, item->ag); |
2a82b8be | 435 | |
2a82b8be DC |
436 | TRACE_FREE(ip->i_mount, ip, item->pip, item->ag, |
437 | xfs_filestream_peek_ag(ip->i_mount, item->ag)); | |
438 | ||
439 | /* | |
440 | * _xfs_filestream_update_ag() always takes a reference on the inode | |
441 | * itself, whether it's a file or a directory. Release it here. | |
442 | * This can result in the inode being freed and so we must | |
443 | * not hold any inode locks when freeing filesstreams objects | |
444 | * otherwise we can deadlock here. | |
445 | */ | |
446 | IRELE(ip); | |
447 | ||
448 | /* | |
449 | * In the case of a regular file, _xfs_filestream_update_ag() also | |
450 | * takes a ref on the parent inode to keep it in-core. Release that | |
451 | * too. | |
452 | */ | |
453 | if (item->pip) | |
454 | IRELE(item->pip); | |
455 | ||
456 | /* Finally, free the memory allocated for the item. */ | |
457 | kmem_zone_free(item_zone, item); | |
458 | } | |
459 | ||
460 | /* | |
461 | * xfs_filestream_init() is called at xfs initialisation time to set up the | |
462 | * memory zone that will be used for filestream data structure allocation. | |
463 | */ | |
464 | int | |
465 | xfs_filestream_init(void) | |
466 | { | |
467 | item_zone = kmem_zone_init(sizeof(fstrm_item_t), "fstrm_item"); | |
9f8868ff CH |
468 | if (!item_zone) |
469 | return -ENOMEM; | |
0b1b213f | 470 | |
9f8868ff | 471 | return 0; |
2a82b8be DC |
472 | } |
473 | ||
474 | /* | |
475 | * xfs_filestream_uninit() is called at xfs termination time to destroy the | |
476 | * memory zone that was used for filestream data structure allocation. | |
477 | */ | |
478 | void | |
479 | xfs_filestream_uninit(void) | |
480 | { | |
2a82b8be DC |
481 | kmem_zone_destroy(item_zone); |
482 | } | |
483 | ||
484 | /* | |
485 | * xfs_filestream_mount() is called when a file system is mounted with the | |
486 | * filestream option. It is responsible for allocating the data structures | |
487 | * needed to track the new file system's file streams. | |
488 | */ | |
489 | int | |
490 | xfs_filestream_mount( | |
491 | xfs_mount_t *mp) | |
492 | { | |
493 | int err; | |
494 | unsigned int lifetime, grp_count; | |
495 | ||
496 | /* | |
497 | * The filestream timer tunable is currently fixed within the range of | |
498 | * one second to four minutes, with five seconds being the default. The | |
499 | * group count is somewhat arbitrary, but it'd be nice to adhere to the | |
500 | * timer tunable to within about 10 percent. This requires at least 10 | |
501 | * groups. | |
502 | */ | |
503 | lifetime = xfs_fstrm_centisecs * 10; | |
504 | grp_count = 10; | |
505 | ||
506 | err = xfs_mru_cache_create(&mp->m_filestream, lifetime, grp_count, | |
bcc7b445 | 507 | xfs_fstrm_free_func); |
2a82b8be DC |
508 | |
509 | return err; | |
510 | } | |
511 | ||
512 | /* | |
513 | * xfs_filestream_unmount() is called when a file system that was mounted with | |
514 | * the filestream option is unmounted. It drains the data structures created | |
515 | * to track the file system's file streams and frees all the memory that was | |
516 | * allocated. | |
517 | */ | |
518 | void | |
519 | xfs_filestream_unmount( | |
520 | xfs_mount_t *mp) | |
521 | { | |
522 | xfs_mru_cache_destroy(mp->m_filestream); | |
523 | } | |
524 | ||
2a82b8be DC |
525 | /* |
526 | * Return the AG of the filestream the file or directory belongs to, or | |
527 | * NULLAGNUMBER otherwise. | |
528 | */ | |
529 | xfs_agnumber_t | |
530 | xfs_filestream_lookup_ag( | |
531 | xfs_inode_t *ip) | |
532 | { | |
22328d71 CH |
533 | struct xfs_mount *mp = ip->i_mount; |
534 | struct xfs_mru_cache_elem *mru; | |
2a82b8be DC |
535 | fstrm_item_t *item; |
536 | xfs_agnumber_t ag; | |
537 | int ref; | |
538 | ||
03209378 | 539 | if (!S_ISREG(ip->i_d.di_mode) && !S_ISDIR(ip->i_d.di_mode)) { |
2a82b8be DC |
540 | ASSERT(0); |
541 | return NULLAGNUMBER; | |
542 | } | |
543 | ||
22328d71 CH |
544 | mru = xfs_mru_cache_lookup(mp->m_filestream, ip->i_ino); |
545 | if (!mru) { | |
2a82b8be DC |
546 | TRACE_LOOKUP(ip->i_mount, ip, NULL, NULLAGNUMBER, 0); |
547 | return NULLAGNUMBER; | |
548 | } | |
549 | ||
22328d71 | 550 | item = container_of(mru, fstrm_item_t, mru); |
2a82b8be DC |
551 | ASSERT(ip == item->ip); |
552 | ag = item->ag; | |
553 | ref = xfs_filestream_peek_ag(ip->i_mount, ag); | |
22328d71 | 554 | xfs_mru_cache_done(mp->m_filestream); |
2a82b8be DC |
555 | |
556 | TRACE_LOOKUP(ip->i_mount, ip, item->pip, ag, ref); | |
557 | return ag; | |
558 | } | |
559 | ||
560 | /* | |
561 | * xfs_filestream_associate() should only be called to associate a regular file | |
562 | * with its parent directory. Calling it with a child directory isn't | |
563 | * appropriate because filestreams don't apply to entire directory hierarchies. | |
564 | * Creating a file in a child directory of an existing filestream directory | |
565 | * starts a new filestream with its own allocation group association. | |
566 | * | |
567 | * Returns < 0 on error, 0 if successful association occurred, > 0 if | |
568 | * we failed to get an association because of locking issues. | |
569 | */ | |
570 | int | |
571 | xfs_filestream_associate( | |
572 | xfs_inode_t *pip, | |
573 | xfs_inode_t *ip) | |
574 | { | |
22328d71 | 575 | struct xfs_mru_cache_elem *mru; |
2a82b8be | 576 | xfs_mount_t *mp; |
2a82b8be DC |
577 | fstrm_item_t *item; |
578 | xfs_agnumber_t ag, rotorstep, startag; | |
579 | int err = 0; | |
580 | ||
03209378 AV |
581 | ASSERT(S_ISDIR(pip->i_d.di_mode)); |
582 | ASSERT(S_ISREG(ip->i_d.di_mode)); | |
583 | if (!S_ISDIR(pip->i_d.di_mode) || !S_ISREG(ip->i_d.di_mode)) | |
2a82b8be DC |
584 | return -EINVAL; |
585 | ||
586 | mp = pip->i_mount; | |
2a82b8be DC |
587 | |
588 | /* | |
589 | * We have a problem, Houston. | |
590 | * | |
591 | * Taking the iolock here violates inode locking order - we already | |
592 | * hold the ilock. Hence if we block getting this lock we may never | |
593 | * wake. Unfortunately, that means if we can't get the lock, we're | |
594 | * screwed in terms of getting a stream association - we can't spin | |
595 | * waiting for the lock because someone else is waiting on the lock we | |
596 | * hold and we cannot drop that as we are in a transaction here. | |
597 | * | |
075fe102 CH |
598 | * Lucky for us, this inversion is not a problem because it's a |
599 | * directory inode that we are trying to lock here. | |
2a82b8be DC |
600 | * |
601 | * So, if we can't get the iolock without sleeping then just give up | |
602 | */ | |
1c1c6ebc | 603 | if (!xfs_ilock_nowait(pip, XFS_IOLOCK_EXCL)) |
2a82b8be | 604 | return 1; |
2a82b8be DC |
605 | |
606 | /* If the parent directory is already in the cache, use its AG. */ | |
22328d71 CH |
607 | mru = xfs_mru_cache_lookup(mp->m_filestream, pip->i_ino); |
608 | if (mru) { | |
609 | item = container_of(mru, fstrm_item_t, mru); | |
610 | ||
2a82b8be DC |
611 | ASSERT(item->ip == pip); |
612 | ag = item->ag; | |
22328d71 | 613 | xfs_mru_cache_done(mp->m_filestream); |
2a82b8be DC |
614 | |
615 | TRACE_LOOKUP(mp, pip, pip, ag, xfs_filestream_peek_ag(mp, ag)); | |
616 | err = _xfs_filestream_update_ag(ip, pip, ag); | |
617 | ||
618 | goto exit; | |
619 | } | |
620 | ||
621 | /* | |
622 | * Set the starting AG using the rotor for inode32, otherwise | |
623 | * use the directory inode's AG. | |
624 | */ | |
625 | if (mp->m_flags & XFS_MOUNT_32BITINODES) { | |
626 | rotorstep = xfs_rotorstep; | |
627 | startag = (mp->m_agfrotor / rotorstep) % mp->m_sb.sb_agcount; | |
628 | mp->m_agfrotor = (mp->m_agfrotor + 1) % | |
629 | (mp->m_sb.sb_agcount * rotorstep); | |
630 | } else | |
631 | startag = XFS_INO_TO_AGNO(mp, pip->i_ino); | |
632 | ||
633 | /* Pick a new AG for the parent inode starting at startag. */ | |
634 | err = _xfs_filestream_pick_ag(mp, startag, &ag, 0, 0); | |
635 | if (err || ag == NULLAGNUMBER) | |
636 | goto exit_did_pick; | |
637 | ||
638 | /* Associate the parent inode with the AG. */ | |
639 | err = _xfs_filestream_update_ag(pip, NULL, ag); | |
640 | if (err) | |
641 | goto exit_did_pick; | |
642 | ||
643 | /* Associate the file inode with the AG. */ | |
644 | err = _xfs_filestream_update_ag(ip, pip, ag); | |
645 | if (err) | |
646 | goto exit_did_pick; | |
647 | ||
648 | TRACE_ASSOCIATE(mp, ip, pip, ag, xfs_filestream_peek_ag(mp, ag)); | |
649 | ||
650 | exit_did_pick: | |
651 | /* | |
652 | * If _xfs_filestream_pick_ag() returned a valid AG, remove the | |
653 | * reference it took on it, since the file and directory will have taken | |
654 | * their own now if they were successfully cached. | |
655 | */ | |
656 | if (ag != NULLAGNUMBER) | |
657 | xfs_filestream_put_ag(mp, ag); | |
658 | ||
659 | exit: | |
660 | xfs_iunlock(pip, XFS_IOLOCK_EXCL); | |
2a82b8be DC |
661 | return -err; |
662 | } | |
663 | ||
664 | /* | |
665 | * Pick a new allocation group for the current file and its file stream. This | |
666 | * function is called by xfs_bmap_filestreams() with the mount point's per-ag | |
667 | * lock held. | |
668 | */ | |
669 | int | |
670 | xfs_filestream_new_ag( | |
68988114 DC |
671 | struct xfs_bmalloca *ap, |
672 | xfs_agnumber_t *agp) | |
2a82b8be | 673 | { |
22328d71 | 674 | struct xfs_mru_cache_elem *mru, *mru2; |
2a82b8be DC |
675 | int flags, err; |
676 | xfs_inode_t *ip, *pip = NULL; | |
677 | xfs_mount_t *mp; | |
2a82b8be DC |
678 | xfs_extlen_t minlen; |
679 | fstrm_item_t *dir, *file; | |
680 | xfs_agnumber_t ag = NULLAGNUMBER; | |
681 | ||
682 | ip = ap->ip; | |
683 | mp = ip->i_mount; | |
3a75667e | 684 | minlen = ap->length; |
2a82b8be DC |
685 | *agp = NULLAGNUMBER; |
686 | ||
687 | /* | |
688 | * Look for the file in the cache, removing it if it's found. Doing | |
689 | * this allows it to be held across the dir lookup that follows. | |
690 | */ | |
22328d71 CH |
691 | mru = xfs_mru_cache_remove(mp->m_filestream, ip->i_ino); |
692 | if (mru) { | |
693 | file = container_of(mru, fstrm_item_t, mru); | |
2a82b8be DC |
694 | ASSERT(ip == file->ip); |
695 | ||
696 | /* Save the file's parent inode and old AG number for later. */ | |
697 | pip = file->pip; | |
698 | ag = file->ag; | |
699 | ||
700 | /* Look for the file's directory in the cache. */ | |
22328d71 CH |
701 | mru2 = xfs_mru_cache_lookup(mp->m_filestream, pip->i_ino); |
702 | if (mru2) { | |
703 | dir = container_of(mru2, fstrm_item_t, mru); | |
2a82b8be DC |
704 | ASSERT(pip == dir->ip); |
705 | ||
706 | /* | |
707 | * If the directory has already moved on to a new AG, | |
708 | * use that AG as the new AG for the file. Don't | |
709 | * forget to twiddle the AG refcounts to match the | |
710 | * movement. | |
711 | */ | |
712 | if (dir->ag != file->ag) { | |
713 | xfs_filestream_put_ag(mp, file->ag); | |
714 | xfs_filestream_get_ag(mp, dir->ag); | |
715 | *agp = file->ag = dir->ag; | |
716 | } | |
717 | ||
22328d71 | 718 | xfs_mru_cache_done(mp->m_filestream); |
2a82b8be DC |
719 | } |
720 | ||
721 | /* | |
722 | * Put the file back in the cache. If this fails, the free | |
723 | * function needs to be called to tidy up in the same way as if | |
724 | * the item had simply expired from the cache. | |
725 | */ | |
22328d71 | 726 | err = xfs_mru_cache_insert(mp->m_filestream, ip->i_ino, mru); |
2a82b8be | 727 | if (err) { |
22328d71 | 728 | xfs_fstrm_free_func(mru); |
2a82b8be DC |
729 | return err; |
730 | } | |
731 | ||
732 | /* | |
733 | * If the file's AG was moved to the directory's new AG, there's | |
734 | * nothing more to be done. | |
735 | */ | |
736 | if (*agp != NULLAGNUMBER) { | |
737 | TRACE_MOVEAG(mp, ip, pip, | |
738 | ag, xfs_filestream_peek_ag(mp, ag), | |
739 | *agp, xfs_filestream_peek_ag(mp, *agp)); | |
740 | return 0; | |
741 | } | |
742 | } | |
743 | ||
744 | /* | |
745 | * If the file's parent directory is known, take its iolock in exclusive | |
746 | * mode to prevent two sibling files from racing each other to migrate | |
747 | * themselves and their parent to different AGs. | |
785ce418 CH |
748 | * |
749 | * Note that we lock the parent directory iolock inside the child | |
750 | * iolock here. That's fine as we never hold both parent and child | |
751 | * iolock in any other place. This is different from the ilock, | |
752 | * which requires locking of the child after the parent for namespace | |
753 | * operations. | |
2a82b8be DC |
754 | */ |
755 | if (pip) | |
785ce418 | 756 | xfs_ilock(pip, XFS_IOLOCK_EXCL | XFS_IOLOCK_PARENT); |
2a82b8be DC |
757 | |
758 | /* | |
759 | * A new AG needs to be found for the file. If the file's parent | |
760 | * directory is also known, it will be moved to the new AG as well to | |
761 | * ensure that files created inside it in future use the new AG. | |
762 | */ | |
763 | ag = (ag == NULLAGNUMBER) ? 0 : (ag + 1) % mp->m_sb.sb_agcount; | |
764 | flags = (ap->userdata ? XFS_PICK_USERDATA : 0) | | |
0937e0fd | 765 | (ap->flist->xbf_low ? XFS_PICK_LOWSPACE : 0); |
2a82b8be DC |
766 | |
767 | err = _xfs_filestream_pick_ag(mp, ag, agp, flags, minlen); | |
768 | if (err || *agp == NULLAGNUMBER) | |
769 | goto exit; | |
770 | ||
771 | /* | |
772 | * If the file wasn't found in the file cache, then its parent directory | |
773 | * inode isn't known. For this to have happened, the file must either | |
774 | * be pre-existing, or it was created long enough ago that its cache | |
775 | * entry has expired. This isn't the sort of usage that the filestreams | |
776 | * allocator is trying to optimise, so there's no point trying to track | |
777 | * its new AG somehow in the filestream data structures. | |
778 | */ | |
779 | if (!pip) { | |
780 | TRACE_ORPHAN(mp, ip, *agp); | |
781 | goto exit; | |
782 | } | |
783 | ||
784 | /* Associate the parent inode with the AG. */ | |
785 | err = _xfs_filestream_update_ag(pip, NULL, *agp); | |
786 | if (err) | |
787 | goto exit; | |
788 | ||
789 | /* Associate the file inode with the AG. */ | |
790 | err = _xfs_filestream_update_ag(ip, pip, *agp); | |
791 | if (err) | |
792 | goto exit; | |
793 | ||
794 | TRACE_MOVEAG(mp, ip, pip, NULLAGNUMBER, 0, | |
795 | *agp, xfs_filestream_peek_ag(mp, *agp)); | |
796 | ||
797 | exit: | |
798 | /* | |
799 | * If _xfs_filestream_pick_ag() returned a valid AG, remove the | |
800 | * reference it took on it, since the file and directory will have taken | |
801 | * their own now if they were successfully cached. | |
802 | */ | |
803 | if (*agp != NULLAGNUMBER) | |
804 | xfs_filestream_put_ag(mp, *agp); | |
805 | else | |
806 | *agp = 0; | |
807 | ||
808 | if (pip) | |
809 | xfs_iunlock(pip, XFS_IOLOCK_EXCL); | |
810 | ||
811 | return err; | |
812 | } | |
813 | ||
814 | /* | |
815 | * Remove an association between an inode and a filestream object. | |
816 | * Typically this is done on last close of an unlinked file. | |
817 | */ | |
818 | void | |
819 | xfs_filestream_deassociate( | |
820 | xfs_inode_t *ip) | |
821 | { | |
22328d71 | 822 | xfs_mru_cache_delete(ip->i_mount->m_filestream, ip->i_ino); |
2a82b8be | 823 | } |