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"
13 #include "xfs_mount.h"
14 #include "xfs_inode.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"
22 struct xfs_fstrm_item {
23 struct xfs_mru_cache_elem mru;
24 xfs_agnumber_t ag; /* AG in use for this directory */
27 enum xfs_fstrm_alloc {
28 XFS_PICK_USERDATA = 1,
29 XFS_PICK_LOWSPACE = 2,
33 * Allocation group filestream associations are tracked with per-ag atomic
34 * counters. These counters allow xfs_filestream_pick_ag() to tell whether a
35 * particular AG already has active filestreams associated with it. The mount
36 * point's m_peraglock is used to protect these counters from per-ag array
37 * re-allocation during a growfs operation. When xfs_growfs_data_private() is
38 * about to reallocate the array, it calls xfs_filestream_flush() with the
39 * m_peraglock held in write mode.
41 * Since xfs_mru_cache_flush() guarantees that all the free functions for all
42 * the cache elements have finished executing before it returns, it's safe for
43 * the free functions to use the atomic counters without m_peraglock protection.
44 * This allows the implementation of xfs_fstrm_free_func() to be agnostic about
45 * whether it was called with the m_peraglock held in read mode, write mode or
46 * not held at all. The race condition this addresses is the following:
48 * - The work queue scheduler fires and pulls a filestream directory cache
49 * element off the LRU end of the cache for deletion, then gets pre-empted.
50 * - A growfs operation grabs the m_peraglock in write mode, flushes all the
51 * remaining items from the cache and reallocates the mount point's per-ag
52 * array, resetting all the counters to zero.
53 * - The work queue thread resumes and calls the free function for the element
54 * it started cleaning up earlier. In the process it decrements the
55 * filestreams counter for an AG that now has no references.
57 * With a shrinkfs feature, the above scenario could panic the system.
59 * All other uses of the following macros should be protected by either the
60 * m_peraglock held in read mode, or the cache's internal locking exposed by the
61 * interval between a call to xfs_mru_cache_lookup() and a call to
62 * xfs_mru_cache_done(). In addition, the m_peraglock must be held in read mode
63 * when new elements are added to the cache.
65 * Combined, these locking rules ensure that no associations will ever exist in
66 * the cache that reference per-ag array elements that have since been
70 xfs_filestream_peek_ag(
74 struct xfs_perag *pag;
77 pag = xfs_perag_get(mp, agno);
78 ret = atomic_read(&pag->pagf_fstrms);
84 xfs_filestream_get_ag(
88 struct xfs_perag *pag;
91 pag = xfs_perag_get(mp, agno);
92 ret = atomic_inc_return(&pag->pagf_fstrms);
98 xfs_filestream_put_ag(
102 struct xfs_perag *pag;
104 pag = xfs_perag_get(mp, agno);
105 atomic_dec(&pag->pagf_fstrms);
112 struct xfs_mru_cache_elem *mru)
114 struct xfs_mount *mp = data;
115 struct xfs_fstrm_item *item =
116 container_of(mru, struct xfs_fstrm_item, mru);
118 xfs_filestream_put_ag(mp, item->ag);
119 trace_xfs_filestream_free(mp, mru->key, item->ag);
125 * Scan the AGs starting at startag looking for an AG that isn't in use and has
126 * at least minlen blocks free.
129 xfs_filestream_pick_ag(
130 struct xfs_inode *ip,
131 xfs_agnumber_t startag,
136 struct xfs_mount *mp = ip->i_mount;
137 struct xfs_fstrm_item *item;
138 struct xfs_perag *pag;
139 xfs_extlen_t longest, free = 0, minfree, maxfree = 0;
140 xfs_agnumber_t ag, max_ag = NULLAGNUMBER;
141 int err, trylock, nscan;
143 ASSERT(S_ISDIR(VFS_I(ip)->i_mode));
145 /* 2% of an AG's blocks must be free for it to be chosen. */
146 minfree = mp->m_sb.sb_agblocks / 50;
151 /* For the first pass, don't sleep trying to init the per-AG. */
152 trylock = XFS_ALLOC_FLAG_TRYLOCK;
154 for (nscan = 0; 1; nscan++) {
155 trace_xfs_filestream_scan(mp, ip->i_ino, ag);
157 pag = xfs_perag_get(mp, ag);
159 if (!pag->pagf_init) {
160 err = xfs_alloc_pagf_init(mp, NULL, ag, trylock);
161 if (err && !trylock) {
167 /* Might fail sometimes during the 1st pass with trylock set. */
171 /* Keep track of the AG with the most free blocks. */
172 if (pag->pagf_freeblks > maxfree) {
173 maxfree = pag->pagf_freeblks;
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.
183 if (xfs_filestream_get_ag(mp, ag) > 1) {
184 xfs_filestream_put_ag(mp, ag);
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))) {
196 /* Break out, retaining the reference on the AG. */
197 free = pag->pagf_freeblks;
203 /* Drop the reference on this AG, it's not usable. */
204 xfs_filestream_put_ag(mp, ag);
207 /* Move to the next AG, wrapping to AG 0 if necessary. */
208 if (++ag >= mp->m_sb.sb_agcount)
211 /* If a full pass of the AGs hasn't been done yet, continue. */
215 /* Allow sleeping in xfs_alloc_pagf_init() on the 2nd pass. */
221 /* Finally, if lowspace wasn't set, set it for the 3rd pass. */
222 if (!(flags & XFS_PICK_LOWSPACE)) {
223 flags |= XFS_PICK_LOWSPACE;
228 * Take the AG with the most free space, regardless of whether
229 * it's already in use by another filestream.
231 if (max_ag != NULLAGNUMBER) {
232 xfs_filestream_get_ag(mp, max_ag);
238 /* take AG 0 if none matched */
239 trace_xfs_filestream_pick(ip, *agp, free, nscan);
244 trace_xfs_filestream_pick(ip, *agp, free, nscan);
246 if (*agp == NULLAGNUMBER)
250 item = kmem_alloc(sizeof(*item), KM_MAYFAIL);
256 err = xfs_mru_cache_insert(mp->m_filestream, ip->i_ino, &item->mru);
268 xfs_filestream_put_ag(mp, *agp);
272 static struct xfs_inode *
273 xfs_filestream_get_parent(
274 struct xfs_inode *ip)
276 struct inode *inode = VFS_I(ip), *dir = NULL;
277 struct dentry *dentry, *parent;
279 dentry = d_find_alias(inode);
283 parent = dget_parent(dentry);
287 dir = igrab(d_inode(parent));
293 return dir ? XFS_I(dir) : NULL;
297 * Find the right allocation group for a file, either by finding an
298 * existing file stream or creating a new one.
300 * Returns NULLAGNUMBER in case of an error.
303 xfs_filestream_lookup_ag(
304 struct xfs_inode *ip)
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;
311 ASSERT(S_ISREG(VFS_I(ip)->i_mode));
313 pip = xfs_filestream_get_parent(ip);
317 mru = xfs_mru_cache_lookup(mp->m_filestream, pip->i_ino);
319 ag = container_of(mru, struct xfs_fstrm_item, mru)->ag;
320 xfs_mru_cache_done(mp->m_filestream);
322 trace_xfs_filestream_lookup(mp, ip->i_ino, ag);
327 * Set the starting AG using the rotor for inode32, otherwise
328 * use the directory inode's AG.
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);
336 startag = XFS_INO_TO_AGNO(mp, pip->i_ino);
338 if (xfs_filestream_pick_ag(pip, startag, &ag, 0, 0))
346 * Pick a new allocation group for the current file and its file stream.
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.
352 xfs_filestream_new_ag(
353 struct xfs_bmalloca *ap,
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;
362 struct xfs_mru_cache_elem *mru;
366 pip = xfs_filestream_get_parent(ip);
370 mru = xfs_mru_cache_remove(mp->m_filestream, pip->i_ino);
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;
377 if (xfs_alloc_is_userdata(ap->datatype))
378 flags |= XFS_PICK_USERDATA;
379 if (ap->tp->t_flags & XFS_TRANS_LOWMODE)
380 flags |= XFS_PICK_LOWSPACE;
382 err = xfs_filestream_pick_ag(pip, startag, agp, flags, minlen);
385 * Only free the item here so we skip over the old AG earlier.
388 xfs_fstrm_free_func(mp, mru);
392 if (*agp == NULLAGNUMBER)
398 xfs_filestream_deassociate(
399 struct xfs_inode *ip)
401 xfs_mru_cache_delete(ip->i_mount->m_filestream, ip->i_ino);
405 xfs_filestream_mount(
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
415 return xfs_mru_cache_create(&mp->m_filestream, mp,
416 xfs_fstrm_centisecs * 10, 10, xfs_fstrm_free_func);
420 xfs_filestream_unmount(
423 xfs_mru_cache_destroy(mp->m_filestream);