shuffle.c 4.6 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182
  1. // SPDX-License-Identifier: GPL-2.0
  2. // Copyright(c) 2018 Intel Corporation. All rights reserved.
  3. #include <linux/mm.h>
  4. #include <linux/init.h>
  5. #include <linux/mmzone.h>
  6. #include <linux/random.h>
  7. #include <linux/moduleparam.h>
  8. #include "internal.h"
  9. #include "shuffle.h"
  10. DEFINE_STATIC_KEY_FALSE(page_alloc_shuffle_key);
  11. static bool shuffle_param;
  12. static __meminit int shuffle_param_set(const char *val,
  13. const struct kernel_param *kp)
  14. {
  15. if (param_set_bool(val, kp))
  16. return -EINVAL;
  17. if (*(bool *)kp->arg)
  18. static_branch_enable(&page_alloc_shuffle_key);
  19. return 0;
  20. }
  21. static const struct kernel_param_ops shuffle_param_ops = {
  22. .set = shuffle_param_set,
  23. .get = param_get_bool,
  24. };
  25. module_param_cb(shuffle, &shuffle_param_ops, &shuffle_param, 0400);
  26. /*
  27. * For two pages to be swapped in the shuffle, they must be free (on a
  28. * 'free_area' lru), have the same order, and have the same migratetype.
  29. */
  30. static struct page * __meminit shuffle_valid_page(struct zone *zone,
  31. unsigned long pfn, int order)
  32. {
  33. struct page *page = pfn_to_online_page(pfn);
  34. /*
  35. * Given we're dealing with randomly selected pfns in a zone we
  36. * need to ask questions like...
  37. */
  38. /* ... is the page managed by the buddy? */
  39. if (!page)
  40. return NULL;
  41. /* ... is the page assigned to the same zone? */
  42. if (page_zone(page) != zone)
  43. return NULL;
  44. /* ...is the page free and currently on a free_area list? */
  45. if (!PageBuddy(page))
  46. return NULL;
  47. /*
  48. * ...is the page on the same list as the page we will
  49. * shuffle it with?
  50. */
  51. if (buddy_order(page) != order)
  52. return NULL;
  53. return page;
  54. }
  55. /*
  56. * Fisher-Yates shuffle the freelist which prescribes iterating through an
  57. * array, pfns in this case, and randomly swapping each entry with another in
  58. * the span, end_pfn - start_pfn.
  59. *
  60. * To keep the implementation simple it does not attempt to correct for sources
  61. * of bias in the distribution, like modulo bias or pseudo-random number
  62. * generator bias. I.e. the expectation is that this shuffling raises the bar
  63. * for attacks that exploit the predictability of page allocations, but need not
  64. * be a perfect shuffle.
  65. */
  66. #define SHUFFLE_RETRY 10
  67. void __meminit __shuffle_zone(struct zone *z)
  68. {
  69. unsigned long i, flags;
  70. unsigned long start_pfn = z->zone_start_pfn;
  71. unsigned long end_pfn = zone_end_pfn(z);
  72. const int order = SHUFFLE_ORDER;
  73. const int order_pages = 1 << order;
  74. spin_lock_irqsave(&z->lock, flags);
  75. start_pfn = ALIGN(start_pfn, order_pages);
  76. for (i = start_pfn; i < end_pfn; i += order_pages) {
  77. unsigned long j;
  78. int migratetype, retry;
  79. struct page *page_i, *page_j;
  80. /*
  81. * We expect page_i, in the sub-range of a zone being added
  82. * (@start_pfn to @end_pfn), to more likely be valid compared to
  83. * page_j randomly selected in the span @zone_start_pfn to
  84. * @spanned_pages.
  85. */
  86. page_i = shuffle_valid_page(z, i, order);
  87. if (!page_i)
  88. continue;
  89. for (retry = 0; retry < SHUFFLE_RETRY; retry++) {
  90. /*
  91. * Pick a random order aligned page in the zone span as
  92. * a swap target. If the selected pfn is a hole, retry
  93. * up to SHUFFLE_RETRY attempts find a random valid pfn
  94. * in the zone.
  95. */
  96. j = z->zone_start_pfn +
  97. ALIGN_DOWN(get_random_long() % z->spanned_pages,
  98. order_pages);
  99. page_j = shuffle_valid_page(z, j, order);
  100. if (page_j && page_j != page_i)
  101. break;
  102. }
  103. if (retry >= SHUFFLE_RETRY) {
  104. pr_debug("%s: failed to swap %#lx\n", __func__, i);
  105. continue;
  106. }
  107. /*
  108. * Each migratetype corresponds to its own list, make sure the
  109. * types match otherwise we're moving pages to lists where they
  110. * do not belong.
  111. */
  112. migratetype = get_pageblock_migratetype(page_i);
  113. if (get_pageblock_migratetype(page_j) != migratetype) {
  114. pr_debug("%s: migratetype mismatch %#lx\n", __func__, i);
  115. continue;
  116. }
  117. list_swap(&page_i->lru, &page_j->lru);
  118. pr_debug("%s: swap: %#lx -> %#lx\n", __func__, i, j);
  119. /* take it easy on the zone lock */
  120. if ((i % (100 * order_pages)) == 0) {
  121. spin_unlock_irqrestore(&z->lock, flags);
  122. cond_resched();
  123. spin_lock_irqsave(&z->lock, flags);
  124. }
  125. }
  126. spin_unlock_irqrestore(&z->lock, flags);
  127. }
  128. /*
  129. * __shuffle_free_memory - reduce the predictability of the page allocator
  130. * @pgdat: node page data
  131. */
  132. void __meminit __shuffle_free_memory(pg_data_t *pgdat)
  133. {
  134. struct zone *z;
  135. for (z = pgdat->node_zones; z < pgdat->node_zones + MAX_NR_ZONES; z++)
  136. shuffle_zone(z);
  137. }
  138. bool shuffle_pick_tail(void)
  139. {
  140. static u64 rand;
  141. static u8 rand_bits;
  142. bool ret;
  143. /*
  144. * The lack of locking is deliberate. If 2 threads race to
  145. * update the rand state it just adds to the entropy.
  146. */
  147. if (rand_bits == 0) {
  148. rand_bits = 64;
  149. rand = get_random_u64();
  150. }
  151. ret = rand & 1;
  152. rand_bits--;
  153. rand >>= 1;
  154. return ret;
  155. }