Commit | Line | Data |
---|---|---|
2a82b8be DC |
1 | /* |
2 | * Copyright (c) 2006-2007 Silicon Graphics, Inc. | |
2cd2ef6a | 3 | * Copyright (c) 2014 Christoph Hellwig. |
2a82b8be DC |
4 | * All Rights Reserved. |
5 | * | |
6 | * This program is free software; you can redistribute it and/or | |
7 | * modify it under the terms of the GNU General Public License as | |
8 | * published by the Free Software Foundation. | |
9 | * | |
10 | * This program is distributed in the hope that it would be useful, | |
11 | * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
12 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | |
13 | * GNU General Public License for more details. | |
14 | * | |
15 | * You should have received a copy of the GNU General Public License | |
16 | * along with this program; if not, write the Free Software Foundation, | |
17 | * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA | |
18 | */ | |
19 | #include "xfs.h" | |
a4fbe6ab | 20 | #include "xfs_format.h" |
239880ef DC |
21 | #include "xfs_log_format.h" |
22 | #include "xfs_trans_resv.h" | |
23 | #include "xfs_ag.h" | |
24 | #include "xfs_sb.h" | |
25 | #include "xfs_mount.h" | |
2a82b8be | 26 | #include "xfs_inum.h" |
2a82b8be | 27 | #include "xfs_inode.h" |
2a82b8be | 28 | #include "xfs_bmap.h" |
68988114 | 29 | #include "xfs_bmap_util.h" |
2a82b8be | 30 | #include "xfs_alloc.h" |
2a82b8be | 31 | #include "xfs_mru_cache.h" |
a4fbe6ab | 32 | #include "xfs_dinode.h" |
2a82b8be | 33 | #include "xfs_filestream.h" |
0b1b213f | 34 | #include "xfs_trace.h" |
2a82b8be | 35 | |
2a82b8be DC |
36 | #define TRACE_AG_SCAN(mp, ag, ag2) |
37 | #define TRACE_AG_PICK1(mp, max_ag, maxfree) | |
38 | #define TRACE_AG_PICK2(mp, ag, ag2, cnt, free, scan, flag) | |
2a82b8be DC |
39 | #define TRACE_FREE(mp, ip, pip, ag, cnt) |
40 | #define TRACE_LOOKUP(mp, ip, pip, ag, cnt) | |
2a82b8be | 41 | |
2cd2ef6a CH |
42 | struct xfs_fstrm_item { |
43 | struct xfs_mru_cache_elem mru; | |
44 | struct xfs_inode *ip; | |
45 | xfs_agnumber_t ag; /* AG in use for this directory */ | |
46 | }; | |
47 | ||
48 | enum xfs_fstrm_alloc { | |
49 | XFS_PICK_USERDATA = 1, | |
50 | XFS_PICK_LOWSPACE = 2, | |
51 | }; | |
2a82b8be | 52 | |
0664ce8d CH |
53 | /* |
54 | * Allocation group filestream associations are tracked with per-ag atomic | |
2cd2ef6a | 55 | * counters. These counters allow xfs_filestream_pick_ag() to tell whether a |
0664ce8d CH |
56 | * particular AG already has active filestreams associated with it. The mount |
57 | * point's m_peraglock is used to protect these counters from per-ag array | |
58 | * re-allocation during a growfs operation. When xfs_growfs_data_private() is | |
59 | * about to reallocate the array, it calls xfs_filestream_flush() with the | |
60 | * m_peraglock held in write mode. | |
61 | * | |
62 | * Since xfs_mru_cache_flush() guarantees that all the free functions for all | |
63 | * the cache elements have finished executing before it returns, it's safe for | |
64 | * the free functions to use the atomic counters without m_peraglock protection. | |
65 | * This allows the implementation of xfs_fstrm_free_func() to be agnostic about | |
66 | * whether it was called with the m_peraglock held in read mode, write mode or | |
67 | * not held at all. The race condition this addresses is the following: | |
68 | * | |
69 | * - The work queue scheduler fires and pulls a filestream directory cache | |
70 | * element off the LRU end of the cache for deletion, then gets pre-empted. | |
71 | * - A growfs operation grabs the m_peraglock in write mode, flushes all the | |
72 | * remaining items from the cache and reallocates the mount point's per-ag | |
73 | * array, resetting all the counters to zero. | |
74 | * - The work queue thread resumes and calls the free function for the element | |
75 | * it started cleaning up earlier. In the process it decrements the | |
76 | * filestreams counter for an AG that now has no references. | |
77 | * | |
78 | * With a shrinkfs feature, the above scenario could panic the system. | |
79 | * | |
80 | * All other uses of the following macros should be protected by either the | |
81 | * m_peraglock held in read mode, or the cache's internal locking exposed by the | |
82 | * interval between a call to xfs_mru_cache_lookup() and a call to | |
83 | * xfs_mru_cache_done(). In addition, the m_peraglock must be held in read mode | |
84 | * when new elements are added to the cache. | |
85 | * | |
86 | * Combined, these locking rules ensure that no associations will ever exist in | |
87 | * the cache that reference per-ag array elements that have since been | |
88 | * reallocated. | |
89 | */ | |
90 | static int | |
91 | xfs_filestream_peek_ag( | |
92 | xfs_mount_t *mp, | |
93 | xfs_agnumber_t agno) | |
94 | { | |
95 | struct xfs_perag *pag; | |
96 | int ret; | |
97 | ||
98 | pag = xfs_perag_get(mp, agno); | |
99 | ret = atomic_read(&pag->pagf_fstrms); | |
100 | xfs_perag_put(pag); | |
101 | return ret; | |
102 | } | |
103 | ||
104 | static int | |
105 | xfs_filestream_get_ag( | |
106 | xfs_mount_t *mp, | |
107 | xfs_agnumber_t agno) | |
108 | { | |
109 | struct xfs_perag *pag; | |
110 | int ret; | |
111 | ||
112 | pag = xfs_perag_get(mp, agno); | |
113 | ret = atomic_inc_return(&pag->pagf_fstrms); | |
114 | xfs_perag_put(pag); | |
115 | return ret; | |
116 | } | |
117 | ||
118 | static void | |
119 | xfs_filestream_put_ag( | |
120 | xfs_mount_t *mp, | |
121 | xfs_agnumber_t agno) | |
122 | { | |
123 | struct xfs_perag *pag; | |
124 | ||
125 | pag = xfs_perag_get(mp, agno); | |
126 | atomic_dec(&pag->pagf_fstrms); | |
127 | xfs_perag_put(pag); | |
128 | } | |
2a82b8be | 129 | |
2cd2ef6a CH |
130 | static void |
131 | xfs_fstrm_free_func( | |
132 | struct xfs_mru_cache_elem *mru) | |
133 | { | |
134 | struct xfs_fstrm_item *item = | |
135 | container_of(mru, struct xfs_fstrm_item, mru); | |
136 | ||
137 | xfs_filestream_put_ag(item->ip->i_mount, item->ag); | |
138 | ||
139 | TRACE_FREE(mp, ip, NULL, item->ag, | |
140 | xfs_filestream_peek_ag(mp, item->ag)); | |
141 | ||
1919adda | 142 | kmem_free(item); |
2cd2ef6a CH |
143 | } |
144 | ||
2a82b8be DC |
145 | /* |
146 | * Scan the AGs starting at startag looking for an AG that isn't in use and has | |
147 | * at least minlen blocks free. | |
148 | */ | |
149 | static int | |
2cd2ef6a CH |
150 | xfs_filestream_pick_ag( |
151 | struct xfs_inode *ip, | |
152 | xfs_agnumber_t startag, | |
153 | xfs_agnumber_t *agp, | |
154 | int flags, | |
155 | xfs_extlen_t minlen) | |
2a82b8be | 156 | { |
2cd2ef6a CH |
157 | struct xfs_mount *mp = ip->i_mount; |
158 | struct xfs_fstrm_item *item; | |
159 | struct xfs_perag *pag; | |
160 | xfs_extlen_t longest, free, minfree, maxfree = 0; | |
161 | xfs_agnumber_t ag, max_ag = NULLAGNUMBER; | |
162 | int streams, max_streams; | |
163 | int err, trylock, nscan; | |
164 | ||
165 | ASSERT(S_ISDIR(ip->i_d.di_mode)); | |
2a82b8be DC |
166 | |
167 | /* 2% of an AG's blocks must be free for it to be chosen. */ | |
168 | minfree = mp->m_sb.sb_agblocks / 50; | |
169 | ||
170 | ag = startag; | |
171 | *agp = NULLAGNUMBER; | |
172 | ||
173 | /* For the first pass, don't sleep trying to init the per-AG. */ | |
174 | trylock = XFS_ALLOC_FLAG_TRYLOCK; | |
175 | ||
176 | for (nscan = 0; 1; nscan++) { | |
4196ac08 DC |
177 | pag = xfs_perag_get(mp, ag); |
178 | TRACE_AG_SCAN(mp, ag, atomic_read(&pag->pagf_fstrms)); | |
2a82b8be DC |
179 | |
180 | if (!pag->pagf_init) { | |
181 | err = xfs_alloc_pagf_init(mp, NULL, ag, trylock); | |
4196ac08 DC |
182 | if (err && !trylock) { |
183 | xfs_perag_put(pag); | |
2a82b8be | 184 | return err; |
4196ac08 | 185 | } |
2a82b8be DC |
186 | } |
187 | ||
188 | /* Might fail sometimes during the 1st pass with trylock set. */ | |
189 | if (!pag->pagf_init) | |
190 | goto next_ag; | |
191 | ||
192 | /* Keep track of the AG with the most free blocks. */ | |
193 | if (pag->pagf_freeblks > maxfree) { | |
194 | maxfree = pag->pagf_freeblks; | |
4196ac08 | 195 | max_streams = atomic_read(&pag->pagf_fstrms); |
2a82b8be DC |
196 | max_ag = ag; |
197 | } | |
198 | ||
199 | /* | |
200 | * The AG reference count does two things: it enforces mutual | |
201 | * exclusion when examining the suitability of an AG in this | |
202 | * loop, and it guards against two filestreams being established | |
203 | * in the same AG as each other. | |
204 | */ | |
205 | if (xfs_filestream_get_ag(mp, ag) > 1) { | |
206 | xfs_filestream_put_ag(mp, ag); | |
207 | goto next_ag; | |
208 | } | |
209 | ||
6cc87645 | 210 | longest = xfs_alloc_longest_free_extent(mp, pag); |
2a82b8be DC |
211 | if (((minlen && longest >= minlen) || |
212 | (!minlen && pag->pagf_freeblks >= minfree)) && | |
213 | (!pag->pagf_metadata || !(flags & XFS_PICK_USERDATA) || | |
214 | (flags & XFS_PICK_LOWSPACE))) { | |
215 | ||
216 | /* Break out, retaining the reference on the AG. */ | |
217 | free = pag->pagf_freeblks; | |
4196ac08 DC |
218 | streams = atomic_read(&pag->pagf_fstrms); |
219 | xfs_perag_put(pag); | |
2a82b8be DC |
220 | *agp = ag; |
221 | break; | |
222 | } | |
223 | ||
224 | /* Drop the reference on this AG, it's not usable. */ | |
225 | xfs_filestream_put_ag(mp, ag); | |
226 | next_ag: | |
4196ac08 | 227 | xfs_perag_put(pag); |
2a82b8be DC |
228 | /* Move to the next AG, wrapping to AG 0 if necessary. */ |
229 | if (++ag >= mp->m_sb.sb_agcount) | |
230 | ag = 0; | |
231 | ||
232 | /* If a full pass of the AGs hasn't been done yet, continue. */ | |
233 | if (ag != startag) | |
234 | continue; | |
235 | ||
236 | /* Allow sleeping in xfs_alloc_pagf_init() on the 2nd pass. */ | |
237 | if (trylock != 0) { | |
238 | trylock = 0; | |
239 | continue; | |
240 | } | |
241 | ||
242 | /* Finally, if lowspace wasn't set, set it for the 3rd pass. */ | |
243 | if (!(flags & XFS_PICK_LOWSPACE)) { | |
244 | flags |= XFS_PICK_LOWSPACE; | |
245 | continue; | |
246 | } | |
247 | ||
248 | /* | |
249 | * Take the AG with the most free space, regardless of whether | |
250 | * it's already in use by another filestream. | |
251 | */ | |
252 | if (max_ag != NULLAGNUMBER) { | |
253 | xfs_filestream_get_ag(mp, max_ag); | |
254 | TRACE_AG_PICK1(mp, max_ag, maxfree); | |
4196ac08 | 255 | streams = max_streams; |
2a82b8be DC |
256 | free = maxfree; |
257 | *agp = max_ag; | |
258 | break; | |
259 | } | |
260 | ||
261 | /* take AG 0 if none matched */ | |
262 | TRACE_AG_PICK1(mp, max_ag, maxfree); | |
263 | *agp = 0; | |
264 | return 0; | |
265 | } | |
266 | ||
4196ac08 | 267 | TRACE_AG_PICK2(mp, startag, *agp, streams, free, nscan, flags); |
2a82b8be | 268 | |
2cd2ef6a | 269 | if (*agp == NULLAGNUMBER) |
2a82b8be | 270 | return 0; |
2a82b8be | 271 | |
2cd2ef6a | 272 | err = ENOMEM; |
1919adda | 273 | item = kmem_alloc(sizeof(*item), KM_MAYFAIL); |
2a82b8be | 274 | if (!item) |
2cd2ef6a | 275 | goto out_put_ag; |
2a82b8be | 276 | |
2cd2ef6a | 277 | item->ag = *agp; |
2a82b8be | 278 | item->ip = ip; |
2a82b8be | 279 | |
22328d71 | 280 | err = xfs_mru_cache_insert(mp->m_filestream, ip->i_ino, &item->mru); |
2a82b8be | 281 | if (err) { |
2cd2ef6a CH |
282 | if (err == EEXIST) |
283 | err = 0; | |
284 | goto out_free_item; | |
2a82b8be DC |
285 | } |
286 | ||
2a82b8be | 287 | return 0; |
2a82b8be | 288 | |
2cd2ef6a | 289 | out_free_item: |
1919adda | 290 | kmem_free(item); |
2cd2ef6a CH |
291 | out_put_ag: |
292 | xfs_filestream_put_ag(mp, *agp); | |
293 | return err; | |
2a82b8be DC |
294 | } |
295 | ||
2cd2ef6a CH |
296 | static struct xfs_inode * |
297 | xfs_filestream_get_parent( | |
298 | struct xfs_inode *ip) | |
2a82b8be | 299 | { |
2cd2ef6a CH |
300 | struct inode *inode = VFS_I(ip), *dir = NULL; |
301 | struct dentry *dentry, *parent; | |
2a82b8be | 302 | |
2cd2ef6a CH |
303 | dentry = d_find_alias(inode); |
304 | if (!dentry) | |
305 | goto out; | |
2a82b8be | 306 | |
2cd2ef6a CH |
307 | parent = dget_parent(dentry); |
308 | if (!parent) | |
309 | goto out_dput; | |
2a82b8be | 310 | |
2cd2ef6a CH |
311 | dir = igrab(parent->d_inode); |
312 | dput(parent); | |
2a82b8be | 313 | |
2cd2ef6a CH |
314 | out_dput: |
315 | dput(dentry); | |
316 | out: | |
317 | return dir ? XFS_I(dir) : NULL; | |
2a82b8be DC |
318 | } |
319 | ||
2a82b8be DC |
320 | /* |
321 | * Return the AG of the filestream the file or directory belongs to, or | |
322 | * NULLAGNUMBER otherwise. | |
323 | */ | |
324 | xfs_agnumber_t | |
325 | xfs_filestream_lookup_ag( | |
2cd2ef6a | 326 | struct xfs_inode *ip) |
2a82b8be | 327 | { |
2cd2ef6a CH |
328 | struct xfs_mount *mp = ip->i_mount; |
329 | struct xfs_fstrm_item *item; | |
330 | struct xfs_inode *pip = NULL; | |
331 | xfs_agnumber_t ag = NULLAGNUMBER; | |
332 | int ref = 0; | |
22328d71 | 333 | struct xfs_mru_cache_elem *mru; |
2a82b8be | 334 | |
2cd2ef6a | 335 | ASSERT(S_ISREG(ip->i_d.di_mode)); |
2a82b8be | 336 | |
2cd2ef6a CH |
337 | pip = xfs_filestream_get_parent(ip); |
338 | if (!pip) | |
339 | goto out; | |
340 | ||
341 | mru = xfs_mru_cache_lookup(mp->m_filestream, pip->i_ino); | |
342 | if (!mru) | |
343 | goto out; | |
344 | ||
345 | item = container_of(mru, struct xfs_fstrm_item, mru); | |
2a82b8be | 346 | |
2a82b8be | 347 | ag = item->ag; |
22328d71 | 348 | xfs_mru_cache_done(mp->m_filestream); |
2a82b8be | 349 | |
2cd2ef6a CH |
350 | ref = xfs_filestream_peek_ag(ip->i_mount, ag); |
351 | out: | |
352 | TRACE_LOOKUP(mp, ip, pip, ag, ref); | |
353 | IRELE(pip); | |
2a82b8be DC |
354 | return ag; |
355 | } | |
356 | ||
357 | /* | |
2cd2ef6a | 358 | * Make sure a directory has a filestream associated with it. |
2a82b8be | 359 | * |
2cd2ef6a CH |
360 | * This is called when creating regular files in an directory that has |
361 | * filestreams enabled, so that a stream is ready by the time we need it | |
362 | * in the allocator for the files inside the directory. | |
2a82b8be DC |
363 | */ |
364 | int | |
365 | xfs_filestream_associate( | |
2cd2ef6a | 366 | struct xfs_inode *pip) |
2a82b8be | 367 | { |
2cd2ef6a | 368 | struct xfs_mount *mp = pip->i_mount; |
22328d71 | 369 | struct xfs_mru_cache_elem *mru; |
2cd2ef6a | 370 | xfs_agnumber_t startag, ag; |
2a82b8be | 371 | |
03209378 | 372 | ASSERT(S_ISDIR(pip->i_d.di_mode)); |
2a82b8be DC |
373 | |
374 | /* | |
2cd2ef6a | 375 | * If the directory already has a file stream associated we're done. |
2a82b8be | 376 | */ |
22328d71 CH |
377 | mru = xfs_mru_cache_lookup(mp->m_filestream, pip->i_ino); |
378 | if (mru) { | |
22328d71 | 379 | xfs_mru_cache_done(mp->m_filestream); |
2cd2ef6a | 380 | return 0; |
2a82b8be DC |
381 | } |
382 | ||
383 | /* | |
384 | * Set the starting AG using the rotor for inode32, otherwise | |
385 | * use the directory inode's AG. | |
386 | */ | |
387 | if (mp->m_flags & XFS_MOUNT_32BITINODES) { | |
2cd2ef6a | 388 | xfs_agnumber_t rotorstep = xfs_rotorstep; |
2a82b8be DC |
389 | startag = (mp->m_agfrotor / rotorstep) % mp->m_sb.sb_agcount; |
390 | mp->m_agfrotor = (mp->m_agfrotor + 1) % | |
391 | (mp->m_sb.sb_agcount * rotorstep); | |
392 | } else | |
393 | startag = XFS_INO_TO_AGNO(mp, pip->i_ino); | |
394 | ||
2cd2ef6a | 395 | return xfs_filestream_pick_ag(pip, startag, &ag, 0, 0); |
2a82b8be DC |
396 | } |
397 | ||
398 | /* | |
2cd2ef6a CH |
399 | * Pick a new allocation group for the current file and its file stream. |
400 | * | |
401 | * This is called when the allocator can't find a suitable extent in the | |
402 | * current AG, and we have to move the stream into a new AG with more space. | |
2a82b8be DC |
403 | */ |
404 | int | |
405 | xfs_filestream_new_ag( | |
68988114 DC |
406 | struct xfs_bmalloca *ap, |
407 | xfs_agnumber_t *agp) | |
2a82b8be | 408 | { |
2cd2ef6a CH |
409 | struct xfs_inode *ip = ap->ip, *pip; |
410 | struct xfs_mount *mp = ip->i_mount; | |
411 | xfs_extlen_t minlen = ap->length; | |
412 | xfs_agnumber_t startag = 0; | |
413 | int flags, err = 0; | |
414 | struct xfs_mru_cache_elem *mru; | |
2a82b8be | 415 | |
2cd2ef6a | 416 | *agp = NULLAGNUMBER; |
2a82b8be | 417 | |
2cd2ef6a CH |
418 | pip = xfs_filestream_get_parent(ip); |
419 | if (!pip) | |
420 | goto exit; | |
2a82b8be | 421 | |
2cd2ef6a CH |
422 | mru = xfs_mru_cache_remove(mp->m_filestream, pip->i_ino); |
423 | if (mru) { | |
424 | struct xfs_fstrm_item *item = | |
425 | container_of(mru, struct xfs_fstrm_item, mru); | |
426 | startag = (item->ag + 1) % mp->m_sb.sb_agcount; | |
2a82b8be DC |
427 | } |
428 | ||
2a82b8be | 429 | flags = (ap->userdata ? XFS_PICK_USERDATA : 0) | |
0937e0fd | 430 | (ap->flist->xbf_low ? XFS_PICK_LOWSPACE : 0); |
2a82b8be | 431 | |
2cd2ef6a | 432 | err = xfs_filestream_pick_ag(pip, startag, agp, flags, minlen); |
2a82b8be DC |
433 | |
434 | /* | |
2cd2ef6a | 435 | * Only free the item here so we skip over the old AG earlier. |
2a82b8be | 436 | */ |
2cd2ef6a CH |
437 | if (mru) |
438 | xfs_fstrm_free_func(mru); | |
2a82b8be | 439 | |
2cd2ef6a CH |
440 | IRELE(pip); |
441 | exit: | |
442 | if (*agp == NULLAGNUMBER) | |
443 | *agp = 0; | |
444 | return err; | |
445 | } | |
2a82b8be | 446 | |
2cd2ef6a CH |
447 | void |
448 | xfs_filestream_deassociate( | |
449 | struct xfs_inode *ip) | |
450 | { | |
451 | xfs_mru_cache_delete(ip->i_mount->m_filestream, ip->i_ino); | |
452 | } | |
2a82b8be | 453 | |
2cd2ef6a CH |
454 | int |
455 | xfs_filestream_mount( | |
456 | xfs_mount_t *mp) | |
457 | { | |
2a82b8be | 458 | /* |
2cd2ef6a CH |
459 | * The filestream timer tunable is currently fixed within the range of |
460 | * one second to four minutes, with five seconds being the default. The | |
461 | * group count is somewhat arbitrary, but it'd be nice to adhere to the | |
462 | * timer tunable to within about 10 percent. This requires at least 10 | |
463 | * groups. | |
2a82b8be | 464 | */ |
2cd2ef6a CH |
465 | return xfs_mru_cache_create(&mp->m_filestream, xfs_fstrm_centisecs * 10, |
466 | 10, xfs_fstrm_free_func); | |
467 | } | |
2a82b8be | 468 | |
2cd2ef6a CH |
469 | void |
470 | xfs_filestream_unmount( | |
471 | xfs_mount_t *mp) | |
472 | { | |
473 | xfs_mru_cache_destroy(mp->m_filestream); | |
474 | } |