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 signle-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_head_page(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         unsigned long flags;
 136
 137         if (IS_ENABLED(CONFIG_SLAB))
 138                 local_irq_save(flags);
 139
 140         *(u8 *)kasan_mem_to_shadow(object) = KASAN_KMALLOC_FREE;
 141         ___cache_free(cache, object, _THIS_IP_);
 142
 143         if (IS_ENABLED(CONFIG_SLAB))
 144                 local_irq_restore(flags);
 145 }
 146
 147 static void qlist_free_all(struct qlist_head *q, struct kmem_cache *cache)
 148 {
 149         struct qlist_node *qlink;
 150
 151         if (unlikely(qlist_empty(q)))
 152                 return;
 153
 154         qlink = q->head;
 155         while (qlink) {
 156                 struct kmem_cache *obj_cache =
 157                         cache ? cache : qlink_to_cache(qlink);
 158                 struct qlist_node *next = qlink->next;
 159
 160                 qlink_free(qlink, obj_cache);
 161                 qlink = next;
 162         }
 163         qlist_init(q);
 164 }
 165
 166 void quarantine_put(struct kmem_cache *cache, void *object)
 167 {
 168         unsigned long flags;
 169         struct qlist_head *q;
 170         struct qlist_head temp = QLIST_INIT;
 171         struct kasan_free_meta *info = get_free_info(cache, object);
 172
 173         /*
 174          * Note: irq must be disabled until after we move the batch to the
 175          * global quarantine. Otherwise quarantine_remove_cache() can miss
 176          * some objects belonging to the cache if they are in our local temp
 177          * list. quarantine_remove_cache() executes on_each_cpu() at the
 178          * beginning which ensures that it either sees the objects in per-cpu
 179          * lists or in the global quarantine.
 180          */
 181         local_irq_save(flags);
 182
 183         q = this_cpu_ptr(&cpu_quarantine);
 184         if (q->offline) {
 185                 local_irq_restore(flags);
 186                 return;
 187         }
 188         qlist_put(q, &info->quarantine_link, cache->size);
 189         if (unlikely(q->bytes > QUARANTINE_PERCPU_SIZE)) {
 190                 qlist_move_all(q, &temp);
 191
 192                 raw_spin_lock(&quarantine_lock);
 193                 WRITE_ONCE(quarantine_size, quarantine_size + temp.bytes);
 194                 qlist_move_all(&temp, &global_quarantine[quarantine_tail]);
 195                 if (global_quarantine[quarantine_tail].bytes >=
 196                                 READ_ONCE(quarantine_batch_size)) {
 197                         int new_tail;
 198
 199                         new_tail = quarantine_tail + 1;
 200                         if (new_tail == QUARANTINE_BATCHES)
 201                                 new_tail = 0;
 202                         if (new_tail != quarantine_head)
 203                                 quarantine_tail = new_tail;
 204                 }
 205                 raw_spin_unlock(&quarantine_lock);
 206         }
 207
 208         local_irq_restore(flags);
 209 }
 210
 211 void quarantine_reduce(void)
 212 {
 213         size_t total_size, new_quarantine_size, percpu_quarantines;
 214         unsigned long flags;
 215         int srcu_idx;
 216         struct qlist_head to_free = QLIST_INIT;
 217
 218         if (likely(READ_ONCE(quarantine_size) <=
 219                    READ_ONCE(quarantine_max_size)))
 220                 return;
 221
 222         /*
 223          * srcu critical section ensures that quarantine_remove_cache()
 224          * will not miss objects belonging to the cache while they are in our
 225          * local to_free list. srcu is chosen because (1) it gives us private
 226          * grace period domain that does not interfere with anything else,
 227          * and (2) it allows synchronize_srcu() to return without waiting
 228          * if there are no pending read critical sections (which is the
 229          * expected case).
 230          */
 231         srcu_idx = srcu_read_lock(&remove_cache_srcu);
 232         raw_spin_lock_irqsave(&quarantine_lock, flags);
 233
 234         /*
 235          * Update quarantine size in case of hotplug. Allocate a fraction of
 236          * the installed memory to quarantine minus per-cpu queue limits.
 237          */
 238         total_size = (totalram_pages() << PAGE_SHIFT) /
 239                 QUARANTINE_FRACTION;
 240         percpu_quarantines = QUARANTINE_PERCPU_SIZE * num_online_cpus();
 241         new_quarantine_size = (total_size < percpu_quarantines) ?
 242                 0 : total_size - percpu_quarantines;
 243         WRITE_ONCE(quarantine_max_size, new_quarantine_size);
 244         /* Aim at consuming at most 1/2 of slots in quarantine. */
 245         WRITE_ONCE(quarantine_batch_size, max((size_t)QUARANTINE_PERCPU_SIZE,
 246                 2 * total_size / QUARANTINE_BATCHES));
 247
 248         if (likely(quarantine_size > quarantine_max_size)) {
 249                 qlist_move_all(&global_quarantine[quarantine_head], &to_free);
 250                 WRITE_ONCE(quarantine_size, quarantine_size - to_free.bytes);
 251                 quarantine_head++;
 252                 if (quarantine_head == QUARANTINE_BATCHES)
 253                         quarantine_head = 0;
 254         }
 255
 256         raw_spin_unlock_irqrestore(&quarantine_lock, flags);
 257
 258         qlist_free_all(&to_free, NULL);
 259         srcu_read_unlock(&remove_cache_srcu, srcu_idx);
 260 }
 261
 262 static void qlist_move_cache(struct qlist_head *from,
 263                                    struct qlist_head *to,
 264                                    struct kmem_cache *cache)
 265 {
 266         struct qlist_node *curr;
 267
 268         if (unlikely(qlist_empty(from)))
 269                 return;
 270
 271         curr = from->head;
 272         qlist_init(from);
 273         while (curr) {
 274                 struct qlist_node *next = curr->next;
 275                 struct kmem_cache *obj_cache = qlink_to_cache(curr);
 276
 277                 if (obj_cache == cache)
 278                         qlist_put(to, curr, obj_cache->size);
 279                 else
 280                         qlist_put(from, curr, obj_cache->size);
 281
 282                 curr = next;
 283         }
 284 }
 285
 286 static void per_cpu_remove_cache(void *arg)
 287 {
 288         struct kmem_cache *cache = arg;
 289         struct qlist_head to_free = QLIST_INIT;
 290         struct qlist_head *q;
 291
 292         q = this_cpu_ptr(&cpu_quarantine);
 293         qlist_move_cache(q, &to_free, cache);
 294         qlist_free_all(&to_free, cache);
 295 }
 296
 297 /* Free all quarantined objects belonging to cache. */
 298 void quarantine_remove_cache(struct kmem_cache *cache)
 299 {
 300         unsigned long flags, i;
 301         struct qlist_head to_free = QLIST_INIT;
 302
 303         /*
 304          * Must be careful to not miss any objects that are being moved from
 305          * per-cpu list to the global quarantine in quarantine_put(),
 306          * nor objects being freed in quarantine_reduce(). on_each_cpu()
 307          * achieves the first goal, while synchronize_srcu() achieves the
 308          * second.
 309          */
 310         on_each_cpu(per_cpu_remove_cache, cache, 1);
 311
 312         raw_spin_lock_irqsave(&quarantine_lock, flags);
 313         for (i = 0; i < QUARANTINE_BATCHES; i++) {
 314                 if (qlist_empty(&global_quarantine[i]))
 315                         continue;
 316                 qlist_move_cache(&global_quarantine[i], &to_free, cache);
 317                 /* Scanning whole quarantine can take a while. */
 318                 raw_spin_unlock_irqrestore(&quarantine_lock, flags);
 319                 cond_resched();
 320                 raw_spin_lock_irqsave(&quarantine_lock, flags);
 321         }
 322         raw_spin_unlock_irqrestore(&quarantine_lock, flags);
 323
 324         qlist_free_all(&to_free, cache);
 325
 326         synchronize_srcu(&remove_cache_srcu);
 327 }
 328
 329 static int kasan_cpu_online(unsigned int cpu)
 330 {
 331         this_cpu_ptr(&cpu_quarantine)->offline = false;
 332         return 0;
 333 }
 334
 335 static int kasan_cpu_offline(unsigned int cpu)
 336 {
 337         struct qlist_head *q;
 338
 339         q = this_cpu_ptr(&cpu_quarantine);
 340         /* Ensure the ordering between the writing to q->offline and
 341          * qlist_free_all. Otherwise, cpu_quarantine may be corrupted
 342          * by interrupt.
 343          */
 344         WRITE_ONCE(q->offline, true);
 345         barrier();
 346         qlist_free_all(q, NULL);
 347         return 0;
 348 }
 349
 350 static int __init kasan_cpu_quarantine_init(void)
 351 {
 352         int ret = 0;
 353
 354         ret = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "mm/kasan:online",
 355                                 kasan_cpu_online, kasan_cpu_offline);
 356         if (ret < 0)
 357                 pr_err("kasan cpu quarantine register failed [%d]\n", ret);
 358         return ret;
 359 }
 360 late_initcall(kasan_cpu_quarantine_init);