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_defer.h"
15 #include "xfs_inode.h"
17 #include "xfs_bmap_util.h"
18 #include "xfs_alloc.h"
19 #include "xfs_mru_cache.h"
20 #include "xfs_filestream.h"
21 #include "xfs_trace.h"
22 #include "xfs_ag_resv.h"
23 #include "xfs_trans.h"
24 #include "xfs_shared.h"
26 struct xfs_fstrm_item {
27 struct xfs_mru_cache_elem mru;
28 xfs_agnumber_t ag; /* AG in use for this directory */
31 enum xfs_fstrm_alloc {
32 XFS_PICK_USERDATA = 1,
33 XFS_PICK_LOWSPACE = 2,
37 * Allocation group filestream associations are tracked with per-ag atomic
38 * counters. These counters allow xfs_filestream_pick_ag() to tell whether a
39 * particular AG already has active filestreams associated with it. The mount
40 * point's m_peraglock is used to protect these counters from per-ag array
41 * re-allocation during a growfs operation. When xfs_growfs_data_private() is
42 * about to reallocate the array, it calls xfs_filestream_flush() with the
43 * m_peraglock held in write mode.
45 * Since xfs_mru_cache_flush() guarantees that all the free functions for all
46 * the cache elements have finished executing before it returns, it's safe for
47 * the free functions to use the atomic counters without m_peraglock protection.
48 * This allows the implementation of xfs_fstrm_free_func() to be agnostic about
49 * whether it was called with the m_peraglock held in read mode, write mode or
50 * not held at all. The race condition this addresses is the following:
52 * - The work queue scheduler fires and pulls a filestream directory cache
53 * element off the LRU end of the cache for deletion, then gets pre-empted.
54 * - A growfs operation grabs the m_peraglock in write mode, flushes all the
55 * remaining items from the cache and reallocates the mount point's per-ag
56 * array, resetting all the counters to zero.
57 * - The work queue thread resumes and calls the free function for the element
58 * it started cleaning up earlier. In the process it decrements the
59 * filestreams counter for an AG that now has no references.
61 * With a shrinkfs feature, the above scenario could panic the system.
63 * All other uses of the following macros should be protected by either the
64 * m_peraglock held in read mode, or the cache's internal locking exposed by the
65 * interval between a call to xfs_mru_cache_lookup() and a call to
66 * xfs_mru_cache_done(). In addition, the m_peraglock must be held in read mode
67 * when new elements are added to the cache.
69 * Combined, these locking rules ensure that no associations will ever exist in
70 * the cache that reference per-ag array elements that have since been
74 xfs_filestream_peek_ag(
78 struct xfs_perag *pag;
81 pag = xfs_perag_get(mp, agno);
82 ret = atomic_read(&pag->pagf_fstrms);
88 xfs_filestream_get_ag(
92 struct xfs_perag *pag;
95 pag = xfs_perag_get(mp, agno);
96 ret = atomic_inc_return(&pag->pagf_fstrms);
102 xfs_filestream_put_ag(
106 struct xfs_perag *pag;
108 pag = xfs_perag_get(mp, agno);
109 atomic_dec(&pag->pagf_fstrms);
116 struct xfs_mru_cache_elem *mru)
118 struct xfs_mount *mp = data;
119 struct xfs_fstrm_item *item =
120 container_of(mru, struct xfs_fstrm_item, mru);
122 xfs_filestream_put_ag(mp, item->ag);
123 trace_xfs_filestream_free(mp, mru->key, item->ag);
129 * Scan the AGs starting at startag looking for an AG that isn't in use and has
130 * at least minlen blocks free.
133 xfs_filestream_pick_ag(
134 struct xfs_inode *ip,
135 xfs_agnumber_t startag,
140 struct xfs_mount *mp = ip->i_mount;
141 struct xfs_fstrm_item *item;
142 struct xfs_perag *pag;
143 xfs_extlen_t longest, free = 0, minfree, maxfree = 0;
144 xfs_agnumber_t ag, max_ag = NULLAGNUMBER;
145 int err, trylock, nscan;
147 ASSERT(S_ISDIR(VFS_I(ip)->i_mode));
149 /* 2% of an AG's blocks must be free for it to be chosen. */
150 minfree = mp->m_sb.sb_agblocks / 50;
155 /* For the first pass, don't sleep trying to init the per-AG. */
156 trylock = XFS_ALLOC_FLAG_TRYLOCK;
158 for (nscan = 0; 1; nscan++) {
159 trace_xfs_filestream_scan(mp, ip->i_ino, ag);
161 pag = xfs_perag_get(mp, ag);
163 if (!pag->pagf_init) {
164 err = xfs_alloc_pagf_init(mp, NULL, ag, trylock);
165 if (err && !trylock) {
171 /* Might fail sometimes during the 1st pass with trylock set. */
175 /* Keep track of the AG with the most free blocks. */
176 if (pag->pagf_freeblks > maxfree) {
177 maxfree = pag->pagf_freeblks;
182 * The AG reference count does two things: it enforces mutual
183 * exclusion when examining the suitability of an AG in this
184 * loop, and it guards against two filestreams being established
185 * in the same AG as each other.
187 if (xfs_filestream_get_ag(mp, ag) > 1) {
188 xfs_filestream_put_ag(mp, ag);
192 longest = xfs_alloc_longest_free_extent(pag,
193 xfs_alloc_min_freelist(mp, pag),
194 xfs_ag_resv_needed(pag, XFS_AG_RESV_NONE));
195 if (((minlen && longest >= minlen) ||
196 (!minlen && pag->pagf_freeblks >= minfree)) &&
197 (!pag->pagf_metadata || !(flags & XFS_PICK_USERDATA) ||
198 (flags & XFS_PICK_LOWSPACE))) {
200 /* Break out, retaining the reference on the AG. */
201 free = pag->pagf_freeblks;
207 /* Drop the reference on this AG, it's not usable. */
208 xfs_filestream_put_ag(mp, ag);
211 /* Move to the next AG, wrapping to AG 0 if necessary. */
212 if (++ag >= mp->m_sb.sb_agcount)
215 /* If a full pass of the AGs hasn't been done yet, continue. */
219 /* Allow sleeping in xfs_alloc_pagf_init() on the 2nd pass. */
225 /* Finally, if lowspace wasn't set, set it for the 3rd pass. */
226 if (!(flags & XFS_PICK_LOWSPACE)) {
227 flags |= XFS_PICK_LOWSPACE;
232 * Take the AG with the most free space, regardless of whether
233 * it's already in use by another filestream.
235 if (max_ag != NULLAGNUMBER) {
236 xfs_filestream_get_ag(mp, max_ag);
242 /* take AG 0 if none matched */
243 trace_xfs_filestream_pick(ip, *agp, free, nscan);
248 trace_xfs_filestream_pick(ip, *agp, free, nscan);
250 if (*agp == NULLAGNUMBER)
254 item = kmem_alloc(sizeof(*item), KM_MAYFAIL);
260 err = xfs_mru_cache_insert(mp->m_filestream, ip->i_ino, &item->mru);
272 xfs_filestream_put_ag(mp, *agp);
276 static struct xfs_inode *
277 xfs_filestream_get_parent(
278 struct xfs_inode *ip)
280 struct inode *inode = VFS_I(ip), *dir = NULL;
281 struct dentry *dentry, *parent;
283 dentry = d_find_alias(inode);
287 parent = dget_parent(dentry);
291 dir = igrab(d_inode(parent));
297 return dir ? XFS_I(dir) : NULL;
301 * Find the right allocation group for a file, either by finding an
302 * existing file stream or creating a new one.
304 * Returns NULLAGNUMBER in case of an error.
307 xfs_filestream_lookup_ag(
308 struct xfs_inode *ip)
310 struct xfs_mount *mp = ip->i_mount;
311 struct xfs_inode *pip = NULL;
312 xfs_agnumber_t startag, ag = NULLAGNUMBER;
313 struct xfs_mru_cache_elem *mru;
315 ASSERT(S_ISREG(VFS_I(ip)->i_mode));
317 pip = xfs_filestream_get_parent(ip);
321 mru = xfs_mru_cache_lookup(mp->m_filestream, pip->i_ino);
323 ag = container_of(mru, struct xfs_fstrm_item, mru)->ag;
324 xfs_mru_cache_done(mp->m_filestream);
326 trace_xfs_filestream_lookup(mp, ip->i_ino, ag);
331 * Set the starting AG using the rotor for inode32, otherwise
332 * use the directory inode's AG.
334 if (mp->m_flags & XFS_MOUNT_32BITINODES) {
335 xfs_agnumber_t rotorstep = xfs_rotorstep;
336 startag = (mp->m_agfrotor / rotorstep) % mp->m_sb.sb_agcount;
337 mp->m_agfrotor = (mp->m_agfrotor + 1) %
338 (mp->m_sb.sb_agcount * rotorstep);
340 startag = XFS_INO_TO_AGNO(mp, pip->i_ino);
342 if (xfs_filestream_pick_ag(pip, startag, &ag, 0, 0))
350 * Pick a new allocation group for the current file and its file stream.
352 * This is called when the allocator can't find a suitable extent in the
353 * current AG, and we have to move the stream into a new AG with more space.
356 xfs_filestream_new_ag(
357 struct xfs_bmalloca *ap,
360 struct xfs_inode *ip = ap->ip, *pip;
361 struct xfs_mount *mp = ip->i_mount;
362 xfs_extlen_t minlen = ap->length;
363 xfs_agnumber_t startag = 0;
366 struct xfs_mru_cache_elem *mru;
370 pip = xfs_filestream_get_parent(ip);
374 mru = xfs_mru_cache_remove(mp->m_filestream, pip->i_ino);
376 struct xfs_fstrm_item *item =
377 container_of(mru, struct xfs_fstrm_item, mru);
378 startag = (item->ag + 1) % mp->m_sb.sb_agcount;
381 if (xfs_alloc_is_userdata(ap->datatype))
382 flags |= XFS_PICK_USERDATA;
383 if (ap->tp->t_flags & XFS_TRANS_LOWMODE)
384 flags |= XFS_PICK_LOWSPACE;
386 err = xfs_filestream_pick_ag(pip, startag, agp, flags, minlen);
389 * Only free the item here so we skip over the old AG earlier.
392 xfs_fstrm_free_func(mp, mru);
396 if (*agp == NULLAGNUMBER)
402 xfs_filestream_deassociate(
403 struct xfs_inode *ip)
405 xfs_mru_cache_delete(ip->i_mount->m_filestream, ip->i_ino);
409 xfs_filestream_mount(
413 * The filestream timer tunable is currently fixed within the range of
414 * one second to four minutes, with five seconds being the default. The
415 * group count is somewhat arbitrary, but it'd be nice to adhere to the
416 * timer tunable to within about 10 percent. This requires at least 10
419 return xfs_mru_cache_create(&mp->m_filestream, mp,
420 xfs_fstrm_centisecs * 10, 10, xfs_fstrm_free_func);
424 xfs_filestream_unmount(
427 xfs_mru_cache_destroy(mp->m_filestream);