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sizeof(*meta)); 159 160 /* 161 * As the object now gets freed from the quarantine, assume that its 162 * free track is no longer valid. 163 */ 164 *(u8 *)kasan_mem_to_shadow(object) = KASAN_SLAB_FREE; 165 166 ___cache_free(cache, object, _THIS_IP_); 167 168 if (IS_ENABLED(CONFIG_SLAB)) 169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