xref: /openbmc/linux/mm/page_idle.c (revision b96fc2f3)
1 #include <linux/init.h>
2 #include <linux/bootmem.h>
3 #include <linux/fs.h>
4 #include <linux/sysfs.h>
5 #include <linux/kobject.h>
6 #include <linux/mm.h>
7 #include <linux/mmzone.h>
8 #include <linux/pagemap.h>
9 #include <linux/rmap.h>
10 #include <linux/mmu_notifier.h>
11 #include <linux/page_ext.h>
12 #include <linux/page_idle.h>
13 
14 #define BITMAP_CHUNK_SIZE	sizeof(u64)
15 #define BITMAP_CHUNK_BITS	(BITMAP_CHUNK_SIZE * BITS_PER_BYTE)
16 
17 /*
18  * Idle page tracking only considers user memory pages, for other types of
19  * pages the idle flag is always unset and an attempt to set it is silently
20  * ignored.
21  *
22  * We treat a page as a user memory page if it is on an LRU list, because it is
23  * always safe to pass such a page to rmap_walk(), which is essential for idle
24  * page tracking. With such an indicator of user pages we can skip isolated
25  * pages, but since there are not usually many of them, it will hardly affect
26  * the overall result.
27  *
28  * This function tries to get a user memory page by pfn as described above.
29  */
30 static struct page *page_idle_get_page(unsigned long pfn)
31 {
32 	struct page *page;
33 	struct zone *zone;
34 
35 	if (!pfn_valid(pfn))
36 		return NULL;
37 
38 	page = pfn_to_page(pfn);
39 	if (!page || !PageLRU(page) ||
40 	    !get_page_unless_zero(page))
41 		return NULL;
42 
43 	zone = page_zone(page);
44 	spin_lock_irq(&zone->lru_lock);
45 	if (unlikely(!PageLRU(page))) {
46 		put_page(page);
47 		page = NULL;
48 	}
49 	spin_unlock_irq(&zone->lru_lock);
50 	return page;
51 }
52 
53 static int page_idle_clear_pte_refs_one(struct page *page,
54 					struct vm_area_struct *vma,
55 					unsigned long addr, void *arg)
56 {
57 	struct mm_struct *mm = vma->vm_mm;
58 	spinlock_t *ptl;
59 	pmd_t *pmd;
60 	pte_t *pte;
61 	bool referenced = false;
62 
63 	if (unlikely(PageTransHuge(page))) {
64 		pmd = page_check_address_pmd(page, mm, addr,
65 					     PAGE_CHECK_ADDRESS_PMD_FLAG, &ptl);
66 		if (pmd) {
67 			referenced = pmdp_clear_young_notify(vma, addr, pmd);
68 			spin_unlock(ptl);
69 		}
70 	} else {
71 		pte = page_check_address(page, mm, addr, &ptl, 0);
72 		if (pte) {
73 			referenced = ptep_clear_young_notify(vma, addr, pte);
74 			pte_unmap_unlock(pte, ptl);
75 		}
76 	}
77 	if (referenced) {
78 		clear_page_idle(page);
79 		/*
80 		 * We cleared the referenced bit in a mapping to this page. To
81 		 * avoid interference with page reclaim, mark it young so that
82 		 * page_referenced() will return > 0.
83 		 */
84 		set_page_young(page);
85 	}
86 	return SWAP_AGAIN;
87 }
88 
89 static void page_idle_clear_pte_refs(struct page *page)
90 {
91 	/*
92 	 * Since rwc.arg is unused, rwc is effectively immutable, so we
93 	 * can make it static const to save some cycles and stack.
94 	 */
95 	static const struct rmap_walk_control rwc = {
96 		.rmap_one = page_idle_clear_pte_refs_one,
97 		.anon_lock = page_lock_anon_vma_read,
98 	};
99 	bool need_lock;
100 
101 	if (!page_mapped(page) ||
102 	    !page_rmapping(page))
103 		return;
104 
105 	need_lock = !PageAnon(page) || PageKsm(page);
106 	if (need_lock && !trylock_page(page))
107 		return;
108 
109 	rmap_walk(page, (struct rmap_walk_control *)&rwc);
110 
111 	if (need_lock)
112 		unlock_page(page);
113 }
114 
115 static ssize_t page_idle_bitmap_read(struct file *file, struct kobject *kobj,
116 				     struct bin_attribute *attr, char *buf,
117 				     loff_t pos, size_t count)
118 {
119 	u64 *out = (u64 *)buf;
120 	struct page *page;
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 = ALIGN(max_pfn, BITMAP_CHUNK_BITS);
134 
135 	for (; pfn < end_pfn; pfn++) {
136 		bit = pfn % BITMAP_CHUNK_BITS;
137 		if (!bit)
138 			*out = 0ULL;
139 		page = page_idle_get_page(pfn);
140 		if (page) {
141 			if (page_is_idle(page)) {
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(page);
148 				if (page_is_idle(page))
149 					*out |= 1ULL << bit;
150 			}
151 			put_page(page);
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 				      struct bin_attribute *attr, char *buf,
162 				      loff_t pos, size_t count)
163 {
164 	const u64 *in = (u64 *)buf;
165 	struct page *page;
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 = ALIGN(max_pfn, BITMAP_CHUNK_BITS);
179 
180 	for (; pfn < end_pfn; pfn++) {
181 		bit = pfn % BITMAP_CHUNK_BITS;
182 		if ((*in >> bit) & 1) {
183 			page = page_idle_get_page(pfn);
184 			if (page) {
185 				page_idle_clear_pte_refs(page);
186 				set_page_idle(page);
187 				put_page(page);
188 			}
189 		}
190 		if (bit == BITMAP_CHUNK_BITS - 1)
191 			in++;
192 		cond_resched();
193 	}
194 	return (char *)in - buf;
195 }
196 
197 static struct bin_attribute page_idle_bitmap_attr =
198 		__BIN_ATTR(bitmap, S_IRUSR | S_IWUSR,
199 			   page_idle_bitmap_read, page_idle_bitmap_write, 0);
200 
201 static struct bin_attribute *page_idle_bin_attrs[] = {
202 	&page_idle_bitmap_attr,
203 	NULL,
204 };
205 
206 static struct attribute_group page_idle_attr_group = {
207 	.bin_attrs = page_idle_bin_attrs,
208 	.name = "page_idle",
209 };
210 
211 #ifndef CONFIG_64BIT
212 static bool need_page_idle(void)
213 {
214 	return true;
215 }
216 struct page_ext_operations page_idle_ops = {
217 	.need = need_page_idle,
218 };
219 #endif
220 
221 static int __init page_idle_init(void)
222 {
223 	int err;
224 
225 	err = sysfs_create_group(mm_kobj, &page_idle_attr_group);
226 	if (err) {
227 		pr_err("page_idle: register sysfs failed\n");
228 		return err;
229 	}
230 	return 0;
231 }
232 subsys_initcall(page_idle_init);
233