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 long delta;
21
22 if (!c->parent)
23 return;
24
25 protected = min(usage, READ_ONCE(c->min));
26 old_protected = atomic_long_read(&c->min_usage);
27 if (protected != old_protected) {
28 old_protected = atomic_long_xchg(&c->min_usage, protected);
29 delta = protected - old_protected;
30 if (delta)
31 atomic_long_add(delta, &c->parent->children_min_usage);
32 }
33
34 protected = min(usage, READ_ONCE(c->low));
35 old_protected = atomic_long_read(&c->low_usage);
36 if (protected != old_protected) {
37 old_protected = atomic_long_xchg(&c->low_usage, protected);
38 delta = protected - old_protected;
39 if (delta)
40 atomic_long_add(delta, &c->parent->children_low_usage);
41 }
42 }
43
44 /**
45 * page_counter_cancel - take pages out of the local counter
46 * @counter: counter
47 * @nr_pages: number of pages to cancel
48 */
49 void page_counter_cancel(struct page_counter *counter, unsigned long nr_pages)
50 {
51 long new;
52
53 new = atomic_long_sub_return(nr_pages, &counter->usage);
54 /* More uncharges than charges? */
55 if (WARN_ONCE(new < 0, "page_counter underflow: %ld nr_pages=%lu\n",
56 new, nr_pages)) {
57 new = 0;
58 atomic_long_set(&counter->usage, new);
59 }
60 propagate_protected_usage(counter, new);
61 }
62
63 /**
64 * page_counter_charge - hierarchically charge pages
65 * @counter: counter
66 * @nr_pages: number of pages to charge
67 *
68 * NOTE: This does not consider any configured counter limits.
69 */
70 void page_counter_charge(struct page_counter *counter, unsigned long nr_pages)
71 {
72 struct page_counter *c;
73
74 for (c = counter; c; c = c->parent) {
75 long new;
76
77 new = atomic_long_add_return(nr_pages, &c->usage);
78 propagate_protected_usage(c, new);
79 /*
80 * This is indeed racy, but we can live with some
81 * inaccuracy in the watermark.
82 */
83 if (new > READ_ONCE(c->watermark))
84 WRITE_ONCE(c->watermark, new);
85 }
86 }
87
88 /**
89 * page_counter_try_charge - try to hierarchically charge pages
90 * @counter: counter
91 * @nr_pages: number of pages to charge
92 * @fail: points first counter to hit its limit, if any
93 *
94 * Returns %true on success, or %false and @fail if the counter or one
95 * of its ancestors has hit its configured limit.
96 */
97 bool page_counter_try_charge(struct page_counter *counter,
98 unsigned long nr_pages,
99 struct page_counter **fail)
100 {
101 struct page_counter *c;
102
103 for (c = counter; c; c = c->parent) {
104 long new;
105 /*
106 * Charge speculatively to avoid an expensive CAS. If
107 * a bigger charge fails, it might falsely lock out a
108 * racing smaller charge and send it into reclaim
109 * early, but the error is limited to the difference
110 * between the two sizes, which is less than 2M/4M in
111 * case of a THP locking out a regular page charge.
112 *
113 * The atomic_long_add_return() implies a full memory
114 * barrier between incrementing the count and reading
115 * the limit. When racing with page_counter_set_max(),
116 * we either see the new limit or the setter sees the
117 * counter has changed and retries.
118 */
119 new = atomic_long_add_return(nr_pages, &c->usage);
120 if (new > c->max) {
121 atomic_long_sub(nr_pages, &c->usage);
122 /*
123 * This is racy, but we can live with some
124 * inaccuracy in the failcnt which is only used
125 * to report stats.
126 */
127 data_race(c->failcnt++);
128 *fail = c;
129 goto failed;
130 }
131 propagate_protected_usage(c, new);
132 /*
133 * Just like with failcnt, we can live with some
134 * inaccuracy in the watermark.
135 */
136 if (new > READ_ONCE(c->watermark))
137 WRITE_ONCE(c->watermark, new);
138 }
139 return true;
140
141 failed:
142 for (c = counter; c != *fail; c = c->parent)
143 page_counter_cancel(c, nr_pages);
144
145 return false;
146 }
147
148 /**
149 * page_counter_uncharge - hierarchically uncharge pages
150 * @counter: counter
151 * @nr_pages: number of pages to uncharge
152 */
153 void page_counter_uncharge(struct page_counter *counter, unsigned long nr_pages)
154 {
155 struct page_counter *c;
156
157 for (c = counter; c; c = c->parent)
158 page_counter_cancel(c, nr_pages);
159 }
160
161 /**
162 * page_counter_set_max - set the maximum number of pages allowed
163 * @counter: counter
164 * @nr_pages: limit to set
165 *
166 * Returns 0 on success, -EBUSY if the current number of pages on the
167 * counter already exceeds the specified limit.
168 *
169 * The caller must serialize invocations on the same counter.
170 */
171 int page_counter_set_max(struct page_counter *counter, unsigned long nr_pages)
172 {
173 for (;;) {
174 unsigned long old;
175 long usage;
176
177 /*
178 * Update the limit while making sure that it's not
179 * below the concurrently-changing counter value.
180 *
181 * The xchg implies two full memory barriers before
182 * and after, so the read-swap-read is ordered and
183 * ensures coherency with page_counter_try_charge():
184 * that function modifies the count before checking
185 * the limit, so if it sees the old limit, we see the
186 * modified counter and retry.
187 */
188 usage = page_counter_read(counter);
189
190 if (usage > nr_pages)
191 return -EBUSY;
192
193 old = xchg(&counter->max, nr_pages);
194
195 if (page_counter_read(counter) <= usage || nr_pages >= old)
196 return 0;
197
198 counter->max = old;
199 cond_resched();
200 }
201 }
202
203 /**
204 * page_counter_set_min - set the amount of protected memory
205 * @counter: counter
206 * @nr_pages: value to set
207 *
208 * The caller must serialize invocations on the same counter.
209 */
210 void page_counter_set_min(struct page_counter *counter, unsigned long nr_pages)
211 {
212 struct page_counter *c;
213
214 WRITE_ONCE(counter->min, nr_pages);
215
216 for (c = counter; c; c = c->parent)
217 propagate_protected_usage(c, atomic_long_read(&c->usage));
218 }
219
220 /**
221 * page_counter_set_low - set the amount of protected memory
222 * @counter: counter
223 * @nr_pages: value to set
224 *
225 * The caller must serialize invocations on the same counter.
226 */
227 void page_counter_set_low(struct page_counter *counter, unsigned long nr_pages)
228 {
229 struct page_counter *c;
230
231 WRITE_ONCE(counter->low, nr_pages);
232
233 for (c = counter; c; c = c->parent)
234 propagate_protected_usage(c, atomic_long_read(&c->usage));
235 }
236
237 /**
238 * page_counter_memparse - memparse() for page counter limits
239 * @buf: string to parse
240 * @max: string meaning maximum possible value
241 * @nr_pages: returns the result in number of pages
242 *
243 * Returns -EINVAL, or 0 and @nr_pages on success. @nr_pages will be
244 * limited to %PAGE_COUNTER_MAX.
245 */
246 int page_counter_memparse(const char *buf, const char *max,
247 unsigned long *nr_pages)
248 {
249 char *end;
250 u64 bytes;
251
252 if (!strcmp(buf, max)) {
253 *nr_pages = PAGE_COUNTER_MAX;
254 return 0;
255 }
256
257 bytes = memparse(buf, &end);
258 if (*end != '\0')
259 return -EINVAL;
260
261 *nr_pages = min(bytes / PAGE_SIZE, (u64)PAGE_COUNTER_MAX);
262
263 return 0;
264 }
265
266
267 /*
268 * This function calculates an individual page counter's effective
269 * protection which is derived from its own memory.min/low, its
270 * parent's and siblings' settings, as well as the actual memory
271 * distribution in the tree.
272 *
273 * The following rules apply to the effective protection values:
274 *
275 * 1. At the first level of reclaim, effective protection is equal to
276 * the declared protection in memory.min and memory.low.
277 *
278 * 2. To enable safe delegation of the protection configuration, at
279 * subsequent levels the effective protection is capped to the
280 * parent's effective protection.
281 *
282 * 3. To make complex and dynamic subtrees easier to configure, the
283 * user is allowed to overcommit the declared protection at a given
284 * level. If that is the case, the parent's effective protection is
285 * distributed to the children in proportion to how much protection
286 * they have declared and how much of it they are utilizing.
287 *
288 * This makes distribution proportional, but also work-conserving:
289 * if one counter claims much more protection than it uses memory,
290 * the unused remainder is available to its siblings.
291 *
292 * 4. Conversely, when the declared protection is undercommitted at a
293 * given level, the distribution of the larger parental protection
294 * budget is NOT proportional. A counter's protection from a sibling
295 * is capped to its own memory.min/low setting.
296 *
297 * 5. However, to allow protecting recursive subtrees from each other
298 * without having to declare each individual counter's fixed share
299 * of the ancestor's claim to protection, any unutilized -
300 * "floating" - protection from up the tree is distributed in
301 * proportion to each counter's *usage*. This makes the protection
302 * neutral wrt sibling cgroups and lets them compete freely over
303 * the shared parental protection budget, but it protects the
304 * subtree as a whole from neighboring subtrees.
305 *
306 * Note that 4. and 5. are not in conflict: 4. is about protecting
307 * against immediate siblings whereas 5. is about protecting against
308 * neighboring subtrees.
309 */
310 static unsigned long effective_protection(unsigned long usage,
311 unsigned long parent_usage,
312 unsigned long setting,
313 unsigned long parent_effective,
314 unsigned long siblings_protected,
315 bool recursive_protection)
316 {
317 unsigned long protected;
318 unsigned long ep;
319
320 protected = min(usage, setting);
321 /*
322 * If all cgroups at this level combined claim and use more
323 * protection than what the parent affords them, distribute
324 * shares in proportion to utilization.
325 *
326 * We are using actual utilization rather than the statically
327 * claimed protection in order to be work-conserving: claimed
328 * but unused protection is available to siblings that would
329 * otherwise get a smaller chunk than what they claimed.
330 */
331 if (siblings_protected > parent_effective)
332 return protected * parent_effective / siblings_protected;
333
334 /*
335 * Ok, utilized protection of all children is within what the
336 * parent affords them, so we know whatever this child claims
337 * and utilizes is effectively protected.
338 *
339 * If there is unprotected usage beyond this value, reclaim
340 * will apply pressure in proportion to that amount.
341 *
342 * If there is unutilized protection, the cgroup will be fully
343 * shielded from reclaim, but we do return a smaller value for
344 * protection than what the group could enjoy in theory. This
345 * is okay. With the overcommit distribution above, effective
346 * protection is always dependent on how memory is actually
347 * consumed among the siblings anyway.
348 */
349 ep = protected;
350
351 /*
352 * If the children aren't claiming (all of) the protection
353 * afforded to them by the parent, distribute the remainder in
354 * proportion to the (unprotected) memory of each cgroup. That
355 * way, cgroups that aren't explicitly prioritized wrt each
356 * other compete freely over the allowance, but they are
357 * collectively protected from neighboring trees.
358 *
359 * We're using unprotected memory for the weight so that if
360 * some cgroups DO claim explicit protection, we don't protect
361 * the same bytes twice.
362 *
363 * Check both usage and parent_usage against the respective
364 * protected values. One should imply the other, but they
365 * aren't read atomically - make sure the division is sane.
366 */
367 if (!recursive_protection)
368 return ep;
369
370 if (parent_effective > siblings_protected &&
371 parent_usage > siblings_protected &&
372 usage > protected) {
373 unsigned long unclaimed;
374
375 unclaimed = parent_effective - siblings_protected;
376 unclaimed *= usage - protected;
377 unclaimed /= parent_usage - siblings_protected;
378
379 ep += unclaimed;
380 }
381
382 return ep;
383 }
384
385
386 /**
387 * page_counter_calculate_protection - check if memory consumption is in the normal range
388 * @root: the top ancestor of the sub-tree being checked
389 * @counter: the page_counter the counter to update
390 * @recursive_protection: Whether to use memory_recursiveprot behavior.
391 *
392 * Calculates elow/emin thresholds for given page_counter.
393 *
394 * WARNING: This function is not stateless! It can only be used as part
395 * of a top-down tree iteration, not for isolated queries.
396 */
397 void page_counter_calculate_protection(struct page_counter *root,
398 struct page_counter *counter,
399 bool recursive_protection)
400 {
401 unsigned long usage, parent_usage;
402 struct page_counter *parent = counter->parent;
403
404 /*
405 * Effective values of the reclaim targets are ignored so they
406 * can be stale. Have a look at mem_cgroup_protection for more
407 * details.
408 * TODO: calculation should be more robust so that we do not need
409 * that special casing.
410 */
411 if (root == counter)
412 return;
413
414 usage = page_counter_read(counter);
415 if (!usage)
416 return;
417
418 if (parent == root) {
419 counter->emin = READ_ONCE(counter->min);
420 counter->elow = READ_ONCE(counter->low);
421 return;
422 }
423
424 parent_usage = page_counter_read(parent);
425
426 WRITE_ONCE(counter->emin, effective_protection(usage, parent_usage,
427 READ_ONCE(counter->min),
428 READ_ONCE(parent->emin),
429 atomic_long_read(&parent->children_min_usage),
430 recursive_protection));
431
432 WRITE_ONCE(counter->elow, effective_protection(usage, parent_usage,
433 READ_ONCE(counter->low),
434 READ_ONCE(parent->elow),
435 atomic_long_read(&parent->children_low_usage),
436 recursive_protection));
437 }
438
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.