hugetlb: do not fail in hugetlb_cgroup_pre_destroy
[~shefty/rdma-dev.git] / mm / hugetlb_cgroup.c
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 container_of(s, struct hugetlb_cgroup, css);
40 }
41
42 static inline
43 struct 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
49 static inline
50 struct 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
56 static inline bool hugetlb_cgroup_is_root(struct hugetlb_cgroup *h_cg)
57 {
58         return (h_cg == root_h_cgroup);
59 }
60
61 static 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
68 static 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
80 static 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
104 static 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
112
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  */
120 static 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);
150 out:
151         return;
152 }
153
154 /*
155  * Force the hugetlb cgroup to empty the hugetlb resources by moving them to
156  * the parent cgroup.
157  */
158 static int hugetlb_cgroup_pre_destroy(struct cgroup *cgroup)
159 {
160         struct hstate *h;
161         struct page *page;
162         int idx = 0;
163
164         do {
165                 for_each_hstate(h) {
166                         spin_lock(&hugetlb_lock);
167                         list_for_each_entry(page, &h->hugepage_activelist, lru)
168                                 hugetlb_cgroup_move_parent(idx, cgroup, page);
169
170                         spin_unlock(&hugetlb_lock);
171                         idx++;
172                 }
173                 cond_resched();
174         } while (hugetlb_cgroup_have_usage(cgroup));
175
176         return 0;
177 }
178
179 int hugetlb_cgroup_charge_cgroup(int idx, unsigned long nr_pages,
180                                  struct hugetlb_cgroup **ptr)
181 {
182         int ret = 0;
183         struct res_counter *fail_res;
184         struct hugetlb_cgroup *h_cg = NULL;
185         unsigned long csize = nr_pages * PAGE_SIZE;
186
187         if (hugetlb_cgroup_disabled())
188                 goto done;
189         /*
190          * We don't charge any cgroup if the compound page have less
191          * than 3 pages.
192          */
193         if (huge_page_order(&hstates[idx]) < HUGETLB_CGROUP_MIN_ORDER)
194                 goto done;
195 again:
196         rcu_read_lock();
197         h_cg = hugetlb_cgroup_from_task(current);
198         if (!css_tryget(&h_cg->css)) {
199                 rcu_read_unlock();
200                 goto again;
201         }
202         rcu_read_unlock();
203
204         ret = res_counter_charge(&h_cg->hugepage[idx], csize, &fail_res);
205         css_put(&h_cg->css);
206 done:
207         *ptr = h_cg;
208         return ret;
209 }
210
211 /* Should be called with hugetlb_lock held */
212 void hugetlb_cgroup_commit_charge(int idx, unsigned long nr_pages,
213                                   struct hugetlb_cgroup *h_cg,
214                                   struct page *page)
215 {
216         if (hugetlb_cgroup_disabled() || !h_cg)
217                 return;
218
219         set_hugetlb_cgroup(page, h_cg);
220         return;
221 }
222
223 /*
224  * Should be called with hugetlb_lock held
225  */
226 void hugetlb_cgroup_uncharge_page(int idx, unsigned long nr_pages,
227                                   struct page *page)
228 {
229         struct hugetlb_cgroup *h_cg;
230         unsigned long csize = nr_pages * PAGE_SIZE;
231
232         if (hugetlb_cgroup_disabled())
233                 return;
234         VM_BUG_ON(!spin_is_locked(&hugetlb_lock));
235         h_cg = hugetlb_cgroup_from_page(page);
236         if (unlikely(!h_cg))
237                 return;
238         set_hugetlb_cgroup(page, NULL);
239         res_counter_uncharge(&h_cg->hugepage[idx], csize);
240         return;
241 }
242
243 void hugetlb_cgroup_uncharge_cgroup(int idx, unsigned long nr_pages,
244                                     struct hugetlb_cgroup *h_cg)
245 {
246         unsigned long csize = nr_pages * PAGE_SIZE;
247
248         if (hugetlb_cgroup_disabled() || !h_cg)
249                 return;
250
251         if (huge_page_order(&hstates[idx]) < HUGETLB_CGROUP_MIN_ORDER)
252                 return;
253
254         res_counter_uncharge(&h_cg->hugepage[idx], csize);
255         return;
256 }
257
258 static ssize_t hugetlb_cgroup_read(struct cgroup *cgroup, struct cftype *cft,
259                                    struct file *file, char __user *buf,
260                                    size_t nbytes, loff_t *ppos)
261 {
262         u64 val;
263         char str[64];
264         int idx, name, len;
265         struct hugetlb_cgroup *h_cg = hugetlb_cgroup_from_cgroup(cgroup);
266
267         idx = MEMFILE_IDX(cft->private);
268         name = MEMFILE_ATTR(cft->private);
269
270         val = res_counter_read_u64(&h_cg->hugepage[idx], name);
271         len = scnprintf(str, sizeof(str), "%llu\n", (unsigned long long)val);
272         return simple_read_from_buffer(buf, nbytes, ppos, str, len);
273 }
274
275 static int hugetlb_cgroup_write(struct cgroup *cgroup, struct cftype *cft,
276                                 const char *buffer)
277 {
278         int idx, name, ret;
279         unsigned long long val;
280         struct hugetlb_cgroup *h_cg = hugetlb_cgroup_from_cgroup(cgroup);
281
282         idx = MEMFILE_IDX(cft->private);
283         name = MEMFILE_ATTR(cft->private);
284
285         switch (name) {
286         case RES_LIMIT:
287                 if (hugetlb_cgroup_is_root(h_cg)) {
288                         /* Can't set limit on root */
289                         ret = -EINVAL;
290                         break;
291                 }
292                 /* This function does all necessary parse...reuse it */
293                 ret = res_counter_memparse_write_strategy(buffer, &val);
294                 if (ret)
295                         break;
296                 ret = res_counter_set_limit(&h_cg->hugepage[idx], val);
297                 break;
298         default:
299                 ret = -EINVAL;
300                 break;
301         }
302         return ret;
303 }
304
305 static int hugetlb_cgroup_reset(struct cgroup *cgroup, unsigned int event)
306 {
307         int idx, name, ret = 0;
308         struct hugetlb_cgroup *h_cg = hugetlb_cgroup_from_cgroup(cgroup);
309
310         idx = MEMFILE_IDX(event);
311         name = MEMFILE_ATTR(event);
312
313         switch (name) {
314         case RES_MAX_USAGE:
315                 res_counter_reset_max(&h_cg->hugepage[idx]);
316                 break;
317         case RES_FAILCNT:
318                 res_counter_reset_failcnt(&h_cg->hugepage[idx]);
319                 break;
320         default:
321                 ret = -EINVAL;
322                 break;
323         }
324         return ret;
325 }
326
327 static char *mem_fmt(char *buf, int size, unsigned long hsize)
328 {
329         if (hsize >= (1UL << 30))
330                 snprintf(buf, size, "%luGB", hsize >> 30);
331         else if (hsize >= (1UL << 20))
332                 snprintf(buf, size, "%luMB", hsize >> 20);
333         else
334                 snprintf(buf, size, "%luKB", hsize >> 10);
335         return buf;
336 }
337
338 int __init hugetlb_cgroup_file_init(int idx)
339 {
340         char buf[32];
341         struct cftype *cft;
342         struct hstate *h = &hstates[idx];
343
344         /* format the size */
345         mem_fmt(buf, 32, huge_page_size(h));
346
347         /* Add the limit file */
348         cft = &h->cgroup_files[0];
349         snprintf(cft->name, MAX_CFTYPE_NAME, "%s.limit_in_bytes", buf);
350         cft->private = MEMFILE_PRIVATE(idx, RES_LIMIT);
351         cft->read = hugetlb_cgroup_read;
352         cft->write_string = hugetlb_cgroup_write;
353
354         /* Add the usage file */
355         cft = &h->cgroup_files[1];
356         snprintf(cft->name, MAX_CFTYPE_NAME, "%s.usage_in_bytes", buf);
357         cft->private = MEMFILE_PRIVATE(idx, RES_USAGE);
358         cft->read = hugetlb_cgroup_read;
359
360         /* Add the MAX usage file */
361         cft = &h->cgroup_files[2];
362         snprintf(cft->name, MAX_CFTYPE_NAME, "%s.max_usage_in_bytes", buf);
363         cft->private = MEMFILE_PRIVATE(idx, RES_MAX_USAGE);
364         cft->trigger = hugetlb_cgroup_reset;
365         cft->read = hugetlb_cgroup_read;
366
367         /* Add the failcntfile */
368         cft = &h->cgroup_files[3];
369         snprintf(cft->name, MAX_CFTYPE_NAME, "%s.failcnt", buf);
370         cft->private  = MEMFILE_PRIVATE(idx, RES_FAILCNT);
371         cft->trigger  = hugetlb_cgroup_reset;
372         cft->read = hugetlb_cgroup_read;
373
374         /* NULL terminate the last cft */
375         cft = &h->cgroup_files[4];
376         memset(cft, 0, sizeof(*cft));
377
378         WARN_ON(cgroup_add_cftypes(&hugetlb_subsys, h->cgroup_files));
379
380         return 0;
381 }
382
383 /*
384  * hugetlb_lock will make sure a parallel cgroup rmdir won't happen
385  * when we migrate hugepages
386  */
387 void hugetlb_cgroup_migrate(struct page *oldhpage, struct page *newhpage)
388 {
389         struct hugetlb_cgroup *h_cg;
390         struct hstate *h = page_hstate(oldhpage);
391
392         if (hugetlb_cgroup_disabled())
393                 return;
394
395         VM_BUG_ON(!PageHuge(oldhpage));
396         spin_lock(&hugetlb_lock);
397         h_cg = hugetlb_cgroup_from_page(oldhpage);
398         set_hugetlb_cgroup(oldhpage, NULL);
399
400         /* move the h_cg details to new cgroup */
401         set_hugetlb_cgroup(newhpage, h_cg);
402         list_move(&newhpage->lru, &h->hugepage_activelist);
403         spin_unlock(&hugetlb_lock);
404         return;
405 }
406
407 struct cgroup_subsys hugetlb_subsys = {
408         .name = "hugetlb",
409         .create     = hugetlb_cgroup_create,
410         .pre_destroy = hugetlb_cgroup_pre_destroy,
411         .destroy    = hugetlb_cgroup_destroy,
412         .subsys_id  = hugetlb_subsys_id,
413 };