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 struct xfs_fstrm_item {
37 struct xfs_mru_cache_elem mru;
39 xfs_agnumber_t ag; /* AG in use for this directory */
42 enum xfs_fstrm_alloc {
43 XFS_PICK_USERDATA = 1,
44 XFS_PICK_LOWSPACE = 2,
48 * Allocation group filestream associations are tracked with per-ag atomic
49 * counters. These counters allow xfs_filestream_pick_ag() to tell whether a
50 * particular AG already has active filestreams associated with it. The mount
51 * point's m_peraglock is used to protect these counters from per-ag array
52 * re-allocation during a growfs operation. When xfs_growfs_data_private() is
53 * about to reallocate the array, it calls xfs_filestream_flush() with the
54 * m_peraglock held in write mode.
56 * Since xfs_mru_cache_flush() guarantees that all the free functions for all
57 * the cache elements have finished executing before it returns, it's safe for
58 * the free functions to use the atomic counters without m_peraglock protection.
59 * This allows the implementation of xfs_fstrm_free_func() to be agnostic about
60 * whether it was called with the m_peraglock held in read mode, write mode or
61 * not held at all. The race condition this addresses is the following:
63 * - The work queue scheduler fires and pulls a filestream directory cache
64 * element off the LRU end of the cache for deletion, then gets pre-empted.
65 * - A growfs operation grabs the m_peraglock in write mode, flushes all the
66 * remaining items from the cache and reallocates the mount point's per-ag
67 * array, resetting all the counters to zero.
68 * - The work queue thread resumes and calls the free function for the element
69 * it started cleaning up earlier. In the process it decrements the
70 * filestreams counter for an AG that now has no references.
72 * With a shrinkfs feature, the above scenario could panic the system.
74 * All other uses of the following macros should be protected by either the
75 * m_peraglock held in read mode, or the cache's internal locking exposed by the
76 * interval between a call to xfs_mru_cache_lookup() and a call to
77 * xfs_mru_cache_done(). In addition, the m_peraglock must be held in read mode
78 * when new elements are added to the cache.
80 * Combined, these locking rules ensure that no associations will ever exist in
81 * the cache that reference per-ag array elements that have since been
85 xfs_filestream_peek_ag(
89 struct xfs_perag *pag;
92 pag = xfs_perag_get(mp, agno);
93 ret = atomic_read(&pag->pagf_fstrms);
99 xfs_filestream_get_ag(
103 struct xfs_perag *pag;
106 pag = xfs_perag_get(mp, agno);
107 ret = atomic_inc_return(&pag->pagf_fstrms);
113 xfs_filestream_put_ag(
117 struct xfs_perag *pag;
119 pag = xfs_perag_get(mp, agno);
120 atomic_dec(&pag->pagf_fstrms);
126 struct xfs_mru_cache_elem *mru)
128 struct xfs_fstrm_item *item =
129 container_of(mru, struct xfs_fstrm_item, mru);
131 xfs_filestream_put_ag(item->ip->i_mount, item->ag);
133 trace_xfs_filestream_free(item->ip, item->ag);
139 * Scan the AGs starting at startag looking for an AG that isn't in use and has
140 * at least minlen blocks free.
143 xfs_filestream_pick_ag(
144 struct xfs_inode *ip,
145 xfs_agnumber_t startag,
150 struct xfs_mount *mp = ip->i_mount;
151 struct xfs_fstrm_item *item;
152 struct xfs_perag *pag;
153 xfs_extlen_t longest, free = 0, minfree, maxfree = 0;
154 xfs_agnumber_t ag, max_ag = NULLAGNUMBER;
155 int err, trylock, nscan;
157 ASSERT(S_ISDIR(ip->i_d.di_mode));
159 /* 2% of an AG's blocks must be free for it to be chosen. */
160 minfree = mp->m_sb.sb_agblocks / 50;
165 /* For the first pass, don't sleep trying to init the per-AG. */
166 trylock = XFS_ALLOC_FLAG_TRYLOCK;
168 for (nscan = 0; 1; nscan++) {
169 trace_xfs_filestream_scan(ip, ag);
171 pag = xfs_perag_get(mp, ag);
173 if (!pag->pagf_init) {
174 err = xfs_alloc_pagf_init(mp, NULL, ag, trylock);
175 if (err && !trylock) {
181 /* Might fail sometimes during the 1st pass with trylock set. */
185 /* Keep track of the AG with the most free blocks. */
186 if (pag->pagf_freeblks > maxfree) {
187 maxfree = pag->pagf_freeblks;
192 * The AG reference count does two things: it enforces mutual
193 * exclusion when examining the suitability of an AG in this
194 * loop, and it guards against two filestreams being established
195 * in the same AG as each other.
197 if (xfs_filestream_get_ag(mp, ag) > 1) {
198 xfs_filestream_put_ag(mp, ag);
202 longest = xfs_alloc_longest_free_extent(mp, pag);
203 if (((minlen && longest >= minlen) ||
204 (!minlen && pag->pagf_freeblks >= minfree)) &&
205 (!pag->pagf_metadata || !(flags & XFS_PICK_USERDATA) ||
206 (flags & XFS_PICK_LOWSPACE))) {
208 /* Break out, retaining the reference on the AG. */
209 free = pag->pagf_freeblks;
215 /* Drop the reference on this AG, it's not usable. */
216 xfs_filestream_put_ag(mp, ag);
219 /* Move to the next AG, wrapping to AG 0 if necessary. */
220 if (++ag >= mp->m_sb.sb_agcount)
223 /* If a full pass of the AGs hasn't been done yet, continue. */
227 /* Allow sleeping in xfs_alloc_pagf_init() on the 2nd pass. */
233 /* Finally, if lowspace wasn't set, set it for the 3rd pass. */
234 if (!(flags & XFS_PICK_LOWSPACE)) {
235 flags |= XFS_PICK_LOWSPACE;
240 * Take the AG with the most free space, regardless of whether
241 * it's already in use by another filestream.
243 if (max_ag != NULLAGNUMBER) {
244 xfs_filestream_get_ag(mp, max_ag);
250 /* take AG 0 if none matched */
251 trace_xfs_filestream_pick(ip, *agp, free, nscan);
256 trace_xfs_filestream_pick(ip, *agp, free, nscan);
258 if (*agp == NULLAGNUMBER)
262 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(parent->d_inode);
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(ip->i_d.di_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(ip, 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;
374 struct xfs_mru_cache_elem *mru;
378 pip = xfs_filestream_get_parent(ip);
382 mru = xfs_mru_cache_remove(mp->m_filestream, pip->i_ino);
384 struct xfs_fstrm_item *item =
385 container_of(mru, struct xfs_fstrm_item, mru);
386 startag = (item->ag + 1) % mp->m_sb.sb_agcount;
389 flags = (ap->userdata ? XFS_PICK_USERDATA : 0) |
390 (ap->flist->xbf_low ? XFS_PICK_LOWSPACE : 0);
392 err = xfs_filestream_pick_ag(pip, startag, agp, flags, minlen);
395 * Only free the item here so we skip over the old AG earlier.
398 xfs_fstrm_free_func(mru);
402 if (*agp == NULLAGNUMBER)
408 xfs_filestream_deassociate(
409 struct xfs_inode *ip)
411 xfs_mru_cache_delete(ip->i_mount->m_filestream, ip->i_ino);
415 xfs_filestream_mount(
419 * The filestream timer tunable is currently fixed within the range of
420 * one second to four minutes, with five seconds being the default. The
421 * group count is somewhat arbitrary, but it'd be nice to adhere to the
422 * timer tunable to within about 10 percent. This requires at least 10
425 return xfs_mru_cache_create(&mp->m_filestream, xfs_fstrm_centisecs * 10,
426 10, xfs_fstrm_free_func);
430 xfs_filestream_unmount(
433 xfs_mru_cache_destroy(mp->m_filestream);