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 struct cgroup_subsys_state *
71hugetlb_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
72{
73	struct hugetlb_cgroup *parent_h_cgroup = hugetlb_cgroup_from_css(parent_css);
74	struct hugetlb_cgroup *h_cgroup;
75	int idx;
76
77	h_cgroup = kzalloc(sizeof(*h_cgroup), GFP_KERNEL);
78	if (!h_cgroup)
79		return ERR_PTR(-ENOMEM);
80
81	if (parent_h_cgroup) {
82		for (idx = 0; idx < HUGE_MAX_HSTATE; idx++)
83			page_counter_init(&h_cgroup->hugepage[idx],
84					  &parent_h_cgroup->hugepage[idx]);
85	} else {
86		root_h_cgroup = h_cgroup;
87		for (idx = 0; idx < HUGE_MAX_HSTATE; idx++)
88			page_counter_init(&h_cgroup->hugepage[idx], NULL);
89	}
90	return &h_cgroup->css;
91}
92
93static void hugetlb_cgroup_css_free(struct cgroup_subsys_state *css)
94{
95	struct hugetlb_cgroup *h_cgroup;
96
97	h_cgroup = hugetlb_cgroup_from_css(css);
98	kfree(h_cgroup);
99}
100
101
102/*
103 * Should be called with hugetlb_lock held.
104 * Since we are holding hugetlb_lock, pages cannot get moved from
105 * active list or uncharged from the cgroup, So no need to get
106 * page reference and test for page active here. This function
107 * cannot fail.
108 */
109static void hugetlb_cgroup_move_parent(int idx, struct hugetlb_cgroup *h_cg,
110				       struct page *page)
111{
112	unsigned int nr_pages;
113	struct page_counter *counter;
114	struct hugetlb_cgroup *page_hcg;
115	struct hugetlb_cgroup *parent = parent_hugetlb_cgroup(h_cg);
116
117	page_hcg = hugetlb_cgroup_from_page(page);
118	/*
119	 * We can have pages in active list without any cgroup
120	 * ie, hugepage with less than 3 pages. We can safely
121	 * ignore those pages.
122	 */
123	if (!page_hcg || page_hcg != h_cg)
124		goto out;
125
126	nr_pages = 1 << compound_order(page);
127	if (!parent) {
128		parent = root_h_cgroup;
129		/* root has no limit */
130		page_counter_charge(&parent->hugepage[idx], nr_pages);
131	}
132	counter = &h_cg->hugepage[idx];
133	/* Take the pages off the local counter */
134	page_counter_cancel(counter, nr_pages);
135
136	set_hugetlb_cgroup(page, parent);
137out:
138	return;
139}
140
141/*
142 * Force the hugetlb cgroup to empty the hugetlb resources by moving them to
143 * the parent cgroup.
144 */
145static void hugetlb_cgroup_css_offline(struct cgroup_subsys_state *css)
146{
147	struct hugetlb_cgroup *h_cg = hugetlb_cgroup_from_css(css);
148	struct hstate *h;
149	struct page *page;
150	int idx = 0;
151
152	do {
153		for_each_hstate(h) {
154			spin_lock(&hugetlb_lock);
155			list_for_each_entry(page, &h->hugepage_activelist, lru)
156				hugetlb_cgroup_move_parent(idx, h_cg, page);
157
158			spin_unlock(&hugetlb_lock);
159			idx++;
160		}
161		cond_resched();
162	} while (hugetlb_cgroup_have_usage(h_cg));
163}
164
165int hugetlb_cgroup_charge_cgroup(int idx, unsigned long nr_pages,
166				 struct hugetlb_cgroup **ptr)
167{
168	int ret = 0;
169	struct page_counter *counter;
170	struct hugetlb_cgroup *h_cg = NULL;
171
172	if (hugetlb_cgroup_disabled())
173		goto done;
174	/*
175	 * We don't charge any cgroup if the compound page have less
176	 * than 3 pages.
177	 */
178	if (huge_page_order(&hstates[idx]) < HUGETLB_CGROUP_MIN_ORDER)
179		goto done;
180again:
181	rcu_read_lock();
182	h_cg = hugetlb_cgroup_from_task(current);
183	if (!css_tryget_online(&h_cg->css)) {
184		rcu_read_unlock();
185		goto again;
186	}
187	rcu_read_unlock();
188
189	ret = page_counter_try_charge(&h_cg->hugepage[idx], nr_pages, &counter);
190	css_put(&h_cg->css);
191done:
192	*ptr = h_cg;
193	return ret;
194}
195
196/* Should be called with hugetlb_lock held */
197void hugetlb_cgroup_commit_charge(int idx, unsigned long nr_pages,
198				  struct hugetlb_cgroup *h_cg,
199				  struct page *page)
200{
201	if (hugetlb_cgroup_disabled() || !h_cg)
202		return;
203
204	set_hugetlb_cgroup(page, h_cg);
205	return;
206}
207
208/*
209 * Should be called with hugetlb_lock held
210 */
211void hugetlb_cgroup_uncharge_page(int idx, unsigned long nr_pages,
212				  struct page *page)
213{
214	struct hugetlb_cgroup *h_cg;
215
216	if (hugetlb_cgroup_disabled())
217		return;
218	lockdep_assert_held(&hugetlb_lock);
219	h_cg = hugetlb_cgroup_from_page(page);
220	if (unlikely(!h_cg))
221		return;
222	set_hugetlb_cgroup(page, NULL);
223	page_counter_uncharge(&h_cg->hugepage[idx], nr_pages);
224	return;
225}
226
227void hugetlb_cgroup_uncharge_cgroup(int idx, unsigned long nr_pages,
228				    struct hugetlb_cgroup *h_cg)
229{
230	if (hugetlb_cgroup_disabled() || !h_cg)
231		return;
232
233	if (huge_page_order(&hstates[idx]) < HUGETLB_CGROUP_MIN_ORDER)
234		return;
235
236	page_counter_uncharge(&h_cg->hugepage[idx], nr_pages);
237	return;
238}
239
240enum {
241	RES_USAGE,
242	RES_LIMIT,
243	RES_MAX_USAGE,
244	RES_FAILCNT,
245};
246
247static u64 hugetlb_cgroup_read_u64(struct cgroup_subsys_state *css,
248				   struct cftype *cft)
249{
250	struct page_counter *counter;
251	struct hugetlb_cgroup *h_cg = hugetlb_cgroup_from_css(css);
252
253	counter = &h_cg->hugepage[MEMFILE_IDX(cft->private)];
254
255	switch (MEMFILE_ATTR(cft->private)) {
256	case RES_USAGE:
257		return (u64)page_counter_read(counter) * PAGE_SIZE;
258	case RES_LIMIT:
259		return (u64)counter->limit * PAGE_SIZE;
260	case RES_MAX_USAGE:
261		return (u64)counter->watermark * PAGE_SIZE;
262	case RES_FAILCNT:
263		return counter->failcnt;
264	default:
265		BUG();
266	}
267}
268
269static DEFINE_MUTEX(hugetlb_limit_mutex);
270
271static ssize_t hugetlb_cgroup_write(struct kernfs_open_file *of,
272				    char *buf, size_t nbytes, loff_t off)
273{
274	int ret, idx;
275	unsigned long nr_pages;
276	struct hugetlb_cgroup *h_cg = hugetlb_cgroup_from_css(of_css(of));
277
278	if (hugetlb_cgroup_is_root(h_cg)) /* Can't set limit on root */
279		return -EINVAL;
280
281	buf = strstrip(buf);
282	ret = page_counter_memparse(buf, "-1", &nr_pages);
283	if (ret)
284		return ret;
285
286	idx = MEMFILE_IDX(of_cft(of)->private);
287
288	switch (MEMFILE_ATTR(of_cft(of)->private)) {
289	case RES_LIMIT:
290		mutex_lock(&hugetlb_limit_mutex);
291		ret = page_counter_limit(&h_cg->hugepage[idx], nr_pages);
292		mutex_unlock(&hugetlb_limit_mutex);
293		break;
294	default:
295		ret = -EINVAL;
296		break;
297	}
298	return ret ?: nbytes;
299}
300
301static ssize_t hugetlb_cgroup_reset(struct kernfs_open_file *of,
302				    char *buf, size_t nbytes, loff_t off)
303{
304	int ret = 0;
305	struct page_counter *counter;
306	struct hugetlb_cgroup *h_cg = hugetlb_cgroup_from_css(of_css(of));
307
308	counter = &h_cg->hugepage[MEMFILE_IDX(of_cft(of)->private)];
309
310	switch (MEMFILE_ATTR(of_cft(of)->private)) {
311	case RES_MAX_USAGE:
312		page_counter_reset_watermark(counter);
313		break;
314	case RES_FAILCNT:
315		counter->failcnt = 0;
316		break;
317	default:
318		ret = -EINVAL;
319		break;
320	}
321	return ret ?: nbytes;
322}
323
324static char *mem_fmt(char *buf, int size, unsigned long hsize)
325{
326	if (hsize >= (1UL << 30))
327		snprintf(buf, size, "%luGB", hsize >> 30);
328	else if (hsize >= (1UL << 20))
329		snprintf(buf, size, "%luMB", hsize >> 20);
330	else
331		snprintf(buf, size, "%luKB", hsize >> 10);
332	return buf;
333}
334
335static void __init __hugetlb_cgroup_file_init(int idx)
336{
337	char buf[32];
338	struct cftype *cft;
339	struct hstate *h = &hstates[idx];
340
341	/* format the size */
342	mem_fmt(buf, 32, huge_page_size(h));
343
344	/* Add the limit file */
345	cft = &h->cgroup_files[0];
346	snprintf(cft->name, MAX_CFTYPE_NAME, "%s.limit_in_bytes", buf);
347	cft->private = MEMFILE_PRIVATE(idx, RES_LIMIT);
348	cft->read_u64 = hugetlb_cgroup_read_u64;
349	cft->write = hugetlb_cgroup_write;
350
351	/* Add the usage file */
352	cft = &h->cgroup_files[1];
353	snprintf(cft->name, MAX_CFTYPE_NAME, "%s.usage_in_bytes", buf);
354	cft->private = MEMFILE_PRIVATE(idx, RES_USAGE);
355	cft->read_u64 = hugetlb_cgroup_read_u64;
356
357	/* Add the MAX usage file */
358	cft = &h->cgroup_files[2];
359	snprintf(cft->name, MAX_CFTYPE_NAME, "%s.max_usage_in_bytes", buf);
360	cft->private = MEMFILE_PRIVATE(idx, RES_MAX_USAGE);
361	cft->write = hugetlb_cgroup_reset;
362	cft->read_u64 = hugetlb_cgroup_read_u64;
363
364	/* Add the failcntfile */
365	cft = &h->cgroup_files[3];
366	snprintf(cft->name, MAX_CFTYPE_NAME, "%s.failcnt", buf);
367	cft->private  = MEMFILE_PRIVATE(idx, RES_FAILCNT);
368	cft->write = hugetlb_cgroup_reset;
369	cft->read_u64 = hugetlb_cgroup_read_u64;
370
371	/* NULL terminate the last cft */
372	cft = &h->cgroup_files[4];
373	memset(cft, 0, sizeof(*cft));
374
375	WARN_ON(cgroup_add_legacy_cftypes(&hugetlb_cgrp_subsys,
376					  h->cgroup_files));
377}
378
379void __init hugetlb_cgroup_file_init(void)
380{
381	struct hstate *h;
382
383	for_each_hstate(h) {
384		/*
385		 * Add cgroup control files only if the huge page consists
386		 * of more than two normal pages. This is because we use
387		 * page[2].lru.next for storing cgroup details.
388		 */
389		if (huge_page_order(h) >= HUGETLB_CGROUP_MIN_ORDER)
390			__hugetlb_cgroup_file_init(hstate_index(h));
391	}
392}
393
394/*
395 * hugetlb_lock will make sure a parallel cgroup rmdir won't happen
396 * when we migrate hugepages
397 */
398void hugetlb_cgroup_migrate(struct page *oldhpage, struct page *newhpage)
399{
400	struct hugetlb_cgroup *h_cg;
401	struct hstate *h = page_hstate(oldhpage);
402
403	if (hugetlb_cgroup_disabled())
404		return;
405
406	VM_BUG_ON_PAGE(!PageHuge(oldhpage), oldhpage);
407	spin_lock(&hugetlb_lock);
408	h_cg = hugetlb_cgroup_from_page(oldhpage);
409	set_hugetlb_cgroup(oldhpage, NULL);
410
411	/* move the h_cg details to new cgroup */
412	set_hugetlb_cgroup(newhpage, h_cg);
413	list_move(&newhpage->lru, &h->hugepage_activelist);
414	spin_unlock(&hugetlb_lock);
415	return;
416}
417
418struct cgroup_subsys hugetlb_cgrp_subsys = {
419	.css_alloc	= hugetlb_cgroup_css_alloc,
420	.css_offline	= hugetlb_cgroup_css_offline,
421	.css_free	= hugetlb_cgroup_css_free,
422};
423