Commit | Line | Data |
---|---|---|
e900a918 DW |
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); | |
e900a918 DW |
13 | |
14 | static bool shuffle_param; | |
e900a918 | 15 | |
85a34107 | 16 | static __meminit int shuffle_param_set(const char *val, |
e900a918 DW |
17 | const struct kernel_param *kp) |
18 | { | |
85a34107 LS |
19 | if (param_set_bool(val, kp)) |
20 | return -EINVAL; | |
21 | if (*(bool *)kp->arg) | |
83919535 | 22 | static_branch_enable(&page_alloc_shuffle_key); |
e900a918 DW |
23 | return 0; |
24 | } | |
85a34107 LS |
25 | |
26 | static const struct kernel_param_ops shuffle_param_ops = { | |
27 | .set = shuffle_param_set, | |
28 | .get = param_get_bool, | |
29 | }; | |
30 | module_param_cb(shuffle, &shuffle_param_ops, &shuffle_param, 0400); | |
e900a918 DW |
31 | |
32 | /* | |
33 | * For two pages to be swapped in the shuffle, they must be free (on a | |
34 | * 'free_area' lru), have the same order, and have the same migratetype. | |
35 | */ | |
4a93025c DH |
36 | static struct page * __meminit shuffle_valid_page(struct zone *zone, |
37 | unsigned long pfn, int order) | |
e900a918 | 38 | { |
4a93025c | 39 | struct page *page = pfn_to_online_page(pfn); |
e900a918 DW |
40 | |
41 | /* | |
42 | * Given we're dealing with randomly selected pfns in a zone we | |
43 | * need to ask questions like... | |
44 | */ | |
45 | ||
4a93025c DH |
46 | /* ... is the page managed by the buddy? */ |
47 | if (!page) | |
e900a918 DW |
48 | return NULL; |
49 | ||
4a93025c DH |
50 | /* ... is the page assigned to the same zone? */ |
51 | if (page_zone(page) != zone) | |
e900a918 DW |
52 | return NULL; |
53 | ||
54 | /* ...is the page free and currently on a free_area list? */ | |
e900a918 DW |
55 | if (!PageBuddy(page)) |
56 | return NULL; | |
57 | ||
58 | /* | |
59 | * ...is the page on the same list as the page we will | |
60 | * shuffle it with? | |
61 | */ | |
ab130f91 | 62 | if (buddy_order(page) != order) |
e900a918 DW |
63 | return NULL; |
64 | ||
65 | return page; | |
66 | } | |
67 | ||
68 | /* | |
69 | * Fisher-Yates shuffle the freelist which prescribes iterating through an | |
70 | * array, pfns in this case, and randomly swapping each entry with another in | |
71 | * the span, end_pfn - start_pfn. | |
72 | * | |
73 | * To keep the implementation simple it does not attempt to correct for sources | |
74 | * of bias in the distribution, like modulo bias or pseudo-random number | |
75 | * generator bias. I.e. the expectation is that this shuffling raises the bar | |
76 | * for attacks that exploit the predictability of page allocations, but need not | |
77 | * be a perfect shuffle. | |
78 | */ | |
79 | #define SHUFFLE_RETRY 10 | |
80 | void __meminit __shuffle_zone(struct zone *z) | |
81 | { | |
82 | unsigned long i, flags; | |
83 | unsigned long start_pfn = z->zone_start_pfn; | |
84 | unsigned long end_pfn = zone_end_pfn(z); | |
85 | const int order = SHUFFLE_ORDER; | |
86 | const int order_pages = 1 << order; | |
87 | ||
88 | spin_lock_irqsave(&z->lock, flags); | |
89 | start_pfn = ALIGN(start_pfn, order_pages); | |
90 | for (i = start_pfn; i < end_pfn; i += order_pages) { | |
91 | unsigned long j; | |
92 | int migratetype, retry; | |
93 | struct page *page_i, *page_j; | |
94 | ||
95 | /* | |
96 | * We expect page_i, in the sub-range of a zone being added | |
97 | * (@start_pfn to @end_pfn), to more likely be valid compared to | |
98 | * page_j randomly selected in the span @zone_start_pfn to | |
99 | * @spanned_pages. | |
100 | */ | |
4a93025c | 101 | page_i = shuffle_valid_page(z, i, order); |
e900a918 DW |
102 | if (!page_i) |
103 | continue; | |
104 | ||
105 | for (retry = 0; retry < SHUFFLE_RETRY; retry++) { | |
106 | /* | |
107 | * Pick a random order aligned page in the zone span as | |
108 | * a swap target. If the selected pfn is a hole, retry | |
109 | * up to SHUFFLE_RETRY attempts find a random valid pfn | |
110 | * in the zone. | |
111 | */ | |
112 | j = z->zone_start_pfn + | |
113 | ALIGN_DOWN(get_random_long() % z->spanned_pages, | |
114 | order_pages); | |
4a93025c | 115 | page_j = shuffle_valid_page(z, j, order); |
e900a918 DW |
116 | if (page_j && page_j != page_i) |
117 | break; | |
118 | } | |
119 | if (retry >= SHUFFLE_RETRY) { | |
120 | pr_debug("%s: failed to swap %#lx\n", __func__, i); | |
121 | continue; | |
122 | } | |
123 | ||
124 | /* | |
125 | * Each migratetype corresponds to its own list, make sure the | |
126 | * types match otherwise we're moving pages to lists where they | |
127 | * do not belong. | |
128 | */ | |
129 | migratetype = get_pageblock_migratetype(page_i); | |
130 | if (get_pageblock_migratetype(page_j) != migratetype) { | |
131 | pr_debug("%s: migratetype mismatch %#lx\n", __func__, i); | |
132 | continue; | |
133 | } | |
134 | ||
135 | list_swap(&page_i->lru, &page_j->lru); | |
136 | ||
137 | pr_debug("%s: swap: %#lx -> %#lx\n", __func__, i, j); | |
138 | ||
139 | /* take it easy on the zone lock */ | |
140 | if ((i % (100 * order_pages)) == 0) { | |
141 | spin_unlock_irqrestore(&z->lock, flags); | |
142 | cond_resched(); | |
143 | spin_lock_irqsave(&z->lock, flags); | |
144 | } | |
145 | } | |
146 | spin_unlock_irqrestore(&z->lock, flags); | |
147 | } | |
148 | ||
845be1cd RD |
149 | /* |
150 | * __shuffle_free_memory - reduce the predictability of the page allocator | |
e900a918 DW |
151 | * @pgdat: node page data |
152 | */ | |
153 | void __meminit __shuffle_free_memory(pg_data_t *pgdat) | |
154 | { | |
155 | struct zone *z; | |
156 | ||
157 | for (z = pgdat->node_zones; z < pgdat->node_zones + MAX_NR_ZONES; z++) | |
158 | shuffle_zone(z); | |
159 | } | |
97500a4a | 160 | |
a2129f24 | 161 | bool shuffle_pick_tail(void) |
97500a4a DW |
162 | { |
163 | static u64 rand; | |
164 | static u8 rand_bits; | |
a2129f24 | 165 | bool ret; |
97500a4a DW |
166 | |
167 | /* | |
168 | * The lack of locking is deliberate. If 2 threads race to | |
169 | * update the rand state it just adds to the entropy. | |
170 | */ | |
171 | if (rand_bits == 0) { | |
172 | rand_bits = 64; | |
173 | rand = get_random_u64(); | |
174 | } | |
175 | ||
a2129f24 AD |
176 | ret = rand & 1; |
177 | ||
97500a4a DW |
178 | rand_bits--; |
179 | rand >>= 1; | |
a2129f24 AD |
180 | |
181 | return ret; | |
97500a4a | 182 | } |