xref: /linux/mm/page_counter.c (revision c5d3cdad688ed75fb311a3a671eb30ba7106d7d3)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Lockless hierarchical page accounting & limiting
4  *
5  * Copyright (C) 2014 Red Hat, Inc., Johannes Weiner
6  */
7 
8 #include <linux/page_counter.h>
9 #include <linux/atomic.h>
10 #include <linux/kernel.h>
11 #include <linux/string.h>
12 #include <linux/sched.h>
13 #include <linux/bug.h>
14 #include <asm/page.h>
15 
16 static void propagate_protected_usage(struct page_counter *c,
17 				      unsigned long usage)
18 {
19 	unsigned long protected, old_protected;
20 	unsigned long low, min;
21 	long delta;
22 
23 	if (!c->parent)
24 		return;
25 
26 	min = READ_ONCE(c->min);
27 	if (min || atomic_long_read(&c->min_usage)) {
28 		protected = min(usage, min);
29 		old_protected = atomic_long_xchg(&c->min_usage, protected);
30 		delta = protected - old_protected;
31 		if (delta)
32 			atomic_long_add(delta, &c->parent->children_min_usage);
33 	}
34 
35 	low = READ_ONCE(c->low);
36 	if (low || atomic_long_read(&c->low_usage)) {
37 		protected = min(usage, low);
38 		old_protected = atomic_long_xchg(&c->low_usage, protected);
39 		delta = protected - old_protected;
40 		if (delta)
41 			atomic_long_add(delta, &c->parent->children_low_usage);
42 	}
43 }
44 
45 /**
46  * page_counter_cancel - take pages out of the local counter
47  * @counter: counter
48  * @nr_pages: number of pages to cancel
49  */
50 void page_counter_cancel(struct page_counter *counter, unsigned long nr_pages)
51 {
52 	long new;
53 
54 	new = atomic_long_sub_return(nr_pages, &counter->usage);
55 	propagate_protected_usage(counter, new);
56 	/* More uncharges than charges? */
57 	WARN_ON_ONCE(new < 0);
58 }
59 
60 /**
61  * page_counter_charge - hierarchically charge pages
62  * @counter: counter
63  * @nr_pages: number of pages to charge
64  *
65  * NOTE: This does not consider any configured counter limits.
66  */
67 void page_counter_charge(struct page_counter *counter, unsigned long nr_pages)
68 {
69 	struct page_counter *c;
70 
71 	for (c = counter; c; c = c->parent) {
72 		long new;
73 
74 		new = atomic_long_add_return(nr_pages, &c->usage);
75 		propagate_protected_usage(counter, new);
76 		/*
77 		 * This is indeed racy, but we can live with some
78 		 * inaccuracy in the watermark.
79 		 */
80 		if (new > c->watermark)
81 			c->watermark = new;
82 	}
83 }
84 
85 /**
86  * page_counter_try_charge - try to hierarchically charge pages
87  * @counter: counter
88  * @nr_pages: number of pages to charge
89  * @fail: points first counter to hit its limit, if any
90  *
91  * Returns %true on success, or %false and @fail if the counter or one
92  * of its ancestors has hit its configured limit.
93  */
94 bool page_counter_try_charge(struct page_counter *counter,
95 			     unsigned long nr_pages,
96 			     struct page_counter **fail)
97 {
98 	struct page_counter *c;
99 
100 	for (c = counter; c; c = c->parent) {
101 		long new;
102 		/*
103 		 * Charge speculatively to avoid an expensive CAS.  If
104 		 * a bigger charge fails, it might falsely lock out a
105 		 * racing smaller charge and send it into reclaim
106 		 * early, but the error is limited to the difference
107 		 * between the two sizes, which is less than 2M/4M in
108 		 * case of a THP locking out a regular page charge.
109 		 *
110 		 * The atomic_long_add_return() implies a full memory
111 		 * barrier between incrementing the count and reading
112 		 * the limit.  When racing with page_counter_limit(),
113 		 * we either see the new limit or the setter sees the
114 		 * counter has changed and retries.
115 		 */
116 		new = atomic_long_add_return(nr_pages, &c->usage);
117 		if (new > c->max) {
118 			atomic_long_sub(nr_pages, &c->usage);
119 			propagate_protected_usage(counter, new);
120 			/*
121 			 * This is racy, but we can live with some
122 			 * inaccuracy in the failcnt.
123 			 */
124 			c->failcnt++;
125 			*fail = c;
126 			goto failed;
127 		}
128 		propagate_protected_usage(counter, new);
129 		/*
130 		 * Just like with failcnt, we can live with some
131 		 * inaccuracy in the watermark.
132 		 */
133 		if (new > c->watermark)
134 			c->watermark = new;
135 	}
136 	return true;
137 
138 failed:
139 	for (c = counter; c != *fail; c = c->parent)
140 		page_counter_cancel(c, nr_pages);
141 
142 	return false;
143 }
144 
145 /**
146  * page_counter_uncharge - hierarchically uncharge pages
147  * @counter: counter
148  * @nr_pages: number of pages to uncharge
149  */
150 void page_counter_uncharge(struct page_counter *counter, unsigned long nr_pages)
151 {
152 	struct page_counter *c;
153 
154 	for (c = counter; c; c = c->parent)
155 		page_counter_cancel(c, nr_pages);
156 }
157 
158 /**
159  * page_counter_set_max - set the maximum number of pages allowed
160  * @counter: counter
161  * @nr_pages: limit to set
162  *
163  * Returns 0 on success, -EBUSY if the current number of pages on the
164  * counter already exceeds the specified limit.
165  *
166  * The caller must serialize invocations on the same counter.
167  */
168 int page_counter_set_max(struct page_counter *counter, unsigned long nr_pages)
169 {
170 	for (;;) {
171 		unsigned long old;
172 		long usage;
173 
174 		/*
175 		 * Update the limit while making sure that it's not
176 		 * below the concurrently-changing counter value.
177 		 *
178 		 * The xchg implies two full memory barriers before
179 		 * and after, so the read-swap-read is ordered and
180 		 * ensures coherency with page_counter_try_charge():
181 		 * that function modifies the count before checking
182 		 * the limit, so if it sees the old limit, we see the
183 		 * modified counter and retry.
184 		 */
185 		usage = atomic_long_read(&counter->usage);
186 
187 		if (usage > nr_pages)
188 			return -EBUSY;
189 
190 		old = xchg(&counter->max, nr_pages);
191 
192 		if (atomic_long_read(&counter->usage) <= usage)
193 			return 0;
194 
195 		counter->max = old;
196 		cond_resched();
197 	}
198 }
199 
200 /**
201  * page_counter_set_min - set the amount of protected memory
202  * @counter: counter
203  * @nr_pages: value to set
204  *
205  * The caller must serialize invocations on the same counter.
206  */
207 void page_counter_set_min(struct page_counter *counter, unsigned long nr_pages)
208 {
209 	struct page_counter *c;
210 
211 	WRITE_ONCE(counter->min, nr_pages);
212 
213 	for (c = counter; c; c = c->parent)
214 		propagate_protected_usage(c, atomic_long_read(&c->usage));
215 }
216 
217 /**
218  * page_counter_set_low - set the amount of protected memory
219  * @counter: counter
220  * @nr_pages: value to set
221  *
222  * The caller must serialize invocations on the same counter.
223  */
224 void page_counter_set_low(struct page_counter *counter, unsigned long nr_pages)
225 {
226 	struct page_counter *c;
227 
228 	WRITE_ONCE(counter->low, nr_pages);
229 
230 	for (c = counter; c; c = c->parent)
231 		propagate_protected_usage(c, atomic_long_read(&c->usage));
232 }
233 
234 /**
235  * page_counter_memparse - memparse() for page counter limits
236  * @buf: string to parse
237  * @max: string meaning maximum possible value
238  * @nr_pages: returns the result in number of pages
239  *
240  * Returns -EINVAL, or 0 and @nr_pages on success.  @nr_pages will be
241  * limited to %PAGE_COUNTER_MAX.
242  */
243 int page_counter_memparse(const char *buf, const char *max,
244 			  unsigned long *nr_pages)
245 {
246 	char *end;
247 	u64 bytes;
248 
249 	if (!strcmp(buf, max)) {
250 		*nr_pages = PAGE_COUNTER_MAX;
251 		return 0;
252 	}
253 
254 	bytes = memparse(buf, &end);
255 	if (*end != '\0')
256 		return -EINVAL;
257 
258 	*nr_pages = min(bytes / PAGE_SIZE, (u64)PAGE_COUNTER_MAX);
259 
260 	return 0;
261 }
262