1 /* 2 * Generic process-grouping system. 3 * 4 * Based originally on the cpuset system, extracted by Paul Menage 5 * Copyright (C) 2006 Google, Inc 6 * 7 * Notifications support 8 * Copyright (C) 2009 Nokia Corporation 9 * Author: Kirill A. Shutemov 10 * 11 * Copyright notices from the original cpuset code: 12 * -------------------------------------------------- 13 * Copyright (C) 2003 BULL SA. 14 * Copyright (C) 2004-2006 Silicon Graphics, Inc. 15 * 16 * Portions derived from Patrick Mochel's sysfs code. 17 * sysfs is Copyright (c) 2001-3 Patrick Mochel 18 * 19 * 2003-10-10 Written by Simon Derr. 20 * 2003-10-22 Updates by Stephen Hemminger. 21 * 2004 May-July Rework by Paul Jackson. 22 * --------------------------------------------------- 23 * 24 * This file is subject to the terms and conditions of the GNU General Public 25 * License. See the file COPYING in the main directory of the Linux 26 * distribution for more details. 27 */ 28 29 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt 30 31 #include "cgroup-internal.h" 32 33 #include <linux/cred.h> 34 #include <linux/errno.h> 35 #include <linux/init_task.h> 36 #include <linux/kernel.h> 37 #include <linux/magic.h> 38 #include <linux/mutex.h> 39 #include <linux/mount.h> 40 #include <linux/pagemap.h> 41 #include <linux/proc_fs.h> 42 #include <linux/rcupdate.h> 43 #include <linux/sched.h> 44 #include <linux/sched/task.h> 45 #include <linux/slab.h> 46 #include <linux/spinlock.h> 47 #include <linux/percpu-rwsem.h> 48 #include <linux/string.h> 49 #include <linux/hashtable.h> 50 #include <linux/idr.h> 51 #include <linux/kthread.h> 52 #include <linux/atomic.h> 53 #include <linux/cpuset.h> 54 #include <linux/proc_ns.h> 55 #include <linux/nsproxy.h> 56 #include <linux/file.h> 57 #include <linux/fs_parser.h> 58 #include <linux/sched/cputime.h> 59 #include <linux/psi.h> 60 #include <net/sock.h> 61 62 #define CREATE_TRACE_POINTS 63 #include <trace/events/cgroup.h> 64 65 #define CGROUP_FILE_NAME_MAX (MAX_CGROUP_TYPE_NAMELEN + \ 66 MAX_CFTYPE_NAME + 2) 67 /* let's not notify more than 100 times per second */ 68 #define CGROUP_FILE_NOTIFY_MIN_INTV DIV_ROUND_UP(HZ, 100) 69 70 /* 71 * cgroup_mutex is the master lock. Any modification to cgroup or its 72 * hierarchy must be performed while holding it. 73 * 74 * css_set_lock protects task->cgroups pointer, the list of css_set 75 * objects, and the chain of tasks off each css_set. 76 * 77 * These locks are exported if CONFIG_PROVE_RCU so that accessors in 78 * cgroup.h can use them for lockdep annotations. 79 */ 80 DEFINE_MUTEX(cgroup_mutex); 81 DEFINE_SPINLOCK(css_set_lock); 82 83 #ifdef CONFIG_PROVE_RCU 84 EXPORT_SYMBOL_GPL(cgroup_mutex); 85 EXPORT_SYMBOL_GPL(css_set_lock); 86 #endif 87 88 DEFINE_SPINLOCK(trace_cgroup_path_lock); 89 char trace_cgroup_path[TRACE_CGROUP_PATH_LEN]; 90 bool cgroup_debug __read_mostly; 91 92 /* 93 * Protects cgroup_idr and css_idr so that IDs can be released without 94 * grabbing cgroup_mutex. 95 */ 96 static DEFINE_SPINLOCK(cgroup_idr_lock); 97 98 /* 99 * Protects cgroup_file->kn for !self csses. It synchronizes notifications 100 * against file removal/re-creation across css hiding. 101 */ 102 static DEFINE_SPINLOCK(cgroup_file_kn_lock); 103 104 DEFINE_PERCPU_RWSEM(cgroup_threadgroup_rwsem); 105 106 #define cgroup_assert_mutex_or_rcu_locked() \ 107 RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \ 108 !lockdep_is_held(&cgroup_mutex), \ 109 "cgroup_mutex or RCU read lock required"); 110 111 /* 112 * cgroup destruction makes heavy use of work items and there can be a lot 113 * of concurrent destructions. Use a separate workqueue so that cgroup 114 * destruction work items don't end up filling up max_active of system_wq 115 * which may lead to deadlock. 116 */ 117 static struct workqueue_struct *cgroup_destroy_wq; 118 119 /* generate an array of cgroup subsystem pointers */ 120 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys, 121 struct cgroup_subsys *cgroup_subsys[] = { 122 #include <linux/cgroup_subsys.h> 123 }; 124 #undef SUBSYS 125 126 /* array of cgroup subsystem names */ 127 #define SUBSYS(_x) [_x ## _cgrp_id] = #_x, 128 static const char *cgroup_subsys_name[] = { 129 #include <linux/cgroup_subsys.h> 130 }; 131 #undef SUBSYS 132 133 /* array of static_keys for cgroup_subsys_enabled() and cgroup_subsys_on_dfl() */ 134 #define SUBSYS(_x) \ 135 DEFINE_STATIC_KEY_TRUE(_x ## _cgrp_subsys_enabled_key); \ 136 DEFINE_STATIC_KEY_TRUE(_x ## _cgrp_subsys_on_dfl_key); \ 137 EXPORT_SYMBOL_GPL(_x ## _cgrp_subsys_enabled_key); \ 138 EXPORT_SYMBOL_GPL(_x ## _cgrp_subsys_on_dfl_key); 139 #include <linux/cgroup_subsys.h> 140 #undef SUBSYS 141 142 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys_enabled_key, 143 static struct static_key_true *cgroup_subsys_enabled_key[] = { 144 #include <linux/cgroup_subsys.h> 145 }; 146 #undef SUBSYS 147 148 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys_on_dfl_key, 149 static struct static_key_true *cgroup_subsys_on_dfl_key[] = { 150 #include <linux/cgroup_subsys.h> 151 }; 152 #undef SUBSYS 153 154 static DEFINE_PER_CPU(struct cgroup_rstat_cpu, cgrp_dfl_root_rstat_cpu); 155 156 /* the default hierarchy */ 157 struct cgroup_root cgrp_dfl_root = { .cgrp.rstat_cpu = &cgrp_dfl_root_rstat_cpu }; 158 EXPORT_SYMBOL_GPL(cgrp_dfl_root); 159 160 /* 161 * The default hierarchy always exists but is hidden until mounted for the 162 * first time. This is for backward compatibility. 163 */ 164 static bool cgrp_dfl_visible; 165 166 /* some controllers are not supported in the default hierarchy */ 167 static u16 cgrp_dfl_inhibit_ss_mask; 168 169 /* some controllers are implicitly enabled on the default hierarchy */ 170 static u16 cgrp_dfl_implicit_ss_mask; 171 172 /* some controllers can be threaded on the default hierarchy */ 173 static u16 cgrp_dfl_threaded_ss_mask; 174 175 /* The list of hierarchy roots */ 176 LIST_HEAD(cgroup_roots); 177 static int cgroup_root_count; 178 179 /* hierarchy ID allocation and mapping, protected by cgroup_mutex */ 180 static DEFINE_IDR(cgroup_hierarchy_idr); 181 182 /* 183 * Assign a monotonically increasing serial number to csses. It guarantees 184 * cgroups with bigger numbers are newer than those with smaller numbers. 185 * Also, as csses are always appended to the parent's ->children list, it 186 * guarantees that sibling csses are always sorted in the ascending serial 187 * number order on the list. Protected by cgroup_mutex. 188 */ 189 static u64 css_serial_nr_next = 1; 190 191 /* 192 * These bitmasks identify subsystems with specific features to avoid 193 * having to do iterative checks repeatedly. 194 */ 195 static u16 have_fork_callback __read_mostly; 196 static u16 have_exit_callback __read_mostly; 197 static u16 have_release_callback __read_mostly; 198 static u16 have_canfork_callback __read_mostly; 199 200 /* cgroup namespace for init task */ 201 struct cgroup_namespace init_cgroup_ns = { 202 .ns.count = REFCOUNT_INIT(2), 203 .user_ns = &init_user_ns, 204 .ns.ops = &cgroupns_operations, 205 .ns.inum = PROC_CGROUP_INIT_INO, 206 .root_cset = &init_css_set, 207 }; 208 209 static struct file_system_type cgroup2_fs_type; 210 static struct cftype cgroup_base_files[]; 211 212 static int cgroup_apply_control(struct cgroup *cgrp); 213 static void cgroup_finalize_control(struct cgroup *cgrp, int ret); 214 static void css_task_iter_skip(struct css_task_iter *it, 215 struct task_struct *task); 216 static int cgroup_destroy_locked(struct cgroup *cgrp); 217 static struct cgroup_subsys_state *css_create(struct cgroup *cgrp, 218 struct cgroup_subsys *ss); 219 static void css_release(struct percpu_ref *ref); 220 static void kill_css(struct cgroup_subsys_state *css); 221 static int cgroup_addrm_files(struct cgroup_subsys_state *css, 222 struct cgroup *cgrp, struct cftype cfts[], 223 bool is_add); 224 225 /** 226 * cgroup_ssid_enabled - cgroup subsys enabled test by subsys ID 227 * @ssid: subsys ID of interest 228 * 229 * cgroup_subsys_enabled() can only be used with literal subsys names which 230 * is fine for individual subsystems but unsuitable for cgroup core. This 231 * is slower static_key_enabled() based test indexed by @ssid. 232 */ 233 bool cgroup_ssid_enabled(int ssid) 234 { 235 if (CGROUP_SUBSYS_COUNT == 0) 236 return false; 237 238 return static_key_enabled(cgroup_subsys_enabled_key[ssid]); 239 } 240 241 /** 242 * cgroup_on_dfl - test whether a cgroup is on the default hierarchy 243 * @cgrp: the cgroup of interest 244 * 245 * The default hierarchy is the v2 interface of cgroup and this function 246 * can be used to test whether a cgroup is on the default hierarchy for 247 * cases where a subsystem should behave differently depending on the 248 * interface version. 249 * 250 * List of changed behaviors: 251 * 252 * - Mount options "noprefix", "xattr", "clone_children", "release_agent" 253 * and "name" are disallowed. 254 * 255 * - When mounting an existing superblock, mount options should match. 256 * 257 * - Remount is disallowed. 258 * 259 * - rename(2) is disallowed. 260 * 261 * - "tasks" is removed. Everything should be at process granularity. Use 262 * "cgroup.procs" instead. 263 * 264 * - "cgroup.procs" is not sorted. pids will be unique unless they got 265 * recycled in-between reads. 266 * 267 * - "release_agent" and "notify_on_release" are removed. Replacement 268 * notification mechanism will be implemented. 269 * 270 * - "cgroup.clone_children" is removed. 271 * 272 * - "cgroup.subtree_populated" is available. Its value is 0 if the cgroup 273 * and its descendants contain no task; otherwise, 1. The file also 274 * generates kernfs notification which can be monitored through poll and 275 * [di]notify when the value of the file changes. 276 * 277 * - cpuset: tasks will be kept in empty cpusets when hotplug happens and 278 * take masks of ancestors with non-empty cpus/mems, instead of being 279 * moved to an ancestor. 280 * 281 * - cpuset: a task can be moved into an empty cpuset, and again it takes 282 * masks of ancestors. 283 * 284 * - blkcg: blk-throttle becomes properly hierarchical. 285 * 286 * - debug: disallowed on the default hierarchy. 287 */ 288 bool cgroup_on_dfl(const struct cgroup *cgrp) 289 { 290 return cgrp->root == &cgrp_dfl_root; 291 } 292 293 /* IDR wrappers which synchronize using cgroup_idr_lock */ 294 static int cgroup_idr_alloc(struct idr *idr, void *ptr, int start, int end, 295 gfp_t gfp_mask) 296 { 297 int ret; 298 299 idr_preload(gfp_mask); 300 spin_lock_bh(&cgroup_idr_lock); 301 ret = idr_alloc(idr, ptr, start, end, gfp_mask & ~__GFP_DIRECT_RECLAIM); 302 spin_unlock_bh(&cgroup_idr_lock); 303 idr_preload_end(); 304 return ret; 305 } 306 307 static void *cgroup_idr_replace(struct idr *idr, void *ptr, int id) 308 { 309 void *ret; 310 311 spin_lock_bh(&cgroup_idr_lock); 312 ret = idr_replace(idr, ptr, id); 313 spin_unlock_bh(&cgroup_idr_lock); 314 return ret; 315 } 316 317 static void cgroup_idr_remove(struct idr *idr, int id) 318 { 319 spin_lock_bh(&cgroup_idr_lock); 320 idr_remove(idr, id); 321 spin_unlock_bh(&cgroup_idr_lock); 322 } 323 324 static bool cgroup_has_tasks(struct cgroup *cgrp) 325 { 326 return cgrp->nr_populated_csets; 327 } 328 329 bool cgroup_is_threaded(struct cgroup *cgrp) 330 { 331 return cgrp->dom_cgrp != cgrp; 332 } 333 334 /* can @cgrp host both domain and threaded children? */ 335 static bool cgroup_is_mixable(struct cgroup *cgrp) 336 { 337 /* 338 * Root isn't under domain level resource control exempting it from 339 * the no-internal-process constraint, so it can serve as a thread 340 * root and a parent of resource domains at the same time. 341 */ 342 return !cgroup_parent(cgrp); 343 } 344 345 /* can @cgrp become a thread root? Should always be true for a thread root */ 346 static bool cgroup_can_be_thread_root(struct cgroup *cgrp) 347 { 348 /* mixables don't care */ 349 if (cgroup_is_mixable(cgrp)) 350 return true; 351 352 /* domain roots can't be nested under threaded */ 353 if (cgroup_is_threaded(cgrp)) 354 return false; 355 356 /* can only have either domain or threaded children */ 357 if (cgrp->nr_populated_domain_children) 358 return false; 359 360 /* and no domain controllers can be enabled */ 361 if (cgrp->subtree_control & ~cgrp_dfl_threaded_ss_mask) 362 return false; 363 364 return true; 365 } 366 367 /* is @cgrp root of a threaded subtree? */ 368 bool cgroup_is_thread_root(struct cgroup *cgrp) 369 { 370 /* thread root should be a domain */ 371 if (cgroup_is_threaded(cgrp)) 372 return false; 373 374 /* a domain w/ threaded children is a thread root */ 375 if (cgrp->nr_threaded_children) 376 return true; 377 378 /* 379 * A domain which has tasks and explicit threaded controllers 380 * enabled is a thread root. 381 */ 382 if (cgroup_has_tasks(cgrp) && 383 (cgrp->subtree_control & cgrp_dfl_threaded_ss_mask)) 384 return true; 385 386 return false; 387 } 388 389 /* a domain which isn't connected to the root w/o brekage can't be used */ 390 static bool cgroup_is_valid_domain(struct cgroup *cgrp) 391 { 392 /* the cgroup itself can be a thread root */ 393 if (cgroup_is_threaded(cgrp)) 394 return false; 395 396 /* but the ancestors can't be unless mixable */ 397 while ((cgrp = cgroup_parent(cgrp))) { 398 if (!cgroup_is_mixable(cgrp) && cgroup_is_thread_root(cgrp)) 399 return false; 400 if (cgroup_is_threaded(cgrp)) 401 return false; 402 } 403 404 return true; 405 } 406 407 /* subsystems visibly enabled on a cgroup */ 408 static u16 cgroup_control(struct cgroup *cgrp) 409 { 410 struct cgroup *parent = cgroup_parent(cgrp); 411 u16 root_ss_mask = cgrp->root->subsys_mask; 412 413 if (parent) { 414 u16 ss_mask = parent->subtree_control; 415 416 /* threaded cgroups can only have threaded controllers */ 417 if (cgroup_is_threaded(cgrp)) 418 ss_mask &= cgrp_dfl_threaded_ss_mask; 419 return ss_mask; 420 } 421 422 if (cgroup_on_dfl(cgrp)) 423 root_ss_mask &= ~(cgrp_dfl_inhibit_ss_mask | 424 cgrp_dfl_implicit_ss_mask); 425 return root_ss_mask; 426 } 427 428 /* subsystems enabled on a cgroup */ 429 static u16 cgroup_ss_mask(struct cgroup *cgrp) 430 { 431 struct cgroup *parent = cgroup_parent(cgrp); 432 433 if (parent) { 434 u16 ss_mask = parent->subtree_ss_mask; 435 436 /* threaded cgroups can only have threaded controllers */ 437 if (cgroup_is_threaded(cgrp)) 438 ss_mask &= cgrp_dfl_threaded_ss_mask; 439 return ss_mask; 440 } 441 442 return cgrp->root->subsys_mask; 443 } 444 445 /** 446 * cgroup_css - obtain a cgroup's css for the specified subsystem 447 * @cgrp: the cgroup of interest 448 * @ss: the subsystem of interest (%NULL returns @cgrp->self) 449 * 450 * Return @cgrp's css (cgroup_subsys_state) associated with @ss. This 451 * function must be called either under cgroup_mutex or rcu_read_lock() and 452 * the caller is responsible for pinning the returned css if it wants to 453 * keep accessing it outside the said locks. This function may return 454 * %NULL if @cgrp doesn't have @subsys_id enabled. 455 */ 456 static struct cgroup_subsys_state *cgroup_css(struct cgroup *cgrp, 457 struct cgroup_subsys *ss) 458 { 459 if (ss) 460 return rcu_dereference_check(cgrp->subsys[ss->id], 461 lockdep_is_held(&cgroup_mutex)); 462 else 463 return &cgrp->self; 464 } 465 466 /** 467 * cgroup_tryget_css - try to get a cgroup's css for the specified subsystem 468 * @cgrp: the cgroup of interest 469 * @ss: the subsystem of interest 470 * 471 * Find and get @cgrp's css assocaited with @ss. If the css doesn't exist 472 * or is offline, %NULL is returned. 473 */ 474 static struct cgroup_subsys_state *cgroup_tryget_css(struct cgroup *cgrp, 475 struct cgroup_subsys *ss) 476 { 477 struct cgroup_subsys_state *css; 478 479 rcu_read_lock(); 480 css = cgroup_css(cgrp, ss); 481 if (css && !css_tryget_online(css)) 482 css = NULL; 483 rcu_read_unlock(); 484 485 return css; 486 } 487 488 /** 489 * cgroup_e_css_by_mask - obtain a cgroup's effective css for the specified ss 490 * @cgrp: the cgroup of interest 491 * @ss: the subsystem of interest (%NULL returns @cgrp->self) 492 * 493 * Similar to cgroup_css() but returns the effective css, which is defined 494 * as the matching css of the nearest ancestor including self which has @ss 495 * enabled. If @ss is associated with the hierarchy @cgrp is on, this 496 * function is guaranteed to return non-NULL css. 497 */ 498 static struct cgroup_subsys_state *cgroup_e_css_by_mask(struct cgroup *cgrp, 499 struct cgroup_subsys *ss) 500 { 501 lockdep_assert_held(&cgroup_mutex); 502 503 if (!ss) 504 return &cgrp->self; 505 506 /* 507 * This function is used while updating css associations and thus 508 * can't test the csses directly. Test ss_mask. 509 */ 510 while (!(cgroup_ss_mask(cgrp) & (1 << ss->id))) { 511 cgrp = cgroup_parent(cgrp); 512 if (!cgrp) 513 return NULL; 514 } 515 516 return cgroup_css(cgrp, ss); 517 } 518 519 /** 520 * cgroup_e_css - obtain a cgroup's effective css for the specified subsystem 521 * @cgrp: the cgroup of interest 522 * @ss: the subsystem of interest 523 * 524 * Find and get the effective css of @cgrp for @ss. The effective css is 525 * defined as the matching css of the nearest ancestor including self which 526 * has @ss enabled. If @ss is not mounted on the hierarchy @cgrp is on, 527 * the root css is returned, so this function always returns a valid css. 528 * 529 * The returned css is not guaranteed to be online, and therefore it is the 530 * callers responsibility to try get a reference for it. 531 */ 532 struct cgroup_subsys_state *cgroup_e_css(struct cgroup *cgrp, 533 struct cgroup_subsys *ss) 534 { 535 struct cgroup_subsys_state *css; 536 537 do { 538 css = cgroup_css(cgrp, ss); 539 540 if (css) 541 return css; 542 cgrp = cgroup_parent(cgrp); 543 } while (cgrp); 544 545 return init_css_set.subsys[ss->id]; 546 } 547 548 /** 549 * cgroup_get_e_css - get a cgroup's effective css for the specified subsystem 550 * @cgrp: the cgroup of interest 551 * @ss: the subsystem of interest 552 * 553 * Find and get the effective css of @cgrp for @ss. The effective css is 554 * defined as the matching css of the nearest ancestor including self which 555 * has @ss enabled. If @ss is not mounted on the hierarchy @cgrp is on, 556 * the root css is returned, so this function always returns a valid css. 557 * The returned css must be put using css_put(). 558 */ 559 struct cgroup_subsys_state *cgroup_get_e_css(struct cgroup *cgrp, 560 struct cgroup_subsys *ss) 561 { 562 struct cgroup_subsys_state *css; 563 564 rcu_read_lock(); 565 566 do { 567 css = cgroup_css(cgrp, ss); 568 569 if (css && css_tryget_online(css)) 570 goto out_unlock; 571 cgrp = cgroup_parent(cgrp); 572 } while (cgrp); 573 574 css = init_css_set.subsys[ss->id]; 575 css_get(css); 576 out_unlock: 577 rcu_read_unlock(); 578 return css; 579 } 580 581 static void cgroup_get_live(struct cgroup *cgrp) 582 { 583 WARN_ON_ONCE(cgroup_is_dead(cgrp)); 584 css_get(&cgrp->self); 585 } 586 587 /** 588 * __cgroup_task_count - count the number of tasks in a cgroup. The caller 589 * is responsible for taking the css_set_lock. 590 * @cgrp: the cgroup in question 591 */ 592 int __cgroup_task_count(const struct cgroup *cgrp) 593 { 594 int count = 0; 595 struct cgrp_cset_link *link; 596 597 lockdep_assert_held(&css_set_lock); 598 599 list_for_each_entry(link, &cgrp->cset_links, cset_link) 600 count += link->cset->nr_tasks; 601 602 return count; 603 } 604 605 /** 606 * cgroup_task_count - count the number of tasks in a cgroup. 607 * @cgrp: the cgroup in question 608 */ 609 int cgroup_task_count(const struct cgroup *cgrp) 610 { 611 int count; 612 613 spin_lock_irq(&css_set_lock); 614 count = __cgroup_task_count(cgrp); 615 spin_unlock_irq(&css_set_lock); 616 617 return count; 618 } 619 620 struct cgroup_subsys_state *of_css(struct kernfs_open_file *of) 621 { 622 struct cgroup *cgrp = of->kn->parent->priv; 623 struct cftype *cft = of_cft(of); 624 625 /* 626 * This is open and unprotected implementation of cgroup_css(). 627 * seq_css() is only called from a kernfs file operation which has 628 * an active reference on the file. Because all the subsystem 629 * files are drained before a css is disassociated with a cgroup, 630 * the matching css from the cgroup's subsys table is guaranteed to 631 * be and stay valid until the enclosing operation is complete. 632 */ 633 if (cft->ss) 634 return rcu_dereference_raw(cgrp->subsys[cft->ss->id]); 635 else 636 return &cgrp->self; 637 } 638 EXPORT_SYMBOL_GPL(of_css); 639 640 /** 641 * for_each_css - iterate all css's of a cgroup 642 * @css: the iteration cursor 643 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end 644 * @cgrp: the target cgroup to iterate css's of 645 * 646 * Should be called under cgroup_[tree_]mutex. 647 */ 648 #define for_each_css(css, ssid, cgrp) \ 649 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \ 650 if (!((css) = rcu_dereference_check( \ 651 (cgrp)->subsys[(ssid)], \ 652 lockdep_is_held(&cgroup_mutex)))) { } \ 653 else 654 655 /** 656 * for_each_e_css - iterate all effective css's of a cgroup 657 * @css: the iteration cursor 658 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end 659 * @cgrp: the target cgroup to iterate css's of 660 * 661 * Should be called under cgroup_[tree_]mutex. 662 */ 663 #define for_each_e_css(css, ssid, cgrp) \ 664 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \ 665 if (!((css) = cgroup_e_css_by_mask(cgrp, \ 666 cgroup_subsys[(ssid)]))) \ 667 ; \ 668 else 669 670 /** 671 * do_each_subsys_mask - filter for_each_subsys with a bitmask 672 * @ss: the iteration cursor 673 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end 674 * @ss_mask: the bitmask 675 * 676 * The block will only run for cases where the ssid-th bit (1 << ssid) of 677 * @ss_mask is set. 678 */ 679 #define do_each_subsys_mask(ss, ssid, ss_mask) do { \ 680 unsigned long __ss_mask = (ss_mask); \ 681 if (!CGROUP_SUBSYS_COUNT) { /* to avoid spurious gcc warning */ \ 682 (ssid) = 0; \ 683 break; \ 684 } \ 685 for_each_set_bit(ssid, &__ss_mask, CGROUP_SUBSYS_COUNT) { \ 686 (ss) = cgroup_subsys[ssid]; \ 687 { 688 689 #define while_each_subsys_mask() \ 690 } \ 691 } \ 692 } while (false) 693 694 /* iterate over child cgrps, lock should be held throughout iteration */ 695 #define cgroup_for_each_live_child(child, cgrp) \ 696 list_for_each_entry((child), &(cgrp)->self.children, self.sibling) \ 697 if (({ lockdep_assert_held(&cgroup_mutex); \ 698 cgroup_is_dead(child); })) \ 699 ; \ 700 else 701 702 /* walk live descendants in pre order */ 703 #define cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) \ 704 css_for_each_descendant_pre((d_css), cgroup_css((cgrp), NULL)) \ 705 if (({ lockdep_assert_held(&cgroup_mutex); \ 706 (dsct) = (d_css)->cgroup; \ 707 cgroup_is_dead(dsct); })) \ 708 ; \ 709 else 710 711 /* walk live descendants in postorder */ 712 #define cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) \ 713 css_for_each_descendant_post((d_css), cgroup_css((cgrp), NULL)) \ 714 if (({ lockdep_assert_held(&cgroup_mutex); \ 715 (dsct) = (d_css)->cgroup; \ 716 cgroup_is_dead(dsct); })) \ 717 ; \ 718 else 719 720 /* 721 * The default css_set - used by init and its children prior to any 722 * hierarchies being mounted. It contains a pointer to the root state 723 * for each subsystem. Also used to anchor the list of css_sets. Not 724 * reference-counted, to improve performance when child cgroups 725 * haven't been created. 726 */ 727 struct css_set init_css_set = { 728 .refcount = REFCOUNT_INIT(1), 729 .dom_cset = &init_css_set, 730 .tasks = LIST_HEAD_INIT(init_css_set.tasks), 731 .mg_tasks = LIST_HEAD_INIT(init_css_set.mg_tasks), 732 .dying_tasks = LIST_HEAD_INIT(init_css_set.dying_tasks), 733 .task_iters = LIST_HEAD_INIT(init_css_set.task_iters), 734 .threaded_csets = LIST_HEAD_INIT(init_css_set.threaded_csets), 735 .cgrp_links = LIST_HEAD_INIT(init_css_set.cgrp_links), 736 .mg_preload_node = LIST_HEAD_INIT(init_css_set.mg_preload_node), 737 .mg_node = LIST_HEAD_INIT(init_css_set.mg_node), 738 739 /* 740 * The following field is re-initialized when this cset gets linked 741 * in cgroup_init(). However, let's initialize the field 742 * statically too so that the default cgroup can be accessed safely 743 * early during boot. 744 */ 745 .dfl_cgrp = &cgrp_dfl_root.cgrp, 746 }; 747 748 static int css_set_count = 1; /* 1 for init_css_set */ 749 750 static bool css_set_threaded(struct css_set *cset) 751 { 752 return cset->dom_cset != cset; 753 } 754 755 /** 756 * css_set_populated - does a css_set contain any tasks? 757 * @cset: target css_set 758 * 759 * css_set_populated() should be the same as !!cset->nr_tasks at steady 760 * state. However, css_set_populated() can be called while a task is being 761 * added to or removed from the linked list before the nr_tasks is 762 * properly updated. Hence, we can't just look at ->nr_tasks here. 763 */ 764 static bool css_set_populated(struct css_set *cset) 765 { 766 lockdep_assert_held(&css_set_lock); 767 768 return !list_empty(&cset->tasks) || !list_empty(&cset->mg_tasks); 769 } 770 771 /** 772 * cgroup_update_populated - update the populated count of a cgroup 773 * @cgrp: the target cgroup 774 * @populated: inc or dec populated count 775 * 776 * One of the css_sets associated with @cgrp is either getting its first 777 * task or losing the last. Update @cgrp->nr_populated_* accordingly. The 778 * count is propagated towards root so that a given cgroup's 779 * nr_populated_children is zero iff none of its descendants contain any 780 * tasks. 781 * 782 * @cgrp's interface file "cgroup.populated" is zero if both 783 * @cgrp->nr_populated_csets and @cgrp->nr_populated_children are zero and 784 * 1 otherwise. When the sum changes from or to zero, userland is notified 785 * that the content of the interface file has changed. This can be used to 786 * detect when @cgrp and its descendants become populated or empty. 787 */ 788 static void cgroup_update_populated(struct cgroup *cgrp, bool populated) 789 { 790 struct cgroup *child = NULL; 791 int adj = populated ? 1 : -1; 792 793 lockdep_assert_held(&css_set_lock); 794 795 do { 796 bool was_populated = cgroup_is_populated(cgrp); 797 798 if (!child) { 799 cgrp->nr_populated_csets += adj; 800 } else { 801 if (cgroup_is_threaded(child)) 802 cgrp->nr_populated_threaded_children += adj; 803 else 804 cgrp->nr_populated_domain_children += adj; 805 } 806 807 if (was_populated == cgroup_is_populated(cgrp)) 808 break; 809 810 cgroup1_check_for_release(cgrp); 811 TRACE_CGROUP_PATH(notify_populated, cgrp, 812 cgroup_is_populated(cgrp)); 813 cgroup_file_notify(&cgrp->events_file); 814 815 child = cgrp; 816 cgrp = cgroup_parent(cgrp); 817 } while (cgrp); 818 } 819 820 /** 821 * css_set_update_populated - update populated state of a css_set 822 * @cset: target css_set 823 * @populated: whether @cset is populated or depopulated 824 * 825 * @cset is either getting the first task or losing the last. Update the 826 * populated counters of all associated cgroups accordingly. 827 */ 828 static void css_set_update_populated(struct css_set *cset, bool populated) 829 { 830 struct cgrp_cset_link *link; 831 832 lockdep_assert_held(&css_set_lock); 833 834 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) 835 cgroup_update_populated(link->cgrp, populated); 836 } 837 838 /* 839 * @task is leaving, advance task iterators which are pointing to it so 840 * that they can resume at the next position. Advancing an iterator might 841 * remove it from the list, use safe walk. See css_task_iter_skip() for 842 * details. 843 */ 844 static void css_set_skip_task_iters(struct css_set *cset, 845 struct task_struct *task) 846 { 847 struct css_task_iter *it, *pos; 848 849 list_for_each_entry_safe(it, pos, &cset->task_iters, iters_node) 850 css_task_iter_skip(it, task); 851 } 852 853 /** 854 * css_set_move_task - move a task from one css_set to another 855 * @task: task being moved 856 * @from_cset: css_set @task currently belongs to (may be NULL) 857 * @to_cset: new css_set @task is being moved to (may be NULL) 858 * @use_mg_tasks: move to @to_cset->mg_tasks instead of ->tasks 859 * 860 * Move @task from @from_cset to @to_cset. If @task didn't belong to any 861 * css_set, @from_cset can be NULL. If @task is being disassociated 862 * instead of moved, @to_cset can be NULL. 863 * 864 * This function automatically handles populated counter updates and 865 * css_task_iter adjustments but the caller is responsible for managing 866 * @from_cset and @to_cset's reference counts. 867 */ 868 static void css_set_move_task(struct task_struct *task, 869 struct css_set *from_cset, struct css_set *to_cset, 870 bool use_mg_tasks) 871 { 872 lockdep_assert_held(&css_set_lock); 873 874 if (to_cset && !css_set_populated(to_cset)) 875 css_set_update_populated(to_cset, true); 876 877 if (from_cset) { 878 WARN_ON_ONCE(list_empty(&task->cg_list)); 879 880 css_set_skip_task_iters(from_cset, task); 881 list_del_init(&task->cg_list); 882 if (!css_set_populated(from_cset)) 883 css_set_update_populated(from_cset, false); 884 } else { 885 WARN_ON_ONCE(!list_empty(&task->cg_list)); 886 } 887 888 if (to_cset) { 889 /* 890 * We are synchronized through cgroup_threadgroup_rwsem 891 * against PF_EXITING setting such that we can't race 892 * against cgroup_exit()/cgroup_free() dropping the css_set. 893 */ 894 WARN_ON_ONCE(task->flags & PF_EXITING); 895 896 cgroup_move_task(task, to_cset); 897 list_add_tail(&task->cg_list, use_mg_tasks ? &to_cset->mg_tasks : 898 &to_cset->tasks); 899 } 900 } 901 902 /* 903 * hash table for cgroup groups. This improves the performance to find 904 * an existing css_set. This hash doesn't (currently) take into 905 * account cgroups in empty hierarchies. 906 */ 907 #define CSS_SET_HASH_BITS 7 908 static DEFINE_HASHTABLE(css_set_table, CSS_SET_HASH_BITS); 909 910 static unsigned long css_set_hash(struct cgroup_subsys_state *css[]) 911 { 912 unsigned long key = 0UL; 913 struct cgroup_subsys *ss; 914 int i; 915 916 for_each_subsys(ss, i) 917 key += (unsigned long)css[i]; 918 key = (key >> 16) ^ key; 919 920 return key; 921 } 922 923 void put_css_set_locked(struct css_set *cset) 924 { 925 struct cgrp_cset_link *link, *tmp_link; 926 struct cgroup_subsys *ss; 927 int ssid; 928 929 lockdep_assert_held(&css_set_lock); 930 931 if (!refcount_dec_and_test(&cset->refcount)) 932 return; 933 934 WARN_ON_ONCE(!list_empty(&cset->threaded_csets)); 935 936 /* This css_set is dead. Unlink it and release cgroup and css refs */ 937 for_each_subsys(ss, ssid) { 938 list_del(&cset->e_cset_node[ssid]); 939 css_put(cset->subsys[ssid]); 940 } 941 hash_del(&cset->hlist); 942 css_set_count--; 943 944 list_for_each_entry_safe(link, tmp_link, &cset->cgrp_links, cgrp_link) { 945 list_del(&link->cset_link); 946 list_del(&link->cgrp_link); 947 if (cgroup_parent(link->cgrp)) 948 cgroup_put(link->cgrp); 949 kfree(link); 950 } 951 952 if (css_set_threaded(cset)) { 953 list_del(&cset->threaded_csets_node); 954 put_css_set_locked(cset->dom_cset); 955 } 956 957 kfree_rcu(cset, rcu_head); 958 } 959 960 /** 961 * compare_css_sets - helper function for find_existing_css_set(). 962 * @cset: candidate css_set being tested 963 * @old_cset: existing css_set for a task 964 * @new_cgrp: cgroup that's being entered by the task 965 * @template: desired set of css pointers in css_set (pre-calculated) 966 * 967 * Returns true if "cset" matches "old_cset" except for the hierarchy 968 * which "new_cgrp" belongs to, for which it should match "new_cgrp". 969 */ 970 static bool compare_css_sets(struct css_set *cset, 971 struct css_set *old_cset, 972 struct cgroup *new_cgrp, 973 struct cgroup_subsys_state *template[]) 974 { 975 struct cgroup *new_dfl_cgrp; 976 struct list_head *l1, *l2; 977 978 /* 979 * On the default hierarchy, there can be csets which are 980 * associated with the same set of cgroups but different csses. 981 * Let's first ensure that csses match. 982 */ 983 if (memcmp(template, cset->subsys, sizeof(cset->subsys))) 984 return false; 985 986 987 /* @cset's domain should match the default cgroup's */ 988 if (cgroup_on_dfl(new_cgrp)) 989 new_dfl_cgrp = new_cgrp; 990 else 991 new_dfl_cgrp = old_cset->dfl_cgrp; 992 993 if (new_dfl_cgrp->dom_cgrp != cset->dom_cset->dfl_cgrp) 994 return false; 995 996 /* 997 * Compare cgroup pointers in order to distinguish between 998 * different cgroups in hierarchies. As different cgroups may 999 * share the same effective css, this comparison is always 1000 * necessary. 1001 */ 1002 l1 = &cset->cgrp_links; 1003 l2 = &old_cset->cgrp_links; 1004 while (1) { 1005 struct cgrp_cset_link *link1, *link2; 1006 struct cgroup *cgrp1, *cgrp2; 1007 1008 l1 = l1->next; 1009 l2 = l2->next; 1010 /* See if we reached the end - both lists are equal length. */ 1011 if (l1 == &cset->cgrp_links) { 1012 BUG_ON(l2 != &old_cset->cgrp_links); 1013 break; 1014 } else { 1015 BUG_ON(l2 == &old_cset->cgrp_links); 1016 } 1017 /* Locate the cgroups associated with these links. */ 1018 link1 = list_entry(l1, struct cgrp_cset_link, cgrp_link); 1019 link2 = list_entry(l2, struct cgrp_cset_link, cgrp_link); 1020 cgrp1 = link1->cgrp; 1021 cgrp2 = link2->cgrp; 1022 /* Hierarchies should be linked in the same order. */ 1023 BUG_ON(cgrp1->root != cgrp2->root); 1024 1025 /* 1026 * If this hierarchy is the hierarchy of the cgroup 1027 * that's changing, then we need to check that this 1028 * css_set points to the new cgroup; if it's any other 1029 * hierarchy, then this css_set should point to the 1030 * same cgroup as the old css_set. 1031 */ 1032 if (cgrp1->root == new_cgrp->root) { 1033 if (cgrp1 != new_cgrp) 1034 return false; 1035 } else { 1036 if (cgrp1 != cgrp2) 1037 return false; 1038 } 1039 } 1040 return true; 1041 } 1042 1043 /** 1044 * find_existing_css_set - init css array and find the matching css_set 1045 * @old_cset: the css_set that we're using before the cgroup transition 1046 * @cgrp: the cgroup that we're moving into 1047 * @template: out param for the new set of csses, should be clear on entry 1048 */ 1049 static struct css_set *find_existing_css_set(struct css_set *old_cset, 1050 struct cgroup *cgrp, 1051 struct cgroup_subsys_state *template[]) 1052 { 1053 struct cgroup_root *root = cgrp->root; 1054 struct cgroup_subsys *ss; 1055 struct css_set *cset; 1056 unsigned long key; 1057 int i; 1058 1059 /* 1060 * Build the set of subsystem state objects that we want to see in the 1061 * new css_set. While subsystems can change globally, the entries here 1062 * won't change, so no need for locking. 1063 */ 1064 for_each_subsys(ss, i) { 1065 if (root->subsys_mask & (1UL << i)) { 1066 /* 1067 * @ss is in this hierarchy, so we want the 1068 * effective css from @cgrp. 1069 */ 1070 template[i] = cgroup_e_css_by_mask(cgrp, ss); 1071 } else { 1072 /* 1073 * @ss is not in this hierarchy, so we don't want 1074 * to change the css. 1075 */ 1076 template[i] = old_cset->subsys[i]; 1077 } 1078 } 1079 1080 key = css_set_hash(template); 1081 hash_for_each_possible(css_set_table, cset, hlist, key) { 1082 if (!compare_css_sets(cset, old_cset, cgrp, template)) 1083 continue; 1084 1085 /* This css_set matches what we need */ 1086 return cset; 1087 } 1088 1089 /* No existing cgroup group matched */ 1090 return NULL; 1091 } 1092 1093 static void free_cgrp_cset_links(struct list_head *links_to_free) 1094 { 1095 struct cgrp_cset_link *link, *tmp_link; 1096 1097 list_for_each_entry_safe(link, tmp_link, links_to_free, cset_link) { 1098 list_del(&link->cset_link); 1099 kfree(link); 1100 } 1101 } 1102 1103 /** 1104 * allocate_cgrp_cset_links - allocate cgrp_cset_links 1105 * @count: the number of links to allocate 1106 * @tmp_links: list_head the allocated links are put on 1107 * 1108 * Allocate @count cgrp_cset_link structures and chain them on @tmp_links 1109 * through ->cset_link. Returns 0 on success or -errno. 1110 */ 1111 static int allocate_cgrp_cset_links(int count, struct list_head *tmp_links) 1112 { 1113 struct cgrp_cset_link *link; 1114 int i; 1115 1116 INIT_LIST_HEAD(tmp_links); 1117 1118 for (i = 0; i < count; i++) { 1119 link = kzalloc(sizeof(*link), GFP_KERNEL); 1120 if (!link) { 1121 free_cgrp_cset_links(tmp_links); 1122 return -ENOMEM; 1123 } 1124 list_add(&link->cset_link, tmp_links); 1125 } 1126 return 0; 1127 } 1128 1129 /** 1130 * link_css_set - a helper function to link a css_set to a cgroup 1131 * @tmp_links: cgrp_cset_link objects allocated by allocate_cgrp_cset_links() 1132 * @cset: the css_set to be linked 1133 * @cgrp: the destination cgroup 1134 */ 1135 static void link_css_set(struct list_head *tmp_links, struct css_set *cset, 1136 struct cgroup *cgrp) 1137 { 1138 struct cgrp_cset_link *link; 1139 1140 BUG_ON(list_empty(tmp_links)); 1141 1142 if (cgroup_on_dfl(cgrp)) 1143 cset->dfl_cgrp = cgrp; 1144 1145 link = list_first_entry(tmp_links, struct cgrp_cset_link, cset_link); 1146 link->cset = cset; 1147 link->cgrp = cgrp; 1148 1149 /* 1150 * Always add links to the tail of the lists so that the lists are 1151 * in chronological order. 1152 */ 1153 list_move_tail(&link->cset_link, &cgrp->cset_links); 1154 list_add_tail(&link->cgrp_link, &cset->cgrp_links); 1155 1156 if (cgroup_parent(cgrp)) 1157 cgroup_get_live(cgrp); 1158 } 1159 1160 /** 1161 * find_css_set - return a new css_set with one cgroup updated 1162 * @old_cset: the baseline css_set 1163 * @cgrp: the cgroup to be updated 1164 * 1165 * Return a new css_set that's equivalent to @old_cset, but with @cgrp 1166 * substituted into the appropriate hierarchy. 1167 */ 1168 static struct css_set *find_css_set(struct css_set *old_cset, 1169 struct cgroup *cgrp) 1170 { 1171 struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT] = { }; 1172 struct css_set *cset; 1173 struct list_head tmp_links; 1174 struct cgrp_cset_link *link; 1175 struct cgroup_subsys *ss; 1176 unsigned long key; 1177 int ssid; 1178 1179 lockdep_assert_held(&cgroup_mutex); 1180 1181 /* First see if we already have a cgroup group that matches 1182 * the desired set */ 1183 spin_lock_irq(&css_set_lock); 1184 cset = find_existing_css_set(old_cset, cgrp, template); 1185 if (cset) 1186 get_css_set(cset); 1187 spin_unlock_irq(&css_set_lock); 1188 1189 if (cset) 1190 return cset; 1191 1192 cset = kzalloc(sizeof(*cset), GFP_KERNEL); 1193 if (!cset) 1194 return NULL; 1195 1196 /* Allocate all the cgrp_cset_link objects that we'll need */ 1197 if (allocate_cgrp_cset_links(cgroup_root_count, &tmp_links) < 0) { 1198 kfree(cset); 1199 return NULL; 1200 } 1201 1202 refcount_set(&cset->refcount, 1); 1203 cset->dom_cset = cset; 1204 INIT_LIST_HEAD(&cset->tasks); 1205 INIT_LIST_HEAD(&cset->mg_tasks); 1206 INIT_LIST_HEAD(&cset->dying_tasks); 1207 INIT_LIST_HEAD(&cset->task_iters); 1208 INIT_LIST_HEAD(&cset->threaded_csets); 1209 INIT_HLIST_NODE(&cset->hlist); 1210 INIT_LIST_HEAD(&cset->cgrp_links); 1211 INIT_LIST_HEAD(&cset->mg_preload_node); 1212 INIT_LIST_HEAD(&cset->mg_node); 1213 1214 /* Copy the set of subsystem state objects generated in 1215 * find_existing_css_set() */ 1216 memcpy(cset->subsys, template, sizeof(cset->subsys)); 1217 1218 spin_lock_irq(&css_set_lock); 1219 /* Add reference counts and links from the new css_set. */ 1220 list_for_each_entry(link, &old_cset->cgrp_links, cgrp_link) { 1221 struct cgroup *c = link->cgrp; 1222 1223 if (c->root == cgrp->root) 1224 c = cgrp; 1225 link_css_set(&tmp_links, cset, c); 1226 } 1227 1228 BUG_ON(!list_empty(&tmp_links)); 1229 1230 css_set_count++; 1231 1232 /* Add @cset to the hash table */ 1233 key = css_set_hash(cset->subsys); 1234 hash_add(css_set_table, &cset->hlist, key); 1235 1236 for_each_subsys(ss, ssid) { 1237 struct cgroup_subsys_state *css = cset->subsys[ssid]; 1238 1239 list_add_tail(&cset->e_cset_node[ssid], 1240 &css->cgroup->e_csets[ssid]); 1241 css_get(css); 1242 } 1243 1244 spin_unlock_irq(&css_set_lock); 1245 1246 /* 1247 * If @cset should be threaded, look up the matching dom_cset and 1248 * link them up. We first fully initialize @cset then look for the 1249 * dom_cset. It's simpler this way and safe as @cset is guaranteed 1250 * to stay empty until we return. 1251 */ 1252 if (cgroup_is_threaded(cset->dfl_cgrp)) { 1253 struct css_set *dcset; 1254 1255 dcset = find_css_set(cset, cset->dfl_cgrp->dom_cgrp); 1256 if (!dcset) { 1257 put_css_set(cset); 1258 return NULL; 1259 } 1260 1261 spin_lock_irq(&css_set_lock); 1262 cset->dom_cset = dcset; 1263 list_add_tail(&cset->threaded_csets_node, 1264 &dcset->threaded_csets); 1265 spin_unlock_irq(&css_set_lock); 1266 } 1267 1268 return cset; 1269 } 1270 1271 struct cgroup_root *cgroup_root_from_kf(struct kernfs_root *kf_root) 1272 { 1273 struct cgroup *root_cgrp = kf_root->kn->priv; 1274 1275 return root_cgrp->root; 1276 } 1277 1278 static int cgroup_init_root_id(struct cgroup_root *root) 1279 { 1280 int id; 1281 1282 lockdep_assert_held(&cgroup_mutex); 1283 1284 id = idr_alloc_cyclic(&cgroup_hierarchy_idr, root, 0, 0, GFP_KERNEL); 1285 if (id < 0) 1286 return id; 1287 1288 root->hierarchy_id = id; 1289 return 0; 1290 } 1291 1292 static void cgroup_exit_root_id(struct cgroup_root *root) 1293 { 1294 lockdep_assert_held(&cgroup_mutex); 1295 1296 idr_remove(&cgroup_hierarchy_idr, root->hierarchy_id); 1297 } 1298 1299 void cgroup_free_root(struct cgroup_root *root) 1300 { 1301 kfree(root); 1302 } 1303 1304 static void cgroup_destroy_root(struct cgroup_root *root) 1305 { 1306 struct cgroup *cgrp = &root->cgrp; 1307 struct cgrp_cset_link *link, *tmp_link; 1308 1309 trace_cgroup_destroy_root(root); 1310 1311 cgroup_lock_and_drain_offline(&cgrp_dfl_root.cgrp); 1312 1313 BUG_ON(atomic_read(&root->nr_cgrps)); 1314 BUG_ON(!list_empty(&cgrp->self.children)); 1315 1316 /* Rebind all subsystems back to the default hierarchy */ 1317 WARN_ON(rebind_subsystems(&cgrp_dfl_root, root->subsys_mask)); 1318 1319 /* 1320 * Release all the links from cset_links to this hierarchy's 1321 * root cgroup 1322 */ 1323 spin_lock_irq(&css_set_lock); 1324 1325 list_for_each_entry_safe(link, tmp_link, &cgrp->cset_links, cset_link) { 1326 list_del(&link->cset_link); 1327 list_del(&link->cgrp_link); 1328 kfree(link); 1329 } 1330 1331 spin_unlock_irq(&css_set_lock); 1332 1333 if (!list_empty(&root->root_list)) { 1334 list_del(&root->root_list); 1335 cgroup_root_count--; 1336 } 1337 1338 cgroup_exit_root_id(root); 1339 1340 mutex_unlock(&cgroup_mutex); 1341 1342 kernfs_destroy_root(root->kf_root); 1343 cgroup_free_root(root); 1344 } 1345 1346 /* 1347 * look up cgroup associated with current task's cgroup namespace on the 1348 * specified hierarchy 1349 */ 1350 static struct cgroup * 1351 current_cgns_cgroup_from_root(struct cgroup_root *root) 1352 { 1353 struct cgroup *res = NULL; 1354 struct css_set *cset; 1355 1356 lockdep_assert_held(&css_set_lock); 1357 1358 rcu_read_lock(); 1359 1360 cset = current->nsproxy->cgroup_ns->root_cset; 1361 if (cset == &init_css_set) { 1362 res = &root->cgrp; 1363 } else if (root == &cgrp_dfl_root) { 1364 res = cset->dfl_cgrp; 1365 } else { 1366 struct cgrp_cset_link *link; 1367 1368 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) { 1369 struct cgroup *c = link->cgrp; 1370 1371 if (c->root == root) { 1372 res = c; 1373 break; 1374 } 1375 } 1376 } 1377 rcu_read_unlock(); 1378 1379 BUG_ON(!res); 1380 return res; 1381 } 1382 1383 /* look up cgroup associated with given css_set on the specified hierarchy */ 1384 static struct cgroup *cset_cgroup_from_root(struct css_set *cset, 1385 struct cgroup_root *root) 1386 { 1387 struct cgroup *res = NULL; 1388 1389 lockdep_assert_held(&cgroup_mutex); 1390 lockdep_assert_held(&css_set_lock); 1391 1392 if (cset == &init_css_set) { 1393 res = &root->cgrp; 1394 } else if (root == &cgrp_dfl_root) { 1395 res = cset->dfl_cgrp; 1396 } else { 1397 struct cgrp_cset_link *link; 1398 1399 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) { 1400 struct cgroup *c = link->cgrp; 1401 1402 if (c->root == root) { 1403 res = c; 1404 break; 1405 } 1406 } 1407 } 1408 1409 BUG_ON(!res); 1410 return res; 1411 } 1412 1413 /* 1414 * Return the cgroup for "task" from the given hierarchy. Must be 1415 * called with cgroup_mutex and css_set_lock held. 1416 */ 1417 struct cgroup *task_cgroup_from_root(struct task_struct *task, 1418 struct cgroup_root *root) 1419 { 1420 /* 1421 * No need to lock the task - since we hold css_set_lock the 1422 * task can't change groups. 1423 */ 1424 return cset_cgroup_from_root(task_css_set(task), root); 1425 } 1426 1427 /* 1428 * A task must hold cgroup_mutex to modify cgroups. 1429 * 1430 * Any task can increment and decrement the count field without lock. 1431 * So in general, code holding cgroup_mutex can't rely on the count 1432 * field not changing. However, if the count goes to zero, then only 1433 * cgroup_attach_task() can increment it again. Because a count of zero 1434 * means that no tasks are currently attached, therefore there is no 1435 * way a task attached to that cgroup can fork (the other way to 1436 * increment the count). So code holding cgroup_mutex can safely 1437 * assume that if the count is zero, it will stay zero. Similarly, if 1438 * a task holds cgroup_mutex on a cgroup with zero count, it 1439 * knows that the cgroup won't be removed, as cgroup_rmdir() 1440 * needs that mutex. 1441 * 1442 * A cgroup can only be deleted if both its 'count' of using tasks 1443 * is zero, and its list of 'children' cgroups is empty. Since all 1444 * tasks in the system use _some_ cgroup, and since there is always at 1445 * least one task in the system (init, pid == 1), therefore, root cgroup 1446 * always has either children cgroups and/or using tasks. So we don't 1447 * need a special hack to ensure that root cgroup cannot be deleted. 1448 * 1449 * P.S. One more locking exception. RCU is used to guard the 1450 * update of a tasks cgroup pointer by cgroup_attach_task() 1451 */ 1452 1453 static struct kernfs_syscall_ops cgroup_kf_syscall_ops; 1454 1455 static char *cgroup_file_name(struct cgroup *cgrp, const struct cftype *cft, 1456 char *buf) 1457 { 1458 struct cgroup_subsys *ss = cft->ss; 1459 1460 if (cft->ss && !(cft->flags & CFTYPE_NO_PREFIX) && 1461 !(cgrp->root->flags & CGRP_ROOT_NOPREFIX)) { 1462 const char *dbg = (cft->flags & CFTYPE_DEBUG) ? ".__DEBUG__." : ""; 1463 1464 snprintf(buf, CGROUP_FILE_NAME_MAX, "%s%s.%s", 1465 dbg, cgroup_on_dfl(cgrp) ? ss->name : ss->legacy_name, 1466 cft->name); 1467 } else { 1468 strscpy(buf, cft->name, CGROUP_FILE_NAME_MAX); 1469 } 1470 return buf; 1471 } 1472 1473 /** 1474 * cgroup_file_mode - deduce file mode of a control file 1475 * @cft: the control file in question 1476 * 1477 * S_IRUGO for read, S_IWUSR for write. 1478 */ 1479 static umode_t cgroup_file_mode(const struct cftype *cft) 1480 { 1481 umode_t mode = 0; 1482 1483 if (cft->read_u64 || cft->read_s64 || cft->seq_show) 1484 mode |= S_IRUGO; 1485 1486 if (cft->write_u64 || cft->write_s64 || cft->write) { 1487 if (cft->flags & CFTYPE_WORLD_WRITABLE) 1488 mode |= S_IWUGO; 1489 else 1490 mode |= S_IWUSR; 1491 } 1492 1493 return mode; 1494 } 1495 1496 /** 1497 * cgroup_calc_subtree_ss_mask - calculate subtree_ss_mask 1498 * @subtree_control: the new subtree_control mask to consider 1499 * @this_ss_mask: available subsystems 1500 * 1501 * On the default hierarchy, a subsystem may request other subsystems to be 1502 * enabled together through its ->depends_on mask. In such cases, more 1503 * subsystems than specified in "cgroup.subtree_control" may be enabled. 1504 * 1505 * This function calculates which subsystems need to be enabled if 1506 * @subtree_control is to be applied while restricted to @this_ss_mask. 1507 */ 1508 static u16 cgroup_calc_subtree_ss_mask(u16 subtree_control, u16 this_ss_mask) 1509 { 1510 u16 cur_ss_mask = subtree_control; 1511 struct cgroup_subsys *ss; 1512 int ssid; 1513 1514 lockdep_assert_held(&cgroup_mutex); 1515 1516 cur_ss_mask |= cgrp_dfl_implicit_ss_mask; 1517 1518 while (true) { 1519 u16 new_ss_mask = cur_ss_mask; 1520 1521 do_each_subsys_mask(ss, ssid, cur_ss_mask) { 1522 new_ss_mask |= ss->depends_on; 1523 } while_each_subsys_mask(); 1524 1525 /* 1526 * Mask out subsystems which aren't available. This can 1527 * happen only if some depended-upon subsystems were bound 1528 * to non-default hierarchies. 1529 */ 1530 new_ss_mask &= this_ss_mask; 1531 1532 if (new_ss_mask == cur_ss_mask) 1533 break; 1534 cur_ss_mask = new_ss_mask; 1535 } 1536 1537 return cur_ss_mask; 1538 } 1539 1540 /** 1541 * cgroup_kn_unlock - unlocking helper for cgroup kernfs methods 1542 * @kn: the kernfs_node being serviced 1543 * 1544 * This helper undoes cgroup_kn_lock_live() and should be invoked before 1545 * the method finishes if locking succeeded. Note that once this function 1546 * returns the cgroup returned by cgroup_kn_lock_live() may become 1547 * inaccessible any time. If the caller intends to continue to access the 1548 * cgroup, it should pin it before invoking this function. 1549 */ 1550 void cgroup_kn_unlock(struct kernfs_node *kn) 1551 { 1552 struct cgroup *cgrp; 1553 1554 if (kernfs_type(kn) == KERNFS_DIR) 1555 cgrp = kn->priv; 1556 else 1557 cgrp = kn->parent->priv; 1558 1559 mutex_unlock(&cgroup_mutex); 1560 1561 kernfs_unbreak_active_protection(kn); 1562 cgroup_put(cgrp); 1563 } 1564 1565 /** 1566 * cgroup_kn_lock_live - locking helper for cgroup kernfs methods 1567 * @kn: the kernfs_node being serviced 1568 * @drain_offline: perform offline draining on the cgroup 1569 * 1570 * This helper is to be used by a cgroup kernfs method currently servicing 1571 * @kn. It breaks the active protection, performs cgroup locking and 1572 * verifies that the associated cgroup is alive. Returns the cgroup if 1573 * alive; otherwise, %NULL. A successful return should be undone by a 1574 * matching cgroup_kn_unlock() invocation. If @drain_offline is %true, the 1575 * cgroup is drained of offlining csses before return. 1576 * 1577 * Any cgroup kernfs method implementation which requires locking the 1578 * associated cgroup should use this helper. It avoids nesting cgroup 1579 * locking under kernfs active protection and allows all kernfs operations 1580 * including self-removal. 1581 */ 1582 struct cgroup *cgroup_kn_lock_live(struct kernfs_node *kn, bool drain_offline) 1583 { 1584 struct cgroup *cgrp; 1585 1586 if (kernfs_type(kn) == KERNFS_DIR) 1587 cgrp = kn->priv; 1588 else 1589 cgrp = kn->parent->priv; 1590 1591 /* 1592 * We're gonna grab cgroup_mutex which nests outside kernfs 1593 * active_ref. cgroup liveliness check alone provides enough 1594 * protection against removal. Ensure @cgrp stays accessible and 1595 * break the active_ref protection. 1596 */ 1597 if (!cgroup_tryget(cgrp)) 1598 return NULL; 1599 kernfs_break_active_protection(kn); 1600 1601 if (drain_offline) 1602 cgroup_lock_and_drain_offline(cgrp); 1603 else 1604 mutex_lock(&cgroup_mutex); 1605 1606 if (!cgroup_is_dead(cgrp)) 1607 return cgrp; 1608 1609 cgroup_kn_unlock(kn); 1610 return NULL; 1611 } 1612 1613 static void cgroup_rm_file(struct cgroup *cgrp, const struct cftype *cft) 1614 { 1615 char name[CGROUP_FILE_NAME_MAX]; 1616 1617 lockdep_assert_held(&cgroup_mutex); 1618 1619 if (cft->file_offset) { 1620 struct cgroup_subsys_state *css = cgroup_css(cgrp, cft->ss); 1621 struct cgroup_file *cfile = (void *)css + cft->file_offset; 1622 1623 spin_lock_irq(&cgroup_file_kn_lock); 1624 cfile->kn = NULL; 1625 spin_unlock_irq(&cgroup_file_kn_lock); 1626 1627 del_timer_sync(&cfile->notify_timer); 1628 } 1629 1630 kernfs_remove_by_name(cgrp->kn, cgroup_file_name(cgrp, cft, name)); 1631 } 1632 1633 /** 1634 * css_clear_dir - remove subsys files in a cgroup directory 1635 * @css: taget css 1636 */ 1637 static void css_clear_dir(struct cgroup_subsys_state *css) 1638 { 1639 struct cgroup *cgrp = css->cgroup; 1640 struct cftype *cfts; 1641 1642 if (!(css->flags & CSS_VISIBLE)) 1643 return; 1644 1645 css->flags &= ~CSS_VISIBLE; 1646 1647 if (!css->ss) { 1648 if (cgroup_on_dfl(cgrp)) 1649 cfts = cgroup_base_files; 1650 else 1651 cfts = cgroup1_base_files; 1652 1653 cgroup_addrm_files(css, cgrp, cfts, false); 1654 } else { 1655 list_for_each_entry(cfts, &css->ss->cfts, node) 1656 cgroup_addrm_files(css, cgrp, cfts, false); 1657 } 1658 } 1659 1660 /** 1661 * css_populate_dir - create subsys files in a cgroup directory 1662 * @css: target css 1663 * 1664 * On failure, no file is added. 1665 */ 1666 static int css_populate_dir(struct cgroup_subsys_state *css) 1667 { 1668 struct cgroup *cgrp = css->cgroup; 1669 struct cftype *cfts, *failed_cfts; 1670 int ret; 1671 1672 if ((css->flags & CSS_VISIBLE) || !cgrp->kn) 1673 return 0; 1674 1675 if (!css->ss) { 1676 if (cgroup_on_dfl(cgrp)) 1677 cfts = cgroup_base_files; 1678 else 1679 cfts = cgroup1_base_files; 1680 1681 ret = cgroup_addrm_files(&cgrp->self, cgrp, cfts, true); 1682 if (ret < 0) 1683 return ret; 1684 } else { 1685 list_for_each_entry(cfts, &css->ss->cfts, node) { 1686 ret = cgroup_addrm_files(css, cgrp, cfts, true); 1687 if (ret < 0) { 1688 failed_cfts = cfts; 1689 goto err; 1690 } 1691 } 1692 } 1693 1694 css->flags |= CSS_VISIBLE; 1695 1696 return 0; 1697 err: 1698 list_for_each_entry(cfts, &css->ss->cfts, node) { 1699 if (cfts == failed_cfts) 1700 break; 1701 cgroup_addrm_files(css, cgrp, cfts, false); 1702 } 1703 return ret; 1704 } 1705 1706 int rebind_subsystems(struct cgroup_root *dst_root, u16 ss_mask) 1707 { 1708 struct cgroup *dcgrp = &dst_root->cgrp; 1709 struct cgroup_subsys *ss; 1710 int ssid, i, ret; 1711 1712 lockdep_assert_held(&cgroup_mutex); 1713 1714 do_each_subsys_mask(ss, ssid, ss_mask) { 1715 /* 1716 * If @ss has non-root csses attached to it, can't move. 1717 * If @ss is an implicit controller, it is exempt from this 1718 * rule and can be stolen. 1719 */ 1720 if (css_next_child(NULL, cgroup_css(&ss->root->cgrp, ss)) && 1721 !ss->implicit_on_dfl) 1722 return -EBUSY; 1723 1724 /* can't move between two non-dummy roots either */ 1725 if (ss->root != &cgrp_dfl_root && dst_root != &cgrp_dfl_root) 1726 return -EBUSY; 1727 } while_each_subsys_mask(); 1728 1729 do_each_subsys_mask(ss, ssid, ss_mask) { 1730 struct cgroup_root *src_root = ss->root; 1731 struct cgroup *scgrp = &src_root->cgrp; 1732 struct cgroup_subsys_state *css = cgroup_css(scgrp, ss); 1733 struct css_set *cset; 1734 1735 WARN_ON(!css || cgroup_css(dcgrp, ss)); 1736 1737 /* disable from the source */ 1738 src_root->subsys_mask &= ~(1 << ssid); 1739 WARN_ON(cgroup_apply_control(scgrp)); 1740 cgroup_finalize_control(scgrp, 0); 1741 1742 /* rebind */ 1743 RCU_INIT_POINTER(scgrp->subsys[ssid], NULL); 1744 rcu_assign_pointer(dcgrp->subsys[ssid], css); 1745 ss->root = dst_root; 1746 css->cgroup = dcgrp; 1747 1748 spin_lock_irq(&css_set_lock); 1749 hash_for_each(css_set_table, i, cset, hlist) 1750 list_move_tail(&cset->e_cset_node[ss->id], 1751 &dcgrp->e_csets[ss->id]); 1752 spin_unlock_irq(&css_set_lock); 1753 1754 /* default hierarchy doesn't enable controllers by default */ 1755 dst_root->subsys_mask |= 1 << ssid; 1756 if (dst_root == &cgrp_dfl_root) { 1757 static_branch_enable(cgroup_subsys_on_dfl_key[ssid]); 1758 } else { 1759 dcgrp->subtree_control |= 1 << ssid; 1760 static_branch_disable(cgroup_subsys_on_dfl_key[ssid]); 1761 } 1762 1763 ret = cgroup_apply_control(dcgrp); 1764 if (ret) 1765 pr_warn("partial failure to rebind %s controller (err=%d)\n", 1766 ss->name, ret); 1767 1768 if (ss->bind) 1769 ss->bind(css); 1770 } while_each_subsys_mask(); 1771 1772 kernfs_activate(dcgrp->kn); 1773 return 0; 1774 } 1775 1776 int cgroup_show_path(struct seq_file *sf, struct kernfs_node *kf_node, 1777 struct kernfs_root *kf_root) 1778 { 1779 int len = 0; 1780 char *buf = NULL; 1781 struct cgroup_root *kf_cgroot = cgroup_root_from_kf(kf_root); 1782 struct cgroup *ns_cgroup; 1783 1784 buf = kmalloc(PATH_MAX, GFP_KERNEL); 1785 if (!buf) 1786 return -ENOMEM; 1787 1788 spin_lock_irq(&css_set_lock); 1789 ns_cgroup = current_cgns_cgroup_from_root(kf_cgroot); 1790 len = kernfs_path_from_node(kf_node, ns_cgroup->kn, buf, PATH_MAX); 1791 spin_unlock_irq(&css_set_lock); 1792 1793 if (len >= PATH_MAX) 1794 len = -ERANGE; 1795 else if (len > 0) { 1796 seq_escape(sf, buf, " \t\n\\"); 1797 len = 0; 1798 } 1799 kfree(buf); 1800 return len; 1801 } 1802 1803 enum cgroup2_param { 1804 Opt_nsdelegate, 1805 Opt_memory_localevents, 1806 Opt_memory_recursiveprot, 1807 nr__cgroup2_params 1808 }; 1809 1810 static const struct fs_parameter_spec cgroup2_fs_parameters[] = { 1811 fsparam_flag("nsdelegate", Opt_nsdelegate), 1812 fsparam_flag("memory_localevents", Opt_memory_localevents), 1813 fsparam_flag("memory_recursiveprot", Opt_memory_recursiveprot), 1814 {} 1815 }; 1816 1817 static int cgroup2_parse_param(struct fs_context *fc, struct fs_parameter *param) 1818 { 1819 struct cgroup_fs_context *ctx = cgroup_fc2context(fc); 1820 struct fs_parse_result result; 1821 int opt; 1822 1823 opt = fs_parse(fc, cgroup2_fs_parameters, param, &result); 1824 if (opt < 0) 1825 return opt; 1826 1827 switch (opt) { 1828 case Opt_nsdelegate: 1829 ctx->flags |= CGRP_ROOT_NS_DELEGATE; 1830 return 0; 1831 case Opt_memory_localevents: 1832 ctx->flags |= CGRP_ROOT_MEMORY_LOCAL_EVENTS; 1833 return 0; 1834 case Opt_memory_recursiveprot: 1835 ctx->flags |= CGRP_ROOT_MEMORY_RECURSIVE_PROT; 1836 return 0; 1837 } 1838 return -EINVAL; 1839 } 1840 1841 static void apply_cgroup_root_flags(unsigned int root_flags) 1842 { 1843 if (current->nsproxy->cgroup_ns == &init_cgroup_ns) { 1844 if (root_flags & CGRP_ROOT_NS_DELEGATE) 1845 cgrp_dfl_root.flags |= CGRP_ROOT_NS_DELEGATE; 1846 else 1847 cgrp_dfl_root.flags &= ~CGRP_ROOT_NS_DELEGATE; 1848 1849 if (root_flags & CGRP_ROOT_MEMORY_LOCAL_EVENTS) 1850 cgrp_dfl_root.flags |= CGRP_ROOT_MEMORY_LOCAL_EVENTS; 1851 else 1852 cgrp_dfl_root.flags &= ~CGRP_ROOT_MEMORY_LOCAL_EVENTS; 1853 1854 if (root_flags & CGRP_ROOT_MEMORY_RECURSIVE_PROT) 1855 cgrp_dfl_root.flags |= CGRP_ROOT_MEMORY_RECURSIVE_PROT; 1856 else 1857 cgrp_dfl_root.flags &= ~CGRP_ROOT_MEMORY_RECURSIVE_PROT; 1858 } 1859 } 1860 1861 static int cgroup_show_options(struct seq_file *seq, struct kernfs_root *kf_root) 1862 { 1863 if (cgrp_dfl_root.flags & CGRP_ROOT_NS_DELEGATE) 1864 seq_puts(seq, ",nsdelegate"); 1865 if (cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_LOCAL_EVENTS) 1866 seq_puts(seq, ",memory_localevents"); 1867 if (cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_RECURSIVE_PROT) 1868 seq_puts(seq, ",memory_recursiveprot"); 1869 return 0; 1870 } 1871 1872 static int cgroup_reconfigure(struct fs_context *fc) 1873 { 1874 struct cgroup_fs_context *ctx = cgroup_fc2context(fc); 1875 1876 apply_cgroup_root_flags(ctx->flags); 1877 return 0; 1878 } 1879 1880 static void init_cgroup_housekeeping(struct cgroup *cgrp) 1881 { 1882 struct cgroup_subsys *ss; 1883 int ssid; 1884 1885 INIT_LIST_HEAD(&cgrp->self.sibling); 1886 INIT_LIST_HEAD(&cgrp->self.children); 1887 INIT_LIST_HEAD(&cgrp->cset_links); 1888 INIT_LIST_HEAD(&cgrp->pidlists); 1889 mutex_init(&cgrp->pidlist_mutex); 1890 cgrp->self.cgroup = cgrp; 1891 cgrp->self.flags |= CSS_ONLINE; 1892 cgrp->dom_cgrp = cgrp; 1893 cgrp->max_descendants = INT_MAX; 1894 cgrp->max_depth = INT_MAX; 1895 INIT_LIST_HEAD(&cgrp->rstat_css_list); 1896 prev_cputime_init(&cgrp->prev_cputime); 1897 1898 for_each_subsys(ss, ssid) 1899 INIT_LIST_HEAD(&cgrp->e_csets[ssid]); 1900 1901 init_waitqueue_head(&cgrp->offline_waitq); 1902 INIT_WORK(&cgrp->release_agent_work, cgroup1_release_agent); 1903 } 1904 1905 void init_cgroup_root(struct cgroup_fs_context *ctx) 1906 { 1907 struct cgroup_root *root = ctx->root; 1908 struct cgroup *cgrp = &root->cgrp; 1909 1910 INIT_LIST_HEAD(&root->root_list); 1911 atomic_set(&root->nr_cgrps, 1); 1912 cgrp->root = root; 1913 init_cgroup_housekeeping(cgrp); 1914 1915 root->flags = ctx->flags; 1916 if (ctx->release_agent) 1917 strscpy(root->release_agent_path, ctx->release_agent, PATH_MAX); 1918 if (ctx->name) 1919 strscpy(root->name, ctx->name, MAX_CGROUP_ROOT_NAMELEN); 1920 if (ctx->cpuset_clone_children) 1921 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags); 1922 } 1923 1924 int cgroup_setup_root(struct cgroup_root *root, u16 ss_mask) 1925 { 1926 LIST_HEAD(tmp_links); 1927 struct cgroup *root_cgrp = &root->cgrp; 1928 struct kernfs_syscall_ops *kf_sops; 1929 struct css_set *cset; 1930 int i, ret; 1931 1932 lockdep_assert_held(&cgroup_mutex); 1933 1934 ret = percpu_ref_init(&root_cgrp->self.refcnt, css_release, 1935 0, GFP_KERNEL); 1936 if (ret) 1937 goto out; 1938 1939 /* 1940 * We're accessing css_set_count without locking css_set_lock here, 1941 * but that's OK - it can only be increased by someone holding 1942 * cgroup_lock, and that's us. Later rebinding may disable 1943 * controllers on the default hierarchy and thus create new csets, 1944 * which can't be more than the existing ones. Allocate 2x. 1945 */ 1946 ret = allocate_cgrp_cset_links(2 * css_set_count, &tmp_links); 1947 if (ret) 1948 goto cancel_ref; 1949 1950 ret = cgroup_init_root_id(root); 1951 if (ret) 1952 goto cancel_ref; 1953 1954 kf_sops = root == &cgrp_dfl_root ? 1955 &cgroup_kf_syscall_ops : &cgroup1_kf_syscall_ops; 1956 1957 root->kf_root = kernfs_create_root(kf_sops, 1958 KERNFS_ROOT_CREATE_DEACTIVATED | 1959 KERNFS_ROOT_SUPPORT_EXPORTOP | 1960 KERNFS_ROOT_SUPPORT_USER_XATTR, 1961 root_cgrp); 1962 if (IS_ERR(root->kf_root)) { 1963 ret = PTR_ERR(root->kf_root); 1964 goto exit_root_id; 1965 } 1966 root_cgrp->kn = root->kf_root->kn; 1967 WARN_ON_ONCE(cgroup_ino(root_cgrp) != 1); 1968 root_cgrp->ancestor_ids[0] = cgroup_id(root_cgrp); 1969 1970 ret = css_populate_dir(&root_cgrp->self); 1971 if (ret) 1972 goto destroy_root; 1973 1974 ret = rebind_subsystems(root, ss_mask); 1975 if (ret) 1976 goto destroy_root; 1977 1978 ret = cgroup_bpf_inherit(root_cgrp); 1979 WARN_ON_ONCE(ret); 1980 1981 trace_cgroup_setup_root(root); 1982 1983 /* 1984 * There must be no failure case after here, since rebinding takes 1985 * care of subsystems' refcounts, which are explicitly dropped in 1986 * the failure exit path. 1987 */ 1988 list_add(&root->root_list, &cgroup_roots); 1989 cgroup_root_count++; 1990 1991 /* 1992 * Link the root cgroup in this hierarchy into all the css_set 1993 * objects. 1994 */ 1995 spin_lock_irq(&css_set_lock); 1996 hash_for_each(css_set_table, i, cset, hlist) { 1997 link_css_set(&tmp_links, cset, root_cgrp); 1998 if (css_set_populated(cset)) 1999 cgroup_update_populated(root_cgrp, true); 2000 } 2001 spin_unlock_irq(&css_set_lock); 2002 2003 BUG_ON(!list_empty(&root_cgrp->self.children)); 2004 BUG_ON(atomic_read(&root->nr_cgrps) != 1); 2005 2006 ret = 0; 2007 goto out; 2008 2009 destroy_root: 2010 kernfs_destroy_root(root->kf_root); 2011 root->kf_root = NULL; 2012 exit_root_id: 2013 cgroup_exit_root_id(root); 2014 cancel_ref: 2015 percpu_ref_exit(&root_cgrp->self.refcnt); 2016 out: 2017 free_cgrp_cset_links(&tmp_links); 2018 return ret; 2019 } 2020 2021 int cgroup_do_get_tree(struct fs_context *fc) 2022 { 2023 struct cgroup_fs_context *ctx = cgroup_fc2context(fc); 2024 int ret; 2025 2026 ctx->kfc.root = ctx->root->kf_root; 2027 if (fc->fs_type == &cgroup2_fs_type) 2028 ctx->kfc.magic = CGROUP2_SUPER_MAGIC; 2029 else 2030 ctx->kfc.magic = CGROUP_SUPER_MAGIC; 2031 ret = kernfs_get_tree(fc); 2032 2033 /* 2034 * In non-init cgroup namespace, instead of root cgroup's dentry, 2035 * we return the dentry corresponding to the cgroupns->root_cgrp. 2036 */ 2037 if (!ret && ctx->ns != &init_cgroup_ns) { 2038 struct dentry *nsdentry; 2039 struct super_block *sb = fc->root->d_sb; 2040 struct cgroup *cgrp; 2041 2042 mutex_lock(&cgroup_mutex); 2043 spin_lock_irq(&css_set_lock); 2044 2045 cgrp = cset_cgroup_from_root(ctx->ns->root_cset, ctx->root); 2046 2047 spin_unlock_irq(&css_set_lock); 2048 mutex_unlock(&cgroup_mutex); 2049 2050 nsdentry = kernfs_node_dentry(cgrp->kn, sb); 2051 dput(fc->root); 2052 if (IS_ERR(nsdentry)) { 2053 deactivate_locked_super(sb); 2054 ret = PTR_ERR(nsdentry); 2055 nsdentry = NULL; 2056 } 2057 fc->root = nsdentry; 2058 } 2059 2060 if (!ctx->kfc.new_sb_created) 2061 cgroup_put(&ctx->root->cgrp); 2062 2063 return ret; 2064 } 2065 2066 /* 2067 * Destroy a cgroup filesystem context. 2068 */ 2069 static void cgroup_fs_context_free(struct fs_context *fc) 2070 { 2071 struct cgroup_fs_context *ctx = cgroup_fc2context(fc); 2072 2073 kfree(ctx->name); 2074 kfree(ctx->release_agent); 2075 put_cgroup_ns(ctx->ns); 2076 kernfs_free_fs_context(fc); 2077 kfree(ctx); 2078 } 2079 2080 static int cgroup_get_tree(struct fs_context *fc) 2081 { 2082 struct cgroup_fs_context *ctx = cgroup_fc2context(fc); 2083 int ret; 2084 2085 cgrp_dfl_visible = true; 2086 cgroup_get_live(&cgrp_dfl_root.cgrp); 2087 ctx->root = &cgrp_dfl_root; 2088 2089 ret = cgroup_do_get_tree(fc); 2090 if (!ret) 2091 apply_cgroup_root_flags(ctx->flags); 2092 return ret; 2093 } 2094 2095 static const struct fs_context_operations cgroup_fs_context_ops = { 2096 .free = cgroup_fs_context_free, 2097 .parse_param = cgroup2_parse_param, 2098 .get_tree = cgroup_get_tree, 2099 .reconfigure = cgroup_reconfigure, 2100 }; 2101 2102 static const struct fs_context_operations cgroup1_fs_context_ops = { 2103 .free = cgroup_fs_context_free, 2104 .parse_param = cgroup1_parse_param, 2105 .get_tree = cgroup1_get_tree, 2106 .reconfigure = cgroup1_reconfigure, 2107 }; 2108 2109 /* 2110 * Initialise the cgroup filesystem creation/reconfiguration context. Notably, 2111 * we select the namespace we're going to use. 2112 */ 2113 static int cgroup_init_fs_context(struct fs_context *fc) 2114 { 2115 struct cgroup_fs_context *ctx; 2116 2117 ctx = kzalloc(sizeof(struct cgroup_fs_context), GFP_KERNEL); 2118 if (!ctx) 2119 return -ENOMEM; 2120 2121 ctx->ns = current->nsproxy->cgroup_ns; 2122 get_cgroup_ns(ctx->ns); 2123 fc->fs_private = &ctx->kfc; 2124 if (fc->fs_type == &cgroup2_fs_type) 2125 fc->ops = &cgroup_fs_context_ops; 2126 else 2127 fc->ops = &cgroup1_fs_context_ops; 2128 put_user_ns(fc->user_ns); 2129 fc->user_ns = get_user_ns(ctx->ns->user_ns); 2130 fc->global = true; 2131 return 0; 2132 } 2133 2134 static void cgroup_kill_sb(struct super_block *sb) 2135 { 2136 struct kernfs_root *kf_root = kernfs_root_from_sb(sb); 2137 struct cgroup_root *root = cgroup_root_from_kf(kf_root); 2138 2139 /* 2140 * If @root doesn't have any children, start killing it. 2141 * This prevents new mounts by disabling percpu_ref_tryget_live(). 2142 * cgroup_mount() may wait for @root's release. 2143 * 2144 * And don't kill the default root. 2145 */ 2146 if (list_empty(&root->cgrp.self.children) && root != &cgrp_dfl_root && 2147 !percpu_ref_is_dying(&root->cgrp.self.refcnt)) 2148 percpu_ref_kill(&root->cgrp.self.refcnt); 2149 cgroup_put(&root->cgrp); 2150 kernfs_kill_sb(sb); 2151 } 2152 2153 struct file_system_type cgroup_fs_type = { 2154 .name = "cgroup", 2155 .init_fs_context = cgroup_init_fs_context, 2156 .parameters = cgroup1_fs_parameters, 2157 .kill_sb = cgroup_kill_sb, 2158 .fs_flags = FS_USERNS_MOUNT, 2159 }; 2160 2161 static struct file_system_type cgroup2_fs_type = { 2162 .name = "cgroup2", 2163 .init_fs_context = cgroup_init_fs_context, 2164 .parameters = cgroup2_fs_parameters, 2165 .kill_sb = cgroup_kill_sb, 2166 .fs_flags = FS_USERNS_MOUNT, 2167 }; 2168 2169 #ifdef CONFIG_CPUSETS 2170 static const struct fs_context_operations cpuset_fs_context_ops = { 2171 .get_tree = cgroup1_get_tree, 2172 .free = cgroup_fs_context_free, 2173 }; 2174 2175 /* 2176 * This is ugly, but preserves the userspace API for existing cpuset 2177 * users. If someone tries to mount the "cpuset" filesystem, we 2178 * silently switch it to mount "cgroup" instead 2179 */ 2180 static int cpuset_init_fs_context(struct fs_context *fc) 2181 { 2182 char *agent = kstrdup("/sbin/cpuset_release_agent", GFP_USER); 2183 struct cgroup_fs_context *ctx; 2184 int err; 2185 2186 err = cgroup_init_fs_context(fc); 2187 if (err) { 2188 kfree(agent); 2189 return err; 2190 } 2191 2192 fc->ops = &cpuset_fs_context_ops; 2193 2194 ctx = cgroup_fc2context(fc); 2195 ctx->subsys_mask = 1 << cpuset_cgrp_id; 2196 ctx->flags |= CGRP_ROOT_NOPREFIX; 2197 ctx->release_agent = agent; 2198 2199 get_filesystem(&cgroup_fs_type); 2200 put_filesystem(fc->fs_type); 2201 fc->fs_type = &cgroup_fs_type; 2202 2203 return 0; 2204 } 2205 2206 static struct file_system_type cpuset_fs_type = { 2207 .name = "cpuset", 2208 .init_fs_context = cpuset_init_fs_context, 2209 .fs_flags = FS_USERNS_MOUNT, 2210 }; 2211 #endif 2212 2213 int cgroup_path_ns_locked(struct cgroup *cgrp, char *buf, size_t buflen, 2214 struct cgroup_namespace *ns) 2215 { 2216 struct cgroup *root = cset_cgroup_from_root(ns->root_cset, cgrp->root); 2217 2218 return kernfs_path_from_node(cgrp->kn, root->kn, buf, buflen); 2219 } 2220 2221 int cgroup_path_ns(struct cgroup *cgrp, char *buf, size_t buflen, 2222 struct cgroup_namespace *ns) 2223 { 2224 int ret; 2225 2226 mutex_lock(&cgroup_mutex); 2227 spin_lock_irq(&css_set_lock); 2228 2229 ret = cgroup_path_ns_locked(cgrp, buf, buflen, ns); 2230 2231 spin_unlock_irq(&css_set_lock); 2232 mutex_unlock(&cgroup_mutex); 2233 2234 return ret; 2235 } 2236 EXPORT_SYMBOL_GPL(cgroup_path_ns); 2237 2238 /** 2239 * task_cgroup_path - cgroup path of a task in the first cgroup hierarchy 2240 * @task: target task 2241 * @buf: the buffer to write the path into 2242 * @buflen: the length of the buffer 2243 * 2244 * Determine @task's cgroup on the first (the one with the lowest non-zero 2245 * hierarchy_id) cgroup hierarchy and copy its path into @buf. This 2246 * function grabs cgroup_mutex and shouldn't be used inside locks used by 2247 * cgroup controller callbacks. 2248 * 2249 * Return value is the same as kernfs_path(). 2250 */ 2251 int task_cgroup_path(struct task_struct *task, char *buf, size_t buflen) 2252 { 2253 struct cgroup_root *root; 2254 struct cgroup *cgrp; 2255 int hierarchy_id = 1; 2256 int ret; 2257 2258 mutex_lock(&cgroup_mutex); 2259 spin_lock_irq(&css_set_lock); 2260 2261 root = idr_get_next(&cgroup_hierarchy_idr, &hierarchy_id); 2262 2263 if (root) { 2264 cgrp = task_cgroup_from_root(task, root); 2265 ret = cgroup_path_ns_locked(cgrp, buf, buflen, &init_cgroup_ns); 2266 } else { 2267 /* if no hierarchy exists, everyone is in "/" */ 2268 ret = strlcpy(buf, "/", buflen); 2269 } 2270 2271 spin_unlock_irq(&css_set_lock); 2272 mutex_unlock(&cgroup_mutex); 2273 return ret; 2274 } 2275 EXPORT_SYMBOL_GPL(task_cgroup_path); 2276 2277 /** 2278 * cgroup_migrate_add_task - add a migration target task to a migration context 2279 * @task: target task 2280 * @mgctx: target migration context 2281 * 2282 * Add @task, which is a migration target, to @mgctx->tset. This function 2283 * becomes noop if @task doesn't need to be migrated. @task's css_set 2284 * should have been added as a migration source and @task->cg_list will be 2285 * moved from the css_set's tasks list to mg_tasks one. 2286 */ 2287 static void cgroup_migrate_add_task(struct task_struct *task, 2288 struct cgroup_mgctx *mgctx) 2289 { 2290 struct css_set *cset; 2291 2292 lockdep_assert_held(&css_set_lock); 2293 2294 /* @task either already exited or can't exit until the end */ 2295 if (task->flags & PF_EXITING) 2296 return; 2297 2298 /* cgroup_threadgroup_rwsem protects racing against forks */ 2299 WARN_ON_ONCE(list_empty(&task->cg_list)); 2300 2301 cset = task_css_set(task); 2302 if (!cset->mg_src_cgrp) 2303 return; 2304 2305 mgctx->tset.nr_tasks++; 2306 2307 list_move_tail(&task->cg_list, &cset->mg_tasks); 2308 if (list_empty(&cset->mg_node)) 2309 list_add_tail(&cset->mg_node, 2310 &mgctx->tset.src_csets); 2311 if (list_empty(&cset->mg_dst_cset->mg_node)) 2312 list_add_tail(&cset->mg_dst_cset->mg_node, 2313 &mgctx->tset.dst_csets); 2314 } 2315 2316 /** 2317 * cgroup_taskset_first - reset taskset and return the first task 2318 * @tset: taskset of interest 2319 * @dst_cssp: output variable for the destination css 2320 * 2321 * @tset iteration is initialized and the first task is returned. 2322 */ 2323 struct task_struct *cgroup_taskset_first(struct cgroup_taskset *tset, 2324 struct cgroup_subsys_state **dst_cssp) 2325 { 2326 tset->cur_cset = list_first_entry(tset->csets, struct css_set, mg_node); 2327 tset->cur_task = NULL; 2328 2329 return cgroup_taskset_next(tset, dst_cssp); 2330 } 2331 2332 /** 2333 * cgroup_taskset_next - iterate to the next task in taskset 2334 * @tset: taskset of interest 2335 * @dst_cssp: output variable for the destination css 2336 * 2337 * Return the next task in @tset. Iteration must have been initialized 2338 * with cgroup_taskset_first(). 2339 */ 2340 struct task_struct *cgroup_taskset_next(struct cgroup_taskset *tset, 2341 struct cgroup_subsys_state **dst_cssp) 2342 { 2343 struct css_set *cset = tset->cur_cset; 2344 struct task_struct *task = tset->cur_task; 2345 2346 while (&cset->mg_node != tset->csets) { 2347 if (!task) 2348 task = list_first_entry(&cset->mg_tasks, 2349 struct task_struct, cg_list); 2350 else 2351 task = list_next_entry(task, cg_list); 2352 2353 if (&task->cg_list != &cset->mg_tasks) { 2354 tset->cur_cset = cset; 2355 tset->cur_task = task; 2356 2357 /* 2358 * This function may be called both before and 2359 * after cgroup_taskset_migrate(). The two cases 2360 * can be distinguished by looking at whether @cset 2361 * has its ->mg_dst_cset set. 2362 */ 2363 if (cset->mg_dst_cset) 2364 *dst_cssp = cset->mg_dst_cset->subsys[tset->ssid]; 2365 else 2366 *dst_cssp = cset->subsys[tset->ssid]; 2367 2368 return task; 2369 } 2370 2371 cset = list_next_entry(cset, mg_node); 2372 task = NULL; 2373 } 2374 2375 return NULL; 2376 } 2377 2378 /** 2379 * cgroup_taskset_migrate - migrate a taskset 2380 * @mgctx: migration context 2381 * 2382 * Migrate tasks in @mgctx as setup by migration preparation functions. 2383 * This function fails iff one of the ->can_attach callbacks fails and 2384 * guarantees that either all or none of the tasks in @mgctx are migrated. 2385 * @mgctx is consumed regardless of success. 2386 */ 2387 static int cgroup_migrate_execute(struct cgroup_mgctx *mgctx) 2388 { 2389 struct cgroup_taskset *tset = &mgctx->tset; 2390 struct cgroup_subsys *ss; 2391 struct task_struct *task, *tmp_task; 2392 struct css_set *cset, *tmp_cset; 2393 int ssid, failed_ssid, ret; 2394 2395 /* check that we can legitimately attach to the cgroup */ 2396 if (tset->nr_tasks) { 2397 do_each_subsys_mask(ss, ssid, mgctx->ss_mask) { 2398 if (ss->can_attach) { 2399 tset->ssid = ssid; 2400 ret = ss->can_attach(tset); 2401 if (ret) { 2402 failed_ssid = ssid; 2403 goto out_cancel_attach; 2404 } 2405 } 2406 } while_each_subsys_mask(); 2407 } 2408 2409 /* 2410 * Now that we're guaranteed success, proceed to move all tasks to 2411 * the new cgroup. There are no failure cases after here, so this 2412 * is the commit point. 2413 */ 2414 spin_lock_irq(&css_set_lock); 2415 list_for_each_entry(cset, &tset->src_csets, mg_node) { 2416 list_for_each_entry_safe(task, tmp_task, &cset->mg_tasks, cg_list) { 2417 struct css_set *from_cset = task_css_set(task); 2418 struct css_set *to_cset = cset->mg_dst_cset; 2419 2420 get_css_set(to_cset); 2421 to_cset->nr_tasks++; 2422 css_set_move_task(task, from_cset, to_cset, true); 2423 from_cset->nr_tasks--; 2424 /* 2425 * If the source or destination cgroup is frozen, 2426 * the task might require to change its state. 2427 */ 2428 cgroup_freezer_migrate_task(task, from_cset->dfl_cgrp, 2429 to_cset->dfl_cgrp); 2430 put_css_set_locked(from_cset); 2431 2432 } 2433 } 2434 spin_unlock_irq(&css_set_lock); 2435 2436 /* 2437 * Migration is committed, all target tasks are now on dst_csets. 2438 * Nothing is sensitive to fork() after this point. Notify 2439 * controllers that migration is complete. 2440 */ 2441 tset->csets = &tset->dst_csets; 2442 2443 if (tset->nr_tasks) { 2444 do_each_subsys_mask(ss, ssid, mgctx->ss_mask) { 2445 if (ss->attach) { 2446 tset->ssid = ssid; 2447 ss->attach(tset); 2448 } 2449 } while_each_subsys_mask(); 2450 } 2451 2452 ret = 0; 2453 goto out_release_tset; 2454 2455 out_cancel_attach: 2456 if (tset->nr_tasks) { 2457 do_each_subsys_mask(ss, ssid, mgctx->ss_mask) { 2458 if (ssid == failed_ssid) 2459 break; 2460 if (ss->cancel_attach) { 2461 tset->ssid = ssid; 2462 ss->cancel_attach(tset); 2463 } 2464 } while_each_subsys_mask(); 2465 } 2466 out_release_tset: 2467 spin_lock_irq(&css_set_lock); 2468 list_splice_init(&tset->dst_csets, &tset->src_csets); 2469 list_for_each_entry_safe(cset, tmp_cset, &tset->src_csets, mg_node) { 2470 list_splice_tail_init(&cset->mg_tasks, &cset->tasks); 2471 list_del_init(&cset->mg_node); 2472 } 2473 spin_unlock_irq(&css_set_lock); 2474 2475 /* 2476 * Re-initialize the cgroup_taskset structure in case it is reused 2477 * again in another cgroup_migrate_add_task()/cgroup_migrate_execute() 2478 * iteration. 2479 */ 2480 tset->nr_tasks = 0; 2481 tset->csets = &tset->src_csets; 2482 return ret; 2483 } 2484 2485 /** 2486 * cgroup_migrate_vet_dst - verify whether a cgroup can be migration destination 2487 * @dst_cgrp: destination cgroup to test 2488 * 2489 * On the default hierarchy, except for the mixable, (possible) thread root 2490 * and threaded cgroups, subtree_control must be zero for migration 2491 * destination cgroups with tasks so that child cgroups don't compete 2492 * against tasks. 2493 */ 2494 int cgroup_migrate_vet_dst(struct cgroup *dst_cgrp) 2495 { 2496 /* v1 doesn't have any restriction */ 2497 if (!cgroup_on_dfl(dst_cgrp)) 2498 return 0; 2499 2500 /* verify @dst_cgrp can host resources */ 2501 if (!cgroup_is_valid_domain(dst_cgrp->dom_cgrp)) 2502 return -EOPNOTSUPP; 2503 2504 /* mixables don't care */ 2505 if (cgroup_is_mixable(dst_cgrp)) 2506 return 0; 2507 2508 /* 2509 * If @dst_cgrp is already or can become a thread root or is 2510 * threaded, it doesn't matter. 2511 */ 2512 if (cgroup_can_be_thread_root(dst_cgrp) || cgroup_is_threaded(dst_cgrp)) 2513 return 0; 2514 2515 /* apply no-internal-process constraint */ 2516 if (dst_cgrp->subtree_control) 2517 return -EBUSY; 2518 2519 return 0; 2520 } 2521 2522 /** 2523 * cgroup_migrate_finish - cleanup after attach 2524 * @mgctx: migration context 2525 * 2526 * Undo cgroup_migrate_add_src() and cgroup_migrate_prepare_dst(). See 2527 * those functions for details. 2528 */ 2529 void cgroup_migrate_finish(struct cgroup_mgctx *mgctx) 2530 { 2531 LIST_HEAD(preloaded); 2532 struct css_set *cset, *tmp_cset; 2533 2534 lockdep_assert_held(&cgroup_mutex); 2535 2536 spin_lock_irq(&css_set_lock); 2537 2538 list_splice_tail_init(&mgctx->preloaded_src_csets, &preloaded); 2539 list_splice_tail_init(&mgctx->preloaded_dst_csets, &preloaded); 2540 2541 list_for_each_entry_safe(cset, tmp_cset, &preloaded, mg_preload_node) { 2542 cset->mg_src_cgrp = NULL; 2543 cset->mg_dst_cgrp = NULL; 2544 cset->mg_dst_cset = NULL; 2545 list_del_init(&cset->mg_preload_node); 2546 put_css_set_locked(cset); 2547 } 2548 2549 spin_unlock_irq(&css_set_lock); 2550 } 2551 2552 /** 2553 * cgroup_migrate_add_src - add a migration source css_set 2554 * @src_cset: the source css_set to add 2555 * @dst_cgrp: the destination cgroup 2556 * @mgctx: migration context 2557 * 2558 * Tasks belonging to @src_cset are about to be migrated to @dst_cgrp. Pin 2559 * @src_cset and add it to @mgctx->src_csets, which should later be cleaned 2560 * up by cgroup_migrate_finish(). 2561 * 2562 * This function may be called without holding cgroup_threadgroup_rwsem 2563 * even if the target is a process. Threads may be created and destroyed 2564 * but as long as cgroup_mutex is not dropped, no new css_set can be put 2565 * into play and the preloaded css_sets are guaranteed to cover all 2566 * migrations. 2567 */ 2568 void cgroup_migrate_add_src(struct css_set *src_cset, 2569 struct cgroup *dst_cgrp, 2570 struct cgroup_mgctx *mgctx) 2571 { 2572 struct cgroup *src_cgrp; 2573 2574 lockdep_assert_held(&cgroup_mutex); 2575 lockdep_assert_held(&css_set_lock); 2576 2577 /* 2578 * If ->dead, @src_set is associated with one or more dead cgroups 2579 * and doesn't contain any migratable tasks. Ignore it early so 2580 * that the rest of migration path doesn't get confused by it. 2581 */ 2582 if (src_cset->dead) 2583 return; 2584 2585 src_cgrp = cset_cgroup_from_root(src_cset, dst_cgrp->root); 2586 2587 if (!list_empty(&src_cset->mg_preload_node)) 2588 return; 2589 2590 WARN_ON(src_cset->mg_src_cgrp); 2591 WARN_ON(src_cset->mg_dst_cgrp); 2592 WARN_ON(!list_empty(&src_cset->mg_tasks)); 2593 WARN_ON(!list_empty(&src_cset->mg_node)); 2594 2595 src_cset->mg_src_cgrp = src_cgrp; 2596 src_cset->mg_dst_cgrp = dst_cgrp; 2597 get_css_set(src_cset); 2598 list_add_tail(&src_cset->mg_preload_node, &mgctx->preloaded_src_csets); 2599 } 2600 2601 /** 2602 * cgroup_migrate_prepare_dst - prepare destination css_sets for migration 2603 * @mgctx: migration context 2604 * 2605 * Tasks are about to be moved and all the source css_sets have been 2606 * preloaded to @mgctx->preloaded_src_csets. This function looks up and 2607 * pins all destination css_sets, links each to its source, and append them 2608 * to @mgctx->preloaded_dst_csets. 2609 * 2610 * This function must be called after cgroup_migrate_add_src() has been 2611 * called on each migration source css_set. After migration is performed 2612 * using cgroup_migrate(), cgroup_migrate_finish() must be called on 2613 * @mgctx. 2614 */ 2615 int cgroup_migrate_prepare_dst(struct cgroup_mgctx *mgctx) 2616 { 2617 struct css_set *src_cset, *tmp_cset; 2618 2619 lockdep_assert_held(&cgroup_mutex); 2620 2621 /* look up the dst cset for each src cset and link it to src */ 2622 list_for_each_entry_safe(src_cset, tmp_cset, &mgctx->preloaded_src_csets, 2623 mg_preload_node) { 2624 struct css_set *dst_cset; 2625 struct cgroup_subsys *ss; 2626 int ssid; 2627 2628 dst_cset = find_css_set(src_cset, src_cset->mg_dst_cgrp); 2629 if (!dst_cset) 2630 return -ENOMEM; 2631 2632 WARN_ON_ONCE(src_cset->mg_dst_cset || dst_cset->mg_dst_cset); 2633 2634 /* 2635 * If src cset equals dst, it's noop. Drop the src. 2636 * cgroup_migrate() will skip the cset too. Note that we 2637 * can't handle src == dst as some nodes are used by both. 2638 */ 2639 if (src_cset == dst_cset) { 2640 src_cset->mg_src_cgrp = NULL; 2641 src_cset->mg_dst_cgrp = NULL; 2642 list_del_init(&src_cset->mg_preload_node); 2643 put_css_set(src_cset); 2644 put_css_set(dst_cset); 2645 continue; 2646 } 2647 2648 src_cset->mg_dst_cset = dst_cset; 2649 2650 if (list_empty(&dst_cset->mg_preload_node)) 2651 list_add_tail(&dst_cset->mg_preload_node, 2652 &mgctx->preloaded_dst_csets); 2653 else 2654 put_css_set(dst_cset); 2655 2656 for_each_subsys(ss, ssid) 2657 if (src_cset->subsys[ssid] != dst_cset->subsys[ssid]) 2658 mgctx->ss_mask |= 1 << ssid; 2659 } 2660 2661 return 0; 2662 } 2663 2664 /** 2665 * cgroup_migrate - migrate a process or task to a cgroup 2666 * @leader: the leader of the process or the task to migrate 2667 * @threadgroup: whether @leader points to the whole process or a single task 2668 * @mgctx: migration context 2669 * 2670 * Migrate a process or task denoted by @leader. If migrating a process, 2671 * the caller must be holding cgroup_threadgroup_rwsem. The caller is also 2672 * responsible for invoking cgroup_migrate_add_src() and 2673 * cgroup_migrate_prepare_dst() on the targets before invoking this 2674 * function and following up with cgroup_migrate_finish(). 2675 * 2676 * As long as a controller's ->can_attach() doesn't fail, this function is 2677 * guaranteed to succeed. This means that, excluding ->can_attach() 2678 * failure, when migrating multiple targets, the success or failure can be 2679 * decided for all targets by invoking group_migrate_prepare_dst() before 2680 * actually starting migrating. 2681 */ 2682 int cgroup_migrate(struct task_struct *leader, bool threadgroup, 2683 struct cgroup_mgctx *mgctx) 2684 { 2685 struct task_struct *task; 2686 2687 /* 2688 * Prevent freeing of tasks while we take a snapshot. Tasks that are 2689 * already PF_EXITING could be freed from underneath us unless we 2690 * take an rcu_read_lock. 2691 */ 2692 spin_lock_irq(&css_set_lock); 2693 rcu_read_lock(); 2694 task = leader; 2695 do { 2696 cgroup_migrate_add_task(task, mgctx); 2697 if (!threadgroup) 2698 break; 2699 } while_each_thread(leader, task); 2700 rcu_read_unlock(); 2701 spin_unlock_irq(&css_set_lock); 2702 2703 return cgroup_migrate_execute(mgctx); 2704 } 2705 2706 /** 2707 * cgroup_attach_task - attach a task or a whole threadgroup to a cgroup 2708 * @dst_cgrp: the cgroup to attach to 2709 * @leader: the task or the leader of the threadgroup to be attached 2710 * @threadgroup: attach the whole threadgroup? 2711 * 2712 * Call holding cgroup_mutex and cgroup_threadgroup_rwsem. 2713 */ 2714 int cgroup_attach_task(struct cgroup *dst_cgrp, struct task_struct *leader, 2715 bool threadgroup) 2716 { 2717 DEFINE_CGROUP_MGCTX(mgctx); 2718 struct task_struct *task; 2719 int ret = 0; 2720 2721 /* look up all src csets */ 2722 spin_lock_irq(&css_set_lock); 2723 rcu_read_lock(); 2724 task = leader; 2725 do { 2726 cgroup_migrate_add_src(task_css_set(task), dst_cgrp, &mgctx); 2727 if (!threadgroup) 2728 break; 2729 } while_each_thread(leader, task); 2730 rcu_read_unlock(); 2731 spin_unlock_irq(&css_set_lock); 2732 2733 /* prepare dst csets and commit */ 2734 ret = cgroup_migrate_prepare_dst(&mgctx); 2735 if (!ret) 2736 ret = cgroup_migrate(leader, threadgroup, &mgctx); 2737 2738 cgroup_migrate_finish(&mgctx); 2739 2740 if (!ret) 2741 TRACE_CGROUP_PATH(attach_task, dst_cgrp, leader, threadgroup); 2742 2743 return ret; 2744 } 2745 2746 struct task_struct *cgroup_procs_write_start(char *buf, bool threadgroup, 2747 bool *locked) 2748 __acquires(&cgroup_threadgroup_rwsem) 2749 { 2750 struct task_struct *tsk; 2751 pid_t pid; 2752 2753 if (kstrtoint(strstrip(buf), 0, &pid) || pid < 0) 2754 return ERR_PTR(-EINVAL); 2755 2756 /* 2757 * If we migrate a single thread, we don't care about threadgroup 2758 * stability. If the thread is `current`, it won't exit(2) under our 2759 * hands or change PID through exec(2). We exclude 2760 * cgroup_update_dfl_csses and other cgroup_{proc,thread}s_write 2761 * callers by cgroup_mutex. 2762 * Therefore, we can skip the global lock. 2763 */ 2764 lockdep_assert_held(&cgroup_mutex); 2765 if (pid || threadgroup) { 2766 percpu_down_write(&cgroup_threadgroup_rwsem); 2767 *locked = true; 2768 } else { 2769 *locked = false; 2770 } 2771 2772 rcu_read_lock(); 2773 if (pid) { 2774 tsk = find_task_by_vpid(pid); 2775 if (!tsk) { 2776 tsk = ERR_PTR(-ESRCH); 2777 goto out_unlock_threadgroup; 2778 } 2779 } else { 2780 tsk = current; 2781 } 2782 2783 if (threadgroup) 2784 tsk = tsk->group_leader; 2785 2786 /* 2787 * kthreads may acquire PF_NO_SETAFFINITY during initialization. 2788 * If userland migrates such a kthread to a non-root cgroup, it can 2789 * become trapped in a cpuset, or RT kthread may be born in a 2790 * cgroup with no rt_runtime allocated. Just say no. 2791 */ 2792 if (tsk->no_cgroup_migration || (tsk->flags & PF_NO_SETAFFINITY)) { 2793 tsk = ERR_PTR(-EINVAL); 2794 goto out_unlock_threadgroup; 2795 } 2796 2797 get_task_struct(tsk); 2798 goto out_unlock_rcu; 2799 2800 out_unlock_threadgroup: 2801 if (*locked) { 2802 percpu_up_write(&cgroup_threadgroup_rwsem); 2803 *locked = false; 2804 } 2805 out_unlock_rcu: 2806 rcu_read_unlock(); 2807 return tsk; 2808 } 2809 2810 void cgroup_procs_write_finish(struct task_struct *task, bool locked) 2811 __releases(&cgroup_threadgroup_rwsem) 2812 { 2813 struct cgroup_subsys *ss; 2814 int ssid; 2815 2816 /* release reference from cgroup_procs_write_start() */ 2817 put_task_struct(task); 2818 2819 if (locked) 2820 percpu_up_write(&cgroup_threadgroup_rwsem); 2821 for_each_subsys(ss, ssid) 2822 if (ss->post_attach) 2823 ss->post_attach(); 2824 } 2825 2826 static void cgroup_print_ss_mask(struct seq_file *seq, u16 ss_mask) 2827 { 2828 struct cgroup_subsys *ss; 2829 bool printed = false; 2830 int ssid; 2831 2832 do_each_subsys_mask(ss, ssid, ss_mask) { 2833 if (printed) 2834 seq_putc(seq, ' '); 2835 seq_puts(seq, ss->name); 2836 printed = true; 2837 } while_each_subsys_mask(); 2838 if (printed) 2839 seq_putc(seq, '\n'); 2840 } 2841 2842 /* show controllers which are enabled from the parent */ 2843 static int cgroup_controllers_show(struct seq_file *seq, void *v) 2844 { 2845 struct cgroup *cgrp = seq_css(seq)->cgroup; 2846 2847 cgroup_print_ss_mask(seq, cgroup_control(cgrp)); 2848 return 0; 2849 } 2850 2851 /* show controllers which are enabled for a given cgroup's children */ 2852 static int cgroup_subtree_control_show(struct seq_file *seq, void *v) 2853 { 2854 struct cgroup *cgrp = seq_css(seq)->cgroup; 2855 2856 cgroup_print_ss_mask(seq, cgrp->subtree_control); 2857 return 0; 2858 } 2859 2860 /** 2861 * cgroup_update_dfl_csses - update css assoc of a subtree in default hierarchy 2862 * @cgrp: root of the subtree to update csses for 2863 * 2864 * @cgrp's control masks have changed and its subtree's css associations 2865 * need to be updated accordingly. This function looks up all css_sets 2866 * which are attached to the subtree, creates the matching updated css_sets 2867 * and migrates the tasks to the new ones. 2868 */ 2869 static int cgroup_update_dfl_csses(struct cgroup *cgrp) 2870 { 2871 DEFINE_CGROUP_MGCTX(mgctx); 2872 struct cgroup_subsys_state *d_css; 2873 struct cgroup *dsct; 2874 struct css_set *src_cset; 2875 int ret; 2876 2877 lockdep_assert_held(&cgroup_mutex); 2878 2879 percpu_down_write(&cgroup_threadgroup_rwsem); 2880 2881 /* look up all csses currently attached to @cgrp's subtree */ 2882 spin_lock_irq(&css_set_lock); 2883 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) { 2884 struct cgrp_cset_link *link; 2885 2886 list_for_each_entry(link, &dsct->cset_links, cset_link) 2887 cgroup_migrate_add_src(link->cset, dsct, &mgctx); 2888 } 2889 spin_unlock_irq(&css_set_lock); 2890 2891 /* NULL dst indicates self on default hierarchy */ 2892 ret = cgroup_migrate_prepare_dst(&mgctx); 2893 if (ret) 2894 goto out_finish; 2895 2896 spin_lock_irq(&css_set_lock); 2897 list_for_each_entry(src_cset, &mgctx.preloaded_src_csets, mg_preload_node) { 2898 struct task_struct *task, *ntask; 2899 2900 /* all tasks in src_csets need to be migrated */ 2901 list_for_each_entry_safe(task, ntask, &src_cset->tasks, cg_list) 2902 cgroup_migrate_add_task(task, &mgctx); 2903 } 2904 spin_unlock_irq(&css_set_lock); 2905 2906 ret = cgroup_migrate_execute(&mgctx); 2907 out_finish: 2908 cgroup_migrate_finish(&mgctx); 2909 percpu_up_write(&cgroup_threadgroup_rwsem); 2910 return ret; 2911 } 2912 2913 /** 2914 * cgroup_lock_and_drain_offline - lock cgroup_mutex and drain offlined csses 2915 * @cgrp: root of the target subtree 2916 * 2917 * Because css offlining is asynchronous, userland may try to re-enable a 2918 * controller while the previous css is still around. This function grabs 2919 * cgroup_mutex and drains the previous css instances of @cgrp's subtree. 2920 */ 2921 void cgroup_lock_and_drain_offline(struct cgroup *cgrp) 2922 __acquires(&cgroup_mutex) 2923 { 2924 struct cgroup *dsct; 2925 struct cgroup_subsys_state *d_css; 2926 struct cgroup_subsys *ss; 2927 int ssid; 2928 2929 restart: 2930 mutex_lock(&cgroup_mutex); 2931 2932 cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) { 2933 for_each_subsys(ss, ssid) { 2934 struct cgroup_subsys_state *css = cgroup_css(dsct, ss); 2935 DEFINE_WAIT(wait); 2936 2937 if (!css || !percpu_ref_is_dying(&css->refcnt)) 2938 continue; 2939 2940 cgroup_get_live(dsct); 2941 prepare_to_wait(&dsct->offline_waitq, &wait, 2942 TASK_UNINTERRUPTIBLE); 2943 2944 mutex_unlock(&cgroup_mutex); 2945 schedule(); 2946 finish_wait(&dsct->offline_waitq, &wait); 2947 2948 cgroup_put(dsct); 2949 goto restart; 2950 } 2951 } 2952 } 2953 2954 /** 2955 * cgroup_save_control - save control masks and dom_cgrp of a subtree 2956 * @cgrp: root of the target subtree 2957 * 2958 * Save ->subtree_control, ->subtree_ss_mask and ->dom_cgrp to the 2959 * respective old_ prefixed fields for @cgrp's subtree including @cgrp 2960 * itself. 2961 */ 2962 static void cgroup_save_control(struct cgroup *cgrp) 2963 { 2964 struct cgroup *dsct; 2965 struct cgroup_subsys_state *d_css; 2966 2967 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) { 2968 dsct->old_subtree_control = dsct->subtree_control; 2969 dsct->old_subtree_ss_mask = dsct->subtree_ss_mask; 2970 dsct->old_dom_cgrp = dsct->dom_cgrp; 2971 } 2972 } 2973 2974 /** 2975 * cgroup_propagate_control - refresh control masks of a subtree 2976 * @cgrp: root of the target subtree 2977 * 2978 * For @cgrp and its subtree, ensure ->subtree_ss_mask matches 2979 * ->subtree_control and propagate controller availability through the 2980 * subtree so that descendants don't have unavailable controllers enabled. 2981 */ 2982 static void cgroup_propagate_control(struct cgroup *cgrp) 2983 { 2984 struct cgroup *dsct; 2985 struct cgroup_subsys_state *d_css; 2986 2987 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) { 2988 dsct->subtree_control &= cgroup_control(dsct); 2989 dsct->subtree_ss_mask = 2990 cgroup_calc_subtree_ss_mask(dsct->subtree_control, 2991 cgroup_ss_mask(dsct)); 2992 } 2993 } 2994 2995 /** 2996 * cgroup_restore_control - restore control masks and dom_cgrp of a subtree 2997 * @cgrp: root of the target subtree 2998 * 2999 * Restore ->subtree_control, ->subtree_ss_mask and ->dom_cgrp from the 3000 * respective old_ prefixed fields for @cgrp's subtree including @cgrp 3001 * itself. 3002 */ 3003 static void cgroup_restore_control(struct cgroup *cgrp) 3004 { 3005 struct cgroup *dsct; 3006 struct cgroup_subsys_state *d_css; 3007 3008 cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) { 3009 dsct->subtree_control = dsct->old_subtree_control; 3010 dsct->subtree_ss_mask = dsct->old_subtree_ss_mask; 3011 dsct->dom_cgrp = dsct->old_dom_cgrp; 3012 } 3013 } 3014 3015 static bool css_visible(struct cgroup_subsys_state *css) 3016 { 3017 struct cgroup_subsys *ss = css->ss; 3018 struct cgroup *cgrp = css->cgroup; 3019 3020 if (cgroup_control(cgrp) & (1 << ss->id)) 3021 return true; 3022 if (!(cgroup_ss_mask(cgrp) & (1 << ss->id))) 3023 return false; 3024 return cgroup_on_dfl(cgrp) && ss->implicit_on_dfl; 3025 } 3026 3027 /** 3028 * cgroup_apply_control_enable - enable or show csses according to control 3029 * @cgrp: root of the target subtree 3030 * 3031 * Walk @cgrp's subtree and create new csses or make the existing ones 3032 * visible. A css is created invisible if it's being implicitly enabled 3033 * through dependency. An invisible css is made visible when the userland 3034 * explicitly enables it. 3035 * 3036 * Returns 0 on success, -errno on failure. On failure, csses which have 3037 * been processed already aren't cleaned up. The caller is responsible for 3038 * cleaning up with cgroup_apply_control_disable(). 3039 */ 3040 static int cgroup_apply_control_enable(struct cgroup *cgrp) 3041 { 3042 struct cgroup *dsct; 3043 struct cgroup_subsys_state *d_css; 3044 struct cgroup_subsys *ss; 3045 int ssid, ret; 3046 3047 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) { 3048 for_each_subsys(ss, ssid) { 3049 struct cgroup_subsys_state *css = cgroup_css(dsct, ss); 3050 3051 if (!(cgroup_ss_mask(dsct) & (1 << ss->id))) 3052 continue; 3053 3054 if (!css) { 3055 css = css_create(dsct, ss); 3056 if (IS_ERR(css)) 3057 return PTR_ERR(css); 3058 } 3059 3060 WARN_ON_ONCE(percpu_ref_is_dying(&css->refcnt)); 3061 3062 if (css_visible(css)) { 3063 ret = css_populate_dir(css); 3064 if (ret) 3065 return ret; 3066 } 3067 } 3068 } 3069 3070 return 0; 3071 } 3072 3073 /** 3074 * cgroup_apply_control_disable - kill or hide csses according to control 3075 * @cgrp: root of the target subtree 3076 * 3077 * Walk @cgrp's subtree and kill and hide csses so that they match 3078 * cgroup_ss_mask() and cgroup_visible_mask(). 3079 * 3080 * A css is hidden when the userland requests it to be disabled while other 3081 * subsystems are still depending on it. The css must not actively control 3082 * resources and be in the vanilla state if it's made visible again later. 3083 * Controllers which may be depended upon should provide ->css_reset() for 3084 * this purpose. 3085 */ 3086 static void cgroup_apply_control_disable(struct cgroup *cgrp) 3087 { 3088 struct cgroup *dsct; 3089 struct cgroup_subsys_state *d_css; 3090 struct cgroup_subsys *ss; 3091 int ssid; 3092 3093 cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) { 3094 for_each_subsys(ss, ssid) { 3095 struct cgroup_subsys_state *css = cgroup_css(dsct, ss); 3096 3097 if (!css) 3098 continue; 3099 3100 WARN_ON_ONCE(percpu_ref_is_dying(&css->refcnt)); 3101 3102 if (css->parent && 3103 !(cgroup_ss_mask(dsct) & (1 << ss->id))) { 3104 kill_css(css); 3105 } else if (!css_visible(css)) { 3106 css_clear_dir(css); 3107 if (ss->css_reset) 3108 ss->css_reset(css); 3109 } 3110 } 3111 } 3112 } 3113 3114 /** 3115 * cgroup_apply_control - apply control mask updates to the subtree 3116 * @cgrp: root of the target subtree 3117 * 3118 * subsystems can be enabled and disabled in a subtree using the following 3119 * steps. 3120 * 3121 * 1. Call cgroup_save_control() to stash the current state. 3122 * 2. Update ->subtree_control masks in the subtree as desired. 3123 * 3. Call cgroup_apply_control() to apply the changes. 3124 * 4. Optionally perform other related operations. 3125 * 5. Call cgroup_finalize_control() to finish up. 3126 * 3127 * This function implements step 3 and propagates the mask changes 3128 * throughout @cgrp's subtree, updates csses accordingly and perform 3129 * process migrations. 3130 */ 3131 static int cgroup_apply_control(struct cgroup *cgrp) 3132 { 3133 int ret; 3134 3135 cgroup_propagate_control(cgrp); 3136 3137 ret = cgroup_apply_control_enable(cgrp); 3138 if (ret) 3139 return ret; 3140 3141 /* 3142 * At this point, cgroup_e_css_by_mask() results reflect the new csses 3143 * making the following cgroup_update_dfl_csses() properly update 3144 * css associations of all tasks in the subtree. 3145 */ 3146 ret = cgroup_update_dfl_csses(cgrp); 3147 if (ret) 3148 return ret; 3149 3150 return 0; 3151 } 3152 3153 /** 3154 * cgroup_finalize_control - finalize control mask update 3155 * @cgrp: root of the target subtree 3156 * @ret: the result of the update 3157 * 3158 * Finalize control mask update. See cgroup_apply_control() for more info. 3159 */ 3160 static void cgroup_finalize_control(struct cgroup *cgrp, int ret) 3161 { 3162 if (ret) { 3163 cgroup_restore_control(cgrp); 3164 cgroup_propagate_control(cgrp); 3165 } 3166 3167 cgroup_apply_control_disable(cgrp); 3168 } 3169 3170 static int cgroup_vet_subtree_control_enable(struct cgroup *cgrp, u16 enable) 3171 { 3172 u16 domain_enable = enable & ~cgrp_dfl_threaded_ss_mask; 3173 3174 /* if nothing is getting enabled, nothing to worry about */ 3175 if (!enable) 3176 return 0; 3177 3178 /* can @cgrp host any resources? */ 3179 if (!cgroup_is_valid_domain(cgrp->dom_cgrp)) 3180 return -EOPNOTSUPP; 3181 3182 /* mixables don't care */ 3183 if (cgroup_is_mixable(cgrp)) 3184 return 0; 3185 3186 if (domain_enable) { 3187 /* can't enable domain controllers inside a thread subtree */ 3188 if (cgroup_is_thread_root(cgrp) || cgroup_is_threaded(cgrp)) 3189 return -EOPNOTSUPP; 3190 } else { 3191 /* 3192 * Threaded controllers can handle internal competitions 3193 * and are always allowed inside a (prospective) thread 3194 * subtree. 3195 */ 3196 if (cgroup_can_be_thread_root(cgrp) || cgroup_is_threaded(cgrp)) 3197 return 0; 3198 } 3199 3200 /* 3201 * Controllers can't be enabled for a cgroup with tasks to avoid 3202 * child cgroups competing against tasks. 3203 */ 3204 if (cgroup_has_tasks(cgrp)) 3205 return -EBUSY; 3206 3207 return 0; 3208 } 3209 3210 /* change the enabled child controllers for a cgroup in the default hierarchy */ 3211 static ssize_t cgroup_subtree_control_write(struct kernfs_open_file *of, 3212 char *buf, size_t nbytes, 3213 loff_t off) 3214 { 3215 u16 enable = 0, disable = 0; 3216 struct cgroup *cgrp, *child; 3217 struct cgroup_subsys *ss; 3218 char *tok; 3219 int ssid, ret; 3220 3221 /* 3222 * Parse input - space separated list of subsystem names prefixed 3223 * with either + or -. 3224 */ 3225 buf = strstrip(buf); 3226 while ((tok = strsep(&buf, " "))) { 3227 if (tok[0] == '\0') 3228 continue; 3229 do_each_subsys_mask(ss, ssid, ~cgrp_dfl_inhibit_ss_mask) { 3230 if (!cgroup_ssid_enabled(ssid) || 3231 strcmp(tok + 1, ss->name)) 3232 continue; 3233 3234 if (*tok == '+') { 3235 enable |= 1 << ssid; 3236 disable &= ~(1 << ssid); 3237 } else if (*tok == '-') { 3238 disable |= 1 << ssid; 3239 enable &= ~(1 << ssid); 3240 } else { 3241 return -EINVAL; 3242 } 3243 break; 3244 } while_each_subsys_mask(); 3245 if (ssid == CGROUP_SUBSYS_COUNT) 3246 return -EINVAL; 3247 } 3248 3249 cgrp = cgroup_kn_lock_live(of->kn, true); 3250 if (!cgrp) 3251 return -ENODEV; 3252 3253 for_each_subsys(ss, ssid) { 3254 if (enable & (1 << ssid)) { 3255 if (cgrp->subtree_control & (1 << ssid)) { 3256 enable &= ~(1 << ssid); 3257 continue; 3258 } 3259 3260 if (!(cgroup_control(cgrp) & (1 << ssid))) { 3261 ret = -ENOENT; 3262 goto out_unlock; 3263 } 3264 } else if (disable & (1 << ssid)) { 3265 if (!(cgrp->subtree_control & (1 << ssid))) { 3266 disable &= ~(1 << ssid); 3267 continue; 3268 } 3269 3270 /* a child has it enabled? */ 3271 cgroup_for_each_live_child(child, cgrp) { 3272 if (child->subtree_control & (1 << ssid)) { 3273 ret = -EBUSY; 3274 goto out_unlock; 3275 } 3276 } 3277 } 3278 } 3279 3280 if (!enable && !disable) { 3281 ret = 0; 3282 goto out_unlock; 3283 } 3284 3285 ret = cgroup_vet_subtree_control_enable(cgrp, enable); 3286 if (ret) 3287 goto out_unlock; 3288 3289 /* save and update control masks and prepare csses */ 3290 cgroup_save_control(cgrp); 3291 3292 cgrp->subtree_control |= enable; 3293 cgrp->subtree_control &= ~disable; 3294 3295 ret = cgroup_apply_control(cgrp); 3296 cgroup_finalize_control(cgrp, ret); 3297 if (ret) 3298 goto out_unlock; 3299 3300 kernfs_activate(cgrp->kn); 3301 out_unlock: 3302 cgroup_kn_unlock(of->kn); 3303 return ret ?: nbytes; 3304 } 3305 3306 /** 3307 * cgroup_enable_threaded - make @cgrp threaded 3308 * @cgrp: the target cgroup 3309 * 3310 * Called when "threaded" is written to the cgroup.type interface file and 3311 * tries to make @cgrp threaded and join the parent's resource domain. 3312 * This function is never called on the root cgroup as cgroup.type doesn't 3313 * exist on it. 3314 */ 3315 static int cgroup_enable_threaded(struct cgroup *cgrp) 3316 { 3317 struct cgroup *parent = cgroup_parent(cgrp); 3318 struct cgroup *dom_cgrp = parent->dom_cgrp; 3319 struct cgroup *dsct; 3320 struct cgroup_subsys_state *d_css; 3321 int ret; 3322 3323 lockdep_assert_held(&cgroup_mutex); 3324 3325 /* noop if already threaded */ 3326 if (cgroup_is_threaded(cgrp)) 3327 return 0; 3328 3329 /* 3330 * If @cgroup is populated or has domain controllers enabled, it 3331 * can't be switched. While the below cgroup_can_be_thread_root() 3332 * test can catch the same conditions, that's only when @parent is 3333 * not mixable, so let's check it explicitly. 3334 */ 3335 if (cgroup_is_populated(cgrp) || 3336 cgrp->subtree_control & ~cgrp_dfl_threaded_ss_mask) 3337 return -EOPNOTSUPP; 3338 3339 /* we're joining the parent's domain, ensure its validity */ 3340 if (!cgroup_is_valid_domain(dom_cgrp) || 3341 !cgroup_can_be_thread_root(dom_cgrp)) 3342 return -EOPNOTSUPP; 3343 3344 /* 3345 * The following shouldn't cause actual migrations and should 3346 * always succeed. 3347 */ 3348 cgroup_save_control(cgrp); 3349 3350 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) 3351 if (dsct == cgrp || cgroup_is_threaded(dsct)) 3352 dsct->dom_cgrp = dom_cgrp; 3353 3354 ret = cgroup_apply_control(cgrp); 3355 if (!ret) 3356 parent->nr_threaded_children++; 3357 3358 cgroup_finalize_control(cgrp, ret); 3359 return ret; 3360 } 3361 3362 static int cgroup_type_show(struct seq_file *seq, void *v) 3363 { 3364 struct cgroup *cgrp = seq_css(seq)->cgroup; 3365 3366 if (cgroup_is_threaded(cgrp)) 3367 seq_puts(seq, "threaded\n"); 3368 else if (!cgroup_is_valid_domain(cgrp)) 3369 seq_puts(seq, "domain invalid\n"); 3370 else if (cgroup_is_thread_root(cgrp)) 3371 seq_puts(seq, "domain threaded\n"); 3372 else 3373 seq_puts(seq, "domain\n"); 3374 3375 return 0; 3376 } 3377 3378 static ssize_t cgroup_type_write(struct kernfs_open_file *of, char *buf, 3379 size_t nbytes, loff_t off) 3380 { 3381 struct cgroup *cgrp; 3382 int ret; 3383 3384 /* only switching to threaded mode is supported */ 3385 if (strcmp(strstrip(buf), "threaded")) 3386 return -EINVAL; 3387 3388 /* drain dying csses before we re-apply (threaded) subtree control */ 3389 cgrp = cgroup_kn_lock_live(of->kn, true); 3390 if (!cgrp) 3391 return -ENOENT; 3392 3393 /* threaded can only be enabled */ 3394 ret = cgroup_enable_threaded(cgrp); 3395 3396 cgroup_kn_unlock(of->kn); 3397 return ret ?: nbytes; 3398 } 3399 3400 static int cgroup_max_descendants_show(struct seq_file *seq, void *v) 3401 { 3402 struct cgroup *cgrp = seq_css(seq)->cgroup; 3403 int descendants = READ_ONCE(cgrp->max_descendants); 3404 3405 if (descendants == INT_MAX) 3406 seq_puts(seq, "max\n"); 3407 else 3408 seq_printf(seq, "%d\n", descendants); 3409 3410 return 0; 3411 } 3412 3413 static ssize_t cgroup_max_descendants_write(struct kernfs_open_file *of, 3414 char *buf, size_t nbytes, loff_t off) 3415 { 3416 struct cgroup *cgrp; 3417 int descendants; 3418 ssize_t ret; 3419 3420 buf = strstrip(buf); 3421 if (!strcmp(buf, "max")) { 3422 descendants = INT_MAX; 3423 } else { 3424 ret = kstrtoint(buf, 0, &descendants); 3425 if (ret) 3426 return ret; 3427 } 3428 3429 if (descendants < 0) 3430 return -ERANGE; 3431 3432 cgrp = cgroup_kn_lock_live(of->kn, false); 3433 if (!cgrp) 3434 return -ENOENT; 3435 3436 cgrp->max_descendants = descendants; 3437 3438 cgroup_kn_unlock(of->kn); 3439 3440 return nbytes; 3441 } 3442 3443 static int cgroup_max_depth_show(struct seq_file *seq, void *v) 3444 { 3445 struct cgroup *cgrp = seq_css(seq)->cgroup; 3446 int depth = READ_ONCE(cgrp->max_depth); 3447 3448 if (depth == INT_MAX) 3449 seq_puts(seq, "max\n"); 3450 else 3451 seq_printf(seq, "%d\n", depth); 3452 3453 return 0; 3454 } 3455 3456 static ssize_t cgroup_max_depth_write(struct kernfs_open_file *of, 3457 char *buf, size_t nbytes, loff_t off) 3458 { 3459 struct cgroup *cgrp; 3460 ssize_t ret; 3461 int depth; 3462 3463 buf = strstrip(buf); 3464 if (!strcmp(buf, "max")) { 3465 depth = INT_MAX; 3466 } else { 3467 ret = kstrtoint(buf, 0, &depth); 3468 if (ret) 3469 return ret; 3470 } 3471 3472 if (depth < 0) 3473 return -ERANGE; 3474 3475 cgrp = cgroup_kn_lock_live(of->kn, false); 3476 if (!cgrp) 3477 return -ENOENT; 3478 3479 cgrp->max_depth = depth; 3480 3481 cgroup_kn_unlock(of->kn); 3482 3483 return nbytes; 3484 } 3485 3486 static int cgroup_events_show(struct seq_file *seq, void *v) 3487 { 3488 struct cgroup *cgrp = seq_css(seq)->cgroup; 3489 3490 seq_printf(seq, "populated %d\n", cgroup_is_populated(cgrp)); 3491 seq_printf(seq, "frozen %d\n", test_bit(CGRP_FROZEN, &cgrp->flags)); 3492 3493 return 0; 3494 } 3495 3496 static int cgroup_stat_show(struct seq_file *seq, void *v) 3497 { 3498 struct cgroup *cgroup = seq_css(seq)->cgroup; 3499 3500 seq_printf(seq, "nr_descendants %d\n", 3501 cgroup->nr_descendants); 3502 seq_printf(seq, "nr_dying_descendants %d\n", 3503 cgroup->nr_dying_descendants); 3504 3505 return 0; 3506 } 3507 3508 static int __maybe_unused cgroup_extra_stat_show(struct seq_file *seq, 3509 struct cgroup *cgrp, int ssid) 3510 { 3511 struct cgroup_subsys *ss = cgroup_subsys[ssid]; 3512 struct cgroup_subsys_state *css; 3513 int ret; 3514 3515 if (!ss->css_extra_stat_show) 3516 return 0; 3517 3518 css = cgroup_tryget_css(cgrp, ss); 3519 if (!css) 3520 return 0; 3521 3522 ret = ss->css_extra_stat_show(seq, css); 3523 css_put(css); 3524 return ret; 3525 } 3526 3527 static int cpu_stat_show(struct seq_file *seq, void *v) 3528 { 3529 struct cgroup __maybe_unused *cgrp = seq_css(seq)->cgroup; 3530 int ret = 0; 3531 3532 cgroup_base_stat_cputime_show(seq); 3533 #ifdef CONFIG_CGROUP_SCHED 3534 ret = cgroup_extra_stat_show(seq, cgrp, cpu_cgrp_id); 3535 #endif 3536 return ret; 3537 } 3538 3539 #ifdef CONFIG_PSI 3540 static int cgroup_io_pressure_show(struct seq_file *seq, void *v) 3541 { 3542 struct cgroup *cgrp = seq_css(seq)->cgroup; 3543 struct psi_group *psi = cgroup_ino(cgrp) == 1 ? &psi_system : &cgrp->psi; 3544 3545 return psi_show(seq, psi, PSI_IO); 3546 } 3547 static int cgroup_memory_pressure_show(struct seq_file *seq, void *v) 3548 { 3549 struct cgroup *cgrp = seq_css(seq)->cgroup; 3550 struct psi_group *psi = cgroup_ino(cgrp) == 1 ? &psi_system : &cgrp->psi; 3551 3552 return psi_show(seq, psi, PSI_MEM); 3553 } 3554 static int cgroup_cpu_pressure_show(struct seq_file *seq, void *v) 3555 { 3556 struct cgroup *cgrp = seq_css(seq)->cgroup; 3557 struct psi_group *psi = cgroup_ino(cgrp) == 1 ? &psi_system : &cgrp->psi; 3558 3559 return psi_show(seq, psi, PSI_CPU); 3560 } 3561 3562 static ssize_t cgroup_pressure_write(struct kernfs_open_file *of, char *buf, 3563 size_t nbytes, enum psi_res res) 3564 { 3565 struct psi_trigger *new; 3566 struct cgroup *cgrp; 3567 3568 cgrp = cgroup_kn_lock_live(of->kn, false); 3569 if (!cgrp) 3570 return -ENODEV; 3571 3572 cgroup_get(cgrp); 3573 cgroup_kn_unlock(of->kn); 3574 3575 new = psi_trigger_create(&cgrp->psi, buf, nbytes, res); 3576 if (IS_ERR(new)) { 3577 cgroup_put(cgrp); 3578 return PTR_ERR(new); 3579 } 3580 3581 psi_trigger_replace(&of->priv, new); 3582 3583 cgroup_put(cgrp); 3584 3585 return nbytes; 3586 } 3587 3588 static ssize_t cgroup_io_pressure_write(struct kernfs_open_file *of, 3589 char *buf, size_t nbytes, 3590 loff_t off) 3591 { 3592 return cgroup_pressure_write(of, buf, nbytes, PSI_IO); 3593 } 3594 3595 static ssize_t cgroup_memory_pressure_write(struct kernfs_open_file *of, 3596 char *buf, size_t nbytes, 3597 loff_t off) 3598 { 3599 return cgroup_pressure_write(of, buf, nbytes, PSI_MEM); 3600 } 3601 3602 static ssize_t cgroup_cpu_pressure_write(struct kernfs_open_file *of, 3603 char *buf, size_t nbytes, 3604 loff_t off) 3605 { 3606 return cgroup_pressure_write(of, buf, nbytes, PSI_CPU); 3607 } 3608 3609 static __poll_t cgroup_pressure_poll(struct kernfs_open_file *of, 3610 poll_table *pt) 3611 { 3612 return psi_trigger_poll(&of->priv, of->file, pt); 3613 } 3614 3615 static void cgroup_pressure_release(struct kernfs_open_file *of) 3616 { 3617 psi_trigger_replace(&of->priv, NULL); 3618 } 3619 #endif /* CONFIG_PSI */ 3620 3621 static int cgroup_freeze_show(struct seq_file *seq, void *v) 3622 { 3623 struct cgroup *cgrp = seq_css(seq)->cgroup; 3624 3625 seq_printf(seq, "%d\n", cgrp->freezer.freeze); 3626 3627 return 0; 3628 } 3629 3630 static ssize_t cgroup_freeze_write(struct kernfs_open_file *of, 3631 char *buf, size_t nbytes, loff_t off) 3632 { 3633 struct cgroup *cgrp; 3634 ssize_t ret; 3635 int freeze; 3636 3637 ret = kstrtoint(strstrip(buf), 0, &freeze); 3638 if (ret) 3639 return ret; 3640 3641 if (freeze < 0 || freeze > 1) 3642 return -ERANGE; 3643 3644 cgrp = cgroup_kn_lock_live(of->kn, false); 3645 if (!cgrp) 3646 return -ENOENT; 3647 3648 cgroup_freeze(cgrp, freeze); 3649 3650 cgroup_kn_unlock(of->kn); 3651 3652 return nbytes; 3653 } 3654 3655 static int cgroup_file_open(struct kernfs_open_file *of) 3656 { 3657 struct cftype *cft = of_cft(of); 3658 3659 if (cft->open) 3660 return cft->open(of); 3661 return 0; 3662 } 3663 3664 static void cgroup_file_release(struct kernfs_open_file *of) 3665 { 3666 struct cftype *cft = of_cft(of); 3667 3668 if (cft->release) 3669 cft->release(of); 3670 } 3671 3672 static ssize_t cgroup_file_write(struct kernfs_open_file *of, char *buf, 3673 size_t nbytes, loff_t off) 3674 { 3675 struct cgroup_namespace *ns = current->nsproxy->cgroup_ns; 3676 struct cgroup *cgrp = of->kn->parent->priv; 3677 struct cftype *cft = of_cft(of); 3678 struct cgroup_subsys_state *css; 3679 int ret; 3680 3681 if (!nbytes) 3682 return 0; 3683 3684 /* 3685 * If namespaces are delegation boundaries, disallow writes to 3686 * files in an non-init namespace root from inside the namespace 3687 * except for the files explicitly marked delegatable - 3688 * cgroup.procs and cgroup.subtree_control. 3689 */ 3690 if ((cgrp->root->flags & CGRP_ROOT_NS_DELEGATE) && 3691 !(cft->flags & CFTYPE_NS_DELEGATABLE) && 3692 ns != &init_cgroup_ns && ns->root_cset->dfl_cgrp == cgrp) 3693 return -EPERM; 3694 3695 if (cft->write) 3696 return cft->write(of, buf, nbytes, off); 3697 3698 /* 3699 * kernfs guarantees that a file isn't deleted with operations in 3700 * flight, which means that the matching css is and stays alive and 3701 * doesn't need to be pinned. The RCU locking is not necessary 3702 * either. It's just for the convenience of using cgroup_css(). 3703 */ 3704 rcu_read_lock(); 3705 css = cgroup_css(cgrp, cft->ss); 3706 rcu_read_unlock(); 3707 3708 if (cft->write_u64) { 3709 unsigned long long v; 3710 ret = kstrtoull(buf, 0, &v); 3711 if (!ret) 3712 ret = cft->write_u64(css, cft, v); 3713 } else if (cft->write_s64) { 3714 long long v; 3715 ret = kstrtoll(buf, 0, &v); 3716 if (!ret) 3717 ret = cft->write_s64(css, cft, v); 3718 } else { 3719 ret = -EINVAL; 3720 } 3721 3722 return ret ?: nbytes; 3723 } 3724 3725 static __poll_t cgroup_file_poll(struct kernfs_open_file *of, poll_table *pt) 3726 { 3727 struct cftype *cft = of_cft(of); 3728 3729 if (cft->poll) 3730 return cft->poll(of, pt); 3731 3732 return kernfs_generic_poll(of, pt); 3733 } 3734 3735 static void *cgroup_seqfile_start(struct seq_file *seq, loff_t *ppos) 3736 { 3737 return seq_cft(seq)->seq_start(seq, ppos); 3738 } 3739 3740 static void *cgroup_seqfile_next(struct seq_file *seq, void *v, loff_t *ppos) 3741 { 3742 return seq_cft(seq)->seq_next(seq, v, ppos); 3743 } 3744 3745 static void cgroup_seqfile_stop(struct seq_file *seq, void *v) 3746 { 3747 if (seq_cft(seq)->seq_stop) 3748 seq_cft(seq)->seq_stop(seq, v); 3749 } 3750 3751 static int cgroup_seqfile_show(struct seq_file *m, void *arg) 3752 { 3753 struct cftype *cft = seq_cft(m); 3754 struct cgroup_subsys_state *css = seq_css(m); 3755 3756 if (cft->seq_show) 3757 return cft->seq_show(m, arg); 3758 3759 if (cft->read_u64) 3760 seq_printf(m, "%llu\n", cft->read_u64(css, cft)); 3761 else if (cft->read_s64) 3762 seq_printf(m, "%lld\n", cft->read_s64(css, cft)); 3763 else 3764 return -EINVAL; 3765 return 0; 3766 } 3767 3768 static struct kernfs_ops cgroup_kf_single_ops = { 3769 .atomic_write_len = PAGE_SIZE, 3770 .open = cgroup_file_open, 3771 .release = cgroup_file_release, 3772 .write = cgroup_file_write, 3773 .poll = cgroup_file_poll, 3774 .seq_show = cgroup_seqfile_show, 3775 }; 3776 3777 static struct kernfs_ops cgroup_kf_ops = { 3778 .atomic_write_len = PAGE_SIZE, 3779 .open = cgroup_file_open, 3780 .release = cgroup_file_release, 3781 .write = cgroup_file_write, 3782 .poll = cgroup_file_poll, 3783 .seq_start = cgroup_seqfile_start, 3784 .seq_next = cgroup_seqfile_next, 3785 .seq_stop = cgroup_seqfile_stop, 3786 .seq_show = cgroup_seqfile_show, 3787 }; 3788 3789 /* set uid and gid of cgroup dirs and files to that of the creator */ 3790 static int cgroup_kn_set_ugid(struct kernfs_node *kn) 3791 { 3792 struct iattr iattr = { .ia_valid = ATTR_UID | ATTR_GID, 3793 .ia_uid = current_fsuid(), 3794 .ia_gid = current_fsgid(), }; 3795 3796 if (uid_eq(iattr.ia_uid, GLOBAL_ROOT_UID) && 3797 gid_eq(iattr.ia_gid, GLOBAL_ROOT_GID)) 3798 return 0; 3799 3800 return kernfs_setattr(kn, &iattr); 3801 } 3802 3803 static void cgroup_file_notify_timer(struct timer_list *timer) 3804 { 3805 cgroup_file_notify(container_of(timer, struct cgroup_file, 3806 notify_timer)); 3807 } 3808 3809 static int cgroup_add_file(struct cgroup_subsys_state *css, struct cgroup *cgrp, 3810 struct cftype *cft) 3811 { 3812 char name[CGROUP_FILE_NAME_MAX]; 3813 struct kernfs_node *kn; 3814 struct lock_class_key *key = NULL; 3815 int ret; 3816 3817 #ifdef CONFIG_DEBUG_LOCK_ALLOC 3818 key = &cft->lockdep_key; 3819 #endif 3820 kn = __kernfs_create_file(cgrp->kn, cgroup_file_name(cgrp, cft, name), 3821 cgroup_file_mode(cft), 3822 GLOBAL_ROOT_UID, GLOBAL_ROOT_GID, 3823 0, cft->kf_ops, cft, 3824 NULL, key); 3825 if (IS_ERR(kn)) 3826 return PTR_ERR(kn); 3827 3828 ret = cgroup_kn_set_ugid(kn); 3829 if (ret) { 3830 kernfs_remove(kn); 3831 return ret; 3832 } 3833 3834 if (cft->file_offset) { 3835 struct cgroup_file *cfile = (void *)css + cft->file_offset; 3836 3837 timer_setup(&cfile->notify_timer, cgroup_file_notify_timer, 0); 3838 3839 spin_lock_irq(&cgroup_file_kn_lock); 3840 cfile->kn = kn; 3841 spin_unlock_irq(&cgroup_file_kn_lock); 3842 } 3843 3844 return 0; 3845 } 3846 3847 /** 3848 * cgroup_addrm_files - add or remove files to a cgroup directory 3849 * @css: the target css 3850 * @cgrp: the target cgroup (usually css->cgroup) 3851 * @cfts: array of cftypes to be added 3852 * @is_add: whether to add or remove 3853 * 3854 * Depending on @is_add, add or remove files defined by @cfts on @cgrp. 3855 * For removals, this function never fails. 3856 */ 3857 static int cgroup_addrm_files(struct cgroup_subsys_state *css, 3858 struct cgroup *cgrp, struct cftype cfts[], 3859 bool is_add) 3860 { 3861 struct cftype *cft, *cft_end = NULL; 3862 int ret = 0; 3863 3864 lockdep_assert_held(&cgroup_mutex); 3865 3866 restart: 3867 for (cft = cfts; cft != cft_end && cft->name[0] != '\0'; cft++) { 3868 /* does cft->flags tell us to skip this file on @cgrp? */ 3869 if ((cft->flags & __CFTYPE_ONLY_ON_DFL) && !cgroup_on_dfl(cgrp)) 3870 continue; 3871 if ((cft->flags & __CFTYPE_NOT_ON_DFL) && cgroup_on_dfl(cgrp)) 3872 continue; 3873 if ((cft->flags & CFTYPE_NOT_ON_ROOT) && !cgroup_parent(cgrp)) 3874 continue; 3875 if ((cft->flags & CFTYPE_ONLY_ON_ROOT) && cgroup_parent(cgrp)) 3876 continue; 3877 if ((cft->flags & CFTYPE_DEBUG) && !cgroup_debug) 3878 continue; 3879 if (is_add) { 3880 ret = cgroup_add_file(css, cgrp, cft); 3881 if (ret) { 3882 pr_warn("%s: failed to add %s, err=%d\n", 3883 __func__, cft->name, ret); 3884 cft_end = cft; 3885 is_add = false; 3886 goto restart; 3887 } 3888 } else { 3889 cgroup_rm_file(cgrp, cft); 3890 } 3891 } 3892 return ret; 3893 } 3894 3895 static int cgroup_apply_cftypes(struct cftype *cfts, bool is_add) 3896 { 3897 struct cgroup_subsys *ss = cfts[0].ss; 3898 struct cgroup *root = &ss->root->cgrp; 3899 struct cgroup_subsys_state *css; 3900 int ret = 0; 3901 3902 lockdep_assert_held(&cgroup_mutex); 3903 3904 /* add/rm files for all cgroups created before */ 3905 css_for_each_descendant_pre(css, cgroup_css(root, ss)) { 3906 struct cgroup *cgrp = css->cgroup; 3907 3908 if (!(css->flags & CSS_VISIBLE)) 3909 continue; 3910 3911 ret = cgroup_addrm_files(css, cgrp, cfts, is_add); 3912 if (ret) 3913 break; 3914 } 3915 3916 if (is_add && !ret) 3917 kernfs_activate(root->kn); 3918 return ret; 3919 } 3920 3921 static void cgroup_exit_cftypes(struct cftype *cfts) 3922 { 3923 struct cftype *cft; 3924 3925 for (cft = cfts; cft->name[0] != '\0'; cft++) { 3926 /* free copy for custom atomic_write_len, see init_cftypes() */ 3927 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE) 3928 kfree(cft->kf_ops); 3929 cft->kf_ops = NULL; 3930 cft->ss = NULL; 3931 3932 /* revert flags set by cgroup core while adding @cfts */ 3933 cft->flags &= ~(__CFTYPE_ONLY_ON_DFL | __CFTYPE_NOT_ON_DFL); 3934 } 3935 } 3936 3937 static int cgroup_init_cftypes(struct cgroup_subsys *ss, struct cftype *cfts) 3938 { 3939 struct cftype *cft; 3940 3941 for (cft = cfts; cft->name[0] != '\0'; cft++) { 3942 struct kernfs_ops *kf_ops; 3943 3944 WARN_ON(cft->ss || cft->kf_ops); 3945 3946 if (cft->seq_start) 3947 kf_ops = &cgroup_kf_ops; 3948 else 3949 kf_ops = &cgroup_kf_single_ops; 3950 3951 /* 3952 * Ugh... if @cft wants a custom max_write_len, we need to 3953 * make a copy of kf_ops to set its atomic_write_len. 3954 */ 3955 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE) { 3956 kf_ops = kmemdup(kf_ops, sizeof(*kf_ops), GFP_KERNEL); 3957 if (!kf_ops) { 3958 cgroup_exit_cftypes(cfts); 3959 return -ENOMEM; 3960 } 3961 kf_ops->atomic_write_len = cft->max_write_len; 3962 } 3963 3964 cft->kf_ops = kf_ops; 3965 cft->ss = ss; 3966 } 3967 3968 return 0; 3969 } 3970 3971 static int cgroup_rm_cftypes_locked(struct cftype *cfts) 3972 { 3973 lockdep_assert_held(&cgroup_mutex); 3974 3975 if (!cfts || !cfts[0].ss) 3976 return -ENOENT; 3977 3978 list_del(&cfts->node); 3979 cgroup_apply_cftypes(cfts, false); 3980 cgroup_exit_cftypes(cfts); 3981 return 0; 3982 } 3983 3984 /** 3985 * cgroup_rm_cftypes - remove an array of cftypes from a subsystem 3986 * @cfts: zero-length name terminated array of cftypes 3987 * 3988 * Unregister @cfts. Files described by @cfts are removed from all 3989 * existing cgroups and all future cgroups won't have them either. This 3990 * function can be called anytime whether @cfts' subsys is attached or not. 3991 * 3992 * Returns 0 on successful unregistration, -ENOENT if @cfts is not 3993 * registered. 3994 */ 3995 int cgroup_rm_cftypes(struct cftype *cfts) 3996 { 3997 int ret; 3998 3999 mutex_lock(&cgroup_mutex); 4000 ret = cgroup_rm_cftypes_locked(cfts); 4001 mutex_unlock(&cgroup_mutex); 4002 return ret; 4003 } 4004 4005 /** 4006 * cgroup_add_cftypes - add an array of cftypes to a subsystem 4007 * @ss: target cgroup subsystem 4008 * @cfts: zero-length name terminated array of cftypes 4009 * 4010 * Register @cfts to @ss. Files described by @cfts are created for all 4011 * existing cgroups to which @ss is attached and all future cgroups will 4012 * have them too. This function can be called anytime whether @ss is 4013 * attached or not. 4014 * 4015 * Returns 0 on successful registration, -errno on failure. Note that this 4016 * function currently returns 0 as long as @cfts registration is successful 4017 * even if some file creation attempts on existing cgroups fail. 4018 */ 4019 static int cgroup_add_cftypes(struct cgroup_subsys *ss, struct cftype *cfts) 4020 { 4021 int ret; 4022 4023 if (!cgroup_ssid_enabled(ss->id)) 4024 return 0; 4025 4026 if (!cfts || cfts[0].name[0] == '\0') 4027 return 0; 4028 4029 ret = cgroup_init_cftypes(ss, cfts); 4030 if (ret) 4031 return ret; 4032 4033 mutex_lock(&cgroup_mutex); 4034 4035 list_add_tail(&cfts->node, &ss->cfts); 4036 ret = cgroup_apply_cftypes(cfts, true); 4037 if (ret) 4038 cgroup_rm_cftypes_locked(cfts); 4039 4040 mutex_unlock(&cgroup_mutex); 4041 return ret; 4042 } 4043 4044 /** 4045 * cgroup_add_dfl_cftypes - add an array of cftypes for default hierarchy 4046 * @ss: target cgroup subsystem 4047 * @cfts: zero-length name terminated array of cftypes 4048 * 4049 * Similar to cgroup_add_cftypes() but the added files are only used for 4050 * the default hierarchy. 4051 */ 4052 int cgroup_add_dfl_cftypes(struct cgroup_subsys *ss, struct cftype *cfts) 4053 { 4054 struct cftype *cft; 4055 4056 for (cft = cfts; cft && cft->name[0] != '\0'; cft++) 4057 cft->flags |= __CFTYPE_ONLY_ON_DFL; 4058 return cgroup_add_cftypes(ss, cfts); 4059 } 4060 4061 /** 4062 * cgroup_add_legacy_cftypes - add an array of cftypes for legacy hierarchies 4063 * @ss: target cgroup subsystem 4064 * @cfts: zero-length name terminated array of cftypes 4065 * 4066 * Similar to cgroup_add_cftypes() but the added files are only used for 4067 * the legacy hierarchies. 4068 */ 4069 int cgroup_add_legacy_cftypes(struct cgroup_subsys *ss, struct cftype *cfts) 4070 { 4071 struct cftype *cft; 4072 4073 for (cft = cfts; cft && cft->name[0] != '\0'; cft++) 4074 cft->flags |= __CFTYPE_NOT_ON_DFL; 4075 return cgroup_add_cftypes(ss, cfts); 4076 } 4077 4078 /** 4079 * cgroup_file_notify - generate a file modified event for a cgroup_file 4080 * @cfile: target cgroup_file 4081 * 4082 * @cfile must have been obtained by setting cftype->file_offset. 4083 */ 4084 void cgroup_file_notify(struct cgroup_file *cfile) 4085 { 4086 unsigned long flags; 4087 4088 spin_lock_irqsave(&cgroup_file_kn_lock, flags); 4089 if (cfile->kn) { 4090 unsigned long last = cfile->notified_at; 4091 unsigned long next = last + CGROUP_FILE_NOTIFY_MIN_INTV; 4092 4093 if (time_in_range(jiffies, last, next)) { 4094 timer_reduce(&cfile->notify_timer, next); 4095 } else { 4096 kernfs_notify(cfile->kn); 4097 cfile->notified_at = jiffies; 4098 } 4099 } 4100 spin_unlock_irqrestore(&cgroup_file_kn_lock, flags); 4101 } 4102 4103 /** 4104 * css_next_child - find the next child of a given css 4105 * @pos: the current position (%NULL to initiate traversal) 4106 * @parent: css whose children to walk 4107 * 4108 * This function returns the next child of @parent and should be called 4109 * under either cgroup_mutex or RCU read lock. The only requirement is 4110 * that @parent and @pos are accessible. The next sibling is guaranteed to 4111 * be returned regardless of their states. 4112 * 4113 * If a subsystem synchronizes ->css_online() and the start of iteration, a 4114 * css which finished ->css_online() is guaranteed to be visible in the 4115 * future iterations and will stay visible until the last reference is put. 4116 * A css which hasn't finished ->css_online() or already finished 4117 * ->css_offline() may show up during traversal. It's each subsystem's 4118 * responsibility to synchronize against on/offlining. 4119 */ 4120 struct cgroup_subsys_state *css_next_child(struct cgroup_subsys_state *pos, 4121 struct cgroup_subsys_state *parent) 4122 { 4123 struct cgroup_subsys_state *next; 4124 4125 cgroup_assert_mutex_or_rcu_locked(); 4126 4127 /* 4128 * @pos could already have been unlinked from the sibling list. 4129 * Once a cgroup is removed, its ->sibling.next is no longer 4130 * updated when its next sibling changes. CSS_RELEASED is set when 4131 * @pos is taken off list, at which time its next pointer is valid, 4132 * and, as releases are serialized, the one pointed to by the next 4133 * pointer is guaranteed to not have started release yet. This 4134 * implies that if we observe !CSS_RELEASED on @pos in this RCU 4135 * critical section, the one pointed to by its next pointer is 4136 * guaranteed to not have finished its RCU grace period even if we 4137 * have dropped rcu_read_lock() in-between iterations. 4138 * 4139 * If @pos has CSS_RELEASED set, its next pointer can't be 4140 * dereferenced; however, as each css is given a monotonically 4141 * increasing unique serial number and always appended to the 4142 * sibling list, the next one can be found by walking the parent's 4143 * children until the first css with higher serial number than 4144 * @pos's. While this path can be slower, it happens iff iteration 4145 * races against release and the race window is very small. 4146 */ 4147 if (!pos) { 4148 next = list_entry_rcu(parent->children.next, struct cgroup_subsys_state, sibling); 4149 } else if (likely(!(pos->flags & CSS_RELEASED))) { 4150 next = list_entry_rcu(pos->sibling.next, struct cgroup_subsys_state, sibling); 4151 } else { 4152 list_for_each_entry_rcu(next, &parent->children, sibling, 4153 lockdep_is_held(&cgroup_mutex)) 4154 if (next->serial_nr > pos->serial_nr) 4155 break; 4156 } 4157 4158 /* 4159 * @next, if not pointing to the head, can be dereferenced and is 4160 * the next sibling. 4161 */ 4162 if (&next->sibling != &parent->children) 4163 return next; 4164 return NULL; 4165 } 4166 4167 /** 4168 * css_next_descendant_pre - find the next descendant for pre-order walk 4169 * @pos: the current position (%NULL to initiate traversal) 4170 * @root: css whose descendants to walk 4171 * 4172 * To be used by css_for_each_descendant_pre(). Find the next descendant 4173 * to visit for pre-order traversal of @root's descendants. @root is 4174 * included in the iteration and the first node to be visited. 4175 * 4176 * While this function requires cgroup_mutex or RCU read locking, it 4177 * doesn't require the whole traversal to be contained in a single critical 4178 * section. This function will return the correct next descendant as long 4179 * as both @pos and @root are accessible and @pos is a descendant of @root. 4180 * 4181 * If a subsystem synchronizes ->css_online() and the start of iteration, a 4182 * css which finished ->css_online() is guaranteed to be visible in the 4183 * future iterations and will stay visible until the last reference is put. 4184 * A css which hasn't finished ->css_online() or already finished 4185 * ->css_offline() may show up during traversal. It's each subsystem's 4186 * responsibility to synchronize against on/offlining. 4187 */ 4188 struct cgroup_subsys_state * 4189 css_next_descendant_pre(struct cgroup_subsys_state *pos, 4190 struct cgroup_subsys_state *root) 4191 { 4192 struct cgroup_subsys_state *next; 4193 4194 cgroup_assert_mutex_or_rcu_locked(); 4195 4196 /* if first iteration, visit @root */ 4197 if (!pos) 4198 return root; 4199 4200 /* visit the first child if exists */ 4201 next = css_next_child(NULL, pos); 4202 if (next) 4203 return next; 4204 4205 /* no child, visit my or the closest ancestor's next sibling */ 4206 while (pos != root) { 4207 next = css_next_child(pos, pos->parent); 4208 if (next) 4209 return next; 4210 pos = pos->parent; 4211 } 4212 4213 return NULL; 4214 } 4215 EXPORT_SYMBOL_GPL(css_next_descendant_pre); 4216 4217 /** 4218 * css_rightmost_descendant - return the rightmost descendant of a css 4219 * @pos: css of interest 4220 * 4221 * Return the rightmost descendant of @pos. If there's no descendant, @pos 4222 * is returned. This can be used during pre-order traversal to skip 4223 * subtree of @pos. 4224 * 4225 * While this function requires cgroup_mutex or RCU read locking, it 4226 * doesn't require the whole traversal to be contained in a single critical 4227 * section. This function will return the correct rightmost descendant as 4228 * long as @pos is accessible. 4229 */ 4230 struct cgroup_subsys_state * 4231 css_rightmost_descendant(struct cgroup_subsys_state *pos) 4232 { 4233 struct cgroup_subsys_state *last, *tmp; 4234 4235 cgroup_assert_mutex_or_rcu_locked(); 4236 4237 do { 4238 last = pos; 4239 /* ->prev isn't RCU safe, walk ->next till the end */ 4240 pos = NULL; 4241 css_for_each_child(tmp, last) 4242 pos = tmp; 4243 } while (pos); 4244 4245 return last; 4246 } 4247 4248 static struct cgroup_subsys_state * 4249 css_leftmost_descendant(struct cgroup_subsys_state *pos) 4250 { 4251 struct cgroup_subsys_state *last; 4252 4253 do { 4254 last = pos; 4255 pos = css_next_child(NULL, pos); 4256 } while (pos); 4257 4258 return last; 4259 } 4260 4261 /** 4262 * css_next_descendant_post - find the next descendant for post-order walk 4263 * @pos: the current position (%NULL to initiate traversal) 4264 * @root: css whose descendants to walk 4265 * 4266 * To be used by css_for_each_descendant_post(). Find the next descendant 4267 * to visit for post-order traversal of @root's descendants. @root is 4268 * included in the iteration and the last node to be visited. 4269 * 4270 * While this function requires cgroup_mutex or RCU read locking, it 4271 * doesn't require the whole traversal to be contained in a single critical 4272 * section. This function will return the correct next descendant as long 4273 * as both @pos and @cgroup are accessible and @pos is a descendant of 4274 * @cgroup. 4275 * 4276 * If a subsystem synchronizes ->css_online() and the start of iteration, a 4277 * css which finished ->css_online() is guaranteed to be visible in the 4278 * future iterations and will stay visible until the last reference is put. 4279 * A css which hasn't finished ->css_online() or already finished 4280 * ->css_offline() may show up during traversal. It's each subsystem's 4281 * responsibility to synchronize against on/offlining. 4282 */ 4283 struct cgroup_subsys_state * 4284 css_next_descendant_post(struct cgroup_subsys_state *pos, 4285 struct cgroup_subsys_state *root) 4286 { 4287 struct cgroup_subsys_state *next; 4288 4289 cgroup_assert_mutex_or_rcu_locked(); 4290 4291 /* if first iteration, visit leftmost descendant which may be @root */ 4292 if (!pos) 4293 return css_leftmost_descendant(root); 4294 4295 /* if we visited @root, we're done */ 4296 if (pos == root) 4297 return NULL; 4298 4299 /* if there's an unvisited sibling, visit its leftmost descendant */ 4300 next = css_next_child(pos, pos->parent); 4301 if (next) 4302 return css_leftmost_descendant(next); 4303 4304 /* no sibling left, visit parent */ 4305 return pos->parent; 4306 } 4307 4308 /** 4309 * css_has_online_children - does a css have online children 4310 * @css: the target css 4311 * 4312 * Returns %true if @css has any online children; otherwise, %false. This 4313 * function can be called from any context but the caller is responsible 4314 * for synchronizing against on/offlining as necessary. 4315 */ 4316 bool css_has_online_children(struct cgroup_subsys_state *css) 4317 { 4318 struct cgroup_subsys_state *child; 4319 bool ret = false; 4320 4321 rcu_read_lock(); 4322 css_for_each_child(child, css) { 4323 if (child->flags & CSS_ONLINE) { 4324 ret = true; 4325 break; 4326 } 4327 } 4328 rcu_read_unlock(); 4329 return ret; 4330 } 4331 4332 static struct css_set *css_task_iter_next_css_set(struct css_task_iter *it) 4333 { 4334 struct list_head *l; 4335 struct cgrp_cset_link *link; 4336 struct css_set *cset; 4337 4338 lockdep_assert_held(&css_set_lock); 4339 4340 /* find the next threaded cset */ 4341 if (it->tcset_pos) { 4342 l = it->tcset_pos->next; 4343 4344 if (l != it->tcset_head) { 4345 it->tcset_pos = l; 4346 return container_of(l, struct css_set, 4347 threaded_csets_node); 4348 } 4349 4350 it->tcset_pos = NULL; 4351 } 4352 4353 /* find the next cset */ 4354 l = it->cset_pos; 4355 l = l->next; 4356 if (l == it->cset_head) { 4357 it->cset_pos = NULL; 4358 return NULL; 4359 } 4360 4361 if (it->ss) { 4362 cset = container_of(l, struct css_set, e_cset_node[it->ss->id]); 4363 } else { 4364 link = list_entry(l, struct cgrp_cset_link, cset_link); 4365 cset = link->cset; 4366 } 4367 4368 it->cset_pos = l; 4369 4370 /* initialize threaded css_set walking */ 4371 if (it->flags & CSS_TASK_ITER_THREADED) { 4372 if (it->cur_dcset) 4373 put_css_set_locked(it->cur_dcset); 4374 it->cur_dcset = cset; 4375 get_css_set(cset); 4376 4377 it->tcset_head = &cset->threaded_csets; 4378 it->tcset_pos = &cset->threaded_csets; 4379 } 4380 4381 return cset; 4382 } 4383 4384 /** 4385 * css_task_iter_advance_css_set - advance a task iterator to the next css_set 4386 * @it: the iterator to advance 4387 * 4388 * Advance @it to the next css_set to walk. 4389 */ 4390 static void css_task_iter_advance_css_set(struct css_task_iter *it) 4391 { 4392 struct css_set *cset; 4393 4394 lockdep_assert_held(&css_set_lock); 4395 4396 /* Advance to the next non-empty css_set and find first non-empty tasks list*/ 4397 while ((cset = css_task_iter_next_css_set(it))) { 4398 if (!list_empty(&cset->tasks)) { 4399 it->cur_tasks_head = &cset->tasks; 4400 break; 4401 } else if (!list_empty(&cset->mg_tasks)) { 4402 it->cur_tasks_head = &cset->mg_tasks; 4403 break; 4404 } else if (!list_empty(&cset->dying_tasks)) { 4405 it->cur_tasks_head = &cset->dying_tasks; 4406 break; 4407 } 4408 } 4409 if (!cset) { 4410 it->task_pos = NULL; 4411 return; 4412 } 4413 it->task_pos = it->cur_tasks_head->next; 4414 4415 /* 4416 * We don't keep css_sets locked across iteration steps and thus 4417 * need to take steps to ensure that iteration can be resumed after 4418 * the lock is re-acquired. Iteration is performed at two levels - 4419 * css_sets and tasks in them. 4420 * 4421 * Once created, a css_set never leaves its cgroup lists, so a 4422 * pinned css_set is guaranteed to stay put and we can resume 4423 * iteration afterwards. 4424 * 4425 * Tasks may leave @cset across iteration steps. This is resolved 4426 * by registering each iterator with the css_set currently being 4427 * walked and making css_set_move_task() advance iterators whose 4428 * next task is leaving. 4429 */ 4430 if (it->cur_cset) { 4431 list_del(&it->iters_node); 4432 put_css_set_locked(it->cur_cset); 4433 } 4434 get_css_set(cset); 4435 it->cur_cset = cset; 4436 list_add(&it->iters_node, &cset->task_iters); 4437 } 4438 4439 static void css_task_iter_skip(struct css_task_iter *it, 4440 struct task_struct *task) 4441 { 4442 lockdep_assert_held(&css_set_lock); 4443 4444 if (it->task_pos == &task->cg_list) { 4445 it->task_pos = it->task_pos->next; 4446 it->flags |= CSS_TASK_ITER_SKIPPED; 4447 } 4448 } 4449 4450 static void css_task_iter_advance(struct css_task_iter *it) 4451 { 4452 struct task_struct *task; 4453 4454 lockdep_assert_held(&css_set_lock); 4455 repeat: 4456 if (it->task_pos) { 4457 /* 4458 * Advance iterator to find next entry. We go through cset 4459 * tasks, mg_tasks and dying_tasks, when consumed we move onto 4460 * the next cset. 4461 */ 4462 if (it->flags & CSS_TASK_ITER_SKIPPED) 4463 it->flags &= ~CSS_TASK_ITER_SKIPPED; 4464 else 4465 it->task_pos = it->task_pos->next; 4466 4467 if (it->task_pos == &it->cur_cset->tasks) { 4468 it->cur_tasks_head = &it->cur_cset->mg_tasks; 4469 it->task_pos = it->cur_tasks_head->next; 4470 } 4471 if (it->task_pos == &it->cur_cset->mg_tasks) { 4472 it->cur_tasks_head = &it->cur_cset->dying_tasks; 4473 it->task_pos = it->cur_tasks_head->next; 4474 } 4475 if (it->task_pos == &it->cur_cset->dying_tasks) 4476 css_task_iter_advance_css_set(it); 4477 } else { 4478 /* called from start, proceed to the first cset */ 4479 css_task_iter_advance_css_set(it); 4480 } 4481 4482 if (!it->task_pos) 4483 return; 4484 4485 task = list_entry(it->task_pos, struct task_struct, cg_list); 4486 4487 if (it->flags & CSS_TASK_ITER_PROCS) { 4488 /* if PROCS, skip over tasks which aren't group leaders */ 4489 if (!thread_group_leader(task)) 4490 goto repeat; 4491 4492 /* and dying leaders w/o live member threads */ 4493 if (it->cur_tasks_head == &it->cur_cset->dying_tasks && 4494 !atomic_read(&task->signal->live)) 4495 goto repeat; 4496 } else { 4497 /* skip all dying ones */ 4498 if (it->cur_tasks_head == &it->cur_cset->dying_tasks) 4499 goto repeat; 4500 } 4501 } 4502 4503 /** 4504 * css_task_iter_start - initiate task iteration 4505 * @css: the css to walk tasks of 4506 * @flags: CSS_TASK_ITER_* flags 4507 * @it: the task iterator to use 4508 * 4509 * Initiate iteration through the tasks of @css. The caller can call 4510 * css_task_iter_next() to walk through the tasks until the function 4511 * returns NULL. On completion of iteration, css_task_iter_end() must be 4512 * called. 4513 */ 4514 void css_task_iter_start(struct cgroup_subsys_state *css, unsigned int flags, 4515 struct css_task_iter *it) 4516 { 4517 memset(it, 0, sizeof(*it)); 4518 4519 spin_lock_irq(&css_set_lock); 4520 4521 it->ss = css->ss; 4522 it->flags = flags; 4523 4524 if (it->ss) 4525 it->cset_pos = &css->cgroup->e_csets[css->ss->id]; 4526 else 4527 it->cset_pos = &css->cgroup->cset_links; 4528 4529 it->cset_head = it->cset_pos; 4530 4531 css_task_iter_advance(it); 4532 4533 spin_unlock_irq(&css_set_lock); 4534 } 4535 4536 /** 4537 * css_task_iter_next - return the next task for the iterator 4538 * @it: the task iterator being iterated 4539 * 4540 * The "next" function for task iteration. @it should have been 4541 * initialized via css_task_iter_start(). Returns NULL when the iteration 4542 * reaches the end. 4543 */ 4544 struct task_struct *css_task_iter_next(struct css_task_iter *it) 4545 { 4546 if (it->cur_task) { 4547 put_task_struct(it->cur_task); 4548 it->cur_task = NULL; 4549 } 4550 4551 spin_lock_irq(&css_set_lock); 4552 4553 /* @it may be half-advanced by skips, finish advancing */ 4554 if (it->flags & CSS_TASK_ITER_SKIPPED) 4555 css_task_iter_advance(it); 4556 4557 if (it->task_pos) { 4558 it->cur_task = list_entry(it->task_pos, struct task_struct, 4559 cg_list); 4560 get_task_struct(it->cur_task); 4561 css_task_iter_advance(it); 4562 } 4563 4564 spin_unlock_irq(&css_set_lock); 4565 4566 return it->cur_task; 4567 } 4568 4569 /** 4570 * css_task_iter_end - finish task iteration 4571 * @it: the task iterator to finish 4572 * 4573 * Finish task iteration started by css_task_iter_start(). 4574 */ 4575 void css_task_iter_end(struct css_task_iter *it) 4576 { 4577 if (it->cur_cset) { 4578 spin_lock_irq(&css_set_lock); 4579 list_del(&it->iters_node); 4580 put_css_set_locked(it->cur_cset); 4581 spin_unlock_irq(&css_set_lock); 4582 } 4583 4584 if (it->cur_dcset) 4585 put_css_set(it->cur_dcset); 4586 4587 if (it->cur_task) 4588 put_task_struct(it->cur_task); 4589 } 4590 4591 static void cgroup_procs_release(struct kernfs_open_file *of) 4592 { 4593 if (of->priv) { 4594 css_task_iter_end(of->priv); 4595 kfree(of->priv); 4596 } 4597 } 4598 4599 static void *cgroup_procs_next(struct seq_file *s, void *v, loff_t *pos) 4600 { 4601 struct kernfs_open_file *of = s->private; 4602 struct css_task_iter *it = of->priv; 4603 4604 if (pos) 4605 (*pos)++; 4606 4607 return css_task_iter_next(it); 4608 } 4609 4610 static void *__cgroup_procs_start(struct seq_file *s, loff_t *pos, 4611 unsigned int iter_flags) 4612 { 4613 struct kernfs_open_file *of = s->private; 4614 struct cgroup *cgrp = seq_css(s)->cgroup; 4615 struct css_task_iter *it = of->priv; 4616 4617 /* 4618 * When a seq_file is seeked, it's always traversed sequentially 4619 * from position 0, so we can simply keep iterating on !0 *pos. 4620 */ 4621 if (!it) { 4622 if (WARN_ON_ONCE((*pos))) 4623 return ERR_PTR(-EINVAL); 4624 4625 it = kzalloc(sizeof(*it), GFP_KERNEL); 4626 if (!it) 4627 return ERR_PTR(-ENOMEM); 4628 of->priv = it; 4629 css_task_iter_start(&cgrp->self, iter_flags, it); 4630 } else if (!(*pos)) { 4631 css_task_iter_end(it); 4632 css_task_iter_start(&cgrp->self, iter_flags, it); 4633 } else 4634 return it->cur_task; 4635 4636 return cgroup_procs_next(s, NULL, NULL); 4637 } 4638 4639 static void *cgroup_procs_start(struct seq_file *s, loff_t *pos) 4640 { 4641 struct cgroup *cgrp = seq_css(s)->cgroup; 4642 4643 /* 4644 * All processes of a threaded subtree belong to the domain cgroup 4645 * of the subtree. Only threads can be distributed across the 4646 * subtree. Reject reads on cgroup.procs in the subtree proper. 4647 * They're always empty anyway. 4648 */ 4649 if (cgroup_is_threaded(cgrp)) 4650 return ERR_PTR(-EOPNOTSUPP); 4651 4652 return __cgroup_procs_start(s, pos, CSS_TASK_ITER_PROCS | 4653 CSS_TASK_ITER_THREADED); 4654 } 4655 4656 static int cgroup_procs_show(struct seq_file *s, void *v) 4657 { 4658 seq_printf(s, "%d\n", task_pid_vnr(v)); 4659 return 0; 4660 } 4661 4662 static int cgroup_may_write(const struct cgroup *cgrp, struct super_block *sb) 4663 { 4664 int ret; 4665 struct inode *inode; 4666 4667 lockdep_assert_held(&cgroup_mutex); 4668 4669 inode = kernfs_get_inode(sb, cgrp->procs_file.kn); 4670 if (!inode) 4671 return -ENOMEM; 4672 4673 ret = inode_permission(inode, MAY_WRITE); 4674 iput(inode); 4675 return ret; 4676 } 4677 4678 static int cgroup_procs_write_permission(struct cgroup *src_cgrp, 4679 struct cgroup *dst_cgrp, 4680 struct super_block *sb) 4681 { 4682 struct cgroup_namespace *ns = current->nsproxy->cgroup_ns; 4683 struct cgroup *com_cgrp = src_cgrp; 4684 int ret; 4685 4686 lockdep_assert_held(&cgroup_mutex); 4687 4688 /* find the common ancestor */ 4689 while (!cgroup_is_descendant(dst_cgrp, com_cgrp)) 4690 com_cgrp = cgroup_parent(com_cgrp); 4691 4692 /* %current should be authorized to migrate to the common ancestor */ 4693 ret = cgroup_may_write(com_cgrp, sb); 4694 if (ret) 4695 return ret; 4696 4697 /* 4698 * If namespaces are delegation boundaries, %current must be able 4699 * to see both source and destination cgroups from its namespace. 4700 */ 4701 if ((cgrp_dfl_root.flags & CGRP_ROOT_NS_DELEGATE) && 4702 (!cgroup_is_descendant(src_cgrp, ns->root_cset->dfl_cgrp) || 4703 !cgroup_is_descendant(dst_cgrp, ns->root_cset->dfl_cgrp))) 4704 return -ENOENT; 4705 4706 return 0; 4707 } 4708 4709 static int cgroup_attach_permissions(struct cgroup *src_cgrp, 4710 struct cgroup *dst_cgrp, 4711 struct super_block *sb, bool threadgroup) 4712 { 4713 int ret = 0; 4714 4715 ret = cgroup_procs_write_permission(src_cgrp, dst_cgrp, sb); 4716 if (ret) 4717 return ret; 4718 4719 ret = cgroup_migrate_vet_dst(dst_cgrp); 4720 if (ret) 4721 return ret; 4722 4723 if (!threadgroup && (src_cgrp->dom_cgrp != dst_cgrp->dom_cgrp)) 4724 ret = -EOPNOTSUPP; 4725 4726 return ret; 4727 } 4728 4729 static ssize_t cgroup_procs_write(struct kernfs_open_file *of, 4730 char *buf, size_t nbytes, loff_t off) 4731 { 4732 struct cgroup *src_cgrp, *dst_cgrp; 4733 struct task_struct *task; 4734 ssize_t ret; 4735 bool locked; 4736 4737 dst_cgrp = cgroup_kn_lock_live(of->kn, false); 4738 if (!dst_cgrp) 4739 return -ENODEV; 4740 4741 task = cgroup_procs_write_start(buf, true, &locked); 4742 ret = PTR_ERR_OR_ZERO(task); 4743 if (ret) 4744 goto out_unlock; 4745 4746 /* find the source cgroup */ 4747 spin_lock_irq(&css_set_lock); 4748 src_cgrp = task_cgroup_from_root(task, &cgrp_dfl_root); 4749 spin_unlock_irq(&css_set_lock); 4750 4751 ret = cgroup_attach_permissions(src_cgrp, dst_cgrp, 4752 of->file->f_path.dentry->d_sb, true); 4753 if (ret) 4754 goto out_finish; 4755 4756 ret = cgroup_attach_task(dst_cgrp, task, true); 4757 4758 out_finish: 4759 cgroup_procs_write_finish(task, locked); 4760 out_unlock: 4761 cgroup_kn_unlock(of->kn); 4762 4763 return ret ?: nbytes; 4764 } 4765 4766 static void *cgroup_threads_start(struct seq_file *s, loff_t *pos) 4767 { 4768 return __cgroup_procs_start(s, pos, 0); 4769 } 4770 4771 static ssize_t cgroup_threads_write(struct kernfs_open_file *of, 4772 char *buf, size_t nbytes, loff_t off) 4773 { 4774 struct cgroup *src_cgrp, *dst_cgrp; 4775 struct task_struct *task; 4776 ssize_t ret; 4777 bool locked; 4778 4779 buf = strstrip(buf); 4780 4781 dst_cgrp = cgroup_kn_lock_live(of->kn, false); 4782 if (!dst_cgrp) 4783 return -ENODEV; 4784 4785 task = cgroup_procs_write_start(buf, false, &locked); 4786 ret = PTR_ERR_OR_ZERO(task); 4787 if (ret) 4788 goto out_unlock; 4789 4790 /* find the source cgroup */ 4791 spin_lock_irq(&css_set_lock); 4792 src_cgrp = task_cgroup_from_root(task, &cgrp_dfl_root); 4793 spin_unlock_irq(&css_set_lock); 4794 4795 /* thread migrations follow the cgroup.procs delegation rule */ 4796 ret = cgroup_attach_permissions(src_cgrp, dst_cgrp, 4797 of->file->f_path.dentry->d_sb, false); 4798 if (ret) 4799 goto out_finish; 4800 4801 ret = cgroup_attach_task(dst_cgrp, task, false); 4802 4803 out_finish: 4804 cgroup_procs_write_finish(task, locked); 4805 out_unlock: 4806 cgroup_kn_unlock(of->kn); 4807 4808 return ret ?: nbytes; 4809 } 4810 4811 /* cgroup core interface files for the default hierarchy */ 4812 static struct cftype cgroup_base_files[] = { 4813 { 4814 .name = "cgroup.type", 4815 .flags = CFTYPE_NOT_ON_ROOT, 4816 .seq_show = cgroup_type_show, 4817 .write = cgroup_type_write, 4818 }, 4819 { 4820 .name = "cgroup.procs", 4821 .flags = CFTYPE_NS_DELEGATABLE, 4822 .file_offset = offsetof(struct cgroup, procs_file), 4823 .release = cgroup_procs_release, 4824 .seq_start = cgroup_procs_start, 4825 .seq_next = cgroup_procs_next, 4826 .seq_show = cgroup_procs_show, 4827 .write = cgroup_procs_write, 4828 }, 4829 { 4830 .name = "cgroup.threads", 4831 .flags = CFTYPE_NS_DELEGATABLE, 4832 .release = cgroup_procs_release, 4833 .seq_start = cgroup_threads_start, 4834 .seq_next = cgroup_procs_next, 4835 .seq_show = cgroup_procs_show, 4836 .write = cgroup_threads_write, 4837 }, 4838 { 4839 .name = "cgroup.controllers", 4840 .seq_show = cgroup_controllers_show, 4841 }, 4842 { 4843 .name = "cgroup.subtree_control", 4844 .flags = CFTYPE_NS_DELEGATABLE, 4845 .seq_show = cgroup_subtree_control_show, 4846 .write = cgroup_subtree_control_write, 4847 }, 4848 { 4849 .name = "cgroup.events", 4850 .flags = CFTYPE_NOT_ON_ROOT, 4851 .file_offset = offsetof(struct cgroup, events_file), 4852 .seq_show = cgroup_events_show, 4853 }, 4854 { 4855 .name = "cgroup.max.descendants", 4856 .seq_show = cgroup_max_descendants_show, 4857 .write = cgroup_max_descendants_write, 4858 }, 4859 { 4860 .name = "cgroup.max.depth", 4861 .seq_show = cgroup_max_depth_show, 4862 .write = cgroup_max_depth_write, 4863 }, 4864 { 4865 .name = "cgroup.stat", 4866 .seq_show = cgroup_stat_show, 4867 }, 4868 { 4869 .name = "cgroup.freeze", 4870 .flags = CFTYPE_NOT_ON_ROOT, 4871 .seq_show = cgroup_freeze_show, 4872 .write = cgroup_freeze_write, 4873 }, 4874 { 4875 .name = "cpu.stat", 4876 .seq_show = cpu_stat_show, 4877 }, 4878 #ifdef CONFIG_PSI 4879 { 4880 .name = "io.pressure", 4881 .seq_show = cgroup_io_pressure_show, 4882 .write = cgroup_io_pressure_write, 4883 .poll = cgroup_pressure_poll, 4884 .release = cgroup_pressure_release, 4885 }, 4886 { 4887 .name = "memory.pressure", 4888 .seq_show = cgroup_memory_pressure_show, 4889 .write = cgroup_memory_pressure_write, 4890 .poll = cgroup_pressure_poll, 4891 .release = cgroup_pressure_release, 4892 }, 4893 { 4894 .name = "cpu.pressure", 4895 .seq_show = cgroup_cpu_pressure_show, 4896 .write = cgroup_cpu_pressure_write, 4897 .poll = cgroup_pressure_poll, 4898 .release = cgroup_pressure_release, 4899 }, 4900 #endif /* CONFIG_PSI */ 4901 { } /* terminate */ 4902 }; 4903 4904 /* 4905 * css destruction is four-stage process. 4906 * 4907 * 1. Destruction starts. Killing of the percpu_ref is initiated. 4908 * Implemented in kill_css(). 4909 * 4910 * 2. When the percpu_ref is confirmed to be visible as killed on all CPUs 4911 * and thus css_tryget_online() is guaranteed to fail, the css can be 4912 * offlined by invoking offline_css(). After offlining, the base ref is 4913 * put. Implemented in css_killed_work_fn(). 4914 * 4915 * 3. When the percpu_ref reaches zero, the only possible remaining 4916 * accessors are inside RCU read sections. css_release() schedules the 4917 * RCU callback. 4918 * 4919 * 4. After the grace period, the css can be freed. Implemented in 4920 * css_free_work_fn(). 4921 * 4922 * It is actually hairier because both step 2 and 4 require process context 4923 * and thus involve punting to css->destroy_work adding two additional 4924 * steps to the already complex sequence. 4925 */ 4926 static void css_free_rwork_fn(struct work_struct *work) 4927 { 4928 struct cgroup_subsys_state *css = container_of(to_rcu_work(work), 4929 struct cgroup_subsys_state, destroy_rwork); 4930 struct cgroup_subsys *ss = css->ss; 4931 struct cgroup *cgrp = css->cgroup; 4932 4933 percpu_ref_exit(&css->refcnt); 4934 4935 if (ss) { 4936 /* css free path */ 4937 struct cgroup_subsys_state *parent = css->parent; 4938 int id = css->id; 4939 4940 ss->css_free(css); 4941 cgroup_idr_remove(&ss->css_idr, id); 4942 cgroup_put(cgrp); 4943 4944 if (parent) 4945 css_put(parent); 4946 } else { 4947 /* cgroup free path */ 4948 atomic_dec(&cgrp->root->nr_cgrps); 4949 cgroup1_pidlist_destroy_all(cgrp); 4950 cancel_work_sync(&cgrp->release_agent_work); 4951 4952 if (cgroup_parent(cgrp)) { 4953 /* 4954 * We get a ref to the parent, and put the ref when 4955 * this cgroup is being freed, so it's guaranteed 4956 * that the parent won't be destroyed before its 4957 * children. 4958 */ 4959 cgroup_put(cgroup_parent(cgrp)); 4960 kernfs_put(cgrp->kn); 4961 psi_cgroup_free(cgrp); 4962 if (cgroup_on_dfl(cgrp)) 4963 cgroup_rstat_exit(cgrp); 4964 kfree(cgrp); 4965 } else { 4966 /* 4967 * This is root cgroup's refcnt reaching zero, 4968 * which indicates that the root should be 4969 * released. 4970 */ 4971 cgroup_destroy_root(cgrp->root); 4972 } 4973 } 4974 } 4975 4976 static void css_release_work_fn(struct work_struct *work) 4977 { 4978 struct cgroup_subsys_state *css = 4979 container_of(work, struct cgroup_subsys_state, destroy_work); 4980 struct cgroup_subsys *ss = css->ss; 4981 struct cgroup *cgrp = css->cgroup; 4982 4983 mutex_lock(&cgroup_mutex); 4984 4985 css->flags |= CSS_RELEASED; 4986 list_del_rcu(&css->sibling); 4987 4988 if (ss) { 4989 /* css release path */ 4990 if (!list_empty(&css->rstat_css_node)) { 4991 cgroup_rstat_flush(cgrp); 4992 list_del_rcu(&css->rstat_css_node); 4993 } 4994 4995 cgroup_idr_replace(&ss->css_idr, NULL, css->id); 4996 if (ss->css_released) 4997 ss->css_released(css); 4998 } else { 4999 struct cgroup *tcgrp; 5000 5001 /* cgroup release path */ 5002 TRACE_CGROUP_PATH(release, cgrp); 5003 5004 if (cgroup_on_dfl(cgrp)) 5005 cgroup_rstat_flush(cgrp); 5006 5007 spin_lock_irq(&css_set_lock); 5008 for (tcgrp = cgroup_parent(cgrp); tcgrp; 5009 tcgrp = cgroup_parent(tcgrp)) 5010 tcgrp->nr_dying_descendants--; 5011 spin_unlock_irq(&css_set_lock); 5012 5013 /* 5014 * There are two control paths which try to determine 5015 * cgroup from dentry without going through kernfs - 5016 * cgroupstats_build() and css_tryget_online_from_dir(). 5017 * Those are supported by RCU protecting clearing of 5018 * cgrp->kn->priv backpointer. 5019 */ 5020 if (cgrp->kn) 5021 RCU_INIT_POINTER(*(void __rcu __force **)&cgrp->kn->priv, 5022 NULL); 5023 } 5024 5025 mutex_unlock(&cgroup_mutex); 5026 5027 INIT_RCU_WORK(&css->destroy_rwork, css_free_rwork_fn); 5028 queue_rcu_work(cgroup_destroy_wq, &css->destroy_rwork); 5029 } 5030 5031 static void css_release(struct percpu_ref *ref) 5032 { 5033 struct cgroup_subsys_state *css = 5034 container_of(ref, struct cgroup_subsys_state, refcnt); 5035 5036 INIT_WORK(&css->destroy_work, css_release_work_fn); 5037 queue_work(cgroup_destroy_wq, &css->destroy_work); 5038 } 5039 5040 static void init_and_link_css(struct cgroup_subsys_state *css, 5041 struct cgroup_subsys *ss, struct cgroup *cgrp) 5042 { 5043 lockdep_assert_held(&cgroup_mutex); 5044 5045 cgroup_get_live(cgrp); 5046 5047 memset(css, 0, sizeof(*css)); 5048 css->cgroup = cgrp; 5049 css->ss = ss; 5050 css->id = -1; 5051 INIT_LIST_HEAD(&css->sibling); 5052 INIT_LIST_HEAD(&css->children); 5053 INIT_LIST_HEAD(&css->rstat_css_node); 5054 css->serial_nr = css_serial_nr_next++; 5055 atomic_set(&css->online_cnt, 0); 5056 5057 if (cgroup_parent(cgrp)) { 5058 css->parent = cgroup_css(cgroup_parent(cgrp), ss); 5059 css_get(css->parent); 5060 } 5061 5062 if (cgroup_on_dfl(cgrp) && ss->css_rstat_flush) 5063 list_add_rcu(&css->rstat_css_node, &cgrp->rstat_css_list); 5064 5065 BUG_ON(cgroup_css(cgrp, ss)); 5066 } 5067 5068 /* invoke ->css_online() on a new CSS and mark it online if successful */ 5069 static int online_css(struct cgroup_subsys_state *css) 5070 { 5071 struct cgroup_subsys *ss = css->ss; 5072 int ret = 0; 5073 5074 lockdep_assert_held(&cgroup_mutex); 5075 5076 if (ss->css_online) 5077 ret = ss->css_online(css); 5078 if (!ret) { 5079 css->flags |= CSS_ONLINE; 5080 rcu_assign_pointer(css->cgroup->subsys[ss->id], css); 5081 5082 atomic_inc(&css->online_cnt); 5083 if (css->parent) 5084 atomic_inc(&css->parent->online_cnt); 5085 } 5086 return ret; 5087 } 5088 5089 /* if the CSS is online, invoke ->css_offline() on it and mark it offline */ 5090 static void offline_css(struct cgroup_subsys_state *css) 5091 { 5092 struct cgroup_subsys *ss = css->ss; 5093 5094 lockdep_assert_held(&cgroup_mutex); 5095 5096 if (!(css->flags & CSS_ONLINE)) 5097 return; 5098 5099 if (ss->css_offline) 5100 ss->css_offline(css); 5101 5102 css->flags &= ~CSS_ONLINE; 5103 RCU_INIT_POINTER(css->cgroup->subsys[ss->id], NULL); 5104 5105 wake_up_all(&css->cgroup->offline_waitq); 5106 } 5107 5108 /** 5109 * css_create - create a cgroup_subsys_state 5110 * @cgrp: the cgroup new css will be associated with 5111 * @ss: the subsys of new css 5112 * 5113 * Create a new css associated with @cgrp - @ss pair. On success, the new 5114 * css is online and installed in @cgrp. This function doesn't create the 5115 * interface files. Returns 0 on success, -errno on failure. 5116 */ 5117 static struct cgroup_subsys_state *css_create(struct cgroup *cgrp, 5118 struct cgroup_subsys *ss) 5119 { 5120 struct cgroup *parent = cgroup_parent(cgrp); 5121 struct cgroup_subsys_state *parent_css = cgroup_css(parent, ss); 5122 struct cgroup_subsys_state *css; 5123 int err; 5124 5125 lockdep_assert_held(&cgroup_mutex); 5126 5127 css = ss->css_alloc(parent_css); 5128 if (!css) 5129 css = ERR_PTR(-ENOMEM); 5130 if (IS_ERR(css)) 5131 return css; 5132 5133 init_and_link_css(css, ss, cgrp); 5134 5135 err = percpu_ref_init(&css->refcnt, css_release, 0, GFP_KERNEL); 5136 if (err) 5137 goto err_free_css; 5138 5139 err = cgroup_idr_alloc(&ss->css_idr, NULL, 2, 0, GFP_KERNEL); 5140 if (err < 0) 5141 goto err_free_css; 5142 css->id = err; 5143 5144 /* @css is ready to be brought online now, make it visible */ 5145 list_add_tail_rcu(&css->sibling, &parent_css->children); 5146 cgroup_idr_replace(&ss->css_idr, css, css->id); 5147 5148 err = online_css(css); 5149 if (err) 5150 goto err_list_del; 5151 5152 return css; 5153 5154 err_list_del: 5155 list_del_rcu(&css->sibling); 5156 err_free_css: 5157 list_del_rcu(&css->rstat_css_node); 5158 INIT_RCU_WORK(&css->destroy_rwork, css_free_rwork_fn); 5159 queue_rcu_work(cgroup_destroy_wq, &css->destroy_rwork); 5160 return ERR_PTR(err); 5161 } 5162 5163 /* 5164 * The returned cgroup is fully initialized including its control mask, but 5165 * it isn't associated with its kernfs_node and doesn't have the control 5166 * mask applied. 5167 */ 5168 static struct cgroup *cgroup_create(struct cgroup *parent, const char *name, 5169 umode_t mode) 5170 { 5171 struct cgroup_root *root = parent->root; 5172 struct cgroup *cgrp, *tcgrp; 5173 struct kernfs_node *kn; 5174 int level = parent->level + 1; 5175 int ret; 5176 5177 /* allocate the cgroup and its ID, 0 is reserved for the root */ 5178 cgrp = kzalloc(struct_size(cgrp, ancestor_ids, (level + 1)), 5179 GFP_KERNEL); 5180 if (!cgrp) 5181 return ERR_PTR(-ENOMEM); 5182 5183 ret = percpu_ref_init(&cgrp->self.refcnt, css_release, 0, GFP_KERNEL); 5184 if (ret) 5185 goto out_free_cgrp; 5186 5187 if (cgroup_on_dfl(parent)) { 5188 ret = cgroup_rstat_init(cgrp); 5189 if (ret) 5190 goto out_cancel_ref; 5191 } 5192 5193 /* create the directory */ 5194 kn = kernfs_create_dir(parent->kn, name, mode, cgrp); 5195 if (IS_ERR(kn)) { 5196 ret = PTR_ERR(kn); 5197 goto out_stat_exit; 5198 } 5199 cgrp->kn = kn; 5200 5201 init_cgroup_housekeeping(cgrp); 5202 5203 cgrp->self.parent = &parent->self; 5204 cgrp->root = root; 5205 cgrp->level = level; 5206 5207 ret = psi_cgroup_alloc(cgrp); 5208 if (ret) 5209 goto out_kernfs_remove; 5210 5211 ret = cgroup_bpf_inherit(cgrp); 5212 if (ret) 5213 goto out_psi_free; 5214 5215 /* 5216 * New cgroup inherits effective freeze counter, and 5217 * if the parent has to be frozen, the child has too. 5218 */ 5219 cgrp->freezer.e_freeze = parent->freezer.e_freeze; 5220 if (cgrp->freezer.e_freeze) { 5221 /* 5222 * Set the CGRP_FREEZE flag, so when a process will be 5223 * attached to the child cgroup, it will become frozen. 5224 * At this point the new cgroup is unpopulated, so we can 5225 * consider it frozen immediately. 5226 */ 5227 set_bit(CGRP_FREEZE, &cgrp->flags); 5228 set_bit(CGRP_FROZEN, &cgrp->flags); 5229 } 5230 5231 spin_lock_irq(&css_set_lock); 5232 for (tcgrp = cgrp; tcgrp; tcgrp = cgroup_parent(tcgrp)) { 5233 cgrp->ancestor_ids[tcgrp->level] = cgroup_id(tcgrp); 5234 5235 if (tcgrp != cgrp) { 5236 tcgrp->nr_descendants++; 5237 5238 /* 5239 * If the new cgroup is frozen, all ancestor cgroups 5240 * get a new frozen descendant, but their state can't 5241 * change because of this. 5242 */ 5243 if (cgrp->freezer.e_freeze) 5244 tcgrp->freezer.nr_frozen_descendants++; 5245 } 5246 } 5247 spin_unlock_irq(&css_set_lock); 5248 5249 if (notify_on_release(parent)) 5250 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags); 5251 5252 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &parent->flags)) 5253 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags); 5254 5255 cgrp->self.serial_nr = css_serial_nr_next++; 5256 5257 /* allocation complete, commit to creation */ 5258 list_add_tail_rcu(&cgrp->self.sibling, &cgroup_parent(cgrp)->self.children); 5259 atomic_inc(&root->nr_cgrps); 5260 cgroup_get_live(parent); 5261 5262 /* 5263 * On the default hierarchy, a child doesn't automatically inherit 5264 * subtree_control from the parent. Each is configured manually. 5265 */ 5266 if (!cgroup_on_dfl(cgrp)) 5267 cgrp->subtree_control = cgroup_control(cgrp); 5268 5269 cgroup_propagate_control(cgrp); 5270 5271 return cgrp; 5272 5273 out_psi_free: 5274 psi_cgroup_free(cgrp); 5275 out_kernfs_remove: 5276 kernfs_remove(cgrp->kn); 5277 out_stat_exit: 5278 if (cgroup_on_dfl(parent)) 5279 cgroup_rstat_exit(cgrp); 5280 out_cancel_ref: 5281 percpu_ref_exit(&cgrp->self.refcnt); 5282 out_free_cgrp: 5283 kfree(cgrp); 5284 return ERR_PTR(ret); 5285 } 5286 5287 static bool cgroup_check_hierarchy_limits(struct cgroup *parent) 5288 { 5289 struct cgroup *cgroup; 5290 int ret = false; 5291 int level = 1; 5292 5293 lockdep_assert_held(&cgroup_mutex); 5294 5295 for (cgroup = parent; cgroup; cgroup = cgroup_parent(cgroup)) { 5296 if (cgroup->nr_descendants >= cgroup->max_descendants) 5297 goto fail; 5298 5299 if (level > cgroup->max_depth) 5300 goto fail; 5301 5302 level++; 5303 } 5304 5305 ret = true; 5306 fail: 5307 return ret; 5308 } 5309 5310 int cgroup_mkdir(struct kernfs_node *parent_kn, const char *name, umode_t mode) 5311 { 5312 struct cgroup *parent, *cgrp; 5313 int ret; 5314 5315 /* do not accept '\n' to prevent making /proc/<pid>/cgroup unparsable */ 5316 if (strchr(name, '\n')) 5317 return -EINVAL; 5318 5319 parent = cgroup_kn_lock_live(parent_kn, false); 5320 if (!parent) 5321 return -ENODEV; 5322 5323 if (!cgroup_check_hierarchy_limits(parent)) { 5324 ret = -EAGAIN; 5325 goto out_unlock; 5326 } 5327 5328 cgrp = cgroup_create(parent, name, mode); 5329 if (IS_ERR(cgrp)) { 5330 ret = PTR_ERR(cgrp); 5331 goto out_unlock; 5332 } 5333 5334 /* 5335 * This extra ref will be put in cgroup_free_fn() and guarantees 5336 * that @cgrp->kn is always accessible. 5337 */ 5338 kernfs_get(cgrp->kn); 5339 5340 ret = cgroup_kn_set_ugid(cgrp->kn); 5341 if (ret) 5342 goto out_destroy; 5343 5344 ret = css_populate_dir(&cgrp->self); 5345 if (ret) 5346 goto out_destroy; 5347 5348 ret = cgroup_apply_control_enable(cgrp); 5349 if (ret) 5350 goto out_destroy; 5351 5352 TRACE_CGROUP_PATH(mkdir, cgrp); 5353 5354 /* let's create and online css's */ 5355 kernfs_activate(cgrp->kn); 5356 5357 ret = 0; 5358 goto out_unlock; 5359 5360 out_destroy: 5361 cgroup_destroy_locked(cgrp); 5362 out_unlock: 5363 cgroup_kn_unlock(parent_kn); 5364 return ret; 5365 } 5366 5367 /* 5368 * This is called when the refcnt of a css is confirmed to be killed. 5369 * css_tryget_online() is now guaranteed to fail. Tell the subsystem to 5370 * initate destruction and put the css ref from kill_css(). 5371 */ 5372 static void css_killed_work_fn(struct work_struct *work) 5373 { 5374 struct cgroup_subsys_state *css = 5375 container_of(work, struct cgroup_subsys_state, destroy_work); 5376 5377 mutex_lock(&cgroup_mutex); 5378 5379 do { 5380 offline_css(css); 5381 css_put(css); 5382 /* @css can't go away while we're holding cgroup_mutex */ 5383 css = css->parent; 5384 } while (css && atomic_dec_and_test(&css->online_cnt)); 5385 5386 mutex_unlock(&cgroup_mutex); 5387 } 5388 5389 /* css kill confirmation processing requires process context, bounce */ 5390 static void css_killed_ref_fn(struct percpu_ref *ref) 5391 { 5392 struct cgroup_subsys_state *css = 5393 container_of(ref, struct cgroup_subsys_state, refcnt); 5394 5395 if (atomic_dec_and_test(&css->online_cnt)) { 5396 INIT_WORK(&css->destroy_work, css_killed_work_fn); 5397 queue_work(cgroup_destroy_wq, &css->destroy_work); 5398 } 5399 } 5400 5401 /** 5402 * kill_css - destroy a css 5403 * @css: css to destroy 5404 * 5405 * This function initiates destruction of @css by removing cgroup interface 5406 * files and putting its base reference. ->css_offline() will be invoked 5407 * asynchronously once css_tryget_online() is guaranteed to fail and when 5408 * the reference count reaches zero, @css will be released. 5409 */ 5410 static void kill_css(struct cgroup_subsys_state *css) 5411 { 5412 lockdep_assert_held(&cgroup_mutex); 5413 5414 if (css->flags & CSS_DYING) 5415 return; 5416 5417 css->flags |= CSS_DYING; 5418 5419 /* 5420 * This must happen before css is disassociated with its cgroup. 5421 * See seq_css() for details. 5422 */ 5423 css_clear_dir(css); 5424 5425 /* 5426 * Killing would put the base ref, but we need to keep it alive 5427 * until after ->css_offline(). 5428 */ 5429 css_get(css); 5430 5431 /* 5432 * cgroup core guarantees that, by the time ->css_offline() is 5433 * invoked, no new css reference will be given out via 5434 * css_tryget_online(). We can't simply call percpu_ref_kill() and 5435 * proceed to offlining css's because percpu_ref_kill() doesn't 5436 * guarantee that the ref is seen as killed on all CPUs on return. 5437 * 5438 * Use percpu_ref_kill_and_confirm() to get notifications as each 5439 * css is confirmed to be seen as killed on all CPUs. 5440 */ 5441 percpu_ref_kill_and_confirm(&css->refcnt, css_killed_ref_fn); 5442 } 5443 5444 /** 5445 * cgroup_destroy_locked - the first stage of cgroup destruction 5446 * @cgrp: cgroup to be destroyed 5447 * 5448 * css's make use of percpu refcnts whose killing latency shouldn't be 5449 * exposed to userland and are RCU protected. Also, cgroup core needs to 5450 * guarantee that css_tryget_online() won't succeed by the time 5451 * ->css_offline() is invoked. To satisfy all the requirements, 5452 * destruction is implemented in the following two steps. 5453 * 5454 * s1. Verify @cgrp can be destroyed and mark it dying. Remove all 5455 * userland visible parts and start killing the percpu refcnts of 5456 * css's. Set up so that the next stage will be kicked off once all 5457 * the percpu refcnts are confirmed to be killed. 5458 * 5459 * s2. Invoke ->css_offline(), mark the cgroup dead and proceed with the 5460 * rest of destruction. Once all cgroup references are gone, the 5461 * cgroup is RCU-freed. 5462 * 5463 * This function implements s1. After this step, @cgrp is gone as far as 5464 * the userland is concerned and a new cgroup with the same name may be 5465 * created. As cgroup doesn't care about the names internally, this 5466 * doesn't cause any problem. 5467 */ 5468 static int cgroup_destroy_locked(struct cgroup *cgrp) 5469 __releases(&cgroup_mutex) __acquires(&cgroup_mutex) 5470 { 5471 struct cgroup *tcgrp, *parent = cgroup_parent(cgrp); 5472 struct cgroup_subsys_state *css; 5473 struct cgrp_cset_link *link; 5474 int ssid; 5475 5476 lockdep_assert_held(&cgroup_mutex); 5477 5478 /* 5479 * Only migration can raise populated from zero and we're already 5480 * holding cgroup_mutex. 5481 */ 5482 if (cgroup_is_populated(cgrp)) 5483 return -EBUSY; 5484 5485 /* 5486 * Make sure there's no live children. We can't test emptiness of 5487 * ->self.children as dead children linger on it while being 5488 * drained; otherwise, "rmdir parent/child parent" may fail. 5489 */ 5490 if (css_has_online_children(&cgrp->self)) 5491 return -EBUSY; 5492 5493 /* 5494 * Mark @cgrp and the associated csets dead. The former prevents 5495 * further task migration and child creation by disabling 5496 * cgroup_lock_live_group(). The latter makes the csets ignored by 5497 * the migration path. 5498 */ 5499 cgrp->self.flags &= ~CSS_ONLINE; 5500 5501 spin_lock_irq(&css_set_lock); 5502 list_for_each_entry(link, &cgrp->cset_links, cset_link) 5503 link->cset->dead = true; 5504 spin_unlock_irq(&css_set_lock); 5505 5506 /* initiate massacre of all css's */ 5507 for_each_css(css, ssid, cgrp) 5508 kill_css(css); 5509 5510 /* clear and remove @cgrp dir, @cgrp has an extra ref on its kn */ 5511 css_clear_dir(&cgrp->self); 5512 kernfs_remove(cgrp->kn); 5513 5514 if (parent && cgroup_is_threaded(cgrp)) 5515 parent->nr_threaded_children--; 5516 5517 spin_lock_irq(&css_set_lock); 5518 for (tcgrp = cgroup_parent(cgrp); tcgrp; tcgrp = cgroup_parent(tcgrp)) { 5519 tcgrp->nr_descendants--; 5520 tcgrp->nr_dying_descendants++; 5521 /* 5522 * If the dying cgroup is frozen, decrease frozen descendants 5523 * counters of ancestor cgroups. 5524 */ 5525 if (test_bit(CGRP_FROZEN, &cgrp->flags)) 5526 tcgrp->freezer.nr_frozen_descendants--; 5527 } 5528 spin_unlock_irq(&css_set_lock); 5529 5530 cgroup1_check_for_release(parent); 5531 5532 cgroup_bpf_offline(cgrp); 5533 5534 /* put the base reference */ 5535 percpu_ref_kill(&cgrp->self.refcnt); 5536 5537 return 0; 5538 }; 5539 5540 int cgroup_rmdir(struct kernfs_node *kn) 5541 { 5542 struct cgroup *cgrp; 5543 int ret = 0; 5544 5545 cgrp = cgroup_kn_lock_live(kn, false); 5546 if (!cgrp) 5547 return 0; 5548 5549 ret = cgroup_destroy_locked(cgrp); 5550 if (!ret) 5551 TRACE_CGROUP_PATH(rmdir, cgrp); 5552 5553 cgroup_kn_unlock(kn); 5554 return ret; 5555 } 5556 5557 static struct kernfs_syscall_ops cgroup_kf_syscall_ops = { 5558 .show_options = cgroup_show_options, 5559 .mkdir = cgroup_mkdir, 5560 .rmdir = cgroup_rmdir, 5561 .show_path = cgroup_show_path, 5562 }; 5563 5564 static void __init cgroup_init_subsys(struct cgroup_subsys *ss, bool early) 5565 { 5566 struct cgroup_subsys_state *css; 5567 5568 pr_debug("Initializing cgroup subsys %s\n", ss->name); 5569 5570 mutex_lock(&cgroup_mutex); 5571 5572 idr_init(&ss->css_idr); 5573 INIT_LIST_HEAD(&ss->cfts); 5574 5575 /* Create the root cgroup state for this subsystem */ 5576 ss->root = &cgrp_dfl_root; 5577 css = ss->css_alloc(cgroup_css(&cgrp_dfl_root.cgrp, ss)); 5578 /* We don't handle early failures gracefully */ 5579 BUG_ON(IS_ERR(css)); 5580 init_and_link_css(css, ss, &cgrp_dfl_root.cgrp); 5581 5582 /* 5583 * Root csses are never destroyed and we can't initialize 5584 * percpu_ref during early init. Disable refcnting. 5585 */ 5586 css->flags |= CSS_NO_REF; 5587 5588 if (early) { 5589 /* allocation can't be done safely during early init */ 5590 css->id = 1; 5591 } else { 5592 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2, GFP_KERNEL); 5593 BUG_ON(css->id < 0); 5594 } 5595 5596 /* Update the init_css_set to contain a subsys 5597 * pointer to this state - since the subsystem is 5598 * newly registered, all tasks and hence the 5599 * init_css_set is in the subsystem's root cgroup. */ 5600 init_css_set.subsys[ss->id] = css; 5601 5602 have_fork_callback |= (bool)ss->fork << ss->id; 5603 have_exit_callback |= (bool)ss->exit << ss->id; 5604 have_release_callback |= (bool)ss->release << ss->id; 5605 have_canfork_callback |= (bool)ss->can_fork << ss->id; 5606 5607 /* At system boot, before all subsystems have been 5608 * registered, no tasks have been forked, so we don't 5609 * need to invoke fork callbacks here. */ 5610 BUG_ON(!list_empty(&init_task.tasks)); 5611 5612 BUG_ON(online_css(css)); 5613 5614 mutex_unlock(&cgroup_mutex); 5615 } 5616 5617 /** 5618 * cgroup_init_early - cgroup initialization at system boot 5619 * 5620 * Initialize cgroups at system boot, and initialize any 5621 * subsystems that request early init. 5622 */ 5623 int __init cgroup_init_early(void) 5624 { 5625 static struct cgroup_fs_context __initdata ctx; 5626 struct cgroup_subsys *ss; 5627 int i; 5628 5629 ctx.root = &cgrp_dfl_root; 5630 init_cgroup_root(&ctx); 5631 cgrp_dfl_root.cgrp.self.flags |= CSS_NO_REF; 5632 5633 RCU_INIT_POINTER(init_task.cgroups, &init_css_set); 5634 5635 for_each_subsys(ss, i) { 5636 WARN(!ss->css_alloc || !ss->css_free || ss->name || ss->id, 5637 "invalid cgroup_subsys %d:%s css_alloc=%p css_free=%p id:name=%d:%s\n", 5638 i, cgroup_subsys_name[i], ss->css_alloc, ss->css_free, 5639 ss->id, ss->name); 5640 WARN(strlen(cgroup_subsys_name[i]) > MAX_CGROUP_TYPE_NAMELEN, 5641 "cgroup_subsys_name %s too long\n", cgroup_subsys_name[i]); 5642 5643 ss->id = i; 5644 ss->name = cgroup_subsys_name[i]; 5645 if (!ss->legacy_name) 5646 ss->legacy_name = cgroup_subsys_name[i]; 5647 5648 if (ss->early_init) 5649 cgroup_init_subsys(ss, true); 5650 } 5651 return 0; 5652 } 5653 5654 static u16 cgroup_disable_mask __initdata; 5655 5656 /** 5657 * cgroup_init - cgroup initialization 5658 * 5659 * Register cgroup filesystem and /proc file, and initialize 5660 * any subsystems that didn't request early init. 5661 */ 5662 int __init cgroup_init(void) 5663 { 5664 struct cgroup_subsys *ss; 5665 int ssid; 5666 5667 BUILD_BUG_ON(CGROUP_SUBSYS_COUNT > 16); 5668 BUG_ON(cgroup_init_cftypes(NULL, cgroup_base_files)); 5669 BUG_ON(cgroup_init_cftypes(NULL, cgroup1_base_files)); 5670 5671 cgroup_rstat_boot(); 5672 5673 /* 5674 * The latency of the synchronize_rcu() is too high for cgroups, 5675 * avoid it at the cost of forcing all readers into the slow path. 5676 */ 5677 rcu_sync_enter_start(&cgroup_threadgroup_rwsem.rss); 5678 5679 get_user_ns(init_cgroup_ns.user_ns); 5680 5681 mutex_lock(&cgroup_mutex); 5682 5683 /* 5684 * Add init_css_set to the hash table so that dfl_root can link to 5685 * it during init. 5686 */ 5687 hash_add(css_set_table, &init_css_set.hlist, 5688 css_set_hash(init_css_set.subsys)); 5689 5690 BUG_ON(cgroup_setup_root(&cgrp_dfl_root, 0)); 5691 5692 mutex_unlock(&cgroup_mutex); 5693 5694 for_each_subsys(ss, ssid) { 5695 if (ss->early_init) { 5696 struct cgroup_subsys_state *css = 5697 init_css_set.subsys[ss->id]; 5698 5699 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2, 5700 GFP_KERNEL); 5701 BUG_ON(css->id < 0); 5702 } else { 5703 cgroup_init_subsys(ss, false); 5704 } 5705 5706 list_add_tail(&init_css_set.e_cset_node[ssid], 5707 &cgrp_dfl_root.cgrp.e_csets[ssid]); 5708 5709 /* 5710 * Setting dfl_root subsys_mask needs to consider the 5711 * disabled flag and cftype registration needs kmalloc, 5712 * both of which aren't available during early_init. 5713 */ 5714 if (cgroup_disable_mask & (1 << ssid)) { 5715 static_branch_disable(cgroup_subsys_enabled_key[ssid]); 5716 printk(KERN_INFO "Disabling %s control group subsystem\n", 5717 ss->name); 5718 continue; 5719 } 5720 5721 if (cgroup1_ssid_disabled(ssid)) 5722 printk(KERN_INFO "Disabling %s control group subsystem in v1 mounts\n", 5723 ss->name); 5724 5725 cgrp_dfl_root.subsys_mask |= 1 << ss->id; 5726 5727 /* implicit controllers must be threaded too */ 5728 WARN_ON(ss->implicit_on_dfl && !ss->threaded); 5729 5730 if (ss->implicit_on_dfl) 5731 cgrp_dfl_implicit_ss_mask |= 1 << ss->id; 5732 else if (!ss->dfl_cftypes) 5733 cgrp_dfl_inhibit_ss_mask |= 1 << ss->id; 5734 5735 if (ss->threaded) 5736 cgrp_dfl_threaded_ss_mask |= 1 << ss->id; 5737 5738 if (ss->dfl_cftypes == ss->legacy_cftypes) { 5739 WARN_ON(cgroup_add_cftypes(ss, ss->dfl_cftypes)); 5740 } else { 5741 WARN_ON(cgroup_add_dfl_cftypes(ss, ss->dfl_cftypes)); 5742 WARN_ON(cgroup_add_legacy_cftypes(ss, ss->legacy_cftypes)); 5743 } 5744 5745 if (ss->bind) 5746 ss->bind(init_css_set.subsys[ssid]); 5747 5748 mutex_lock(&cgroup_mutex); 5749 css_populate_dir(init_css_set.subsys[ssid]); 5750 mutex_unlock(&cgroup_mutex); 5751 } 5752 5753 /* init_css_set.subsys[] has been updated, re-hash */ 5754 hash_del(&init_css_set.hlist); 5755 hash_add(css_set_table, &init_css_set.hlist, 5756 css_set_hash(init_css_set.subsys)); 5757 5758 WARN_ON(sysfs_create_mount_point(fs_kobj, "cgroup")); 5759 WARN_ON(register_filesystem(&cgroup_fs_type)); 5760 WARN_ON(register_filesystem(&cgroup2_fs_type)); 5761 WARN_ON(!proc_create_single("cgroups", 0, NULL, proc_cgroupstats_show)); 5762 #ifdef CONFIG_CPUSETS 5763 WARN_ON(register_filesystem(&cpuset_fs_type)); 5764 #endif 5765 5766 return 0; 5767 } 5768 5769 static int __init cgroup_wq_init(void) 5770 { 5771 /* 5772 * There isn't much point in executing destruction path in 5773 * parallel. Good chunk is serialized with cgroup_mutex anyway. 5774 * Use 1 for @max_active. 5775 * 5776 * We would prefer to do this in cgroup_init() above, but that 5777 * is called before init_workqueues(): so leave this until after. 5778 */ 5779 cgroup_destroy_wq = alloc_workqueue("cgroup_destroy", 0, 1); 5780 BUG_ON(!cgroup_destroy_wq); 5781 return 0; 5782 } 5783 core_initcall(cgroup_wq_init); 5784 5785 void cgroup_path_from_kernfs_id(u64 id, char *buf, size_t buflen) 5786 { 5787 struct kernfs_node *kn; 5788 5789 kn = kernfs_find_and_get_node_by_id(cgrp_dfl_root.kf_root, id); 5790 if (!kn) 5791 return; 5792 kernfs_path(kn, buf, buflen); 5793 kernfs_put(kn); 5794 } 5795 5796 /* 5797 * proc_cgroup_show() 5798 * - Print task's cgroup paths into seq_file, one line for each hierarchy 5799 * - Used for /proc/<pid>/cgroup. 5800 */ 5801 int proc_cgroup_show(struct seq_file *m, struct pid_namespace *ns, 5802 struct pid *pid, struct task_struct *tsk) 5803 { 5804 char *buf; 5805 int retval; 5806 struct cgroup_root *root; 5807 5808 retval = -ENOMEM; 5809 buf = kmalloc(PATH_MAX, GFP_KERNEL); 5810 if (!buf) 5811 goto out; 5812 5813 mutex_lock(&cgroup_mutex); 5814 spin_lock_irq(&css_set_lock); 5815 5816 for_each_root(root) { 5817 struct cgroup_subsys *ss; 5818 struct cgroup *cgrp; 5819 int ssid, count = 0; 5820 5821 if (root == &cgrp_dfl_root && !cgrp_dfl_visible) 5822 continue; 5823 5824 seq_printf(m, "%d:", root->hierarchy_id); 5825 if (root != &cgrp_dfl_root) 5826 for_each_subsys(ss, ssid) 5827 if (root->subsys_mask & (1 << ssid)) 5828 seq_printf(m, "%s%s", count++ ? "," : "", 5829 ss->legacy_name); 5830 if (strlen(root->name)) 5831 seq_printf(m, "%sname=%s", count ? "," : "", 5832 root->name); 5833 seq_putc(m, ':'); 5834 5835 cgrp = task_cgroup_from_root(tsk, root); 5836 5837 /* 5838 * On traditional hierarchies, all zombie tasks show up as 5839 * belonging to the root cgroup. On the default hierarchy, 5840 * while a zombie doesn't show up in "cgroup.procs" and 5841 * thus can't be migrated, its /proc/PID/cgroup keeps 5842 * reporting the cgroup it belonged to before exiting. If 5843 * the cgroup is removed before the zombie is reaped, 5844 * " (deleted)" is appended to the cgroup path. 5845 */ 5846 if (cgroup_on_dfl(cgrp) || !(tsk->flags & PF_EXITING)) { 5847 retval = cgroup_path_ns_locked(cgrp, buf, PATH_MAX, 5848 current->nsproxy->cgroup_ns); 5849 if (retval >= PATH_MAX) 5850 retval = -ENAMETOOLONG; 5851 if (retval < 0) 5852 goto out_unlock; 5853 5854 seq_puts(m, buf); 5855 } else { 5856 seq_puts(m, "/"); 5857 } 5858 5859 if (cgroup_on_dfl(cgrp) && cgroup_is_dead(cgrp)) 5860 seq_puts(m, " (deleted)\n"); 5861 else 5862 seq_putc(m, '\n'); 5863 } 5864 5865 retval = 0; 5866 out_unlock: 5867 spin_unlock_irq(&css_set_lock); 5868 mutex_unlock(&cgroup_mutex); 5869 kfree(buf); 5870 out: 5871 return retval; 5872 } 5873 5874 /** 5875 * cgroup_fork - initialize cgroup related fields during copy_process() 5876 * @child: pointer to task_struct of forking parent process. 5877 * 5878 * A task is associated with the init_css_set until cgroup_post_fork() 5879 * attaches it to the target css_set. 5880 */ 5881 void cgroup_fork(struct task_struct *child) 5882 { 5883 RCU_INIT_POINTER(child->cgroups, &init_css_set); 5884 INIT_LIST_HEAD(&child->cg_list); 5885 } 5886 5887 static struct cgroup *cgroup_get_from_file(struct file *f) 5888 { 5889 struct cgroup_subsys_state *css; 5890 struct cgroup *cgrp; 5891 5892 css = css_tryget_online_from_dir(f->f_path.dentry, NULL); 5893 if (IS_ERR(css)) 5894 return ERR_CAST(css); 5895 5896 cgrp = css->cgroup; 5897 if (!cgroup_on_dfl(cgrp)) { 5898 cgroup_put(cgrp); 5899 return ERR_PTR(-EBADF); 5900 } 5901 5902 return cgrp; 5903 } 5904 5905 /** 5906 * cgroup_css_set_fork - find or create a css_set for a child process 5907 * @kargs: the arguments passed to create the child process 5908 * 5909 * This functions finds or creates a new css_set which the child 5910 * process will be attached to in cgroup_post_fork(). By default, 5911 * the child process will be given the same css_set as its parent. 5912 * 5913 * If CLONE_INTO_CGROUP is specified this function will try to find an 5914 * existing css_set which includes the requested cgroup and if not create 5915 * a new css_set that the child will be attached to later. If this function 5916 * succeeds it will hold cgroup_threadgroup_rwsem on return. If 5917 * CLONE_INTO_CGROUP is requested this function will grab cgroup mutex 5918 * before grabbing cgroup_threadgroup_rwsem and will hold a reference 5919 * to the target cgroup. 5920 */ 5921 static int cgroup_css_set_fork(struct kernel_clone_args *kargs) 5922 __acquires(&cgroup_mutex) __acquires(&cgroup_threadgroup_rwsem) 5923 { 5924 int ret; 5925 struct cgroup *dst_cgrp = NULL; 5926 struct css_set *cset; 5927 struct super_block *sb; 5928 struct file *f; 5929 5930 if (kargs->flags & CLONE_INTO_CGROUP) 5931 mutex_lock(&cgroup_mutex); 5932 5933 cgroup_threadgroup_change_begin(current); 5934 5935 spin_lock_irq(&css_set_lock); 5936 cset = task_css_set(current); 5937 get_css_set(cset); 5938 spin_unlock_irq(&css_set_lock); 5939 5940 if (!(kargs->flags & CLONE_INTO_CGROUP)) { 5941 kargs->cset = cset; 5942 return 0; 5943 } 5944 5945 f = fget_raw(kargs->cgroup); 5946 if (!f) { 5947 ret = -EBADF; 5948 goto err; 5949 } 5950 sb = f->f_path.dentry->d_sb; 5951 5952 dst_cgrp = cgroup_get_from_file(f); 5953 if (IS_ERR(dst_cgrp)) { 5954 ret = PTR_ERR(dst_cgrp); 5955 dst_cgrp = NULL; 5956 goto err; 5957 } 5958 5959 if (cgroup_is_dead(dst_cgrp)) { 5960 ret = -ENODEV; 5961 goto err; 5962 } 5963 5964 /* 5965 * Verify that we the target cgroup is writable for us. This is 5966 * usually done by the vfs layer but since we're not going through 5967 * the vfs layer here we need to do it "manually". 5968 */ 5969 ret = cgroup_may_write(dst_cgrp, sb); 5970 if (ret) 5971 goto err; 5972 5973 ret = cgroup_attach_permissions(cset->dfl_cgrp, dst_cgrp, sb, 5974 !(kargs->flags & CLONE_THREAD)); 5975 if (ret) 5976 goto err; 5977 5978 kargs->cset = find_css_set(cset, dst_cgrp); 5979 if (!kargs->cset) { 5980 ret = -ENOMEM; 5981 goto err; 5982 } 5983 5984 put_css_set(cset); 5985 fput(f); 5986 kargs->cgrp = dst_cgrp; 5987 return ret; 5988 5989 err: 5990 cgroup_threadgroup_change_end(current); 5991 mutex_unlock(&cgroup_mutex); 5992 if (f) 5993 fput(f); 5994 if (dst_cgrp) 5995 cgroup_put(dst_cgrp); 5996 put_css_set(cset); 5997 if (kargs->cset) 5998 put_css_set(kargs->cset); 5999 return ret; 6000 } 6001 6002 /** 6003 * cgroup_css_set_put_fork - drop references we took during fork 6004 * @kargs: the arguments passed to create the child process 6005 * 6006 * Drop references to the prepared css_set and target cgroup if 6007 * CLONE_INTO_CGROUP was requested. 6008 */ 6009 static void cgroup_css_set_put_fork(struct kernel_clone_args *kargs) 6010 __releases(&cgroup_threadgroup_rwsem) __releases(&cgroup_mutex) 6011 { 6012 cgroup_threadgroup_change_end(current); 6013 6014 if (kargs->flags & CLONE_INTO_CGROUP) { 6015 struct cgroup *cgrp = kargs->cgrp; 6016 struct css_set *cset = kargs->cset; 6017 6018 mutex_unlock(&cgroup_mutex); 6019 6020 if (cset) { 6021 put_css_set(cset); 6022 kargs->cset = NULL; 6023 } 6024 6025 if (cgrp) { 6026 cgroup_put(cgrp); 6027 kargs->cgrp = NULL; 6028 } 6029 } 6030 } 6031 6032 /** 6033 * cgroup_can_fork - called on a new task before the process is exposed 6034 * @child: the child process 6035 * 6036 * This prepares a new css_set for the child process which the child will 6037 * be attached to in cgroup_post_fork(). 6038 * This calls the subsystem can_fork() callbacks. If the cgroup_can_fork() 6039 * callback returns an error, the fork aborts with that error code. This 6040 * allows for a cgroup subsystem to conditionally allow or deny new forks. 6041 */ 6042 int cgroup_can_fork(struct task_struct *child, struct kernel_clone_args *kargs) 6043 { 6044 struct cgroup_subsys *ss; 6045 int i, j, ret; 6046 6047 ret = cgroup_css_set_fork(kargs); 6048 if (ret) 6049 return ret; 6050 6051 do_each_subsys_mask(ss, i, have_canfork_callback) { 6052 ret = ss->can_fork(child, kargs->cset); 6053 if (ret) 6054 goto out_revert; 6055 } while_each_subsys_mask(); 6056 6057 return 0; 6058 6059 out_revert: 6060 for_each_subsys(ss, j) { 6061 if (j >= i) 6062 break; 6063 if (ss->cancel_fork) 6064 ss->cancel_fork(child, kargs->cset); 6065 } 6066 6067 cgroup_css_set_put_fork(kargs); 6068 6069 return ret; 6070 } 6071 6072 /** 6073 * cgroup_cancel_fork - called if a fork failed after cgroup_can_fork() 6074 * @child: the child process 6075 * @kargs: the arguments passed to create the child process 6076 * 6077 * This calls the cancel_fork() callbacks if a fork failed *after* 6078 * cgroup_can_fork() succeded and cleans up references we took to 6079 * prepare a new css_set for the child process in cgroup_can_fork(). 6080 */ 6081 void cgroup_cancel_fork(struct task_struct *child, 6082 struct kernel_clone_args *kargs) 6083 { 6084 struct cgroup_subsys *ss; 6085 int i; 6086 6087 for_each_subsys(ss, i) 6088 if (ss->cancel_fork) 6089 ss->cancel_fork(child, kargs->cset); 6090 6091 cgroup_css_set_put_fork(kargs); 6092 } 6093 6094 /** 6095 * cgroup_post_fork - finalize cgroup setup for the child process 6096 * @child: the child process 6097 * 6098 * Attach the child process to its css_set calling the subsystem fork() 6099 * callbacks. 6100 */ 6101 void cgroup_post_fork(struct task_struct *child, 6102 struct kernel_clone_args *kargs) 6103 __releases(&cgroup_threadgroup_rwsem) __releases(&cgroup_mutex) 6104 { 6105 struct cgroup_subsys *ss; 6106 struct css_set *cset; 6107 int i; 6108 6109 cset = kargs->cset; 6110 kargs->cset = NULL; 6111 6112 spin_lock_irq(&css_set_lock); 6113 6114 /* init tasks are special, only link regular threads */ 6115 if (likely(child->pid)) { 6116 WARN_ON_ONCE(!list_empty(&child->cg_list)); 6117 cset->nr_tasks++; 6118 css_set_move_task(child, NULL, cset, false); 6119 } else { 6120 put_css_set(cset); 6121 cset = NULL; 6122 } 6123 6124 /* 6125 * If the cgroup has to be frozen, the new task has too. Let's set 6126 * the JOBCTL_TRAP_FREEZE jobctl bit to get the task into the 6127 * frozen state. 6128 */ 6129 if (unlikely(cgroup_task_freeze(child))) { 6130 spin_lock(&child->sighand->siglock); 6131 WARN_ON_ONCE(child->frozen); 6132 child->jobctl |= JOBCTL_TRAP_FREEZE; 6133 spin_unlock(&child->sighand->siglock); 6134 6135 /* 6136 * Calling cgroup_update_frozen() isn't required here, 6137 * because it will be called anyway a bit later from 6138 * do_freezer_trap(). So we avoid cgroup's transient switch 6139 * from the frozen state and back. 6140 */ 6141 } 6142 6143 spin_unlock_irq(&css_set_lock); 6144 6145 /* 6146 * Call ss->fork(). This must happen after @child is linked on 6147 * css_set; otherwise, @child might change state between ->fork() 6148 * and addition to css_set. 6149 */ 6150 do_each_subsys_mask(ss, i, have_fork_callback) { 6151 ss->fork(child); 6152 } while_each_subsys_mask(); 6153 6154 /* Make the new cset the root_cset of the new cgroup namespace. */ 6155 if (kargs->flags & CLONE_NEWCGROUP) { 6156 struct css_set *rcset = child->nsproxy->cgroup_ns->root_cset; 6157 6158 get_css_set(cset); 6159 child->nsproxy->cgroup_ns->root_cset = cset; 6160 put_css_set(rcset); 6161 } 6162 6163 cgroup_css_set_put_fork(kargs); 6164 } 6165 6166 /** 6167 * cgroup_exit - detach cgroup from exiting task 6168 * @tsk: pointer to task_struct of exiting process 6169 * 6170 * Description: Detach cgroup from @tsk. 6171 * 6172 */ 6173 void cgroup_exit(struct task_struct *tsk) 6174 { 6175 struct cgroup_subsys *ss; 6176 struct css_set *cset; 6177 int i; 6178 6179 spin_lock_irq(&css_set_lock); 6180 6181 WARN_ON_ONCE(list_empty(&tsk->cg_list)); 6182 cset = task_css_set(tsk); 6183 css_set_move_task(tsk, cset, NULL, false); 6184 list_add_tail(&tsk->cg_list, &cset->dying_tasks); 6185 cset->nr_tasks--; 6186 6187 WARN_ON_ONCE(cgroup_task_frozen(tsk)); 6188 if (unlikely(cgroup_task_freeze(tsk))) 6189 cgroup_update_frozen(task_dfl_cgroup(tsk)); 6190 6191 spin_unlock_irq(&css_set_lock); 6192 6193 /* see cgroup_post_fork() for details */ 6194 do_each_subsys_mask(ss, i, have_exit_callback) { 6195 ss->exit(tsk); 6196 } while_each_subsys_mask(); 6197 } 6198 6199 void cgroup_release(struct task_struct *task) 6200 { 6201 struct cgroup_subsys *ss; 6202 int ssid; 6203 6204 do_each_subsys_mask(ss, ssid, have_release_callback) { 6205 ss->release(task); 6206 } while_each_subsys_mask(); 6207 6208 spin_lock_irq(&css_set_lock); 6209 css_set_skip_task_iters(task_css_set(task), task); 6210 list_del_init(&task->cg_list); 6211 spin_unlock_irq(&css_set_lock); 6212 } 6213 6214 void cgroup_free(struct task_struct *task) 6215 { 6216 struct css_set *cset = task_css_set(task); 6217 put_css_set(cset); 6218 } 6219 6220 static int __init cgroup_disable(char *str) 6221 { 6222 struct cgroup_subsys *ss; 6223 char *token; 6224 int i; 6225 6226 while ((token = strsep(&str, ",")) != NULL) { 6227 if (!*token) 6228 continue; 6229 6230 for_each_subsys(ss, i) { 6231 if (strcmp(token, ss->name) && 6232 strcmp(token, ss->legacy_name)) 6233 continue; 6234 cgroup_disable_mask |= 1 << i; 6235 } 6236 } 6237 return 1; 6238 } 6239 __setup("cgroup_disable=", cgroup_disable); 6240 6241 void __init __weak enable_debug_cgroup(void) { } 6242 6243 static int __init enable_cgroup_debug(char *str) 6244 { 6245 cgroup_debug = true; 6246 enable_debug_cgroup(); 6247 return 1; 6248 } 6249 __setup("cgroup_debug", enable_cgroup_debug); 6250 6251 /** 6252 * css_tryget_online_from_dir - get corresponding css from a cgroup dentry 6253 * @dentry: directory dentry of interest 6254 * @ss: subsystem of interest 6255 * 6256 * If @dentry is a directory for a cgroup which has @ss enabled on it, try 6257 * to get the corresponding css and return it. If such css doesn't exist 6258 * or can't be pinned, an ERR_PTR value is returned. 6259 */ 6260 struct cgroup_subsys_state *css_tryget_online_from_dir(struct dentry *dentry, 6261 struct cgroup_subsys *ss) 6262 { 6263 struct kernfs_node *kn = kernfs_node_from_dentry(dentry); 6264 struct file_system_type *s_type = dentry->d_sb->s_type; 6265 struct cgroup_subsys_state *css = NULL; 6266 struct cgroup *cgrp; 6267 6268 /* is @dentry a cgroup dir? */ 6269 if ((s_type != &cgroup_fs_type && s_type != &cgroup2_fs_type) || 6270 !kn || kernfs_type(kn) != KERNFS_DIR) 6271 return ERR_PTR(-EBADF); 6272 6273 rcu_read_lock(); 6274 6275 /* 6276 * This path doesn't originate from kernfs and @kn could already 6277 * have been or be removed at any point. @kn->priv is RCU 6278 * protected for this access. See css_release_work_fn() for details. 6279 */ 6280 cgrp = rcu_dereference(*(void __rcu __force **)&kn->priv); 6281 if (cgrp) 6282 css = cgroup_css(cgrp, ss); 6283 6284 if (!css || !css_tryget_online(css)) 6285 css = ERR_PTR(-ENOENT); 6286 6287 rcu_read_unlock(); 6288 return css; 6289 } 6290 6291 /** 6292 * css_from_id - lookup css by id 6293 * @id: the cgroup id 6294 * @ss: cgroup subsys to be looked into 6295 * 6296 * Returns the css if there's valid one with @id, otherwise returns NULL. 6297 * Should be called under rcu_read_lock(). 6298 */ 6299 struct cgroup_subsys_state *css_from_id(int id, struct cgroup_subsys *ss) 6300 { 6301 WARN_ON_ONCE(!rcu_read_lock_held()); 6302 return idr_find(&ss->css_idr, id); 6303 } 6304 6305 /** 6306 * cgroup_get_from_path - lookup and get a cgroup from its default hierarchy path 6307 * @path: path on the default hierarchy 6308 * 6309 * Find the cgroup at @path on the default hierarchy, increment its 6310 * reference count and return it. Returns pointer to the found cgroup on 6311 * success, ERR_PTR(-ENOENT) if @path doesn't exist and ERR_PTR(-ENOTDIR) 6312 * if @path points to a non-directory. 6313 */ 6314 struct cgroup *cgroup_get_from_path(const char *path) 6315 { 6316 struct kernfs_node *kn; 6317 struct cgroup *cgrp; 6318 6319 mutex_lock(&cgroup_mutex); 6320 6321 kn = kernfs_walk_and_get(cgrp_dfl_root.cgrp.kn, path); 6322 if (kn) { 6323 if (kernfs_type(kn) == KERNFS_DIR) { 6324 cgrp = kn->priv; 6325 cgroup_get_live(cgrp); 6326 } else { 6327 cgrp = ERR_PTR(-ENOTDIR); 6328 } 6329 kernfs_put(kn); 6330 } else { 6331 cgrp = ERR_PTR(-ENOENT); 6332 } 6333 6334 mutex_unlock(&cgroup_mutex); 6335 return cgrp; 6336 } 6337 EXPORT_SYMBOL_GPL(cgroup_get_from_path); 6338 6339 /** 6340 * cgroup_get_from_fd - get a cgroup pointer from a fd 6341 * @fd: fd obtained by open(cgroup2_dir) 6342 * 6343 * Find the cgroup from a fd which should be obtained 6344 * by opening a cgroup directory. Returns a pointer to the 6345 * cgroup on success. ERR_PTR is returned if the cgroup 6346 * cannot be found. 6347 */ 6348 struct cgroup *cgroup_get_from_fd(int fd) 6349 { 6350 struct cgroup *cgrp; 6351 struct file *f; 6352 6353 f = fget_raw(fd); 6354 if (!f) 6355 return ERR_PTR(-EBADF); 6356 6357 cgrp = cgroup_get_from_file(f); 6358 fput(f); 6359 return cgrp; 6360 } 6361 EXPORT_SYMBOL_GPL(cgroup_get_from_fd); 6362 6363 static u64 power_of_ten(int power) 6364 { 6365 u64 v = 1; 6366 while (power--) 6367 v *= 10; 6368 return v; 6369 } 6370 6371 /** 6372 * cgroup_parse_float - parse a floating number 6373 * @input: input string 6374 * @dec_shift: number of decimal digits to shift 6375 * @v: output 6376 * 6377 * Parse a decimal floating point number in @input and store the result in 6378 * @v with decimal point right shifted @dec_shift times. For example, if 6379 * @input is "12.3456" and @dec_shift is 3, *@v will be set to 12345. 6380 * Returns 0 on success, -errno otherwise. 6381 * 6382 * There's nothing cgroup specific about this function except that it's 6383 * currently the only user. 6384 */ 6385 int cgroup_parse_float(const char *input, unsigned dec_shift, s64 *v) 6386 { 6387 s64 whole, frac = 0; 6388 int fstart = 0, fend = 0, flen; 6389 6390 if (!sscanf(input, "%lld.%n%lld%n", &whole, &fstart, &frac, &fend)) 6391 return -EINVAL; 6392 if (frac < 0) 6393 return -EINVAL; 6394 6395 flen = fend > fstart ? fend - fstart : 0; 6396 if (flen < dec_shift) 6397 frac *= power_of_ten(dec_shift - flen); 6398 else 6399 frac = DIV_ROUND_CLOSEST_ULL(frac, power_of_ten(flen - dec_shift)); 6400 6401 *v = whole * power_of_ten(dec_shift) + frac; 6402 return 0; 6403 } 6404 6405 /* 6406 * sock->sk_cgrp_data handling. For more info, see sock_cgroup_data 6407 * definition in cgroup-defs.h. 6408 */ 6409 #ifdef CONFIG_SOCK_CGROUP_DATA 6410 6411 #if defined(CONFIG_CGROUP_NET_PRIO) || defined(CONFIG_CGROUP_NET_CLASSID) 6412 6413 DEFINE_SPINLOCK(cgroup_sk_update_lock); 6414 static bool cgroup_sk_alloc_disabled __read_mostly; 6415 6416 void cgroup_sk_alloc_disable(void) 6417 { 6418 if (cgroup_sk_alloc_disabled) 6419 return; 6420 pr_info("cgroup: disabling cgroup2 socket matching due to net_prio or net_cls activation\n"); 6421 cgroup_sk_alloc_disabled = true; 6422 } 6423 6424 #else 6425 6426 #define cgroup_sk_alloc_disabled false 6427 6428 #endif 6429 6430 void cgroup_sk_alloc(struct sock_cgroup_data *skcd) 6431 { 6432 if (cgroup_sk_alloc_disabled) { 6433 skcd->no_refcnt = 1; 6434 return; 6435 } 6436 6437 /* Don't associate the sock with unrelated interrupted task's cgroup. */ 6438 if (in_interrupt()) 6439 return; 6440 6441 rcu_read_lock(); 6442 6443 while (true) { 6444 struct css_set *cset; 6445 6446 cset = task_css_set(current); 6447 if (likely(cgroup_tryget(cset->dfl_cgrp))) { 6448 skcd->val = (unsigned long)cset->dfl_cgrp; 6449 cgroup_bpf_get(cset->dfl_cgrp); 6450 break; 6451 } 6452 cpu_relax(); 6453 } 6454 6455 rcu_read_unlock(); 6456 } 6457 6458 void cgroup_sk_clone(struct sock_cgroup_data *skcd) 6459 { 6460 if (skcd->val) { 6461 if (skcd->no_refcnt) 6462 return; 6463 /* 6464 * We might be cloning a socket which is left in an empty 6465 * cgroup and the cgroup might have already been rmdir'd. 6466 * Don't use cgroup_get_live(). 6467 */ 6468 cgroup_get(sock_cgroup_ptr(skcd)); 6469 cgroup_bpf_get(sock_cgroup_ptr(skcd)); 6470 } 6471 } 6472 6473 void cgroup_sk_free(struct sock_cgroup_data *skcd) 6474 { 6475 struct cgroup *cgrp = sock_cgroup_ptr(skcd); 6476 6477 if (skcd->no_refcnt) 6478 return; 6479 cgroup_bpf_put(cgrp); 6480 cgroup_put(cgrp); 6481 } 6482 6483 #endif /* CONFIG_SOCK_CGROUP_DATA */ 6484 6485 #ifdef CONFIG_CGROUP_BPF 6486 int cgroup_bpf_attach(struct cgroup *cgrp, 6487 struct bpf_prog *prog, struct bpf_prog *replace_prog, 6488 struct bpf_cgroup_link *link, 6489 enum bpf_attach_type type, 6490 u32 flags) 6491 { 6492 int ret; 6493 6494 mutex_lock(&cgroup_mutex); 6495 ret = __cgroup_bpf_attach(cgrp, prog, replace_prog, link, type, flags); 6496 mutex_unlock(&cgroup_mutex); 6497 return ret; 6498 } 6499 6500 int cgroup_bpf_detach(struct cgroup *cgrp, struct bpf_prog *prog, 6501 enum bpf_attach_type type) 6502 { 6503 int ret; 6504 6505 mutex_lock(&cgroup_mutex); 6506 ret = __cgroup_bpf_detach(cgrp, prog, NULL, type); 6507 mutex_unlock(&cgroup_mutex); 6508 return ret; 6509 } 6510 6511 int cgroup_bpf_query(struct cgroup *cgrp, const union bpf_attr *attr, 6512 union bpf_attr __user *uattr) 6513 { 6514 int ret; 6515 6516 mutex_lock(&cgroup_mutex); 6517 ret = __cgroup_bpf_query(cgrp, attr, uattr); 6518 mutex_unlock(&cgroup_mutex); 6519 return ret; 6520 } 6521 #endif /* CONFIG_CGROUP_BPF */ 6522 6523 #ifdef CONFIG_SYSFS 6524 static ssize_t show_delegatable_files(struct cftype *files, char *buf, 6525 ssize_t size, const char *prefix) 6526 { 6527 struct cftype *cft; 6528 ssize_t ret = 0; 6529 6530 for (cft = files; cft && cft->name[0] != '\0'; cft++) { 6531 if (!(cft->flags & CFTYPE_NS_DELEGATABLE)) 6532 continue; 6533 6534 if (prefix) 6535 ret += snprintf(buf + ret, size - ret, "%s.", prefix); 6536 6537 ret += snprintf(buf + ret, size - ret, "%s\n", cft->name); 6538 6539 if (WARN_ON(ret >= size)) 6540 break; 6541 } 6542 6543 return ret; 6544 } 6545 6546 static ssize_t delegate_show(struct kobject *kobj, struct kobj_attribute *attr, 6547 char *buf) 6548 { 6549 struct cgroup_subsys *ss; 6550 int ssid; 6551 ssize_t ret = 0; 6552 6553 ret = show_delegatable_files(cgroup_base_files, buf, PAGE_SIZE - ret, 6554 NULL); 6555 6556 for_each_subsys(ss, ssid) 6557 ret += show_delegatable_files(ss->dfl_cftypes, buf + ret, 6558 PAGE_SIZE - ret, 6559 cgroup_subsys_name[ssid]); 6560 6561 return ret; 6562 } 6563 static struct kobj_attribute cgroup_delegate_attr = __ATTR_RO(delegate); 6564 6565 static ssize_t features_show(struct kobject *kobj, struct kobj_attribute *attr, 6566 char *buf) 6567 { 6568 return snprintf(buf, PAGE_SIZE, 6569 "nsdelegate\n" 6570 "memory_localevents\n" 6571 "memory_recursiveprot\n"); 6572 } 6573 static struct kobj_attribute cgroup_features_attr = __ATTR_RO(features); 6574 6575 static struct attribute *cgroup_sysfs_attrs[] = { 6576 &cgroup_delegate_attr.attr, 6577 &cgroup_features_attr.attr, 6578 NULL, 6579 }; 6580 6581 static const struct attribute_group cgroup_sysfs_attr_group = { 6582 .attrs = cgroup_sysfs_attrs, 6583 .name = "cgroup", 6584 }; 6585 6586 static int __init cgroup_sysfs_init(void) 6587 { 6588 return sysfs_create_group(kernel_kobj, &cgroup_sysfs_attr_group); 6589 } 6590 subsys_initcall(cgroup_sysfs_init); 6591 6592 #endif /* CONFIG_SYSFS */ 6593