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 * This program is free software; you can redistribute it and/or
 11 * modify it under the terms of the GNU General Public License
 12 * version 2 as published by the Free Software Foundation.
 13 *
 14 * This program is distributed in the hope that it will be useful, but
 15 * WITHOUT ANY WARRANTY; without even the implied warranty of
 16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
 17 * General Public License for more details.
 18 *
 19 */
 20
 21#include <linux/gfp.h>
 22#include <linux/hash.h>
 23#include <linux/kernel.h>
 24#include <linux/mm.h>
 25#include <linux/percpu.h>
 26#include <linux/printk.h>
 27#include <linux/shrinker.h>
 28#include <linux/slab.h>
 29#include <linux/srcu.h>
 30#include <linux/string.h>
 31#include <linux/types.h>
 32
 33#include "../slab.h"
 34#include "kasan.h"
 35
 36/* Data structure and operations for quarantine queues. */
 37
 38/*
 39 * Each queue is a signle-linked list, which also stores the total size of
 40 * objects inside of it.
 41 */
 42struct qlist_head {
 43	struct qlist_node *head;
 44	struct qlist_node *tail;
 45	size_t bytes;
 46};
 47
 48#define QLIST_INIT { NULL, NULL, 0 }
 49
 50static bool qlist_empty(struct qlist_head *q)
 51{
 52	return !q->head;
 53}
 54
 55static void qlist_init(struct qlist_head *q)
 56{
 57	q->head = q->tail = NULL;
 58	q->bytes = 0;
 59}
 60
 61static void qlist_put(struct qlist_head *q, struct qlist_node *qlink,
 62		size_t size)
 63{
 64	if (unlikely(qlist_empty(q)))
 65		q->head = qlink;
 66	else
 67		q->tail->next = qlink;
 68	q->tail = qlink;
 69	qlink->next = NULL;
 70	q->bytes += size;
 71}
 72
 73static void qlist_move_all(struct qlist_head *from, struct qlist_head *to)
 74{
 75	if (unlikely(qlist_empty(from)))
 76		return;
 77
 78	if (qlist_empty(to)) {
 79		*to = *from;
 80		qlist_init(from);
 81		return;
 82	}
 83
 84	to->tail->next = from->head;
 85	to->tail = from->tail;
 86	to->bytes += from->bytes;
 87
 88	qlist_init(from);
 89}
 90
 91#define QUARANTINE_PERCPU_SIZE (1 << 20)
 92#define QUARANTINE_BATCHES \
 93	(1024 > 4 * CONFIG_NR_CPUS ? 1024 : 4 * CONFIG_NR_CPUS)
 94
 95/*
 96 * The object quarantine consists of per-cpu queues and a global queue,
 97 * guarded by quarantine_lock.
 98 */
 99static DEFINE_PER_CPU(struct qlist_head, cpu_quarantine);
100
101/* Round-robin FIFO array of batches. */
102static struct qlist_head global_quarantine[QUARANTINE_BATCHES];
103static int quarantine_head;
104static int quarantine_tail;
105/* Total size of all objects in global_quarantine across all batches. */
106static unsigned long quarantine_size;
107static DEFINE_RAW_SPINLOCK(quarantine_lock);
108DEFINE_STATIC_SRCU(remove_cache_srcu);
109
110/* Maximum size of the global queue. */
111static unsigned long quarantine_max_size;
112
113/*
114 * Target size of a batch in global_quarantine.
115 * Usually equal to QUARANTINE_PERCPU_SIZE unless we have too much RAM.
116 */
117static unsigned long quarantine_batch_size;
118
119/*
120 * The fraction of physical memory the quarantine is allowed to occupy.
121 * Quarantine doesn't support memory shrinker with SLAB allocator, so we keep
122 * the ratio low to avoid OOM.
123 */
124#define QUARANTINE_FRACTION 32
125
126static struct kmem_cache *qlink_to_cache(struct qlist_node *qlink)
127{
128	return virt_to_head_page(qlink)->slab_cache;
129}
130
131static void *qlink_to_object(struct qlist_node *qlink, struct kmem_cache *cache)
132{
133	struct kasan_free_meta *free_info =
134		container_of(qlink, struct kasan_free_meta,
135			     quarantine_link);
136
137	return ((void *)free_info) - cache->kasan_info.free_meta_offset;
138}
139
140static void qlink_free(struct qlist_node *qlink, struct kmem_cache *cache)
141{
142	void *object = qlink_to_object(qlink, cache);
143	unsigned long flags;
144
145	if (IS_ENABLED(CONFIG_SLAB))
146		local_irq_save(flags);
147
148	*(u8 *)kasan_mem_to_shadow(object) = KASAN_KMALLOC_FREE;
149	___cache_free(cache, object, _THIS_IP_);
150
151	if (IS_ENABLED(CONFIG_SLAB))
152		local_irq_restore(flags);
153}
154
155static void qlist_free_all(struct qlist_head *q, struct kmem_cache *cache)
156{
157	struct qlist_node *qlink;
158
159	if (unlikely(qlist_empty(q)))
160		return;
161
162	qlink = q->head;
163	while (qlink) {
164		struct kmem_cache *obj_cache =
165			cache ? cache :	qlink_to_cache(qlink);
166		struct qlist_node *next = qlink->next;
167
168		qlink_free(qlink, obj_cache);
169		qlink = next;
170	}
171	qlist_init(q);
172}
173
174void quarantine_put(struct kasan_free_meta *info, struct kmem_cache *cache)
175{
176	unsigned long flags;
177	struct qlist_head *q;
178	struct qlist_head temp = QLIST_INIT;
179
180	/*
181	 * Note: irq must be disabled until after we move the batch to the
182	 * global quarantine. Otherwise quarantine_remove_cache() can miss
183	 * some objects belonging to the cache if they are in our local temp
184	 * list. quarantine_remove_cache() executes on_each_cpu() at the
185	 * beginning which ensures that it either sees the objects in per-cpu
186	 * lists or in the global quarantine.
187	 */
188	local_irq_save(flags);
189
190	q = this_cpu_ptr(&cpu_quarantine);
191	qlist_put(q, &info->quarantine_link, cache->size);
192	if (unlikely(q->bytes > QUARANTINE_PERCPU_SIZE)) {
193		qlist_move_all(q, &temp);
194
195		raw_spin_lock(&quarantine_lock);
196		WRITE_ONCE(quarantine_size, quarantine_size + temp.bytes);
197		qlist_move_all(&temp, &global_quarantine[quarantine_tail]);
198		if (global_quarantine[quarantine_tail].bytes >=
199				READ_ONCE(quarantine_batch_size)) {
200			int new_tail;
201
202			new_tail = quarantine_tail + 1;
203			if (new_tail == QUARANTINE_BATCHES)
204				new_tail = 0;
205			if (new_tail != quarantine_head)
206				quarantine_tail = new_tail;
207		}
208		raw_spin_unlock(&quarantine_lock);
209	}
210
211	local_irq_restore(flags);
212}
213
214void quarantine_reduce(void)
215{
216	size_t total_size, new_quarantine_size, percpu_quarantines;
217	unsigned long flags;
218	int srcu_idx;
219	struct qlist_head to_free = QLIST_INIT;
220
221	if (likely(READ_ONCE(quarantine_size) <=
222		   READ_ONCE(quarantine_max_size)))
223		return;
224
225	/*
226	 * srcu critical section ensures that quarantine_remove_cache()
227	 * will not miss objects belonging to the cache while they are in our
228	 * local to_free list. srcu is chosen because (1) it gives us private
229	 * grace period domain that does not interfere with anything else,
230	 * and (2) it allows synchronize_srcu() to return without waiting
231	 * if there are no pending read critical sections (which is the
232	 * expected case).
233	 */
234	srcu_idx = srcu_read_lock(&remove_cache_srcu);
235	raw_spin_lock_irqsave(&quarantine_lock, flags);
236
237	/*
238	 * Update quarantine size in case of hotplug. Allocate a fraction of
239	 * the installed memory to quarantine minus per-cpu queue limits.
240	 */
241	total_size = (totalram_pages() << PAGE_SHIFT) /
242		QUARANTINE_FRACTION;
243	percpu_quarantines = QUARANTINE_PERCPU_SIZE * num_online_cpus();
244	new_quarantine_size = (total_size < percpu_quarantines) ?
245		0 : total_size - percpu_quarantines;
246	WRITE_ONCE(quarantine_max_size, new_quarantine_size);
247	/* Aim at consuming at most 1/2 of slots in quarantine. */
248	WRITE_ONCE(quarantine_batch_size, max((size_t)QUARANTINE_PERCPU_SIZE,
249		2 * total_size / QUARANTINE_BATCHES));
250
251	if (likely(quarantine_size > quarantine_max_size)) {
252		qlist_move_all(&global_quarantine[quarantine_head], &to_free);
253		WRITE_ONCE(quarantine_size, quarantine_size - to_free.bytes);
254		quarantine_head++;
255		if (quarantine_head == QUARANTINE_BATCHES)
256			quarantine_head = 0;
257	}
258
259	raw_spin_unlock_irqrestore(&quarantine_lock, flags);
260
261	qlist_free_all(&to_free, NULL);
262	srcu_read_unlock(&remove_cache_srcu, srcu_idx);
263}
264
265static void qlist_move_cache(struct qlist_head *from,
266				   struct qlist_head *to,
267				   struct kmem_cache *cache)
268{
269	struct qlist_node *curr;
270
271	if (unlikely(qlist_empty(from)))
272		return;
273
274	curr = from->head;
275	qlist_init(from);
276	while (curr) {
277		struct qlist_node *next = curr->next;
278		struct kmem_cache *obj_cache = qlink_to_cache(curr);
279
280		if (obj_cache == cache)
281			qlist_put(to, curr, obj_cache->size);
282		else
283			qlist_put(from, curr, obj_cache->size);
284
285		curr = next;
286	}
287}
288
289static void per_cpu_remove_cache(void *arg)
290{
291	struct kmem_cache *cache = arg;
292	struct qlist_head to_free = QLIST_INIT;
293	struct qlist_head *q;
294
295	q = this_cpu_ptr(&cpu_quarantine);
296	qlist_move_cache(q, &to_free, cache);
297	qlist_free_all(&to_free, cache);
298}
299
300/* Free all quarantined objects belonging to cache. */
301void quarantine_remove_cache(struct kmem_cache *cache)
302{
303	unsigned long flags, i;
304	struct qlist_head to_free = QLIST_INIT;
305
306	/*
307	 * Must be careful to not miss any objects that are being moved from
308	 * per-cpu list to the global quarantine in quarantine_put(),
309	 * nor objects being freed in quarantine_reduce(). on_each_cpu()
310	 * achieves the first goal, while synchronize_srcu() achieves the
311	 * second.
312	 */
313	on_each_cpu(per_cpu_remove_cache, cache, 1);
314
315	raw_spin_lock_irqsave(&quarantine_lock, flags);
316	for (i = 0; i < QUARANTINE_BATCHES; i++) {
317		if (qlist_empty(&global_quarantine[i]))
318			continue;
319		qlist_move_cache(&global_quarantine[i], &to_free, cache);
320		/* Scanning whole quarantine can take a while. */
321		raw_spin_unlock_irqrestore(&quarantine_lock, flags);
322		cond_resched();
323		raw_spin_lock_irqsave(&quarantine_lock, flags);
324	}
325	raw_spin_unlock_irqrestore(&quarantine_lock, flags);
326
327	qlist_free_all(&to_free, cache);
328
329	synchronize_srcu(&remove_cache_srcu);
330}