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"
24 #include "xfs_mount.h"
25 #include "xfs_defer.h"
26 #include "xfs_inode.h"
28 #include "xfs_bmap_util.h"
29 #include "xfs_alloc.h"
30 #include "xfs_mru_cache.h"
31 #include "xfs_filestream.h"
32 #include "xfs_trace.h"
33 #include "xfs_ag_resv.h"
35 struct xfs_fstrm_item {
36 struct xfs_mru_cache_elem mru;
37 xfs_agnumber_t ag; /* AG in use for this directory */
40 enum xfs_fstrm_alloc {
41 XFS_PICK_USERDATA = 1,
42 XFS_PICK_LOWSPACE = 2,
46 * Allocation group filestream associations are tracked with per-ag atomic
47 * counters. These counters allow xfs_filestream_pick_ag() to tell whether a
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.
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:
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.
70 * With a shrinkfs feature, the above scenario could panic the system.
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.
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
83 xfs_filestream_peek_ag(
87 struct xfs_perag *pag;
90 pag = xfs_perag_get(mp, agno);
91 ret = atomic_read(&pag->pagf_fstrms);
97 xfs_filestream_get_ag(
101 struct xfs_perag *pag;
104 pag = xfs_perag_get(mp, agno);
105 ret = atomic_inc_return(&pag->pagf_fstrms);
111 xfs_filestream_put_ag(
115 struct xfs_perag *pag;
117 pag = xfs_perag_get(mp, agno);
118 atomic_dec(&pag->pagf_fstrms);
125 struct xfs_mru_cache_elem *mru)
127 struct xfs_mount *mp = data;
128 struct xfs_fstrm_item *item =
129 container_of(mru, struct xfs_fstrm_item, mru);
131 xfs_filestream_put_ag(mp, item->ag);
132 trace_xfs_filestream_free(mp, mru->key, item->ag);
138 * Scan the AGs starting at startag looking for an AG that isn't in use and has
139 * at least minlen blocks free.
142 xfs_filestream_pick_ag(
143 struct xfs_inode *ip,
144 xfs_agnumber_t startag,
149 struct xfs_mount *mp = ip->i_mount;
150 struct xfs_fstrm_item *item;
151 struct xfs_perag *pag;
152 xfs_extlen_t longest, free = 0, minfree, maxfree = 0;
153 xfs_agnumber_t ag, max_ag = NULLAGNUMBER;
154 int err, trylock, nscan;
156 ASSERT(S_ISDIR(VFS_I(ip)->i_mode));
158 /* 2% of an AG's blocks must be free for it to be chosen. */
159 minfree = mp->m_sb.sb_agblocks / 50;
164 /* For the first pass, don't sleep trying to init the per-AG. */
165 trylock = XFS_ALLOC_FLAG_TRYLOCK;
167 for (nscan = 0; 1; nscan++) {
168 trace_xfs_filestream_scan(mp, ip->i_ino, ag);
170 pag = xfs_perag_get(mp, ag);
172 if (!pag->pagf_init) {
173 err = xfs_alloc_pagf_init(mp, NULL, ag, trylock);
174 if (err && !trylock) {
180 /* Might fail sometimes during the 1st pass with trylock set. */
184 /* Keep track of the AG with the most free blocks. */
185 if (pag->pagf_freeblks > maxfree) {
186 maxfree = pag->pagf_freeblks;
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.
196 if (xfs_filestream_get_ag(mp, ag) > 1) {
197 xfs_filestream_put_ag(mp, ag);
201 longest = xfs_alloc_longest_free_extent(pag,
202 xfs_alloc_min_freelist(mp, pag),
203 xfs_ag_resv_needed(pag, XFS_AG_RESV_NONE));
204 if (((minlen && longest >= minlen) ||
205 (!minlen && pag->pagf_freeblks >= minfree)) &&
206 (!pag->pagf_metadata || !(flags & XFS_PICK_USERDATA) ||
207 (flags & XFS_PICK_LOWSPACE))) {
209 /* Break out, retaining the reference on the AG. */
210 free = pag->pagf_freeblks;
216 /* Drop the reference on this AG, it's not usable. */
217 xfs_filestream_put_ag(mp, ag);
220 /* Move to the next AG, wrapping to AG 0 if necessary. */
221 if (++ag >= mp->m_sb.sb_agcount)
224 /* If a full pass of the AGs hasn't been done yet, continue. */
228 /* Allow sleeping in xfs_alloc_pagf_init() on the 2nd pass. */
234 /* Finally, if lowspace wasn't set, set it for the 3rd pass. */
235 if (!(flags & XFS_PICK_LOWSPACE)) {
236 flags |= XFS_PICK_LOWSPACE;
241 * Take the AG with the most free space, regardless of whether
242 * it's already in use by another filestream.
244 if (max_ag != NULLAGNUMBER) {
245 xfs_filestream_get_ag(mp, max_ag);
251 /* take AG 0 if none matched */
252 trace_xfs_filestream_pick(ip, *agp, free, nscan);
257 trace_xfs_filestream_pick(ip, *agp, free, nscan);
259 if (*agp == NULLAGNUMBER)
263 item = kmem_alloc(sizeof(*item), KM_MAYFAIL);
269 err = xfs_mru_cache_insert(mp->m_filestream, ip->i_ino, &item->mru);
281 xfs_filestream_put_ag(mp, *agp);
285 static struct xfs_inode *
286 xfs_filestream_get_parent(
287 struct xfs_inode *ip)
289 struct inode *inode = VFS_I(ip), *dir = NULL;
290 struct dentry *dentry, *parent;
292 dentry = d_find_alias(inode);
296 parent = dget_parent(dentry);
300 dir = igrab(d_inode(parent));
306 return dir ? XFS_I(dir) : NULL;
310 * Find the right allocation group for a file, either by finding an
311 * existing file stream or creating a new one.
313 * Returns NULLAGNUMBER in case of an error.
316 xfs_filestream_lookup_ag(
317 struct xfs_inode *ip)
319 struct xfs_mount *mp = ip->i_mount;
320 struct xfs_inode *pip = NULL;
321 xfs_agnumber_t startag, ag = NULLAGNUMBER;
322 struct xfs_mru_cache_elem *mru;
324 ASSERT(S_ISREG(VFS_I(ip)->i_mode));
326 pip = xfs_filestream_get_parent(ip);
330 mru = xfs_mru_cache_lookup(mp->m_filestream, pip->i_ino);
332 ag = container_of(mru, struct xfs_fstrm_item, mru)->ag;
333 xfs_mru_cache_done(mp->m_filestream);
335 trace_xfs_filestream_lookup(mp, ip->i_ino, ag);
340 * Set the starting AG using the rotor for inode32, otherwise
341 * use the directory inode's AG.
343 if (mp->m_flags & XFS_MOUNT_32BITINODES) {
344 xfs_agnumber_t rotorstep = xfs_rotorstep;
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);
349 startag = XFS_INO_TO_AGNO(mp, pip->i_ino);
351 if (xfs_filestream_pick_ag(pip, startag, &ag, 0, 0))
359 * Pick a new allocation group for the current file and its file stream.
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.
365 xfs_filestream_new_ag(
366 struct xfs_bmalloca *ap,
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;
375 struct xfs_mru_cache_elem *mru;
379 pip = xfs_filestream_get_parent(ip);
383 mru = xfs_mru_cache_remove(mp->m_filestream, pip->i_ino);
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;
390 if (xfs_alloc_is_userdata(ap->datatype))
391 flags |= XFS_PICK_USERDATA;
392 if (ap->dfops->dop_low)
393 flags |= XFS_PICK_LOWSPACE;
395 err = xfs_filestream_pick_ag(pip, startag, agp, flags, minlen);
398 * Only free the item here so we skip over the old AG earlier.
401 xfs_fstrm_free_func(mp, mru);
405 if (*agp == NULLAGNUMBER)
411 xfs_filestream_deassociate(
412 struct xfs_inode *ip)
414 xfs_mru_cache_delete(ip->i_mount->m_filestream, ip->i_ino);
418 xfs_filestream_mount(
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
428 return xfs_mru_cache_create(&mp->m_filestream, mp,
429 xfs_fstrm_centisecs * 10, 10, xfs_fstrm_free_func);
433 xfs_filestream_unmount(
436 xfs_mru_cache_destroy(mp->m_filestream);