Loading...
1/*
2 * Copyright (c) 2006-2007 Silicon Graphics, Inc.
3 * All Rights Reserved.
4 *
5 * This program is free software; you can redistribute it and/or
6 * modify it under the terms of the GNU General Public License as
7 * published by the Free Software Foundation.
8 *
9 * This program is distributed in the hope that it would be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
13 *
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, write the Free Software Foundation,
16 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA
17 */
18#include "xfs.h"
19#include "xfs_bmap_btree.h"
20#include "xfs_inum.h"
21#include "xfs_dinode.h"
22#include "xfs_inode.h"
23#include "xfs_ag.h"
24#include "xfs_log.h"
25#include "xfs_trans.h"
26#include "xfs_sb.h"
27#include "xfs_mount.h"
28#include "xfs_bmap.h"
29#include "xfs_alloc.h"
30#include "xfs_utils.h"
31#include "xfs_mru_cache.h"
32#include "xfs_filestream.h"
33#include "xfs_trace.h"
34
35#ifdef XFS_FILESTREAMS_TRACE
36
37ktrace_t *xfs_filestreams_trace_buf;
38
39STATIC void
40xfs_filestreams_trace(
41 xfs_mount_t *mp, /* mount point */
42 int type, /* type of trace */
43 const char *func, /* source function */
44 int line, /* source line number */
45 __psunsigned_t arg0,
46 __psunsigned_t arg1,
47 __psunsigned_t arg2,
48 __psunsigned_t arg3,
49 __psunsigned_t arg4,
50 __psunsigned_t arg5)
51{
52 ktrace_enter(xfs_filestreams_trace_buf,
53 (void *)(__psint_t)(type | (line << 16)),
54 (void *)func,
55 (void *)(__psunsigned_t)current_pid(),
56 (void *)mp,
57 (void *)(__psunsigned_t)arg0,
58 (void *)(__psunsigned_t)arg1,
59 (void *)(__psunsigned_t)arg2,
60 (void *)(__psunsigned_t)arg3,
61 (void *)(__psunsigned_t)arg4,
62 (void *)(__psunsigned_t)arg5,
63 NULL, NULL, NULL, NULL, NULL, NULL);
64}
65
66#define TRACE0(mp,t) TRACE6(mp,t,0,0,0,0,0,0)
67#define TRACE1(mp,t,a0) TRACE6(mp,t,a0,0,0,0,0,0)
68#define TRACE2(mp,t,a0,a1) TRACE6(mp,t,a0,a1,0,0,0,0)
69#define TRACE3(mp,t,a0,a1,a2) TRACE6(mp,t,a0,a1,a2,0,0,0)
70#define TRACE4(mp,t,a0,a1,a2,a3) TRACE6(mp,t,a0,a1,a2,a3,0,0)
71#define TRACE5(mp,t,a0,a1,a2,a3,a4) TRACE6(mp,t,a0,a1,a2,a3,a4,0)
72#define TRACE6(mp,t,a0,a1,a2,a3,a4,a5) \
73 xfs_filestreams_trace(mp, t, __func__, __LINE__, \
74 (__psunsigned_t)a0, (__psunsigned_t)a1, \
75 (__psunsigned_t)a2, (__psunsigned_t)a3, \
76 (__psunsigned_t)a4, (__psunsigned_t)a5)
77
78#define TRACE_AG_SCAN(mp, ag, ag2) \
79 TRACE2(mp, XFS_FSTRM_KTRACE_AGSCAN, ag, ag2);
80#define TRACE_AG_PICK1(mp, max_ag, maxfree) \
81 TRACE2(mp, XFS_FSTRM_KTRACE_AGPICK1, max_ag, maxfree);
82#define TRACE_AG_PICK2(mp, ag, ag2, cnt, free, scan, flag) \
83 TRACE6(mp, XFS_FSTRM_KTRACE_AGPICK2, ag, ag2, \
84 cnt, free, scan, flag)
85#define TRACE_UPDATE(mp, ip, ag, cnt, ag2, cnt2) \
86 TRACE5(mp, XFS_FSTRM_KTRACE_UPDATE, ip, ag, cnt, ag2, cnt2)
87#define TRACE_FREE(mp, ip, pip, ag, cnt) \
88 TRACE4(mp, XFS_FSTRM_KTRACE_FREE, ip, pip, ag, cnt)
89#define TRACE_LOOKUP(mp, ip, pip, ag, cnt) \
90 TRACE4(mp, XFS_FSTRM_KTRACE_ITEM_LOOKUP, ip, pip, ag, cnt)
91#define TRACE_ASSOCIATE(mp, ip, pip, ag, cnt) \
92 TRACE4(mp, XFS_FSTRM_KTRACE_ASSOCIATE, ip, pip, ag, cnt)
93#define TRACE_MOVEAG(mp, ip, pip, oag, ocnt, nag, ncnt) \
94 TRACE6(mp, XFS_FSTRM_KTRACE_MOVEAG, ip, pip, oag, ocnt, nag, ncnt)
95#define TRACE_ORPHAN(mp, ip, ag) \
96 TRACE2(mp, XFS_FSTRM_KTRACE_ORPHAN, ip, ag);
97
98
99#else
100#define TRACE_AG_SCAN(mp, ag, ag2)
101#define TRACE_AG_PICK1(mp, max_ag, maxfree)
102#define TRACE_AG_PICK2(mp, ag, ag2, cnt, free, scan, flag)
103#define TRACE_UPDATE(mp, ip, ag, cnt, ag2, cnt2)
104#define TRACE_FREE(mp, ip, pip, ag, cnt)
105#define TRACE_LOOKUP(mp, ip, pip, ag, cnt)
106#define TRACE_ASSOCIATE(mp, ip, pip, ag, cnt)
107#define TRACE_MOVEAG(mp, ip, pip, oag, ocnt, nag, ncnt)
108#define TRACE_ORPHAN(mp, ip, ag)
109#endif
110
111static kmem_zone_t *item_zone;
112
113/*
114 * Structure for associating a file or a directory with an allocation group.
115 * The parent directory pointer is only needed for files, but since there will
116 * generally be vastly more files than directories in the cache, using the same
117 * data structure simplifies the code with very little memory overhead.
118 */
119typedef struct fstrm_item
120{
121 xfs_agnumber_t ag; /* AG currently in use for the file/directory. */
122 xfs_inode_t *ip; /* inode self-pointer. */
123 xfs_inode_t *pip; /* Parent directory inode pointer. */
124} fstrm_item_t;
125
126/*
127 * Allocation group filestream associations are tracked with per-ag atomic
128 * counters. These counters allow _xfs_filestream_pick_ag() to tell whether a
129 * particular AG already has active filestreams associated with it. The mount
130 * point's m_peraglock is used to protect these counters from per-ag array
131 * re-allocation during a growfs operation. When xfs_growfs_data_private() is
132 * about to reallocate the array, it calls xfs_filestream_flush() with the
133 * m_peraglock held in write mode.
134 *
135 * Since xfs_mru_cache_flush() guarantees that all the free functions for all
136 * the cache elements have finished executing before it returns, it's safe for
137 * the free functions to use the atomic counters without m_peraglock protection.
138 * This allows the implementation of xfs_fstrm_free_func() to be agnostic about
139 * whether it was called with the m_peraglock held in read mode, write mode or
140 * not held at all. The race condition this addresses is the following:
141 *
142 * - The work queue scheduler fires and pulls a filestream directory cache
143 * element off the LRU end of the cache for deletion, then gets pre-empted.
144 * - A growfs operation grabs the m_peraglock in write mode, flushes all the
145 * remaining items from the cache and reallocates the mount point's per-ag
146 * array, resetting all the counters to zero.
147 * - The work queue thread resumes and calls the free function for the element
148 * it started cleaning up earlier. In the process it decrements the
149 * filestreams counter for an AG that now has no references.
150 *
151 * With a shrinkfs feature, the above scenario could panic the system.
152 *
153 * All other uses of the following macros should be protected by either the
154 * m_peraglock held in read mode, or the cache's internal locking exposed by the
155 * interval between a call to xfs_mru_cache_lookup() and a call to
156 * xfs_mru_cache_done(). In addition, the m_peraglock must be held in read mode
157 * when new elements are added to the cache.
158 *
159 * Combined, these locking rules ensure that no associations will ever exist in
160 * the cache that reference per-ag array elements that have since been
161 * reallocated.
162 */
163static int
164xfs_filestream_peek_ag(
165 xfs_mount_t *mp,
166 xfs_agnumber_t agno)
167{
168 struct xfs_perag *pag;
169 int ret;
170
171 pag = xfs_perag_get(mp, agno);
172 ret = atomic_read(&pag->pagf_fstrms);
173 xfs_perag_put(pag);
174 return ret;
175}
176
177static int
178xfs_filestream_get_ag(
179 xfs_mount_t *mp,
180 xfs_agnumber_t agno)
181{
182 struct xfs_perag *pag;
183 int ret;
184
185 pag = xfs_perag_get(mp, agno);
186 ret = atomic_inc_return(&pag->pagf_fstrms);
187 xfs_perag_put(pag);
188 return ret;
189}
190
191static void
192xfs_filestream_put_ag(
193 xfs_mount_t *mp,
194 xfs_agnumber_t agno)
195{
196 struct xfs_perag *pag;
197
198 pag = xfs_perag_get(mp, agno);
199 atomic_dec(&pag->pagf_fstrms);
200 xfs_perag_put(pag);
201}
202
203/*
204 * Scan the AGs starting at startag looking for an AG that isn't in use and has
205 * at least minlen blocks free.
206 */
207static int
208_xfs_filestream_pick_ag(
209 xfs_mount_t *mp,
210 xfs_agnumber_t startag,
211 xfs_agnumber_t *agp,
212 int flags,
213 xfs_extlen_t minlen)
214{
215 int streams, max_streams;
216 int err, trylock, nscan;
217 xfs_extlen_t longest, free, minfree, maxfree = 0;
218 xfs_agnumber_t ag, max_ag = NULLAGNUMBER;
219 struct xfs_perag *pag;
220
221 /* 2% of an AG's blocks must be free for it to be chosen. */
222 minfree = mp->m_sb.sb_agblocks / 50;
223
224 ag = startag;
225 *agp = NULLAGNUMBER;
226
227 /* For the first pass, don't sleep trying to init the per-AG. */
228 trylock = XFS_ALLOC_FLAG_TRYLOCK;
229
230 for (nscan = 0; 1; nscan++) {
231 pag = xfs_perag_get(mp, ag);
232 TRACE_AG_SCAN(mp, ag, atomic_read(&pag->pagf_fstrms));
233
234 if (!pag->pagf_init) {
235 err = xfs_alloc_pagf_init(mp, NULL, ag, trylock);
236 if (err && !trylock) {
237 xfs_perag_put(pag);
238 return err;
239 }
240 }
241
242 /* Might fail sometimes during the 1st pass with trylock set. */
243 if (!pag->pagf_init)
244 goto next_ag;
245
246 /* Keep track of the AG with the most free blocks. */
247 if (pag->pagf_freeblks > maxfree) {
248 maxfree = pag->pagf_freeblks;
249 max_streams = atomic_read(&pag->pagf_fstrms);
250 max_ag = ag;
251 }
252
253 /*
254 * The AG reference count does two things: it enforces mutual
255 * exclusion when examining the suitability of an AG in this
256 * loop, and it guards against two filestreams being established
257 * in the same AG as each other.
258 */
259 if (xfs_filestream_get_ag(mp, ag) > 1) {
260 xfs_filestream_put_ag(mp, ag);
261 goto next_ag;
262 }
263
264 longest = xfs_alloc_longest_free_extent(mp, pag);
265 if (((minlen && longest >= minlen) ||
266 (!minlen && pag->pagf_freeblks >= minfree)) &&
267 (!pag->pagf_metadata || !(flags & XFS_PICK_USERDATA) ||
268 (flags & XFS_PICK_LOWSPACE))) {
269
270 /* Break out, retaining the reference on the AG. */
271 free = pag->pagf_freeblks;
272 streams = atomic_read(&pag->pagf_fstrms);
273 xfs_perag_put(pag);
274 *agp = ag;
275 break;
276 }
277
278 /* Drop the reference on this AG, it's not usable. */
279 xfs_filestream_put_ag(mp, ag);
280next_ag:
281 xfs_perag_put(pag);
282 /* Move to the next AG, wrapping to AG 0 if necessary. */
283 if (++ag >= mp->m_sb.sb_agcount)
284 ag = 0;
285
286 /* If a full pass of the AGs hasn't been done yet, continue. */
287 if (ag != startag)
288 continue;
289
290 /* Allow sleeping in xfs_alloc_pagf_init() on the 2nd pass. */
291 if (trylock != 0) {
292 trylock = 0;
293 continue;
294 }
295
296 /* Finally, if lowspace wasn't set, set it for the 3rd pass. */
297 if (!(flags & XFS_PICK_LOWSPACE)) {
298 flags |= XFS_PICK_LOWSPACE;
299 continue;
300 }
301
302 /*
303 * Take the AG with the most free space, regardless of whether
304 * it's already in use by another filestream.
305 */
306 if (max_ag != NULLAGNUMBER) {
307 xfs_filestream_get_ag(mp, max_ag);
308 TRACE_AG_PICK1(mp, max_ag, maxfree);
309 streams = max_streams;
310 free = maxfree;
311 *agp = max_ag;
312 break;
313 }
314
315 /* take AG 0 if none matched */
316 TRACE_AG_PICK1(mp, max_ag, maxfree);
317 *agp = 0;
318 return 0;
319 }
320
321 TRACE_AG_PICK2(mp, startag, *agp, streams, free, nscan, flags);
322
323 return 0;
324}
325
326/*
327 * Set the allocation group number for a file or a directory, updating inode
328 * references and per-AG references as appropriate.
329 */
330static int
331_xfs_filestream_update_ag(
332 xfs_inode_t *ip,
333 xfs_inode_t *pip,
334 xfs_agnumber_t ag)
335{
336 int err = 0;
337 xfs_mount_t *mp;
338 xfs_mru_cache_t *cache;
339 fstrm_item_t *item;
340 xfs_agnumber_t old_ag;
341 xfs_inode_t *old_pip;
342
343 /*
344 * Either ip is a regular file and pip is a directory, or ip is a
345 * directory and pip is NULL.
346 */
347 ASSERT(ip && ((S_ISREG(ip->i_d.di_mode) && pip &&
348 S_ISDIR(pip->i_d.di_mode)) ||
349 (S_ISDIR(ip->i_d.di_mode) && !pip)));
350
351 mp = ip->i_mount;
352 cache = mp->m_filestream;
353
354 item = xfs_mru_cache_lookup(cache, ip->i_ino);
355 if (item) {
356 ASSERT(item->ip == ip);
357 old_ag = item->ag;
358 item->ag = ag;
359 old_pip = item->pip;
360 item->pip = pip;
361 xfs_mru_cache_done(cache);
362
363 /*
364 * If the AG has changed, drop the old ref and take a new one,
365 * effectively transferring the reference from old to new AG.
366 */
367 if (ag != old_ag) {
368 xfs_filestream_put_ag(mp, old_ag);
369 xfs_filestream_get_ag(mp, ag);
370 }
371
372 /*
373 * If ip is a file and its pip has changed, drop the old ref and
374 * take a new one.
375 */
376 if (pip && pip != old_pip) {
377 IRELE(old_pip);
378 IHOLD(pip);
379 }
380
381 TRACE_UPDATE(mp, ip, old_ag, xfs_filestream_peek_ag(mp, old_ag),
382 ag, xfs_filestream_peek_ag(mp, ag));
383 return 0;
384 }
385
386 item = kmem_zone_zalloc(item_zone, KM_MAYFAIL);
387 if (!item)
388 return ENOMEM;
389
390 item->ag = ag;
391 item->ip = ip;
392 item->pip = pip;
393
394 err = xfs_mru_cache_insert(cache, ip->i_ino, item);
395 if (err) {
396 kmem_zone_free(item_zone, item);
397 return err;
398 }
399
400 /* Take a reference on the AG. */
401 xfs_filestream_get_ag(mp, ag);
402
403 /*
404 * Take a reference on the inode itself regardless of whether it's a
405 * regular file or a directory.
406 */
407 IHOLD(ip);
408
409 /*
410 * In the case of a regular file, take a reference on the parent inode
411 * as well to ensure it remains in-core.
412 */
413 if (pip)
414 IHOLD(pip);
415
416 TRACE_UPDATE(mp, ip, ag, xfs_filestream_peek_ag(mp, ag),
417 ag, xfs_filestream_peek_ag(mp, ag));
418
419 return 0;
420}
421
422/* xfs_fstrm_free_func(): callback for freeing cached stream items. */
423STATIC void
424xfs_fstrm_free_func(
425 unsigned long ino,
426 void *data)
427{
428 fstrm_item_t *item = (fstrm_item_t *)data;
429 xfs_inode_t *ip = item->ip;
430
431 ASSERT(ip->i_ino == ino);
432
433 xfs_iflags_clear(ip, XFS_IFILESTREAM);
434
435 /* Drop the reference taken on the AG when the item was added. */
436 xfs_filestream_put_ag(ip->i_mount, item->ag);
437
438 TRACE_FREE(ip->i_mount, ip, item->pip, item->ag,
439 xfs_filestream_peek_ag(ip->i_mount, item->ag));
440
441 /*
442 * _xfs_filestream_update_ag() always takes a reference on the inode
443 * itself, whether it's a file or a directory. Release it here.
444 * This can result in the inode being freed and so we must
445 * not hold any inode locks when freeing filesstreams objects
446 * otherwise we can deadlock here.
447 */
448 IRELE(ip);
449
450 /*
451 * In the case of a regular file, _xfs_filestream_update_ag() also
452 * takes a ref on the parent inode to keep it in-core. Release that
453 * too.
454 */
455 if (item->pip)
456 IRELE(item->pip);
457
458 /* Finally, free the memory allocated for the item. */
459 kmem_zone_free(item_zone, item);
460}
461
462/*
463 * xfs_filestream_init() is called at xfs initialisation time to set up the
464 * memory zone that will be used for filestream data structure allocation.
465 */
466int
467xfs_filestream_init(void)
468{
469 item_zone = kmem_zone_init(sizeof(fstrm_item_t), "fstrm_item");
470 if (!item_zone)
471 return -ENOMEM;
472
473 return 0;
474}
475
476/*
477 * xfs_filestream_uninit() is called at xfs termination time to destroy the
478 * memory zone that was used for filestream data structure allocation.
479 */
480void
481xfs_filestream_uninit(void)
482{
483 kmem_zone_destroy(item_zone);
484}
485
486/*
487 * xfs_filestream_mount() is called when a file system is mounted with the
488 * filestream option. It is responsible for allocating the data structures
489 * needed to track the new file system's file streams.
490 */
491int
492xfs_filestream_mount(
493 xfs_mount_t *mp)
494{
495 int err;
496 unsigned int lifetime, grp_count;
497
498 /*
499 * The filestream timer tunable is currently fixed within the range of
500 * one second to four minutes, with five seconds being the default. The
501 * group count is somewhat arbitrary, but it'd be nice to adhere to the
502 * timer tunable to within about 10 percent. This requires at least 10
503 * groups.
504 */
505 lifetime = xfs_fstrm_centisecs * 10;
506 grp_count = 10;
507
508 err = xfs_mru_cache_create(&mp->m_filestream, lifetime, grp_count,
509 xfs_fstrm_free_func);
510
511 return err;
512}
513
514/*
515 * xfs_filestream_unmount() is called when a file system that was mounted with
516 * the filestream option is unmounted. It drains the data structures created
517 * to track the file system's file streams and frees all the memory that was
518 * allocated.
519 */
520void
521xfs_filestream_unmount(
522 xfs_mount_t *mp)
523{
524 xfs_mru_cache_destroy(mp->m_filestream);
525}
526
527/*
528 * Return the AG of the filestream the file or directory belongs to, or
529 * NULLAGNUMBER otherwise.
530 */
531xfs_agnumber_t
532xfs_filestream_lookup_ag(
533 xfs_inode_t *ip)
534{
535 xfs_mru_cache_t *cache;
536 fstrm_item_t *item;
537 xfs_agnumber_t ag;
538 int ref;
539
540 if (!S_ISREG(ip->i_d.di_mode) && !S_ISDIR(ip->i_d.di_mode)) {
541 ASSERT(0);
542 return NULLAGNUMBER;
543 }
544
545 cache = ip->i_mount->m_filestream;
546 item = xfs_mru_cache_lookup(cache, ip->i_ino);
547 if (!item) {
548 TRACE_LOOKUP(ip->i_mount, ip, NULL, NULLAGNUMBER, 0);
549 return NULLAGNUMBER;
550 }
551
552 ASSERT(ip == item->ip);
553 ag = item->ag;
554 ref = xfs_filestream_peek_ag(ip->i_mount, ag);
555 xfs_mru_cache_done(cache);
556
557 TRACE_LOOKUP(ip->i_mount, ip, item->pip, ag, ref);
558 return ag;
559}
560
561/*
562 * xfs_filestream_associate() should only be called to associate a regular file
563 * with its parent directory. Calling it with a child directory isn't
564 * appropriate because filestreams don't apply to entire directory hierarchies.
565 * Creating a file in a child directory of an existing filestream directory
566 * starts a new filestream with its own allocation group association.
567 *
568 * Returns < 0 on error, 0 if successful association occurred, > 0 if
569 * we failed to get an association because of locking issues.
570 */
571int
572xfs_filestream_associate(
573 xfs_inode_t *pip,
574 xfs_inode_t *ip)
575{
576 xfs_mount_t *mp;
577 xfs_mru_cache_t *cache;
578 fstrm_item_t *item;
579 xfs_agnumber_t ag, rotorstep, startag;
580 int err = 0;
581
582 ASSERT(S_ISDIR(pip->i_d.di_mode));
583 ASSERT(S_ISREG(ip->i_d.di_mode));
584 if (!S_ISDIR(pip->i_d.di_mode) || !S_ISREG(ip->i_d.di_mode))
585 return -EINVAL;
586
587 mp = pip->i_mount;
588 cache = mp->m_filestream;
589
590 /*
591 * We have a problem, Houston.
592 *
593 * Taking the iolock here violates inode locking order - we already
594 * hold the ilock. Hence if we block getting this lock we may never
595 * wake. Unfortunately, that means if we can't get the lock, we're
596 * screwed in terms of getting a stream association - we can't spin
597 * waiting for the lock because someone else is waiting on the lock we
598 * hold and we cannot drop that as we are in a transaction here.
599 *
600 * Lucky for us, this inversion is not a problem because it's a
601 * directory inode that we are trying to lock here.
602 *
603 * So, if we can't get the iolock without sleeping then just give up
604 */
605 if (!xfs_ilock_nowait(pip, XFS_IOLOCK_EXCL))
606 return 1;
607
608 /* If the parent directory is already in the cache, use its AG. */
609 item = xfs_mru_cache_lookup(cache, pip->i_ino);
610 if (item) {
611 ASSERT(item->ip == pip);
612 ag = item->ag;
613 xfs_mru_cache_done(cache);
614
615 TRACE_LOOKUP(mp, pip, pip, ag, xfs_filestream_peek_ag(mp, ag));
616 err = _xfs_filestream_update_ag(ip, pip, ag);
617
618 goto exit;
619 }
620
621 /*
622 * Set the starting AG using the rotor for inode32, otherwise
623 * use the directory inode's AG.
624 */
625 if (mp->m_flags & XFS_MOUNT_32BITINODES) {
626 rotorstep = xfs_rotorstep;
627 startag = (mp->m_agfrotor / rotorstep) % mp->m_sb.sb_agcount;
628 mp->m_agfrotor = (mp->m_agfrotor + 1) %
629 (mp->m_sb.sb_agcount * rotorstep);
630 } else
631 startag = XFS_INO_TO_AGNO(mp, pip->i_ino);
632
633 /* Pick a new AG for the parent inode starting at startag. */
634 err = _xfs_filestream_pick_ag(mp, startag, &ag, 0, 0);
635 if (err || ag == NULLAGNUMBER)
636 goto exit_did_pick;
637
638 /* Associate the parent inode with the AG. */
639 err = _xfs_filestream_update_ag(pip, NULL, ag);
640 if (err)
641 goto exit_did_pick;
642
643 /* Associate the file inode with the AG. */
644 err = _xfs_filestream_update_ag(ip, pip, ag);
645 if (err)
646 goto exit_did_pick;
647
648 TRACE_ASSOCIATE(mp, ip, pip, ag, xfs_filestream_peek_ag(mp, ag));
649
650exit_did_pick:
651 /*
652 * If _xfs_filestream_pick_ag() returned a valid AG, remove the
653 * reference it took on it, since the file and directory will have taken
654 * their own now if they were successfully cached.
655 */
656 if (ag != NULLAGNUMBER)
657 xfs_filestream_put_ag(mp, ag);
658
659exit:
660 xfs_iunlock(pip, XFS_IOLOCK_EXCL);
661 return -err;
662}
663
664/*
665 * Pick a new allocation group for the current file and its file stream. This
666 * function is called by xfs_bmap_filestreams() with the mount point's per-ag
667 * lock held.
668 */
669int
670xfs_filestream_new_ag(
671 xfs_bmalloca_t *ap,
672 xfs_agnumber_t *agp)
673{
674 int flags, err;
675 xfs_inode_t *ip, *pip = NULL;
676 xfs_mount_t *mp;
677 xfs_mru_cache_t *cache;
678 xfs_extlen_t minlen;
679 fstrm_item_t *dir, *file;
680 xfs_agnumber_t ag = NULLAGNUMBER;
681
682 ip = ap->ip;
683 mp = ip->i_mount;
684 cache = mp->m_filestream;
685 minlen = ap->alen;
686 *agp = NULLAGNUMBER;
687
688 /*
689 * Look for the file in the cache, removing it if it's found. Doing
690 * this allows it to be held across the dir lookup that follows.
691 */
692 file = xfs_mru_cache_remove(cache, ip->i_ino);
693 if (file) {
694 ASSERT(ip == file->ip);
695
696 /* Save the file's parent inode and old AG number for later. */
697 pip = file->pip;
698 ag = file->ag;
699
700 /* Look for the file's directory in the cache. */
701 dir = xfs_mru_cache_lookup(cache, pip->i_ino);
702 if (dir) {
703 ASSERT(pip == dir->ip);
704
705 /*
706 * If the directory has already moved on to a new AG,
707 * use that AG as the new AG for the file. Don't
708 * forget to twiddle the AG refcounts to match the
709 * movement.
710 */
711 if (dir->ag != file->ag) {
712 xfs_filestream_put_ag(mp, file->ag);
713 xfs_filestream_get_ag(mp, dir->ag);
714 *agp = file->ag = dir->ag;
715 }
716
717 xfs_mru_cache_done(cache);
718 }
719
720 /*
721 * Put the file back in the cache. If this fails, the free
722 * function needs to be called to tidy up in the same way as if
723 * the item had simply expired from the cache.
724 */
725 err = xfs_mru_cache_insert(cache, ip->i_ino, file);
726 if (err) {
727 xfs_fstrm_free_func(ip->i_ino, file);
728 return err;
729 }
730
731 /*
732 * If the file's AG was moved to the directory's new AG, there's
733 * nothing more to be done.
734 */
735 if (*agp != NULLAGNUMBER) {
736 TRACE_MOVEAG(mp, ip, pip,
737 ag, xfs_filestream_peek_ag(mp, ag),
738 *agp, xfs_filestream_peek_ag(mp, *agp));
739 return 0;
740 }
741 }
742
743 /*
744 * If the file's parent directory is known, take its iolock in exclusive
745 * mode to prevent two sibling files from racing each other to migrate
746 * themselves and their parent to different AGs.
747 *
748 * Note that we lock the parent directory iolock inside the child
749 * iolock here. That's fine as we never hold both parent and child
750 * iolock in any other place. This is different from the ilock,
751 * which requires locking of the child after the parent for namespace
752 * operations.
753 */
754 if (pip)
755 xfs_ilock(pip, XFS_IOLOCK_EXCL | XFS_IOLOCK_PARENT);
756
757 /*
758 * A new AG needs to be found for the file. If the file's parent
759 * directory is also known, it will be moved to the new AG as well to
760 * ensure that files created inside it in future use the new AG.
761 */
762 ag = (ag == NULLAGNUMBER) ? 0 : (ag + 1) % mp->m_sb.sb_agcount;
763 flags = (ap->userdata ? XFS_PICK_USERDATA : 0) |
764 (ap->low ? XFS_PICK_LOWSPACE : 0);
765
766 err = _xfs_filestream_pick_ag(mp, ag, agp, flags, minlen);
767 if (err || *agp == NULLAGNUMBER)
768 goto exit;
769
770 /*
771 * If the file wasn't found in the file cache, then its parent directory
772 * inode isn't known. For this to have happened, the file must either
773 * be pre-existing, or it was created long enough ago that its cache
774 * entry has expired. This isn't the sort of usage that the filestreams
775 * allocator is trying to optimise, so there's no point trying to track
776 * its new AG somehow in the filestream data structures.
777 */
778 if (!pip) {
779 TRACE_ORPHAN(mp, ip, *agp);
780 goto exit;
781 }
782
783 /* Associate the parent inode with the AG. */
784 err = _xfs_filestream_update_ag(pip, NULL, *agp);
785 if (err)
786 goto exit;
787
788 /* Associate the file inode with the AG. */
789 err = _xfs_filestream_update_ag(ip, pip, *agp);
790 if (err)
791 goto exit;
792
793 TRACE_MOVEAG(mp, ip, pip, NULLAGNUMBER, 0,
794 *agp, xfs_filestream_peek_ag(mp, *agp));
795
796exit:
797 /*
798 * If _xfs_filestream_pick_ag() returned a valid AG, remove the
799 * reference it took on it, since the file and directory will have taken
800 * their own now if they were successfully cached.
801 */
802 if (*agp != NULLAGNUMBER)
803 xfs_filestream_put_ag(mp, *agp);
804 else
805 *agp = 0;
806
807 if (pip)
808 xfs_iunlock(pip, XFS_IOLOCK_EXCL);
809
810 return err;
811}
812
813/*
814 * Remove an association between an inode and a filestream object.
815 * Typically this is done on last close of an unlinked file.
816 */
817void
818xfs_filestream_deassociate(
819 xfs_inode_t *ip)
820{
821 xfs_mru_cache_t *cache = ip->i_mount->m_filestream;
822
823 xfs_mru_cache_delete(cache, ip->i_ino);
824}
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}