2 * Copyright (c) 2006-2007 Silicon Graphics, Inc.
3 * Copyright (c) 2014 Christoph Hellwig.
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.
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.
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
20 #include "xfs_format.h"
21 #include "xfs_log_format.h"
22 #include "xfs_trans_resv.h"
25 #include "xfs_mount.h"
27 #include "xfs_inode.h"
29 #include "xfs_bmap_util.h"
30 #include "xfs_alloc.h"
31 #include "xfs_mru_cache.h"
32 #include "xfs_dinode.h"
33 #include "xfs_filestream.h"
34 #include "xfs_trace.h"
36 #define TRACE_AG_SCAN(mp, ag, ag2)
37 #define TRACE_AG_PICK1(mp, max_ag, maxfree)
38 #define TRACE_AG_PICK2(mp, ag, ag2, cnt, free, scan, flag)
39 #define TRACE_FREE(mp, ip, pip, ag, cnt)
40 #define TRACE_LOOKUP(mp, ip, pip, ag, cnt)
42 struct xfs_fstrm_item {
43 struct xfs_mru_cache_elem mru;
45 xfs_agnumber_t ag; /* AG in use for this directory */
48 enum xfs_fstrm_alloc {
49 XFS_PICK_USERDATA = 1,
50 XFS_PICK_LOWSPACE = 2,
54 * Allocation group filestream associations are tracked with per-ag atomic
55 * counters. These counters allow xfs_filestream_pick_ag() to tell whether a
56 * particular AG already has active filestreams associated with it. The mount
57 * point's m_peraglock is used to protect these counters from per-ag array
58 * re-allocation during a growfs operation. When xfs_growfs_data_private() is
59 * about to reallocate the array, it calls xfs_filestream_flush() with the
60 * m_peraglock held in write mode.
62 * Since xfs_mru_cache_flush() guarantees that all the free functions for all
63 * the cache elements have finished executing before it returns, it's safe for
64 * the free functions to use the atomic counters without m_peraglock protection.
65 * This allows the implementation of xfs_fstrm_free_func() to be agnostic about
66 * whether it was called with the m_peraglock held in read mode, write mode or
67 * not held at all. The race condition this addresses is the following:
69 * - The work queue scheduler fires and pulls a filestream directory cache
70 * element off the LRU end of the cache for deletion, then gets pre-empted.
71 * - A growfs operation grabs the m_peraglock in write mode, flushes all the
72 * remaining items from the cache and reallocates the mount point's per-ag
73 * array, resetting all the counters to zero.
74 * - The work queue thread resumes and calls the free function for the element
75 * it started cleaning up earlier. In the process it decrements the
76 * filestreams counter for an AG that now has no references.
78 * With a shrinkfs feature, the above scenario could panic the system.
80 * All other uses of the following macros should be protected by either the
81 * m_peraglock held in read mode, or the cache's internal locking exposed by the
82 * interval between a call to xfs_mru_cache_lookup() and a call to
83 * xfs_mru_cache_done(). In addition, the m_peraglock must be held in read mode
84 * when new elements are added to the cache.
86 * Combined, these locking rules ensure that no associations will ever exist in
87 * the cache that reference per-ag array elements that have since been
91 xfs_filestream_peek_ag(
95 struct xfs_perag *pag;
98 pag = xfs_perag_get(mp, agno);
99 ret = atomic_read(&pag->pagf_fstrms);
105 xfs_filestream_get_ag(
109 struct xfs_perag *pag;
112 pag = xfs_perag_get(mp, agno);
113 ret = atomic_inc_return(&pag->pagf_fstrms);
119 xfs_filestream_put_ag(
123 struct xfs_perag *pag;
125 pag = xfs_perag_get(mp, agno);
126 atomic_dec(&pag->pagf_fstrms);
132 struct xfs_mru_cache_elem *mru)
134 struct xfs_fstrm_item *item =
135 container_of(mru, struct xfs_fstrm_item, mru);
137 xfs_filestream_put_ag(item->ip->i_mount, item->ag);
139 TRACE_FREE(mp, ip, NULL, item->ag,
140 xfs_filestream_peek_ag(mp, item->ag));
146 * Scan the AGs starting at startag looking for an AG that isn't in use and has
147 * at least minlen blocks free.
150 xfs_filestream_pick_ag(
151 struct xfs_inode *ip,
152 xfs_agnumber_t startag,
157 struct xfs_mount *mp = ip->i_mount;
158 struct xfs_fstrm_item *item;
159 struct xfs_perag *pag;
160 xfs_extlen_t longest, free, minfree, maxfree = 0;
161 xfs_agnumber_t ag, max_ag = NULLAGNUMBER;
162 int streams, max_streams;
163 int err, trylock, nscan;
165 ASSERT(S_ISDIR(ip->i_d.di_mode));
167 /* 2% of an AG's blocks must be free for it to be chosen. */
168 minfree = mp->m_sb.sb_agblocks / 50;
173 /* For the first pass, don't sleep trying to init the per-AG. */
174 trylock = XFS_ALLOC_FLAG_TRYLOCK;
176 for (nscan = 0; 1; nscan++) {
177 pag = xfs_perag_get(mp, ag);
178 TRACE_AG_SCAN(mp, ag, atomic_read(&pag->pagf_fstrms));
180 if (!pag->pagf_init) {
181 err = xfs_alloc_pagf_init(mp, NULL, ag, trylock);
182 if (err && !trylock) {
188 /* Might fail sometimes during the 1st pass with trylock set. */
192 /* Keep track of the AG with the most free blocks. */
193 if (pag->pagf_freeblks > maxfree) {
194 maxfree = pag->pagf_freeblks;
195 max_streams = atomic_read(&pag->pagf_fstrms);
200 * The AG reference count does two things: it enforces mutual
201 * exclusion when examining the suitability of an AG in this
202 * loop, and it guards against two filestreams being established
203 * in the same AG as each other.
205 if (xfs_filestream_get_ag(mp, ag) > 1) {
206 xfs_filestream_put_ag(mp, ag);
210 longest = xfs_alloc_longest_free_extent(mp, pag);
211 if (((minlen && longest >= minlen) ||
212 (!minlen && pag->pagf_freeblks >= minfree)) &&
213 (!pag->pagf_metadata || !(flags & XFS_PICK_USERDATA) ||
214 (flags & XFS_PICK_LOWSPACE))) {
216 /* Break out, retaining the reference on the AG. */
217 free = pag->pagf_freeblks;
218 streams = atomic_read(&pag->pagf_fstrms);
224 /* Drop the reference on this AG, it's not usable. */
225 xfs_filestream_put_ag(mp, ag);
228 /* Move to the next AG, wrapping to AG 0 if necessary. */
229 if (++ag >= mp->m_sb.sb_agcount)
232 /* If a full pass of the AGs hasn't been done yet, continue. */
236 /* Allow sleeping in xfs_alloc_pagf_init() on the 2nd pass. */
242 /* Finally, if lowspace wasn't set, set it for the 3rd pass. */
243 if (!(flags & XFS_PICK_LOWSPACE)) {
244 flags |= XFS_PICK_LOWSPACE;
249 * Take the AG with the most free space, regardless of whether
250 * it's already in use by another filestream.
252 if (max_ag != NULLAGNUMBER) {
253 xfs_filestream_get_ag(mp, max_ag);
254 TRACE_AG_PICK1(mp, max_ag, maxfree);
255 streams = max_streams;
261 /* take AG 0 if none matched */
262 TRACE_AG_PICK1(mp, max_ag, maxfree);
267 TRACE_AG_PICK2(mp, startag, *agp, streams, free, nscan, flags);
269 if (*agp == NULLAGNUMBER)
273 item = kmem_alloc(sizeof(*item), KM_MAYFAIL);
280 err = xfs_mru_cache_insert(mp->m_filestream, ip->i_ino, &item->mru);
292 xfs_filestream_put_ag(mp, *agp);
296 static struct xfs_inode *
297 xfs_filestream_get_parent(
298 struct xfs_inode *ip)
300 struct inode *inode = VFS_I(ip), *dir = NULL;
301 struct dentry *dentry, *parent;
303 dentry = d_find_alias(inode);
307 parent = dget_parent(dentry);
311 dir = igrab(parent->d_inode);
317 return dir ? XFS_I(dir) : NULL;
321 * Find the right allocation group for a file, either by finding an
322 * existing file stream or creating a new one.
324 * Returns NULLAGNUMBER in case of an error.
327 xfs_filestream_lookup_ag(
328 struct xfs_inode *ip)
330 struct xfs_mount *mp = ip->i_mount;
331 struct xfs_inode *pip = NULL;
332 xfs_agnumber_t startag, ag = NULLAGNUMBER;
334 struct xfs_mru_cache_elem *mru;
336 ASSERT(S_ISREG(ip->i_d.di_mode));
338 pip = xfs_filestream_get_parent(ip);
342 mru = xfs_mru_cache_lookup(mp->m_filestream, pip->i_ino);
344 ag = container_of(mru, struct xfs_fstrm_item, mru)->ag;
345 xfs_mru_cache_done(mp->m_filestream);
347 ref = xfs_filestream_peek_ag(ip->i_mount, ag);
348 TRACE_LOOKUP(mp, ip, pip, ag, ref);
353 * Set the starting AG using the rotor for inode32, otherwise
354 * use the directory inode's AG.
356 if (mp->m_flags & XFS_MOUNT_32BITINODES) {
357 xfs_agnumber_t rotorstep = xfs_rotorstep;
358 startag = (mp->m_agfrotor / rotorstep) % mp->m_sb.sb_agcount;
359 mp->m_agfrotor = (mp->m_agfrotor + 1) %
360 (mp->m_sb.sb_agcount * rotorstep);
362 startag = XFS_INO_TO_AGNO(mp, pip->i_ino);
364 if (xfs_filestream_pick_ag(pip, startag, &ag, 0, 0))
372 * Pick a new allocation group for the current file and its file stream.
374 * This is called when the allocator can't find a suitable extent in the
375 * current AG, and we have to move the stream into a new AG with more space.
378 xfs_filestream_new_ag(
379 struct xfs_bmalloca *ap,
382 struct xfs_inode *ip = ap->ip, *pip;
383 struct xfs_mount *mp = ip->i_mount;
384 xfs_extlen_t minlen = ap->length;
385 xfs_agnumber_t startag = 0;
387 struct xfs_mru_cache_elem *mru;
391 pip = xfs_filestream_get_parent(ip);
395 mru = xfs_mru_cache_remove(mp->m_filestream, pip->i_ino);
397 struct xfs_fstrm_item *item =
398 container_of(mru, struct xfs_fstrm_item, mru);
399 startag = (item->ag + 1) % mp->m_sb.sb_agcount;
402 flags = (ap->userdata ? XFS_PICK_USERDATA : 0) |
403 (ap->flist->xbf_low ? XFS_PICK_LOWSPACE : 0);
405 err = xfs_filestream_pick_ag(pip, startag, agp, flags, minlen);
408 * Only free the item here so we skip over the old AG earlier.
411 xfs_fstrm_free_func(mru);
415 if (*agp == NULLAGNUMBER)
421 xfs_filestream_deassociate(
422 struct xfs_inode *ip)
424 xfs_mru_cache_delete(ip->i_mount->m_filestream, ip->i_ino);
428 xfs_filestream_mount(
432 * The filestream timer tunable is currently fixed within the range of
433 * one second to four minutes, with five seconds being the default. The
434 * group count is somewhat arbitrary, but it'd be nice to adhere to the
435 * timer tunable to within about 10 percent. This requires at least 10
438 return xfs_mru_cache_create(&mp->m_filestream, xfs_fstrm_centisecs * 10,
439 10, xfs_fstrm_free_func);
443 xfs_filestream_unmount(
446 xfs_mru_cache_destroy(mp->m_filestream);