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 static struct hugetlb_cgroup *root_h_cgroup __read_mostly; 34 35 static inline 36 struct hugetlb_cgroup *hugetlb_cgroup_from_css(struct cgroup_subsys_state *s) 37 { 38 return s ? container_of(s, struct hugetlb_cgroup, css) : NULL; 39 } 40 41 static inline 42 struct hugetlb_cgroup *hugetlb_cgroup_from_task(struct task_struct *task) 43 { 44 return hugetlb_cgroup_from_css(task_css(task, hugetlb_cgrp_id)); 45 } 46 47 static inline bool hugetlb_cgroup_is_root(struct hugetlb_cgroup *h_cg) 48 { 49 return (h_cg == root_h_cgroup); 50 } 51 52 static inline struct hugetlb_cgroup * 53 parent_hugetlb_cgroup(struct hugetlb_cgroup *h_cg) 54 { 55 return hugetlb_cgroup_from_css(h_cg->css.parent); 56 } 57 58 static inline bool hugetlb_cgroup_have_usage(struct hugetlb_cgroup *h_cg) 59 { 60 int idx; 61 62 for (idx = 0; idx < hugetlb_max_hstate; idx++) { 63 if ((res_counter_read_u64(&h_cg->hugepage[idx], RES_USAGE)) > 0) 64 return true; 65 } 66 return false; 67 } 68 69 static struct cgroup_subsys_state * 70 hugetlb_cgroup_css_alloc(struct cgroup_subsys_state *parent_css) 71 { 72 struct hugetlb_cgroup *parent_h_cgroup = hugetlb_cgroup_from_css(parent_css); 73 struct hugetlb_cgroup *h_cgroup; 74 int idx; 75 76 h_cgroup = kzalloc(sizeof(*h_cgroup), GFP_KERNEL); 77 if (!h_cgroup) 78 return ERR_PTR(-ENOMEM); 79 80 if (parent_h_cgroup) { 81 for (idx = 0; idx < HUGE_MAX_HSTATE; idx++) 82 res_counter_init(&h_cgroup->hugepage[idx], 83 &parent_h_cgroup->hugepage[idx]); 84 } else { 85 root_h_cgroup = h_cgroup; 86 for (idx = 0; idx < HUGE_MAX_HSTATE; idx++) 87 res_counter_init(&h_cgroup->hugepage[idx], NULL); 88 } 89 return &h_cgroup->css; 90 } 91 92 static void hugetlb_cgroup_css_free(struct cgroup_subsys_state *css) 93 { 94 struct hugetlb_cgroup *h_cgroup; 95 96 h_cgroup = hugetlb_cgroup_from_css(css); 97 kfree(h_cgroup); 98 } 99 100 101 /* 102 * Should be called with hugetlb_lock held. 103 * Since we are holding hugetlb_lock, pages cannot get moved from 104 * active list or uncharged from the cgroup, So no need to get 105 * page reference and test for page active here. This function 106 * cannot fail. 107 */ 108 static void hugetlb_cgroup_move_parent(int idx, struct hugetlb_cgroup *h_cg, 109 struct page *page) 110 { 111 int csize; 112 struct res_counter *counter; 113 struct res_counter *fail_res; 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 csize = PAGE_SIZE << compound_order(page); 127 if (!parent) { 128 parent = root_h_cgroup; 129 /* root has no limit */ 130 res_counter_charge_nofail(&parent->hugepage[idx], 131 csize, &fail_res); 132 } 133 counter = &h_cg->hugepage[idx]; 134 res_counter_uncharge_until(counter, counter->parent, csize); 135 136 set_hugetlb_cgroup(page, parent); 137 out: 138 return; 139 } 140 141 /* 142 * Force the hugetlb cgroup to empty the hugetlb resources by moving them to 143 * the parent cgroup. 144 */ 145 static 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 } 164 165 int hugetlb_cgroup_charge_cgroup(int idx, unsigned long nr_pages, 166 struct hugetlb_cgroup **ptr) 167 { 168 int ret = 0; 169 struct res_counter *fail_res; 170 struct hugetlb_cgroup *h_cg = NULL; 171 unsigned long csize = nr_pages * PAGE_SIZE; 172 173 if (hugetlb_cgroup_disabled()) 174 goto done; 175 /* 176 * We don't charge any cgroup if the compound page have less 177 * than 3 pages. 178 */ 179 if (huge_page_order(&hstates[idx]) < HUGETLB_CGROUP_MIN_ORDER) 180 goto done; 181 again: 182 rcu_read_lock(); 183 h_cg = hugetlb_cgroup_from_task(current); 184 if (!css_tryget_online(&h_cg->css)) { 185 rcu_read_unlock(); 186 goto again; 187 } 188 rcu_read_unlock(); 189 190 ret = res_counter_charge(&h_cg->hugepage[idx], csize, &fail_res); 191 css_put(&h_cg->css); 192 done: 193 *ptr = h_cg; 194 return ret; 195 } 196 197 /* Should be called with hugetlb_lock held */ 198 void hugetlb_cgroup_commit_charge(int idx, unsigned long nr_pages, 199 struct hugetlb_cgroup *h_cg, 200 struct page *page) 201 { 202 if (hugetlb_cgroup_disabled() || !h_cg) 203 return; 204 205 set_hugetlb_cgroup(page, h_cg); 206 return; 207 } 208 209 /* 210 * Should be called with hugetlb_lock held 211 */ 212 void hugetlb_cgroup_uncharge_page(int idx, unsigned long nr_pages, 213 struct page *page) 214 { 215 struct hugetlb_cgroup *h_cg; 216 unsigned long csize = nr_pages * PAGE_SIZE; 217 218 if (hugetlb_cgroup_disabled()) 219 return; 220 lockdep_assert_held(&hugetlb_lock); 221 h_cg = hugetlb_cgroup_from_page(page); 222 if (unlikely(!h_cg)) 223 return; 224 set_hugetlb_cgroup(page, NULL); 225 res_counter_uncharge(&h_cg->hugepage[idx], csize); 226 return; 227 } 228 229 void hugetlb_cgroup_uncharge_cgroup(int idx, unsigned long nr_pages, 230 struct hugetlb_cgroup *h_cg) 231 { 232 unsigned long csize = nr_pages * PAGE_SIZE; 233 234 if (hugetlb_cgroup_disabled() || !h_cg) 235 return; 236 237 if (huge_page_order(&hstates[idx]) < HUGETLB_CGROUP_MIN_ORDER) 238 return; 239 240 res_counter_uncharge(&h_cg->hugepage[idx], csize); 241 return; 242 } 243 244 static u64 hugetlb_cgroup_read_u64(struct cgroup_subsys_state *css, 245 struct cftype *cft) 246 { 247 int idx, name; 248 struct hugetlb_cgroup *h_cg = hugetlb_cgroup_from_css(css); 249 250 idx = MEMFILE_IDX(cft->private); 251 name = MEMFILE_ATTR(cft->private); 252 253 return res_counter_read_u64(&h_cg->hugepage[idx], name); 254 } 255 256 static ssize_t hugetlb_cgroup_write(struct kernfs_open_file *of, 257 char *buf, size_t nbytes, loff_t off) 258 { 259 int idx, name, ret; 260 unsigned long long val; 261 struct hugetlb_cgroup *h_cg = hugetlb_cgroup_from_css(of_css(of)); 262 263 buf = strstrip(buf); 264 idx = MEMFILE_IDX(of_cft(of)->private); 265 name = MEMFILE_ATTR(of_cft(of)->private); 266 267 switch (name) { 268 case RES_LIMIT: 269 if (hugetlb_cgroup_is_root(h_cg)) { 270 /* Can't set limit on root */ 271 ret = -EINVAL; 272 break; 273 } 274 /* This function does all necessary parse...reuse it */ 275 ret = res_counter_memparse_write_strategy(buf, &val); 276 if (ret) 277 break; 278 val = ALIGN(val, 1ULL << huge_page_shift(&hstates[idx])); 279 ret = res_counter_set_limit(&h_cg->hugepage[idx], val); 280 break; 281 default: 282 ret = -EINVAL; 283 break; 284 } 285 return ret ?: nbytes; 286 } 287 288 static ssize_t hugetlb_cgroup_reset(struct kernfs_open_file *of, 289 char *buf, size_t nbytes, loff_t off) 290 { 291 int idx, name, ret = 0; 292 struct hugetlb_cgroup *h_cg = hugetlb_cgroup_from_css(of_css(of)); 293 294 idx = MEMFILE_IDX(of_cft(of)->private); 295 name = MEMFILE_ATTR(of_cft(of)->private); 296 297 switch (name) { 298 case RES_MAX_USAGE: 299 res_counter_reset_max(&h_cg->hugepage[idx]); 300 break; 301 case RES_FAILCNT: 302 res_counter_reset_failcnt(&h_cg->hugepage[idx]); 303 break; 304 default: 305 ret = -EINVAL; 306 break; 307 } 308 return ret ?: nbytes; 309 } 310 311 static char *mem_fmt(char *buf, int size, unsigned long hsize) 312 { 313 if (hsize >= (1UL << 30)) 314 snprintf(buf, size, "%luGB", hsize >> 30); 315 else if (hsize >= (1UL << 20)) 316 snprintf(buf, size, "%luMB", hsize >> 20); 317 else 318 snprintf(buf, size, "%luKB", hsize >> 10); 319 return buf; 320 } 321 322 static void __init __hugetlb_cgroup_file_init(int idx) 323 { 324 char buf[32]; 325 struct cftype *cft; 326 struct hstate *h = &hstates[idx]; 327 328 /* format the size */ 329 mem_fmt(buf, 32, huge_page_size(h)); 330 331 /* Add the limit file */ 332 cft = &h->cgroup_files[0]; 333 snprintf(cft->name, MAX_CFTYPE_NAME, "%s.limit_in_bytes", buf); 334 cft->private = MEMFILE_PRIVATE(idx, RES_LIMIT); 335 cft->read_u64 = hugetlb_cgroup_read_u64; 336 cft->write = hugetlb_cgroup_write; 337 338 /* Add the usage file */ 339 cft = &h->cgroup_files[1]; 340 snprintf(cft->name, MAX_CFTYPE_NAME, "%s.usage_in_bytes", buf); 341 cft->private = MEMFILE_PRIVATE(idx, RES_USAGE); 342 cft->read_u64 = hugetlb_cgroup_read_u64; 343 344 /* Add the MAX usage file */ 345 cft = &h->cgroup_files[2]; 346 snprintf(cft->name, MAX_CFTYPE_NAME, "%s.max_usage_in_bytes", buf); 347 cft->private = MEMFILE_PRIVATE(idx, RES_MAX_USAGE); 348 cft->write = hugetlb_cgroup_reset; 349 cft->read_u64 = hugetlb_cgroup_read_u64; 350 351 /* Add the failcntfile */ 352 cft = &h->cgroup_files[3]; 353 snprintf(cft->name, MAX_CFTYPE_NAME, "%s.failcnt", buf); 354 cft->private = MEMFILE_PRIVATE(idx, RES_FAILCNT); 355 cft->write = hugetlb_cgroup_reset; 356 cft->read_u64 = hugetlb_cgroup_read_u64; 357 358 /* NULL terminate the last cft */ 359 cft = &h->cgroup_files[4]; 360 memset(cft, 0, sizeof(*cft)); 361 362 WARN_ON(cgroup_add_legacy_cftypes(&hugetlb_cgrp_subsys, 363 h->cgroup_files)); 364 } 365 366 void __init hugetlb_cgroup_file_init(void) 367 { 368 struct hstate *h; 369 370 for_each_hstate(h) { 371 /* 372 * Add cgroup control files only if the huge page consists 373 * of more than two normal pages. This is because we use 374 * page[2].lru.next for storing cgroup details. 375 */ 376 if (huge_page_order(h) >= HUGETLB_CGROUP_MIN_ORDER) 377 __hugetlb_cgroup_file_init(hstate_index(h)); 378 } 379 } 380 381 /* 382 * hugetlb_lock will make sure a parallel cgroup rmdir won't happen 383 * when we migrate hugepages 384 */ 385 void hugetlb_cgroup_migrate(struct page *oldhpage, struct page *newhpage) 386 { 387 struct hugetlb_cgroup *h_cg; 388 struct hstate *h = page_hstate(oldhpage); 389 390 if (hugetlb_cgroup_disabled()) 391 return; 392 393 VM_BUG_ON_PAGE(!PageHuge(oldhpage), oldhpage); 394 spin_lock(&hugetlb_lock); 395 h_cg = hugetlb_cgroup_from_page(oldhpage); 396 set_hugetlb_cgroup(oldhpage, NULL); 397 398 /* move the h_cg details to new cgroup */ 399 set_hugetlb_cgroup(newhpage, h_cg); 400 list_move(&newhpage->lru, &h->hugepage_activelist); 401 spin_unlock(&hugetlb_lock); 402 return; 403 } 404 405 struct cgroup_subsys hugetlb_cgrp_subsys = { 406 .css_alloc = hugetlb_cgroup_css_alloc, 407 .css_offline = hugetlb_cgroup_css_offline, 408 .css_free = hugetlb_cgroup_css_free, 409 }; 410