Linux Audio

Check our new training course

Loading...
v4.17
 
  1/*
  2 * Copyright (c) 2006-2007 Silicon Graphics, Inc.
  3 * Copyright (c) 2014 Christoph Hellwig.
  4 * All Rights Reserved.
  5 *
  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.
  9 *
 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.
 14 *
 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
 18 */
 19#include "xfs.h"
 
 20#include "xfs_format.h"
 21#include "xfs_log_format.h"
 22#include "xfs_trans_resv.h"
 23#include "xfs_sb.h"
 24#include "xfs_mount.h"
 25#include "xfs_defer.h"
 26#include "xfs_inode.h"
 27#include "xfs_bmap.h"
 28#include "xfs_bmap_util.h"
 29#include "xfs_alloc.h"
 30#include "xfs_mru_cache.h"
 31#include "xfs_filestream.h"
 32#include "xfs_trace.h"
 33#include "xfs_ag_resv.h"
 
 
 34
 35struct xfs_fstrm_item {
 36	struct xfs_mru_cache_elem	mru;
 37	xfs_agnumber_t			ag; /* AG in use for this directory */
 38};
 39
 40enum xfs_fstrm_alloc {
 41	XFS_PICK_USERDATA = 1,
 42	XFS_PICK_LOWSPACE = 2,
 43};
 44
 45/*
 46 * Allocation group filestream associations are tracked with per-ag atomic
 47 * counters.  These counters allow xfs_filestream_pick_ag() to tell whether a
 48 * particular AG already has active filestreams associated with it. The mount
 49 * point's m_peraglock is used to protect these counters from per-ag array
 50 * re-allocation during a growfs operation.  When xfs_growfs_data_private() is
 51 * about to reallocate the array, it calls xfs_filestream_flush() with the
 52 * m_peraglock held in write mode.
 53 *
 54 * Since xfs_mru_cache_flush() guarantees that all the free functions for all
 55 * the cache elements have finished executing before it returns, it's safe for
 56 * the free functions to use the atomic counters without m_peraglock protection.
 57 * This allows the implementation of xfs_fstrm_free_func() to be agnostic about
 58 * whether it was called with the m_peraglock held in read mode, write mode or
 59 * not held at all.  The race condition this addresses is the following:
 60 *
 61 *  - The work queue scheduler fires and pulls a filestream directory cache
 62 *    element off the LRU end of the cache for deletion, then gets pre-empted.
 63 *  - A growfs operation grabs the m_peraglock in write mode, flushes all the
 64 *    remaining items from the cache and reallocates the mount point's per-ag
 65 *    array, resetting all the counters to zero.
 66 *  - The work queue thread resumes and calls the free function for the element
 67 *    it started cleaning up earlier.  In the process it decrements the
 68 *    filestreams counter for an AG that now has no references.
 69 *
 70 * With a shrinkfs feature, the above scenario could panic the system.
 71 *
 72 * All other uses of the following macros should be protected by either the
 73 * m_peraglock held in read mode, or the cache's internal locking exposed by the
 74 * interval between a call to xfs_mru_cache_lookup() and a call to
 75 * xfs_mru_cache_done().  In addition, the m_peraglock must be held in read mode
 76 * when new elements are added to the cache.
 77 *
 78 * Combined, these locking rules ensure that no associations will ever exist in
 79 * the cache that reference per-ag array elements that have since been
 80 * reallocated.
 81 */
 82int
 83xfs_filestream_peek_ag(
 84	xfs_mount_t	*mp,
 85	xfs_agnumber_t	agno)
 86{
 87	struct xfs_perag *pag;
 88	int		ret;
 89
 90	pag = xfs_perag_get(mp, agno);
 91	ret = atomic_read(&pag->pagf_fstrms);
 92	xfs_perag_put(pag);
 93	return ret;
 94}
 95
 96static int
 97xfs_filestream_get_ag(
 98	xfs_mount_t	*mp,
 99	xfs_agnumber_t	agno)
100{
101	struct xfs_perag *pag;
102	int		ret;
103
104	pag = xfs_perag_get(mp, agno);
105	ret = atomic_inc_return(&pag->pagf_fstrms);
106	xfs_perag_put(pag);
107	return ret;
108}
109
110static void
111xfs_filestream_put_ag(
112	xfs_mount_t	*mp,
113	xfs_agnumber_t	agno)
114{
115	struct xfs_perag *pag;
116
117	pag = xfs_perag_get(mp, agno);
118	atomic_dec(&pag->pagf_fstrms);
119	xfs_perag_put(pag);
120}
121
122static void
123xfs_fstrm_free_func(
124	void			*data,
125	struct xfs_mru_cache_elem *mru)
126{
127	struct xfs_mount	*mp = data;
128	struct xfs_fstrm_item	*item =
129		container_of(mru, struct xfs_fstrm_item, mru);
130
131	xfs_filestream_put_ag(mp, item->ag);
132	trace_xfs_filestream_free(mp, mru->key, item->ag);
133
134	kmem_free(item);
135}
136
137/*
138 * Scan the AGs starting at startag looking for an AG that isn't in use and has
139 * at least minlen blocks free.
140 */
141static int
142xfs_filestream_pick_ag(
143	struct xfs_inode	*ip,
144	xfs_agnumber_t		startag,
145	xfs_agnumber_t		*agp,
146	int			flags,
147	xfs_extlen_t		minlen)
148{
149	struct xfs_mount	*mp = ip->i_mount;
150	struct xfs_fstrm_item	*item;
151	struct xfs_perag	*pag;
152	xfs_extlen_t		longest, free = 0, minfree, maxfree = 0;
153	xfs_agnumber_t		ag, max_ag = NULLAGNUMBER;
154	int			err, trylock, nscan;
155
156	ASSERT(S_ISDIR(VFS_I(ip)->i_mode));
157
158	/* 2% of an AG's blocks must be free for it to be chosen. */
159	minfree = mp->m_sb.sb_agblocks / 50;
160
161	ag = startag;
162	*agp = NULLAGNUMBER;
163
164	/* For the first pass, don't sleep trying to init the per-AG. */
165	trylock = XFS_ALLOC_FLAG_TRYLOCK;
166
167	for (nscan = 0; 1; nscan++) {
168		trace_xfs_filestream_scan(mp, ip->i_ino, ag);
169
170		pag = xfs_perag_get(mp, ag);
171
172		if (!pag->pagf_init) {
173			err = xfs_alloc_pagf_init(mp, NULL, ag, trylock);
174			if (err && !trylock) {
175				xfs_perag_put(pag);
176				return err;
 
 
 
177			}
178		}
179
180		/* Might fail sometimes during the 1st pass with trylock set. */
181		if (!pag->pagf_init)
182			goto next_ag;
183
184		/* Keep track of the AG with the most free blocks. */
185		if (pag->pagf_freeblks > maxfree) {
186			maxfree = pag->pagf_freeblks;
187			max_ag = ag;
188		}
189
190		/*
191		 * The AG reference count does two things: it enforces mutual
192		 * exclusion when examining the suitability of an AG in this
193		 * loop, and it guards against two filestreams being established
194		 * in the same AG as each other.
195		 */
196		if (xfs_filestream_get_ag(mp, ag) > 1) {
197			xfs_filestream_put_ag(mp, ag);
198			goto next_ag;
199		}
200
201		longest = xfs_alloc_longest_free_extent(pag,
202				xfs_alloc_min_freelist(mp, pag),
203				xfs_ag_resv_needed(pag, XFS_AG_RESV_NONE));
204		if (((minlen && longest >= minlen) ||
205		     (!minlen && pag->pagf_freeblks >= minfree)) &&
206		    (!pag->pagf_metadata || !(flags & XFS_PICK_USERDATA) ||
207		     (flags & XFS_PICK_LOWSPACE))) {
208
209			/* Break out, retaining the reference on the AG. */
210			free = pag->pagf_freeblks;
211			xfs_perag_put(pag);
212			*agp = ag;
213			break;
214		}
215
216		/* Drop the reference on this AG, it's not usable. */
217		xfs_filestream_put_ag(mp, ag);
218next_ag:
219		xfs_perag_put(pag);
220		/* Move to the next AG, wrapping to AG 0 if necessary. */
221		if (++ag >= mp->m_sb.sb_agcount)
222			ag = 0;
223
224		/* If a full pass of the AGs hasn't been done yet, continue. */
225		if (ag != startag)
226			continue;
227
228		/* Allow sleeping in xfs_alloc_pagf_init() on the 2nd pass. */
229		if (trylock != 0) {
230			trylock = 0;
231			continue;
232		}
233
234		/* Finally, if lowspace wasn't set, set it for the 3rd pass. */
235		if (!(flags & XFS_PICK_LOWSPACE)) {
236			flags |= XFS_PICK_LOWSPACE;
237			continue;
238		}
239
240		/*
241		 * Take the AG with the most free space, regardless of whether
242		 * it's already in use by another filestream.
243		 */
244		if (max_ag != NULLAGNUMBER) {
245			xfs_filestream_get_ag(mp, max_ag);
246			free = maxfree;
247			*agp = max_ag;
248			break;
249		}
250
251		/* take AG 0 if none matched */
252		trace_xfs_filestream_pick(ip, *agp, free, nscan);
253		*agp = 0;
254		return 0;
255	}
256
257	trace_xfs_filestream_pick(ip, *agp, free, nscan);
258
259	if (*agp == NULLAGNUMBER)
260		return 0;
261
262	err = -ENOMEM;
263	item = kmem_alloc(sizeof(*item), KM_MAYFAIL);
264	if (!item)
265		goto out_put_ag;
266
267	item->ag = *agp;
268
269	err = xfs_mru_cache_insert(mp->m_filestream, ip->i_ino, &item->mru);
270	if (err) {
271		if (err == -EEXIST)
272			err = 0;
273		goto out_free_item;
274	}
275
276	return 0;
277
278out_free_item:
279	kmem_free(item);
280out_put_ag:
281	xfs_filestream_put_ag(mp, *agp);
282	return err;
283}
284
285static struct xfs_inode *
286xfs_filestream_get_parent(
287	struct xfs_inode	*ip)
288{
289	struct inode		*inode = VFS_I(ip), *dir = NULL;
290	struct dentry		*dentry, *parent;
291
292	dentry = d_find_alias(inode);
293	if (!dentry)
294		goto out;
295
296	parent = dget_parent(dentry);
297	if (!parent)
298		goto out_dput;
299
300	dir = igrab(d_inode(parent));
301	dput(parent);
302
303out_dput:
304	dput(dentry);
305out:
306	return dir ? XFS_I(dir) : NULL;
307}
308
309/*
310 * Find the right allocation group for a file, either by finding an
311 * existing file stream or creating a new one.
312 *
313 * Returns NULLAGNUMBER in case of an error.
314 */
315xfs_agnumber_t
316xfs_filestream_lookup_ag(
317	struct xfs_inode	*ip)
318{
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;
323
324	ASSERT(S_ISREG(VFS_I(ip)->i_mode));
325
326	pip = xfs_filestream_get_parent(ip);
327	if (!pip)
328		return NULLAGNUMBER;
329
330	mru = xfs_mru_cache_lookup(mp->m_filestream, pip->i_ino);
331	if (mru) {
332		ag = container_of(mru, struct xfs_fstrm_item, mru)->ag;
333		xfs_mru_cache_done(mp->m_filestream);
334
335		trace_xfs_filestream_lookup(mp, ip->i_ino, ag);
336		goto out;
337	}
338
339	/*
340	 * Set the starting AG using the rotor for inode32, otherwise
341	 * use the directory inode's AG.
342	 */
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);
348	} else
349		startag = XFS_INO_TO_AGNO(mp, pip->i_ino);
350
351	if (xfs_filestream_pick_ag(pip, startag, &ag, 0, 0))
352		ag = NULLAGNUMBER;
353out:
354	IRELE(pip);
355	return ag;
356}
357
358/*
359 * Pick a new allocation group for the current file and its file stream.
360 *
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.
363 */
364int
365xfs_filestream_new_ag(
366	struct xfs_bmalloca	*ap,
367	xfs_agnumber_t		*agp)
368{
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;
373	int			flags = 0;
374	int			err = 0;
375	struct xfs_mru_cache_elem *mru;
376
377	*agp = NULLAGNUMBER;
378
379	pip = xfs_filestream_get_parent(ip);
380	if (!pip)
381		goto exit;
382
383	mru = xfs_mru_cache_remove(mp->m_filestream, pip->i_ino);
384	if (mru) {
385		struct xfs_fstrm_item *item =
386			container_of(mru, struct xfs_fstrm_item, mru);
387		startag = (item->ag + 1) % mp->m_sb.sb_agcount;
388	}
389
390	if (xfs_alloc_is_userdata(ap->datatype))
391		flags |= XFS_PICK_USERDATA;
392	if (ap->dfops->dop_low)
393		flags |= XFS_PICK_LOWSPACE;
394
395	err = xfs_filestream_pick_ag(pip, startag, agp, flags, minlen);
396
397	/*
398	 * Only free the item here so we skip over the old AG earlier.
399	 */
400	if (mru)
401		xfs_fstrm_free_func(mp, mru);
402
403	IRELE(pip);
404exit:
405	if (*agp == NULLAGNUMBER)
406		*agp = 0;
407	return err;
408}
409
410void
411xfs_filestream_deassociate(
412	struct xfs_inode	*ip)
413{
414	xfs_mru_cache_delete(ip->i_mount->m_filestream, ip->i_ino);
415}
416
417int
418xfs_filestream_mount(
419	xfs_mount_t	*mp)
420{
421	/*
422	 * The filestream timer tunable is currently fixed within the range of
423	 * one second to four minutes, with five seconds being the default.  The
424	 * group count is somewhat arbitrary, but it'd be nice to adhere to the
425	 * timer tunable to within about 10 percent.  This requires at least 10
426	 * groups.
427	 */
428	return xfs_mru_cache_create(&mp->m_filestream, mp,
429			xfs_fstrm_centisecs * 10, 10, xfs_fstrm_free_func);
430}
431
432void
433xfs_filestream_unmount(
434	xfs_mount_t	*mp)
435{
436	xfs_mru_cache_destroy(mp->m_filestream);
437}
v5.9
  1// SPDX-License-Identifier: GPL-2.0
  2/*
  3 * Copyright (c) 2006-2007 Silicon Graphics, Inc.
  4 * Copyright (c) 2014 Christoph Hellwig.
  5 * All Rights Reserved.
 
 
 
 
 
 
 
 
 
 
 
 
 
  6 */
  7#include "xfs.h"
  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_sb.h"
 13#include "xfs_mount.h"
 
 14#include "xfs_inode.h"
 15#include "xfs_bmap.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"
 21#include "xfs_filestream.h"
 22
 23struct xfs_fstrm_item {
 24	struct xfs_mru_cache_elem	mru;
 25	xfs_agnumber_t			ag; /* AG in use for this directory */
 26};
 27
 28enum xfs_fstrm_alloc {
 29	XFS_PICK_USERDATA = 1,
 30	XFS_PICK_LOWSPACE = 2,
 31};
 32
 33/*
 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. The mount
 37 * point's m_peraglock is used to protect these counters from per-ag array
 38 * re-allocation during a growfs operation.  When xfs_growfs_data_private() is
 39 * about to reallocate the array, it calls xfs_filestream_flush() with the
 40 * m_peraglock held in write mode.
 41 *
 42 * Since xfs_mru_cache_flush() guarantees that all the free functions for all
 43 * the cache elements have finished executing before it returns, it's safe for
 44 * the free functions to use the atomic counters without m_peraglock protection.
 45 * This allows the implementation of xfs_fstrm_free_func() to be agnostic about
 46 * whether it was called with the m_peraglock held in read mode, write mode or
 47 * not held at all.  The race condition this addresses is the following:
 48 *
 49 *  - The work queue scheduler fires and pulls a filestream directory cache
 50 *    element off the LRU end of the cache for deletion, then gets pre-empted.
 51 *  - A growfs operation grabs the m_peraglock in write mode, flushes all the
 52 *    remaining items from the cache and reallocates the mount point's per-ag
 53 *    array, resetting all the counters to zero.
 54 *  - The work queue thread resumes and calls the free function for the element
 55 *    it started cleaning up earlier.  In the process it decrements the
 56 *    filestreams counter for an AG that now has no references.
 57 *
 58 * With a shrinkfs feature, the above scenario could panic the system.
 59 *
 60 * All other uses of the following macros should be protected by either the
 61 * m_peraglock held in read mode, or the cache's internal locking exposed by the
 62 * interval between a call to xfs_mru_cache_lookup() and a call to
 63 * xfs_mru_cache_done().  In addition, the m_peraglock must be held in read mode
 64 * when new elements are added to the cache.
 65 *
 66 * Combined, these locking rules ensure that no associations will ever exist in
 67 * the cache that reference per-ag array elements that have since been
 68 * reallocated.
 69 */
 70int
 71xfs_filestream_peek_ag(
 72	xfs_mount_t	*mp,
 73	xfs_agnumber_t	agno)
 74{
 75	struct xfs_perag *pag;
 76	int		ret;
 77
 78	pag = xfs_perag_get(mp, agno);
 79	ret = atomic_read(&pag->pagf_fstrms);
 80	xfs_perag_put(pag);
 81	return ret;
 82}
 83
 84static int
 85xfs_filestream_get_ag(
 86	xfs_mount_t	*mp,
 87	xfs_agnumber_t	agno)
 88{
 89	struct xfs_perag *pag;
 90	int		ret;
 91
 92	pag = xfs_perag_get(mp, agno);
 93	ret = atomic_inc_return(&pag->pagf_fstrms);
 94	xfs_perag_put(pag);
 95	return ret;
 96}
 97
 98static void
 99xfs_filestream_put_ag(
100	xfs_mount_t	*mp,
101	xfs_agnumber_t	agno)
102{
103	struct xfs_perag *pag;
104
105	pag = xfs_perag_get(mp, agno);
106	atomic_dec(&pag->pagf_fstrms);
107	xfs_perag_put(pag);
108}
109
110static void
111xfs_fstrm_free_func(
112	void			*data,
113	struct xfs_mru_cache_elem *mru)
114{
115	struct xfs_mount	*mp = data;
116	struct xfs_fstrm_item	*item =
117		container_of(mru, struct xfs_fstrm_item, mru);
118
119	xfs_filestream_put_ag(mp, item->ag);
120	trace_xfs_filestream_free(mp, mru->key, item->ag);
121
122	kmem_free(item);
123}
124
125/*
126 * Scan the AGs starting at startag looking for an AG that isn't in use and has
127 * at least minlen blocks free.
128 */
129static int
130xfs_filestream_pick_ag(
131	struct xfs_inode	*ip,
132	xfs_agnumber_t		startag,
133	xfs_agnumber_t		*agp,
134	int			flags,
135	xfs_extlen_t		minlen)
136{
137	struct xfs_mount	*mp = ip->i_mount;
138	struct xfs_fstrm_item	*item;
139	struct xfs_perag	*pag;
140	xfs_extlen_t		longest, free = 0, minfree, maxfree = 0;
141	xfs_agnumber_t		ag, max_ag = NULLAGNUMBER;
142	int			err, trylock, nscan;
143
144	ASSERT(S_ISDIR(VFS_I(ip)->i_mode));
145
146	/* 2% of an AG's blocks must be free for it to be chosen. */
147	minfree = mp->m_sb.sb_agblocks / 50;
148
149	ag = startag;
150	*agp = NULLAGNUMBER;
151
152	/* For the first pass, don't sleep trying to init the per-AG. */
153	trylock = XFS_ALLOC_FLAG_TRYLOCK;
154
155	for (nscan = 0; 1; nscan++) {
156		trace_xfs_filestream_scan(mp, ip->i_ino, ag);
157
158		pag = xfs_perag_get(mp, ag);
159
160		if (!pag->pagf_init) {
161			err = xfs_alloc_pagf_init(mp, NULL, ag, trylock);
162			if (err) {
163				xfs_perag_put(pag);
164				if (err != -EAGAIN)
165					return err;
166				/* Couldn't lock the AGF, skip this AG. */
167				continue;
168			}
169		}
170
 
 
 
 
171		/* Keep track of the AG with the most free blocks. */
172		if (pag->pagf_freeblks > maxfree) {
173			maxfree = pag->pagf_freeblks;
174			max_ag = ag;
175		}
176
177		/*
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.
182		 */
183		if (xfs_filestream_get_ag(mp, ag) > 1) {
184			xfs_filestream_put_ag(mp, ag);
185			goto next_ag;
186		}
187
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))) {
195
196			/* Break out, retaining the reference on the AG. */
197			free = pag->pagf_freeblks;
198			xfs_perag_put(pag);
199			*agp = ag;
200			break;
201		}
202
203		/* Drop the reference on this AG, it's not usable. */
204		xfs_filestream_put_ag(mp, ag);
205next_ag:
206		xfs_perag_put(pag);
207		/* Move to the next AG, wrapping to AG 0 if necessary. */
208		if (++ag >= mp->m_sb.sb_agcount)
209			ag = 0;
210
211		/* If a full pass of the AGs hasn't been done yet, continue. */
212		if (ag != startag)
213			continue;
214
215		/* Allow sleeping in xfs_alloc_pagf_init() on the 2nd pass. */
216		if (trylock != 0) {
217			trylock = 0;
218			continue;
219		}
220
221		/* Finally, if lowspace wasn't set, set it for the 3rd pass. */
222		if (!(flags & XFS_PICK_LOWSPACE)) {
223			flags |= XFS_PICK_LOWSPACE;
224			continue;
225		}
226
227		/*
228		 * Take the AG with the most free space, regardless of whether
229		 * it's already in use by another filestream.
230		 */
231		if (max_ag != NULLAGNUMBER) {
232			xfs_filestream_get_ag(mp, max_ag);
233			free = maxfree;
234			*agp = max_ag;
235			break;
236		}
237
238		/* take AG 0 if none matched */
239		trace_xfs_filestream_pick(ip, *agp, free, nscan);
240		*agp = 0;
241		return 0;
242	}
243
244	trace_xfs_filestream_pick(ip, *agp, free, nscan);
245
246	if (*agp == NULLAGNUMBER)
247		return 0;
248
249	err = -ENOMEM;
250	item = kmem_alloc(sizeof(*item), KM_MAYFAIL);
251	if (!item)
252		goto out_put_ag;
253
254	item->ag = *agp;
255
256	err = xfs_mru_cache_insert(mp->m_filestream, ip->i_ino, &item->mru);
257	if (err) {
258		if (err == -EEXIST)
259			err = 0;
260		goto out_free_item;
261	}
262
263	return 0;
264
265out_free_item:
266	kmem_free(item);
267out_put_ag:
268	xfs_filestream_put_ag(mp, *agp);
269	return err;
270}
271
272static struct xfs_inode *
273xfs_filestream_get_parent(
274	struct xfs_inode	*ip)
275{
276	struct inode		*inode = VFS_I(ip), *dir = NULL;
277	struct dentry		*dentry, *parent;
278
279	dentry = d_find_alias(inode);
280	if (!dentry)
281		goto out;
282
283	parent = dget_parent(dentry);
284	if (!parent)
285		goto out_dput;
286
287	dir = igrab(d_inode(parent));
288	dput(parent);
289
290out_dput:
291	dput(dentry);
292out:
293	return dir ? XFS_I(dir) : NULL;
294}
295
296/*
297 * Find the right allocation group for a file, either by finding an
298 * existing file stream or creating a new one.
299 *
300 * Returns NULLAGNUMBER in case of an error.
301 */
302xfs_agnumber_t
303xfs_filestream_lookup_ag(
304	struct xfs_inode	*ip)
305{
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;
310
311	ASSERT(S_ISREG(VFS_I(ip)->i_mode));
312
313	pip = xfs_filestream_get_parent(ip);
314	if (!pip)
315		return NULLAGNUMBER;
316
317	mru = xfs_mru_cache_lookup(mp->m_filestream, pip->i_ino);
318	if (mru) {
319		ag = container_of(mru, struct xfs_fstrm_item, mru)->ag;
320		xfs_mru_cache_done(mp->m_filestream);
321
322		trace_xfs_filestream_lookup(mp, ip->i_ino, ag);
323		goto out;
324	}
325
326	/*
327	 * Set the starting AG using the rotor for inode32, otherwise
328	 * use the directory inode's AG.
329	 */
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);
335	} else
336		startag = XFS_INO_TO_AGNO(mp, pip->i_ino);
337
338	if (xfs_filestream_pick_ag(pip, startag, &ag, 0, 0))
339		ag = NULLAGNUMBER;
340out:
341	xfs_irele(pip);
342	return ag;
343}
344
345/*
346 * Pick a new allocation group for the current file and its file stream.
347 *
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.
350 */
351int
352xfs_filestream_new_ag(
353	struct xfs_bmalloca	*ap,
354	xfs_agnumber_t		*agp)
355{
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;
360	int			flags = 0;
361	int			err = 0;
362	struct xfs_mru_cache_elem *mru;
363
364	*agp = NULLAGNUMBER;
365
366	pip = xfs_filestream_get_parent(ip);
367	if (!pip)
368		goto exit;
369
370	mru = xfs_mru_cache_remove(mp->m_filestream, pip->i_ino);
371	if (mru) {
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;
375	}
376
377	if (ap->datatype & XFS_ALLOC_USERDATA)
378		flags |= XFS_PICK_USERDATA;
379	if (ap->tp->t_flags & XFS_TRANS_LOWMODE)
380		flags |= XFS_PICK_LOWSPACE;
381
382	err = xfs_filestream_pick_ag(pip, startag, agp, flags, minlen);
383
384	/*
385	 * Only free the item here so we skip over the old AG earlier.
386	 */
387	if (mru)
388		xfs_fstrm_free_func(mp, mru);
389
390	xfs_irele(pip);
391exit:
392	if (*agp == NULLAGNUMBER)
393		*agp = 0;
394	return err;
395}
396
397void
398xfs_filestream_deassociate(
399	struct xfs_inode	*ip)
400{
401	xfs_mru_cache_delete(ip->i_mount->m_filestream, ip->i_ino);
402}
403
404int
405xfs_filestream_mount(
406	xfs_mount_t	*mp)
407{
408	/*
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
413	 * groups.
414	 */
415	return xfs_mru_cache_create(&mp->m_filestream, mp,
416			xfs_fstrm_centisecs * 10, 10, xfs_fstrm_free_func);
417}
418
419void
420xfs_filestream_unmount(
421	xfs_mount_t	*mp)
422{
423	xfs_mru_cache_destroy(mp->m_filestream);
424}