Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (c) 2006-2007 Silicon Graphics, Inc.
4 * Copyright (c) 2014 Christoph Hellwig.
5 * All Rights Reserved.
6 */
7#include "xfs.h"
8#include "xfs_shared.h"
9#include "xfs_format.h"
10#include "xfs_log_format.h"
11#include "xfs_trans_resv.h"
12#include "xfs_mount.h"
13#include "xfs_inode.h"
14#include "xfs_bmap.h"
15#include "xfs_bmap_util.h"
16#include "xfs_alloc.h"
17#include "xfs_mru_cache.h"
18#include "xfs_trace.h"
19#include "xfs_ag.h"
20#include "xfs_ag_resv.h"
21#include "xfs_trans.h"
22#include "xfs_filestream.h"
23
24struct xfs_fstrm_item {
25 struct xfs_mru_cache_elem mru;
26 struct xfs_perag *pag; /* AG in use for this directory */
27};
28
29enum xfs_fstrm_alloc {
30 XFS_PICK_USERDATA = 1,
31 XFS_PICK_LOWSPACE = 2,
32};
33
34static void
35xfs_fstrm_free_func(
36 void *data,
37 struct xfs_mru_cache_elem *mru)
38{
39 struct xfs_fstrm_item *item =
40 container_of(mru, struct xfs_fstrm_item, mru);
41 struct xfs_perag *pag = item->pag;
42
43 trace_xfs_filestream_free(pag, mru->key);
44 atomic_dec(&pag->pagf_fstrms);
45 xfs_perag_rele(pag);
46
47 kfree(item);
48}
49
50/*
51 * Scan the AGs starting at start_agno looking for an AG that isn't in use and
52 * has at least minlen blocks free. If no AG is found to match the allocation
53 * requirements, pick the AG with the most free space in it.
54 */
55static int
56xfs_filestream_pick_ag(
57 struct xfs_alloc_arg *args,
58 xfs_ino_t pino,
59 xfs_agnumber_t start_agno,
60 int flags,
61 xfs_extlen_t *longest)
62{
63 struct xfs_mount *mp = args->mp;
64 struct xfs_perag *pag;
65 struct xfs_perag *max_pag = NULL;
66 xfs_extlen_t minlen = *longest;
67 xfs_extlen_t minfree, maxfree = 0;
68 xfs_agnumber_t agno;
69 bool first_pass = true;
70
71 /* 2% of an AG's blocks must be free for it to be chosen. */
72 minfree = mp->m_sb.sb_agblocks / 50;
73
74restart:
75 for_each_perag_wrap(mp, start_agno, agno, pag) {
76 int err;
77
78 trace_xfs_filestream_scan(pag, pino);
79
80 *longest = 0;
81 err = xfs_bmap_longest_free_extent(pag, NULL, longest);
82 if (err) {
83 if (err == -EAGAIN) {
84 /* Couldn't lock the AGF, skip this AG. */
85 err = 0;
86 continue;
87 }
88 xfs_perag_rele(pag);
89 if (max_pag)
90 xfs_perag_rele(max_pag);
91 return err;
92 }
93
94 /* Keep track of the AG with the most free blocks. */
95 if (pag->pagf_freeblks > maxfree) {
96 maxfree = pag->pagf_freeblks;
97 if (max_pag)
98 xfs_perag_rele(max_pag);
99 atomic_inc(&pag_group(pag)->xg_active_ref);
100 max_pag = pag;
101 }
102
103 /*
104 * The AG reference count does two things: it enforces mutual
105 * exclusion when examining the suitability of an AG in this
106 * loop, and it guards against two filestreams being established
107 * in the same AG as each other.
108 */
109 if (atomic_inc_return(&pag->pagf_fstrms) <= 1) {
110 if (((minlen && *longest >= minlen) ||
111 (!minlen && pag->pagf_freeblks >= minfree)) &&
112 (!xfs_perag_prefers_metadata(pag) ||
113 !(flags & XFS_PICK_USERDATA) ||
114 (flags & XFS_PICK_LOWSPACE))) {
115 /* Break out, retaining the reference on the AG. */
116 if (max_pag)
117 xfs_perag_rele(max_pag);
118 goto done;
119 }
120 }
121
122 /* Drop the reference on this AG, it's not usable. */
123 atomic_dec(&pag->pagf_fstrms);
124 }
125
126 /*
127 * Allow a second pass to give xfs_bmap_longest_free_extent() another
128 * attempt at locking AGFs that it might have skipped over before we
129 * fail.
130 */
131 if (first_pass) {
132 first_pass = false;
133 goto restart;
134 }
135
136 /*
137 * We must be low on data space, so run a final lowspace optimised
138 * selection pass if we haven't already.
139 */
140 if (!(flags & XFS_PICK_LOWSPACE)) {
141 flags |= XFS_PICK_LOWSPACE;
142 goto restart;
143 }
144
145 /*
146 * No unassociated AGs are available, so select the AG with the most
147 * free space, regardless of whether it's already in use by another
148 * filestream. It none suit, just use whatever AG we can grab.
149 */
150 if (!max_pag) {
151 for_each_perag_wrap(args->mp, 0, start_agno, pag) {
152 max_pag = pag;
153 break;
154 }
155
156 /* Bail if there are no AGs at all to select from. */
157 if (!max_pag)
158 return -ENOSPC;
159 }
160
161 pag = max_pag;
162 atomic_inc(&pag->pagf_fstrms);
163done:
164 trace_xfs_filestream_pick(pag, pino);
165 args->pag = pag;
166 return 0;
167}
168
169static struct xfs_inode *
170xfs_filestream_get_parent(
171 struct xfs_inode *ip)
172{
173 struct inode *inode = VFS_I(ip), *dir = NULL;
174 struct dentry *dentry, *parent;
175
176 dentry = d_find_alias(inode);
177 if (!dentry)
178 goto out;
179
180 parent = dget_parent(dentry);
181 if (!parent)
182 goto out_dput;
183
184 dir = igrab(d_inode(parent));
185 dput(parent);
186
187out_dput:
188 dput(dentry);
189out:
190 return dir ? XFS_I(dir) : NULL;
191}
192
193/*
194 * Lookup the mru cache for an existing association. If one exists and we can
195 * use it, return with an active perag reference indicating that the allocation
196 * will proceed with that association.
197 *
198 * If we have no association, or we cannot use the current one and have to
199 * destroy it, return with longest = 0 to tell the caller to create a new
200 * association.
201 */
202static int
203xfs_filestream_lookup_association(
204 struct xfs_bmalloca *ap,
205 struct xfs_alloc_arg *args,
206 xfs_ino_t pino,
207 xfs_extlen_t *longest)
208{
209 struct xfs_mount *mp = args->mp;
210 struct xfs_perag *pag;
211 struct xfs_mru_cache_elem *mru;
212 int error = 0;
213
214 *longest = 0;
215 mru = xfs_mru_cache_lookup(mp->m_filestream, pino);
216 if (!mru)
217 return 0;
218 /*
219 * Grab the pag and take an extra active reference for the caller whilst
220 * the mru item cannot go away. This means we'll pin the perag with
221 * the reference we get here even if the filestreams association is torn
222 * down immediately after we mark the lookup as done.
223 */
224 pag = container_of(mru, struct xfs_fstrm_item, mru)->pag;
225 atomic_inc(&pag_group(pag)->xg_active_ref);
226 xfs_mru_cache_done(mp->m_filestream);
227
228 trace_xfs_filestream_lookup(pag, ap->ip->i_ino);
229
230 ap->blkno = xfs_agbno_to_fsb(pag, 0);
231 xfs_bmap_adjacent(ap);
232
233 /*
234 * If there is very little free space before we start a filestreams
235 * allocation, we're almost guaranteed to fail to find a large enough
236 * free space available so just use the cached AG.
237 */
238 if (ap->tp->t_flags & XFS_TRANS_LOWMODE) {
239 *longest = 1;
240 goto out_done;
241 }
242
243 error = xfs_bmap_longest_free_extent(pag, args->tp, longest);
244 if (error == -EAGAIN)
245 error = 0;
246 if (error || *longest < args->maxlen) {
247 /* We aren't going to use this perag */
248 *longest = 0;
249 xfs_perag_rele(pag);
250 return error;
251 }
252
253out_done:
254 args->pag = pag;
255 return 0;
256}
257
258static int
259xfs_filestream_create_association(
260 struct xfs_bmalloca *ap,
261 struct xfs_alloc_arg *args,
262 xfs_ino_t pino,
263 xfs_extlen_t *longest)
264{
265 struct xfs_mount *mp = args->mp;
266 struct xfs_mru_cache_elem *mru;
267 struct xfs_fstrm_item *item;
268 xfs_agnumber_t agno = XFS_INO_TO_AGNO(mp, pino);
269 int flags = 0;
270 int error;
271
272 /* Changing parent AG association now, so remove the existing one. */
273 mru = xfs_mru_cache_remove(mp->m_filestream, pino);
274 if (mru) {
275 struct xfs_fstrm_item *item =
276 container_of(mru, struct xfs_fstrm_item, mru);
277
278 agno = (pag_agno(item->pag) + 1) % mp->m_sb.sb_agcount;
279 xfs_fstrm_free_func(mp, mru);
280 } else if (xfs_is_inode32(mp)) {
281 xfs_agnumber_t rotorstep = xfs_rotorstep;
282
283 agno = (mp->m_agfrotor / rotorstep) % mp->m_sb.sb_agcount;
284 mp->m_agfrotor = (mp->m_agfrotor + 1) %
285 (mp->m_sb.sb_agcount * rotorstep);
286 }
287
288 ap->blkno = XFS_AGB_TO_FSB(args->mp, agno, 0);
289 xfs_bmap_adjacent(ap);
290
291 if (ap->datatype & XFS_ALLOC_USERDATA)
292 flags |= XFS_PICK_USERDATA;
293 if (ap->tp->t_flags & XFS_TRANS_LOWMODE)
294 flags |= XFS_PICK_LOWSPACE;
295
296 *longest = ap->length;
297 error = xfs_filestream_pick_ag(args, pino, agno, flags, longest);
298 if (error)
299 return error;
300
301 /*
302 * We are going to use this perag now, so create an assoication for it.
303 * xfs_filestream_pick_ag() has already bumped the perag fstrms counter
304 * for us, so all we need to do here is take another active reference to
305 * the perag for the cached association.
306 *
307 * If we fail to store the association, we need to drop the fstrms
308 * counter as well as drop the perag reference we take here for the
309 * item. We do not need to return an error for this failure - as long as
310 * we return a referenced AG, the allocation can still go ahead just
311 * fine.
312 */
313 item = kmalloc(sizeof(*item), GFP_KERNEL | __GFP_RETRY_MAYFAIL);
314 if (!item)
315 goto out_put_fstrms;
316
317 atomic_inc(&pag_group(args->pag)->xg_active_ref);
318 item->pag = args->pag;
319 error = xfs_mru_cache_insert(mp->m_filestream, pino, &item->mru);
320 if (error)
321 goto out_free_item;
322 return 0;
323
324out_free_item:
325 xfs_perag_rele(item->pag);
326 kfree(item);
327out_put_fstrms:
328 atomic_dec(&args->pag->pagf_fstrms);
329 return 0;
330}
331
332/*
333 * Search for an allocation group with a single extent large enough for
334 * the request. First we look for an existing association and use that if it
335 * is found. Otherwise, we create a new association by selecting an AG that fits
336 * the allocation criteria.
337 *
338 * We return with a referenced perag in args->pag to indicate which AG we are
339 * allocating into or an error with no references held.
340 */
341int
342xfs_filestream_select_ag(
343 struct xfs_bmalloca *ap,
344 struct xfs_alloc_arg *args,
345 xfs_extlen_t *longest)
346{
347 struct xfs_inode *pip;
348 xfs_ino_t ino = 0;
349 int error = 0;
350
351 *longest = 0;
352 args->total = ap->total;
353 pip = xfs_filestream_get_parent(ap->ip);
354 if (pip) {
355 ino = pip->i_ino;
356 error = xfs_filestream_lookup_association(ap, args, ino,
357 longest);
358 xfs_irele(pip);
359 if (error)
360 return error;
361 if (*longest >= args->maxlen)
362 goto out_select;
363 if (ap->tp->t_flags & XFS_TRANS_LOWMODE)
364 goto out_select;
365 }
366
367 error = xfs_filestream_create_association(ap, args, ino, longest);
368 if (error)
369 return error;
370
371out_select:
372 ap->blkno = xfs_agbno_to_fsb(args->pag, 0);
373 return 0;
374}
375
376void
377xfs_filestream_deassociate(
378 struct xfs_inode *ip)
379{
380 xfs_mru_cache_delete(ip->i_mount->m_filestream, ip->i_ino);
381}
382
383int
384xfs_filestream_mount(
385 xfs_mount_t *mp)
386{
387 /*
388 * The filestream timer tunable is currently fixed within the range of
389 * one second to four minutes, with five seconds being the default. The
390 * group count is somewhat arbitrary, but it'd be nice to adhere to the
391 * timer tunable to within about 10 percent. This requires at least 10
392 * groups.
393 */
394 return xfs_mru_cache_create(&mp->m_filestream, mp,
395 xfs_fstrm_centisecs * 10, 10, xfs_fstrm_free_func);
396}
397
398void
399xfs_filestream_unmount(
400 xfs_mount_t *mp)
401{
402 xfs_mru_cache_destroy(mp->m_filestream);
403}
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (c) 2006-2007 Silicon Graphics, Inc.
4 * Copyright (c) 2014 Christoph Hellwig.
5 * All Rights Reserved.
6 */
7#include "xfs.h"
8#include "xfs_shared.h"
9#include "xfs_format.h"
10#include "xfs_log_format.h"
11#include "xfs_trans_resv.h"
12#include "xfs_sb.h"
13#include "xfs_mount.h"
14#include "xfs_inode.h"
15#include "xfs_bmap.h"
16#include "xfs_alloc.h"
17#include "xfs_mru_cache.h"
18#include "xfs_trace.h"
19#include "xfs_ag_resv.h"
20#include "xfs_trans.h"
21#include "xfs_filestream.h"
22
23struct xfs_fstrm_item {
24 struct xfs_mru_cache_elem mru;
25 xfs_agnumber_t ag; /* AG in use for this directory */
26};
27
28enum xfs_fstrm_alloc {
29 XFS_PICK_USERDATA = 1,
30 XFS_PICK_LOWSPACE = 2,
31};
32
33/*
34 * Allocation group filestream associations are tracked with per-ag atomic
35 * counters. These counters allow xfs_filestream_pick_ag() to tell whether a
36 * particular AG already has active filestreams associated with it. The mount
37 * point's m_peraglock is used to protect these counters from per-ag array
38 * re-allocation during a growfs operation. When xfs_growfs_data_private() is
39 * about to reallocate the array, it calls xfs_filestream_flush() with the
40 * m_peraglock held in write mode.
41 *
42 * Since xfs_mru_cache_flush() guarantees that all the free functions for all
43 * the cache elements have finished executing before it returns, it's safe for
44 * the free functions to use the atomic counters without m_peraglock protection.
45 * This allows the implementation of xfs_fstrm_free_func() to be agnostic about
46 * whether it was called with the m_peraglock held in read mode, write mode or
47 * not held at all. The race condition this addresses is the following:
48 *
49 * - The work queue scheduler fires and pulls a filestream directory cache
50 * element off the LRU end of the cache for deletion, then gets pre-empted.
51 * - A growfs operation grabs the m_peraglock in write mode, flushes all the
52 * remaining items from the cache and reallocates the mount point's per-ag
53 * array, resetting all the counters to zero.
54 * - The work queue thread resumes and calls the free function for the element
55 * it started cleaning up earlier. In the process it decrements the
56 * filestreams counter for an AG that now has no references.
57 *
58 * With a shrinkfs feature, the above scenario could panic the system.
59 *
60 * All other uses of the following macros should be protected by either the
61 * m_peraglock held in read mode, or the cache's internal locking exposed by the
62 * interval between a call to xfs_mru_cache_lookup() and a call to
63 * xfs_mru_cache_done(). In addition, the m_peraglock must be held in read mode
64 * when new elements are added to the cache.
65 *
66 * Combined, these locking rules ensure that no associations will ever exist in
67 * the cache that reference per-ag array elements that have since been
68 * reallocated.
69 */
70int
71xfs_filestream_peek_ag(
72 xfs_mount_t *mp,
73 xfs_agnumber_t agno)
74{
75 struct xfs_perag *pag;
76 int ret;
77
78 pag = xfs_perag_get(mp, agno);
79 ret = atomic_read(&pag->pagf_fstrms);
80 xfs_perag_put(pag);
81 return ret;
82}
83
84static int
85xfs_filestream_get_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_inc_return(&pag->pagf_fstrms);
94 xfs_perag_put(pag);
95 return ret;
96}
97
98static void
99xfs_filestream_put_ag(
100 xfs_mount_t *mp,
101 xfs_agnumber_t agno)
102{
103 struct xfs_perag *pag;
104
105 pag = xfs_perag_get(mp, agno);
106 atomic_dec(&pag->pagf_fstrms);
107 xfs_perag_put(pag);
108}
109
110static void
111xfs_fstrm_free_func(
112 void *data,
113 struct xfs_mru_cache_elem *mru)
114{
115 struct xfs_mount *mp = data;
116 struct xfs_fstrm_item *item =
117 container_of(mru, struct xfs_fstrm_item, mru);
118
119 xfs_filestream_put_ag(mp, item->ag);
120 trace_xfs_filestream_free(mp, mru->key, item->ag);
121
122 kmem_free(item);
123}
124
125/*
126 * Scan the AGs starting at startag looking for an AG that isn't in use and has
127 * at least minlen blocks free.
128 */
129static int
130xfs_filestream_pick_ag(
131 struct xfs_inode *ip,
132 xfs_agnumber_t startag,
133 xfs_agnumber_t *agp,
134 int flags,
135 xfs_extlen_t minlen)
136{
137 struct xfs_mount *mp = ip->i_mount;
138 struct xfs_fstrm_item *item;
139 struct xfs_perag *pag;
140 xfs_extlen_t longest, free = 0, minfree, maxfree = 0;
141 xfs_agnumber_t ag, max_ag = NULLAGNUMBER;
142 int err, trylock, nscan;
143
144 ASSERT(S_ISDIR(VFS_I(ip)->i_mode));
145
146 /* 2% of an AG's blocks must be free for it to be chosen. */
147 minfree = mp->m_sb.sb_agblocks / 50;
148
149 ag = startag;
150 *agp = NULLAGNUMBER;
151
152 /* For the first pass, don't sleep trying to init the per-AG. */
153 trylock = XFS_ALLOC_FLAG_TRYLOCK;
154
155 for (nscan = 0; 1; nscan++) {
156 trace_xfs_filestream_scan(mp, ip->i_ino, ag);
157
158 pag = xfs_perag_get(mp, ag);
159
160 if (!pag->pagf_init) {
161 err = xfs_alloc_pagf_init(mp, NULL, ag, trylock);
162 if (err) {
163 xfs_perag_put(pag);
164 if (err != -EAGAIN)
165 return err;
166 /* Couldn't lock the AGF, skip this AG. */
167 continue;
168 }
169 }
170
171 /* Keep track of the AG with the most free blocks. */
172 if (pag->pagf_freeblks > maxfree) {
173 maxfree = pag->pagf_freeblks;
174 max_ag = ag;
175 }
176
177 /*
178 * The AG reference count does two things: it enforces mutual
179 * exclusion when examining the suitability of an AG in this
180 * loop, and it guards against two filestreams being established
181 * in the same AG as each other.
182 */
183 if (xfs_filestream_get_ag(mp, ag) > 1) {
184 xfs_filestream_put_ag(mp, ag);
185 goto next_ag;
186 }
187
188 longest = xfs_alloc_longest_free_extent(pag,
189 xfs_alloc_min_freelist(mp, pag),
190 xfs_ag_resv_needed(pag, XFS_AG_RESV_NONE));
191 if (((minlen && longest >= minlen) ||
192 (!minlen && pag->pagf_freeblks >= minfree)) &&
193 (!pag->pagf_metadata || !(flags & XFS_PICK_USERDATA) ||
194 (flags & XFS_PICK_LOWSPACE))) {
195
196 /* Break out, retaining the reference on the AG. */
197 free = pag->pagf_freeblks;
198 xfs_perag_put(pag);
199 *agp = ag;
200 break;
201 }
202
203 /* Drop the reference on this AG, it's not usable. */
204 xfs_filestream_put_ag(mp, ag);
205next_ag:
206 xfs_perag_put(pag);
207 /* Move to the next AG, wrapping to AG 0 if necessary. */
208 if (++ag >= mp->m_sb.sb_agcount)
209 ag = 0;
210
211 /* If a full pass of the AGs hasn't been done yet, continue. */
212 if (ag != startag)
213 continue;
214
215 /* Allow sleeping in xfs_alloc_pagf_init() on the 2nd pass. */
216 if (trylock != 0) {
217 trylock = 0;
218 continue;
219 }
220
221 /* Finally, if lowspace wasn't set, set it for the 3rd pass. */
222 if (!(flags & XFS_PICK_LOWSPACE)) {
223 flags |= XFS_PICK_LOWSPACE;
224 continue;
225 }
226
227 /*
228 * Take the AG with the most free space, regardless of whether
229 * it's already in use by another filestream.
230 */
231 if (max_ag != NULLAGNUMBER) {
232 xfs_filestream_get_ag(mp, max_ag);
233 free = maxfree;
234 *agp = max_ag;
235 break;
236 }
237
238 /* take AG 0 if none matched */
239 trace_xfs_filestream_pick(ip, *agp, free, nscan);
240 *agp = 0;
241 return 0;
242 }
243
244 trace_xfs_filestream_pick(ip, *agp, free, nscan);
245
246 if (*agp == NULLAGNUMBER)
247 return 0;
248
249 err = -ENOMEM;
250 item = kmem_alloc(sizeof(*item), KM_MAYFAIL);
251 if (!item)
252 goto out_put_ag;
253
254 item->ag = *agp;
255
256 err = xfs_mru_cache_insert(mp->m_filestream, ip->i_ino, &item->mru);
257 if (err) {
258 if (err == -EEXIST)
259 err = 0;
260 goto out_free_item;
261 }
262
263 return 0;
264
265out_free_item:
266 kmem_free(item);
267out_put_ag:
268 xfs_filestream_put_ag(mp, *agp);
269 return err;
270}
271
272static struct xfs_inode *
273xfs_filestream_get_parent(
274 struct xfs_inode *ip)
275{
276 struct inode *inode = VFS_I(ip), *dir = NULL;
277 struct dentry *dentry, *parent;
278
279 dentry = d_find_alias(inode);
280 if (!dentry)
281 goto out;
282
283 parent = dget_parent(dentry);
284 if (!parent)
285 goto out_dput;
286
287 dir = igrab(d_inode(parent));
288 dput(parent);
289
290out_dput:
291 dput(dentry);
292out:
293 return dir ? XFS_I(dir) : NULL;
294}
295
296/*
297 * Find the right allocation group for a file, either by finding an
298 * existing file stream or creating a new one.
299 *
300 * Returns NULLAGNUMBER in case of an error.
301 */
302xfs_agnumber_t
303xfs_filestream_lookup_ag(
304 struct xfs_inode *ip)
305{
306 struct xfs_mount *mp = ip->i_mount;
307 struct xfs_inode *pip = NULL;
308 xfs_agnumber_t startag, ag = NULLAGNUMBER;
309 struct xfs_mru_cache_elem *mru;
310
311 ASSERT(S_ISREG(VFS_I(ip)->i_mode));
312
313 pip = xfs_filestream_get_parent(ip);
314 if (!pip)
315 return NULLAGNUMBER;
316
317 mru = xfs_mru_cache_lookup(mp->m_filestream, pip->i_ino);
318 if (mru) {
319 ag = container_of(mru, struct xfs_fstrm_item, mru)->ag;
320 xfs_mru_cache_done(mp->m_filestream);
321
322 trace_xfs_filestream_lookup(mp, ip->i_ino, ag);
323 goto out;
324 }
325
326 /*
327 * Set the starting AG using the rotor for inode32, otherwise
328 * use the directory inode's AG.
329 */
330 if (mp->m_flags & XFS_MOUNT_32BITINODES) {
331 xfs_agnumber_t rotorstep = xfs_rotorstep;
332 startag = (mp->m_agfrotor / rotorstep) % mp->m_sb.sb_agcount;
333 mp->m_agfrotor = (mp->m_agfrotor + 1) %
334 (mp->m_sb.sb_agcount * rotorstep);
335 } else
336 startag = XFS_INO_TO_AGNO(mp, pip->i_ino);
337
338 if (xfs_filestream_pick_ag(pip, startag, &ag, 0, 0))
339 ag = NULLAGNUMBER;
340out:
341 xfs_irele(pip);
342 return ag;
343}
344
345/*
346 * Pick a new allocation group for the current file and its file stream.
347 *
348 * This is called when the allocator can't find a suitable extent in the
349 * current AG, and we have to move the stream into a new AG with more space.
350 */
351int
352xfs_filestream_new_ag(
353 struct xfs_bmalloca *ap,
354 xfs_agnumber_t *agp)
355{
356 struct xfs_inode *ip = ap->ip, *pip;
357 struct xfs_mount *mp = ip->i_mount;
358 xfs_extlen_t minlen = ap->length;
359 xfs_agnumber_t startag = 0;
360 int flags = 0;
361 int err = 0;
362 struct xfs_mru_cache_elem *mru;
363
364 *agp = NULLAGNUMBER;
365
366 pip = xfs_filestream_get_parent(ip);
367 if (!pip)
368 goto exit;
369
370 mru = xfs_mru_cache_remove(mp->m_filestream, pip->i_ino);
371 if (mru) {
372 struct xfs_fstrm_item *item =
373 container_of(mru, struct xfs_fstrm_item, mru);
374 startag = (item->ag + 1) % mp->m_sb.sb_agcount;
375 }
376
377 if (ap->datatype & XFS_ALLOC_USERDATA)
378 flags |= XFS_PICK_USERDATA;
379 if (ap->tp->t_flags & XFS_TRANS_LOWMODE)
380 flags |= XFS_PICK_LOWSPACE;
381
382 err = xfs_filestream_pick_ag(pip, startag, agp, flags, minlen);
383
384 /*
385 * Only free the item here so we skip over the old AG earlier.
386 */
387 if (mru)
388 xfs_fstrm_free_func(mp, mru);
389
390 xfs_irele(pip);
391exit:
392 if (*agp == NULLAGNUMBER)
393 *agp = 0;
394 return err;
395}
396
397void
398xfs_filestream_deassociate(
399 struct xfs_inode *ip)
400{
401 xfs_mru_cache_delete(ip->i_mount->m_filestream, ip->i_ino);
402}
403
404int
405xfs_filestream_mount(
406 xfs_mount_t *mp)
407{
408 /*
409 * The filestream timer tunable is currently fixed within the range of
410 * one second to four minutes, with five seconds being the default. The
411 * group count is somewhat arbitrary, but it'd be nice to adhere to the
412 * timer tunable to within about 10 percent. This requires at least 10
413 * groups.
414 */
415 return xfs_mru_cache_create(&mp->m_filestream, mp,
416 xfs_fstrm_centisecs * 10, 10, xfs_fstrm_free_func);
417}
418
419void
420xfs_filestream_unmount(
421 xfs_mount_t *mp)
422{
423 xfs_mru_cache_destroy(mp->m_filestream);
424}