Merge tag 'for_linus_stable' of git://git.kernel.org/pub/scm/linux/kernel/git/tytso...
[linux-2.6-block.git] / fs / xfs / xfs_filestream.c
CommitLineData
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"
239880ef
DC
23#include "xfs_sb.h"
24#include "xfs_mount.h"
3ab78df2 25#include "xfs_defer.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
DC
30#include "xfs_mru_cache.h"
31#include "xfs_filestream.h"
0b1b213f 32#include "xfs_trace.h"
3fd129b6 33#include "xfs_ag_resv.h"
2a82b8be 34
2cd2ef6a
CH
35struct xfs_fstrm_item {
36 struct xfs_mru_cache_elem mru;
2cd2ef6a
CH
37 xfs_agnumber_t ag; /* AG in use for this directory */
38};
39
40enum xfs_fstrm_alloc {
41 XFS_PICK_USERDATA = 1,
42 XFS_PICK_LOWSPACE = 2,
43};
2a82b8be 44
0664ce8d
CH
45/*
46 * Allocation group filestream associations are tracked with per-ag atomic
2cd2ef6a 47 * counters. These counters allow xfs_filestream_pick_ag() to tell whether a
0664ce8d
CH
48 * particular AG already has active filestreams associated with it. The mount
49 * point's m_peraglock is used to protect these counters from per-ag array
50 * re-allocation during a growfs operation. When xfs_growfs_data_private() is
51 * about to reallocate the array, it calls xfs_filestream_flush() with the
52 * m_peraglock held in write mode.
53 *
54 * Since xfs_mru_cache_flush() guarantees that all the free functions for all
55 * the cache elements have finished executing before it returns, it's safe for
56 * the free functions to use the atomic counters without m_peraglock protection.
57 * This allows the implementation of xfs_fstrm_free_func() to be agnostic about
58 * whether it was called with the m_peraglock held in read mode, write mode or
59 * not held at all. The race condition this addresses is the following:
60 *
61 * - The work queue scheduler fires and pulls a filestream directory cache
62 * element off the LRU end of the cache for deletion, then gets pre-empted.
63 * - A growfs operation grabs the m_peraglock in write mode, flushes all the
64 * remaining items from the cache and reallocates the mount point's per-ag
65 * array, resetting all the counters to zero.
66 * - The work queue thread resumes and calls the free function for the element
67 * it started cleaning up earlier. In the process it decrements the
68 * filestreams counter for an AG that now has no references.
69 *
70 * With a shrinkfs feature, the above scenario could panic the system.
71 *
72 * All other uses of the following macros should be protected by either the
73 * m_peraglock held in read mode, or the cache's internal locking exposed by the
74 * interval between a call to xfs_mru_cache_lookup() and a call to
75 * xfs_mru_cache_done(). In addition, the m_peraglock must be held in read mode
76 * when new elements are added to the cache.
77 *
78 * Combined, these locking rules ensure that no associations will ever exist in
79 * the cache that reference per-ag array elements that have since been
80 * reallocated.
81 */
b94acd47 82int
0664ce8d
CH
83xfs_filestream_peek_ag(
84 xfs_mount_t *mp,
85 xfs_agnumber_t agno)
86{
87 struct xfs_perag *pag;
88 int ret;
89
90 pag = xfs_perag_get(mp, agno);
91 ret = atomic_read(&pag->pagf_fstrms);
92 xfs_perag_put(pag);
93 return ret;
94}
95
96static int
97xfs_filestream_get_ag(
98 xfs_mount_t *mp,
99 xfs_agnumber_t agno)
100{
101 struct xfs_perag *pag;
102 int ret;
103
104 pag = xfs_perag_get(mp, agno);
105 ret = atomic_inc_return(&pag->pagf_fstrms);
106 xfs_perag_put(pag);
107 return ret;
108}
109
110static void
111xfs_filestream_put_ag(
112 xfs_mount_t *mp,
113 xfs_agnumber_t agno)
114{
115 struct xfs_perag *pag;
116
117 pag = xfs_perag_get(mp, agno);
118 atomic_dec(&pag->pagf_fstrms);
119 xfs_perag_put(pag);
120}
2a82b8be 121
2cd2ef6a
CH
122static void
123xfs_fstrm_free_func(
7fcd3efa 124 void *data,
2cd2ef6a
CH
125 struct xfs_mru_cache_elem *mru)
126{
7fcd3efa 127 struct xfs_mount *mp = data;
2cd2ef6a
CH
128 struct xfs_fstrm_item *item =
129 container_of(mru, struct xfs_fstrm_item, mru);
130
7fcd3efa
CH
131 xfs_filestream_put_ag(mp, item->ag);
132 trace_xfs_filestream_free(mp, mru->key, item->ag);
2cd2ef6a 133
1919adda 134 kmem_free(item);
2cd2ef6a
CH
135}
136
2a82b8be
DC
137/*
138 * Scan the AGs starting at startag looking for an AG that isn't in use and has
139 * at least minlen blocks free.
140 */
141static int
2cd2ef6a
CH
142xfs_filestream_pick_ag(
143 struct xfs_inode *ip,
144 xfs_agnumber_t startag,
145 xfs_agnumber_t *agp,
146 int flags,
147 xfs_extlen_t minlen)
2a82b8be 148{
2cd2ef6a
CH
149 struct xfs_mount *mp = ip->i_mount;
150 struct xfs_fstrm_item *item;
151 struct xfs_perag *pag;
b94acd47 152 xfs_extlen_t longest, free = 0, minfree, maxfree = 0;
2cd2ef6a 153 xfs_agnumber_t ag, max_ag = NULLAGNUMBER;
2cd2ef6a
CH
154 int err, trylock, nscan;
155
c19b3b05 156 ASSERT(S_ISDIR(VFS_I(ip)->i_mode));
2a82b8be
DC
157
158 /* 2% of an AG's blocks must be free for it to be chosen. */
159 minfree = mp->m_sb.sb_agblocks / 50;
160
161 ag = startag;
162 *agp = NULLAGNUMBER;
163
164 /* For the first pass, don't sleep trying to init the per-AG. */
165 trylock = XFS_ALLOC_FLAG_TRYLOCK;
166
167 for (nscan = 0; 1; nscan++) {
7fcd3efa 168 trace_xfs_filestream_scan(mp, ip->i_ino, ag);
b94acd47 169
4196ac08 170 pag = xfs_perag_get(mp, ag);
2a82b8be
DC
171
172 if (!pag->pagf_init) {
173 err = xfs_alloc_pagf_init(mp, NULL, ag, trylock);
4196ac08
DC
174 if (err && !trylock) {
175 xfs_perag_put(pag);
2a82b8be 176 return err;
4196ac08 177 }
2a82b8be
DC
178 }
179
180 /* Might fail sometimes during the 1st pass with trylock set. */
181 if (!pag->pagf_init)
182 goto next_ag;
183
184 /* Keep track of the AG with the most free blocks. */
185 if (pag->pagf_freeblks > maxfree) {
186 maxfree = pag->pagf_freeblks;
187 max_ag = ag;
188 }
189
190 /*
191 * The AG reference count does two things: it enforces mutual
192 * exclusion when examining the suitability of an AG in this
193 * loop, and it guards against two filestreams being established
194 * in the same AG as each other.
195 */
196 if (xfs_filestream_get_ag(mp, ag) > 1) {
197 xfs_filestream_put_ag(mp, ag);
198 goto next_ag;
199 }
200
a1f69417 201 longest = xfs_alloc_longest_free_extent(pag,
3fd129b6
DW
202 xfs_alloc_min_freelist(mp, pag),
203 xfs_ag_resv_needed(pag, XFS_AG_RESV_NONE));
2a82b8be
DC
204 if (((minlen && longest >= minlen) ||
205 (!minlen && pag->pagf_freeblks >= minfree)) &&
206 (!pag->pagf_metadata || !(flags & XFS_PICK_USERDATA) ||
207 (flags & XFS_PICK_LOWSPACE))) {
208
209 /* Break out, retaining the reference on the AG. */
210 free = pag->pagf_freeblks;
4196ac08 211 xfs_perag_put(pag);
2a82b8be
DC
212 *agp = ag;
213 break;
214 }
215
216 /* Drop the reference on this AG, it's not usable. */
217 xfs_filestream_put_ag(mp, ag);
218next_ag:
4196ac08 219 xfs_perag_put(pag);
2a82b8be
DC
220 /* Move to the next AG, wrapping to AG 0 if necessary. */
221 if (++ag >= mp->m_sb.sb_agcount)
222 ag = 0;
223
224 /* If a full pass of the AGs hasn't been done yet, continue. */
225 if (ag != startag)
226 continue;
227
228 /* Allow sleeping in xfs_alloc_pagf_init() on the 2nd pass. */
229 if (trylock != 0) {
230 trylock = 0;
231 continue;
232 }
233
234 /* Finally, if lowspace wasn't set, set it for the 3rd pass. */
235 if (!(flags & XFS_PICK_LOWSPACE)) {
236 flags |= XFS_PICK_LOWSPACE;
237 continue;
238 }
239
240 /*
241 * Take the AG with the most free space, regardless of whether
242 * it's already in use by another filestream.
243 */
244 if (max_ag != NULLAGNUMBER) {
245 xfs_filestream_get_ag(mp, max_ag);
2a82b8be
DC
246 free = maxfree;
247 *agp = max_ag;
248 break;
249 }
250
251 /* take AG 0 if none matched */
b94acd47 252 trace_xfs_filestream_pick(ip, *agp, free, nscan);
2a82b8be
DC
253 *agp = 0;
254 return 0;
255 }
256
b94acd47 257 trace_xfs_filestream_pick(ip, *agp, free, nscan);
2a82b8be 258
2cd2ef6a 259 if (*agp == NULLAGNUMBER)
2a82b8be 260 return 0;
2a82b8be 261
2451337d 262 err = -ENOMEM;
1919adda 263 item = kmem_alloc(sizeof(*item), KM_MAYFAIL);
2a82b8be 264 if (!item)
2cd2ef6a 265 goto out_put_ag;
2a82b8be 266
2cd2ef6a 267 item->ag = *agp;
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 278out_free_item:
1919adda 279 kmem_free(item);
2cd2ef6a
CH
280out_put_ag:
281 xfs_filestream_put_ag(mp, *agp);
282 return err;
2a82b8be
DC
283}
284
2cd2ef6a
CH
285static struct xfs_inode *
286xfs_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
2b0143b5 300 dir = igrab(d_inode(parent));
2cd2ef6a 301 dput(parent);
2a82b8be 302
2cd2ef6a
CH
303out_dput:
304 dput(dentry);
305out:
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 */
315xfs_agnumber_t
316xfs_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
c19b3b05 324 ASSERT(S_ISREG(VFS_I(ip)->i_mode));
2a82b8be 325
2cd2ef6a
CH
326 pip = xfs_filestream_get_parent(ip);
327 if (!pip)
b26384dc 328 return NULLAGNUMBER;
2cd2ef6a
CH
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
7fcd3efa 335 trace_xfs_filestream_lookup(mp, ip->i_ino, 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;
353out:
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 */
364int
365xfs_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;
292378ed
DC
373 int flags = 0;
374 int err = 0;
2cd2ef6a 375 struct xfs_mru_cache_elem *mru;
2a82b8be 376
2cd2ef6a 377 *agp = NULLAGNUMBER;
2a82b8be 378
2cd2ef6a
CH
379 pip = xfs_filestream_get_parent(ip);
380 if (!pip)
381 goto exit;
2a82b8be 382
2cd2ef6a
CH
383 mru = xfs_mru_cache_remove(mp->m_filestream, pip->i_ino);
384 if (mru) {
385 struct xfs_fstrm_item *item =
386 container_of(mru, struct xfs_fstrm_item, mru);
387 startag = (item->ag + 1) % mp->m_sb.sb_agcount;
2a82b8be
DC
388 }
389
292378ed
DC
390 if (xfs_alloc_is_userdata(ap->datatype))
391 flags |= XFS_PICK_USERDATA;
392 if (ap->dfops->dop_low)
393 flags |= XFS_PICK_LOWSPACE;
2a82b8be 394
2cd2ef6a 395 err = xfs_filestream_pick_ag(pip, startag, agp, flags, minlen);
2a82b8be
DC
396
397 /*
2cd2ef6a 398 * Only free the item here so we skip over the old AG earlier.
2a82b8be 399 */
2cd2ef6a 400 if (mru)
7fcd3efa 401 xfs_fstrm_free_func(mp, mru);
2a82b8be 402
2cd2ef6a
CH
403 IRELE(pip);
404exit:
405 if (*agp == NULLAGNUMBER)
406 *agp = 0;
407 return err;
408}
2a82b8be 409
2cd2ef6a
CH
410void
411xfs_filestream_deassociate(
412 struct xfs_inode *ip)
413{
414 xfs_mru_cache_delete(ip->i_mount->m_filestream, ip->i_ino);
415}
2a82b8be 416
2cd2ef6a
CH
417int
418xfs_filestream_mount(
419 xfs_mount_t *mp)
420{
2a82b8be 421 /*
2cd2ef6a
CH
422 * The filestream timer tunable is currently fixed within the range of
423 * one second to four minutes, with five seconds being the default. The
424 * group count is somewhat arbitrary, but it'd be nice to adhere to the
425 * timer tunable to within about 10 percent. This requires at least 10
426 * groups.
2a82b8be 427 */
7fcd3efa
CH
428 return xfs_mru_cache_create(&mp->m_filestream, mp,
429 xfs_fstrm_centisecs * 10, 10, xfs_fstrm_free_func);
2cd2ef6a 430}
2a82b8be 431
2cd2ef6a
CH
432void
433xfs_filestream_unmount(
434 xfs_mount_t *mp)
435{
436 xfs_mru_cache_destroy(mp->m_filestream);
437}