1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (c) 2006-2007 Silicon Graphics, Inc.
4 * Copyright (c) 2014 Christoph Hellwig.
8 #include "xfs_format.h"
9 #include "xfs_log_format.h"
10 #include "xfs_trans_resv.h"
12 #include "xfs_mount.h"
13 #include "xfs_defer.h"
14 #include "xfs_inode.h"
16 #include "xfs_bmap_util.h"
17 #include "xfs_alloc.h"
18 #include "xfs_mru_cache.h"
19 #include "xfs_filestream.h"
20 #include "xfs_trace.h"
21 #include "xfs_ag_resv.h"
22 #include "xfs_trans.h"
23 #include "xfs_shared.h"
25 struct xfs_fstrm_item {
26 struct xfs_mru_cache_elem mru;
27 xfs_agnumber_t ag; /* AG in use for this directory */
30 enum xfs_fstrm_alloc {
31 XFS_PICK_USERDATA = 1,
32 XFS_PICK_LOWSPACE = 2,
36 * Allocation group filestream associations are tracked with per-ag atomic
37 * counters. These counters allow xfs_filestream_pick_ag() to tell whether a
38 * particular AG already has active filestreams associated with it. The mount
39 * point's m_peraglock is used to protect these counters from per-ag array
40 * re-allocation during a growfs operation. When xfs_growfs_data_private() is
41 * about to reallocate the array, it calls xfs_filestream_flush() with the
42 * m_peraglock held in write mode.
44 * Since xfs_mru_cache_flush() guarantees that all the free functions for all
45 * the cache elements have finished executing before it returns, it's safe for
46 * the free functions to use the atomic counters without m_peraglock protection.
47 * This allows the implementation of xfs_fstrm_free_func() to be agnostic about
48 * whether it was called with the m_peraglock held in read mode, write mode or
49 * not held at all. The race condition this addresses is the following:
51 * - The work queue scheduler fires and pulls a filestream directory cache
52 * element off the LRU end of the cache for deletion, then gets pre-empted.
53 * - A growfs operation grabs the m_peraglock in write mode, flushes all the
54 * remaining items from the cache and reallocates the mount point's per-ag
55 * array, resetting all the counters to zero.
56 * - The work queue thread resumes and calls the free function for the element
57 * it started cleaning up earlier. In the process it decrements the
58 * filestreams counter for an AG that now has no references.
60 * With a shrinkfs feature, the above scenario could panic the system.
62 * All other uses of the following macros should be protected by either the
63 * m_peraglock held in read mode, or the cache's internal locking exposed by the
64 * interval between a call to xfs_mru_cache_lookup() and a call to
65 * xfs_mru_cache_done(). In addition, the m_peraglock must be held in read mode
66 * when new elements are added to the cache.
68 * Combined, these locking rules ensure that no associations will ever exist in
69 * the cache that reference per-ag array elements that have since been
73 xfs_filestream_peek_ag(
77 struct xfs_perag *pag;
80 pag = xfs_perag_get(mp, agno);
81 ret = atomic_read(&pag->pagf_fstrms);
87 xfs_filestream_get_ag(
91 struct xfs_perag *pag;
94 pag = xfs_perag_get(mp, agno);
95 ret = atomic_inc_return(&pag->pagf_fstrms);
101 xfs_filestream_put_ag(
105 struct xfs_perag *pag;
107 pag = xfs_perag_get(mp, agno);
108 atomic_dec(&pag->pagf_fstrms);
115 struct xfs_mru_cache_elem *mru)
117 struct xfs_mount *mp = data;
118 struct xfs_fstrm_item *item =
119 container_of(mru, struct xfs_fstrm_item, mru);
121 xfs_filestream_put_ag(mp, item->ag);
122 trace_xfs_filestream_free(mp, mru->key, item->ag);
128 * Scan the AGs starting at startag looking for an AG that isn't in use and has
129 * at least minlen blocks free.
132 xfs_filestream_pick_ag(
133 struct xfs_inode *ip,
134 xfs_agnumber_t startag,
139 struct xfs_mount *mp = ip->i_mount;
140 struct xfs_fstrm_item *item;
141 struct xfs_perag *pag;
142 xfs_extlen_t longest, free = 0, minfree, maxfree = 0;
143 xfs_agnumber_t ag, max_ag = NULLAGNUMBER;
144 int err, trylock, nscan;
146 ASSERT(S_ISDIR(VFS_I(ip)->i_mode));
148 /* 2% of an AG's blocks must be free for it to be chosen. */
149 minfree = mp->m_sb.sb_agblocks / 50;
154 /* For the first pass, don't sleep trying to init the per-AG. */
155 trylock = XFS_ALLOC_FLAG_TRYLOCK;
157 for (nscan = 0; 1; nscan++) {
158 trace_xfs_filestream_scan(mp, ip->i_ino, ag);
160 pag = xfs_perag_get(mp, ag);
162 if (!pag->pagf_init) {
163 err = xfs_alloc_pagf_init(mp, NULL, ag, trylock);
164 if (err && !trylock) {
170 /* Might fail sometimes during the 1st pass with trylock set. */
174 /* Keep track of the AG with the most free blocks. */
175 if (pag->pagf_freeblks > maxfree) {
176 maxfree = pag->pagf_freeblks;
181 * The AG reference count does two things: it enforces mutual
182 * exclusion when examining the suitability of an AG in this
183 * loop, and it guards against two filestreams being established
184 * in the same AG as each other.
186 if (xfs_filestream_get_ag(mp, ag) > 1) {
187 xfs_filestream_put_ag(mp, ag);
191 longest = xfs_alloc_longest_free_extent(pag,
192 xfs_alloc_min_freelist(mp, pag),
193 xfs_ag_resv_needed(pag, XFS_AG_RESV_NONE));
194 if (((minlen && longest >= minlen) ||
195 (!minlen && pag->pagf_freeblks >= minfree)) &&
196 (!pag->pagf_metadata || !(flags & XFS_PICK_USERDATA) ||
197 (flags & XFS_PICK_LOWSPACE))) {
199 /* Break out, retaining the reference on the AG. */
200 free = pag->pagf_freeblks;
206 /* Drop the reference on this AG, it's not usable. */
207 xfs_filestream_put_ag(mp, ag);
210 /* Move to the next AG, wrapping to AG 0 if necessary. */
211 if (++ag >= mp->m_sb.sb_agcount)
214 /* If a full pass of the AGs hasn't been done yet, continue. */
218 /* Allow sleeping in xfs_alloc_pagf_init() on the 2nd pass. */
224 /* Finally, if lowspace wasn't set, set it for the 3rd pass. */
225 if (!(flags & XFS_PICK_LOWSPACE)) {
226 flags |= XFS_PICK_LOWSPACE;
231 * Take the AG with the most free space, regardless of whether
232 * it's already in use by another filestream.
234 if (max_ag != NULLAGNUMBER) {
235 xfs_filestream_get_ag(mp, max_ag);
241 /* take AG 0 if none matched */
242 trace_xfs_filestream_pick(ip, *agp, free, nscan);
247 trace_xfs_filestream_pick(ip, *agp, free, nscan);
249 if (*agp == NULLAGNUMBER)
253 item = kmem_alloc(sizeof(*item), KM_MAYFAIL);
259 err = xfs_mru_cache_insert(mp->m_filestream, ip->i_ino, &item->mru);
271 xfs_filestream_put_ag(mp, *agp);
275 static struct xfs_inode *
276 xfs_filestream_get_parent(
277 struct xfs_inode *ip)
279 struct inode *inode = VFS_I(ip), *dir = NULL;
280 struct dentry *dentry, *parent;
282 dentry = d_find_alias(inode);
286 parent = dget_parent(dentry);
290 dir = igrab(d_inode(parent));
296 return dir ? XFS_I(dir) : NULL;
300 * Find the right allocation group for a file, either by finding an
301 * existing file stream or creating a new one.
303 * Returns NULLAGNUMBER in case of an error.
306 xfs_filestream_lookup_ag(
307 struct xfs_inode *ip)
309 struct xfs_mount *mp = ip->i_mount;
310 struct xfs_inode *pip = NULL;
311 xfs_agnumber_t startag, ag = NULLAGNUMBER;
312 struct xfs_mru_cache_elem *mru;
314 ASSERT(S_ISREG(VFS_I(ip)->i_mode));
316 pip = xfs_filestream_get_parent(ip);
320 mru = xfs_mru_cache_lookup(mp->m_filestream, pip->i_ino);
322 ag = container_of(mru, struct xfs_fstrm_item, mru)->ag;
323 xfs_mru_cache_done(mp->m_filestream);
325 trace_xfs_filestream_lookup(mp, ip->i_ino, ag);
330 * Set the starting AG using the rotor for inode32, otherwise
331 * use the directory inode's AG.
333 if (mp->m_flags & XFS_MOUNT_32BITINODES) {
334 xfs_agnumber_t rotorstep = xfs_rotorstep;
335 startag = (mp->m_agfrotor / rotorstep) % mp->m_sb.sb_agcount;
336 mp->m_agfrotor = (mp->m_agfrotor + 1) %
337 (mp->m_sb.sb_agcount * rotorstep);
339 startag = XFS_INO_TO_AGNO(mp, pip->i_ino);
341 if (xfs_filestream_pick_ag(pip, startag, &ag, 0, 0))
349 * Pick a new allocation group for the current file and its file stream.
351 * This is called when the allocator can't find a suitable extent in the
352 * current AG, and we have to move the stream into a new AG with more space.
355 xfs_filestream_new_ag(
356 struct xfs_bmalloca *ap,
359 struct xfs_inode *ip = ap->ip, *pip;
360 struct xfs_mount *mp = ip->i_mount;
361 xfs_extlen_t minlen = ap->length;
362 xfs_agnumber_t startag = 0;
365 struct xfs_mru_cache_elem *mru;
369 pip = xfs_filestream_get_parent(ip);
373 mru = xfs_mru_cache_remove(mp->m_filestream, pip->i_ino);
375 struct xfs_fstrm_item *item =
376 container_of(mru, struct xfs_fstrm_item, mru);
377 startag = (item->ag + 1) % mp->m_sb.sb_agcount;
380 if (xfs_alloc_is_userdata(ap->datatype))
381 flags |= XFS_PICK_USERDATA;
382 if (ap->tp->t_flags & XFS_TRANS_LOWMODE)
383 flags |= XFS_PICK_LOWSPACE;
385 err = xfs_filestream_pick_ag(pip, startag, agp, flags, minlen);
388 * Only free the item here so we skip over the old AG earlier.
391 xfs_fstrm_free_func(mp, mru);
395 if (*agp == NULLAGNUMBER)
401 xfs_filestream_deassociate(
402 struct xfs_inode *ip)
404 xfs_mru_cache_delete(ip->i_mount->m_filestream, ip->i_ino);
408 xfs_filestream_mount(
412 * The filestream timer tunable is currently fixed within the range of
413 * one second to four minutes, with five seconds being the default. The
414 * group count is somewhat arbitrary, but it'd be nice to adhere to the
415 * timer tunable to within about 10 percent. This requires at least 10
418 return xfs_mru_cache_create(&mp->m_filestream, mp,
419 xfs_fstrm_centisecs * 10, 10, xfs_fstrm_free_func);
423 xfs_filestream_unmount(
426 xfs_mru_cache_destroy(mp->m_filestream);