xref: /openbmc/linux/mm/page_counter.c (revision b737eecd)
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(c, new);
76 		/*
77 		 * This is indeed racy, but we can live with some
78 		 * inaccuracy in the watermark.
79 		 */
80 		if (new > READ_ONCE(c->watermark))
81 			WRITE_ONCE(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_set_max(),
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(c, new);
120 			/*
121 			 * This is racy, but we can live with some
122 			 * inaccuracy in the failcnt which is only used
123 			 * to report stats.
124 			 */
125 			data_race(c->failcnt++);
126 			*fail = c;
127 			goto failed;
128 		}
129 		propagate_protected_usage(c, new);
130 		/*
131 		 * Just like with failcnt, we can live with some
132 		 * inaccuracy in the watermark.
133 		 */
134 		if (new > READ_ONCE(c->watermark))
135 			WRITE_ONCE(c->watermark, new);
136 	}
137 	return true;
138 
139 failed:
140 	for (c = counter; c != *fail; c = c->parent)
141 		page_counter_cancel(c, nr_pages);
142 
143 	return false;
144 }
145 
146 /**
147  * page_counter_uncharge - hierarchically uncharge pages
148  * @counter: counter
149  * @nr_pages: number of pages to uncharge
150  */
151 void page_counter_uncharge(struct page_counter *counter, unsigned long nr_pages)
152 {
153 	struct page_counter *c;
154 
155 	for (c = counter; c; c = c->parent)
156 		page_counter_cancel(c, nr_pages);
157 }
158 
159 /**
160  * page_counter_set_max - set the maximum number of pages allowed
161  * @counter: counter
162  * @nr_pages: limit to set
163  *
164  * Returns 0 on success, -EBUSY if the current number of pages on the
165  * counter already exceeds the specified limit.
166  *
167  * The caller must serialize invocations on the same counter.
168  */
169 int page_counter_set_max(struct page_counter *counter, unsigned long nr_pages)
170 {
171 	for (;;) {
172 		unsigned long old;
173 		long usage;
174 
175 		/*
176 		 * Update the limit while making sure that it's not
177 		 * below the concurrently-changing counter value.
178 		 *
179 		 * The xchg implies two full memory barriers before
180 		 * and after, so the read-swap-read is ordered and
181 		 * ensures coherency with page_counter_try_charge():
182 		 * that function modifies the count before checking
183 		 * the limit, so if it sees the old limit, we see the
184 		 * modified counter and retry.
185 		 */
186 		usage = page_counter_read(counter);
187 
188 		if (usage > nr_pages)
189 			return -EBUSY;
190 
191 		old = xchg(&counter->max, nr_pages);
192 
193 		if (page_counter_read(counter) <= usage)
194 			return 0;
195 
196 		counter->max = old;
197 		cond_resched();
198 	}
199 }
200 
201 /**
202  * page_counter_set_min - set the amount of protected memory
203  * @counter: counter
204  * @nr_pages: value to set
205  *
206  * The caller must serialize invocations on the same counter.
207  */
208 void page_counter_set_min(struct page_counter *counter, unsigned long nr_pages)
209 {
210 	struct page_counter *c;
211 
212 	WRITE_ONCE(counter->min, nr_pages);
213 
214 	for (c = counter; c; c = c->parent)
215 		propagate_protected_usage(c, atomic_long_read(&c->usage));
216 }
217 
218 /**
219  * page_counter_set_low - set the amount of protected memory
220  * @counter: counter
221  * @nr_pages: value to set
222  *
223  * The caller must serialize invocations on the same counter.
224  */
225 void page_counter_set_low(struct page_counter *counter, unsigned long nr_pages)
226 {
227 	struct page_counter *c;
228 
229 	WRITE_ONCE(counter->low, nr_pages);
230 
231 	for (c = counter; c; c = c->parent)
232 		propagate_protected_usage(c, atomic_long_read(&c->usage));
233 }
234 
235 /**
236  * page_counter_memparse - memparse() for page counter limits
237  * @buf: string to parse
238  * @max: string meaning maximum possible value
239  * @nr_pages: returns the result in number of pages
240  *
241  * Returns -EINVAL, or 0 and @nr_pages on success.  @nr_pages will be
242  * limited to %PAGE_COUNTER_MAX.
243  */
244 int page_counter_memparse(const char *buf, const char *max,
245 			  unsigned long *nr_pages)
246 {
247 	char *end;
248 	u64 bytes;
249 
250 	if (!strcmp(buf, max)) {
251 		*nr_pages = PAGE_COUNTER_MAX;
252 		return 0;
253 	}
254 
255 	bytes = memparse(buf, &end);
256 	if (*end != '\0')
257 		return -EINVAL;
258 
259 	*nr_pages = min(bytes / PAGE_SIZE, (u64)PAGE_COUNTER_MAX);
260 
261 	return 0;
262 }
263