mm/shuffle.c

   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 static unsigned long shuffle_state __ro_after_init;
  14
  15 /*
  16  * Depending on the architecture, module parameter parsing may run
  17  * before, or after the cache detection. SHUFFLE_FORCE_DISABLE prevents,
  18  * or reverts the enabling of the shuffle implementation. SHUFFLE_ENABLE
  19  * attempts to turn on the implementation, but aborts if it finds
  20  * SHUFFLE_FORCE_DISABLE already set.
  21  */
  22 __meminit void page_alloc_shuffle(enum mm_shuffle_ctl ctl)
  23 {
  24         if (ctl == SHUFFLE_FORCE_DISABLE)
  25                 set_bit(SHUFFLE_FORCE_DISABLE, &shuffle_state);
  26
  27         if (test_bit(SHUFFLE_FORCE_DISABLE, &shuffle_state)) {
  28                 if (test_and_clear_bit(SHUFFLE_ENABLE, &shuffle_state))
  29                         static_branch_disable(&page_alloc_shuffle_key);
  30         } else if (ctl == SHUFFLE_ENABLE
  31                         && !test_and_set_bit(SHUFFLE_ENABLE, &shuffle_state))
  32                 static_branch_enable(&page_alloc_shuffle_key);
  33 }
  34
  35 static bool shuffle_param;
  36 static int shuffle_show(char *buffer, const struct kernel_param *kp)
  37 {
  38         return sprintf(buffer, "%c\n", test_bit(SHUFFLE_ENABLE, &shuffle_state)
  39                         ? 'Y' : 'N');
  40 }
  41
  42 static __meminit int shuffle_store(const char *val,
  43                 const struct kernel_param *kp)
  44 {
  45         int rc = param_set_bool(val, kp);
  46
  47         if (rc < 0)
  48                 return rc;
  49         if (shuffle_param)
  50                 page_alloc_shuffle(SHUFFLE_ENABLE);
  51         else
  52                 page_alloc_shuffle(SHUFFLE_FORCE_DISABLE);
  53         return 0;
  54 }
  55 module_param_call(shuffle, shuffle_store, shuffle_show, &shuffle_param, 0400);
  56
  57 /*
  58  * For two pages to be swapped in the shuffle, they must be free (on a
  59  * 'free_area' lru), have the same order, and have the same migratetype.
  60  */
  61 static struct page * __meminit shuffle_valid_page(unsigned long pfn, int order)
  62 {
  63         struct page *page;
  64
  65         /*
  66          * Given we're dealing with randomly selected pfns in a zone we
  67          * need to ask questions like...
  68          */
  69
  70         /* ...is the pfn even in the memmap? */
  71         if (!pfn_valid_within(pfn))
  72                 return NULL;
  73
  74         /* ...is the pfn in a present section or a hole? */
  75         if (!pfn_in_present_section(pfn))
  76                 return NULL;
  77
  78         /* ...is the page free and currently on a free_area list? */
  79         page = pfn_to_page(pfn);
  80         if (!PageBuddy(page))
  81                 return NULL;
  82
  83         /*
  84          * ...is the page on the same list as the page we will
  85          * shuffle it with?
  86          */
  87         if (page_order(page) != order)
  88                 return NULL;
  89
  90         return page;
  91 }
  92
  93 /*
  94  * Fisher-Yates shuffle the freelist which prescribes iterating through an
  95  * array, pfns in this case, and randomly swapping each entry with another in
  96  * the span, end_pfn - start_pfn.
  97  *
  98  * To keep the implementation simple it does not attempt to correct for sources
  99  * of bias in the distribution, like modulo bias or pseudo-random number
 100  * generator bias. I.e. the expectation is that this shuffling raises the bar
 101  * for attacks that exploit the predictability of page allocations, but need not
 102  * be a perfect shuffle.
 103  */
 104 #define SHUFFLE_RETRY 10
 105 void __meminit __shuffle_zone(struct zone *z)
 106 {
 107         unsigned long i, flags;
 108         unsigned long start_pfn = z->zone_start_pfn;
 109         unsigned long end_pfn = zone_end_pfn(z);
 110         const int order = SHUFFLE_ORDER;
 111         const int order_pages = 1 << order;
 112
 113         spin_lock_irqsave(&z->lock, flags);
 114         start_pfn = ALIGN(start_pfn, order_pages);
 115         for (i = start_pfn; i < end_pfn; i += order_pages) {
 116                 unsigned long j;
 117                 int migratetype, retry;
 118                 struct page *page_i, *page_j;
 119
 120                 /*
 121                  * We expect page_i, in the sub-range of a zone being added
 122                  * (@start_pfn to @end_pfn), to more likely be valid compared to
 123                  * page_j randomly selected in the span @zone_start_pfn to
 124                  * @spanned_pages.
 125                  */
 126                 page_i = shuffle_valid_page(i, order);
 127                 if (!page_i)
 128                         continue;
 129
 130                 for (retry = 0; retry < SHUFFLE_RETRY; retry++) {
 131                         /*
 132                          * Pick a random order aligned page in the zone span as
 133                          * a swap target. If the selected pfn is a hole, retry
 134                          * up to SHUFFLE_RETRY attempts find a random valid pfn
 135                          * in the zone.
 136                          */
 137                         j = z->zone_start_pfn +
 138                                 ALIGN_DOWN(get_random_long() % z->spanned_pages,
 139                                                 order_pages);
 140                         page_j = shuffle_valid_page(j, order);
 141                         if (page_j && page_j != page_i)
 142                                 break;
 143                 }
 144                 if (retry >= SHUFFLE_RETRY) {
 145                         pr_debug("%s: failed to swap %#lx\n", __func__, i);
 146                         continue;
 147                 }
 148
 149                 /*
 150                  * Each migratetype corresponds to its own list, make sure the
 151                  * types match otherwise we're moving pages to lists where they
 152                  * do not belong.
 153                  */
 154                 migratetype = get_pageblock_migratetype(page_i);
 155                 if (get_pageblock_migratetype(page_j) != migratetype) {
 156                         pr_debug("%s: migratetype mismatch %#lx\n", __func__, i);
 157                         continue;
 158                 }
 159
 160                 list_swap(&page_i->lru, &page_j->lru);
 161
 162                 pr_debug("%s: swap: %#lx -> %#lx\n", __func__, i, j);
 163
 164                 /* take it easy on the zone lock */
 165                 if ((i % (100 * order_pages)) == 0) {
 166                         spin_unlock_irqrestore(&z->lock, flags);
 167                         cond_resched();
 168                         spin_lock_irqsave(&z->lock, flags);
 169                 }
 170         }
 171         spin_unlock_irqrestore(&z->lock, flags);
 172 }
 173
 174 /**
 175  * shuffle_free_memory - reduce the predictability of the page allocator
 176  * @pgdat: node page data
 177  */
 178 void __meminit __shuffle_free_memory(pg_data_t *pgdat)
 179 {
 180         struct zone *z;
 181
 182         for (z = pgdat->node_zones; z < pgdat->node_zones + MAX_NR_ZONES; z++)
 183                 shuffle_zone(z);
 184 }
 185
 186 bool shuffle_pick_tail(void)
 187 {
 188         static u64 rand;
 189         static u8 rand_bits;
 190         bool ret;
 191
 192         /*
 193          * The lack of locking is deliberate. If 2 threads race to
 194          * update the rand state it just adds to the entropy.
 195          */
 196         if (rand_bits == 0) {
 197                 rand_bits = 64;
 198                 rand = get_random_u64();
 199         }
 200
 201         ret = rand & 1;
 202
 203         rand_bits--;
 204         rand >>= 1;
 205
 206         return ret;
 207 }