1// SPDX-License-Identifier: GPL-2.0-only
  2/*
  3 * Copyright (C) 2018 HUAWEI, Inc.
  4 *             https://www.huawei.com/
  5 */
  6#include "internal.h"
  7#include <linux/pagevec.h>
  8
  9struct page *erofs_allocpage(struct list_head *pool, gfp_t gfp)
 10{
 11	struct page *page;
 12
 13	if (!list_empty(pool)) {
 14		page = lru_to_page(pool);
 15		DBG_BUGON(page_ref_count(page) != 1);
 16		list_del(&page->lru);
 17	} else {
 18		page = alloc_page(gfp);
 19	}
 20	return page;
 21}
 22
 23#ifdef CONFIG_EROFS_FS_ZIP
 24/* global shrink count (for all mounted EROFS instances) */
 25static atomic_long_t erofs_global_shrink_cnt;
 26
 27static int erofs_workgroup_get(struct erofs_workgroup *grp)
 28{
 29	int o;
 30
 31repeat:
 32	o = erofs_wait_on_workgroup_freezed(grp);
 33	if (o <= 0)
 34		return -1;
 35
 36	if (atomic_cmpxchg(&grp->refcount, o, o + 1) != o)
 37		goto repeat;
 38
 39	/* decrease refcount paired by erofs_workgroup_put */
 40	if (o == 1)
 41		atomic_long_dec(&erofs_global_shrink_cnt);
 42	return 0;
 43}
 44
 45struct erofs_workgroup *erofs_find_workgroup(struct super_block *sb,
 46					     pgoff_t index)
 47{
 48	struct erofs_sb_info *sbi = EROFS_SB(sb);
 49	struct erofs_workgroup *grp;
 50
 51repeat:
 52	rcu_read_lock();
 53	grp = xa_load(&sbi->managed_pslots, index);
 54	if (grp) {
 55		if (erofs_workgroup_get(grp)) {
 56			/* prefer to relax rcu read side */
 57			rcu_read_unlock();
 58			goto repeat;
 59		}
 60
 61		DBG_BUGON(index != grp->index);
 62	}
 63	rcu_read_unlock();
 64	return grp;
 65}
 66
 67struct erofs_workgroup *erofs_insert_workgroup(struct super_block *sb,
 68					       struct erofs_workgroup *grp)
 69{
 70	struct erofs_sb_info *const sbi = EROFS_SB(sb);
 71	struct erofs_workgroup *pre;
 72
 73	/*
 74	 * Bump up a reference count before making this visible
 75	 * to others for the XArray in order to avoid potential
 76	 * UAF without serialized by xa_lock.
 77	 */
 78	atomic_inc(&grp->refcount);
 79
 80repeat:
 81	xa_lock(&sbi->managed_pslots);
 82	pre = __xa_cmpxchg(&sbi->managed_pslots, grp->index,
 83			   NULL, grp, GFP_NOFS);
 84	if (pre) {
 85		if (xa_is_err(pre)) {
 86			pre = ERR_PTR(xa_err(pre));
 87		} else if (erofs_workgroup_get(pre)) {
 88			/* try to legitimize the current in-tree one */
 89			xa_unlock(&sbi->managed_pslots);
 90			cond_resched();
 91			goto repeat;
 92		}
 93		atomic_dec(&grp->refcount);
 94		grp = pre;
 95	}
 96	xa_unlock(&sbi->managed_pslots);
 97	return grp;
 98}
 99
100static void  __erofs_workgroup_free(struct erofs_workgroup *grp)
101{
102	atomic_long_dec(&erofs_global_shrink_cnt);
103	erofs_workgroup_free_rcu(grp);
104}
105
106int erofs_workgroup_put(struct erofs_workgroup *grp)
107{
108	int count = atomic_dec_return(&grp->refcount);
109
110	if (count == 1)
111		atomic_long_inc(&erofs_global_shrink_cnt);
112	else if (!count)
113		__erofs_workgroup_free(grp);
114	return count;
115}
116
117static bool erofs_try_to_release_workgroup(struct erofs_sb_info *sbi,
118					   struct erofs_workgroup *grp)
119{
120	/*
121	 * If managed cache is on, refcount of workgroups
122	 * themselves could be < 0 (freezed). In other words,
123	 * there is no guarantee that all refcounts > 0.
124	 */
125	if (!erofs_workgroup_try_to_freeze(grp, 1))
126		return false;
127
128	/*
129	 * Note that all cached pages should be unattached
130	 * before deleted from the XArray. Otherwise some
131	 * cached pages could be still attached to the orphan
132	 * old workgroup when the new one is available in the tree.
133	 */
134	if (erofs_try_to_free_all_cached_pages(sbi, grp)) {
135		erofs_workgroup_unfreeze(grp, 1);
136		return false;
137	}
138
139	/*
140	 * It's impossible to fail after the workgroup is freezed,
141	 * however in order to avoid some race conditions, add a
142	 * DBG_BUGON to observe this in advance.
143	 */
144	DBG_BUGON(xa_erase(&sbi->managed_pslots, grp->index) != grp);
145
146	/* last refcount should be connected with its managed pslot.  */
147	erofs_workgroup_unfreeze(grp, 0);
148	__erofs_workgroup_free(grp);
149	return true;
150}
151
152static unsigned long erofs_shrink_workstation(struct erofs_sb_info *sbi,
153					      unsigned long nr_shrink)
154{
155	struct erofs_workgroup *grp;
156	unsigned int freed = 0;
157	unsigned long index;
158
159	xa_for_each(&sbi->managed_pslots, index, grp) {
160		/* try to shrink each valid workgroup */
161		if (!erofs_try_to_release_workgroup(sbi, grp))
162			continue;
163
164		++freed;
165		if (!--nr_shrink)
166			break;
167	}
168	return freed;
169}
170
171/* protected by 'erofs_sb_list_lock' */
172static unsigned int shrinker_run_no;
173
174/* protects the mounted 'erofs_sb_list' */
175static DEFINE_SPINLOCK(erofs_sb_list_lock);
176static LIST_HEAD(erofs_sb_list);
177
178void erofs_shrinker_register(struct super_block *sb)
179{
180	struct erofs_sb_info *sbi = EROFS_SB(sb);
181
182	mutex_init(&sbi->umount_mutex);
183
184	spin_lock(&erofs_sb_list_lock);
185	list_add(&sbi->list, &erofs_sb_list);
186	spin_unlock(&erofs_sb_list_lock);
187}
188
189void erofs_shrinker_unregister(struct super_block *sb)
190{
191	struct erofs_sb_info *const sbi = EROFS_SB(sb);
192
193	mutex_lock(&sbi->umount_mutex);
194	/* clean up all remaining workgroups in memory */
195	erofs_shrink_workstation(sbi, ~0UL);
196
197	spin_lock(&erofs_sb_list_lock);
198	list_del(&sbi->list);
199	spin_unlock(&erofs_sb_list_lock);
200	mutex_unlock(&sbi->umount_mutex);
201}
202
203static unsigned long erofs_shrink_count(struct shrinker *shrink,
204					struct shrink_control *sc)
205{
206	return atomic_long_read(&erofs_global_shrink_cnt);
207}
208
209static unsigned long erofs_shrink_scan(struct shrinker *shrink,
210				       struct shrink_control *sc)
211{
212	struct erofs_sb_info *sbi;
213	struct list_head *p;
214
215	unsigned long nr = sc->nr_to_scan;
216	unsigned int run_no;
217	unsigned long freed = 0;
218
219	spin_lock(&erofs_sb_list_lock);
220	do {
221		run_no = ++shrinker_run_no;
222	} while (run_no == 0);
223
224	/* Iterate over all mounted superblocks and try to shrink them */
225	p = erofs_sb_list.next;
226	while (p != &erofs_sb_list) {
227		sbi = list_entry(p, struct erofs_sb_info, list);
228
229		/*
230		 * We move the ones we do to the end of the list, so we stop
231		 * when we see one we have already done.
232		 */
233		if (sbi->shrinker_run_no == run_no)
234			break;
235
236		if (!mutex_trylock(&sbi->umount_mutex)) {
237			p = p->next;
238			continue;
239		}
240
241		spin_unlock(&erofs_sb_list_lock);
242		sbi->shrinker_run_no = run_no;
243
244		freed += erofs_shrink_workstation(sbi, nr - freed);
245
246		spin_lock(&erofs_sb_list_lock);
247		/* Get the next list element before we move this one */
248		p = p->next;
249
250		/*
251		 * Move this one to the end of the list to provide some
252		 * fairness.
253		 */
254		list_move_tail(&sbi->list, &erofs_sb_list);
255		mutex_unlock(&sbi->umount_mutex);
256
257		if (freed >= nr)
258			break;
259	}
260	spin_unlock(&erofs_sb_list_lock);
261	return freed;
262}
263
264static struct shrinker erofs_shrinker_info = {
265	.scan_objects = erofs_shrink_scan,
266	.count_objects = erofs_shrink_count,
267	.seeks = DEFAULT_SEEKS,
268};
269
270int __init erofs_init_shrinker(void)
271{
272	return register_shrinker(&erofs_shrinker_info);
273}
274
275void erofs_exit_shrinker(void)
276{
277	unregister_shrinker(&erofs_shrinker_info);
278}
279#endif	/* !CONFIG_EROFS_FS_ZIP */