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 #ifdef CONFIG_PREEMPT_RT
 103 struct cpu_shrink_qlist {
 104         raw_spinlock_t lock;
 105         struct qlist_head qlist;
 106 };
 107
 108 static DEFINE_PER_CPU(struct cpu_shrink_qlist, shrink_qlist) = {
 109         .lock = __RAW_SPIN_LOCK_UNLOCKED(shrink_qlist.lock),
 110 };
 111 #endif
 112
 113 /* Maximum size of the global queue. */
 114 static unsigned long quarantine_max_size;
 115
 116 /*
 117  * Target size of a batch in global_quarantine.
 118  * Usually equal to QUARANTINE_PERCPU_SIZE unless we have too much RAM.
 119  */
 120 static unsigned long quarantine_batch_size;
 121
 122 /*
 123  * The fraction of physical memory the quarantine is allowed to occupy.
 124  * Quarantine doesn't support memory shrinker with SLAB allocator, so we keep
 125  * the ratio low to avoid OOM.
 126  */
 127 #define QUARANTINE_FRACTION 32
 128
 129 static struct kmem_cache *qlink_to_cache(struct qlist_node *qlink)
 130 {
 131         return virt_to_slab(qlink)->slab_cache;
 132 }
 133
 134 static void *qlink_to_object(struct qlist_node *qlink, struct kmem_cache *cache)
 135 {
 136         struct kasan_free_meta *free_info =
 137                 container_of(qlink, struct kasan_free_meta,
 138                              quarantine_link);
 139
 140         return ((void *)free_info) - cache->kasan_info.free_meta_offset;
 141 }
 142
 143 static void qlink_free(struct qlist_node *qlink, struct kmem_cache *cache)
 144 {
 145         void *object = qlink_to_object(qlink, cache);
 146         struct kasan_free_meta *meta = kasan_get_free_meta(cache, object);
 147         unsigned long flags;
 148
 149         if (IS_ENABLED(CONFIG_SLAB))
 150                 local_irq_save(flags);
 151
 152         /*
 153          * If init_on_free is enabled and KASAN's free metadata is stored in
 154          * the object, zero the metadata. Otherwise, the object's memory will
 155          * not be properly zeroed, as KASAN saves the metadata after the slab
 156          * allocator zeroes the object.
 157          */
 158         if (slab_want_init_on_free(cache) &&
 159             cache->kasan_info.free_meta_offset == 0)
 160                 memzero_explicit(meta, sizeof(*meta));
 161
 162         /*
 163          * As the object now gets freed from the quarantine, assume that its
 164          * free track is no longer valid.
 165          */
 166         *(u8 *)kasan_mem_to_shadow(object) = KASAN_SLAB_FREE;
 167
 168         ___cache_free(cache, object, _THIS_IP_);
 169
 170         if (IS_ENABLED(CONFIG_SLAB))
 171                 local_irq_restore(flags);
 172 }
 173
 174 static void qlist_free_all(struct qlist_head *q, struct kmem_cache *cache)
 175 {
 176         struct qlist_node *qlink;
 177
 178         if (unlikely(qlist_empty(q)))
 179                 return;
 180
 181         qlink = q->head;
 182         while (qlink) {
 183                 struct kmem_cache *obj_cache =
 184                         cache ? cache : qlink_to_cache(qlink);
 185                 struct qlist_node *next = qlink->next;
 186
 187                 qlink_free(qlink, obj_cache);
 188                 qlink = next;
 189         }
 190         qlist_init(q);
 191 }
 192
 193 bool kasan_quarantine_put(struct kmem_cache *cache, void *object)
 194 {
 195         unsigned long flags;
 196         struct qlist_head *q;
 197         struct qlist_head temp = QLIST_INIT;
 198         struct kasan_free_meta *meta = kasan_get_free_meta(cache, object);
 199
 200         /*
 201          * If there's no metadata for this object, don't put it into
 202          * quarantine.
 203          */
 204         if (!meta)
 205                 return false;
 206
 207         /*
 208          * Note: irq must be disabled until after we move the batch to the
 209          * global quarantine. Otherwise kasan_quarantine_remove_cache() can
 210          * miss some objects belonging to the cache if they are in our local
 211          * temp list. kasan_quarantine_remove_cache() executes on_each_cpu()
 212          * at the beginning which ensures that it either sees the objects in
 213          * per-cpu lists or in the global quarantine.
 214          */
 215         local_irq_save(flags);
 216
 217         q = this_cpu_ptr(&cpu_quarantine);
 218         if (q->offline) {
 219                 local_irq_restore(flags);
 220                 return false;
 221         }
 222         qlist_put(q, &meta->quarantine_link, cache->size);
 223         if (unlikely(q->bytes > QUARANTINE_PERCPU_SIZE)) {
 224                 qlist_move_all(q, &temp);
 225
 226                 raw_spin_lock(&quarantine_lock);
 227                 WRITE_ONCE(quarantine_size, quarantine_size + temp.bytes);
 228                 qlist_move_all(&temp, &global_quarantine[quarantine_tail]);
 229                 if (global_quarantine[quarantine_tail].bytes >=
 230                                 READ_ONCE(quarantine_batch_size)) {
 231                         int new_tail;
 232
 233                         new_tail = quarantine_tail + 1;
 234                         if (new_tail == QUARANTINE_BATCHES)
 235                                 new_tail = 0;
 236                         if (new_tail != quarantine_head)
 237                                 quarantine_tail = new_tail;
 238                 }
 239                 raw_spin_unlock(&quarantine_lock);
 240         }
 241
 242         local_irq_restore(flags);
 243
 244         return true;
 245 }
 246
 247 void kasan_quarantine_reduce(void)
 248 {
 249         size_t total_size, new_quarantine_size, percpu_quarantines;
 250         unsigned long flags;
 251         int srcu_idx;
 252         struct qlist_head to_free = QLIST_INIT;
 253
 254         if (likely(READ_ONCE(quarantine_size) <=
 255                    READ_ONCE(quarantine_max_size)))
 256                 return;
 257
 258         /*
 259          * srcu critical section ensures that kasan_quarantine_remove_cache()
 260          * will not miss objects belonging to the cache while they are in our
 261          * local to_free list. srcu is chosen because (1) it gives us private
 262          * grace period domain that does not interfere with anything else,
 263          * and (2) it allows synchronize_srcu() to return without waiting
 264          * if there are no pending read critical sections (which is the
 265          * expected case).
 266          */
 267         srcu_idx = srcu_read_lock(&remove_cache_srcu);
 268         raw_spin_lock_irqsave(&quarantine_lock, flags);
 269
 270         /*
 271          * Update quarantine size in case of hotplug. Allocate a fraction of
 272          * the installed memory to quarantine minus per-cpu queue limits.
 273          */
 274         total_size = (totalram_pages() << PAGE_SHIFT) /
 275                 QUARANTINE_FRACTION;
 276         percpu_quarantines = QUARANTINE_PERCPU_SIZE * num_online_cpus();
 277         new_quarantine_size = (total_size < percpu_quarantines) ?
 278                 0 : total_size - percpu_quarantines;
 279         WRITE_ONCE(quarantine_max_size, new_quarantine_size);
 280         /* Aim at consuming at most 1/2 of slots in quarantine. */
 281         WRITE_ONCE(quarantine_batch_size, max((size_t)QUARANTINE_PERCPU_SIZE,
 282                 2 * total_size / QUARANTINE_BATCHES));
 283
 284         if (likely(quarantine_size > quarantine_max_size)) {
 285                 qlist_move_all(&global_quarantine[quarantine_head], &to_free);
 286                 WRITE_ONCE(quarantine_size, quarantine_size - to_free.bytes);
 287                 quarantine_head++;
 288                 if (quarantine_head == QUARANTINE_BATCHES)
 289                         quarantine_head = 0;
 290         }
 291
 292         raw_spin_unlock_irqrestore(&quarantine_lock, flags);
 293
 294         qlist_free_all(&to_free, NULL);
 295         srcu_read_unlock(&remove_cache_srcu, srcu_idx);
 296 }
 297
 298 static void qlist_move_cache(struct qlist_head *from,
 299                                    struct qlist_head *to,
 300                                    struct kmem_cache *cache)
 301 {
 302         struct qlist_node *curr;
 303
 304         if (unlikely(qlist_empty(from)))
 305                 return;
 306
 307         curr = from->head;
 308         qlist_init(from);
 309         while (curr) {
 310                 struct qlist_node *next = curr->next;
 311                 struct kmem_cache *obj_cache = qlink_to_cache(curr);
 312
 313                 if (obj_cache == cache)
 314                         qlist_put(to, curr, obj_cache->size);
 315                 else
 316                         qlist_put(from, curr, obj_cache->size);
 317
 318                 curr = next;
 319         }
 320 }
 321
 322 #ifndef CONFIG_PREEMPT_RT
 323 static void __per_cpu_remove_cache(struct qlist_head *q, void *arg)
 324 {
 325         struct kmem_cache *cache = arg;
 326         struct qlist_head to_free = QLIST_INIT;
 327
 328         qlist_move_cache(q, &to_free, cache);
 329         qlist_free_all(&to_free, cache);
 330 }
 331 #else
 332 static void __per_cpu_remove_cache(struct qlist_head *q, void *arg)
 333 {
 334         struct kmem_cache *cache = arg;
 335         unsigned long flags;
 336         struct cpu_shrink_qlist *sq;
 337
 338         sq = this_cpu_ptr(&shrink_qlist);
 339         raw_spin_lock_irqsave(&sq->lock, flags);
 340         qlist_move_cache(q, &sq->qlist, cache);
 341         raw_spin_unlock_irqrestore(&sq->lock, flags);
 342 }
 343 #endif
 344
 345 static void per_cpu_remove_cache(void *arg)
 346 {
 347         struct qlist_head *q;
 348
 349         q = this_cpu_ptr(&cpu_quarantine);
 350         /*
 351          * Ensure the ordering between the writing to q->offline and
 352          * per_cpu_remove_cache.  Prevent cpu_quarantine from being corrupted
 353          * by interrupt.
 354          */
 355         if (READ_ONCE(q->offline))
 356                 return;
 357         __per_cpu_remove_cache(q, arg);
 358 }
 359
 360 /* Free all quarantined objects belonging to cache. */
 361 void kasan_quarantine_remove_cache(struct kmem_cache *cache)
 362 {
 363         unsigned long flags, i;
 364         struct qlist_head to_free = QLIST_INIT;
 365
 366         /*
 367          * Must be careful to not miss any objects that are being moved from
 368          * per-cpu list to the global quarantine in kasan_quarantine_put(),
 369          * nor objects being freed in kasan_quarantine_reduce(). on_each_cpu()
 370          * achieves the first goal, while synchronize_srcu() achieves the
 371          * second.
 372          */
 373         on_each_cpu(per_cpu_remove_cache, cache, 1);
 374
 375 #ifdef CONFIG_PREEMPT_RT
 376         {
 377                 int cpu;
 378                 struct cpu_shrink_qlist *sq;
 379
 380                 for_each_online_cpu(cpu) {
 381                         sq = per_cpu_ptr(&shrink_qlist, cpu);
 382                         raw_spin_lock_irqsave(&sq->lock, flags);
 383                         qlist_move_cache(&sq->qlist, &to_free, cache);
 384                         raw_spin_unlock_irqrestore(&sq->lock, flags);
 385                 }
 386                 qlist_free_all(&to_free, cache);
 387         }
 388 #endif
 389
 390         raw_spin_lock_irqsave(&quarantine_lock, flags);
 391         for (i = 0; i < QUARANTINE_BATCHES; i++) {
 392                 if (qlist_empty(&global_quarantine[i]))
 393                         continue;
 394                 qlist_move_cache(&global_quarantine[i], &to_free, cache);
 395                 /* Scanning whole quarantine can take a while. */
 396                 raw_spin_unlock_irqrestore(&quarantine_lock, flags);
 397                 cond_resched();
 398                 raw_spin_lock_irqsave(&quarantine_lock, flags);
 399         }
 400         raw_spin_unlock_irqrestore(&quarantine_lock, flags);
 401
 402         qlist_free_all(&to_free, cache);
 403
 404         synchronize_srcu(&remove_cache_srcu);
 405 }
 406
 407 static int kasan_cpu_online(unsigned int cpu)
 408 {
 409         this_cpu_ptr(&cpu_quarantine)->offline = false;
 410         return 0;
 411 }
 412
 413 static int kasan_cpu_offline(unsigned int cpu)
 414 {
 415         struct qlist_head *q;
 416
 417         q = this_cpu_ptr(&cpu_quarantine);
 418         /* Ensure the ordering between the writing to q->offline and
 419          * qlist_free_all. Otherwise, cpu_quarantine may be corrupted
 420          * by interrupt.
 421          */
 422         WRITE_ONCE(q->offline, true);
 423         barrier();
 424         qlist_free_all(q, NULL);
 425         return 0;
 426 }
 427
 428 static int __init kasan_cpu_quarantine_init(void)
 429 {
 430         int ret = 0;
 431
 432         ret = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "mm/kasan:online",
 433                                 kasan_cpu_online, kasan_cpu_offline);
 434         if (ret < 0)
 435                 pr_err("kasan cpu quarantine register failed [%d]\n", ret);
 436         return ret;
 437 }
 438 late_initcall(kasan_cpu_quarantine_init);