]>
Commit | Line | Data |
---|---|---|
2bc64a20 AK |
1 | /* |
2 | * | |
3 | * Copyright IBM Corporation, 2012 | |
4 | * Author Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com> | |
5 | * | |
faced7e0 GS |
6 | * Cgroup v2 |
7 | * Copyright (C) 2019 Red Hat, Inc. | |
8 | * Author: Giuseppe Scrivano <gscrivan@redhat.com> | |
9 | * | |
2bc64a20 AK |
10 | * This program is free software; you can redistribute it and/or modify it |
11 | * under the terms of version 2.1 of the GNU Lesser General Public License | |
12 | * as published by the Free Software Foundation. | |
13 | * | |
14 | * This program is distributed in the hope that it would be useful, but | |
15 | * WITHOUT ANY WARRANTY; without even the implied warranty of | |
16 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. | |
17 | * | |
18 | */ | |
19 | ||
20 | #include <linux/cgroup.h> | |
71f87bee | 21 | #include <linux/page_counter.h> |
2bc64a20 AK |
22 | #include <linux/slab.h> |
23 | #include <linux/hugetlb.h> | |
24 | #include <linux/hugetlb_cgroup.h> | |
25 | ||
faced7e0 GS |
26 | enum hugetlb_memory_event { |
27 | HUGETLB_MAX, | |
28 | HUGETLB_NR_MEMORY_EVENTS, | |
29 | }; | |
30 | ||
2bc64a20 AK |
31 | struct hugetlb_cgroup { |
32 | struct cgroup_subsys_state css; | |
faced7e0 | 33 | |
2bc64a20 AK |
34 | /* |
35 | * the counter to account for hugepages from hugetlb. | |
36 | */ | |
71f87bee | 37 | struct page_counter hugepage[HUGE_MAX_HSTATE]; |
faced7e0 | 38 | |
cdc2fcfe MA |
39 | /* |
40 | * the counter to account for hugepage reservations from hugetlb. | |
41 | */ | |
42 | struct page_counter rsvd_hugepage[HUGE_MAX_HSTATE]; | |
43 | ||
faced7e0 GS |
44 | atomic_long_t events[HUGE_MAX_HSTATE][HUGETLB_NR_MEMORY_EVENTS]; |
45 | atomic_long_t events_local[HUGE_MAX_HSTATE][HUGETLB_NR_MEMORY_EVENTS]; | |
46 | ||
47 | /* Handle for "hugetlb.events" */ | |
48 | struct cgroup_file events_file[HUGE_MAX_HSTATE]; | |
49 | ||
50 | /* Handle for "hugetlb.events.local" */ | |
51 | struct cgroup_file events_local_file[HUGE_MAX_HSTATE]; | |
2bc64a20 AK |
52 | }; |
53 | ||
abb8206c AK |
54 | #define MEMFILE_PRIVATE(x, val) (((x) << 16) | (val)) |
55 | #define MEMFILE_IDX(val) (((val) >> 16) & 0xffff) | |
56 | #define MEMFILE_ATTR(val) ((val) & 0xffff) | |
57 | ||
faced7e0 GS |
58 | #define hugetlb_cgroup_from_counter(counter, idx) \ |
59 | container_of(counter, struct hugetlb_cgroup, hugepage[idx]) | |
60 | ||
2bc64a20 AK |
61 | static struct hugetlb_cgroup *root_h_cgroup __read_mostly; |
62 | ||
cdc2fcfe | 63 | static inline struct page_counter * |
1adc4d41 MA |
64 | __hugetlb_cgroup_counter_from_cgroup(struct hugetlb_cgroup *h_cg, int idx, |
65 | bool rsvd) | |
cdc2fcfe MA |
66 | { |
67 | if (rsvd) | |
68 | return &h_cg->rsvd_hugepage[idx]; | |
69 | return &h_cg->hugepage[idx]; | |
70 | } | |
71 | ||
1adc4d41 MA |
72 | static inline struct page_counter * |
73 | hugetlb_cgroup_counter_from_cgroup(struct hugetlb_cgroup *h_cg, int idx) | |
74 | { | |
75 | return __hugetlb_cgroup_counter_from_cgroup(h_cg, idx, false); | |
76 | } | |
77 | ||
78 | static inline struct page_counter * | |
79 | hugetlb_cgroup_counter_from_cgroup_rsvd(struct hugetlb_cgroup *h_cg, int idx) | |
80 | { | |
81 | return __hugetlb_cgroup_counter_from_cgroup(h_cg, idx, true); | |
82 | } | |
83 | ||
2bc64a20 AK |
84 | static inline |
85 | struct hugetlb_cgroup *hugetlb_cgroup_from_css(struct cgroup_subsys_state *s) | |
86 | { | |
a7c6d554 | 87 | return s ? container_of(s, struct hugetlb_cgroup, css) : NULL; |
2bc64a20 AK |
88 | } |
89 | ||
2bc64a20 AK |
90 | static inline |
91 | struct hugetlb_cgroup *hugetlb_cgroup_from_task(struct task_struct *task) | |
92 | { | |
073219e9 | 93 | return hugetlb_cgroup_from_css(task_css(task, hugetlb_cgrp_id)); |
2bc64a20 AK |
94 | } |
95 | ||
96 | static inline bool hugetlb_cgroup_is_root(struct hugetlb_cgroup *h_cg) | |
97 | { | |
98 | return (h_cg == root_h_cgroup); | |
99 | } | |
100 | ||
3f798518 TH |
101 | static inline struct hugetlb_cgroup * |
102 | parent_hugetlb_cgroup(struct hugetlb_cgroup *h_cg) | |
2bc64a20 | 103 | { |
5c9d535b | 104 | return hugetlb_cgroup_from_css(h_cg->css.parent); |
2bc64a20 AK |
105 | } |
106 | ||
3f798518 | 107 | static inline bool hugetlb_cgroup_have_usage(struct hugetlb_cgroup *h_cg) |
2bc64a20 AK |
108 | { |
109 | int idx; | |
2bc64a20 AK |
110 | |
111 | for (idx = 0; idx < hugetlb_max_hstate; idx++) { | |
1adc4d41 MA |
112 | if (page_counter_read( |
113 | hugetlb_cgroup_counter_from_cgroup(h_cg, idx)) || | |
114 | page_counter_read(hugetlb_cgroup_counter_from_cgroup_rsvd( | |
115 | h_cg, idx))) { | |
2bc64a20 | 116 | return true; |
1adc4d41 | 117 | } |
2bc64a20 AK |
118 | } |
119 | return false; | |
120 | } | |
121 | ||
297880f4 DR |
122 | static void hugetlb_cgroup_init(struct hugetlb_cgroup *h_cgroup, |
123 | struct hugetlb_cgroup *parent_h_cgroup) | |
124 | { | |
125 | int idx; | |
126 | ||
127 | for (idx = 0; idx < HUGE_MAX_HSTATE; idx++) { | |
1adc4d41 MA |
128 | struct page_counter *fault_parent = NULL; |
129 | struct page_counter *rsvd_parent = NULL; | |
297880f4 DR |
130 | unsigned long limit; |
131 | int ret; | |
132 | ||
1adc4d41 MA |
133 | if (parent_h_cgroup) { |
134 | fault_parent = hugetlb_cgroup_counter_from_cgroup( | |
135 | parent_h_cgroup, idx); | |
136 | rsvd_parent = hugetlb_cgroup_counter_from_cgroup_rsvd( | |
137 | parent_h_cgroup, idx); | |
138 | } | |
139 | page_counter_init(hugetlb_cgroup_counter_from_cgroup(h_cgroup, | |
140 | idx), | |
141 | fault_parent); | |
142 | page_counter_init( | |
143 | hugetlb_cgroup_counter_from_cgroup_rsvd(h_cgroup, idx), | |
144 | rsvd_parent); | |
297880f4 DR |
145 | |
146 | limit = round_down(PAGE_COUNTER_MAX, | |
147 | 1 << huge_page_order(&hstates[idx])); | |
1adc4d41 MA |
148 | |
149 | ret = page_counter_set_max( | |
150 | hugetlb_cgroup_counter_from_cgroup(h_cgroup, idx), | |
151 | limit); | |
152 | VM_BUG_ON(ret); | |
153 | ret = page_counter_set_max( | |
154 | hugetlb_cgroup_counter_from_cgroup_rsvd(h_cgroup, idx), | |
155 | limit); | |
297880f4 DR |
156 | VM_BUG_ON(ret); |
157 | } | |
158 | } | |
159 | ||
eb95419b TH |
160 | static struct cgroup_subsys_state * |
161 | hugetlb_cgroup_css_alloc(struct cgroup_subsys_state *parent_css) | |
2bc64a20 | 162 | { |
eb95419b TH |
163 | struct hugetlb_cgroup *parent_h_cgroup = hugetlb_cgroup_from_css(parent_css); |
164 | struct hugetlb_cgroup *h_cgroup; | |
2bc64a20 AK |
165 | |
166 | h_cgroup = kzalloc(sizeof(*h_cgroup), GFP_KERNEL); | |
167 | if (!h_cgroup) | |
168 | return ERR_PTR(-ENOMEM); | |
169 | ||
297880f4 | 170 | if (!parent_h_cgroup) |
2bc64a20 | 171 | root_h_cgroup = h_cgroup; |
297880f4 DR |
172 | |
173 | hugetlb_cgroup_init(h_cgroup, parent_h_cgroup); | |
2bc64a20 AK |
174 | return &h_cgroup->css; |
175 | } | |
176 | ||
eb95419b | 177 | static void hugetlb_cgroup_css_free(struct cgroup_subsys_state *css) |
2bc64a20 AK |
178 | { |
179 | struct hugetlb_cgroup *h_cgroup; | |
180 | ||
eb95419b | 181 | h_cgroup = hugetlb_cgroup_from_css(css); |
2bc64a20 AK |
182 | kfree(h_cgroup); |
183 | } | |
184 | ||
da1def55 AK |
185 | /* |
186 | * Should be called with hugetlb_lock held. | |
187 | * Since we are holding hugetlb_lock, pages cannot get moved from | |
188 | * active list or uncharged from the cgroup, So no need to get | |
189 | * page reference and test for page active here. This function | |
190 | * cannot fail. | |
191 | */ | |
3f798518 | 192 | static void hugetlb_cgroup_move_parent(int idx, struct hugetlb_cgroup *h_cg, |
da1def55 AK |
193 | struct page *page) |
194 | { | |
71f87bee JW |
195 | unsigned int nr_pages; |
196 | struct page_counter *counter; | |
da1def55 | 197 | struct hugetlb_cgroup *page_hcg; |
3f798518 | 198 | struct hugetlb_cgroup *parent = parent_hugetlb_cgroup(h_cg); |
da1def55 AK |
199 | |
200 | page_hcg = hugetlb_cgroup_from_page(page); | |
201 | /* | |
202 | * We can have pages in active list without any cgroup | |
203 | * ie, hugepage with less than 3 pages. We can safely | |
204 | * ignore those pages. | |
205 | */ | |
206 | if (!page_hcg || page_hcg != h_cg) | |
207 | goto out; | |
208 | ||
d8c6546b | 209 | nr_pages = compound_nr(page); |
da1def55 AK |
210 | if (!parent) { |
211 | parent = root_h_cgroup; | |
212 | /* root has no limit */ | |
71f87bee | 213 | page_counter_charge(&parent->hugepage[idx], nr_pages); |
da1def55 AK |
214 | } |
215 | counter = &h_cg->hugepage[idx]; | |
71f87bee JW |
216 | /* Take the pages off the local counter */ |
217 | page_counter_cancel(counter, nr_pages); | |
da1def55 AK |
218 | |
219 | set_hugetlb_cgroup(page, parent); | |
220 | out: | |
221 | return; | |
222 | } | |
223 | ||
224 | /* | |
225 | * Force the hugetlb cgroup to empty the hugetlb resources by moving them to | |
226 | * the parent cgroup. | |
227 | */ | |
eb95419b | 228 | static void hugetlb_cgroup_css_offline(struct cgroup_subsys_state *css) |
2bc64a20 | 229 | { |
eb95419b | 230 | struct hugetlb_cgroup *h_cg = hugetlb_cgroup_from_css(css); |
da1def55 AK |
231 | struct hstate *h; |
232 | struct page *page; | |
9d093cb1 | 233 | int idx = 0; |
da1def55 AK |
234 | |
235 | do { | |
da1def55 AK |
236 | for_each_hstate(h) { |
237 | spin_lock(&hugetlb_lock); | |
238 | list_for_each_entry(page, &h->hugepage_activelist, lru) | |
3f798518 | 239 | hugetlb_cgroup_move_parent(idx, h_cg, page); |
da1def55 AK |
240 | |
241 | spin_unlock(&hugetlb_lock); | |
242 | idx++; | |
243 | } | |
244 | cond_resched(); | |
3f798518 | 245 | } while (hugetlb_cgroup_have_usage(h_cg)); |
2bc64a20 AK |
246 | } |
247 | ||
faced7e0 GS |
248 | static inline void hugetlb_event(struct hugetlb_cgroup *hugetlb, int idx, |
249 | enum hugetlb_memory_event event) | |
250 | { | |
251 | atomic_long_inc(&hugetlb->events_local[idx][event]); | |
252 | cgroup_file_notify(&hugetlb->events_local_file[idx]); | |
253 | ||
254 | do { | |
255 | atomic_long_inc(&hugetlb->events[idx][event]); | |
256 | cgroup_file_notify(&hugetlb->events_file[idx]); | |
257 | } while ((hugetlb = parent_hugetlb_cgroup(hugetlb)) && | |
258 | !hugetlb_cgroup_is_root(hugetlb)); | |
259 | } | |
260 | ||
1adc4d41 MA |
261 | static int __hugetlb_cgroup_charge_cgroup(int idx, unsigned long nr_pages, |
262 | struct hugetlb_cgroup **ptr, | |
263 | bool rsvd) | |
6d76dcf4 AK |
264 | { |
265 | int ret = 0; | |
71f87bee | 266 | struct page_counter *counter; |
6d76dcf4 | 267 | struct hugetlb_cgroup *h_cg = NULL; |
6d76dcf4 AK |
268 | |
269 | if (hugetlb_cgroup_disabled()) | |
270 | goto done; | |
271 | /* | |
272 | * We don't charge any cgroup if the compound page have less | |
273 | * than 3 pages. | |
274 | */ | |
275 | if (huge_page_order(&hstates[idx]) < HUGETLB_CGROUP_MIN_ORDER) | |
276 | goto done; | |
277 | again: | |
278 | rcu_read_lock(); | |
279 | h_cg = hugetlb_cgroup_from_task(current); | |
0362f326 | 280 | if (!css_tryget(&h_cg->css)) { |
6d76dcf4 AK |
281 | rcu_read_unlock(); |
282 | goto again; | |
283 | } | |
284 | rcu_read_unlock(); | |
285 | ||
1adc4d41 MA |
286 | if (!page_counter_try_charge( |
287 | __hugetlb_cgroup_counter_from_cgroup(h_cg, idx, rsvd), | |
288 | nr_pages, &counter)) { | |
6071ca52 | 289 | ret = -ENOMEM; |
726b7bbe | 290 | hugetlb_event(h_cg, idx, HUGETLB_MAX); |
1adc4d41 MA |
291 | css_put(&h_cg->css); |
292 | goto done; | |
faced7e0 | 293 | } |
1adc4d41 MA |
294 | /* Reservations take a reference to the css because they do not get |
295 | * reparented. | |
296 | */ | |
297 | if (!rsvd) | |
298 | css_put(&h_cg->css); | |
6d76dcf4 AK |
299 | done: |
300 | *ptr = h_cg; | |
301 | return ret; | |
302 | } | |
303 | ||
1adc4d41 MA |
304 | int hugetlb_cgroup_charge_cgroup(int idx, unsigned long nr_pages, |
305 | struct hugetlb_cgroup **ptr) | |
306 | { | |
307 | return __hugetlb_cgroup_charge_cgroup(idx, nr_pages, ptr, false); | |
308 | } | |
309 | ||
310 | int hugetlb_cgroup_charge_cgroup_rsvd(int idx, unsigned long nr_pages, | |
311 | struct hugetlb_cgroup **ptr) | |
312 | { | |
313 | return __hugetlb_cgroup_charge_cgroup(idx, nr_pages, ptr, true); | |
314 | } | |
315 | ||
94ae8ba7 | 316 | /* Should be called with hugetlb_lock held */ |
1adc4d41 MA |
317 | static void __hugetlb_cgroup_commit_charge(int idx, unsigned long nr_pages, |
318 | struct hugetlb_cgroup *h_cg, | |
319 | struct page *page, bool rsvd) | |
6d76dcf4 AK |
320 | { |
321 | if (hugetlb_cgroup_disabled() || !h_cg) | |
322 | return; | |
323 | ||
1adc4d41 | 324 | __set_hugetlb_cgroup(page, h_cg, rsvd); |
6d76dcf4 AK |
325 | return; |
326 | } | |
327 | ||
1adc4d41 MA |
328 | void hugetlb_cgroup_commit_charge(int idx, unsigned long nr_pages, |
329 | struct hugetlb_cgroup *h_cg, | |
330 | struct page *page) | |
331 | { | |
332 | __hugetlb_cgroup_commit_charge(idx, nr_pages, h_cg, page, false); | |
333 | } | |
334 | ||
335 | void hugetlb_cgroup_commit_charge_rsvd(int idx, unsigned long nr_pages, | |
336 | struct hugetlb_cgroup *h_cg, | |
337 | struct page *page) | |
338 | { | |
339 | __hugetlb_cgroup_commit_charge(idx, nr_pages, h_cg, page, true); | |
340 | } | |
341 | ||
6d76dcf4 AK |
342 | /* |
343 | * Should be called with hugetlb_lock held | |
344 | */ | |
1adc4d41 MA |
345 | static void __hugetlb_cgroup_uncharge_page(int idx, unsigned long nr_pages, |
346 | struct page *page, bool rsvd) | |
6d76dcf4 AK |
347 | { |
348 | struct hugetlb_cgroup *h_cg; | |
6d76dcf4 AK |
349 | |
350 | if (hugetlb_cgroup_disabled()) | |
351 | return; | |
7ea8574e | 352 | lockdep_assert_held(&hugetlb_lock); |
1adc4d41 | 353 | h_cg = __hugetlb_cgroup_from_page(page, rsvd); |
6d76dcf4 AK |
354 | if (unlikely(!h_cg)) |
355 | return; | |
1adc4d41 MA |
356 | __set_hugetlb_cgroup(page, NULL, rsvd); |
357 | ||
358 | page_counter_uncharge(__hugetlb_cgroup_counter_from_cgroup(h_cg, idx, | |
359 | rsvd), | |
360 | nr_pages); | |
361 | ||
362 | if (rsvd) | |
363 | css_put(&h_cg->css); | |
364 | ||
6d76dcf4 AK |
365 | return; |
366 | } | |
367 | ||
1adc4d41 MA |
368 | void hugetlb_cgroup_uncharge_page(int idx, unsigned long nr_pages, |
369 | struct page *page) | |
370 | { | |
371 | __hugetlb_cgroup_uncharge_page(idx, nr_pages, page, false); | |
372 | } | |
373 | ||
374 | void hugetlb_cgroup_uncharge_page_rsvd(int idx, unsigned long nr_pages, | |
375 | struct page *page) | |
376 | { | |
377 | __hugetlb_cgroup_uncharge_page(idx, nr_pages, page, true); | |
378 | } | |
379 | ||
380 | static void __hugetlb_cgroup_uncharge_cgroup(int idx, unsigned long nr_pages, | |
381 | struct hugetlb_cgroup *h_cg, | |
382 | bool rsvd) | |
6d76dcf4 | 383 | { |
6d76dcf4 AK |
384 | if (hugetlb_cgroup_disabled() || !h_cg) |
385 | return; | |
386 | ||
387 | if (huge_page_order(&hstates[idx]) < HUGETLB_CGROUP_MIN_ORDER) | |
388 | return; | |
389 | ||
1adc4d41 MA |
390 | page_counter_uncharge(__hugetlb_cgroup_counter_from_cgroup(h_cg, idx, |
391 | rsvd), | |
392 | nr_pages); | |
393 | ||
394 | if (rsvd) | |
395 | css_put(&h_cg->css); | |
396 | } | |
397 | ||
398 | void hugetlb_cgroup_uncharge_cgroup(int idx, unsigned long nr_pages, | |
399 | struct hugetlb_cgroup *h_cg) | |
400 | { | |
401 | __hugetlb_cgroup_uncharge_cgroup(idx, nr_pages, h_cg, false); | |
402 | } | |
403 | ||
404 | void hugetlb_cgroup_uncharge_cgroup_rsvd(int idx, unsigned long nr_pages, | |
405 | struct hugetlb_cgroup *h_cg) | |
406 | { | |
407 | __hugetlb_cgroup_uncharge_cgroup(idx, nr_pages, h_cg, true); | |
408 | } | |
409 | ||
410 | void hugetlb_cgroup_uncharge_counter(struct page_counter *p, | |
411 | unsigned long nr_pages, | |
412 | struct cgroup_subsys_state *css) | |
413 | { | |
414 | if (hugetlb_cgroup_disabled() || !p || !css) | |
415 | return; | |
416 | ||
417 | page_counter_uncharge(p, nr_pages); | |
418 | css_put(css); | |
6d76dcf4 AK |
419 | } |
420 | ||
71f87bee JW |
421 | enum { |
422 | RES_USAGE, | |
cdc2fcfe | 423 | RES_RSVD_USAGE, |
71f87bee | 424 | RES_LIMIT, |
cdc2fcfe | 425 | RES_RSVD_LIMIT, |
71f87bee | 426 | RES_MAX_USAGE, |
cdc2fcfe | 427 | RES_RSVD_MAX_USAGE, |
71f87bee | 428 | RES_FAILCNT, |
cdc2fcfe | 429 | RES_RSVD_FAILCNT, |
71f87bee JW |
430 | }; |
431 | ||
716f479d TH |
432 | static u64 hugetlb_cgroup_read_u64(struct cgroup_subsys_state *css, |
433 | struct cftype *cft) | |
abb8206c | 434 | { |
71f87bee | 435 | struct page_counter *counter; |
cdc2fcfe | 436 | struct page_counter *rsvd_counter; |
182446d0 | 437 | struct hugetlb_cgroup *h_cg = hugetlb_cgroup_from_css(css); |
abb8206c | 438 | |
71f87bee | 439 | counter = &h_cg->hugepage[MEMFILE_IDX(cft->private)]; |
cdc2fcfe | 440 | rsvd_counter = &h_cg->rsvd_hugepage[MEMFILE_IDX(cft->private)]; |
abb8206c | 441 | |
71f87bee JW |
442 | switch (MEMFILE_ATTR(cft->private)) { |
443 | case RES_USAGE: | |
444 | return (u64)page_counter_read(counter) * PAGE_SIZE; | |
cdc2fcfe MA |
445 | case RES_RSVD_USAGE: |
446 | return (u64)page_counter_read(rsvd_counter) * PAGE_SIZE; | |
71f87bee | 447 | case RES_LIMIT: |
bbec2e15 | 448 | return (u64)counter->max * PAGE_SIZE; |
cdc2fcfe MA |
449 | case RES_RSVD_LIMIT: |
450 | return (u64)rsvd_counter->max * PAGE_SIZE; | |
71f87bee JW |
451 | case RES_MAX_USAGE: |
452 | return (u64)counter->watermark * PAGE_SIZE; | |
cdc2fcfe MA |
453 | case RES_RSVD_MAX_USAGE: |
454 | return (u64)rsvd_counter->watermark * PAGE_SIZE; | |
71f87bee JW |
455 | case RES_FAILCNT: |
456 | return counter->failcnt; | |
cdc2fcfe MA |
457 | case RES_RSVD_FAILCNT: |
458 | return rsvd_counter->failcnt; | |
71f87bee JW |
459 | default: |
460 | BUG(); | |
461 | } | |
abb8206c AK |
462 | } |
463 | ||
faced7e0 GS |
464 | static int hugetlb_cgroup_read_u64_max(struct seq_file *seq, void *v) |
465 | { | |
466 | int idx; | |
467 | u64 val; | |
468 | struct cftype *cft = seq_cft(seq); | |
469 | unsigned long limit; | |
470 | struct page_counter *counter; | |
471 | struct hugetlb_cgroup *h_cg = hugetlb_cgroup_from_css(seq_css(seq)); | |
472 | ||
473 | idx = MEMFILE_IDX(cft->private); | |
474 | counter = &h_cg->hugepage[idx]; | |
475 | ||
476 | limit = round_down(PAGE_COUNTER_MAX, | |
477 | 1 << huge_page_order(&hstates[idx])); | |
478 | ||
479 | switch (MEMFILE_ATTR(cft->private)) { | |
cdc2fcfe MA |
480 | case RES_RSVD_USAGE: |
481 | counter = &h_cg->rsvd_hugepage[idx]; | |
482 | /* Fall through. */ | |
faced7e0 GS |
483 | case RES_USAGE: |
484 | val = (u64)page_counter_read(counter); | |
485 | seq_printf(seq, "%llu\n", val * PAGE_SIZE); | |
486 | break; | |
cdc2fcfe MA |
487 | case RES_RSVD_LIMIT: |
488 | counter = &h_cg->rsvd_hugepage[idx]; | |
489 | /* Fall through. */ | |
faced7e0 GS |
490 | case RES_LIMIT: |
491 | val = (u64)counter->max; | |
492 | if (val == limit) | |
493 | seq_puts(seq, "max\n"); | |
494 | else | |
495 | seq_printf(seq, "%llu\n", val * PAGE_SIZE); | |
496 | break; | |
497 | default: | |
498 | BUG(); | |
499 | } | |
500 | ||
501 | return 0; | |
502 | } | |
503 | ||
71f87bee JW |
504 | static DEFINE_MUTEX(hugetlb_limit_mutex); |
505 | ||
451af504 | 506 | static ssize_t hugetlb_cgroup_write(struct kernfs_open_file *of, |
faced7e0 GS |
507 | char *buf, size_t nbytes, loff_t off, |
508 | const char *max) | |
abb8206c | 509 | { |
71f87bee JW |
510 | int ret, idx; |
511 | unsigned long nr_pages; | |
451af504 | 512 | struct hugetlb_cgroup *h_cg = hugetlb_cgroup_from_css(of_css(of)); |
cdc2fcfe | 513 | bool rsvd = false; |
abb8206c | 514 | |
71f87bee JW |
515 | if (hugetlb_cgroup_is_root(h_cg)) /* Can't set limit on root */ |
516 | return -EINVAL; | |
517 | ||
451af504 | 518 | buf = strstrip(buf); |
faced7e0 | 519 | ret = page_counter_memparse(buf, max, &nr_pages); |
71f87bee JW |
520 | if (ret) |
521 | return ret; | |
522 | ||
451af504 | 523 | idx = MEMFILE_IDX(of_cft(of)->private); |
297880f4 | 524 | nr_pages = round_down(nr_pages, 1 << huge_page_order(&hstates[idx])); |
abb8206c | 525 | |
71f87bee | 526 | switch (MEMFILE_ATTR(of_cft(of)->private)) { |
cdc2fcfe MA |
527 | case RES_RSVD_LIMIT: |
528 | rsvd = true; | |
529 | /* Fall through. */ | |
abb8206c | 530 | case RES_LIMIT: |
71f87bee | 531 | mutex_lock(&hugetlb_limit_mutex); |
cdc2fcfe | 532 | ret = page_counter_set_max( |
1adc4d41 | 533 | __hugetlb_cgroup_counter_from_cgroup(h_cg, idx, rsvd), |
cdc2fcfe | 534 | nr_pages); |
71f87bee | 535 | mutex_unlock(&hugetlb_limit_mutex); |
abb8206c AK |
536 | break; |
537 | default: | |
538 | ret = -EINVAL; | |
539 | break; | |
540 | } | |
451af504 | 541 | return ret ?: nbytes; |
abb8206c AK |
542 | } |
543 | ||
faced7e0 GS |
544 | static ssize_t hugetlb_cgroup_write_legacy(struct kernfs_open_file *of, |
545 | char *buf, size_t nbytes, loff_t off) | |
546 | { | |
547 | return hugetlb_cgroup_write(of, buf, nbytes, off, "-1"); | |
548 | } | |
549 | ||
550 | static ssize_t hugetlb_cgroup_write_dfl(struct kernfs_open_file *of, | |
551 | char *buf, size_t nbytes, loff_t off) | |
552 | { | |
553 | return hugetlb_cgroup_write(of, buf, nbytes, off, "max"); | |
554 | } | |
555 | ||
6770c64e TH |
556 | static ssize_t hugetlb_cgroup_reset(struct kernfs_open_file *of, |
557 | char *buf, size_t nbytes, loff_t off) | |
abb8206c | 558 | { |
71f87bee | 559 | int ret = 0; |
cdc2fcfe | 560 | struct page_counter *counter, *rsvd_counter; |
6770c64e | 561 | struct hugetlb_cgroup *h_cg = hugetlb_cgroup_from_css(of_css(of)); |
abb8206c | 562 | |
71f87bee | 563 | counter = &h_cg->hugepage[MEMFILE_IDX(of_cft(of)->private)]; |
cdc2fcfe | 564 | rsvd_counter = &h_cg->rsvd_hugepage[MEMFILE_IDX(of_cft(of)->private)]; |
abb8206c | 565 | |
71f87bee | 566 | switch (MEMFILE_ATTR(of_cft(of)->private)) { |
abb8206c | 567 | case RES_MAX_USAGE: |
71f87bee | 568 | page_counter_reset_watermark(counter); |
abb8206c | 569 | break; |
cdc2fcfe MA |
570 | case RES_RSVD_MAX_USAGE: |
571 | page_counter_reset_watermark(rsvd_counter); | |
572 | break; | |
abb8206c | 573 | case RES_FAILCNT: |
71f87bee | 574 | counter->failcnt = 0; |
abb8206c | 575 | break; |
cdc2fcfe MA |
576 | case RES_RSVD_FAILCNT: |
577 | rsvd_counter->failcnt = 0; | |
578 | break; | |
abb8206c AK |
579 | default: |
580 | ret = -EINVAL; | |
581 | break; | |
582 | } | |
6770c64e | 583 | return ret ?: nbytes; |
abb8206c AK |
584 | } |
585 | ||
586 | static char *mem_fmt(char *buf, int size, unsigned long hsize) | |
587 | { | |
588 | if (hsize >= (1UL << 30)) | |
589 | snprintf(buf, size, "%luGB", hsize >> 30); | |
590 | else if (hsize >= (1UL << 20)) | |
591 | snprintf(buf, size, "%luMB", hsize >> 20); | |
592 | else | |
593 | snprintf(buf, size, "%luKB", hsize >> 10); | |
594 | return buf; | |
595 | } | |
596 | ||
faced7e0 GS |
597 | static int __hugetlb_events_show(struct seq_file *seq, bool local) |
598 | { | |
599 | int idx; | |
600 | long max; | |
601 | struct cftype *cft = seq_cft(seq); | |
602 | struct hugetlb_cgroup *h_cg = hugetlb_cgroup_from_css(seq_css(seq)); | |
603 | ||
604 | idx = MEMFILE_IDX(cft->private); | |
605 | ||
606 | if (local) | |
607 | max = atomic_long_read(&h_cg->events_local[idx][HUGETLB_MAX]); | |
608 | else | |
609 | max = atomic_long_read(&h_cg->events[idx][HUGETLB_MAX]); | |
610 | ||
611 | seq_printf(seq, "max %lu\n", max); | |
612 | ||
613 | return 0; | |
614 | } | |
615 | ||
616 | static int hugetlb_events_show(struct seq_file *seq, void *v) | |
617 | { | |
618 | return __hugetlb_events_show(seq, false); | |
619 | } | |
620 | ||
621 | static int hugetlb_events_local_show(struct seq_file *seq, void *v) | |
622 | { | |
623 | return __hugetlb_events_show(seq, true); | |
624 | } | |
625 | ||
626 | static void __init __hugetlb_cgroup_file_dfl_init(int idx) | |
abb8206c AK |
627 | { |
628 | char buf[32]; | |
629 | struct cftype *cft; | |
630 | struct hstate *h = &hstates[idx]; | |
631 | ||
632 | /* format the size */ | |
cdc2fcfe | 633 | mem_fmt(buf, sizeof(buf), huge_page_size(h)); |
abb8206c AK |
634 | |
635 | /* Add the limit file */ | |
faced7e0 GS |
636 | cft = &h->cgroup_files_dfl[0]; |
637 | snprintf(cft->name, MAX_CFTYPE_NAME, "%s.max", buf); | |
638 | cft->private = MEMFILE_PRIVATE(idx, RES_LIMIT); | |
639 | cft->seq_show = hugetlb_cgroup_read_u64_max; | |
640 | cft->write = hugetlb_cgroup_write_dfl; | |
641 | cft->flags = CFTYPE_NOT_ON_ROOT; | |
642 | ||
cdc2fcfe | 643 | /* Add the reservation limit file */ |
faced7e0 | 644 | cft = &h->cgroup_files_dfl[1]; |
cdc2fcfe MA |
645 | snprintf(cft->name, MAX_CFTYPE_NAME, "%s.rsvd.max", buf); |
646 | cft->private = MEMFILE_PRIVATE(idx, RES_RSVD_LIMIT); | |
647 | cft->seq_show = hugetlb_cgroup_read_u64_max; | |
648 | cft->write = hugetlb_cgroup_write_dfl; | |
649 | cft->flags = CFTYPE_NOT_ON_ROOT; | |
650 | ||
651 | /* Add the current usage file */ | |
652 | cft = &h->cgroup_files_dfl[2]; | |
faced7e0 GS |
653 | snprintf(cft->name, MAX_CFTYPE_NAME, "%s.current", buf); |
654 | cft->private = MEMFILE_PRIVATE(idx, RES_USAGE); | |
655 | cft->seq_show = hugetlb_cgroup_read_u64_max; | |
656 | cft->flags = CFTYPE_NOT_ON_ROOT; | |
657 | ||
cdc2fcfe MA |
658 | /* Add the current reservation usage file */ |
659 | cft = &h->cgroup_files_dfl[3]; | |
660 | snprintf(cft->name, MAX_CFTYPE_NAME, "%s.rsvd.current", buf); | |
661 | cft->private = MEMFILE_PRIVATE(idx, RES_RSVD_USAGE); | |
662 | cft->seq_show = hugetlb_cgroup_read_u64_max; | |
663 | cft->flags = CFTYPE_NOT_ON_ROOT; | |
664 | ||
faced7e0 | 665 | /* Add the events file */ |
cdc2fcfe | 666 | cft = &h->cgroup_files_dfl[4]; |
faced7e0 GS |
667 | snprintf(cft->name, MAX_CFTYPE_NAME, "%s.events", buf); |
668 | cft->private = MEMFILE_PRIVATE(idx, 0); | |
669 | cft->seq_show = hugetlb_events_show; | |
670 | cft->file_offset = offsetof(struct hugetlb_cgroup, events_file[idx]), | |
671 | cft->flags = CFTYPE_NOT_ON_ROOT; | |
672 | ||
673 | /* Add the events.local file */ | |
cdc2fcfe | 674 | cft = &h->cgroup_files_dfl[5]; |
faced7e0 GS |
675 | snprintf(cft->name, MAX_CFTYPE_NAME, "%s.events.local", buf); |
676 | cft->private = MEMFILE_PRIVATE(idx, 0); | |
677 | cft->seq_show = hugetlb_events_local_show; | |
678 | cft->file_offset = offsetof(struct hugetlb_cgroup, | |
679 | events_local_file[idx]), | |
680 | cft->flags = CFTYPE_NOT_ON_ROOT; | |
681 | ||
682 | /* NULL terminate the last cft */ | |
cdc2fcfe | 683 | cft = &h->cgroup_files_dfl[6]; |
faced7e0 GS |
684 | memset(cft, 0, sizeof(*cft)); |
685 | ||
686 | WARN_ON(cgroup_add_dfl_cftypes(&hugetlb_cgrp_subsys, | |
687 | h->cgroup_files_dfl)); | |
688 | } | |
689 | ||
690 | static void __init __hugetlb_cgroup_file_legacy_init(int idx) | |
691 | { | |
692 | char buf[32]; | |
693 | struct cftype *cft; | |
694 | struct hstate *h = &hstates[idx]; | |
695 | ||
696 | /* format the size */ | |
cdc2fcfe | 697 | mem_fmt(buf, sizeof(buf), huge_page_size(h)); |
faced7e0 GS |
698 | |
699 | /* Add the limit file */ | |
700 | cft = &h->cgroup_files_legacy[0]; | |
abb8206c AK |
701 | snprintf(cft->name, MAX_CFTYPE_NAME, "%s.limit_in_bytes", buf); |
702 | cft->private = MEMFILE_PRIVATE(idx, RES_LIMIT); | |
716f479d | 703 | cft->read_u64 = hugetlb_cgroup_read_u64; |
faced7e0 | 704 | cft->write = hugetlb_cgroup_write_legacy; |
abb8206c | 705 | |
cdc2fcfe | 706 | /* Add the reservation limit file */ |
faced7e0 | 707 | cft = &h->cgroup_files_legacy[1]; |
cdc2fcfe MA |
708 | snprintf(cft->name, MAX_CFTYPE_NAME, "%s.rsvd.limit_in_bytes", buf); |
709 | cft->private = MEMFILE_PRIVATE(idx, RES_RSVD_LIMIT); | |
710 | cft->read_u64 = hugetlb_cgroup_read_u64; | |
711 | cft->write = hugetlb_cgroup_write_legacy; | |
712 | ||
713 | /* Add the usage file */ | |
714 | cft = &h->cgroup_files_legacy[2]; | |
abb8206c AK |
715 | snprintf(cft->name, MAX_CFTYPE_NAME, "%s.usage_in_bytes", buf); |
716 | cft->private = MEMFILE_PRIVATE(idx, RES_USAGE); | |
716f479d | 717 | cft->read_u64 = hugetlb_cgroup_read_u64; |
abb8206c | 718 | |
cdc2fcfe MA |
719 | /* Add the reservation usage file */ |
720 | cft = &h->cgroup_files_legacy[3]; | |
721 | snprintf(cft->name, MAX_CFTYPE_NAME, "%s.rsvd.usage_in_bytes", buf); | |
722 | cft->private = MEMFILE_PRIVATE(idx, RES_RSVD_USAGE); | |
723 | cft->read_u64 = hugetlb_cgroup_read_u64; | |
724 | ||
abb8206c | 725 | /* Add the MAX usage file */ |
cdc2fcfe | 726 | cft = &h->cgroup_files_legacy[4]; |
abb8206c AK |
727 | snprintf(cft->name, MAX_CFTYPE_NAME, "%s.max_usage_in_bytes", buf); |
728 | cft->private = MEMFILE_PRIVATE(idx, RES_MAX_USAGE); | |
6770c64e | 729 | cft->write = hugetlb_cgroup_reset; |
716f479d | 730 | cft->read_u64 = hugetlb_cgroup_read_u64; |
abb8206c | 731 | |
cdc2fcfe MA |
732 | /* Add the MAX reservation usage file */ |
733 | cft = &h->cgroup_files_legacy[5]; | |
734 | snprintf(cft->name, MAX_CFTYPE_NAME, "%s.rsvd.max_usage_in_bytes", buf); | |
735 | cft->private = MEMFILE_PRIVATE(idx, RES_RSVD_MAX_USAGE); | |
736 | cft->write = hugetlb_cgroup_reset; | |
737 | cft->read_u64 = hugetlb_cgroup_read_u64; | |
738 | ||
abb8206c | 739 | /* Add the failcntfile */ |
cdc2fcfe | 740 | cft = &h->cgroup_files_legacy[6]; |
abb8206c | 741 | snprintf(cft->name, MAX_CFTYPE_NAME, "%s.failcnt", buf); |
cdc2fcfe MA |
742 | cft->private = MEMFILE_PRIVATE(idx, RES_FAILCNT); |
743 | cft->write = hugetlb_cgroup_reset; | |
744 | cft->read_u64 = hugetlb_cgroup_read_u64; | |
745 | ||
746 | /* Add the reservation failcntfile */ | |
747 | cft = &h->cgroup_files_legacy[7]; | |
748 | snprintf(cft->name, MAX_CFTYPE_NAME, "%s.rsvd.failcnt", buf); | |
749 | cft->private = MEMFILE_PRIVATE(idx, RES_RSVD_FAILCNT); | |
6770c64e | 750 | cft->write = hugetlb_cgroup_reset; |
716f479d | 751 | cft->read_u64 = hugetlb_cgroup_read_u64; |
abb8206c AK |
752 | |
753 | /* NULL terminate the last cft */ | |
cdc2fcfe | 754 | cft = &h->cgroup_files_legacy[8]; |
abb8206c AK |
755 | memset(cft, 0, sizeof(*cft)); |
756 | ||
2cf669a5 | 757 | WARN_ON(cgroup_add_legacy_cftypes(&hugetlb_cgrp_subsys, |
faced7e0 GS |
758 | h->cgroup_files_legacy)); |
759 | } | |
760 | ||
761 | static void __init __hugetlb_cgroup_file_init(int idx) | |
762 | { | |
763 | __hugetlb_cgroup_file_dfl_init(idx); | |
764 | __hugetlb_cgroup_file_legacy_init(idx); | |
7179e7bf JW |
765 | } |
766 | ||
767 | void __init hugetlb_cgroup_file_init(void) | |
768 | { | |
769 | struct hstate *h; | |
770 | ||
771 | for_each_hstate(h) { | |
772 | /* | |
773 | * Add cgroup control files only if the huge page consists | |
774 | * of more than two normal pages. This is because we use | |
1d798ca3 | 775 | * page[2].private for storing cgroup details. |
7179e7bf JW |
776 | */ |
777 | if (huge_page_order(h) >= HUGETLB_CGROUP_MIN_ORDER) | |
778 | __hugetlb_cgroup_file_init(hstate_index(h)); | |
779 | } | |
abb8206c AK |
780 | } |
781 | ||
75754681 AK |
782 | /* |
783 | * hugetlb_lock will make sure a parallel cgroup rmdir won't happen | |
784 | * when we migrate hugepages | |
785 | */ | |
8e6ac7fa AK |
786 | void hugetlb_cgroup_migrate(struct page *oldhpage, struct page *newhpage) |
787 | { | |
788 | struct hugetlb_cgroup *h_cg; | |
1adc4d41 | 789 | struct hugetlb_cgroup *h_cg_rsvd; |
94ae8ba7 | 790 | struct hstate *h = page_hstate(oldhpage); |
8e6ac7fa AK |
791 | |
792 | if (hugetlb_cgroup_disabled()) | |
793 | return; | |
794 | ||
309381fe | 795 | VM_BUG_ON_PAGE(!PageHuge(oldhpage), oldhpage); |
8e6ac7fa AK |
796 | spin_lock(&hugetlb_lock); |
797 | h_cg = hugetlb_cgroup_from_page(oldhpage); | |
1adc4d41 | 798 | h_cg_rsvd = hugetlb_cgroup_from_page_rsvd(oldhpage); |
8e6ac7fa | 799 | set_hugetlb_cgroup(oldhpage, NULL); |
9808895e | 800 | set_hugetlb_cgroup_rsvd(oldhpage, NULL); |
8e6ac7fa AK |
801 | |
802 | /* move the h_cg details to new cgroup */ | |
9808895e | 803 | set_hugetlb_cgroup(newhpage, h_cg); |
1adc4d41 | 804 | set_hugetlb_cgroup_rsvd(newhpage, h_cg_rsvd); |
94ae8ba7 | 805 | list_move(&newhpage->lru, &h->hugepage_activelist); |
8e6ac7fa | 806 | spin_unlock(&hugetlb_lock); |
8e6ac7fa AK |
807 | return; |
808 | } | |
809 | ||
faced7e0 GS |
810 | static struct cftype hugetlb_files[] = { |
811 | {} /* terminate */ | |
812 | }; | |
813 | ||
073219e9 | 814 | struct cgroup_subsys hugetlb_cgrp_subsys = { |
92fb9748 TH |
815 | .css_alloc = hugetlb_cgroup_css_alloc, |
816 | .css_offline = hugetlb_cgroup_css_offline, | |
817 | .css_free = hugetlb_cgroup_css_free, | |
faced7e0 GS |
818 | .dfl_cftypes = hugetlb_files, |
819 | .legacy_cftypes = hugetlb_files, | |
2bc64a20 | 820 | }; |