1 // SPDX-License-Identifier: GPL-2.0
3 * Lockless hierarchical page accounting & limiting
5 * Copyright (C) 2014 Red Hat, Inc., Johannes Weiner
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>
16 static void propagate_protected_usage(struct page_counter *c,
19 unsigned long protected, old_protected;
25 if (c->min || atomic_long_read(&c->min_usage)) {
26 protected = min(usage, c->min);
27 old_protected = atomic_long_xchg(&c->min_usage, protected);
28 delta = protected - old_protected;
30 atomic_long_add(delta, &c->parent->children_min_usage);
33 if (c->low || atomic_long_read(&c->low_usage)) {
34 protected = min(usage, c->low);
35 old_protected = atomic_long_xchg(&c->low_usage, protected);
36 delta = protected - old_protected;
38 atomic_long_add(delta, &c->parent->children_low_usage);
43 * page_counter_cancel - take pages out of the local counter
45 * @nr_pages: number of pages to cancel
47 void page_counter_cancel(struct page_counter *counter, unsigned long nr_pages)
51 new = atomic_long_sub_return(nr_pages, &counter->usage);
52 propagate_protected_usage(counter, new);
53 /* More uncharges than charges? */
54 WARN_ON_ONCE(new < 0);
58 * page_counter_charge - hierarchically charge pages
60 * @nr_pages: number of pages to charge
62 * NOTE: This does not consider any configured counter limits.
64 void page_counter_charge(struct page_counter *counter, unsigned long nr_pages)
66 struct page_counter *c;
68 for (c = counter; c; c = c->parent) {
71 new = atomic_long_add_return(nr_pages, &c->usage);
72 propagate_protected_usage(counter, new);
74 * This is indeed racy, but we can live with some
75 * inaccuracy in the watermark.
77 if (new > c->watermark)
83 * page_counter_try_charge - try to hierarchically charge pages
85 * @nr_pages: number of pages to charge
86 * @fail: points first counter to hit its limit, if any
88 * Returns %true on success, or %false and @fail if the counter or one
89 * of its ancestors has hit its configured limit.
91 bool page_counter_try_charge(struct page_counter *counter,
92 unsigned long nr_pages,
93 struct page_counter **fail)
95 struct page_counter *c;
97 for (c = counter; c; c = c->parent) {
100 * Charge speculatively to avoid an expensive CAS. If
101 * a bigger charge fails, it might falsely lock out a
102 * racing smaller charge and send it into reclaim
103 * early, but the error is limited to the difference
104 * between the two sizes, which is less than 2M/4M in
105 * case of a THP locking out a regular page charge.
107 * The atomic_long_add_return() implies a full memory
108 * barrier between incrementing the count and reading
109 * the limit. When racing with page_counter_limit(),
110 * we either see the new limit or the setter sees the
111 * counter has changed and retries.
113 new = atomic_long_add_return(nr_pages, &c->usage);
115 atomic_long_sub(nr_pages, &c->usage);
116 propagate_protected_usage(counter, new);
118 * This is racy, but we can live with some
119 * inaccuracy in the failcnt.
125 propagate_protected_usage(counter, new);
127 * Just like with failcnt, we can live with some
128 * inaccuracy in the watermark.
130 if (new > c->watermark)
136 for (c = counter; c != *fail; c = c->parent)
137 page_counter_cancel(c, nr_pages);
143 * page_counter_uncharge - hierarchically uncharge pages
145 * @nr_pages: number of pages to uncharge
147 void page_counter_uncharge(struct page_counter *counter, unsigned long nr_pages)
149 struct page_counter *c;
151 for (c = counter; c; c = c->parent)
152 page_counter_cancel(c, nr_pages);
156 * page_counter_set_max - set the maximum number of pages allowed
158 * @nr_pages: limit to set
160 * Returns 0 on success, -EBUSY if the current number of pages on the
161 * counter already exceeds the specified limit.
163 * The caller must serialize invocations on the same counter.
165 int page_counter_set_max(struct page_counter *counter, unsigned long nr_pages)
172 * Update the limit while making sure that it's not
173 * below the concurrently-changing counter value.
175 * The xchg implies two full memory barriers before
176 * and after, so the read-swap-read is ordered and
177 * ensures coherency with page_counter_try_charge():
178 * that function modifies the count before checking
179 * the limit, so if it sees the old limit, we see the
180 * modified counter and retry.
182 usage = atomic_long_read(&counter->usage);
184 if (usage > nr_pages)
187 old = xchg(&counter->max, nr_pages);
189 if (atomic_long_read(&counter->usage) <= usage)
198 * page_counter_set_min - set the amount of protected memory
200 * @nr_pages: value to set
202 * The caller must serialize invocations on the same counter.
204 void page_counter_set_min(struct page_counter *counter, unsigned long nr_pages)
206 struct page_counter *c;
208 counter->min = nr_pages;
210 for (c = counter; c; c = c->parent)
211 propagate_protected_usage(c, atomic_long_read(&c->usage));
215 * page_counter_set_low - set the amount of protected memory
217 * @nr_pages: value to set
219 * The caller must serialize invocations on the same counter.
221 void page_counter_set_low(struct page_counter *counter, unsigned long nr_pages)
223 struct page_counter *c;
225 counter->low = nr_pages;
227 for (c = counter; c; c = c->parent)
228 propagate_protected_usage(c, atomic_long_read(&c->usage));
232 * page_counter_memparse - memparse() for page counter limits
233 * @buf: string to parse
234 * @max: string meaning maximum possible value
235 * @nr_pages: returns the result in number of pages
237 * Returns -EINVAL, or 0 and @nr_pages on success. @nr_pages will be
238 * limited to %PAGE_COUNTER_MAX.
240 int page_counter_memparse(const char *buf, const char *max,
241 unsigned long *nr_pages)
246 if (!strcmp(buf, max)) {
247 *nr_pages = PAGE_COUNTER_MAX;
251 bytes = memparse(buf, &end);
255 *nr_pages = min(bytes / PAGE_SIZE, (u64)PAGE_COUNTER_MAX);