drivers/staging/erofs/utils.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   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 <gaoxiang25@huawei.com>
   8  */
   9 #include "internal.h"
  10 #include <linux/pagevec.h>
  11
  12 struct page *erofs_allocpage(struct list_head *pool, gfp_t gfp, bool nofail)
  13 {
  14         struct page *page;
  15
  16         if (!list_empty(pool)) {
  17                 page = lru_to_page(pool);
  18                 DBG_BUGON(page_ref_count(page) != 1);
  19                 list_del(&page->lru);
  20         } else {
  21                 page = alloc_pages(gfp | (nofail ? __GFP_NOFAIL : 0), 0);
  22         }
  23         return page;
  24 }
  25
  26 #if (EROFS_PCPUBUF_NR_PAGES > 0)
  27 static struct {
  28         u8 data[PAGE_SIZE * EROFS_PCPUBUF_NR_PAGES];
  29 } ____cacheline_aligned_in_smp erofs_pcpubuf[NR_CPUS];
  30
  31 void *erofs_get_pcpubuf(unsigned int pagenr)
  32 {
  33         preempt_disable();
  34         return &erofs_pcpubuf[smp_processor_id()].data[pagenr * PAGE_SIZE];
  35 }
  36 #endif
  37
  38 #ifdef CONFIG_EROFS_FS_ZIP
  39 /* global shrink count (for all mounted EROFS instances) */
  40 static atomic_long_t erofs_global_shrink_cnt;
  41
  42 #define __erofs_workgroup_get(grp)      atomic_inc(&(grp)->refcount)
  43 #define __erofs_workgroup_put(grp)      atomic_dec(&(grp)->refcount)
  44
  45 static int erofs_workgroup_get(struct erofs_workgroup *grp)
  46 {
  47         int o;
  48
  49 repeat:
  50         o = erofs_wait_on_workgroup_freezed(grp);
  51         if (unlikely(o <= 0))
  52                 return -1;
  53
  54         if (unlikely(atomic_cmpxchg(&grp->refcount, o, o + 1) != o))
  55                 goto repeat;
  56
  57         /* decrease refcount paired by erofs_workgroup_put */
  58         if (unlikely(o == 1))
  59                 atomic_long_dec(&erofs_global_shrink_cnt);
  60         return 0;
  61 }
  62
  63 struct erofs_workgroup *erofs_find_workgroup(struct super_block *sb,
  64                                              pgoff_t index, bool *tag)
  65 {
  66         struct erofs_sb_info *sbi = EROFS_SB(sb);
  67         struct erofs_workgroup *grp;
  68
  69 repeat:
  70         rcu_read_lock();
  71         grp = radix_tree_lookup(&sbi->workstn_tree, index);
  72         if (grp) {
  73                 *tag = xa_pointer_tag(grp);
  74                 grp = xa_untag_pointer(grp);
  75
  76                 if (erofs_workgroup_get(grp)) {
  77                         /* prefer to relax rcu read side */
  78                         rcu_read_unlock();
  79                         goto repeat;
  80                 }
  81
  82                 DBG_BUGON(index != grp->index);
  83         }
  84         rcu_read_unlock();
  85         return grp;
  86 }
  87
  88 int erofs_register_workgroup(struct super_block *sb,
  89                              struct erofs_workgroup *grp,
  90                              bool tag)
  91 {
  92         struct erofs_sb_info *sbi;
  93         int err;
  94
  95         /* grp shouldn't be broken or used before */
  96         if (unlikely(atomic_read(&grp->refcount) != 1)) {
  97                 DBG_BUGON(1);
  98                 return -EINVAL;
  99         }
 100
 101         err = radix_tree_preload(GFP_NOFS);
 102         if (err)
 103                 return err;
 104
 105         sbi = EROFS_SB(sb);
 106         xa_lock(&sbi->workstn_tree);
 107
 108         grp = xa_tag_pointer(grp, tag);
 109
 110         /*
 111          * Bump up reference count before making this workgroup
 112          * visible to other users in order to avoid potential UAF
 113          * without serialized by workstn_lock.
 114          */
 115         __erofs_workgroup_get(grp);
 116
 117         err = radix_tree_insert(&sbi->workstn_tree, grp->index, grp);
 118         if (unlikely(err))
 119                 /*
 120                  * it's safe to decrease since the workgroup isn't visible
 121                  * and refcount >= 2 (cannot be freezed).
 122                  */
 123                 __erofs_workgroup_put(grp);
 124
 125         xa_unlock(&sbi->workstn_tree);
 126         radix_tree_preload_end();
 127         return err;
 128 }
 129
 130 static void  __erofs_workgroup_free(struct erofs_workgroup *grp)
 131 {
 132         atomic_long_dec(&erofs_global_shrink_cnt);
 133         erofs_workgroup_free_rcu(grp);
 134 }
 135
 136 int erofs_workgroup_put(struct erofs_workgroup *grp)
 137 {
 138         int count = atomic_dec_return(&grp->refcount);
 139
 140         if (count == 1)
 141                 atomic_long_inc(&erofs_global_shrink_cnt);
 142         else if (!count)
 143                 __erofs_workgroup_free(grp);
 144         return count;
 145 }
 146
 147 static void erofs_workgroup_unfreeze_final(struct erofs_workgroup *grp)
 148 {
 149         erofs_workgroup_unfreeze(grp, 0);
 150         __erofs_workgroup_free(grp);
 151 }
 152
 153 static bool erofs_try_to_release_workgroup(struct erofs_sb_info *sbi,
 154                                            struct erofs_workgroup *grp,
 155                                            bool cleanup)
 156 {
 157         /*
 158          * If managed cache is on, refcount of workgroups
 159          * themselves could be < 0 (freezed). In other words,
 160          * there is no guarantee that all refcounts > 0.
 161          */
 162         if (!erofs_workgroup_try_to_freeze(grp, 1))
 163                 return false;
 164
 165         /*
 166          * Note that all cached pages should be unattached
 167          * before deleted from the radix tree. Otherwise some
 168          * cached pages could be still attached to the orphan
 169          * old workgroup when the new one is available in the tree.
 170          */
 171         if (erofs_try_to_free_all_cached_pages(sbi, grp)) {
 172                 erofs_workgroup_unfreeze(grp, 1);
 173                 return false;
 174         }
 175
 176         /*
 177          * It's impossible to fail after the workgroup is freezed,
 178          * however in order to avoid some race conditions, add a
 179          * DBG_BUGON to observe this in advance.
 180          */
 181         DBG_BUGON(xa_untag_pointer(radix_tree_delete(&sbi->workstn_tree,
 182                                                      grp->index)) != grp);
 183
 184         /*
 185          * If managed cache is on, last refcount should indicate
 186          * the related workstation.
 187          */
 188         erofs_workgroup_unfreeze_final(grp);
 189         return true;
 190 }
 191
 192 static unsigned long erofs_shrink_workstation(struct erofs_sb_info *sbi,
 193                                               unsigned long nr_shrink,
 194                                               bool cleanup)
 195 {
 196         pgoff_t first_index = 0;
 197         void *batch[PAGEVEC_SIZE];
 198         unsigned int freed = 0;
 199
 200         int i, found;
 201 repeat:
 202         xa_lock(&sbi->workstn_tree);
 203
 204         found = radix_tree_gang_lookup(&sbi->workstn_tree,
 205                                        batch, first_index, PAGEVEC_SIZE);
 206
 207         for (i = 0; i < found; ++i) {
 208                 struct erofs_workgroup *grp = xa_untag_pointer(batch[i]);
 209
 210                 first_index = grp->index + 1;
 211
 212                 /* try to shrink each valid workgroup */
 213                 if (!erofs_try_to_release_workgroup(sbi, grp, cleanup))
 214                         continue;
 215
 216                 ++freed;
 217                 if (unlikely(!--nr_shrink))
 218                         break;
 219         }
 220         xa_unlock(&sbi->workstn_tree);
 221
 222         if (i && nr_shrink)
 223                 goto repeat;
 224         return freed;
 225 }
 226
 227 /* protected by 'erofs_sb_list_lock' */
 228 static unsigned int shrinker_run_no;
 229
 230 /* protects the mounted 'erofs_sb_list' */
 231 static DEFINE_SPINLOCK(erofs_sb_list_lock);
 232 static LIST_HEAD(erofs_sb_list);
 233
 234 void erofs_shrinker_register(struct super_block *sb)
 235 {
 236         struct erofs_sb_info *sbi = EROFS_SB(sb);
 237
 238         mutex_init(&sbi->umount_mutex);
 239
 240         spin_lock(&erofs_sb_list_lock);
 241         list_add(&sbi->list, &erofs_sb_list);
 242         spin_unlock(&erofs_sb_list_lock);
 243 }
 244
 245 void erofs_shrinker_unregister(struct super_block *sb)
 246 {
 247         struct erofs_sb_info *const sbi = EROFS_SB(sb);
 248
 249         mutex_lock(&sbi->umount_mutex);
 250         erofs_shrink_workstation(sbi, ~0UL, true);
 251
 252         spin_lock(&erofs_sb_list_lock);
 253         list_del(&sbi->list);
 254         spin_unlock(&erofs_sb_list_lock);
 255         mutex_unlock(&sbi->umount_mutex);
 256 }
 257
 258 static unsigned long erofs_shrink_count(struct shrinker *shrink,
 259                                         struct shrink_control *sc)
 260 {
 261         return atomic_long_read(&erofs_global_shrink_cnt);
 262 }
 263
 264 static unsigned long erofs_shrink_scan(struct shrinker *shrink,
 265                                        struct shrink_control *sc)
 266 {
 267         struct erofs_sb_info *sbi;
 268         struct list_head *p;
 269
 270         unsigned long nr = sc->nr_to_scan;
 271         unsigned int run_no;
 272         unsigned long freed = 0;
 273
 274         spin_lock(&erofs_sb_list_lock);
 275         do {
 276                 run_no = ++shrinker_run_no;
 277         } while (run_no == 0);
 278
 279         /* Iterate over all mounted superblocks and try to shrink them */
 280         p = erofs_sb_list.next;
 281         while (p != &erofs_sb_list) {
 282                 sbi = list_entry(p, struct erofs_sb_info, list);
 283
 284                 /*
 285                  * We move the ones we do to the end of the list, so we stop
 286                  * when we see one we have already done.
 287                  */
 288                 if (sbi->shrinker_run_no == run_no)
 289                         break;
 290
 291                 if (!mutex_trylock(&sbi->umount_mutex)) {
 292                         p = p->next;
 293                         continue;
 294                 }
 295
 296                 spin_unlock(&erofs_sb_list_lock);
 297                 sbi->shrinker_run_no = run_no;
 298
 299                 freed += erofs_shrink_workstation(sbi, nr, false);
 300
 301                 spin_lock(&erofs_sb_list_lock);
 302                 /* Get the next list element before we move this one */
 303                 p = p->next;
 304
 305                 /*
 306                  * Move this one to the end of the list to provide some
 307                  * fairness.
 308                  */
 309                 list_move_tail(&sbi->list, &erofs_sb_list);
 310                 mutex_unlock(&sbi->umount_mutex);
 311
 312                 if (freed >= nr)
 313                         break;
 314         }
 315         spin_unlock(&erofs_sb_list_lock);
 316         return freed;
 317 }
 318
 319 static struct shrinker erofs_shrinker_info = {
 320         .scan_objects = erofs_shrink_scan,
 321         .count_objects = erofs_shrink_count,
 322         .seeks = DEFAULT_SEEKS,
 323 };
 324
 325 int __init erofs_init_shrinker(void)
 326 {
 327         return register_shrinker(&erofs_shrinker_info);
 328 }
 329
 330 void erofs_exit_shrinker(void)
 331 {
 332         unregister_shrinker(&erofs_shrinker_info);
 333 }
 334 #endif  /* !CONFIG_EROFS_FS_ZIP */
 335