xref: /openbmc/linux/mm/kasan/quarantine.c (revision e5bd61e8)
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_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 	/*
141 	 * As the object now gets freed from the quarantine, assume that its
142 	 * free track is no longer valid.
143 	 */
144 	*(u8 *)kasan_mem_to_shadow(object) = KASAN_KMALLOC_FREE;
145 
146 	___cache_free(cache, object, _THIS_IP_);
147 
148 	if (IS_ENABLED(CONFIG_SLAB))
149 		local_irq_restore(flags);
150 }
151 
152 static void qlist_free_all(struct qlist_head *q, struct kmem_cache *cache)
153 {
154 	struct qlist_node *qlink;
155 
156 	if (unlikely(qlist_empty(q)))
157 		return;
158 
159 	qlink = q->head;
160 	while (qlink) {
161 		struct kmem_cache *obj_cache =
162 			cache ? cache :	qlink_to_cache(qlink);
163 		struct qlist_node *next = qlink->next;
164 
165 		qlink_free(qlink, obj_cache);
166 		qlink = next;
167 	}
168 	qlist_init(q);
169 }
170 
171 bool kasan_quarantine_put(struct kmem_cache *cache, void *object)
172 {
173 	unsigned long flags;
174 	struct qlist_head *q;
175 	struct qlist_head temp = QLIST_INIT;
176 	struct kasan_free_meta *meta = kasan_get_free_meta(cache, object);
177 
178 	/*
179 	 * If there's no metadata for this object, don't put it into
180 	 * quarantine.
181 	 */
182 	if (!meta)
183 		return false;
184 
185 	/*
186 	 * Note: irq must be disabled until after we move the batch to the
187 	 * global quarantine. Otherwise kasan_quarantine_remove_cache() can
188 	 * miss some objects belonging to the cache if they are in our local
189 	 * temp list. kasan_quarantine_remove_cache() executes on_each_cpu()
190 	 * at the beginning which ensures that it either sees the objects in
191 	 * per-cpu lists or in the global quarantine.
192 	 */
193 	local_irq_save(flags);
194 
195 	q = this_cpu_ptr(&cpu_quarantine);
196 	if (q->offline) {
197 		local_irq_restore(flags);
198 		return false;
199 	}
200 	qlist_put(q, &meta->quarantine_link, cache->size);
201 	if (unlikely(q->bytes > QUARANTINE_PERCPU_SIZE)) {
202 		qlist_move_all(q, &temp);
203 
204 		raw_spin_lock(&quarantine_lock);
205 		WRITE_ONCE(quarantine_size, quarantine_size + temp.bytes);
206 		qlist_move_all(&temp, &global_quarantine[quarantine_tail]);
207 		if (global_quarantine[quarantine_tail].bytes >=
208 				READ_ONCE(quarantine_batch_size)) {
209 			int new_tail;
210 
211 			new_tail = quarantine_tail + 1;
212 			if (new_tail == QUARANTINE_BATCHES)
213 				new_tail = 0;
214 			if (new_tail != quarantine_head)
215 				quarantine_tail = new_tail;
216 		}
217 		raw_spin_unlock(&quarantine_lock);
218 	}
219 
220 	local_irq_restore(flags);
221 
222 	return true;
223 }
224 
225 void kasan_quarantine_reduce(void)
226 {
227 	size_t total_size, new_quarantine_size, percpu_quarantines;
228 	unsigned long flags;
229 	int srcu_idx;
230 	struct qlist_head to_free = QLIST_INIT;
231 
232 	if (likely(READ_ONCE(quarantine_size) <=
233 		   READ_ONCE(quarantine_max_size)))
234 		return;
235 
236 	/*
237 	 * srcu critical section ensures that kasan_quarantine_remove_cache()
238 	 * will not miss objects belonging to the cache while they are in our
239 	 * local to_free list. srcu is chosen because (1) it gives us private
240 	 * grace period domain that does not interfere with anything else,
241 	 * and (2) it allows synchronize_srcu() to return without waiting
242 	 * if there are no pending read critical sections (which is the
243 	 * expected case).
244 	 */
245 	srcu_idx = srcu_read_lock(&remove_cache_srcu);
246 	raw_spin_lock_irqsave(&quarantine_lock, flags);
247 
248 	/*
249 	 * Update quarantine size in case of hotplug. Allocate a fraction of
250 	 * the installed memory to quarantine minus per-cpu queue limits.
251 	 */
252 	total_size = (totalram_pages() << PAGE_SHIFT) /
253 		QUARANTINE_FRACTION;
254 	percpu_quarantines = QUARANTINE_PERCPU_SIZE * num_online_cpus();
255 	new_quarantine_size = (total_size < percpu_quarantines) ?
256 		0 : total_size - percpu_quarantines;
257 	WRITE_ONCE(quarantine_max_size, new_quarantine_size);
258 	/* Aim at consuming at most 1/2 of slots in quarantine. */
259 	WRITE_ONCE(quarantine_batch_size, max((size_t)QUARANTINE_PERCPU_SIZE,
260 		2 * total_size / QUARANTINE_BATCHES));
261 
262 	if (likely(quarantine_size > quarantine_max_size)) {
263 		qlist_move_all(&global_quarantine[quarantine_head], &to_free);
264 		WRITE_ONCE(quarantine_size, quarantine_size - to_free.bytes);
265 		quarantine_head++;
266 		if (quarantine_head == QUARANTINE_BATCHES)
267 			quarantine_head = 0;
268 	}
269 
270 	raw_spin_unlock_irqrestore(&quarantine_lock, flags);
271 
272 	qlist_free_all(&to_free, NULL);
273 	srcu_read_unlock(&remove_cache_srcu, srcu_idx);
274 }
275 
276 static void qlist_move_cache(struct qlist_head *from,
277 				   struct qlist_head *to,
278 				   struct kmem_cache *cache)
279 {
280 	struct qlist_node *curr;
281 
282 	if (unlikely(qlist_empty(from)))
283 		return;
284 
285 	curr = from->head;
286 	qlist_init(from);
287 	while (curr) {
288 		struct qlist_node *next = curr->next;
289 		struct kmem_cache *obj_cache = qlink_to_cache(curr);
290 
291 		if (obj_cache == cache)
292 			qlist_put(to, curr, obj_cache->size);
293 		else
294 			qlist_put(from, curr, obj_cache->size);
295 
296 		curr = next;
297 	}
298 }
299 
300 static void per_cpu_remove_cache(void *arg)
301 {
302 	struct kmem_cache *cache = arg;
303 	struct qlist_head to_free = QLIST_INIT;
304 	struct qlist_head *q;
305 
306 	q = this_cpu_ptr(&cpu_quarantine);
307 	qlist_move_cache(q, &to_free, cache);
308 	qlist_free_all(&to_free, cache);
309 }
310 
311 /* Free all quarantined objects belonging to cache. */
312 void kasan_quarantine_remove_cache(struct kmem_cache *cache)
313 {
314 	unsigned long flags, i;
315 	struct qlist_head to_free = QLIST_INIT;
316 
317 	/*
318 	 * Must be careful to not miss any objects that are being moved from
319 	 * per-cpu list to the global quarantine in kasan_quarantine_put(),
320 	 * nor objects being freed in kasan_quarantine_reduce(). on_each_cpu()
321 	 * achieves the first goal, while synchronize_srcu() achieves the
322 	 * second.
323 	 */
324 	on_each_cpu(per_cpu_remove_cache, cache, 1);
325 
326 	raw_spin_lock_irqsave(&quarantine_lock, flags);
327 	for (i = 0; i < QUARANTINE_BATCHES; i++) {
328 		if (qlist_empty(&global_quarantine[i]))
329 			continue;
330 		qlist_move_cache(&global_quarantine[i], &to_free, cache);
331 		/* Scanning whole quarantine can take a while. */
332 		raw_spin_unlock_irqrestore(&quarantine_lock, flags);
333 		cond_resched();
334 		raw_spin_lock_irqsave(&quarantine_lock, flags);
335 	}
336 	raw_spin_unlock_irqrestore(&quarantine_lock, flags);
337 
338 	qlist_free_all(&to_free, cache);
339 
340 	synchronize_srcu(&remove_cache_srcu);
341 }
342 
343 static int kasan_cpu_online(unsigned int cpu)
344 {
345 	this_cpu_ptr(&cpu_quarantine)->offline = false;
346 	return 0;
347 }
348 
349 static int kasan_cpu_offline(unsigned int cpu)
350 {
351 	struct qlist_head *q;
352 
353 	q = this_cpu_ptr(&cpu_quarantine);
354 	/* Ensure the ordering between the writing to q->offline and
355 	 * qlist_free_all. Otherwise, cpu_quarantine may be corrupted
356 	 * by interrupt.
357 	 */
358 	WRITE_ONCE(q->offline, true);
359 	barrier();
360 	qlist_free_all(q, NULL);
361 	return 0;
362 }
363 
364 static int __init kasan_cpu_quarantine_init(void)
365 {
366 	int ret = 0;
367 
368 	ret = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "mm/kasan:online",
369 				kasan_cpu_online, kasan_cpu_offline);
370 	if (ret < 0)
371 		pr_err("kasan cpu quarantine register failed [%d]\n", ret);
372 	return ret;
373 }
374 late_initcall(kasan_cpu_quarantine_init);
375