blob: 680e4819e0775ed50bdb7e928b62cf44a9d005aa [file] [log] [blame]
Aneesh Kumar K.V2bc64a22012-07-31 16:42:12 -07001/*
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/slab.h>
18#include <linux/hugetlb.h>
19#include <linux/hugetlb_cgroup.h>
20
21struct hugetlb_cgroup {
22 struct cgroup_subsys_state css;
23 /*
24 * the counter to account for hugepages from hugetlb.
25 */
26 struct res_counter hugepage[HUGE_MAX_HSTATE];
27};
28
Aneesh Kumar K.Vabb82062012-07-31 16:42:24 -070029#define MEMFILE_PRIVATE(x, val) (((x) << 16) | (val))
30#define MEMFILE_IDX(val) (((val) >> 16) & 0xffff)
31#define MEMFILE_ATTR(val) ((val) & 0xffff)
32
Aneesh Kumar K.V2bc64a22012-07-31 16:42:12 -070033struct cgroup_subsys hugetlb_subsys __read_mostly;
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 container_of(s, struct hugetlb_cgroup, css);
40}
41
42static inline
43struct hugetlb_cgroup *hugetlb_cgroup_from_cgroup(struct cgroup *cgroup)
44{
45 return hugetlb_cgroup_from_css(cgroup_subsys_state(cgroup,
46 hugetlb_subsys_id));
47}
48
49static inline
50struct hugetlb_cgroup *hugetlb_cgroup_from_task(struct task_struct *task)
51{
52 return hugetlb_cgroup_from_css(task_subsys_state(task,
53 hugetlb_subsys_id));
54}
55
56static inline bool hugetlb_cgroup_is_root(struct hugetlb_cgroup *h_cg)
57{
58 return (h_cg == root_h_cgroup);
59}
60
61static inline struct hugetlb_cgroup *parent_hugetlb_cgroup(struct cgroup *cg)
62{
63 if (!cg->parent)
64 return NULL;
65 return hugetlb_cgroup_from_cgroup(cg->parent);
66}
67
68static inline bool hugetlb_cgroup_have_usage(struct cgroup *cg)
69{
70 int idx;
71 struct hugetlb_cgroup *h_cg = hugetlb_cgroup_from_cgroup(cg);
72
73 for (idx = 0; idx < hugetlb_max_hstate; idx++) {
74 if ((res_counter_read_u64(&h_cg->hugepage[idx], RES_USAGE)) > 0)
75 return true;
76 }
77 return false;
78}
79
80static struct cgroup_subsys_state *hugetlb_cgroup_create(struct cgroup *cgroup)
81{
82 int idx;
83 struct cgroup *parent_cgroup;
84 struct hugetlb_cgroup *h_cgroup, *parent_h_cgroup;
85
86 h_cgroup = kzalloc(sizeof(*h_cgroup), GFP_KERNEL);
87 if (!h_cgroup)
88 return ERR_PTR(-ENOMEM);
89
90 parent_cgroup = cgroup->parent;
91 if (parent_cgroup) {
92 parent_h_cgroup = hugetlb_cgroup_from_cgroup(parent_cgroup);
93 for (idx = 0; idx < HUGE_MAX_HSTATE; idx++)
94 res_counter_init(&h_cgroup->hugepage[idx],
95 &parent_h_cgroup->hugepage[idx]);
96 } else {
97 root_h_cgroup = h_cgroup;
98 for (idx = 0; idx < HUGE_MAX_HSTATE; idx++)
99 res_counter_init(&h_cgroup->hugepage[idx], NULL);
100 }
101 return &h_cgroup->css;
102}
103
104static void hugetlb_cgroup_destroy(struct cgroup *cgroup)
105{
106 struct hugetlb_cgroup *h_cgroup;
107
108 h_cgroup = hugetlb_cgroup_from_cgroup(cgroup);
109 kfree(h_cgroup);
110}
111
Aneesh Kumar K.Vda1def52012-07-31 16:42:21 -0700112
113/*
114 * Should be called with hugetlb_lock held.
115 * Since we are holding hugetlb_lock, pages cannot get moved from
116 * active list or uncharged from the cgroup, So no need to get
117 * page reference and test for page active here. This function
118 * cannot fail.
119 */
120static void hugetlb_cgroup_move_parent(int idx, struct cgroup *cgroup,
121 struct page *page)
122{
123 int csize;
124 struct res_counter *counter;
125 struct res_counter *fail_res;
126 struct hugetlb_cgroup *page_hcg;
127 struct hugetlb_cgroup *h_cg = hugetlb_cgroup_from_cgroup(cgroup);
128 struct hugetlb_cgroup *parent = parent_hugetlb_cgroup(cgroup);
129
130 page_hcg = hugetlb_cgroup_from_page(page);
131 /*
132 * We can have pages in active list without any cgroup
133 * ie, hugepage with less than 3 pages. We can safely
134 * ignore those pages.
135 */
136 if (!page_hcg || page_hcg != h_cg)
137 goto out;
138
139 csize = PAGE_SIZE << compound_order(page);
140 if (!parent) {
141 parent = root_h_cgroup;
142 /* root has no limit */
143 res_counter_charge_nofail(&parent->hugepage[idx],
144 csize, &fail_res);
145 }
146 counter = &h_cg->hugepage[idx];
147 res_counter_uncharge_until(counter, counter->parent, csize);
148
149 set_hugetlb_cgroup(page, parent);
150out:
151 return;
152}
153
154/*
155 * Force the hugetlb cgroup to empty the hugetlb resources by moving them to
156 * the parent cgroup.
157 */
Aneesh Kumar K.V2bc64a22012-07-31 16:42:12 -0700158static int hugetlb_cgroup_pre_destroy(struct cgroup *cgroup)
159{
Aneesh Kumar K.Vda1def52012-07-31 16:42:21 -0700160 struct hstate *h;
161 struct page *page;
162 int ret = 0, idx = 0;
163
164 do {
165 if (cgroup_task_count(cgroup) ||
166 !list_empty(&cgroup->children)) {
167 ret = -EBUSY;
168 goto out;
169 }
170 for_each_hstate(h) {
171 spin_lock(&hugetlb_lock);
172 list_for_each_entry(page, &h->hugepage_activelist, lru)
173 hugetlb_cgroup_move_parent(idx, cgroup, page);
174
175 spin_unlock(&hugetlb_lock);
176 idx++;
177 }
178 cond_resched();
179 } while (hugetlb_cgroup_have_usage(cgroup));
180out:
181 return ret;
Aneesh Kumar K.V2bc64a22012-07-31 16:42:12 -0700182}
183
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -0700184int hugetlb_cgroup_charge_cgroup(int idx, unsigned long nr_pages,
185 struct hugetlb_cgroup **ptr)
186{
187 int ret = 0;
188 struct res_counter *fail_res;
189 struct hugetlb_cgroup *h_cg = NULL;
190 unsigned long csize = nr_pages * PAGE_SIZE;
191
192 if (hugetlb_cgroup_disabled())
193 goto done;
194 /*
195 * We don't charge any cgroup if the compound page have less
196 * than 3 pages.
197 */
198 if (huge_page_order(&hstates[idx]) < HUGETLB_CGROUP_MIN_ORDER)
199 goto done;
200again:
201 rcu_read_lock();
202 h_cg = hugetlb_cgroup_from_task(current);
203 if (!css_tryget(&h_cg->css)) {
204 rcu_read_unlock();
205 goto again;
206 }
207 rcu_read_unlock();
208
209 ret = res_counter_charge(&h_cg->hugepage[idx], csize, &fail_res);
210 css_put(&h_cg->css);
211done:
212 *ptr = h_cg;
213 return ret;
214}
215
216void hugetlb_cgroup_commit_charge(int idx, unsigned long nr_pages,
217 struct hugetlb_cgroup *h_cg,
218 struct page *page)
219{
220 if (hugetlb_cgroup_disabled() || !h_cg)
221 return;
222
223 spin_lock(&hugetlb_lock);
224 set_hugetlb_cgroup(page, h_cg);
225 spin_unlock(&hugetlb_lock);
226 return;
227}
228
229/*
230 * Should be called with hugetlb_lock held
231 */
232void hugetlb_cgroup_uncharge_page(int idx, unsigned long nr_pages,
233 struct page *page)
234{
235 struct hugetlb_cgroup *h_cg;
236 unsigned long csize = nr_pages * PAGE_SIZE;
237
238 if (hugetlb_cgroup_disabled())
239 return;
240 VM_BUG_ON(!spin_is_locked(&hugetlb_lock));
241 h_cg = hugetlb_cgroup_from_page(page);
242 if (unlikely(!h_cg))
243 return;
244 set_hugetlb_cgroup(page, NULL);
245 res_counter_uncharge(&h_cg->hugepage[idx], csize);
246 return;
247}
248
249void hugetlb_cgroup_uncharge_cgroup(int idx, unsigned long nr_pages,
250 struct hugetlb_cgroup *h_cg)
251{
252 unsigned long csize = nr_pages * PAGE_SIZE;
253
254 if (hugetlb_cgroup_disabled() || !h_cg)
255 return;
256
257 if (huge_page_order(&hstates[idx]) < HUGETLB_CGROUP_MIN_ORDER)
258 return;
259
260 res_counter_uncharge(&h_cg->hugepage[idx], csize);
261 return;
262}
263
Aneesh Kumar K.Vabb82062012-07-31 16:42:24 -0700264static ssize_t hugetlb_cgroup_read(struct cgroup *cgroup, struct cftype *cft,
265 struct file *file, char __user *buf,
266 size_t nbytes, loff_t *ppos)
267{
268 u64 val;
269 char str[64];
270 int idx, name, len;
271 struct hugetlb_cgroup *h_cg = hugetlb_cgroup_from_cgroup(cgroup);
272
273 idx = MEMFILE_IDX(cft->private);
274 name = MEMFILE_ATTR(cft->private);
275
276 val = res_counter_read_u64(&h_cg->hugepage[idx], name);
277 len = scnprintf(str, sizeof(str), "%llu\n", (unsigned long long)val);
278 return simple_read_from_buffer(buf, nbytes, ppos, str, len);
279}
280
281static int hugetlb_cgroup_write(struct cgroup *cgroup, struct cftype *cft,
282 const char *buffer)
283{
284 int idx, name, ret;
285 unsigned long long val;
286 struct hugetlb_cgroup *h_cg = hugetlb_cgroup_from_cgroup(cgroup);
287
288 idx = MEMFILE_IDX(cft->private);
289 name = MEMFILE_ATTR(cft->private);
290
291 switch (name) {
292 case RES_LIMIT:
293 if (hugetlb_cgroup_is_root(h_cg)) {
294 /* Can't set limit on root */
295 ret = -EINVAL;
296 break;
297 }
298 /* This function does all necessary parse...reuse it */
299 ret = res_counter_memparse_write_strategy(buffer, &val);
300 if (ret)
301 break;
302 ret = res_counter_set_limit(&h_cg->hugepage[idx], val);
303 break;
304 default:
305 ret = -EINVAL;
306 break;
307 }
308 return ret;
309}
310
311static int hugetlb_cgroup_reset(struct cgroup *cgroup, unsigned int event)
312{
313 int idx, name, ret = 0;
314 struct hugetlb_cgroup *h_cg = hugetlb_cgroup_from_cgroup(cgroup);
315
316 idx = MEMFILE_IDX(event);
317 name = MEMFILE_ATTR(event);
318
319 switch (name) {
320 case RES_MAX_USAGE:
321 res_counter_reset_max(&h_cg->hugepage[idx]);
322 break;
323 case RES_FAILCNT:
324 res_counter_reset_failcnt(&h_cg->hugepage[idx]);
325 break;
326 default:
327 ret = -EINVAL;
328 break;
329 }
330 return ret;
331}
332
333static char *mem_fmt(char *buf, int size, unsigned long hsize)
334{
335 if (hsize >= (1UL << 30))
336 snprintf(buf, size, "%luGB", hsize >> 30);
337 else if (hsize >= (1UL << 20))
338 snprintf(buf, size, "%luMB", hsize >> 20);
339 else
340 snprintf(buf, size, "%luKB", hsize >> 10);
341 return buf;
342}
343
344int __init hugetlb_cgroup_file_init(int idx)
345{
346 char buf[32];
347 struct cftype *cft;
348 struct hstate *h = &hstates[idx];
349
350 /* format the size */
351 mem_fmt(buf, 32, huge_page_size(h));
352
353 /* Add the limit file */
354 cft = &h->cgroup_files[0];
355 snprintf(cft->name, MAX_CFTYPE_NAME, "%s.limit_in_bytes", buf);
356 cft->private = MEMFILE_PRIVATE(idx, RES_LIMIT);
357 cft->read = hugetlb_cgroup_read;
358 cft->write_string = hugetlb_cgroup_write;
359
360 /* Add the usage file */
361 cft = &h->cgroup_files[1];
362 snprintf(cft->name, MAX_CFTYPE_NAME, "%s.usage_in_bytes", buf);
363 cft->private = MEMFILE_PRIVATE(idx, RES_USAGE);
364 cft->read = hugetlb_cgroup_read;
365
366 /* Add the MAX usage file */
367 cft = &h->cgroup_files[2];
368 snprintf(cft->name, MAX_CFTYPE_NAME, "%s.max_usage_in_bytes", buf);
369 cft->private = MEMFILE_PRIVATE(idx, RES_MAX_USAGE);
370 cft->trigger = hugetlb_cgroup_reset;
371 cft->read = hugetlb_cgroup_read;
372
373 /* Add the failcntfile */
374 cft = &h->cgroup_files[3];
375 snprintf(cft->name, MAX_CFTYPE_NAME, "%s.failcnt", buf);
376 cft->private = MEMFILE_PRIVATE(idx, RES_FAILCNT);
377 cft->trigger = hugetlb_cgroup_reset;
378 cft->read = hugetlb_cgroup_read;
379
380 /* NULL terminate the last cft */
381 cft = &h->cgroup_files[4];
382 memset(cft, 0, sizeof(*cft));
383
384 WARN_ON(cgroup_add_cftypes(&hugetlb_subsys, h->cgroup_files));
385
386 return 0;
387}
388
Aneesh Kumar K.V8e6ac7f2012-07-31 16:42:27 -0700389void hugetlb_cgroup_migrate(struct page *oldhpage, struct page *newhpage)
390{
391 struct hugetlb_cgroup *h_cg;
392
393 if (hugetlb_cgroup_disabled())
394 return;
395
396 VM_BUG_ON(!PageHuge(oldhpage));
397 spin_lock(&hugetlb_lock);
398 h_cg = hugetlb_cgroup_from_page(oldhpage);
399 set_hugetlb_cgroup(oldhpage, NULL);
400 cgroup_exclude_rmdir(&h_cg->css);
401
402 /* move the h_cg details to new cgroup */
403 set_hugetlb_cgroup(newhpage, h_cg);
404 spin_unlock(&hugetlb_lock);
405 cgroup_release_and_wakeup_rmdir(&h_cg->css);
406 return;
407}
408
Aneesh Kumar K.V2bc64a22012-07-31 16:42:12 -0700409struct cgroup_subsys hugetlb_subsys = {
410 .name = "hugetlb",
411 .create = hugetlb_cgroup_create,
412 .pre_destroy = hugetlb_cgroup_pre_destroy,
413 .destroy = hugetlb_cgroup_destroy,
414 .subsys_id = hugetlb_subsys_id,
415};