mm/kasan/quarantine.c

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