xref: /openbmc/linux/mm/hugetlb_cgroup.c (revision 089a49b6)
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/slab.h>
18 #include <linux/hugetlb.h>
19 #include <linux/hugetlb_cgroup.h>
20 
21 struct 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 
29 #define MEMFILE_PRIVATE(x, val)	(((x) << 16) | (val))
30 #define MEMFILE_IDX(val)	(((val) >> 16) & 0xffff)
31 #define MEMFILE_ATTR(val)	((val) & 0xffff)
32 
33 struct cgroup_subsys hugetlb_subsys __read_mostly;
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_subsys_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(css_parent(&h_cg->css));
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 ((res_counter_read_u64(&h_cg->hugepage[idx], RES_USAGE)) > 0)
65 			return true;
66 	}
67 	return false;
68 }
69 
70 static struct cgroup_subsys_state *
71 hugetlb_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 			res_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 			res_counter_init(&h_cgroup->hugepage[idx], NULL);
89 	}
90 	return &h_cgroup->css;
91 }
92 
93 static 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  */
109 static void hugetlb_cgroup_move_parent(int idx, struct hugetlb_cgroup *h_cg,
110 				       struct page *page)
111 {
112 	int csize;
113 	struct res_counter *counter;
114 	struct res_counter *fail_res;
115 	struct hugetlb_cgroup *page_hcg;
116 	struct hugetlb_cgroup *parent = parent_hugetlb_cgroup(h_cg);
117 
118 	page_hcg = hugetlb_cgroup_from_page(page);
119 	/*
120 	 * We can have pages in active list without any cgroup
121 	 * ie, hugepage with less than 3 pages. We can safely
122 	 * ignore those pages.
123 	 */
124 	if (!page_hcg || page_hcg != h_cg)
125 		goto out;
126 
127 	csize = PAGE_SIZE << compound_order(page);
128 	if (!parent) {
129 		parent = root_h_cgroup;
130 		/* root has no limit */
131 		res_counter_charge_nofail(&parent->hugepage[idx],
132 					  csize, &fail_res);
133 	}
134 	counter = &h_cg->hugepage[idx];
135 	res_counter_uncharge_until(counter, counter->parent, csize);
136 
137 	set_hugetlb_cgroup(page, parent);
138 out:
139 	return;
140 }
141 
142 /*
143  * Force the hugetlb cgroup to empty the hugetlb resources by moving them to
144  * the parent cgroup.
145  */
146 static void hugetlb_cgroup_css_offline(struct cgroup_subsys_state *css)
147 {
148 	struct hugetlb_cgroup *h_cg = hugetlb_cgroup_from_css(css);
149 	struct hstate *h;
150 	struct page *page;
151 	int idx = 0;
152 
153 	do {
154 		for_each_hstate(h) {
155 			spin_lock(&hugetlb_lock);
156 			list_for_each_entry(page, &h->hugepage_activelist, lru)
157 				hugetlb_cgroup_move_parent(idx, h_cg, page);
158 
159 			spin_unlock(&hugetlb_lock);
160 			idx++;
161 		}
162 		cond_resched();
163 	} while (hugetlb_cgroup_have_usage(h_cg));
164 }
165 
166 int hugetlb_cgroup_charge_cgroup(int idx, unsigned long nr_pages,
167 				 struct hugetlb_cgroup **ptr)
168 {
169 	int ret = 0;
170 	struct res_counter *fail_res;
171 	struct hugetlb_cgroup *h_cg = NULL;
172 	unsigned long csize = nr_pages * PAGE_SIZE;
173 
174 	if (hugetlb_cgroup_disabled())
175 		goto done;
176 	/*
177 	 * We don't charge any cgroup if the compound page have less
178 	 * than 3 pages.
179 	 */
180 	if (huge_page_order(&hstates[idx]) < HUGETLB_CGROUP_MIN_ORDER)
181 		goto done;
182 again:
183 	rcu_read_lock();
184 	h_cg = hugetlb_cgroup_from_task(current);
185 	if (!css_tryget(&h_cg->css)) {
186 		rcu_read_unlock();
187 		goto again;
188 	}
189 	rcu_read_unlock();
190 
191 	ret = res_counter_charge(&h_cg->hugepage[idx], csize, &fail_res);
192 	css_put(&h_cg->css);
193 done:
194 	*ptr = h_cg;
195 	return ret;
196 }
197 
198 /* Should be called with hugetlb_lock held */
199 void hugetlb_cgroup_commit_charge(int idx, unsigned long nr_pages,
200 				  struct hugetlb_cgroup *h_cg,
201 				  struct page *page)
202 {
203 	if (hugetlb_cgroup_disabled() || !h_cg)
204 		return;
205 
206 	set_hugetlb_cgroup(page, h_cg);
207 	return;
208 }
209 
210 /*
211  * Should be called with hugetlb_lock held
212  */
213 void hugetlb_cgroup_uncharge_page(int idx, unsigned long nr_pages,
214 				  struct page *page)
215 {
216 	struct hugetlb_cgroup *h_cg;
217 	unsigned long csize = nr_pages * PAGE_SIZE;
218 
219 	if (hugetlb_cgroup_disabled())
220 		return;
221 	VM_BUG_ON(!spin_is_locked(&hugetlb_lock));
222 	h_cg = hugetlb_cgroup_from_page(page);
223 	if (unlikely(!h_cg))
224 		return;
225 	set_hugetlb_cgroup(page, NULL);
226 	res_counter_uncharge(&h_cg->hugepage[idx], csize);
227 	return;
228 }
229 
230 void hugetlb_cgroup_uncharge_cgroup(int idx, unsigned long nr_pages,
231 				    struct hugetlb_cgroup *h_cg)
232 {
233 	unsigned long csize = nr_pages * PAGE_SIZE;
234 
235 	if (hugetlb_cgroup_disabled() || !h_cg)
236 		return;
237 
238 	if (huge_page_order(&hstates[idx]) < HUGETLB_CGROUP_MIN_ORDER)
239 		return;
240 
241 	res_counter_uncharge(&h_cg->hugepage[idx], csize);
242 	return;
243 }
244 
245 static ssize_t hugetlb_cgroup_read(struct cgroup_subsys_state *css,
246 				   struct cftype *cft, struct file *file,
247 				   char __user *buf, size_t nbytes,
248 				   loff_t *ppos)
249 {
250 	u64 val;
251 	char str[64];
252 	int idx, name, len;
253 	struct hugetlb_cgroup *h_cg = hugetlb_cgroup_from_css(css);
254 
255 	idx = MEMFILE_IDX(cft->private);
256 	name = MEMFILE_ATTR(cft->private);
257 
258 	val = res_counter_read_u64(&h_cg->hugepage[idx], name);
259 	len = scnprintf(str, sizeof(str), "%llu\n", (unsigned long long)val);
260 	return simple_read_from_buffer(buf, nbytes, ppos, str, len);
261 }
262 
263 static int hugetlb_cgroup_write(struct cgroup_subsys_state *css,
264 				struct cftype *cft, const char *buffer)
265 {
266 	int idx, name, ret;
267 	unsigned long long val;
268 	struct hugetlb_cgroup *h_cg = hugetlb_cgroup_from_css(css);
269 
270 	idx = MEMFILE_IDX(cft->private);
271 	name = MEMFILE_ATTR(cft->private);
272 
273 	switch (name) {
274 	case RES_LIMIT:
275 		if (hugetlb_cgroup_is_root(h_cg)) {
276 			/* Can't set limit on root */
277 			ret = -EINVAL;
278 			break;
279 		}
280 		/* This function does all necessary parse...reuse it */
281 		ret = res_counter_memparse_write_strategy(buffer, &val);
282 		if (ret)
283 			break;
284 		ret = res_counter_set_limit(&h_cg->hugepage[idx], val);
285 		break;
286 	default:
287 		ret = -EINVAL;
288 		break;
289 	}
290 	return ret;
291 }
292 
293 static int hugetlb_cgroup_reset(struct cgroup_subsys_state *css,
294 				unsigned int event)
295 {
296 	int idx, name, ret = 0;
297 	struct hugetlb_cgroup *h_cg = hugetlb_cgroup_from_css(css);
298 
299 	idx = MEMFILE_IDX(event);
300 	name = MEMFILE_ATTR(event);
301 
302 	switch (name) {
303 	case RES_MAX_USAGE:
304 		res_counter_reset_max(&h_cg->hugepage[idx]);
305 		break;
306 	case RES_FAILCNT:
307 		res_counter_reset_failcnt(&h_cg->hugepage[idx]);
308 		break;
309 	default:
310 		ret = -EINVAL;
311 		break;
312 	}
313 	return ret;
314 }
315 
316 static char *mem_fmt(char *buf, int size, unsigned long hsize)
317 {
318 	if (hsize >= (1UL << 30))
319 		snprintf(buf, size, "%luGB", hsize >> 30);
320 	else if (hsize >= (1UL << 20))
321 		snprintf(buf, size, "%luMB", hsize >> 20);
322 	else
323 		snprintf(buf, size, "%luKB", hsize >> 10);
324 	return buf;
325 }
326 
327 static void __init __hugetlb_cgroup_file_init(int idx)
328 {
329 	char buf[32];
330 	struct cftype *cft;
331 	struct hstate *h = &hstates[idx];
332 
333 	/* format the size */
334 	mem_fmt(buf, 32, huge_page_size(h));
335 
336 	/* Add the limit file */
337 	cft = &h->cgroup_files[0];
338 	snprintf(cft->name, MAX_CFTYPE_NAME, "%s.limit_in_bytes", buf);
339 	cft->private = MEMFILE_PRIVATE(idx, RES_LIMIT);
340 	cft->read = hugetlb_cgroup_read;
341 	cft->write_string = hugetlb_cgroup_write;
342 
343 	/* Add the usage file */
344 	cft = &h->cgroup_files[1];
345 	snprintf(cft->name, MAX_CFTYPE_NAME, "%s.usage_in_bytes", buf);
346 	cft->private = MEMFILE_PRIVATE(idx, RES_USAGE);
347 	cft->read = hugetlb_cgroup_read;
348 
349 	/* Add the MAX usage file */
350 	cft = &h->cgroup_files[2];
351 	snprintf(cft->name, MAX_CFTYPE_NAME, "%s.max_usage_in_bytes", buf);
352 	cft->private = MEMFILE_PRIVATE(idx, RES_MAX_USAGE);
353 	cft->trigger = hugetlb_cgroup_reset;
354 	cft->read = hugetlb_cgroup_read;
355 
356 	/* Add the failcntfile */
357 	cft = &h->cgroup_files[3];
358 	snprintf(cft->name, MAX_CFTYPE_NAME, "%s.failcnt", buf);
359 	cft->private  = MEMFILE_PRIVATE(idx, RES_FAILCNT);
360 	cft->trigger  = hugetlb_cgroup_reset;
361 	cft->read = hugetlb_cgroup_read;
362 
363 	/* NULL terminate the last cft */
364 	cft = &h->cgroup_files[4];
365 	memset(cft, 0, sizeof(*cft));
366 
367 	WARN_ON(cgroup_add_cftypes(&hugetlb_subsys, h->cgroup_files));
368 
369 	return;
370 }
371 
372 void __init hugetlb_cgroup_file_init(void)
373 {
374 	struct hstate *h;
375 
376 	for_each_hstate(h) {
377 		/*
378 		 * Add cgroup control files only if the huge page consists
379 		 * of more than two normal pages. This is because we use
380 		 * page[2].lru.next for storing cgroup details.
381 		 */
382 		if (huge_page_order(h) >= HUGETLB_CGROUP_MIN_ORDER)
383 			__hugetlb_cgroup_file_init(hstate_index(h));
384 	}
385 }
386 
387 /*
388  * hugetlb_lock will make sure a parallel cgroup rmdir won't happen
389  * when we migrate hugepages
390  */
391 void hugetlb_cgroup_migrate(struct page *oldhpage, struct page *newhpage)
392 {
393 	struct hugetlb_cgroup *h_cg;
394 	struct hstate *h = page_hstate(oldhpage);
395 
396 	if (hugetlb_cgroup_disabled())
397 		return;
398 
399 	VM_BUG_ON(!PageHuge(oldhpage));
400 	spin_lock(&hugetlb_lock);
401 	h_cg = hugetlb_cgroup_from_page(oldhpage);
402 	set_hugetlb_cgroup(oldhpage, NULL);
403 
404 	/* move the h_cg details to new cgroup */
405 	set_hugetlb_cgroup(newhpage, h_cg);
406 	list_move(&newhpage->lru, &h->hugepage_activelist);
407 	spin_unlock(&hugetlb_lock);
408 	return;
409 }
410 
411 struct cgroup_subsys hugetlb_subsys = {
412 	.name = "hugetlb",
413 	.css_alloc	= hugetlb_cgroup_css_alloc,
414 	.css_offline	= hugetlb_cgroup_css_offline,
415 	.css_free	= hugetlb_cgroup_css_free,
416 	.subsys_id	= hugetlb_subsys_id,
417 };
418