mm/kasan/quarantine.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * KASAN quarantine.
   4  *
   5  * Author: Alexander Potapenko <glider@google.com>
   6  * Copyright (C) 2016 Google, Inc.
   7  *
   8  * Based on code by Dmitry Chernenkov.
   9  *
  10  * This program is free software; you can redistribute it and/or
  11  * modify it under the terms of the GNU General Public License
  12  * version 2 as published by the Free Software Foundation.
  13  *
  14  * This program is distributed in the hope that it will be useful, but
  15  * WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * General Public License for more details.
  18  *
  19  */
  20
  21 #include <linux/gfp.h>
  22 #include <linux/hash.h>
  23 #include <linux/kernel.h>
  24 #include <linux/mm.h>
  25 #include <linux/percpu.h>
  26 #include <linux/printk.h>
  27 #include <linux/shrinker.h>
  28 #include <linux/slab.h>
  29 #include <linux/srcu.h>
  30 #include <linux/string.h>
  31 #include <linux/types.h>
  32
  33 #include "../slab.h"
  34 #include "kasan.h"
  35
  36 /* Data structure and operations for quarantine queues. */
  37
  38 /*
  39  * Each queue is a signle-linked list, which also stores the total size of
  40  * objects inside of it.
  41  */
  42 struct qlist_head {
  43         struct qlist_node *head;
  44         struct qlist_node *tail;
  45         size_t bytes;
  46 };
  47
  48 #define QLIST_INIT { NULL, NULL, 0 }
  49
  50 static bool qlist_empty(struct qlist_head *q)
  51 {
  52         return !q->head;
  53 }
  54
  55 static void qlist_init(struct qlist_head *q)
  56 {
  57         q->head = q->tail = NULL;
  58         q->bytes = 0;
  59 }
  60
  61 static void qlist_put(struct qlist_head *q, struct qlist_node *qlink,
  62                 size_t size)
  63 {
  64         if (unlikely(qlist_empty(q)))
  65                 q->head = qlink;
  66         else
  67                 q->tail->next = qlink;
  68         q->tail = qlink;
  69         qlink->next = NULL;
  70         q->bytes += size;
  71 }
  72
  73 static void qlist_move_all(struct qlist_head *from, struct qlist_head *to)
  74 {
  75         if (unlikely(qlist_empty(from)))
  76                 return;
  77
  78         if (qlist_empty(to)) {
  79                 *to = *from;
  80                 qlist_init(from);
  81                 return;
  82         }
  83
  84         to->tail->next = from->head;
  85         to->tail = from->tail;
  86         to->bytes += from->bytes;
  87
  88         qlist_init(from);
  89 }
  90
  91 #define QUARANTINE_PERCPU_SIZE (1 << 20)
  92 #define QUARANTINE_BATCHES \
  93         (1024 > 4 * CONFIG_NR_CPUS ? 1024 : 4 * CONFIG_NR_CPUS)
  94
  95 /*
  96  * The object quarantine consists of per-cpu queues and a global queue,
  97  * guarded by quarantine_lock.
  98  */
  99 static DEFINE_PER_CPU(struct qlist_head, cpu_quarantine);
 100
 101 /* Round-robin FIFO array of batches. */
 102 static struct qlist_head global_quarantine[QUARANTINE_BATCHES];
 103 static int quarantine_head;
 104 static int quarantine_tail;
 105 /* Total size of all objects in global_quarantine across all batches. */
 106 static unsigned long quarantine_size;
 107 static DEFINE_RAW_SPINLOCK(quarantine_lock);
 108 DEFINE_STATIC_SRCU(remove_cache_srcu);
 109
 110 /* Maximum size of the global queue. */
 111 static unsigned long quarantine_max_size;
 112
 113 /*
 114  * Target size of a batch in global_quarantine.
 115  * Usually equal to QUARANTINE_PERCPU_SIZE unless we have too much RAM.
 116  */
 117 static unsigned long quarantine_batch_size;
 118
 119 /*
 120  * The fraction of physical memory the quarantine is allowed to occupy.
 121  * Quarantine doesn't support memory shrinker with SLAB allocator, so we keep
 122  * the ratio low to avoid OOM.
 123  */
 124 #define QUARANTINE_FRACTION 32
 125
 126 static struct kmem_cache *qlink_to_cache(struct qlist_node *qlink)
 127 {
 128         return virt_to_head_page(qlink)->slab_cache;
 129 }
 130
 131 static void *qlink_to_object(struct qlist_node *qlink, struct kmem_cache *cache)
 132 {
 133         struct kasan_free_meta *free_info =
 134                 container_of(qlink, struct kasan_free_meta,
 135                              quarantine_link);
 136
 137         return ((void *)free_info) - cache->kasan_info.free_meta_offset;
 138 }
 139
 140 static void qlink_free(struct qlist_node *qlink, struct kmem_cache *cache)
 141 {
 142         void *object = qlink_to_object(qlink, cache);
 143         unsigned long flags;
 144
 145         if (IS_ENABLED(CONFIG_SLAB))
 146                 local_irq_save(flags);
 147
 148         ___cache_free(cache, object, _THIS_IP_);
 149
 150         if (IS_ENABLED(CONFIG_SLAB))
 151                 local_irq_restore(flags);
 152 }
 153
 154 static void qlist_free_all(struct qlist_head *q, struct kmem_cache *cache)
 155 {
 156         struct qlist_node *qlink;
 157
 158         if (unlikely(qlist_empty(q)))
 159                 return;
 160
 161         qlink = q->head;
 162         while (qlink) {
 163                 struct kmem_cache *obj_cache =
 164                         cache ? cache : qlink_to_cache(qlink);
 165                 struct qlist_node *next = qlink->next;
 166
 167                 qlink_free(qlink, obj_cache);
 168                 qlink = next;
 169         }
 170         qlist_init(q);
 171 }
 172
 173 void quarantine_put(struct kasan_free_meta *info, struct kmem_cache *cache)
 174 {
 175         unsigned long flags;
 176         struct qlist_head *q;
 177         struct qlist_head temp = QLIST_INIT;
 178
 179         /*
 180          * Note: irq must be disabled until after we move the batch to the
 181          * global quarantine. Otherwise quarantine_remove_cache() can miss
 182          * some objects belonging to the cache if they are in our local temp
 183          * list. quarantine_remove_cache() executes on_each_cpu() at the
 184          * beginning which ensures that it either sees the objects in per-cpu
 185          * lists or in the global quarantine.
 186          */
 187         local_irq_save(flags);
 188
 189         q = this_cpu_ptr(&cpu_quarantine);
 190         qlist_put(q, &info->quarantine_link, cache->size);
 191         if (unlikely(q->bytes > QUARANTINE_PERCPU_SIZE)) {
 192                 qlist_move_all(q, &temp);
 193
 194                 raw_spin_lock(&quarantine_lock);
 195                 WRITE_ONCE(quarantine_size, quarantine_size + temp.bytes);
 196                 qlist_move_all(&temp, &global_quarantine[quarantine_tail]);
 197                 if (global_quarantine[quarantine_tail].bytes >=
 198                                 READ_ONCE(quarantine_batch_size)) {
 199                         int new_tail;
 200
 201                         new_tail = quarantine_tail + 1;
 202                         if (new_tail == QUARANTINE_BATCHES)
 203                                 new_tail = 0;
 204                         if (new_tail != quarantine_head)
 205                                 quarantine_tail = new_tail;
 206                 }
 207                 raw_spin_unlock(&quarantine_lock);
 208         }
 209
 210         local_irq_restore(flags);
 211 }
 212
 213 void quarantine_reduce(void)
 214 {
 215         size_t total_size, new_quarantine_size, percpu_quarantines;
 216         unsigned long flags;
 217         int srcu_idx;
 218         struct qlist_head to_free = QLIST_INIT;
 219
 220         if (likely(READ_ONCE(quarantine_size) <=
 221                    READ_ONCE(quarantine_max_size)))
 222                 return;
 223
 224         /*
 225          * srcu critical section ensures that quarantine_remove_cache()
 226          * will not miss objects belonging to the cache while they are in our
 227          * local to_free list. srcu is chosen because (1) it gives us private
 228          * grace period domain that does not interfere with anything else,
 229          * and (2) it allows synchronize_srcu() to return without waiting
 230          * if there are no pending read critical sections (which is the
 231          * expected case).
 232          */
 233         srcu_idx = srcu_read_lock(&remove_cache_srcu);
 234         raw_spin_lock_irqsave(&quarantine_lock, flags);
 235
 236         /*
 237          * Update quarantine size in case of hotplug. Allocate a fraction of
 238          * the installed memory to quarantine minus per-cpu queue limits.
 239          */
 240         total_size = (totalram_pages() << PAGE_SHIFT) /
 241                 QUARANTINE_FRACTION;
 242         percpu_quarantines = QUARANTINE_PERCPU_SIZE * num_online_cpus();
 243         new_quarantine_size = (total_size < percpu_quarantines) ?
 244                 0 : total_size - percpu_quarantines;
 245         WRITE_ONCE(quarantine_max_size, new_quarantine_size);
 246         /* Aim at consuming at most 1/2 of slots in quarantine. */
 247         WRITE_ONCE(quarantine_batch_size, max((size_t)QUARANTINE_PERCPU_SIZE,
 248                 2 * total_size / QUARANTINE_BATCHES));
 249
 250         if (likely(quarantine_size > quarantine_max_size)) {
 251                 qlist_move_all(&global_quarantine[quarantine_head], &to_free);
 252                 WRITE_ONCE(quarantine_size, quarantine_size - to_free.bytes);
 253                 quarantine_head++;
 254                 if (quarantine_head == QUARANTINE_BATCHES)
 255                         quarantine_head = 0;
 256         }
 257
 258         raw_spin_unlock_irqrestore(&quarantine_lock, flags);
 259
 260         qlist_free_all(&to_free, NULL);
 261         srcu_read_unlock(&remove_cache_srcu, srcu_idx);
 262 }
 263
 264 static void qlist_move_cache(struct qlist_head *from,
 265                                    struct qlist_head *to,
 266                                    struct kmem_cache *cache)
 267 {
 268         struct qlist_node *curr;
 269
 270         if (unlikely(qlist_empty(from)))
 271                 return;
 272
 273         curr = from->head;
 274         qlist_init(from);
 275         while (curr) {
 276                 struct qlist_node *next = curr->next;
 277                 struct kmem_cache *obj_cache = qlink_to_cache(curr);
 278
 279                 if (obj_cache == cache)
 280                         qlist_put(to, curr, obj_cache->size);
 281                 else
 282                         qlist_put(from, curr, obj_cache->size);
 283
 284                 curr = next;
 285         }
 286 }
 287
 288 static void per_cpu_remove_cache(void *arg)
 289 {
 290         struct kmem_cache *cache = arg;
 291         struct qlist_head to_free = QLIST_INIT;
 292         struct qlist_head *q;
 293
 294         q = this_cpu_ptr(&cpu_quarantine);
 295         qlist_move_cache(q, &to_free, cache);
 296         qlist_free_all(&to_free, cache);
 297 }
 298
 299 /* Free all quarantined objects belonging to cache. */
 300 void quarantine_remove_cache(struct kmem_cache *cache)
 301 {
 302         unsigned long flags, i;
 303         struct qlist_head to_free = QLIST_INIT;
 304
 305         /*
 306          * Must be careful to not miss any objects that are being moved from
 307          * per-cpu list to the global quarantine in quarantine_put(),
 308          * nor objects being freed in quarantine_reduce(). on_each_cpu()
 309          * achieves the first goal, while synchronize_srcu() achieves the
 310          * second.
 311          */
 312         on_each_cpu(per_cpu_remove_cache, cache, 1);
 313
 314         raw_spin_lock_irqsave(&quarantine_lock, flags);
 315         for (i = 0; i < QUARANTINE_BATCHES; i++) {
 316                 if (qlist_empty(&global_quarantine[i]))
 317                         continue;
 318                 qlist_move_cache(&global_quarantine[i], &to_free, cache);
 319                 /* Scanning whole quarantine can take a while. */
 320                 raw_spin_unlock_irqrestore(&quarantine_lock, flags);
 321                 cond_resched();
 322                 raw_spin_lock_irqsave(&quarantine_lock, flags);
 323         }
 324         raw_spin_unlock_irqrestore(&quarantine_lock, flags);
 325
 326         qlist_free_all(&to_free, cache);
 327
 328         synchronize_srcu(&remove_cache_srcu);
 329 }