1// SPDX-License-Identifier: GPL-2.0-only
  2/*
  3 * Copyright (C) 2018 HUAWEI, Inc.
  4 *             https://www.huawei.com/
  5 * Created by Gao Xiang <gaoxiang25@huawei.com>
  6 */
  7#include "internal.h"
  8#include <linux/pagevec.h>
  9
 10struct page *erofs_allocpage(struct list_head *pool, gfp_t gfp)
 11{
 12	struct page *page;
 13
 14	if (!list_empty(pool)) {
 15		page = lru_to_page(pool);
 16		DBG_BUGON(page_ref_count(page) != 1);
 17		list_del(&page->lru);
 18	} else {
 19		page = alloc_page(gfp);
 20	}
 21	return page;
 22}
 23
 24#if (EROFS_PCPUBUF_NR_PAGES > 0)
 25static struct {
 26	u8 data[PAGE_SIZE * EROFS_PCPUBUF_NR_PAGES];
 27} ____cacheline_aligned_in_smp erofs_pcpubuf[NR_CPUS];
 28
 29void *erofs_get_pcpubuf(unsigned int pagenr)
 30{
 31	preempt_disable();
 32	return &erofs_pcpubuf[smp_processor_id()].data[pagenr * PAGE_SIZE];
 33}
 34#endif
 35
 36#ifdef CONFIG_EROFS_FS_ZIP
 37/* global shrink count (for all mounted EROFS instances) */
 38static atomic_long_t erofs_global_shrink_cnt;
 39
 40static int erofs_workgroup_get(struct erofs_workgroup *grp)
 41{
 42	int o;
 43
 44repeat:
 45	o = erofs_wait_on_workgroup_freezed(grp);
 46	if (o <= 0)
 47		return -1;
 48
 49	if (atomic_cmpxchg(&grp->refcount, o, o + 1) != o)
 50		goto repeat;
 51
 52	/* decrease refcount paired by erofs_workgroup_put */
 53	if (o == 1)
 54		atomic_long_dec(&erofs_global_shrink_cnt);
 55	return 0;
 56}
 57
 58struct erofs_workgroup *erofs_find_workgroup(struct super_block *sb,
 59					     pgoff_t index)
 60{
 61	struct erofs_sb_info *sbi = EROFS_SB(sb);
 62	struct erofs_workgroup *grp;
 63
 64repeat:
 65	rcu_read_lock();
 66	grp = xa_load(&sbi->managed_pslots, index);
 67	if (grp) {
 68		if (erofs_workgroup_get(grp)) {
 69			/* prefer to relax rcu read side */
 70			rcu_read_unlock();
 71			goto repeat;
 72		}
 73
 74		DBG_BUGON(index != grp->index);
 75	}
 76	rcu_read_unlock();
 77	return grp;
 78}
 79
 80struct erofs_workgroup *erofs_insert_workgroup(struct super_block *sb,
 81					       struct erofs_workgroup *grp)
 82{
 83	struct erofs_sb_info *const sbi = EROFS_SB(sb);
 84	struct erofs_workgroup *pre;
 85
 86	/*
 87	 * Bump up a reference count before making this visible
 88	 * to others for the XArray in order to avoid potential
 89	 * UAF without serialized by xa_lock.
 90	 */
 91	atomic_inc(&grp->refcount);
 92
 93repeat:
 94	xa_lock(&sbi->managed_pslots);
 95	pre = __xa_cmpxchg(&sbi->managed_pslots, grp->index,
 96			   NULL, grp, GFP_NOFS);
 97	if (pre) {
 98		if (xa_is_err(pre)) {
 99			pre = ERR_PTR(xa_err(pre));
100		} else if (erofs_workgroup_get(pre)) {
101			/* try to legitimize the current in-tree one */
102			xa_unlock(&sbi->managed_pslots);
103			cond_resched();
104			goto repeat;
105		}
106		atomic_dec(&grp->refcount);
107		grp = pre;
108	}
109	xa_unlock(&sbi->managed_pslots);
110	return grp;
111}
112
113static void  __erofs_workgroup_free(struct erofs_workgroup *grp)
114{
115	atomic_long_dec(&erofs_global_shrink_cnt);
116	erofs_workgroup_free_rcu(grp);
117}
118
119int erofs_workgroup_put(struct erofs_workgroup *grp)
120{
121	int count = atomic_dec_return(&grp->refcount);
122
123	if (count == 1)
124		atomic_long_inc(&erofs_global_shrink_cnt);
125	else if (!count)
126		__erofs_workgroup_free(grp);
127	return count;
128}
129
130static bool erofs_try_to_release_workgroup(struct erofs_sb_info *sbi,
131					   struct erofs_workgroup *grp)
132{
133	/*
134	 * If managed cache is on, refcount of workgroups
135	 * themselves could be < 0 (freezed). In other words,
136	 * there is no guarantee that all refcounts > 0.
137	 */
138	if (!erofs_workgroup_try_to_freeze(grp, 1))
139		return false;
140
141	/*
142	 * Note that all cached pages should be unattached
143	 * before deleted from the XArray. Otherwise some
144	 * cached pages could be still attached to the orphan
145	 * old workgroup when the new one is available in the tree.
146	 */
147	if (erofs_try_to_free_all_cached_pages(sbi, grp)) {
148		erofs_workgroup_unfreeze(grp, 1);
149		return false;
150	}
151
152	/*
153	 * It's impossible to fail after the workgroup is freezed,
154	 * however in order to avoid some race conditions, add a
155	 * DBG_BUGON to observe this in advance.
156	 */
157	DBG_BUGON(xa_erase(&sbi->managed_pslots, grp->index) != grp);
158
159	/* last refcount should be connected with its managed pslot.  */
160	erofs_workgroup_unfreeze(grp, 0);
161	__erofs_workgroup_free(grp);
162	return true;
163}
164
165static unsigned long erofs_shrink_workstation(struct erofs_sb_info *sbi,
166					      unsigned long nr_shrink)
167{
168	struct erofs_workgroup *grp;
169	unsigned int freed = 0;
170	unsigned long index;
171
172	xa_for_each(&sbi->managed_pslots, index, grp) {
173		/* try to shrink each valid workgroup */
174		if (!erofs_try_to_release_workgroup(sbi, grp))
175			continue;
176
177		++freed;
178		if (!--nr_shrink)
179			break;
180	}
181	return freed;
182}
183
184/* protected by 'erofs_sb_list_lock' */
185static unsigned int shrinker_run_no;
186
187/* protects the mounted 'erofs_sb_list' */
188static DEFINE_SPINLOCK(erofs_sb_list_lock);
189static LIST_HEAD(erofs_sb_list);
190
191void erofs_shrinker_register(struct super_block *sb)
192{
193	struct erofs_sb_info *sbi = EROFS_SB(sb);
194
195	mutex_init(&sbi->umount_mutex);
196
197	spin_lock(&erofs_sb_list_lock);
198	list_add(&sbi->list, &erofs_sb_list);
199	spin_unlock(&erofs_sb_list_lock);
200}
201
202void erofs_shrinker_unregister(struct super_block *sb)
203{
204	struct erofs_sb_info *const sbi = EROFS_SB(sb);
205
206	mutex_lock(&sbi->umount_mutex);
207	/* clean up all remaining workgroups in memory */
208	erofs_shrink_workstation(sbi, ~0UL);
209
210	spin_lock(&erofs_sb_list_lock);
211	list_del(&sbi->list);
212	spin_unlock(&erofs_sb_list_lock);
213	mutex_unlock(&sbi->umount_mutex);
214}
215
216static unsigned long erofs_shrink_count(struct shrinker *shrink,
217					struct shrink_control *sc)
218{
219	return atomic_long_read(&erofs_global_shrink_cnt);
220}
221
222static unsigned long erofs_shrink_scan(struct shrinker *shrink,
223				       struct shrink_control *sc)
224{
225	struct erofs_sb_info *sbi;
226	struct list_head *p;
227
228	unsigned long nr = sc->nr_to_scan;
229	unsigned int run_no;
230	unsigned long freed = 0;
231
232	spin_lock(&erofs_sb_list_lock);
233	do {
234		run_no = ++shrinker_run_no;
235	} while (run_no == 0);
236
237	/* Iterate over all mounted superblocks and try to shrink them */
238	p = erofs_sb_list.next;
239	while (p != &erofs_sb_list) {
240		sbi = list_entry(p, struct erofs_sb_info, list);
241
242		/*
243		 * We move the ones we do to the end of the list, so we stop
244		 * when we see one we have already done.
245		 */
246		if (sbi->shrinker_run_no == run_no)
247			break;
248
249		if (!mutex_trylock(&sbi->umount_mutex)) {
250			p = p->next;
251			continue;
252		}
253
254		spin_unlock(&erofs_sb_list_lock);
255		sbi->shrinker_run_no = run_no;
256
257		freed += erofs_shrink_workstation(sbi, nr - freed);
258
259		spin_lock(&erofs_sb_list_lock);
260		/* Get the next list element before we move this one */
261		p = p->next;
262
263		/*
264		 * Move this one to the end of the list to provide some
265		 * fairness.
266		 */
267		list_move_tail(&sbi->list, &erofs_sb_list);
268		mutex_unlock(&sbi->umount_mutex);
269
270		if (freed >= nr)
271			break;
272	}
273	spin_unlock(&erofs_sb_list_lock);
274	return freed;
275}
276
277static struct shrinker erofs_shrinker_info = {
278	.scan_objects = erofs_shrink_scan,
279	.count_objects = erofs_shrink_count,
280	.seeks = DEFAULT_SEEKS,
281};
282
283int __init erofs_init_shrinker(void)
284{
285	return register_shrinker(&erofs_shrinker_info);
286}
287
288void erofs_exit_shrinker(void)
289{
290	unregister_shrinker(&erofs_shrinker_info);
291}
292#endif	/* !CONFIG_EROFS_FS_ZIP */
293