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
  11 #include <linux/gfp.h>
  12 #include <linux/hash.h>
  13 #include <linux/kernel.h>
  14 #include <linux/mm.h>
  15 #include <linux/percpu.h>
  16 #include <linux/printk.h>
  17 #include <linux/shrinker.h>
  18 #include <linux/slab.h>
  19 #include <linux/srcu.h>
  20 #include <linux/string.h>
  21 #include <linux/types.h>
  22 #include <linux/cpuhotplug.h>
  23
  24 #include "../slab.h"
  25 #include "kasan.h"
  26
  27 /* Data structure and operations for quarantine queues. */
  28
  29 /*
  30  * Each queue is a single-linked list, which also stores the total size of
  31  * objects inside of it.
  32  */
  33 struct qlist_head {
  34         struct qlist_node *head;
  35         struct qlist_node *tail;
  36         size_t bytes;
  37         bool offline;
  38 };
  39
  40 #define QLIST_INIT { NULL, NULL, 0 }
  41
  42 static bool qlist_empty(struct qlist_head *q)
  43 {
  44         return !q->head;
  45 }
  46
  47 static void qlist_init(struct qlist_head *q)
  48 {
  49         q->head = q->tail = NULL;
  50         q->bytes = 0;
  51 }
  52
  53 static void qlist_put(struct qlist_head *q, struct qlist_node *qlink,
  54                 size_t size)
  55 {
  56         if (unlikely(qlist_empty(q)))
  57                 q->head = qlink;
  58         else
  59                 q->tail->next = qlink;
  60         q->tail = qlink;
  61         qlink->next = NULL;
  62         q->bytes += size;
  63 }
  64
  65 static void qlist_move_all(struct qlist_head *from, struct qlist_head *to)
  66 {
  67         if (unlikely(qlist_empty(from)))
  68                 return;
  69
  70         if (qlist_empty(to)) {
  71                 *to = *from;
  72                 qlist_init(from);
  73                 return;
  74         }
  75
  76         to->tail->next = from->head;
  77         to->tail = from->tail;
  78         to->bytes += from->bytes;
  79
  80         qlist_init(from);
  81 }
  82
  83 #define QUARANTINE_PERCPU_SIZE (1 << 20)
  84 #define QUARANTINE_BATCHES \
  85         (1024 > 4 * CONFIG_NR_CPUS ? 1024 : 4 * CONFIG_NR_CPUS)
  86
  87 /*
  88  * The object quarantine consists of per-cpu queues and a global queue,
  89  * guarded by quarantine_lock.
  90  */
  91 static DEFINE_PER_CPU(struct qlist_head, cpu_quarantine);
  92
  93 /* Round-robin FIFO array of batches. */
  94 static struct qlist_head global_quarantine[QUARANTINE_BATCHES];
  95 static int quarantine_head;
  96 static int quarantine_tail;
  97 /* Total size of all objects in global_quarantine across all batches. */
  98 static unsigned long quarantine_size;
  99 static DEFINE_RAW_SPINLOCK(quarantine_lock);
 100 DEFINE_STATIC_SRCU(remove_cache_srcu);
 101
 102 /* Maximum size of the global queue. */
 103 static unsigned long quarantine_max_size;
 104
 105 /*
 106  * Target size of a batch in global_quarantine.
 107  * Usually equal to QUARANTINE_PERCPU_SIZE unless we have too much RAM.
 108  */
 109 static unsigned long quarantine_batch_size;
 110
 111 /*
 112  * The fraction of physical memory the quarantine is allowed to occupy.
 113  * Quarantine doesn't support memory shrinker with SLAB allocator, so we keep
 114  * the ratio low to avoid OOM.
 115  */
 116 #define QUARANTINE_FRACTION 32
 117
 118 static struct kmem_cache *qlink_to_cache(struct qlist_node *qlink)
 119 {
 120         return virt_to_slab(qlink)->slab_cache;
 121 }
 122
 123 static void *qlink_to_object(struct qlist_node *qlink, struct kmem_cache *cache)
 124 {
 125         struct kasan_free_meta *free_info =
 126                 container_of(qlink, struct kasan_free_meta,
 127                              quarantine_link);
 128
 129         return ((void *)free_info) - cache->kasan_info.free_meta_offset;
 130 }
 131
 132 static void qlink_free(struct qlist_node *qlink, struct kmem_cache *cache)
 133 {
 134         void *object = qlink_to_object(qlink, cache);
 135         struct kasan_free_meta *meta = kasan_get_free_meta(cache, object);
 136         unsigned long flags;
 137
 138         if (IS_ENABLED(CONFIG_SLAB))
 139                 local_irq_save(flags);
 140
 141         /*
 142          * If init_on_free is enabled and KASAN's free metadata is stored in
 143          * the object, zero the metadata. Otherwise, the object's memory will
 144          * not be properly zeroed, as KASAN saves the metadata after the slab
 145          * allocator zeroes the object.
 146          */
 147         if (slab_want_init_on_free(cache) &&
 148             cache->kasan_info.free_meta_offset == 0)
 149                 memzero_explicit(meta, sizeof(*meta));
 150
 151         /*
 152          * As the object now gets freed from the quarantine, assume that its
 153          * free track is no longer valid.
 154          */
 155         *(u8 *)kasan_mem_to_shadow(object) = KASAN_KMALLOC_FREE;
 156
 157         ___cache_free(cache, object, _THIS_IP_);
 158
 159         if (IS_ENABLED(CONFIG_SLAB))
 160                 local_irq_restore(flags);
 161 }
 162
 163 static void qlist_free_all(struct qlist_head *q, struct kmem_cache *cache)
 164 {
 165         struct qlist_node *qlink;
 166
 167         if (unlikely(qlist_empty(q)))
 168                 return;
 169
 170         qlink = q->head;
 171         while (qlink) {
 172                 struct kmem_cache *obj_cache =
 173                         cache ? cache : qlink_to_cache(qlink);
 174                 struct qlist_node *next = qlink->next;
 175
 176                 qlink_free(qlink, obj_cache);
 177                 qlink = next;
 178         }
 179         qlist_init(q);
 180 }
 181
 182 bool kasan_quarantine_put(struct kmem_cache *cache, void *object)
 183 {
 184         unsigned long flags;
 185         struct qlist_head *q;
 186         struct qlist_head temp = QLIST_INIT;
 187         struct kasan_free_meta *meta = kasan_get_free_meta(cache, object);
 188
 189         /*
 190          * If there's no metadata for this object, don't put it into
 191          * quarantine.
 192          */
 193         if (!meta)
 194                 return false;
 195
 196         /*
 197          * Note: irq must be disabled until after we move the batch to the
 198          * global quarantine. Otherwise kasan_quarantine_remove_cache() can
 199          * miss some objects belonging to the cache if they are in our local
 200          * temp list. kasan_quarantine_remove_cache() executes on_each_cpu()
 201          * at the beginning which ensures that it either sees the objects in
 202          * per-cpu lists or in the global quarantine.
 203          */
 204         local_irq_save(flags);
 205
 206         q = this_cpu_ptr(&cpu_quarantine);
 207         if (q->offline) {
 208                 local_irq_restore(flags);
 209                 return false;
 210         }
 211         qlist_put(q, &meta->quarantine_link, cache->size);
 212         if (unlikely(q->bytes > QUARANTINE_PERCPU_SIZE)) {
 213                 qlist_move_all(q, &temp);
 214
 215                 raw_spin_lock(&quarantine_lock);
 216                 WRITE_ONCE(quarantine_size, quarantine_size + temp.bytes);
 217                 qlist_move_all(&temp, &global_quarantine[quarantine_tail]);
 218                 if (global_quarantine[quarantine_tail].bytes >=
 219                                 READ_ONCE(quarantine_batch_size)) {
 220                         int new_tail;
 221
 222                         new_tail = quarantine_tail + 1;
 223                         if (new_tail == QUARANTINE_BATCHES)
 224                                 new_tail = 0;
 225                         if (new_tail != quarantine_head)
 226                                 quarantine_tail = new_tail;
 227                 }
 228                 raw_spin_unlock(&quarantine_lock);
 229         }
 230
 231         local_irq_restore(flags);
 232
 233         return true;
 234 }
 235
 236 void kasan_quarantine_reduce(void)
 237 {
 238         size_t total_size, new_quarantine_size, percpu_quarantines;
 239         unsigned long flags;
 240         int srcu_idx;
 241         struct qlist_head to_free = QLIST_INIT;
 242
 243         if (likely(READ_ONCE(quarantine_size) <=
 244                    READ_ONCE(quarantine_max_size)))
 245                 return;
 246
 247         /*
 248          * srcu critical section ensures that kasan_quarantine_remove_cache()
 249          * will not miss objects belonging to the cache while they are in our
 250          * local to_free list. srcu is chosen because (1) it gives us private
 251          * grace period domain that does not interfere with anything else,
 252          * and (2) it allows synchronize_srcu() to return without waiting
 253          * if there are no pending read critical sections (which is the
 254          * expected case).
 255          */
 256         srcu_idx = srcu_read_lock(&remove_cache_srcu);
 257         raw_spin_lock_irqsave(&quarantine_lock, flags);
 258
 259         /*
 260          * Update quarantine size in case of hotplug. Allocate a fraction of
 261          * the installed memory to quarantine minus per-cpu queue limits.
 262          */
 263         total_size = (totalram_pages() << PAGE_SHIFT) /
 264                 QUARANTINE_FRACTION;
 265         percpu_quarantines = QUARANTINE_PERCPU_SIZE * num_online_cpus();
 266         new_quarantine_size = (total_size < percpu_quarantines) ?
 267                 0 : total_size - percpu_quarantines;
 268         WRITE_ONCE(quarantine_max_size, new_quarantine_size);
 269         /* Aim at consuming at most 1/2 of slots in quarantine. */
 270         WRITE_ONCE(quarantine_batch_size, max((size_t)QUARANTINE_PERCPU_SIZE,
 271                 2 * total_size / QUARANTINE_BATCHES));
 272
 273         if (likely(quarantine_size > quarantine_max_size)) {
 274                 qlist_move_all(&global_quarantine[quarantine_head], &to_free);
 275                 WRITE_ONCE(quarantine_size, quarantine_size - to_free.bytes);
 276                 quarantine_head++;
 277                 if (quarantine_head == QUARANTINE_BATCHES)
 278                         quarantine_head = 0;
 279         }
 280
 281         raw_spin_unlock_irqrestore(&quarantine_lock, flags);
 282
 283         qlist_free_all(&to_free, NULL);
 284         srcu_read_unlock(&remove_cache_srcu, srcu_idx);
 285 }
 286
 287 static void qlist_move_cache(struct qlist_head *from,
 288                                    struct qlist_head *to,
 289                                    struct kmem_cache *cache)
 290 {
 291         struct qlist_node *curr;
 292
 293         if (unlikely(qlist_empty(from)))
 294                 return;
 295
 296         curr = from->head;
 297         qlist_init(from);
 298         while (curr) {
 299                 struct qlist_node *next = curr->next;
 300                 struct kmem_cache *obj_cache = qlink_to_cache(curr);
 301
 302                 if (obj_cache == cache)
 303                         qlist_put(to, curr, obj_cache->size);
 304                 else
 305                         qlist_put(from, curr, obj_cache->size);
 306
 307                 curr = next;
 308         }
 309 }
 310
 311 static void per_cpu_remove_cache(void *arg)
 312 {
 313         struct kmem_cache *cache = arg;
 314         struct qlist_head to_free = QLIST_INIT;
 315         struct qlist_head *q;
 316
 317         q = this_cpu_ptr(&cpu_quarantine);
 318         /*
 319          * Ensure the ordering between the writing to q->offline and
 320          * per_cpu_remove_cache.  Prevent cpu_quarantine from being corrupted
 321          * by interrupt.
 322          */
 323         if (READ_ONCE(q->offline))
 324                 return;
 325         qlist_move_cache(q, &to_free, cache);
 326         qlist_free_all(&to_free, cache);
 327 }
 328
 329 /* Free all quarantined objects belonging to cache. */
 330 void kasan_quarantine_remove_cache(struct kmem_cache *cache)
 331 {
 332         unsigned long flags, i;
 333         struct qlist_head to_free = QLIST_INIT;
 334
 335         /*
 336          * Must be careful to not miss any objects that are being moved from
 337          * per-cpu list to the global quarantine in kasan_quarantine_put(),
 338          * nor objects being freed in kasan_quarantine_reduce(). on_each_cpu()
 339          * achieves the first goal, while synchronize_srcu() achieves the
 340          * second.
 341          */
 342         on_each_cpu(per_cpu_remove_cache, cache, 1);
 343
 344         raw_spin_lock_irqsave(&quarantine_lock, flags);
 345         for (i = 0; i < QUARANTINE_BATCHES; i++) {
 346                 if (qlist_empty(&global_quarantine[i]))
 347                         continue;
 348                 qlist_move_cache(&global_quarantine[i], &to_free, cache);
 349                 /* Scanning whole quarantine can take a while. */
 350                 raw_spin_unlock_irqrestore(&quarantine_lock, flags);
 351                 cond_resched();
 352                 raw_spin_lock_irqsave(&quarantine_lock, flags);
 353         }
 354         raw_spin_unlock_irqrestore(&quarantine_lock, flags);
 355
 356         qlist_free_all(&to_free, cache);
 357
 358         synchronize_srcu(&remove_cache_srcu);
 359 }
 360
 361 static int kasan_cpu_online(unsigned int cpu)
 362 {
 363         this_cpu_ptr(&cpu_quarantine)->offline = false;
 364         return 0;
 365 }
 366
 367 static int kasan_cpu_offline(unsigned int cpu)
 368 {
 369         struct qlist_head *q;
 370
 371         q = this_cpu_ptr(&cpu_quarantine);
 372         /* Ensure the ordering between the writing to q->offline and
 373          * qlist_free_all. Otherwise, cpu_quarantine may be corrupted
 374          * by interrupt.
 375          */
 376         WRITE_ONCE(q->offline, true);
 377         barrier();
 378         qlist_free_all(q, NULL);
 379         return 0;
 380 }
 381
 382 static int __init kasan_cpu_quarantine_init(void)
 383 {
 384         int ret = 0;
 385
 386         ret = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "mm/kasan:online",
 387                                 kasan_cpu_online, kasan_cpu_offline);
 388         if (ret < 0)
 389                 pr_err("kasan cpu quarantine register failed [%d]\n", ret);
 390         return ret;
 391 }
 392 late_initcall(kasan_cpu_quarantine_init);