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 #include <linux/cpuhotplug.h>
  33
  34 #include "../slab.h"
  35 #include "kasan.h"
  36
  37 /* Data structure and operations for quarantine queues. */
  38
  39 /*
  40  * Each queue is a signle-linked list, which also stores the total size of
  41  * objects inside of it.
  42  */
  43 struct qlist_head {
  44         struct qlist_node *head;
  45         struct qlist_node *tail;
  46         size_t bytes;
  47         bool offline;
  48 };
  49
  50 #define QLIST_INIT { NULL, NULL, 0 }
  51
  52 static bool qlist_empty(struct qlist_head *q)
  53 {
  54         return !q->head;
  55 }
  56
  57 static void qlist_init(struct qlist_head *q)
  58 {
  59         q->head = q->tail = NULL;
  60         q->bytes = 0;
  61 }
  62
  63 static void qlist_put(struct qlist_head *q, struct qlist_node *qlink,
  64                 size_t size)
  65 {
  66         if (unlikely(qlist_empty(q)))
  67                 q->head = qlink;
  68         else
  69                 q->tail->next = qlink;
  70         q->tail = qlink;
  71         qlink->next = NULL;
  72         q->bytes += size;
  73 }
  74
  75 static void qlist_move_all(struct qlist_head *from, struct qlist_head *to)
  76 {
  77         if (unlikely(qlist_empty(from)))
  78                 return;
  79
  80         if (qlist_empty(to)) {
  81                 *to = *from;
  82                 qlist_init(from);
  83                 return;
  84         }
  85
  86         to->tail->next = from->head;
  87         to->tail = from->tail;
  88         to->bytes += from->bytes;
  89
  90         qlist_init(from);
  91 }
  92
  93 #define QUARANTINE_PERCPU_SIZE (1 << 20)
  94 #define QUARANTINE_BATCHES \
  95         (1024 > 4 * CONFIG_NR_CPUS ? 1024 : 4 * CONFIG_NR_CPUS)
  96
  97 /*
  98  * The object quarantine consists of per-cpu queues and a global queue,
  99  * guarded by quarantine_lock.
 100  */
 101 static DEFINE_PER_CPU(struct qlist_head, cpu_quarantine);
 102
 103 /* Round-robin FIFO array of batches. */
 104 static struct qlist_head global_quarantine[QUARANTINE_BATCHES];
 105 static int quarantine_head;
 106 static int quarantine_tail;
 107 /* Total size of all objects in global_quarantine across all batches. */
 108 static unsigned long quarantine_size;
 109 static DEFINE_RAW_SPINLOCK(quarantine_lock);
 110 DEFINE_STATIC_SRCU(remove_cache_srcu);
 111
 112 /* Maximum size of the global queue. */
 113 static unsigned long quarantine_max_size;
 114
 115 /*
 116  * Target size of a batch in global_quarantine.
 117  * Usually equal to QUARANTINE_PERCPU_SIZE unless we have too much RAM.
 118  */
 119 static unsigned long quarantine_batch_size;
 120
 121 /*
 122  * The fraction of physical memory the quarantine is allowed to occupy.
 123  * Quarantine doesn't support memory shrinker with SLAB allocator, so we keep
 124  * the ratio low to avoid OOM.
 125  */
 126 #define QUARANTINE_FRACTION 32
 127
 128 static struct kmem_cache *qlink_to_cache(struct qlist_node *qlink)
 129 {
 130         return virt_to_head_page(qlink)->slab_cache;
 131 }
 132
 133 static void *qlink_to_object(struct qlist_node *qlink, struct kmem_cache *cache)
 134 {
 135         struct kasan_free_meta *free_info =
 136                 container_of(qlink, struct kasan_free_meta,
 137                              quarantine_link);
 138
 139         return ((void *)free_info) - cache->kasan_info.free_meta_offset;
 140 }
 141
 142 static void qlink_free(struct qlist_node *qlink, struct kmem_cache *cache)
 143 {
 144         void *object = qlink_to_object(qlink, cache);
 145         unsigned long flags;
 146
 147         if (IS_ENABLED(CONFIG_SLAB))
 148                 local_irq_save(flags);
 149
 150         *(u8 *)kasan_mem_to_shadow(object) = KASAN_KMALLOC_FREE;
 151         ___cache_free(cache, object, _THIS_IP_);
 152
 153         if (IS_ENABLED(CONFIG_SLAB))
 154                 local_irq_restore(flags);
 155 }
 156
 157 static void qlist_free_all(struct qlist_head *q, struct kmem_cache *cache)
 158 {
 159         struct qlist_node *qlink;
 160
 161         if (unlikely(qlist_empty(q)))
 162                 return;
 163
 164         qlink = q->head;
 165         while (qlink) {
 166                 struct kmem_cache *obj_cache =
 167                         cache ? cache : qlink_to_cache(qlink);
 168                 struct qlist_node *next = qlink->next;
 169
 170                 qlink_free(qlink, obj_cache);
 171                 qlink = next;
 172         }
 173         qlist_init(q);
 174 }
 175
 176 void quarantine_put(struct kasan_free_meta *info, struct kmem_cache *cache)
 177 {
 178         unsigned long flags;
 179         struct qlist_head *q;
 180         struct qlist_head temp = QLIST_INIT;
 181
 182         /*
 183          * Note: irq must be disabled until after we move the batch to the
 184          * global quarantine. Otherwise quarantine_remove_cache() can miss
 185          * some objects belonging to the cache if they are in our local temp
 186          * list. quarantine_remove_cache() executes on_each_cpu() at the
 187          * beginning which ensures that it either sees the objects in per-cpu
 188          * lists or in the global quarantine.
 189          */
 190         local_irq_save(flags);
 191
 192         q = this_cpu_ptr(&cpu_quarantine);
 193         if (q->offline) {
 194                 local_irq_restore(flags);
 195                 return;
 196         }
 197         qlist_put(q, &info->quarantine_link, cache->size);
 198         if (unlikely(q->bytes > QUARANTINE_PERCPU_SIZE)) {
 199                 qlist_move_all(q, &temp);
 200
 201                 raw_spin_lock(&quarantine_lock);
 202                 WRITE_ONCE(quarantine_size, quarantine_size + temp.bytes);
 203                 qlist_move_all(&temp, &global_quarantine[quarantine_tail]);
 204                 if (global_quarantine[quarantine_tail].bytes >=
 205                                 READ_ONCE(quarantine_batch_size)) {
 206                         int new_tail;
 207
 208                         new_tail = quarantine_tail + 1;
 209                         if (new_tail == QUARANTINE_BATCHES)
 210                                 new_tail = 0;
 211                         if (new_tail != quarantine_head)
 212                                 quarantine_tail = new_tail;
 213                 }
 214                 raw_spin_unlock(&quarantine_lock);
 215         }
 216
 217         local_irq_restore(flags);
 218 }
 219
 220 void quarantine_reduce(void)
 221 {
 222         size_t total_size, new_quarantine_size, percpu_quarantines;
 223         unsigned long flags;
 224         int srcu_idx;
 225         struct qlist_head to_free = QLIST_INIT;
 226
 227         if (likely(READ_ONCE(quarantine_size) <=
 228                    READ_ONCE(quarantine_max_size)))
 229                 return;
 230
 231         /*
 232          * srcu critical section ensures that quarantine_remove_cache()
 233          * will not miss objects belonging to the cache while they are in our
 234          * local to_free list. srcu is chosen because (1) it gives us private
 235          * grace period domain that does not interfere with anything else,
 236          * and (2) it allows synchronize_srcu() to return without waiting
 237          * if there are no pending read critical sections (which is the
 238          * expected case).
 239          */
 240         srcu_idx = srcu_read_lock(&remove_cache_srcu);
 241         raw_spin_lock_irqsave(&quarantine_lock, flags);
 242
 243         /*
 244          * Update quarantine size in case of hotplug. Allocate a fraction of
 245          * the installed memory to quarantine minus per-cpu queue limits.
 246          */
 247         total_size = (totalram_pages() << PAGE_SHIFT) /
 248                 QUARANTINE_FRACTION;
 249         percpu_quarantines = QUARANTINE_PERCPU_SIZE * num_online_cpus();
 250         new_quarantine_size = (total_size < percpu_quarantines) ?
 251                 0 : total_size - percpu_quarantines;
 252         WRITE_ONCE(quarantine_max_size, new_quarantine_size);
 253         /* Aim at consuming at most 1/2 of slots in quarantine. */
 254         WRITE_ONCE(quarantine_batch_size, max((size_t)QUARANTINE_PERCPU_SIZE,
 255                 2 * total_size / QUARANTINE_BATCHES));
 256
 257         if (likely(quarantine_size > quarantine_max_size)) {
 258                 qlist_move_all(&global_quarantine[quarantine_head], &to_free);
 259                 WRITE_ONCE(quarantine_size, quarantine_size - to_free.bytes);
 260                 quarantine_head++;
 261                 if (quarantine_head == QUARANTINE_BATCHES)
 262                         quarantine_head = 0;
 263         }
 264
 265         raw_spin_unlock_irqrestore(&quarantine_lock, flags);
 266
 267         qlist_free_all(&to_free, NULL);
 268         srcu_read_unlock(&remove_cache_srcu, srcu_idx);
 269 }
 270
 271 static void qlist_move_cache(struct qlist_head *from,
 272                                    struct qlist_head *to,
 273                                    struct kmem_cache *cache)
 274 {
 275         struct qlist_node *curr;
 276
 277         if (unlikely(qlist_empty(from)))
 278                 return;
 279
 280         curr = from->head;
 281         qlist_init(from);
 282         while (curr) {
 283                 struct qlist_node *next = curr->next;
 284                 struct kmem_cache *obj_cache = qlink_to_cache(curr);
 285
 286                 if (obj_cache == cache)
 287                         qlist_put(to, curr, obj_cache->size);
 288                 else
 289                         qlist_put(from, curr, obj_cache->size);
 290
 291                 curr = next;
 292         }
 293 }
 294
 295 static void per_cpu_remove_cache(void *arg)
 296 {
 297         struct kmem_cache *cache = arg;
 298         struct qlist_head to_free = QLIST_INIT;
 299         struct qlist_head *q;
 300
 301         q = this_cpu_ptr(&cpu_quarantine);
 302         qlist_move_cache(q, &to_free, cache);
 303         qlist_free_all(&to_free, cache);
 304 }
 305
 306 /* Free all quarantined objects belonging to cache. */
 307 void quarantine_remove_cache(struct kmem_cache *cache)
 308 {
 309         unsigned long flags, i;
 310         struct qlist_head to_free = QLIST_INIT;
 311
 312         /*
 313          * Must be careful to not miss any objects that are being moved from
 314          * per-cpu list to the global quarantine in quarantine_put(),
 315          * nor objects being freed in quarantine_reduce(). on_each_cpu()
 316          * achieves the first goal, while synchronize_srcu() achieves the
 317          * second.
 318          */
 319         on_each_cpu(per_cpu_remove_cache, cache, 1);
 320
 321         raw_spin_lock_irqsave(&quarantine_lock, flags);
 322         for (i = 0; i < QUARANTINE_BATCHES; i++) {
 323                 if (qlist_empty(&global_quarantine[i]))
 324                         continue;
 325                 qlist_move_cache(&global_quarantine[i], &to_free, cache);
 326                 /* Scanning whole quarantine can take a while. */
 327                 raw_spin_unlock_irqrestore(&quarantine_lock, flags);
 328                 cond_resched();
 329                 raw_spin_lock_irqsave(&quarantine_lock, flags);
 330         }
 331         raw_spin_unlock_irqrestore(&quarantine_lock, flags);
 332
 333         qlist_free_all(&to_free, cache);
 334
 335         synchronize_srcu(&remove_cache_srcu);
 336 }
 337
 338 static int kasan_cpu_online(unsigned int cpu)
 339 {
 340         this_cpu_ptr(&cpu_quarantine)->offline = false;
 341         return 0;
 342 }
 343
 344 static int kasan_cpu_offline(unsigned int cpu)
 345 {
 346         struct qlist_head *q;
 347
 348         q = this_cpu_ptr(&cpu_quarantine);
 349         /* Ensure the ordering between the writing to q->offline and
 350          * qlist_free_all. Otherwise, cpu_quarantine may be corrupted
 351          * by interrupt.
 352          */
 353         WRITE_ONCE(q->offline, true);
 354         barrier();
 355         qlist_free_all(q, NULL);
 356         return 0;
 357 }
 358
 359 static int __init kasan_cpu_quarantine_init(void)
 360 {
 361         int ret = 0;
 362
 363         ret = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "mm/kasan:online",
 364                                 kasan_cpu_online, kasan_cpu_offline);
 365         if (ret < 0)
 366                 pr_err("kasan cpu quarantine register failed [%d]\n", ret);
 367         return ret;
 368 }
 369 late_initcall(kasan_cpu_quarantine_init);