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#define pr_fmt(fmt) "kasan: " fmt
12
13#include <linux/gfp.h>
14#include <linux/hash.h>
15#include <linux/kernel.h>
16#include <linux/mm.h>
17#include <linux/percpu.h>
18#include <linux/printk.h>
19#include <linux/shrinker.h>
20#include <linux/slab.h>
21#include <linux/srcu.h>
22#include <linux/string.h>
23#include <linux/types.h>
24#include <linux/cpuhotplug.h>
25
26#include "../slab.h"
27#include "kasan.h"
28
29/* Data structure and operations for quarantine queues. */
30
31/*
32 * Each queue is a single-linked list, which also stores the total size of
33 * objects inside of it.
34 */
35struct qlist_head {
36	struct qlist_node *head;
37	struct qlist_node *tail;
38	size_t bytes;
39	bool offline;
40};
41
42#define QLIST_INIT { NULL, NULL, 0 }
43
44static bool qlist_empty(struct qlist_head *q)
45{
46	return !q->head;
47}
48
49static void qlist_init(struct qlist_head *q)
50{
51	q->head = q->tail = NULL;
52	q->bytes = 0;
53}
54
55static void qlist_put(struct qlist_head *q, struct qlist_node *qlink,
56		size_t size)
57{
58	if (unlikely(qlist_empty(q)))
59		q->head = qlink;
60	else
61		q->tail->next = qlink;
62	q->tail = qlink;
63	qlink->next = NULL;
64	q->bytes += size;
65}
66
67static void qlist_move_all(struct qlist_head *from, struct qlist_head *to)
68{
69	if (unlikely(qlist_empty(from)))
70		return;
71
72	if (qlist_empty(to)) {
73		*to = *from;
74		qlist_init(from);
75		return;
76	}
77
78	to->tail->next = from->head;
79	to->tail = from->tail;
80	to->bytes += from->bytes;
81
82	qlist_init(from);
83}
84
85#define QUARANTINE_PERCPU_SIZE (1 << 20)
86#define QUARANTINE_BATCHES \
87	(1024 > 4 * CONFIG_NR_CPUS ? 1024 : 4 * CONFIG_NR_CPUS)
88
89/*
90 * The object quarantine consists of per-cpu queues and a global queue,
91 * guarded by quarantine_lock.
92 */
93static DEFINE_PER_CPU(struct qlist_head, cpu_quarantine);
94
95/* Round-robin FIFO array of batches. */
96static struct qlist_head global_quarantine[QUARANTINE_BATCHES];
97static int quarantine_head;
98static int quarantine_tail;
99/* Total size of all objects in global_quarantine across all batches. */
100static unsigned long quarantine_size;
101static DEFINE_RAW_SPINLOCK(quarantine_lock);
102DEFINE_STATIC_SRCU(remove_cache_srcu);
103
104struct cpu_shrink_qlist {
105	raw_spinlock_t lock;
106	struct qlist_head qlist;
107};
108
109static DEFINE_PER_CPU(struct cpu_shrink_qlist, shrink_qlist) = {
110	.lock = __RAW_SPIN_LOCK_UNLOCKED(shrink_qlist.lock),
111};
112
113/* Maximum size of the global queue. */
114static 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 */
120static 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
129static struct kmem_cache *qlink_to_cache(struct qlist_node *qlink)
130{
131	return virt_to_slab(qlink)->slab_cache;
132}
133
134static 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
143static 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 *free_meta = kasan_get_free_meta(cache, object);
147
148	/*
149	 * Note: Keep per-object metadata to allow KASAN print stack traces for
150	 * use-after-free-before-realloc bugs.
151	 */
152
153	/*
154	 * If init_on_free is enabled and KASAN's free metadata is stored in
155	 * the object, zero the metadata. Otherwise, the object's memory will
156	 * not be properly zeroed, as KASAN saves the metadata after the slab
157	 * allocator zeroes the object.
158	 */
159	if (slab_want_init_on_free(cache) &&
160	    cache->kasan_info.free_meta_offset == 0)
161		memzero_explicit(free_meta, sizeof(*free_meta));
162
163	___cache_free(cache, object, _THIS_IP_);
164}
165
166static void qlist_free_all(struct qlist_head *q, struct kmem_cache *cache)
167{
168	struct qlist_node *qlink;
169
170	if (unlikely(qlist_empty(q)))
171		return;
172
173	qlink = q->head;
174	while (qlink) {
175		struct kmem_cache *obj_cache =
176			cache ? cache :	qlink_to_cache(qlink);
177		struct qlist_node *next = qlink->next;
178
179		qlink_free(qlink, obj_cache);
180		qlink = next;
181	}
182	qlist_init(q);
183}
184
185bool kasan_quarantine_put(struct kmem_cache *cache, void *object)
186{
187	unsigned long flags;
188	struct qlist_head *q;
189	struct qlist_head temp = QLIST_INIT;
190	struct kasan_free_meta *meta = kasan_get_free_meta(cache, object);
191
192	/*
193	 * If there's no metadata for this object, don't put it into
194	 * quarantine.
195	 */
196	if (!meta)
197		return false;
198
199	/*
200	 * Note: irq must be disabled until after we move the batch to the
201	 * global quarantine. Otherwise kasan_quarantine_remove_cache() can
202	 * miss some objects belonging to the cache if they are in our local
203	 * temp list. kasan_quarantine_remove_cache() executes on_each_cpu()
204	 * at the beginning which ensures that it either sees the objects in
205	 * per-cpu lists or in the global quarantine.
206	 */
207	local_irq_save(flags);
208
209	q = this_cpu_ptr(&cpu_quarantine);
210	if (q->offline) {
211		local_irq_restore(flags);
212		return false;
213	}
214	qlist_put(q, &meta->quarantine_link, cache->size);
215	if (unlikely(q->bytes > QUARANTINE_PERCPU_SIZE)) {
216		qlist_move_all(q, &temp);
217
218		raw_spin_lock(&quarantine_lock);
219		WRITE_ONCE(quarantine_size, quarantine_size + temp.bytes);
220		qlist_move_all(&temp, &global_quarantine[quarantine_tail]);
221		if (global_quarantine[quarantine_tail].bytes >=
222				READ_ONCE(quarantine_batch_size)) {
223			int new_tail;
224
225			new_tail = quarantine_tail + 1;
226			if (new_tail == QUARANTINE_BATCHES)
227				new_tail = 0;
228			if (new_tail != quarantine_head)
229				quarantine_tail = new_tail;
230		}
231		raw_spin_unlock(&quarantine_lock);
232	}
233
234	local_irq_restore(flags);
235
236	return true;
237}
238
239void kasan_quarantine_reduce(void)
240{
241	size_t total_size, new_quarantine_size, percpu_quarantines;
242	unsigned long flags;
243	int srcu_idx;
244	struct qlist_head to_free = QLIST_INIT;
245
246	if (likely(READ_ONCE(quarantine_size) <=
247		   READ_ONCE(quarantine_max_size)))
248		return;
249
250	/*
251	 * srcu critical section ensures that kasan_quarantine_remove_cache()
252	 * will not miss objects belonging to the cache while they are in our
253	 * local to_free list. srcu is chosen because (1) it gives us private
254	 * grace period domain that does not interfere with anything else,
255	 * and (2) it allows synchronize_srcu() to return without waiting
256	 * if there are no pending read critical sections (which is the
257	 * expected case).
258	 */
259	srcu_idx = srcu_read_lock(&remove_cache_srcu);
260	raw_spin_lock_irqsave(&quarantine_lock, flags);
261
262	/*
263	 * Update quarantine size in case of hotplug. Allocate a fraction of
264	 * the installed memory to quarantine minus per-cpu queue limits.
265	 */
266	total_size = (totalram_pages() << PAGE_SHIFT) /
267		QUARANTINE_FRACTION;
268	percpu_quarantines = QUARANTINE_PERCPU_SIZE * num_online_cpus();
269	new_quarantine_size = (total_size < percpu_quarantines) ?
270		0 : total_size - percpu_quarantines;
271	WRITE_ONCE(quarantine_max_size, new_quarantine_size);
272	/* Aim at consuming at most 1/2 of slots in quarantine. */
273	WRITE_ONCE(quarantine_batch_size, max((size_t)QUARANTINE_PERCPU_SIZE,
274		2 * total_size / QUARANTINE_BATCHES));
275
276	if (likely(quarantine_size > quarantine_max_size)) {
277		qlist_move_all(&global_quarantine[quarantine_head], &to_free);
278		WRITE_ONCE(quarantine_size, quarantine_size - to_free.bytes);
279		quarantine_head++;
280		if (quarantine_head == QUARANTINE_BATCHES)
281			quarantine_head = 0;
282	}
283
284	raw_spin_unlock_irqrestore(&quarantine_lock, flags);
285
286	qlist_free_all(&to_free, NULL);
287	srcu_read_unlock(&remove_cache_srcu, srcu_idx);
288}
289
290static void qlist_move_cache(struct qlist_head *from,
291				   struct qlist_head *to,
292				   struct kmem_cache *cache)
293{
294	struct qlist_node *curr;
295
296	if (unlikely(qlist_empty(from)))
297		return;
298
299	curr = from->head;
300	qlist_init(from);
301	while (curr) {
302		struct qlist_node *next = curr->next;
303		struct kmem_cache *obj_cache = qlink_to_cache(curr);
304
305		if (obj_cache == cache)
306			qlist_put(to, curr, obj_cache->size);
307		else
308			qlist_put(from, curr, obj_cache->size);
309
310		curr = next;
311	}
312}
313
314static void __per_cpu_remove_cache(struct qlist_head *q, void *arg)
315{
316	struct kmem_cache *cache = arg;
317	unsigned long flags;
318	struct cpu_shrink_qlist *sq;
319
320	sq = this_cpu_ptr(&shrink_qlist);
321	raw_spin_lock_irqsave(&sq->lock, flags);
322	qlist_move_cache(q, &sq->qlist, cache);
323	raw_spin_unlock_irqrestore(&sq->lock, flags);
324}
325
326static void per_cpu_remove_cache(void *arg)
327{
328	struct qlist_head *q;
329
330	q = this_cpu_ptr(&cpu_quarantine);
331	/*
332	 * Ensure the ordering between the writing to q->offline and
333	 * per_cpu_remove_cache.  Prevent cpu_quarantine from being corrupted
334	 * by interrupt.
335	 */
336	if (READ_ONCE(q->offline))
337		return;
338	__per_cpu_remove_cache(q, arg);
339}
340
341/* Free all quarantined objects belonging to cache. */
342void kasan_quarantine_remove_cache(struct kmem_cache *cache)
343{
344	unsigned long flags, i;
345	struct qlist_head to_free = QLIST_INIT;
346	int cpu;
347	struct cpu_shrink_qlist *sq;
348
349	/*
350	 * Must be careful to not miss any objects that are being moved from
351	 * per-cpu list to the global quarantine in kasan_quarantine_put(),
352	 * nor objects being freed in kasan_quarantine_reduce(). on_each_cpu()
353	 * achieves the first goal, while synchronize_srcu() achieves the
354	 * second.
355	 */
356	on_each_cpu(per_cpu_remove_cache, cache, 1);
357
358	for_each_online_cpu(cpu) {
359		sq = per_cpu_ptr(&shrink_qlist, cpu);
360		raw_spin_lock_irqsave(&sq->lock, flags);
361		qlist_move_cache(&sq->qlist, &to_free, cache);
362		raw_spin_unlock_irqrestore(&sq->lock, flags);
363	}
364	qlist_free_all(&to_free, cache);
365
366	raw_spin_lock_irqsave(&quarantine_lock, flags);
367	for (i = 0; i < QUARANTINE_BATCHES; i++) {
368		if (qlist_empty(&global_quarantine[i]))
369			continue;
370		qlist_move_cache(&global_quarantine[i], &to_free, cache);
371		/* Scanning whole quarantine can take a while. */
372		raw_spin_unlock_irqrestore(&quarantine_lock, flags);
373		cond_resched();
374		raw_spin_lock_irqsave(&quarantine_lock, flags);
375	}
376	raw_spin_unlock_irqrestore(&quarantine_lock, flags);
377
378	qlist_free_all(&to_free, cache);
379
380	synchronize_srcu(&remove_cache_srcu);
381}
382
383static int kasan_cpu_online(unsigned int cpu)
384{
385	this_cpu_ptr(&cpu_quarantine)->offline = false;
386	return 0;
387}
388
389static int kasan_cpu_offline(unsigned int cpu)
390{
391	struct qlist_head *q;
392
393	q = this_cpu_ptr(&cpu_quarantine);
394	/* Ensure the ordering between the writing to q->offline and
395	 * qlist_free_all. Otherwise, cpu_quarantine may be corrupted
396	 * by interrupt.
397	 */
398	WRITE_ONCE(q->offline, true);
399	barrier();
400	qlist_free_all(q, NULL);
401	return 0;
402}
403
404static int __init kasan_cpu_quarantine_init(void)
405{
406	int ret = 0;
407
408	ret = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "mm/kasan:online",
409				kasan_cpu_online, kasan_cpu_offline);
410	if (ret < 0)
411		pr_err("cpu quarantine register failed [%d]\n", ret);
412	return ret;
413}
414late_initcall(kasan_cpu_quarantine_init);
415