1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (c) 2006-2007 Silicon Graphics, Inc.
4 * Copyright (c) 2014 Christoph Hellwig.
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_mount.h"
13 #include "xfs_inode.h"
15 #include "xfs_alloc.h"
16 #include "xfs_mru_cache.h"
17 #include "xfs_trace.h"
19 #include "xfs_ag_resv.h"
20 #include "xfs_trans.h"
21 #include "xfs_filestream.h"
23 struct xfs_fstrm_item {
24 struct xfs_mru_cache_elem mru;
25 xfs_agnumber_t ag; /* AG in use for this directory */
28 enum xfs_fstrm_alloc {
29 XFS_PICK_USERDATA = 1,
30 XFS_PICK_LOWSPACE = 2,
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.
39 xfs_filestream_peek_ag(
43 struct xfs_perag *pag;
46 pag = xfs_perag_get(mp, agno);
47 ret = atomic_read(&pag->pagf_fstrms);
53 xfs_filestream_get_ag(
57 struct xfs_perag *pag;
60 pag = xfs_perag_get(mp, agno);
61 ret = atomic_inc_return(&pag->pagf_fstrms);
67 xfs_filestream_put_ag(
71 struct xfs_perag *pag;
73 pag = xfs_perag_get(mp, agno);
74 atomic_dec(&pag->pagf_fstrms);
81 struct xfs_mru_cache_elem *mru)
83 struct xfs_mount *mp = data;
84 struct xfs_fstrm_item *item =
85 container_of(mru, struct xfs_fstrm_item, mru);
87 xfs_filestream_put_ag(mp, item->ag);
88 trace_xfs_filestream_free(mp, mru->key, item->ag);
94 * Scan the AGs starting at startag looking for an AG that isn't in use and has
95 * at least minlen blocks free.
98 xfs_filestream_pick_ag(
100 xfs_agnumber_t startag,
105 struct xfs_mount *mp = ip->i_mount;
106 struct xfs_fstrm_item *item;
107 struct xfs_perag *pag;
108 xfs_extlen_t longest, free = 0, minfree, maxfree = 0;
109 xfs_agnumber_t ag, max_ag = NULLAGNUMBER;
110 int err, trylock, nscan;
112 ASSERT(S_ISDIR(VFS_I(ip)->i_mode));
114 /* 2% of an AG's blocks must be free for it to be chosen. */
115 minfree = mp->m_sb.sb_agblocks / 50;
120 /* For the first pass, don't sleep trying to init the per-AG. */
121 trylock = XFS_ALLOC_FLAG_TRYLOCK;
123 for (nscan = 0; 1; nscan++) {
124 trace_xfs_filestream_scan(mp, ip->i_ino, ag);
126 pag = xfs_perag_get(mp, ag);
128 if (!pag->pagf_init) {
129 err = xfs_alloc_pagf_init(mp, NULL, ag, trylock);
134 /* Couldn't lock the AGF, skip this AG. */
139 /* Keep track of the AG with the most free blocks. */
140 if (pag->pagf_freeblks > maxfree) {
141 maxfree = pag->pagf_freeblks;
146 * The AG reference count does two things: it enforces mutual
147 * exclusion when examining the suitability of an AG in this
148 * loop, and it guards against two filestreams being established
149 * in the same AG as each other.
151 if (xfs_filestream_get_ag(mp, ag) > 1) {
152 xfs_filestream_put_ag(mp, ag);
156 longest = xfs_alloc_longest_free_extent(pag,
157 xfs_alloc_min_freelist(mp, pag),
158 xfs_ag_resv_needed(pag, XFS_AG_RESV_NONE));
159 if (((minlen && longest >= minlen) ||
160 (!minlen && pag->pagf_freeblks >= minfree)) &&
161 (!pag->pagf_metadata || !(flags & XFS_PICK_USERDATA) ||
162 (flags & XFS_PICK_LOWSPACE))) {
164 /* Break out, retaining the reference on the AG. */
165 free = pag->pagf_freeblks;
171 /* Drop the reference on this AG, it's not usable. */
172 xfs_filestream_put_ag(mp, ag);
175 /* Move to the next AG, wrapping to AG 0 if necessary. */
176 if (++ag >= mp->m_sb.sb_agcount)
179 /* If a full pass of the AGs hasn't been done yet, continue. */
183 /* Allow sleeping in xfs_alloc_pagf_init() on the 2nd pass. */
189 /* Finally, if lowspace wasn't set, set it for the 3rd pass. */
190 if (!(flags & XFS_PICK_LOWSPACE)) {
191 flags |= XFS_PICK_LOWSPACE;
196 * Take the AG with the most free space, regardless of whether
197 * it's already in use by another filestream.
199 if (max_ag != NULLAGNUMBER) {
200 xfs_filestream_get_ag(mp, max_ag);
206 /* take AG 0 if none matched */
207 trace_xfs_filestream_pick(ip, *agp, free, nscan);
212 trace_xfs_filestream_pick(ip, *agp, free, nscan);
214 if (*agp == NULLAGNUMBER)
218 item = kmem_alloc(sizeof(*item), KM_MAYFAIL);
224 err = xfs_mru_cache_insert(mp->m_filestream, ip->i_ino, &item->mru);
236 xfs_filestream_put_ag(mp, *agp);
240 static struct xfs_inode *
241 xfs_filestream_get_parent(
242 struct xfs_inode *ip)
244 struct inode *inode = VFS_I(ip), *dir = NULL;
245 struct dentry *dentry, *parent;
247 dentry = d_find_alias(inode);
251 parent = dget_parent(dentry);
255 dir = igrab(d_inode(parent));
261 return dir ? XFS_I(dir) : NULL;
265 * Find the right allocation group for a file, either by finding an
266 * existing file stream or creating a new one.
268 * Returns NULLAGNUMBER in case of an error.
271 xfs_filestream_lookup_ag(
272 struct xfs_inode *ip)
274 struct xfs_mount *mp = ip->i_mount;
275 struct xfs_inode *pip = NULL;
276 xfs_agnumber_t startag, ag = NULLAGNUMBER;
277 struct xfs_mru_cache_elem *mru;
279 ASSERT(S_ISREG(VFS_I(ip)->i_mode));
281 pip = xfs_filestream_get_parent(ip);
285 mru = xfs_mru_cache_lookup(mp->m_filestream, pip->i_ino);
287 ag = container_of(mru, struct xfs_fstrm_item, mru)->ag;
288 xfs_mru_cache_done(mp->m_filestream);
290 trace_xfs_filestream_lookup(mp, ip->i_ino, ag);
295 * Set the starting AG using the rotor for inode32, otherwise
296 * use the directory inode's AG.
298 if (xfs_is_inode32(mp)) {
299 xfs_agnumber_t rotorstep = xfs_rotorstep;
300 startag = (mp->m_agfrotor / rotorstep) % mp->m_sb.sb_agcount;
301 mp->m_agfrotor = (mp->m_agfrotor + 1) %
302 (mp->m_sb.sb_agcount * rotorstep);
304 startag = XFS_INO_TO_AGNO(mp, pip->i_ino);
306 if (xfs_filestream_pick_ag(pip, startag, &ag, 0, 0))
314 * Pick a new allocation group for the current file and its file stream.
316 * This is called when the allocator can't find a suitable extent in the
317 * current AG, and we have to move the stream into a new AG with more space.
320 xfs_filestream_new_ag(
321 struct xfs_bmalloca *ap,
324 struct xfs_inode *ip = ap->ip, *pip;
325 struct xfs_mount *mp = ip->i_mount;
326 xfs_extlen_t minlen = ap->length;
327 xfs_agnumber_t startag = 0;
330 struct xfs_mru_cache_elem *mru;
334 pip = xfs_filestream_get_parent(ip);
338 mru = xfs_mru_cache_remove(mp->m_filestream, pip->i_ino);
340 struct xfs_fstrm_item *item =
341 container_of(mru, struct xfs_fstrm_item, mru);
342 startag = (item->ag + 1) % mp->m_sb.sb_agcount;
345 if (ap->datatype & XFS_ALLOC_USERDATA)
346 flags |= XFS_PICK_USERDATA;
347 if (ap->tp->t_flags & XFS_TRANS_LOWMODE)
348 flags |= XFS_PICK_LOWSPACE;
350 err = xfs_filestream_pick_ag(pip, startag, agp, flags, minlen);
353 * Only free the item here so we skip over the old AG earlier.
356 xfs_fstrm_free_func(mp, mru);
360 if (*agp == NULLAGNUMBER)
366 xfs_filestream_deassociate(
367 struct xfs_inode *ip)
369 xfs_mru_cache_delete(ip->i_mount->m_filestream, ip->i_ino);
373 xfs_filestream_mount(
377 * The filestream timer tunable is currently fixed within the range of
378 * one second to four minutes, with five seconds being the default. The
379 * group count is somewhat arbitrary, but it'd be nice to adhere to the
380 * timer tunable to within about 10 percent. This requires at least 10
383 return xfs_mru_cache_create(&mp->m_filestream, mp,
384 xfs_fstrm_centisecs * 10, 10, xfs_fstrm_free_func);
388 xfs_filestream_unmount(
391 xfs_mru_cache_destroy(mp->m_filestream);