xref: /openbmc/linux/mm/page_counter.c (revision c0e297dc)
1 /*
2  * Lockless hierarchical page accounting & limiting
3  *
4  * Copyright (C) 2014 Red Hat, Inc., Johannes Weiner
5  */
6 
7 #include <linux/page_counter.h>
8 #include <linux/atomic.h>
9 #include <linux/kernel.h>
10 #include <linux/string.h>
11 #include <linux/sched.h>
12 #include <linux/bug.h>
13 #include <asm/page.h>
14 
15 /**
16  * page_counter_cancel - take pages out of the local counter
17  * @counter: counter
18  * @nr_pages: number of pages to cancel
19  */
20 void page_counter_cancel(struct page_counter *counter, unsigned long nr_pages)
21 {
22 	long new;
23 
24 	new = atomic_long_sub_return(nr_pages, &counter->count);
25 	/* More uncharges than charges? */
26 	WARN_ON_ONCE(new < 0);
27 }
28 
29 /**
30  * page_counter_charge - hierarchically charge pages
31  * @counter: counter
32  * @nr_pages: number of pages to charge
33  *
34  * NOTE: This does not consider any configured counter limits.
35  */
36 void page_counter_charge(struct page_counter *counter, unsigned long nr_pages)
37 {
38 	struct page_counter *c;
39 
40 	for (c = counter; c; c = c->parent) {
41 		long new;
42 
43 		new = atomic_long_add_return(nr_pages, &c->count);
44 		/*
45 		 * This is indeed racy, but we can live with some
46 		 * inaccuracy in the watermark.
47 		 */
48 		if (new > c->watermark)
49 			c->watermark = new;
50 	}
51 }
52 
53 /**
54  * page_counter_try_charge - try to hierarchically charge pages
55  * @counter: counter
56  * @nr_pages: number of pages to charge
57  * @fail: points first counter to hit its limit, if any
58  *
59  * Returns 0 on success, or -ENOMEM and @fail if the counter or one of
60  * its ancestors has hit its configured limit.
61  */
62 int page_counter_try_charge(struct page_counter *counter,
63 			    unsigned long nr_pages,
64 			    struct page_counter **fail)
65 {
66 	struct page_counter *c;
67 
68 	for (c = counter; c; c = c->parent) {
69 		long new;
70 		/*
71 		 * Charge speculatively to avoid an expensive CAS.  If
72 		 * a bigger charge fails, it might falsely lock out a
73 		 * racing smaller charge and send it into reclaim
74 		 * early, but the error is limited to the difference
75 		 * between the two sizes, which is less than 2M/4M in
76 		 * case of a THP locking out a regular page charge.
77 		 *
78 		 * The atomic_long_add_return() implies a full memory
79 		 * barrier between incrementing the count and reading
80 		 * the limit.  When racing with page_counter_limit(),
81 		 * we either see the new limit or the setter sees the
82 		 * counter has changed and retries.
83 		 */
84 		new = atomic_long_add_return(nr_pages, &c->count);
85 		if (new > c->limit) {
86 			atomic_long_sub(nr_pages, &c->count);
87 			/*
88 			 * This is racy, but we can live with some
89 			 * inaccuracy in the failcnt.
90 			 */
91 			c->failcnt++;
92 			*fail = c;
93 			goto failed;
94 		}
95 		/*
96 		 * Just like with failcnt, we can live with some
97 		 * inaccuracy in the watermark.
98 		 */
99 		if (new > c->watermark)
100 			c->watermark = new;
101 	}
102 	return 0;
103 
104 failed:
105 	for (c = counter; c != *fail; c = c->parent)
106 		page_counter_cancel(c, nr_pages);
107 
108 	return -ENOMEM;
109 }
110 
111 /**
112  * page_counter_uncharge - hierarchically uncharge pages
113  * @counter: counter
114  * @nr_pages: number of pages to uncharge
115  */
116 void page_counter_uncharge(struct page_counter *counter, unsigned long nr_pages)
117 {
118 	struct page_counter *c;
119 
120 	for (c = counter; c; c = c->parent)
121 		page_counter_cancel(c, nr_pages);
122 }
123 
124 /**
125  * page_counter_limit - limit the number of pages allowed
126  * @counter: counter
127  * @limit: limit to set
128  *
129  * Returns 0 on success, -EBUSY if the current number of pages on the
130  * counter already exceeds the specified limit.
131  *
132  * The caller must serialize invocations on the same counter.
133  */
134 int page_counter_limit(struct page_counter *counter, unsigned long limit)
135 {
136 	for (;;) {
137 		unsigned long old;
138 		long count;
139 
140 		/*
141 		 * Update the limit while making sure that it's not
142 		 * below the concurrently-changing counter value.
143 		 *
144 		 * The xchg implies two full memory barriers before
145 		 * and after, so the read-swap-read is ordered and
146 		 * ensures coherency with page_counter_try_charge():
147 		 * that function modifies the count before checking
148 		 * the limit, so if it sees the old limit, we see the
149 		 * modified counter and retry.
150 		 */
151 		count = atomic_long_read(&counter->count);
152 
153 		if (count > limit)
154 			return -EBUSY;
155 
156 		old = xchg(&counter->limit, limit);
157 
158 		if (atomic_long_read(&counter->count) <= count)
159 			return 0;
160 
161 		counter->limit = old;
162 		cond_resched();
163 	}
164 }
165 
166 /**
167  * page_counter_memparse - memparse() for page counter limits
168  * @buf: string to parse
169  * @max: string meaning maximum possible value
170  * @nr_pages: returns the result in number of pages
171  *
172  * Returns -EINVAL, or 0 and @nr_pages on success.  @nr_pages will be
173  * limited to %PAGE_COUNTER_MAX.
174  */
175 int page_counter_memparse(const char *buf, const char *max,
176 			  unsigned long *nr_pages)
177 {
178 	char *end;
179 	u64 bytes;
180 
181 	if (!strcmp(buf, max)) {
182 		*nr_pages = PAGE_COUNTER_MAX;
183 		return 0;
184 	}
185 
186 	bytes = memparse(buf, &end);
187 	if (*end != '\0')
188 		return -EINVAL;
189 
190 	*nr_pages = min(bytes / PAGE_SIZE, (u64)PAGE_COUNTER_MAX);
191 
192 	return 0;
193 }
194