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;
shuffle_show(char * buffer,const struct kernel_param * kp)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
shuffle_store(const char * val,const struct kernel_param * kp)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 */
shuffle_valid_page(struct zone * zone,unsigned long pfn,int order)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
__shuffle_zone(struct zone * z)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 */
__shuffle_free_memory(pg_data_t * pgdat)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
shuffle_pick_tail(void)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