xref: /linux/mm/page_idle.c (revision badfa4361cb116fd9af71aaa2ea470236a8aa25b)
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 folio *page_idle_get_folio(unsigned long pfn)
35 {
36 	struct page *page = pfn_to_online_page(pfn);
37 	struct folio *folio;
38 
39 	if (!page || PageTail(page))
40 		return NULL;
41 
42 	folio = page_folio(page);
43 	if (!folio_test_lru(folio) || !folio_try_get(folio))
44 		return NULL;
45 	if (unlikely(page_folio(page) != folio || !folio_test_lru(folio))) {
46 		folio_put(folio);
47 		folio = NULL;
48 	}
49 	return folio;
50 }
51 
52 static bool page_idle_clear_pte_refs_one(struct folio *folio,
53 					struct vm_area_struct *vma,
54 					unsigned long addr, void *arg)
55 {
56 	DEFINE_FOLIO_VMA_WALK(pvmw, folio, vma, addr, 0);
57 	bool referenced = false;
58 
59 	while (page_vma_mapped_walk(&pvmw)) {
60 		addr = pvmw.address;
61 		if (pvmw.pte) {
62 			/*
63 			 * For PTE-mapped THP, one sub page is referenced,
64 			 * the whole THP is referenced.
65 			 *
66 			 * PFN swap PTEs, such as device-exclusive ones, that
67 			 * actually map pages are "old" from a CPU perspective.
68 			 * The MMU notifier takes care of any device aspects.
69 			 */
70 			if (likely(pte_present(ptep_get(pvmw.pte))))
71 				referenced |= ptep_test_and_clear_young(vma, addr, pvmw.pte);
72 			referenced |= mmu_notifier_clear_young(vma->vm_mm, addr, addr + PAGE_SIZE);
73 		} else if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE)) {
74 			if (pmdp_clear_young_notify(vma, addr, pvmw.pmd))
75 				referenced = true;
76 		} else {
77 			/* unexpected pmd-mapped page? */
78 			WARN_ON_ONCE(1);
79 		}
80 	}
81 
82 	if (referenced) {
83 		folio_clear_idle(folio);
84 		/*
85 		 * We cleared the referenced bit in a mapping to this page. To
86 		 * avoid interference with page reclaim, mark it young so that
87 		 * folio_referenced() will return > 0.
88 		 */
89 		folio_set_young(folio);
90 	}
91 	return true;
92 }
93 
94 static void page_idle_clear_pte_refs(struct folio *folio)
95 {
96 	/*
97 	 * Since rwc.try_lock is unused, rwc is effectively immutable, so we
98 	 * can make it static to save some cycles and stack.
99 	 */
100 	static struct rmap_walk_control rwc = {
101 		.rmap_one = page_idle_clear_pte_refs_one,
102 		.anon_lock = folio_lock_anon_vma_read,
103 	};
104 
105 	if (!folio_mapped(folio) || !folio_raw_mapping(folio))
106 		return;
107 
108 	if (!folio_trylock(folio))
109 		return;
110 
111 	rmap_walk(folio, &rwc);
112 	folio_unlock(folio);
113 }
114 
115 static ssize_t page_idle_bitmap_read(struct file *file, struct kobject *kobj,
116 				     const struct bin_attribute *attr, char *buf,
117 				     loff_t pos, size_t count)
118 {
119 	u64 *out = (u64 *)buf;
120 	struct folio *folio;
121 	unsigned long pfn, end_pfn;
122 	int bit;
123 
124 	if (pos % BITMAP_CHUNK_SIZE || count % BITMAP_CHUNK_SIZE)
125 		return -EINVAL;
126 
127 	pfn = pos * BITS_PER_BYTE;
128 	if (pfn >= max_pfn)
129 		return 0;
130 
131 	end_pfn = pfn + count * BITS_PER_BYTE;
132 	if (end_pfn > max_pfn)
133 		end_pfn = max_pfn;
134 
135 	for (; pfn < end_pfn; pfn++) {
136 		bit = pfn % BITMAP_CHUNK_BITS;
137 		if (!bit)
138 			*out = 0ULL;
139 		folio = page_idle_get_folio(pfn);
140 		if (folio) {
141 			if (folio_test_idle(folio)) {
142 				/*
143 				 * The page might have been referenced via a
144 				 * pte, in which case it is not idle. Clear
145 				 * refs and recheck.
146 				 */
147 				page_idle_clear_pte_refs(folio);
148 				if (folio_test_idle(folio))
149 					*out |= 1ULL << bit;
150 			}
151 			folio_put(folio);
152 		}
153 		if (bit == BITMAP_CHUNK_BITS - 1)
154 			out++;
155 		cond_resched();
156 	}
157 	return (char *)out - buf;
158 }
159 
160 static ssize_t page_idle_bitmap_write(struct file *file, struct kobject *kobj,
161 				      const struct bin_attribute *attr, char *buf,
162 				      loff_t pos, size_t count)
163 {
164 	const u64 *in = (u64 *)buf;
165 	struct folio *folio;
166 	unsigned long pfn, end_pfn;
167 	int bit;
168 
169 	if (pos % BITMAP_CHUNK_SIZE || count % BITMAP_CHUNK_SIZE)
170 		return -EINVAL;
171 
172 	pfn = pos * BITS_PER_BYTE;
173 	if (pfn >= max_pfn)
174 		return -ENXIO;
175 
176 	end_pfn = pfn + count * BITS_PER_BYTE;
177 	if (end_pfn > max_pfn)
178 		end_pfn = max_pfn;
179 
180 	for (; pfn < end_pfn; pfn++) {
181 		bit = pfn % BITMAP_CHUNK_BITS;
182 		if ((*in >> bit) & 1) {
183 			folio = page_idle_get_folio(pfn);
184 			if (folio) {
185 				page_idle_clear_pte_refs(folio);
186 				folio_set_idle(folio);
187 				folio_put(folio);
188 			}
189 		}
190 		if (bit == BITMAP_CHUNK_BITS - 1)
191 			in++;
192 		cond_resched();
193 	}
194 	return (char *)in - buf;
195 }
196 
197 static const struct bin_attribute page_idle_bitmap_attr =
198 		__BIN_ATTR(bitmap, 0600,
199 			   page_idle_bitmap_read, page_idle_bitmap_write, 0);
200 
201 static const struct bin_attribute *const page_idle_bin_attrs[] = {
202 	&page_idle_bitmap_attr,
203 	NULL,
204 };
205 
206 static const struct attribute_group page_idle_attr_group = {
207 	.bin_attrs = page_idle_bin_attrs,
208 	.name = "page_idle",
209 };
210 
211 static int __init page_idle_init(void)
212 {
213 	int err;
214 
215 	err = sysfs_create_group(mm_kobj, &page_idle_attr_group);
216 	if (err) {
217 		pr_err("page_idle: register sysfs failed\n");
218 		return err;
219 	}
220 	return 0;
221 }
222 subsys_initcall(page_idle_init);
223