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/mm.h>
8 #include <linux/mmzone.h>
9 #include <linux/pagemap.h>
10 #include <linux/rmap.h>
11 #include <linux/mmu_notifier.h>
12 #include <linux/page_ext.h>
13 #include <linux/page_idle.h>
14
15 #define BITMAP_CHUNK_SIZE sizeof(u64)
16 #define BITMAP_CHUNK_BITS (BITMAP_CHUNK_SIZE * BITS_PER_BYTE)
17
18 /*
19 * Idle page tracking only considers user memory pages, for other types of
20 * pages the idle flag is always unset and an attempt to set it is silently
21 * ignored.
22 *
23 * We treat a page as a user memory page if it is on an LRU list, because it is
24 * always safe to pass such a page to rmap_walk(), which is essential for idle
25 * page tracking. With such an indicator of user pages we can skip isolated
26 * pages, but since there are not usually many of them, it will hardly affect
27 * the overall result.
28 *
29 * This function tries to get a user memory page by pfn as described above.
30 */
page_idle_get_page(unsigned long pfn)31 static struct page *page_idle_get_page(unsigned long pfn)
32 {
33 struct page *page;
34 pg_data_t *pgdat;
35
36 if (!pfn_valid(pfn))
37 return NULL;
38
39 page = pfn_to_page(pfn);
40 if (!page || !PageLRU(page) ||
41 !get_page_unless_zero(page))
42 return NULL;
43
44 pgdat = page_pgdat(page);
45 spin_lock_irq(&pgdat->lru_lock);
46 if (unlikely(!PageLRU(page))) {
47 put_page(page);
48 page = NULL;
49 }
50 spin_unlock_irq(&pgdat->lru_lock);
51 return page;
52 }
53
page_idle_clear_pte_refs_one(struct page * page,struct vm_area_struct * vma,unsigned long addr,void * arg)54 static bool page_idle_clear_pte_refs_one(struct page *page,
55 struct vm_area_struct *vma,
56 unsigned long addr, void *arg)
57 {
58 struct page_vma_mapped_walk pvmw = {
59 .page = page,
60 .vma = vma,
61 .address = addr,
62 };
63 bool referenced = false;
64
65 while (page_vma_mapped_walk(&pvmw)) {
66 addr = pvmw.address;
67 if (pvmw.pte) {
68 /*
69 * For PTE-mapped THP, one sub page is referenced,
70 * the whole THP is referenced.
71 */
72 if (ptep_clear_young_notify(vma, addr, pvmw.pte))
73 referenced = true;
74 } else if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE)) {
75 if (pmdp_clear_young_notify(vma, addr, pvmw.pmd))
76 referenced = true;
77 } else {
78 /* unexpected pmd-mapped page? */
79 WARN_ON_ONCE(1);
80 }
81 }
82
83 if (referenced) {
84 clear_page_idle(page);
85 /*
86 * We cleared the referenced bit in a mapping to this page. To
87 * avoid interference with page reclaim, mark it young so that
88 * page_referenced() will return > 0.
89 */
90 set_page_young(page);
91 }
92 return true;
93 }
94
page_idle_clear_pte_refs(struct page * page)95 static void page_idle_clear_pte_refs(struct page *page)
96 {
97 /*
98 * Since rwc.arg is unused, rwc is effectively immutable, so we
99 * can make it static const to save some cycles and stack.
100 */
101 static const struct rmap_walk_control rwc = {
102 .rmap_one = page_idle_clear_pte_refs_one,
103 .anon_lock = page_lock_anon_vma_read,
104 };
105 bool need_lock;
106
107 if (!page_mapped(page) ||
108 !page_rmapping(page))
109 return;
110
111 need_lock = !PageAnon(page) || PageKsm(page);
112 if (need_lock && !trylock_page(page))
113 return;
114
115 rmap_walk(page, (struct rmap_walk_control *)&rwc);
116
117 if (need_lock)
118 unlock_page(page);
119 }
120
page_idle_bitmap_read(struct file * file,struct kobject * kobj,struct bin_attribute * attr,char * buf,loff_t pos,size_t count)121 static ssize_t page_idle_bitmap_read(struct file *file, struct kobject *kobj,
122 struct bin_attribute *attr, char *buf,
123 loff_t pos, size_t count)
124 {
125 u64 *out = (u64 *)buf;
126 struct page *page;
127 unsigned long pfn, end_pfn;
128 int bit;
129
130 if (pos % BITMAP_CHUNK_SIZE || count % BITMAP_CHUNK_SIZE)
131 return -EINVAL;
132
133 pfn = pos * BITS_PER_BYTE;
134 if (pfn >= max_pfn)
135 return 0;
136
137 end_pfn = pfn + count * BITS_PER_BYTE;
138 if (end_pfn > max_pfn)
139 end_pfn = max_pfn;
140
141 for (; pfn < end_pfn; pfn++) {
142 bit = pfn % BITMAP_CHUNK_BITS;
143 if (!bit)
144 *out = 0ULL;
145 page = page_idle_get_page(pfn);
146 if (page) {
147 if (page_is_idle(page)) {
148 /*
149 * The page might have been referenced via a
150 * pte, in which case it is not idle. Clear
151 * refs and recheck.
152 */
153 page_idle_clear_pte_refs(page);
154 if (page_is_idle(page))
155 *out |= 1ULL << bit;
156 }
157 put_page(page);
158 }
159 if (bit == BITMAP_CHUNK_BITS - 1)
160 out++;
161 cond_resched();
162 }
163 return (char *)out - buf;
164 }
165
page_idle_bitmap_write(struct file * file,struct kobject * kobj,struct bin_attribute * attr,char * buf,loff_t pos,size_t count)166 static ssize_t page_idle_bitmap_write(struct file *file, struct kobject *kobj,
167 struct bin_attribute *attr, char *buf,
168 loff_t pos, size_t count)
169 {
170 const u64 *in = (u64 *)buf;
171 struct page *page;
172 unsigned long pfn, end_pfn;
173 int bit;
174
175 if (pos % BITMAP_CHUNK_SIZE || count % BITMAP_CHUNK_SIZE)
176 return -EINVAL;
177
178 pfn = pos * BITS_PER_BYTE;
179 if (pfn >= max_pfn)
180 return -ENXIO;
181
182 end_pfn = pfn + count * BITS_PER_BYTE;
183 if (end_pfn > max_pfn)
184 end_pfn = max_pfn;
185
186 for (; pfn < end_pfn; pfn++) {
187 bit = pfn % BITMAP_CHUNK_BITS;
188 if ((*in >> bit) & 1) {
189 page = page_idle_get_page(pfn);
190 if (page) {
191 page_idle_clear_pte_refs(page);
192 set_page_idle(page);
193 put_page(page);
194 }
195 }
196 if (bit == BITMAP_CHUNK_BITS - 1)
197 in++;
198 cond_resched();
199 }
200 return (char *)in - buf;
201 }
202
203 static struct bin_attribute page_idle_bitmap_attr =
204 __BIN_ATTR(bitmap, 0600,
205 page_idle_bitmap_read, page_idle_bitmap_write, 0);
206
207 static struct bin_attribute *page_idle_bin_attrs[] = {
208 &page_idle_bitmap_attr,
209 NULL,
210 };
211
212 static const struct attribute_group page_idle_attr_group = {
213 .bin_attrs = page_idle_bin_attrs,
214 .name = "page_idle",
215 };
216
217 #ifndef CONFIG_64BIT
need_page_idle(void)218 static bool need_page_idle(void)
219 {
220 return true;
221 }
222 struct page_ext_operations page_idle_ops = {
223 .need = need_page_idle,
224 };
225 #endif
226
page_idle_init(void)227 static int __init page_idle_init(void)
228 {
229 int err;
230
231 err = sysfs_create_group(mm_kobj, &page_idle_attr_group);
232 if (err) {
233 pr_err("page_idle: register sysfs failed\n");
234 return err;
235 }
236 return 0;
237 }
238 subsys_initcall(page_idle_init);
239