drivers/staging/erofs/utils.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * linux/drivers/staging/erofs/utils.c
   4  *
   5  * Copyright (C) 2018 HUAWEI, Inc.
   6  *             http://www.huawei.com/
   7  * Created by Gao Xiang <[email protected]>
   8  *
   9  * This file is subject to the terms and conditions of the GNU General Public
  10  * License.  See the file COPYING in the main directory of the Linux
  11  * distribution for more details.
  12  */
  13
  14 #include "internal.h"
  15 #include <linux/pagevec.h>
  16
  17 struct page *erofs_allocpage(struct list_head *pool, gfp_t gfp)
  18 {
  19         struct page *page;
  20
  21         if (!list_empty(pool)) {
  22                 page = lru_to_page(pool);
  23                 list_del(&page->lru);
  24         } else {
  25                 page = alloc_pages(gfp | __GFP_NOFAIL, 0);
  26         }
  27         return page;
  28 }
  29
  30 /* global shrink count (for all mounted EROFS instances) */
  31 static atomic_long_t erofs_global_shrink_cnt;
  32
  33 #ifdef CONFIG_EROFS_FS_ZIP
  34 #define __erofs_workgroup_get(grp)      atomic_inc(&(grp)->refcount)
  35 #define __erofs_workgroup_put(grp)      atomic_dec(&(grp)->refcount)
  36
  37 static int erofs_workgroup_get(struct erofs_workgroup *grp)
  38 {
  39         int o;
  40
  41 repeat:
  42         o = erofs_wait_on_workgroup_freezed(grp);
  43         if (unlikely(o <= 0))
  44                 return -1;
  45
  46         if (unlikely(atomic_cmpxchg(&grp->refcount, o, o + 1) != o))
  47                 goto repeat;
  48
  49         /* decrease refcount paired by erofs_workgroup_put */
  50         if (unlikely(o == 1))
  51                 atomic_long_dec(&erofs_global_shrink_cnt);
  52         return 0;
  53 }
  54
  55 struct erofs_workgroup *erofs_find_workgroup(struct super_block *sb,
  56                                              pgoff_t index, bool *tag)
  57 {
  58         struct erofs_sb_info *sbi = EROFS_SB(sb);
  59         struct erofs_workgroup *grp;
  60
  61 repeat:
  62         rcu_read_lock();
  63         grp = radix_tree_lookup(&sbi->workstn_tree, index);
  64         if (grp != NULL) {
  65                 *tag = xa_pointer_tag(grp);
  66                 grp = xa_untag_pointer(grp);
  67
  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
  80 int erofs_register_workgroup(struct super_block *sb,
  81                              struct erofs_workgroup *grp,
  82                              bool tag)
  83 {
  84         struct erofs_sb_info *sbi;
  85         int err;
  86
  87         /* grp shouldn't be broken or used before */
  88         if (unlikely(atomic_read(&grp->refcount) != 1)) {
  89                 DBG_BUGON(1);
  90                 return -EINVAL;
  91         }
  92
  93         err = radix_tree_preload(GFP_NOFS);
  94         if (err)
  95                 return err;
  96
  97         sbi = EROFS_SB(sb);
  98         erofs_workstn_lock(sbi);
  99
 100         grp = xa_tag_pointer(grp, tag);
 101
 102         /*
 103          * Bump up reference count before making this workgroup
 104          * visible to other users in order to avoid potential UAF
 105          * without serialized by erofs_workstn_lock.
 106          */
 107         __erofs_workgroup_get(grp);
 108
 109         err = radix_tree_insert(&sbi->workstn_tree,
 110                                 grp->index, grp);
 111         if (unlikely(err))
 112                 /*
 113                  * it's safe to decrease since the workgroup isn't visible
 114                  * and refcount >= 2 (cannot be freezed).
 115                  */
 116                 __erofs_workgroup_put(grp);
 117
 118         erofs_workstn_unlock(sbi);
 119         radix_tree_preload_end();
 120         return err;
 121 }
 122
 123 static void  __erofs_workgroup_free(struct erofs_workgroup *grp)
 124 {
 125         atomic_long_dec(&erofs_global_shrink_cnt);
 126         erofs_workgroup_free_rcu(grp);
 127 }
 128
 129 int erofs_workgroup_put(struct erofs_workgroup *grp)
 130 {
 131         int count = atomic_dec_return(&grp->refcount);
 132
 133         if (count == 1)
 134                 atomic_long_inc(&erofs_global_shrink_cnt);
 135         else if (!count)
 136                 __erofs_workgroup_free(grp);
 137         return count;
 138 }
 139
 140 #ifdef EROFS_FS_HAS_MANAGED_CACHE
 141 /* for cache-managed case, customized reclaim paths exist */
 142 static void erofs_workgroup_unfreeze_final(struct erofs_workgroup *grp)
 143 {
 144         erofs_workgroup_unfreeze(grp, 0);
 145         __erofs_workgroup_free(grp);
 146 }
 147
 148 static bool erofs_try_to_release_workgroup(struct erofs_sb_info *sbi,
 149                                            struct erofs_workgroup *grp,
 150                                            bool cleanup)
 151 {
 152         /*
 153          * for managed cache enabled, the refcount of workgroups
 154          * themselves could be < 0 (freezed). So there is no guarantee
 155          * that all refcount > 0 if managed cache is enabled.
 156          */
 157         if (!erofs_workgroup_try_to_freeze(grp, 1))
 158                 return false;
 159
 160         /*
 161          * note that all cached pages should be unlinked
 162          * before delete it from the radix tree.
 163          * Otherwise some cached pages of an orphan old workgroup
 164          * could be still linked after the new one is available.
 165          */
 166         if (erofs_try_to_free_all_cached_pages(sbi, grp)) {
 167                 erofs_workgroup_unfreeze(grp, 1);
 168                 return false;
 169         }
 170
 171         /*
 172          * it is impossible to fail after the workgroup is freezed,
 173          * however in order to avoid some race conditions, add a
 174          * DBG_BUGON to observe this in advance.
 175          */
 176         DBG_BUGON(xa_untag_pointer(radix_tree_delete(&sbi->workstn_tree,
 177                                                      grp->index)) != grp);
 178
 179         /*
 180          * if managed cache is enable, the last refcount
 181          * should indicate the related workstation.
 182          */
 183         erofs_workgroup_unfreeze_final(grp);
 184         return true;
 185 }
 186
 187 #else
 188 /* for nocache case, no customized reclaim path at all */
 189 static bool erofs_try_to_release_workgroup(struct erofs_sb_info *sbi,
 190                                            struct erofs_workgroup *grp,
 191                                            bool cleanup)
 192 {
 193         int cnt = atomic_read(&grp->refcount);
 194
 195         DBG_BUGON(cnt <= 0);
 196         DBG_BUGON(cleanup && cnt != 1);
 197
 198         if (cnt > 1)
 199                 return false;
 200
 201         DBG_BUGON(xa_untag_pointer(radix_tree_delete(&sbi->workstn_tree,
 202                                                      grp->index)) != grp);
 203
 204         /* (rarely) could be grabbed again when freeing */
 205         erofs_workgroup_put(grp);
 206         return true;
 207 }
 208
 209 #endif
 210
 211 unsigned long erofs_shrink_workstation(struct erofs_sb_info *sbi,
 212                                        unsigned long nr_shrink,
 213                                        bool cleanup)
 214 {
 215         pgoff_t first_index = 0;
 216         void *batch[PAGEVEC_SIZE];
 217         unsigned int freed = 0;
 218
 219         int i, found;
 220 repeat:
 221         erofs_workstn_lock(sbi);
 222
 223         found = radix_tree_gang_lookup(&sbi->workstn_tree,
 224                 batch, first_index, PAGEVEC_SIZE);
 225
 226         for (i = 0; i < found; ++i) {
 227                 struct erofs_workgroup *grp = xa_untag_pointer(batch[i]);
 228
 229                 first_index = grp->index + 1;
 230
 231                 /* try to shrink each valid workgroup */
 232                 if (!erofs_try_to_release_workgroup(sbi, grp, cleanup))
 233                         continue;
 234
 235                 ++freed;
 236                 if (unlikely(!--nr_shrink))
 237                         break;
 238         }
 239         erofs_workstn_unlock(sbi);
 240
 241         if (i && nr_shrink)
 242                 goto repeat;
 243         return freed;
 244 }
 245
 246 #endif
 247
 248 /* protected by 'erofs_sb_list_lock' */
 249 static unsigned int shrinker_run_no;
 250
 251 /* protects the mounted 'erofs_sb_list' */
 252 static DEFINE_SPINLOCK(erofs_sb_list_lock);
 253 static LIST_HEAD(erofs_sb_list);
 254
 255 void erofs_register_super(struct super_block *sb)
 256 {
 257         struct erofs_sb_info *sbi = EROFS_SB(sb);
 258
 259         mutex_init(&sbi->umount_mutex);
 260
 261         spin_lock(&erofs_sb_list_lock);
 262         list_add(&sbi->list, &erofs_sb_list);
 263         spin_unlock(&erofs_sb_list_lock);
 264 }
 265
 266 void erofs_unregister_super(struct super_block *sb)
 267 {
 268         spin_lock(&erofs_sb_list_lock);
 269         list_del(&EROFS_SB(sb)->list);
 270         spin_unlock(&erofs_sb_list_lock);
 271 }
 272
 273 static unsigned long erofs_shrink_count(struct shrinker *shrink,
 274                                         struct shrink_control *sc)
 275 {
 276         return atomic_long_read(&erofs_global_shrink_cnt);
 277 }
 278
 279 static unsigned long erofs_shrink_scan(struct shrinker *shrink,
 280                                        struct shrink_control *sc)
 281 {
 282         struct erofs_sb_info *sbi;
 283         struct list_head *p;
 284
 285         unsigned long nr = sc->nr_to_scan;
 286         unsigned int run_no;
 287         unsigned long freed = 0;
 288
 289         spin_lock(&erofs_sb_list_lock);
 290         do
 291                 run_no = ++shrinker_run_no;
 292         while (run_no == 0);
 293
 294         /* Iterate over all mounted superblocks and try to shrink them */
 295         p = erofs_sb_list.next;
 296         while (p != &erofs_sb_list) {
 297                 sbi = list_entry(p, struct erofs_sb_info, list);
 298
 299                 /*
 300                  * We move the ones we do to the end of the list, so we stop
 301                  * when we see one we have already done.
 302                  */
 303                 if (sbi->shrinker_run_no == run_no)
 304                         break;
 305
 306                 if (!mutex_trylock(&sbi->umount_mutex)) {
 307                         p = p->next;
 308                         continue;
 309                 }
 310
 311                 spin_unlock(&erofs_sb_list_lock);
 312                 sbi->shrinker_run_no = run_no;
 313
 314 #ifdef CONFIG_EROFS_FS_ZIP
 315                 freed += erofs_shrink_workstation(sbi, nr, false);
 316 #endif
 317
 318                 spin_lock(&erofs_sb_list_lock);
 319                 /* Get the next list element before we move this one */
 320                 p = p->next;
 321
 322                 /*
 323                  * Move this one to the end of the list to provide some
 324                  * fairness.
 325                  */
 326                 list_move_tail(&sbi->list, &erofs_sb_list);
 327                 mutex_unlock(&sbi->umount_mutex);
 328
 329                 if (freed >= nr)
 330                         break;
 331         }
 332         spin_unlock(&erofs_sb_list_lock);
 333         return freed;
 334 }
 335
 336 struct shrinker erofs_shrinker_info = {
 337         .scan_objects = erofs_shrink_scan,
 338         .count_objects = erofs_shrink_count,
 339         .seeks = DEFAULT_SEEKS,
 340 };
 341