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