xref: /linux/mm/shuffle.c (revision 6c363eafc4d637ac4bd83d4a7dd06dd3cfbe7c5f)
1 // SPDX-License-Identifier: GPL-2.0
2 // Copyright(c) 2018 Intel Corporation. All rights reserved.
3 
4 #include <linux/mm.h>
5 #include <linux/init.h>
6 #include <linux/mmzone.h>
7 #include <linux/random.h>
8 #include <linux/moduleparam.h>
9 #include "internal.h"
10 #include "shuffle.h"
11 
12 DEFINE_STATIC_KEY_FALSE(page_alloc_shuffle_key);
13 
14 static bool shuffle_param;
15 static int shuffle_show(char *buffer, const struct kernel_param *kp)
16 {
17 	return sprintf(buffer, "%c\n", shuffle_param ? 'Y' : 'N');
18 }
19 
20 static __meminit int shuffle_store(const char *val,
21 		const struct kernel_param *kp)
22 {
23 	int rc = param_set_bool(val, kp);
24 
25 	if (rc < 0)
26 		return rc;
27 	if (shuffle_param)
28 		static_branch_enable(&page_alloc_shuffle_key);
29 	return 0;
30 }
31 module_param_call(shuffle, shuffle_store, shuffle_show, &shuffle_param, 0400);
32 
33 /*
34  * For two pages to be swapped in the shuffle, they must be free (on a
35  * 'free_area' lru), have the same order, and have the same migratetype.
36  */
37 static struct page * __meminit shuffle_valid_page(struct zone *zone,
38 						  unsigned long pfn, int order)
39 {
40 	struct page *page = pfn_to_online_page(pfn);
41 
42 	/*
43 	 * Given we're dealing with randomly selected pfns in a zone we
44 	 * need to ask questions like...
45 	 */
46 
47 	/* ... is the page managed by the buddy? */
48 	if (!page)
49 		return NULL;
50 
51 	/* ... is the page assigned to the same zone? */
52 	if (page_zone(page) != zone)
53 		return NULL;
54 
55 	/* ...is the page free and currently on a free_area list? */
56 	if (!PageBuddy(page))
57 		return NULL;
58 
59 	/*
60 	 * ...is the page on the same list as the page we will
61 	 * shuffle it with?
62 	 */
63 	if (buddy_order(page) != order)
64 		return NULL;
65 
66 	return page;
67 }
68 
69 /*
70  * Fisher-Yates shuffle the freelist which prescribes iterating through an
71  * array, pfns in this case, and randomly swapping each entry with another in
72  * the span, end_pfn - start_pfn.
73  *
74  * To keep the implementation simple it does not attempt to correct for sources
75  * of bias in the distribution, like modulo bias or pseudo-random number
76  * generator bias. I.e. the expectation is that this shuffling raises the bar
77  * for attacks that exploit the predictability of page allocations, but need not
78  * be a perfect shuffle.
79  */
80 #define SHUFFLE_RETRY 10
81 void __meminit __shuffle_zone(struct zone *z)
82 {
83 	unsigned long i, flags;
84 	unsigned long start_pfn = z->zone_start_pfn;
85 	unsigned long end_pfn = zone_end_pfn(z);
86 	const int order = SHUFFLE_ORDER;
87 	const int order_pages = 1 << order;
88 
89 	spin_lock_irqsave(&z->lock, flags);
90 	start_pfn = ALIGN(start_pfn, order_pages);
91 	for (i = start_pfn; i < end_pfn; i += order_pages) {
92 		unsigned long j;
93 		int migratetype, retry;
94 		struct page *page_i, *page_j;
95 
96 		/*
97 		 * We expect page_i, in the sub-range of a zone being added
98 		 * (@start_pfn to @end_pfn), to more likely be valid compared to
99 		 * page_j randomly selected in the span @zone_start_pfn to
100 		 * @spanned_pages.
101 		 */
102 		page_i = shuffle_valid_page(z, i, order);
103 		if (!page_i)
104 			continue;
105 
106 		for (retry = 0; retry < SHUFFLE_RETRY; retry++) {
107 			/*
108 			 * Pick a random order aligned page in the zone span as
109 			 * a swap target. If the selected pfn is a hole, retry
110 			 * up to SHUFFLE_RETRY attempts find a random valid pfn
111 			 * in the zone.
112 			 */
113 			j = z->zone_start_pfn +
114 				ALIGN_DOWN(get_random_long() % z->spanned_pages,
115 						order_pages);
116 			page_j = shuffle_valid_page(z, j, order);
117 			if (page_j && page_j != page_i)
118 				break;
119 		}
120 		if (retry >= SHUFFLE_RETRY) {
121 			pr_debug("%s: failed to swap %#lx\n", __func__, i);
122 			continue;
123 		}
124 
125 		/*
126 		 * Each migratetype corresponds to its own list, make sure the
127 		 * types match otherwise we're moving pages to lists where they
128 		 * do not belong.
129 		 */
130 		migratetype = get_pageblock_migratetype(page_i);
131 		if (get_pageblock_migratetype(page_j) != migratetype) {
132 			pr_debug("%s: migratetype mismatch %#lx\n", __func__, i);
133 			continue;
134 		}
135 
136 		list_swap(&page_i->lru, &page_j->lru);
137 
138 		pr_debug("%s: swap: %#lx -> %#lx\n", __func__, i, j);
139 
140 		/* take it easy on the zone lock */
141 		if ((i % (100 * order_pages)) == 0) {
142 			spin_unlock_irqrestore(&z->lock, flags);
143 			cond_resched();
144 			spin_lock_irqsave(&z->lock, flags);
145 		}
146 	}
147 	spin_unlock_irqrestore(&z->lock, flags);
148 }
149 
150 /**
151  * shuffle_free_memory - reduce the predictability of the page allocator
152  * @pgdat: node page data
153  */
154 void __meminit __shuffle_free_memory(pg_data_t *pgdat)
155 {
156 	struct zone *z;
157 
158 	for (z = pgdat->node_zones; z < pgdat->node_zones + MAX_NR_ZONES; z++)
159 		shuffle_zone(z);
160 }
161 
162 bool shuffle_pick_tail(void)
163 {
164 	static u64 rand;
165 	static u8 rand_bits;
166 	bool ret;
167 
168 	/*
169 	 * The lack of locking is deliberate. If 2 threads race to
170 	 * update the rand state it just adds to the entropy.
171 	 */
172 	if (rand_bits == 0) {
173 		rand_bits = 64;
174 		rand = get_random_u64();
175 	}
176 
177 	ret = rand & 1;
178 
179 	rand_bits--;
180 	rand >>= 1;
181 
182 	return ret;
183 }
184