1 // SPDX-License-Identifier: GPL-2.0 2 #include <linux/init.h> 3 #include <linux/memblock.h> 4 #include <linux/fs.h> 5 #include <linux/sysfs.h> 6 #include <linux/kobject.h> 7 #include <linux/memory_hotplug.h> 8 #include <linux/mm.h> 9 #include <linux/mmzone.h> 10 #include <linux/pagemap.h> 11 #include <linux/rmap.h> 12 #include <linux/mmu_notifier.h> 13 #include <linux/page_ext.h> 14 #include <linux/page_idle.h> 15 16 #include "internal.h" 17 18 #define BITMAP_CHUNK_SIZE sizeof(u64) 19 #define BITMAP_CHUNK_BITS (BITMAP_CHUNK_SIZE * BITS_PER_BYTE) 20 21 /* 22 * Idle page tracking only considers user memory pages, for other types of 23 * pages the idle flag is always unset and an attempt to set it is silently 24 * ignored. 25 * 26 * We treat a page as a user memory page if it is on an LRU list, because it is 27 * always safe to pass such a page to rmap_walk(), which is essential for idle 28 * page tracking. With such an indicator of user pages we can skip isolated 29 * pages, but since there are not usually many of them, it will hardly affect 30 * the overall result. 31 * 32 * This function tries to get a user memory page by pfn as described above. 33 */ 34 static struct page *page_idle_get_page(unsigned long pfn) 35 { 36 struct page *page = pfn_to_online_page(pfn); 37 38 if (!page || !PageLRU(page) || 39 !get_page_unless_zero(page)) 40 return NULL; 41 42 if (unlikely(!PageLRU(page))) { 43 put_page(page); 44 page = NULL; 45 } 46 return page; 47 } 48 49 static bool page_idle_clear_pte_refs_one(struct folio *folio, 50 struct vm_area_struct *vma, 51 unsigned long addr, void *arg) 52 { 53 DEFINE_FOLIO_VMA_WALK(pvmw, folio, vma, addr, 0); 54 bool referenced = false; 55 56 while (page_vma_mapped_walk(&pvmw)) { 57 addr = pvmw.address; 58 if (pvmw.pte) { 59 /* 60 * For PTE-mapped THP, one sub page is referenced, 61 * the whole THP is referenced. 62 */ 63 if (ptep_clear_young_notify(vma, addr, pvmw.pte)) 64 referenced = true; 65 } else if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE)) { 66 if (pmdp_clear_young_notify(vma, addr, pvmw.pmd)) 67 referenced = true; 68 } else { 69 /* unexpected pmd-mapped page? */ 70 WARN_ON_ONCE(1); 71 } 72 } 73 74 if (referenced) { 75 folio_clear_idle(folio); 76 /* 77 * We cleared the referenced bit in a mapping to this page. To 78 * avoid interference with page reclaim, mark it young so that 79 * folio_referenced() will return > 0. 80 */ 81 folio_set_young(folio); 82 } 83 return true; 84 } 85 86 static void page_idle_clear_pte_refs(struct page *page) 87 { 88 struct folio *folio = page_folio(page); 89 /* 90 * Since rwc.arg is unused, rwc is effectively immutable, so we 91 * can make it static const to save some cycles and stack. 92 */ 93 static const struct rmap_walk_control rwc = { 94 .rmap_one = page_idle_clear_pte_refs_one, 95 .anon_lock = folio_lock_anon_vma_read, 96 }; 97 bool need_lock; 98 99 if (!folio_mapped(folio) || !folio_raw_mapping(folio)) 100 return; 101 102 need_lock = !folio_test_anon(folio) || folio_test_ksm(folio); 103 if (need_lock && !folio_trylock(folio)) 104 return; 105 106 rmap_walk(folio, &rwc); 107 108 if (need_lock) 109 folio_unlock(folio); 110 } 111 112 static ssize_t page_idle_bitmap_read(struct file *file, struct kobject *kobj, 113 struct bin_attribute *attr, char *buf, 114 loff_t pos, size_t count) 115 { 116 u64 *out = (u64 *)buf; 117 struct page *page; 118 unsigned long pfn, end_pfn; 119 int bit; 120 121 if (pos % BITMAP_CHUNK_SIZE || count % BITMAP_CHUNK_SIZE) 122 return -EINVAL; 123 124 pfn = pos * BITS_PER_BYTE; 125 if (pfn >= max_pfn) 126 return 0; 127 128 end_pfn = pfn + count * BITS_PER_BYTE; 129 if (end_pfn > max_pfn) 130 end_pfn = max_pfn; 131 132 for (; pfn < end_pfn; pfn++) { 133 bit = pfn % BITMAP_CHUNK_BITS; 134 if (!bit) 135 *out = 0ULL; 136 page = page_idle_get_page(pfn); 137 if (page) { 138 if (page_is_idle(page)) { 139 /* 140 * The page might have been referenced via a 141 * pte, in which case it is not idle. Clear 142 * refs and recheck. 143 */ 144 page_idle_clear_pte_refs(page); 145 if (page_is_idle(page)) 146 *out |= 1ULL << bit; 147 } 148 put_page(page); 149 } 150 if (bit == BITMAP_CHUNK_BITS - 1) 151 out++; 152 cond_resched(); 153 } 154 return (char *)out - buf; 155 } 156 157 static ssize_t page_idle_bitmap_write(struct file *file, struct kobject *kobj, 158 struct bin_attribute *attr, char *buf, 159 loff_t pos, size_t count) 160 { 161 const u64 *in = (u64 *)buf; 162 struct page *page; 163 unsigned long pfn, end_pfn; 164 int bit; 165 166 if (pos % BITMAP_CHUNK_SIZE || count % BITMAP_CHUNK_SIZE) 167 return -EINVAL; 168 169 pfn = pos * BITS_PER_BYTE; 170 if (pfn >= max_pfn) 171 return -ENXIO; 172 173 end_pfn = pfn + count * BITS_PER_BYTE; 174 if (end_pfn > max_pfn) 175 end_pfn = max_pfn; 176 177 for (; pfn < end_pfn; pfn++) { 178 bit = pfn % BITMAP_CHUNK_BITS; 179 if ((*in >> bit) & 1) { 180 page = page_idle_get_page(pfn); 181 if (page) { 182 page_idle_clear_pte_refs(page); 183 set_page_idle(page); 184 put_page(page); 185 } 186 } 187 if (bit == BITMAP_CHUNK_BITS - 1) 188 in++; 189 cond_resched(); 190 } 191 return (char *)in - buf; 192 } 193 194 static struct bin_attribute page_idle_bitmap_attr = 195 __BIN_ATTR(bitmap, 0600, 196 page_idle_bitmap_read, page_idle_bitmap_write, 0); 197 198 static struct bin_attribute *page_idle_bin_attrs[] = { 199 &page_idle_bitmap_attr, 200 NULL, 201 }; 202 203 static const struct attribute_group page_idle_attr_group = { 204 .bin_attrs = page_idle_bin_attrs, 205 .name = "page_idle", 206 }; 207 208 static int __init page_idle_init(void) 209 { 210 int err; 211 212 err = sysfs_create_group(mm_kobj, &page_idle_attr_group); 213 if (err) { 214 pr_err("page_idle: register sysfs failed\n"); 215 return err; 216 } 217 return 0; 218 } 219 subsys_initcall(page_idle_init); 220