fs/erofs/utils.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Copyright (C) 2018 HUAWEI, Inc.
   4  *             http://www.huawei.com/
   5  * Created by Gao Xiang <gaoxiang25@huawei.com>
   6  */
   7 #include "internal.h"
   8 #include <linux/pagevec.h>
   9
  10 struct 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)
  25 static struct {
  26         u8 data[PAGE_SIZE * EROFS_PCPUBUF_NR_PAGES];
  27 } ____cacheline_aligned_in_smp erofs_pcpubuf[NR_CPUS];
  28
  29 void *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) */
  38 static atomic_long_t erofs_global_shrink_cnt;
  39
  40 #define __erofs_workgroup_get(grp)      atomic_inc(&(grp)->refcount)
  41 #define __erofs_workgroup_put(grp)      atomic_dec(&(grp)->refcount)
  42
  43 static int erofs_workgroup_get(struct erofs_workgroup *grp)
  44 {
  45         int o;
  46
  47 repeat:
  48         o = erofs_wait_on_workgroup_freezed(grp);
  49         if (o <= 0)
  50                 return -1;
  51
  52         if (atomic_cmpxchg(&grp->refcount, o, o + 1) != o)
  53                 goto repeat;
  54
  55         /* decrease refcount paired by erofs_workgroup_put */
  56         if (o == 1)
  57                 atomic_long_dec(&erofs_global_shrink_cnt);
  58         return 0;
  59 }
  60
  61 struct erofs_workgroup *erofs_find_workgroup(struct super_block *sb,
  62                                              pgoff_t index, bool *tag)
  63 {
  64         struct erofs_sb_info *sbi = EROFS_SB(sb);
  65         struct erofs_workgroup *grp;
  66
  67 repeat:
  68         rcu_read_lock();
  69         grp = radix_tree_lookup(&sbi->workstn_tree, index);
  70         if (grp) {
  71                 *tag = xa_pointer_tag(grp);
  72                 grp = xa_untag_pointer(grp);
  73
  74                 if (erofs_workgroup_get(grp)) {
  75                         /* prefer to relax rcu read side */
  76                         rcu_read_unlock();
  77                         goto repeat;
  78                 }
  79
  80                 DBG_BUGON(index != grp->index);
  81         }
  82         rcu_read_unlock();
  83         return grp;
  84 }
  85
  86 int erofs_register_workgroup(struct super_block *sb,
  87                              struct erofs_workgroup *grp,
  88                              bool tag)
  89 {
  90         struct erofs_sb_info *sbi;
  91         int err;
  92
  93         /* grp shouldn't be broken or used before */
  94         if (atomic_read(&grp->refcount) != 1) {
  95                 DBG_BUGON(1);
  96                 return -EINVAL;
  97         }
  98
  99         err = radix_tree_preload(GFP_NOFS);
 100         if (err)
 101                 return err;
 102
 103         sbi = EROFS_SB(sb);
 104         xa_lock(&sbi->workstn_tree);
 105
 106         grp = xa_tag_pointer(grp, tag);
 107
 108         /*
 109          * Bump up reference count before making this workgroup
 110          * visible to other users in order to avoid potential UAF
 111          * without serialized by workstn_lock.
 112          */
 113         __erofs_workgroup_get(grp);
 114
 115         err = radix_tree_insert(&sbi->workstn_tree, grp->index, grp);
 116         if (err)
 117                 /*
 118                  * it's safe to decrease since the workgroup isn't visible
 119                  * and refcount >= 2 (cannot be freezed).
 120                  */
 121                 __erofs_workgroup_put(grp);
 122
 123         xa_unlock(&sbi->workstn_tree);
 124         radix_tree_preload_end();
 125         return err;
 126 }
 127
 128 static void  __erofs_workgroup_free(struct erofs_workgroup *grp)
 129 {
 130         atomic_long_dec(&erofs_global_shrink_cnt);
 131         erofs_workgroup_free_rcu(grp);
 132 }
 133
 134 int erofs_workgroup_put(struct erofs_workgroup *grp)
 135 {
 136         int count = atomic_dec_return(&grp->refcount);
 137
 138         if (count == 1)
 139                 atomic_long_inc(&erofs_global_shrink_cnt);
 140         else if (!count)
 141                 __erofs_workgroup_free(grp);
 142         return count;
 143 }
 144
 145 static void erofs_workgroup_unfreeze_final(struct erofs_workgroup *grp)
 146 {
 147         erofs_workgroup_unfreeze(grp, 0);
 148         __erofs_workgroup_free(grp);
 149 }
 150
 151 static bool erofs_try_to_release_workgroup(struct erofs_sb_info *sbi,
 152                                            struct erofs_workgroup *grp)
 153 {
 154         /*
 155          * If managed cache is on, refcount of workgroups
 156          * themselves could be < 0 (freezed). In other words,
 157          * there is no guarantee that all refcounts > 0.
 158          */
 159         if (!erofs_workgroup_try_to_freeze(grp, 1))
 160                 return false;
 161
 162         /*
 163          * Note that all cached pages should be unattached
 164          * before deleted from the radix tree. Otherwise some
 165          * cached pages could be still attached to the orphan
 166          * old workgroup when the new one is available in the tree.
 167          */
 168         if (erofs_try_to_free_all_cached_pages(sbi, grp)) {
 169                 erofs_workgroup_unfreeze(grp, 1);
 170                 return false;
 171         }
 172
 173         /*
 174          * It's impossible to fail after the workgroup is freezed,
 175          * however in order to avoid some race conditions, add a
 176          * DBG_BUGON to observe this in advance.
 177          */
 178         DBG_BUGON(xa_untag_pointer(radix_tree_delete(&sbi->workstn_tree,
 179                                                      grp->index)) != grp);
 180
 181         /*
 182          * If managed cache is on, last refcount should indicate
 183          * the related workstation.
 184          */
 185         erofs_workgroup_unfreeze_final(grp);
 186         return true;
 187 }
 188
 189 static unsigned long erofs_shrink_workstation(struct erofs_sb_info *sbi,
 190                                               unsigned long nr_shrink)
 191 {
 192         pgoff_t first_index = 0;
 193         void *batch[PAGEVEC_SIZE];
 194         unsigned int freed = 0;
 195
 196         int i, found;
 197 repeat:
 198         xa_lock(&sbi->workstn_tree);
 199
 200         found = radix_tree_gang_lookup(&sbi->workstn_tree,
 201                                        batch, first_index, PAGEVEC_SIZE);
 202
 203         for (i = 0; i < found; ++i) {
 204                 struct erofs_workgroup *grp = xa_untag_pointer(batch[i]);
 205
 206                 first_index = grp->index + 1;
 207
 208                 /* try to shrink each valid workgroup */
 209                 if (!erofs_try_to_release_workgroup(sbi, grp))
 210                         continue;
 211
 212                 ++freed;
 213                 if (!--nr_shrink)
 214                         break;
 215         }
 216         xa_unlock(&sbi->workstn_tree);
 217
 218         if (i && nr_shrink)
 219                 goto repeat;
 220         return freed;
 221 }
 222
 223 /* protected by 'erofs_sb_list_lock' */
 224 static unsigned int shrinker_run_no;
 225
 226 /* protects the mounted 'erofs_sb_list' */
 227 static DEFINE_SPINLOCK(erofs_sb_list_lock);
 228 static LIST_HEAD(erofs_sb_list);
 229
 230 void erofs_shrinker_register(struct super_block *sb)
 231 {
 232         struct erofs_sb_info *sbi = EROFS_SB(sb);
 233
 234         mutex_init(&sbi->umount_mutex);
 235
 236         spin_lock(&erofs_sb_list_lock);
 237         list_add(&sbi->list, &erofs_sb_list);
 238         spin_unlock(&erofs_sb_list_lock);
 239 }
 240
 241 void erofs_shrinker_unregister(struct super_block *sb)
 242 {
 243         struct erofs_sb_info *const sbi = EROFS_SB(sb);
 244
 245         mutex_lock(&sbi->umount_mutex);
 246         /* clean up all remaining workgroups in memory */
 247         erofs_shrink_workstation(sbi, ~0UL);
 248
 249         spin_lock(&erofs_sb_list_lock);
 250         list_del(&sbi->list);
 251         spin_unlock(&erofs_sb_list_lock);
 252         mutex_unlock(&sbi->umount_mutex);
 253 }
 254
 255 static unsigned long erofs_shrink_count(struct shrinker *shrink,
 256                                         struct shrink_control *sc)
 257 {
 258         return atomic_long_read(&erofs_global_shrink_cnt);
 259 }
 260
 261 static unsigned long erofs_shrink_scan(struct shrinker *shrink,
 262                                        struct shrink_control *sc)
 263 {
 264         struct erofs_sb_info *sbi;
 265         struct list_head *p;
 266
 267         unsigned long nr = sc->nr_to_scan;
 268         unsigned int run_no;
 269         unsigned long freed = 0;
 270
 271         spin_lock(&erofs_sb_list_lock);
 272         do {
 273                 run_no = ++shrinker_run_no;
 274         } while (run_no == 0);
 275
 276         /* Iterate over all mounted superblocks and try to shrink them */
 277         p = erofs_sb_list.next;
 278         while (p != &erofs_sb_list) {
 279                 sbi = list_entry(p, struct erofs_sb_info, list);
 280
 281                 /*
 282                  * We move the ones we do to the end of the list, so we stop
 283                  * when we see one we have already done.
 284                  */
 285                 if (sbi->shrinker_run_no == run_no)
 286                         break;
 287
 288                 if (!mutex_trylock(&sbi->umount_mutex)) {
 289                         p = p->next;
 290                         continue;
 291                 }
 292
 293                 spin_unlock(&erofs_sb_list_lock);
 294                 sbi->shrinker_run_no = run_no;
 295
 296                 freed += erofs_shrink_workstation(sbi, nr);
 297
 298                 spin_lock(&erofs_sb_list_lock);
 299                 /* Get the next list element before we move this one */
 300                 p = p->next;
 301
 302                 /*
 303                  * Move this one to the end of the list to provide some
 304                  * fairness.
 305                  */
 306                 list_move_tail(&sbi->list, &erofs_sb_list);
 307                 mutex_unlock(&sbi->umount_mutex);
 308
 309                 if (freed >= nr)
 310                         break;
 311         }
 312         spin_unlock(&erofs_sb_list_lock);
 313         return freed;
 314 }
 315
 316 static struct shrinker erofs_shrinker_info = {
 317         .scan_objects = erofs_shrink_scan,
 318         .count_objects = erofs_shrink_count,
 319         .seeks = DEFAULT_SEEKS,
 320 };
 321
 322 int __init erofs_init_shrinker(void)
 323 {
 324         return register_shrinker(&erofs_shrinker_info);
 325 }
 326
 327 void erofs_exit_shrinker(void)
 328 {
 329         unregister_shrinker(&erofs_shrinker_info);
 330 }
 331 #endif  /* !CONFIG_EROFS_FS_ZIP */
 332