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 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)); 180 out: 181 return ret; 182 } 183 184 int 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; 200 again: 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); 211 done: 212 *ptr = h_cg; 213 return ret; 214 } 215 216 /* Should be called with hugetlb_lock held */ 217 void hugetlb_cgroup_commit_charge(int idx, unsigned long nr_pages, 218 struct hugetlb_cgroup *h_cg, 219 struct page *page) 220 { 221 if (hugetlb_cgroup_disabled() || !h_cg) 222 return; 223 224 set_hugetlb_cgroup(page, h_cg); 225 return; 226 } 227 228 /* 229 * Should be called with hugetlb_lock held 230 */ 231 void hugetlb_cgroup_uncharge_page(int idx, unsigned long nr_pages, 232 struct page *page) 233 { 234 struct hugetlb_cgroup *h_cg; 235 unsigned long csize = nr_pages * PAGE_SIZE; 236 237 if (hugetlb_cgroup_disabled()) 238 return; 239 VM_BUG_ON(!spin_is_locked(&hugetlb_lock)); 240 h_cg = hugetlb_cgroup_from_page(page); 241 if (unlikely(!h_cg)) 242 return; 243 set_hugetlb_cgroup(page, NULL); 244 res_counter_uncharge(&h_cg->hugepage[idx], csize); 245 return; 246 } 247 248 void hugetlb_cgroup_uncharge_cgroup(int idx, unsigned long nr_pages, 249 struct hugetlb_cgroup *h_cg) 250 { 251 unsigned long csize = nr_pages * PAGE_SIZE; 252 253 if (hugetlb_cgroup_disabled() || !h_cg) 254 return; 255 256 if (huge_page_order(&hstates[idx]) < HUGETLB_CGROUP_MIN_ORDER) 257 return; 258 259 res_counter_uncharge(&h_cg->hugepage[idx], csize); 260 return; 261 } 262 263 static ssize_t hugetlb_cgroup_read(struct cgroup *cgroup, struct cftype *cft, 264 struct file *file, char __user *buf, 265 size_t nbytes, loff_t *ppos) 266 { 267 u64 val; 268 char str[64]; 269 int idx, name, len; 270 struct hugetlb_cgroup *h_cg = hugetlb_cgroup_from_cgroup(cgroup); 271 272 idx = MEMFILE_IDX(cft->private); 273 name = MEMFILE_ATTR(cft->private); 274 275 val = res_counter_read_u64(&h_cg->hugepage[idx], name); 276 len = scnprintf(str, sizeof(str), "%llu\n", (unsigned long long)val); 277 return simple_read_from_buffer(buf, nbytes, ppos, str, len); 278 } 279 280 static int hugetlb_cgroup_write(struct cgroup *cgroup, struct cftype *cft, 281 const char *buffer) 282 { 283 int idx, name, ret; 284 unsigned long long val; 285 struct hugetlb_cgroup *h_cg = hugetlb_cgroup_from_cgroup(cgroup); 286 287 idx = MEMFILE_IDX(cft->private); 288 name = MEMFILE_ATTR(cft->private); 289 290 switch (name) { 291 case RES_LIMIT: 292 if (hugetlb_cgroup_is_root(h_cg)) { 293 /* Can't set limit on root */ 294 ret = -EINVAL; 295 break; 296 } 297 /* This function does all necessary parse...reuse it */ 298 ret = res_counter_memparse_write_strategy(buffer, &val); 299 if (ret) 300 break; 301 ret = res_counter_set_limit(&h_cg->hugepage[idx], val); 302 break; 303 default: 304 ret = -EINVAL; 305 break; 306 } 307 return ret; 308 } 309 310 static int hugetlb_cgroup_reset(struct cgroup *cgroup, unsigned int event) 311 { 312 int idx, name, ret = 0; 313 struct hugetlb_cgroup *h_cg = hugetlb_cgroup_from_cgroup(cgroup); 314 315 idx = MEMFILE_IDX(event); 316 name = MEMFILE_ATTR(event); 317 318 switch (name) { 319 case RES_MAX_USAGE: 320 res_counter_reset_max(&h_cg->hugepage[idx]); 321 break; 322 case RES_FAILCNT: 323 res_counter_reset_failcnt(&h_cg->hugepage[idx]); 324 break; 325 default: 326 ret = -EINVAL; 327 break; 328 } 329 return ret; 330 } 331 332 static char *mem_fmt(char *buf, int size, unsigned long hsize) 333 { 334 if (hsize >= (1UL << 30)) 335 snprintf(buf, size, "%luGB", hsize >> 30); 336 else if (hsize >= (1UL << 20)) 337 snprintf(buf, size, "%luMB", hsize >> 20); 338 else 339 snprintf(buf, size, "%luKB", hsize >> 10); 340 return buf; 341 } 342 343 int __init hugetlb_cgroup_file_init(int idx) 344 { 345 char buf[32]; 346 struct cftype *cft; 347 struct hstate *h = &hstates[idx]; 348 349 /* format the size */ 350 mem_fmt(buf, 32, huge_page_size(h)); 351 352 /* Add the limit file */ 353 cft = &h->cgroup_files[0]; 354 snprintf(cft->name, MAX_CFTYPE_NAME, "%s.limit_in_bytes", buf); 355 cft->private = MEMFILE_PRIVATE(idx, RES_LIMIT); 356 cft->read = hugetlb_cgroup_read; 357 cft->write_string = hugetlb_cgroup_write; 358 359 /* Add the usage file */ 360 cft = &h->cgroup_files[1]; 361 snprintf(cft->name, MAX_CFTYPE_NAME, "%s.usage_in_bytes", buf); 362 cft->private = MEMFILE_PRIVATE(idx, RES_USAGE); 363 cft->read = hugetlb_cgroup_read; 364 365 /* Add the MAX usage file */ 366 cft = &h->cgroup_files[2]; 367 snprintf(cft->name, MAX_CFTYPE_NAME, "%s.max_usage_in_bytes", buf); 368 cft->private = MEMFILE_PRIVATE(idx, RES_MAX_USAGE); 369 cft->trigger = hugetlb_cgroup_reset; 370 cft->read = hugetlb_cgroup_read; 371 372 /* Add the failcntfile */ 373 cft = &h->cgroup_files[3]; 374 snprintf(cft->name, MAX_CFTYPE_NAME, "%s.failcnt", buf); 375 cft->private = MEMFILE_PRIVATE(idx, RES_FAILCNT); 376 cft->trigger = hugetlb_cgroup_reset; 377 cft->read = hugetlb_cgroup_read; 378 379 /* NULL terminate the last cft */ 380 cft = &h->cgroup_files[4]; 381 memset(cft, 0, sizeof(*cft)); 382 383 WARN_ON(cgroup_add_cftypes(&hugetlb_subsys, h->cgroup_files)); 384 385 return 0; 386 } 387 388 /* 389 * hugetlb_lock will make sure a parallel cgroup rmdir won't happen 390 * when we migrate hugepages 391 */ 392 void hugetlb_cgroup_migrate(struct page *oldhpage, struct page *newhpage) 393 { 394 struct hugetlb_cgroup *h_cg; 395 struct hstate *h = page_hstate(oldhpage); 396 397 if (hugetlb_cgroup_disabled()) 398 return; 399 400 VM_BUG_ON(!PageHuge(oldhpage)); 401 spin_lock(&hugetlb_lock); 402 h_cg = hugetlb_cgroup_from_page(oldhpage); 403 set_hugetlb_cgroup(oldhpage, NULL); 404 405 /* move the h_cg details to new cgroup */ 406 set_hugetlb_cgroup(newhpage, h_cg); 407 list_move(&newhpage->lru, &h->hugepage_activelist); 408 spin_unlock(&hugetlb_lock); 409 return; 410 } 411 412 struct cgroup_subsys hugetlb_subsys = { 413 .name = "hugetlb", 414 .create = hugetlb_cgroup_create, 415 .pre_destroy = hugetlb_cgroup_pre_destroy, 416 .destroy = hugetlb_cgroup_destroy, 417 .subsys_id = hugetlb_subsys_id, 418 }; 419