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 static int erofs_workgroup_get(struct erofs_workgroup *grp)
  41 {
  42         int o;
  43
  44 repeat:
  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
  58 struct 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
  64 repeat:
  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
  80 struct 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
  93 repeat:
  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
 113 static 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
 119 int 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
 130 static void erofs_workgroup_unfreeze_final(struct erofs_workgroup *grp)
 131 {
 132         erofs_workgroup_unfreeze(grp, 0);
 133         __erofs_workgroup_free(grp);
 134 }
 135
 136 static bool erofs_try_to_release_workgroup(struct erofs_sb_info *sbi,
 137                                            struct erofs_workgroup *grp)
 138 {
 139         /*
 140          * If managed cache is on, refcount of workgroups
 141          * themselves could be < 0 (freezed). In other words,
 142          * there is no guarantee that all refcounts > 0.
 143          */
 144         if (!erofs_workgroup_try_to_freeze(grp, 1))
 145                 return false;
 146
 147         /*
 148          * Note that all cached pages should be unattached
 149          * before deleted from the XArray. Otherwise some
 150          * cached pages could be still attached to the orphan
 151          * old workgroup when the new one is available in the tree.
 152          */
 153         if (erofs_try_to_free_all_cached_pages(sbi, grp)) {
 154                 erofs_workgroup_unfreeze(grp, 1);
 155                 return false;
 156         }
 157
 158         /*
 159          * It's impossible to fail after the workgroup is freezed,
 160          * however in order to avoid some race conditions, add a
 161          * DBG_BUGON to observe this in advance.
 162          */
 163         DBG_BUGON(xa_erase(&sbi->managed_pslots, grp->index) != grp);
 164
 165         /*
 166          * If managed cache is on, last refcount should indicate
 167          * the related workstation.
 168          */
 169         erofs_workgroup_unfreeze_final(grp);
 170         return true;
 171 }
 172
 173 static unsigned long erofs_shrink_workstation(struct erofs_sb_info *sbi,
 174                                               unsigned long nr_shrink)
 175 {
 176         struct erofs_workgroup *grp;
 177         unsigned int freed = 0;
 178         unsigned long index;
 179
 180         xa_for_each(&sbi->managed_pslots, index, grp) {
 181                 /* try to shrink each valid workgroup */
 182                 if (!erofs_try_to_release_workgroup(sbi, grp))
 183                         continue;
 184
 185                 ++freed;
 186                 if (!--nr_shrink)
 187                         break;
 188         }
 189         return freed;
 190 }
 191
 192 /* protected by 'erofs_sb_list_lock' */
 193 static unsigned int shrinker_run_no;
 194
 195 /* protects the mounted 'erofs_sb_list' */
 196 static DEFINE_SPINLOCK(erofs_sb_list_lock);
 197 static LIST_HEAD(erofs_sb_list);
 198
 199 void erofs_shrinker_register(struct super_block *sb)
 200 {
 201         struct erofs_sb_info *sbi = EROFS_SB(sb);
 202
 203         mutex_init(&sbi->umount_mutex);
 204
 205         spin_lock(&erofs_sb_list_lock);
 206         list_add(&sbi->list, &erofs_sb_list);
 207         spin_unlock(&erofs_sb_list_lock);
 208 }
 209
 210 void erofs_shrinker_unregister(struct super_block *sb)
 211 {
 212         struct erofs_sb_info *const sbi = EROFS_SB(sb);
 213
 214         mutex_lock(&sbi->umount_mutex);
 215         /* clean up all remaining workgroups in memory */
 216         erofs_shrink_workstation(sbi, ~0UL);
 217
 218         spin_lock(&erofs_sb_list_lock);
 219         list_del(&sbi->list);
 220         spin_unlock(&erofs_sb_list_lock);
 221         mutex_unlock(&sbi->umount_mutex);
 222 }
 223
 224 static unsigned long erofs_shrink_count(struct shrinker *shrink,
 225                                         struct shrink_control *sc)
 226 {
 227         return atomic_long_read(&erofs_global_shrink_cnt);
 228 }
 229
 230 static unsigned long erofs_shrink_scan(struct shrinker *shrink,
 231                                        struct shrink_control *sc)
 232 {
 233         struct erofs_sb_info *sbi;
 234         struct list_head *p;
 235
 236         unsigned long nr = sc->nr_to_scan;
 237         unsigned int run_no;
 238         unsigned long freed = 0;
 239
 240         spin_lock(&erofs_sb_list_lock);
 241         do {
 242                 run_no = ++shrinker_run_no;
 243         } while (run_no == 0);
 244
 245         /* Iterate over all mounted superblocks and try to shrink them */
 246         p = erofs_sb_list.next;
 247         while (p != &erofs_sb_list) {
 248                 sbi = list_entry(p, struct erofs_sb_info, list);
 249
 250                 /*
 251                  * We move the ones we do to the end of the list, so we stop
 252                  * when we see one we have already done.
 253                  */
 254                 if (sbi->shrinker_run_no == run_no)
 255                         break;
 256
 257                 if (!mutex_trylock(&sbi->umount_mutex)) {
 258                         p = p->next;
 259                         continue;
 260                 }
 261
 262                 spin_unlock(&erofs_sb_list_lock);
 263                 sbi->shrinker_run_no = run_no;
 264
 265                 freed += erofs_shrink_workstation(sbi, nr - freed);
 266
 267                 spin_lock(&erofs_sb_list_lock);
 268                 /* Get the next list element before we move this one */
 269                 p = p->next;
 270
 271                 /*
 272                  * Move this one to the end of the list to provide some
 273                  * fairness.
 274                  */
 275                 list_move_tail(&sbi->list, &erofs_sb_list);
 276                 mutex_unlock(&sbi->umount_mutex);
 277
 278                 if (freed >= nr)
 279                         break;
 280         }
 281         spin_unlock(&erofs_sb_list_lock);
 282         return freed;
 283 }
 284
 285 static struct shrinker erofs_shrinker_info = {
 286         .scan_objects = erofs_shrink_scan,
 287         .count_objects = erofs_shrink_count,
 288         .seeks = DEFAULT_SEEKS,
 289 };
 290
 291 int __init erofs_init_shrinker(void)
 292 {
 293         return register_shrinker(&erofs_shrinker_info);
 294 }
 295
 296 void erofs_exit_shrinker(void)
 297 {
 298         unregister_shrinker(&erofs_shrinker_info);
 299 }
 300 #endif  /* !CONFIG_EROFS_FS_ZIP */
 301