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
22 struct xfs_fstrm_item {
23 struct xfs_mru_cache_elem mru;
24 xfs_agnumber_t ag; /* AG in use for this directory */
25 };
26
27 enum xfs_fstrm_alloc {
28 XFS_PICK_USERDATA = 1,
29 XFS_PICK_LOWSPACE = 2,
30 };
31
32 /*
33 * Allocation group filestream associations are tracked with per-ag atomic
34 * counters. These counters allow xfs_filestream_pick_ag() to tell whether a
35 * particular AG already has active filestreams associated with it. The mount
36 * point's m_peraglock is used to protect these counters from per-ag array
37 * re-allocation during a growfs operation. When xfs_growfs_data_private() is
38 * about to reallocate the array, it calls xfs_filestream_flush() with the
39 * m_peraglock held in write mode.
40 *
41 * Since xfs_mru_cache_flush() guarantees that all the free functions for all
42 * the cache elements have finished executing before it returns, it's safe for
43 * the free functions to use the atomic counters without m_peraglock protection.
44 * This allows the implementation of xfs_fstrm_free_func() to be agnostic about
45 * whether it was called with the m_peraglock held in read mode, write mode or
46 * not held at all. The race condition this addresses is the following:
47 *
48 * - The work queue scheduler fires and pulls a filestream directory cache
49 * element off the LRU end of the cache for deletion, then gets pre-empted.
50 * - A growfs operation grabs the m_peraglock in write mode, flushes all the
51 * remaining items from the cache and reallocates the mount point's per-ag
52 * array, resetting all the counters to zero.
53 * - The work queue thread resumes and calls the free function for the element
54 * it started cleaning up earlier. In the process it decrements the
55 * filestreams counter for an AG that now has no references.
56 *
57 * With a shrinkfs feature, the above scenario could panic the system.
58 *
59 * All other uses of the following macros should be protected by either the
60 * m_peraglock held in read mode, or the cache's internal locking exposed by the
61 * interval between a call to xfs_mru_cache_lookup() and a call to
62 * xfs_mru_cache_done(). In addition, the m_peraglock must be held in read mode
63 * when new elements are added to the cache.
64 *
65 * Combined, these locking rules ensure that no associations will ever exist in
66 * the cache that reference per-ag array elements that have since been
67 * reallocated.
68 */
69 int
70 xfs_filestream_peek_ag(
71 xfs_mount_t *mp,
72 xfs_agnumber_t agno)
73 {
74 struct xfs_perag *pag;
75 int ret;
76
77 pag = xfs_perag_get(mp, agno);
78 ret = atomic_read(&pag->pagf_fstrms);
79 xfs_perag_put(pag);
80 return ret;
81 }
82
83 static int
84 xfs_filestream_get_ag(
85 xfs_mount_t *mp,
86 xfs_agnumber_t agno)
87 {
88 struct xfs_perag *pag;
89 int ret;
90
91 pag = xfs_perag_get(mp, agno);
92 ret = atomic_inc_return(&pag->pagf_fstrms);
93 xfs_perag_put(pag);
94 return ret;
95 }
96
97 static void
98 xfs_filestream_put_ag(
99 xfs_mount_t *mp,
100 xfs_agnumber_t agno)
101 {
102 struct xfs_perag *pag;
103
104 pag = xfs_perag_get(mp, agno);
105 atomic_dec(&pag->pagf_fstrms);
106 xfs_perag_put(pag);
107 }
108
109 static void
110 xfs_fstrm_free_func(
111 void *data,
112 struct xfs_mru_cache_elem *mru)
113 {
114 struct xfs_mount *mp = data;
115 struct xfs_fstrm_item *item =
116 container_of(mru, struct xfs_fstrm_item, mru);
117
118 xfs_filestream_put_ag(mp, item->ag);
119 trace_xfs_filestream_free(mp, mru->key, item->ag);
120
121 kmem_free(item);
122 }
123
124 /*
125 * Scan the AGs starting at startag looking for an AG that isn't in use and has
126 * at least minlen blocks free.
127 */
128 static int
129 xfs_filestream_pick_ag(
130 struct xfs_inode *ip,
131 xfs_agnumber_t startag,
132 xfs_agnumber_t *agp,
133 int flags,
134 xfs_extlen_t minlen)
135 {
136 struct xfs_mount *mp = ip->i_mount;
137 struct xfs_fstrm_item *item;
138 struct xfs_perag *pag;
139 xfs_extlen_t longest, free = 0, minfree, maxfree = 0;
140 xfs_agnumber_t ag, max_ag = NULLAGNUMBER;
141 int err, trylock, nscan;
142
143 ASSERT(S_ISDIR(VFS_I(ip)->i_mode));
144
145 /* 2% of an AG's blocks must be free for it to be chosen. */
146 minfree = mp->m_sb.sb_agblocks / 50;
147
148 ag = startag;
149 *agp = NULLAGNUMBER;
150
151 /* For the first pass, don't sleep trying to init the per-AG. */
152 trylock = XFS_ALLOC_FLAG_TRYLOCK;
153
154 for (nscan = 0; 1; nscan++) {
155 trace_xfs_filestream_scan(mp, ip->i_ino, ag);
156
157 pag = xfs_perag_get(mp, ag);
158
159 if (!pag->pagf_init) {
160 err = xfs_alloc_pagf_init(mp, NULL, ag, trylock);
161 if (err && !trylock) {
162 xfs_perag_put(pag);
163 return err;
164 }
165 }
166
167 /* Might fail sometimes during the 1st pass with trylock set. */
168 if (!pag->pagf_init)
169 goto next_ag;
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);
205 next_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
265 out_free_item:
266 kmem_free(item);
267 out_put_ag:
268 xfs_filestream_put_ag(mp, *agp);
269 return err;
270 }
271
272 static struct xfs_inode *
273 xfs_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
290 out_dput:
291 dput(dentry);
292 out:
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 */
302 xfs_agnumber_t
303 xfs_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;
340 out:
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 */
351 int
352 xfs_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 (xfs_alloc_is_userdata(ap->datatype))
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);
391 exit:
392 if (*agp == NULLAGNUMBER)
393 *agp = 0;
394 return err;
395 }
396
397 void
398 xfs_filestream_deassociate(
399 struct xfs_inode *ip)
400 {
401 xfs_mru_cache_delete(ip->i_mount->m_filestream, ip->i_ino);
402 }
403
404 int
405 xfs_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
419 void
420 xfs_filestream_unmount(
421 xfs_mount_t *mp)
422 {
423 xfs_mru_cache_destroy(mp->m_filestream);
424 }