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 | |
2cd2ef6a CH |
36 | struct xfs_fstrm_item { |
37 | struct xfs_mru_cache_elem mru; | |
38 | struct xfs_inode *ip; | |
39 | xfs_agnumber_t ag; /* AG in use for this directory */ | |
40 | }; | |
41 | ||
42 | enum xfs_fstrm_alloc { | |
43 | XFS_PICK_USERDATA = 1, | |
44 | XFS_PICK_LOWSPACE = 2, | |
45 | }; | |
2a82b8be | 46 | |
0664ce8d CH |
47 | /* |
48 | * Allocation group filestream associations are tracked with per-ag atomic | |
2cd2ef6a | 49 | * counters. These counters allow xfs_filestream_pick_ag() to tell whether a |
0664ce8d CH |
50 | * particular AG already has active filestreams associated with it. The mount |
51 | * point's m_peraglock is used to protect these counters from per-ag array | |
52 | * re-allocation during a growfs operation. When xfs_growfs_data_private() is | |
53 | * about to reallocate the array, it calls xfs_filestream_flush() with the | |
54 | * m_peraglock held in write mode. | |
55 | * | |
56 | * Since xfs_mru_cache_flush() guarantees that all the free functions for all | |
57 | * the cache elements have finished executing before it returns, it's safe for | |
58 | * the free functions to use the atomic counters without m_peraglock protection. | |
59 | * This allows the implementation of xfs_fstrm_free_func() to be agnostic about | |
60 | * whether it was called with the m_peraglock held in read mode, write mode or | |
61 | * not held at all. The race condition this addresses is the following: | |
62 | * | |
63 | * - The work queue scheduler fires and pulls a filestream directory cache | |
64 | * element off the LRU end of the cache for deletion, then gets pre-empted. | |
65 | * - A growfs operation grabs the m_peraglock in write mode, flushes all the | |
66 | * remaining items from the cache and reallocates the mount point's per-ag | |
67 | * array, resetting all the counters to zero. | |
68 | * - The work queue thread resumes and calls the free function for the element | |
69 | * it started cleaning up earlier. In the process it decrements the | |
70 | * filestreams counter for an AG that now has no references. | |
71 | * | |
72 | * With a shrinkfs feature, the above scenario could panic the system. | |
73 | * | |
74 | * All other uses of the following macros should be protected by either the | |
75 | * m_peraglock held in read mode, or the cache's internal locking exposed by the | |
76 | * interval between a call to xfs_mru_cache_lookup() and a call to | |
77 | * xfs_mru_cache_done(). In addition, the m_peraglock must be held in read mode | |
78 | * when new elements are added to the cache. | |
79 | * | |
80 | * Combined, these locking rules ensure that no associations will ever exist in | |
81 | * the cache that reference per-ag array elements that have since been | |
82 | * reallocated. | |
83 | */ | |
b94acd47 | 84 | int |
0664ce8d CH |
85 | xfs_filestream_peek_ag( |
86 | xfs_mount_t *mp, | |
87 | xfs_agnumber_t agno) | |
88 | { | |
89 | struct xfs_perag *pag; | |
90 | int ret; | |
91 | ||
92 | pag = xfs_perag_get(mp, agno); | |
93 | ret = atomic_read(&pag->pagf_fstrms); | |
94 | xfs_perag_put(pag); | |
95 | return ret; | |
96 | } | |
97 | ||
98 | static int | |
99 | xfs_filestream_get_ag( | |
100 | xfs_mount_t *mp, | |
101 | xfs_agnumber_t agno) | |
102 | { | |
103 | struct xfs_perag *pag; | |
104 | int ret; | |
105 | ||
106 | pag = xfs_perag_get(mp, agno); | |
107 | ret = atomic_inc_return(&pag->pagf_fstrms); | |
108 | xfs_perag_put(pag); | |
109 | return ret; | |
110 | } | |
111 | ||
112 | static void | |
113 | xfs_filestream_put_ag( | |
114 | xfs_mount_t *mp, | |
115 | xfs_agnumber_t agno) | |
116 | { | |
117 | struct xfs_perag *pag; | |
118 | ||
119 | pag = xfs_perag_get(mp, agno); | |
120 | atomic_dec(&pag->pagf_fstrms); | |
121 | xfs_perag_put(pag); | |
122 | } | |
2a82b8be | 123 | |
2cd2ef6a CH |
124 | static void |
125 | xfs_fstrm_free_func( | |
126 | struct xfs_mru_cache_elem *mru) | |
127 | { | |
128 | struct xfs_fstrm_item *item = | |
129 | container_of(mru, struct xfs_fstrm_item, mru); | |
130 | ||
131 | xfs_filestream_put_ag(item->ip->i_mount, item->ag); | |
132 | ||
b94acd47 | 133 | trace_xfs_filestream_free(item->ip, item->ag); |
2cd2ef6a | 134 | |
1919adda | 135 | kmem_free(item); |
2cd2ef6a CH |
136 | } |
137 | ||
2a82b8be DC |
138 | /* |
139 | * Scan the AGs starting at startag looking for an AG that isn't in use and has | |
140 | * at least minlen blocks free. | |
141 | */ | |
142 | static int | |
2cd2ef6a CH |
143 | xfs_filestream_pick_ag( |
144 | struct xfs_inode *ip, | |
145 | xfs_agnumber_t startag, | |
146 | xfs_agnumber_t *agp, | |
147 | int flags, | |
148 | xfs_extlen_t minlen) | |
2a82b8be | 149 | { |
2cd2ef6a CH |
150 | struct xfs_mount *mp = ip->i_mount; |
151 | struct xfs_fstrm_item *item; | |
152 | struct xfs_perag *pag; | |
b94acd47 | 153 | xfs_extlen_t longest, free = 0, minfree, maxfree = 0; |
2cd2ef6a | 154 | xfs_agnumber_t ag, max_ag = NULLAGNUMBER; |
2cd2ef6a CH |
155 | int err, trylock, nscan; |
156 | ||
157 | ASSERT(S_ISDIR(ip->i_d.di_mode)); | |
2a82b8be DC |
158 | |
159 | /* 2% of an AG's blocks must be free for it to be chosen. */ | |
160 | minfree = mp->m_sb.sb_agblocks / 50; | |
161 | ||
162 | ag = startag; | |
163 | *agp = NULLAGNUMBER; | |
164 | ||
165 | /* For the first pass, don't sleep trying to init the per-AG. */ | |
166 | trylock = XFS_ALLOC_FLAG_TRYLOCK; | |
167 | ||
168 | for (nscan = 0; 1; nscan++) { | |
b94acd47 CH |
169 | trace_xfs_filestream_scan(ip, ag); |
170 | ||
4196ac08 | 171 | pag = xfs_perag_get(mp, ag); |
2a82b8be DC |
172 | |
173 | if (!pag->pagf_init) { | |
174 | err = xfs_alloc_pagf_init(mp, NULL, ag, trylock); | |
4196ac08 DC |
175 | if (err && !trylock) { |
176 | xfs_perag_put(pag); | |
2a82b8be | 177 | return err; |
4196ac08 | 178 | } |
2a82b8be DC |
179 | } |
180 | ||
181 | /* Might fail sometimes during the 1st pass with trylock set. */ | |
182 | if (!pag->pagf_init) | |
183 | goto next_ag; | |
184 | ||
185 | /* Keep track of the AG with the most free blocks. */ | |
186 | if (pag->pagf_freeblks > maxfree) { | |
187 | maxfree = pag->pagf_freeblks; | |
188 | max_ag = ag; | |
189 | } | |
190 | ||
191 | /* | |
192 | * The AG reference count does two things: it enforces mutual | |
193 | * exclusion when examining the suitability of an AG in this | |
194 | * loop, and it guards against two filestreams being established | |
195 | * in the same AG as each other. | |
196 | */ | |
197 | if (xfs_filestream_get_ag(mp, ag) > 1) { | |
198 | xfs_filestream_put_ag(mp, ag); | |
199 | goto next_ag; | |
200 | } | |
201 | ||
6cc87645 | 202 | longest = xfs_alloc_longest_free_extent(mp, pag); |
2a82b8be DC |
203 | if (((minlen && longest >= minlen) || |
204 | (!minlen && pag->pagf_freeblks >= minfree)) && | |
205 | (!pag->pagf_metadata || !(flags & XFS_PICK_USERDATA) || | |
206 | (flags & XFS_PICK_LOWSPACE))) { | |
207 | ||
208 | /* Break out, retaining the reference on the AG. */ | |
209 | free = pag->pagf_freeblks; | |
4196ac08 | 210 | xfs_perag_put(pag); |
2a82b8be DC |
211 | *agp = ag; |
212 | break; | |
213 | } | |
214 | ||
215 | /* Drop the reference on this AG, it's not usable. */ | |
216 | xfs_filestream_put_ag(mp, ag); | |
217 | next_ag: | |
4196ac08 | 218 | xfs_perag_put(pag); |
2a82b8be DC |
219 | /* Move to the next AG, wrapping to AG 0 if necessary. */ |
220 | if (++ag >= mp->m_sb.sb_agcount) | |
221 | ag = 0; | |
222 | ||
223 | /* If a full pass of the AGs hasn't been done yet, continue. */ | |
224 | if (ag != startag) | |
225 | continue; | |
226 | ||
227 | /* Allow sleeping in xfs_alloc_pagf_init() on the 2nd pass. */ | |
228 | if (trylock != 0) { | |
229 | trylock = 0; | |
230 | continue; | |
231 | } | |
232 | ||
233 | /* Finally, if lowspace wasn't set, set it for the 3rd pass. */ | |
234 | if (!(flags & XFS_PICK_LOWSPACE)) { | |
235 | flags |= XFS_PICK_LOWSPACE; | |
236 | continue; | |
237 | } | |
238 | ||
239 | /* | |
240 | * Take the AG with the most free space, regardless of whether | |
241 | * it's already in use by another filestream. | |
242 | */ | |
243 | if (max_ag != NULLAGNUMBER) { | |
244 | xfs_filestream_get_ag(mp, max_ag); | |
2a82b8be DC |
245 | free = maxfree; |
246 | *agp = max_ag; | |
247 | break; | |
248 | } | |
249 | ||
250 | /* take AG 0 if none matched */ | |
b94acd47 | 251 | trace_xfs_filestream_pick(ip, *agp, free, nscan); |
2a82b8be DC |
252 | *agp = 0; |
253 | return 0; | |
254 | } | |
255 | ||
b94acd47 | 256 | trace_xfs_filestream_pick(ip, *agp, free, nscan); |
2a82b8be | 257 | |
2cd2ef6a | 258 | if (*agp == NULLAGNUMBER) |
2a82b8be | 259 | return 0; |
2a82b8be | 260 | |
2451337d | 261 | err = -ENOMEM; |
1919adda | 262 | item = kmem_alloc(sizeof(*item), KM_MAYFAIL); |
2a82b8be | 263 | if (!item) |
2cd2ef6a | 264 | goto out_put_ag; |
2a82b8be | 265 | |
2cd2ef6a | 266 | item->ag = *agp; |
2a82b8be | 267 | item->ip = ip; |
2a82b8be | 268 | |
22328d71 | 269 | err = xfs_mru_cache_insert(mp->m_filestream, ip->i_ino, &item->mru); |
2a82b8be | 270 | if (err) { |
2451337d | 271 | if (err == -EEXIST) |
2cd2ef6a CH |
272 | err = 0; |
273 | goto out_free_item; | |
2a82b8be DC |
274 | } |
275 | ||
2a82b8be | 276 | return 0; |
2a82b8be | 277 | |
2cd2ef6a | 278 | out_free_item: |
1919adda | 279 | kmem_free(item); |
2cd2ef6a CH |
280 | out_put_ag: |
281 | xfs_filestream_put_ag(mp, *agp); | |
282 | return err; | |
2a82b8be DC |
283 | } |
284 | ||
2cd2ef6a CH |
285 | static struct xfs_inode * |
286 | xfs_filestream_get_parent( | |
287 | struct xfs_inode *ip) | |
2a82b8be | 288 | { |
2cd2ef6a CH |
289 | struct inode *inode = VFS_I(ip), *dir = NULL; |
290 | struct dentry *dentry, *parent; | |
2a82b8be | 291 | |
2cd2ef6a CH |
292 | dentry = d_find_alias(inode); |
293 | if (!dentry) | |
294 | goto out; | |
2a82b8be | 295 | |
2cd2ef6a CH |
296 | parent = dget_parent(dentry); |
297 | if (!parent) | |
298 | goto out_dput; | |
2a82b8be | 299 | |
2cd2ef6a CH |
300 | dir = igrab(parent->d_inode); |
301 | dput(parent); | |
2a82b8be | 302 | |
2cd2ef6a CH |
303 | out_dput: |
304 | dput(dentry); | |
305 | out: | |
306 | return dir ? XFS_I(dir) : NULL; | |
2a82b8be DC |
307 | } |
308 | ||
2a82b8be | 309 | /* |
3b8d9076 CH |
310 | * Find the right allocation group for a file, either by finding an |
311 | * existing file stream or creating a new one. | |
312 | * | |
313 | * Returns NULLAGNUMBER in case of an error. | |
2a82b8be DC |
314 | */ |
315 | xfs_agnumber_t | |
316 | xfs_filestream_lookup_ag( | |
2cd2ef6a | 317 | struct xfs_inode *ip) |
2a82b8be | 318 | { |
2cd2ef6a | 319 | struct xfs_mount *mp = ip->i_mount; |
2cd2ef6a | 320 | struct xfs_inode *pip = NULL; |
3b8d9076 | 321 | xfs_agnumber_t startag, ag = NULLAGNUMBER; |
22328d71 | 322 | struct xfs_mru_cache_elem *mru; |
2a82b8be | 323 | |
2cd2ef6a | 324 | ASSERT(S_ISREG(ip->i_d.di_mode)); |
2a82b8be | 325 | |
2cd2ef6a CH |
326 | pip = xfs_filestream_get_parent(ip); |
327 | if (!pip) | |
328 | goto out; | |
329 | ||
330 | mru = xfs_mru_cache_lookup(mp->m_filestream, pip->i_ino); | |
22328d71 | 331 | if (mru) { |
3b8d9076 | 332 | ag = container_of(mru, struct xfs_fstrm_item, mru)->ag; |
22328d71 | 333 | xfs_mru_cache_done(mp->m_filestream); |
3b8d9076 | 334 | |
b94acd47 | 335 | trace_xfs_filestream_lookup(ip, ag); |
3b8d9076 | 336 | goto out; |
2a82b8be DC |
337 | } |
338 | ||
339 | /* | |
340 | * Set the starting AG using the rotor for inode32, otherwise | |
341 | * use the directory inode's AG. | |
342 | */ | |
343 | if (mp->m_flags & XFS_MOUNT_32BITINODES) { | |
2cd2ef6a | 344 | xfs_agnumber_t rotorstep = xfs_rotorstep; |
2a82b8be DC |
345 | startag = (mp->m_agfrotor / rotorstep) % mp->m_sb.sb_agcount; |
346 | mp->m_agfrotor = (mp->m_agfrotor + 1) % | |
347 | (mp->m_sb.sb_agcount * rotorstep); | |
348 | } else | |
349 | startag = XFS_INO_TO_AGNO(mp, pip->i_ino); | |
350 | ||
3b8d9076 CH |
351 | if (xfs_filestream_pick_ag(pip, startag, &ag, 0, 0)) |
352 | ag = NULLAGNUMBER; | |
353 | out: | |
354 | IRELE(pip); | |
355 | return ag; | |
2a82b8be DC |
356 | } |
357 | ||
358 | /* | |
2cd2ef6a CH |
359 | * Pick a new allocation group for the current file and its file stream. |
360 | * | |
361 | * This is called when the allocator can't find a suitable extent in the | |
362 | * current AG, and we have to move the stream into a new AG with more space. | |
2a82b8be DC |
363 | */ |
364 | int | |
365 | xfs_filestream_new_ag( | |
68988114 DC |
366 | struct xfs_bmalloca *ap, |
367 | xfs_agnumber_t *agp) | |
2a82b8be | 368 | { |
2cd2ef6a CH |
369 | struct xfs_inode *ip = ap->ip, *pip; |
370 | struct xfs_mount *mp = ip->i_mount; | |
371 | xfs_extlen_t minlen = ap->length; | |
372 | xfs_agnumber_t startag = 0; | |
373 | int flags, err = 0; | |
374 | struct xfs_mru_cache_elem *mru; | |
2a82b8be | 375 | |
2cd2ef6a | 376 | *agp = NULLAGNUMBER; |
2a82b8be | 377 | |
2cd2ef6a CH |
378 | pip = xfs_filestream_get_parent(ip); |
379 | if (!pip) | |
380 | goto exit; | |
2a82b8be | 381 | |
2cd2ef6a CH |
382 | mru = xfs_mru_cache_remove(mp->m_filestream, pip->i_ino); |
383 | if (mru) { | |
384 | struct xfs_fstrm_item *item = | |
385 | container_of(mru, struct xfs_fstrm_item, mru); | |
386 | startag = (item->ag + 1) % mp->m_sb.sb_agcount; | |
2a82b8be DC |
387 | } |
388 | ||
2a82b8be | 389 | flags = (ap->userdata ? XFS_PICK_USERDATA : 0) | |
0937e0fd | 390 | (ap->flist->xbf_low ? XFS_PICK_LOWSPACE : 0); |
2a82b8be | 391 | |
2cd2ef6a | 392 | err = xfs_filestream_pick_ag(pip, startag, agp, flags, minlen); |
2a82b8be DC |
393 | |
394 | /* | |
2cd2ef6a | 395 | * Only free the item here so we skip over the old AG earlier. |
2a82b8be | 396 | */ |
2cd2ef6a CH |
397 | if (mru) |
398 | xfs_fstrm_free_func(mru); | |
2a82b8be | 399 | |
2cd2ef6a CH |
400 | IRELE(pip); |
401 | exit: | |
402 | if (*agp == NULLAGNUMBER) | |
403 | *agp = 0; | |
404 | return err; | |
405 | } | |
2a82b8be | 406 | |
2cd2ef6a CH |
407 | void |
408 | xfs_filestream_deassociate( | |
409 | struct xfs_inode *ip) | |
410 | { | |
411 | xfs_mru_cache_delete(ip->i_mount->m_filestream, ip->i_ino); | |
412 | } | |
2a82b8be | 413 | |
2cd2ef6a CH |
414 | int |
415 | xfs_filestream_mount( | |
416 | xfs_mount_t *mp) | |
417 | { | |
2a82b8be | 418 | /* |
2cd2ef6a CH |
419 | * The filestream timer tunable is currently fixed within the range of |
420 | * one second to four minutes, with five seconds being the default. The | |
421 | * group count is somewhat arbitrary, but it'd be nice to adhere to the | |
422 | * timer tunable to within about 10 percent. This requires at least 10 | |
423 | * groups. | |
2a82b8be | 424 | */ |
2cd2ef6a CH |
425 | return xfs_mru_cache_create(&mp->m_filestream, xfs_fstrm_centisecs * 10, |
426 | 10, xfs_fstrm_free_func); | |
427 | } | |
2a82b8be | 428 | |
2cd2ef6a CH |
429 | void |
430 | xfs_filestream_unmount( | |
431 | xfs_mount_t *mp) | |
432 | { | |
433 | xfs_mru_cache_destroy(mp->m_filestream); | |
434 | } |