blob: eec1150125b9857f3f84185d1d5beeff64ec034f
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 | |
22 | struct 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 | |
34 | static struct hugetlb_cgroup *root_h_cgroup __read_mostly; |
35 | |
36 | static inline |
37 | struct 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 | |
42 | static inline |
43 | struct 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 | |
48 | static inline bool hugetlb_cgroup_is_root(struct hugetlb_cgroup *h_cg) |
49 | { |
50 | return (h_cg == root_h_cgroup); |
51 | } |
52 | |
53 | static inline struct hugetlb_cgroup * |
54 | parent_hugetlb_cgroup(struct hugetlb_cgroup *h_cg) |
55 | { |
56 | return hugetlb_cgroup_from_css(h_cg->css.parent); |
57 | } |
58 | |
59 | static 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 | |
70 | static 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_limit(counter, limit); |
88 | VM_BUG_ON(ret); |
89 | } |
90 | } |
91 | |
92 | static struct cgroup_subsys_state * |
93 | hugetlb_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 | |
109 | static 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 | */ |
125 | static 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 = 1 << compound_order(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); |
153 | out: |
154 | return; |
155 | } |
156 | |
157 | /* |
158 | * Force the hugetlb cgroup to empty the hugetlb resources by moving them to |
159 | * the parent cgroup. |
160 | */ |
161 | static 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 | |
181 | int 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; |
196 | again: |
197 | rcu_read_lock(); |
198 | h_cg = hugetlb_cgroup_from_task(current); |
199 | if (!css_tryget_online(&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); |
208 | done: |
209 | *ptr = h_cg; |
210 | return ret; |
211 | } |
212 | |
213 | /* Should be called with hugetlb_lock held */ |
214 | void 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 | */ |
228 | void 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 | |
244 | void 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 | |
257 | enum { |
258 | RES_USAGE, |
259 | RES_LIMIT, |
260 | RES_MAX_USAGE, |
261 | RES_FAILCNT, |
262 | }; |
263 | |
264 | static 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->limit * 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 | |
286 | static DEFINE_MUTEX(hugetlb_limit_mutex); |
287 | |
288 | static 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_limit(&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 | |
319 | static 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 | |
342 | static 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 | |
353 | static 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 | |
397 | void __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 | */ |
416 | void 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 | |
436 | struct 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 | }; |
441 |