1 // SPDX-License-Identifier: GPL-2.0-only 2 #include "cgroup-internal.h" 3 4 #include <linux/ctype.h> 5 #include <linux/kmod.h> 6 #include <linux/sort.h> 7 #include <linux/delay.h> 8 #include <linux/mm.h> 9 #include <linux/sched/signal.h> 10 #include <linux/sched/task.h> 11 #include <linux/magic.h> 12 #include <linux/slab.h> 13 #include <linux/vmalloc.h> 14 #include <linux/delayacct.h> 15 #include <linux/pid_namespace.h> 16 #include <linux/cgroupstats.h> 17 #include <linux/fs_parser.h> 18 19 #include <trace/events/cgroup.h> 20 21 /* 22 * pidlists linger the following amount before being destroyed. The goal 23 * is avoiding frequent destruction in the middle of consecutive read calls 24 * Expiring in the middle is a performance problem not a correctness one. 25 * 1 sec should be enough. 26 */ 27 #define CGROUP_PIDLIST_DESTROY_DELAY HZ 28 29 /* Controllers blocked by the commandline in v1 */ 30 static u16 cgroup_no_v1_mask; 31 32 /* disable named v1 mounts */ 33 static bool cgroup_no_v1_named; 34 35 /* 36 * pidlist destructions need to be flushed on cgroup destruction. Use a 37 * separate workqueue as flush domain. 38 */ 39 static struct workqueue_struct *cgroup_pidlist_destroy_wq; 40 41 /* protects cgroup_subsys->release_agent_path */ 42 static DEFINE_SPINLOCK(release_agent_path_lock); 43 44 bool cgroup1_ssid_disabled(int ssid) 45 { 46 return cgroup_no_v1_mask & (1 << ssid); 47 } 48 49 /** 50 * cgroup_attach_task_all - attach task 'tsk' to all cgroups of task 'from' 51 * @from: attach to all cgroups of a given task 52 * @tsk: the task to be attached 53 */ 54 int cgroup_attach_task_all(struct task_struct *from, struct task_struct *tsk) 55 { 56 struct cgroup_root *root; 57 int retval = 0; 58 59 mutex_lock(&cgroup_mutex); 60 percpu_down_write(&cgroup_threadgroup_rwsem); 61 for_each_root(root) { 62 struct cgroup *from_cgrp; 63 64 if (root == &cgrp_dfl_root) 65 continue; 66 67 spin_lock_irq(&css_set_lock); 68 from_cgrp = task_cgroup_from_root(from, root); 69 spin_unlock_irq(&css_set_lock); 70 71 retval = cgroup_attach_task(from_cgrp, tsk, false); 72 if (retval) 73 break; 74 } 75 percpu_up_write(&cgroup_threadgroup_rwsem); 76 mutex_unlock(&cgroup_mutex); 77 78 return retval; 79 } 80 EXPORT_SYMBOL_GPL(cgroup_attach_task_all); 81 82 /** 83 * cgroup_trasnsfer_tasks - move tasks from one cgroup to another 84 * @to: cgroup to which the tasks will be moved 85 * @from: cgroup in which the tasks currently reside 86 * 87 * Locking rules between cgroup_post_fork() and the migration path 88 * guarantee that, if a task is forking while being migrated, the new child 89 * is guaranteed to be either visible in the source cgroup after the 90 * parent's migration is complete or put into the target cgroup. No task 91 * can slip out of migration through forking. 92 */ 93 int cgroup_transfer_tasks(struct cgroup *to, struct cgroup *from) 94 { 95 DEFINE_CGROUP_MGCTX(mgctx); 96 struct cgrp_cset_link *link; 97 struct css_task_iter it; 98 struct task_struct *task; 99 int ret; 100 101 if (cgroup_on_dfl(to)) 102 return -EINVAL; 103 104 ret = cgroup_migrate_vet_dst(to); 105 if (ret) 106 return ret; 107 108 mutex_lock(&cgroup_mutex); 109 110 percpu_down_write(&cgroup_threadgroup_rwsem); 111 112 /* all tasks in @from are being moved, all csets are source */ 113 spin_lock_irq(&css_set_lock); 114 list_for_each_entry(link, &from->cset_links, cset_link) 115 cgroup_migrate_add_src(link->cset, to, &mgctx); 116 spin_unlock_irq(&css_set_lock); 117 118 ret = cgroup_migrate_prepare_dst(&mgctx); 119 if (ret) 120 goto out_err; 121 122 /* 123 * Migrate tasks one-by-one until @from is empty. This fails iff 124 * ->can_attach() fails. 125 */ 126 do { 127 css_task_iter_start(&from->self, 0, &it); 128 129 do { 130 task = css_task_iter_next(&it); 131 } while (task && (task->flags & PF_EXITING)); 132 133 if (task) 134 get_task_struct(task); 135 css_task_iter_end(&it); 136 137 if (task) { 138 ret = cgroup_migrate(task, false, &mgctx); 139 if (!ret) 140 TRACE_CGROUP_PATH(transfer_tasks, to, task, false); 141 put_task_struct(task); 142 } 143 } while (task && !ret); 144 out_err: 145 cgroup_migrate_finish(&mgctx); 146 percpu_up_write(&cgroup_threadgroup_rwsem); 147 mutex_unlock(&cgroup_mutex); 148 return ret; 149 } 150 151 /* 152 * Stuff for reading the 'tasks'/'procs' files. 153 * 154 * Reading this file can return large amounts of data if a cgroup has 155 * *lots* of attached tasks. So it may need several calls to read(), 156 * but we cannot guarantee that the information we produce is correct 157 * unless we produce it entirely atomically. 158 * 159 */ 160 161 /* which pidlist file are we talking about? */ 162 enum cgroup_filetype { 163 CGROUP_FILE_PROCS, 164 CGROUP_FILE_TASKS, 165 }; 166 167 /* 168 * A pidlist is a list of pids that virtually represents the contents of one 169 * of the cgroup files ("procs" or "tasks"). We keep a list of such pidlists, 170 * a pair (one each for procs, tasks) for each pid namespace that's relevant 171 * to the cgroup. 172 */ 173 struct cgroup_pidlist { 174 /* 175 * used to find which pidlist is wanted. doesn't change as long as 176 * this particular list stays in the list. 177 */ 178 struct { enum cgroup_filetype type; struct pid_namespace *ns; } key; 179 /* array of xids */ 180 pid_t *list; 181 /* how many elements the above list has */ 182 int length; 183 /* each of these stored in a list by its cgroup */ 184 struct list_head links; 185 /* pointer to the cgroup we belong to, for list removal purposes */ 186 struct cgroup *owner; 187 /* for delayed destruction */ 188 struct delayed_work destroy_dwork; 189 }; 190 191 /* 192 * Used to destroy all pidlists lingering waiting for destroy timer. None 193 * should be left afterwards. 194 */ 195 void cgroup1_pidlist_destroy_all(struct cgroup *cgrp) 196 { 197 struct cgroup_pidlist *l, *tmp_l; 198 199 mutex_lock(&cgrp->pidlist_mutex); 200 list_for_each_entry_safe(l, tmp_l, &cgrp->pidlists, links) 201 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork, 0); 202 mutex_unlock(&cgrp->pidlist_mutex); 203 204 flush_workqueue(cgroup_pidlist_destroy_wq); 205 BUG_ON(!list_empty(&cgrp->pidlists)); 206 } 207 208 static void cgroup_pidlist_destroy_work_fn(struct work_struct *work) 209 { 210 struct delayed_work *dwork = to_delayed_work(work); 211 struct cgroup_pidlist *l = container_of(dwork, struct cgroup_pidlist, 212 destroy_dwork); 213 struct cgroup_pidlist *tofree = NULL; 214 215 mutex_lock(&l->owner->pidlist_mutex); 216 217 /* 218 * Destroy iff we didn't get queued again. The state won't change 219 * as destroy_dwork can only be queued while locked. 220 */ 221 if (!delayed_work_pending(dwork)) { 222 list_del(&l->links); 223 kvfree(l->list); 224 put_pid_ns(l->key.ns); 225 tofree = l; 226 } 227 228 mutex_unlock(&l->owner->pidlist_mutex); 229 kfree(tofree); 230 } 231 232 /* 233 * pidlist_uniq - given a kmalloc()ed list, strip out all duplicate entries 234 * Returns the number of unique elements. 235 */ 236 static int pidlist_uniq(pid_t *list, int length) 237 { 238 int src, dest = 1; 239 240 /* 241 * we presume the 0th element is unique, so i starts at 1. trivial 242 * edge cases first; no work needs to be done for either 243 */ 244 if (length == 0 || length == 1) 245 return length; 246 /* src and dest walk down the list; dest counts unique elements */ 247 for (src = 1; src < length; src++) { 248 /* find next unique element */ 249 while (list[src] == list[src-1]) { 250 src++; 251 if (src == length) 252 goto after; 253 } 254 /* dest always points to where the next unique element goes */ 255 list[dest] = list[src]; 256 dest++; 257 } 258 after: 259 return dest; 260 } 261 262 /* 263 * The two pid files - task and cgroup.procs - guaranteed that the result 264 * is sorted, which forced this whole pidlist fiasco. As pid order is 265 * different per namespace, each namespace needs differently sorted list, 266 * making it impossible to use, for example, single rbtree of member tasks 267 * sorted by task pointer. As pidlists can be fairly large, allocating one 268 * per open file is dangerous, so cgroup had to implement shared pool of 269 * pidlists keyed by cgroup and namespace. 270 */ 271 static int cmppid(const void *a, const void *b) 272 { 273 return *(pid_t *)a - *(pid_t *)b; 274 } 275 276 static struct cgroup_pidlist *cgroup_pidlist_find(struct cgroup *cgrp, 277 enum cgroup_filetype type) 278 { 279 struct cgroup_pidlist *l; 280 /* don't need task_nsproxy() if we're looking at ourself */ 281 struct pid_namespace *ns = task_active_pid_ns(current); 282 283 lockdep_assert_held(&cgrp->pidlist_mutex); 284 285 list_for_each_entry(l, &cgrp->pidlists, links) 286 if (l->key.type == type && l->key.ns == ns) 287 return l; 288 return NULL; 289 } 290 291 /* 292 * find the appropriate pidlist for our purpose (given procs vs tasks) 293 * returns with the lock on that pidlist already held, and takes care 294 * of the use count, or returns NULL with no locks held if we're out of 295 * memory. 296 */ 297 static struct cgroup_pidlist *cgroup_pidlist_find_create(struct cgroup *cgrp, 298 enum cgroup_filetype type) 299 { 300 struct cgroup_pidlist *l; 301 302 lockdep_assert_held(&cgrp->pidlist_mutex); 303 304 l = cgroup_pidlist_find(cgrp, type); 305 if (l) 306 return l; 307 308 /* entry not found; create a new one */ 309 l = kzalloc(sizeof(struct cgroup_pidlist), GFP_KERNEL); 310 if (!l) 311 return l; 312 313 INIT_DELAYED_WORK(&l->destroy_dwork, cgroup_pidlist_destroy_work_fn); 314 l->key.type = type; 315 /* don't need task_nsproxy() if we're looking at ourself */ 316 l->key.ns = get_pid_ns(task_active_pid_ns(current)); 317 l->owner = cgrp; 318 list_add(&l->links, &cgrp->pidlists); 319 return l; 320 } 321 322 /* 323 * Load a cgroup's pidarray with either procs' tgids or tasks' pids 324 */ 325 static int pidlist_array_load(struct cgroup *cgrp, enum cgroup_filetype type, 326 struct cgroup_pidlist **lp) 327 { 328 pid_t *array; 329 int length; 330 int pid, n = 0; /* used for populating the array */ 331 struct css_task_iter it; 332 struct task_struct *tsk; 333 struct cgroup_pidlist *l; 334 335 lockdep_assert_held(&cgrp->pidlist_mutex); 336 337 /* 338 * If cgroup gets more users after we read count, we won't have 339 * enough space - tough. This race is indistinguishable to the 340 * caller from the case that the additional cgroup users didn't 341 * show up until sometime later on. 342 */ 343 length = cgroup_task_count(cgrp); 344 array = kvmalloc_array(length, sizeof(pid_t), GFP_KERNEL); 345 if (!array) 346 return -ENOMEM; 347 /* now, populate the array */ 348 css_task_iter_start(&cgrp->self, 0, &it); 349 while ((tsk = css_task_iter_next(&it))) { 350 if (unlikely(n == length)) 351 break; 352 /* get tgid or pid for procs or tasks file respectively */ 353 if (type == CGROUP_FILE_PROCS) 354 pid = task_tgid_vnr(tsk); 355 else 356 pid = task_pid_vnr(tsk); 357 if (pid > 0) /* make sure to only use valid results */ 358 array[n++] = pid; 359 } 360 css_task_iter_end(&it); 361 length = n; 362 /* now sort & (if procs) strip out duplicates */ 363 sort(array, length, sizeof(pid_t), cmppid, NULL); 364 if (type == CGROUP_FILE_PROCS) 365 length = pidlist_uniq(array, length); 366 367 l = cgroup_pidlist_find_create(cgrp, type); 368 if (!l) { 369 kvfree(array); 370 return -ENOMEM; 371 } 372 373 /* store array, freeing old if necessary */ 374 kvfree(l->list); 375 l->list = array; 376 l->length = length; 377 *lp = l; 378 return 0; 379 } 380 381 /* 382 * seq_file methods for the tasks/procs files. The seq_file position is the 383 * next pid to display; the seq_file iterator is a pointer to the pid 384 * in the cgroup->l->list array. 385 */ 386 387 static void *cgroup_pidlist_start(struct seq_file *s, loff_t *pos) 388 { 389 /* 390 * Initially we receive a position value that corresponds to 391 * one more than the last pid shown (or 0 on the first call or 392 * after a seek to the start). Use a binary-search to find the 393 * next pid to display, if any 394 */ 395 struct kernfs_open_file *of = s->private; 396 struct cgroup *cgrp = seq_css(s)->cgroup; 397 struct cgroup_pidlist *l; 398 enum cgroup_filetype type = seq_cft(s)->private; 399 int index = 0, pid = *pos; 400 int *iter, ret; 401 402 mutex_lock(&cgrp->pidlist_mutex); 403 404 /* 405 * !NULL @of->priv indicates that this isn't the first start() 406 * after open. If the matching pidlist is around, we can use that. 407 * Look for it. Note that @of->priv can't be used directly. It 408 * could already have been destroyed. 409 */ 410 if (of->priv) 411 of->priv = cgroup_pidlist_find(cgrp, type); 412 413 /* 414 * Either this is the first start() after open or the matching 415 * pidlist has been destroyed inbetween. Create a new one. 416 */ 417 if (!of->priv) { 418 ret = pidlist_array_load(cgrp, type, 419 (struct cgroup_pidlist **)&of->priv); 420 if (ret) 421 return ERR_PTR(ret); 422 } 423 l = of->priv; 424 425 if (pid) { 426 int end = l->length; 427 428 while (index < end) { 429 int mid = (index + end) / 2; 430 if (l->list[mid] == pid) { 431 index = mid; 432 break; 433 } else if (l->list[mid] <= pid) 434 index = mid + 1; 435 else 436 end = mid; 437 } 438 } 439 /* If we're off the end of the array, we're done */ 440 if (index >= l->length) 441 return NULL; 442 /* Update the abstract position to be the actual pid that we found */ 443 iter = l->list + index; 444 *pos = *iter; 445 return iter; 446 } 447 448 static void cgroup_pidlist_stop(struct seq_file *s, void *v) 449 { 450 struct kernfs_open_file *of = s->private; 451 struct cgroup_pidlist *l = of->priv; 452 453 if (l) 454 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork, 455 CGROUP_PIDLIST_DESTROY_DELAY); 456 mutex_unlock(&seq_css(s)->cgroup->pidlist_mutex); 457 } 458 459 static void *cgroup_pidlist_next(struct seq_file *s, void *v, loff_t *pos) 460 { 461 struct kernfs_open_file *of = s->private; 462 struct cgroup_pidlist *l = of->priv; 463 pid_t *p = v; 464 pid_t *end = l->list + l->length; 465 /* 466 * Advance to the next pid in the array. If this goes off the 467 * end, we're done 468 */ 469 p++; 470 if (p >= end) { 471 (*pos)++; 472 return NULL; 473 } else { 474 *pos = *p; 475 return p; 476 } 477 } 478 479 static int cgroup_pidlist_show(struct seq_file *s, void *v) 480 { 481 seq_printf(s, "%d\n", *(int *)v); 482 483 return 0; 484 } 485 486 static ssize_t __cgroup1_procs_write(struct kernfs_open_file *of, 487 char *buf, size_t nbytes, loff_t off, 488 bool threadgroup) 489 { 490 struct cgroup *cgrp; 491 struct task_struct *task; 492 const struct cred *cred, *tcred; 493 ssize_t ret; 494 bool locked; 495 496 cgrp = cgroup_kn_lock_live(of->kn, false); 497 if (!cgrp) 498 return -ENODEV; 499 500 task = cgroup_procs_write_start(buf, threadgroup, &locked); 501 ret = PTR_ERR_OR_ZERO(task); 502 if (ret) 503 goto out_unlock; 504 505 /* 506 * Even if we're attaching all tasks in the thread group, we only 507 * need to check permissions on one of them. 508 */ 509 cred = current_cred(); 510 tcred = get_task_cred(task); 511 if (!uid_eq(cred->euid, GLOBAL_ROOT_UID) && 512 !uid_eq(cred->euid, tcred->uid) && 513 !uid_eq(cred->euid, tcred->suid)) 514 ret = -EACCES; 515 put_cred(tcred); 516 if (ret) 517 goto out_finish; 518 519 ret = cgroup_attach_task(cgrp, task, threadgroup); 520 521 out_finish: 522 cgroup_procs_write_finish(task, locked); 523 out_unlock: 524 cgroup_kn_unlock(of->kn); 525 526 return ret ?: nbytes; 527 } 528 529 static ssize_t cgroup1_procs_write(struct kernfs_open_file *of, 530 char *buf, size_t nbytes, loff_t off) 531 { 532 return __cgroup1_procs_write(of, buf, nbytes, off, true); 533 } 534 535 static ssize_t cgroup1_tasks_write(struct kernfs_open_file *of, 536 char *buf, size_t nbytes, loff_t off) 537 { 538 return __cgroup1_procs_write(of, buf, nbytes, off, false); 539 } 540 541 static ssize_t cgroup_release_agent_write(struct kernfs_open_file *of, 542 char *buf, size_t nbytes, loff_t off) 543 { 544 struct cgroup *cgrp; 545 546 BUILD_BUG_ON(sizeof(cgrp->root->release_agent_path) < PATH_MAX); 547 548 cgrp = cgroup_kn_lock_live(of->kn, false); 549 if (!cgrp) 550 return -ENODEV; 551 spin_lock(&release_agent_path_lock); 552 strlcpy(cgrp->root->release_agent_path, strstrip(buf), 553 sizeof(cgrp->root->release_agent_path)); 554 spin_unlock(&release_agent_path_lock); 555 cgroup_kn_unlock(of->kn); 556 return nbytes; 557 } 558 559 static int cgroup_release_agent_show(struct seq_file *seq, void *v) 560 { 561 struct cgroup *cgrp = seq_css(seq)->cgroup; 562 563 spin_lock(&release_agent_path_lock); 564 seq_puts(seq, cgrp->root->release_agent_path); 565 spin_unlock(&release_agent_path_lock); 566 seq_putc(seq, '\n'); 567 return 0; 568 } 569 570 static int cgroup_sane_behavior_show(struct seq_file *seq, void *v) 571 { 572 seq_puts(seq, "0\n"); 573 return 0; 574 } 575 576 static u64 cgroup_read_notify_on_release(struct cgroup_subsys_state *css, 577 struct cftype *cft) 578 { 579 return notify_on_release(css->cgroup); 580 } 581 582 static int cgroup_write_notify_on_release(struct cgroup_subsys_state *css, 583 struct cftype *cft, u64 val) 584 { 585 if (val) 586 set_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags); 587 else 588 clear_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags); 589 return 0; 590 } 591 592 static u64 cgroup_clone_children_read(struct cgroup_subsys_state *css, 593 struct cftype *cft) 594 { 595 return test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags); 596 } 597 598 static int cgroup_clone_children_write(struct cgroup_subsys_state *css, 599 struct cftype *cft, u64 val) 600 { 601 if (val) 602 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags); 603 else 604 clear_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags); 605 return 0; 606 } 607 608 /* cgroup core interface files for the legacy hierarchies */ 609 struct cftype cgroup1_base_files[] = { 610 { 611 .name = "cgroup.procs", 612 .seq_start = cgroup_pidlist_start, 613 .seq_next = cgroup_pidlist_next, 614 .seq_stop = cgroup_pidlist_stop, 615 .seq_show = cgroup_pidlist_show, 616 .private = CGROUP_FILE_PROCS, 617 .write = cgroup1_procs_write, 618 }, 619 { 620 .name = "cgroup.clone_children", 621 .read_u64 = cgroup_clone_children_read, 622 .write_u64 = cgroup_clone_children_write, 623 }, 624 { 625 .name = "cgroup.sane_behavior", 626 .flags = CFTYPE_ONLY_ON_ROOT, 627 .seq_show = cgroup_sane_behavior_show, 628 }, 629 { 630 .name = "tasks", 631 .seq_start = cgroup_pidlist_start, 632 .seq_next = cgroup_pidlist_next, 633 .seq_stop = cgroup_pidlist_stop, 634 .seq_show = cgroup_pidlist_show, 635 .private = CGROUP_FILE_TASKS, 636 .write = cgroup1_tasks_write, 637 }, 638 { 639 .name = "notify_on_release", 640 .read_u64 = cgroup_read_notify_on_release, 641 .write_u64 = cgroup_write_notify_on_release, 642 }, 643 { 644 .name = "release_agent", 645 .flags = CFTYPE_ONLY_ON_ROOT, 646 .seq_show = cgroup_release_agent_show, 647 .write = cgroup_release_agent_write, 648 .max_write_len = PATH_MAX - 1, 649 }, 650 { } /* terminate */ 651 }; 652 653 /* Display information about each subsystem and each hierarchy */ 654 int proc_cgroupstats_show(struct seq_file *m, void *v) 655 { 656 struct cgroup_subsys *ss; 657 int i; 658 659 seq_puts(m, "#subsys_name\thierarchy\tnum_cgroups\tenabled\n"); 660 /* 661 * ideally we don't want subsystems moving around while we do this. 662 * cgroup_mutex is also necessary to guarantee an atomic snapshot of 663 * subsys/hierarchy state. 664 */ 665 mutex_lock(&cgroup_mutex); 666 667 for_each_subsys(ss, i) 668 seq_printf(m, "%s\t%d\t%d\t%d\n", 669 ss->legacy_name, ss->root->hierarchy_id, 670 atomic_read(&ss->root->nr_cgrps), 671 cgroup_ssid_enabled(i)); 672 673 mutex_unlock(&cgroup_mutex); 674 return 0; 675 } 676 677 /** 678 * cgroupstats_build - build and fill cgroupstats 679 * @stats: cgroupstats to fill information into 680 * @dentry: A dentry entry belonging to the cgroup for which stats have 681 * been requested. 682 * 683 * Build and fill cgroupstats so that taskstats can export it to user 684 * space. 685 */ 686 int cgroupstats_build(struct cgroupstats *stats, struct dentry *dentry) 687 { 688 struct kernfs_node *kn = kernfs_node_from_dentry(dentry); 689 struct cgroup *cgrp; 690 struct css_task_iter it; 691 struct task_struct *tsk; 692 693 /* it should be kernfs_node belonging to cgroupfs and is a directory */ 694 if (dentry->d_sb->s_type != &cgroup_fs_type || !kn || 695 kernfs_type(kn) != KERNFS_DIR) 696 return -EINVAL; 697 698 mutex_lock(&cgroup_mutex); 699 700 /* 701 * We aren't being called from kernfs and there's no guarantee on 702 * @kn->priv's validity. For this and css_tryget_online_from_dir(), 703 * @kn->priv is RCU safe. Let's do the RCU dancing. 704 */ 705 rcu_read_lock(); 706 cgrp = rcu_dereference(*(void __rcu __force **)&kn->priv); 707 if (!cgrp || cgroup_is_dead(cgrp)) { 708 rcu_read_unlock(); 709 mutex_unlock(&cgroup_mutex); 710 return -ENOENT; 711 } 712 rcu_read_unlock(); 713 714 css_task_iter_start(&cgrp->self, 0, &it); 715 while ((tsk = css_task_iter_next(&it))) { 716 switch (tsk->state) { 717 case TASK_RUNNING: 718 stats->nr_running++; 719 break; 720 case TASK_INTERRUPTIBLE: 721 stats->nr_sleeping++; 722 break; 723 case TASK_UNINTERRUPTIBLE: 724 stats->nr_uninterruptible++; 725 break; 726 case TASK_STOPPED: 727 stats->nr_stopped++; 728 break; 729 default: 730 if (tsk->in_iowait) 731 stats->nr_io_wait++; 732 break; 733 } 734 } 735 css_task_iter_end(&it); 736 737 mutex_unlock(&cgroup_mutex); 738 return 0; 739 } 740 741 void cgroup1_check_for_release(struct cgroup *cgrp) 742 { 743 if (notify_on_release(cgrp) && !cgroup_is_populated(cgrp) && 744 !css_has_online_children(&cgrp->self) && !cgroup_is_dead(cgrp)) 745 schedule_work(&cgrp->release_agent_work); 746 } 747 748 /* 749 * Notify userspace when a cgroup is released, by running the 750 * configured release agent with the name of the cgroup (path 751 * relative to the root of cgroup file system) as the argument. 752 * 753 * Most likely, this user command will try to rmdir this cgroup. 754 * 755 * This races with the possibility that some other task will be 756 * attached to this cgroup before it is removed, or that some other 757 * user task will 'mkdir' a child cgroup of this cgroup. That's ok. 758 * The presumed 'rmdir' will fail quietly if this cgroup is no longer 759 * unused, and this cgroup will be reprieved from its death sentence, 760 * to continue to serve a useful existence. Next time it's released, 761 * we will get notified again, if it still has 'notify_on_release' set. 762 * 763 * The final arg to call_usermodehelper() is UMH_WAIT_EXEC, which 764 * means only wait until the task is successfully execve()'d. The 765 * separate release agent task is forked by call_usermodehelper(), 766 * then control in this thread returns here, without waiting for the 767 * release agent task. We don't bother to wait because the caller of 768 * this routine has no use for the exit status of the release agent 769 * task, so no sense holding our caller up for that. 770 */ 771 void cgroup1_release_agent(struct work_struct *work) 772 { 773 struct cgroup *cgrp = 774 container_of(work, struct cgroup, release_agent_work); 775 char *pathbuf, *agentbuf; 776 char *argv[3], *envp[3]; 777 int ret; 778 779 /* snoop agent path and exit early if empty */ 780 if (!cgrp->root->release_agent_path[0]) 781 return; 782 783 /* prepare argument buffers */ 784 pathbuf = kmalloc(PATH_MAX, GFP_KERNEL); 785 agentbuf = kmalloc(PATH_MAX, GFP_KERNEL); 786 if (!pathbuf || !agentbuf) 787 goto out_free; 788 789 spin_lock(&release_agent_path_lock); 790 strlcpy(agentbuf, cgrp->root->release_agent_path, PATH_MAX); 791 spin_unlock(&release_agent_path_lock); 792 if (!agentbuf[0]) 793 goto out_free; 794 795 ret = cgroup_path_ns(cgrp, pathbuf, PATH_MAX, &init_cgroup_ns); 796 if (ret < 0 || ret >= PATH_MAX) 797 goto out_free; 798 799 argv[0] = agentbuf; 800 argv[1] = pathbuf; 801 argv[2] = NULL; 802 803 /* minimal command environment */ 804 envp[0] = "HOME=/"; 805 envp[1] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin"; 806 envp[2] = NULL; 807 808 call_usermodehelper(argv[0], argv, envp, UMH_WAIT_EXEC); 809 out_free: 810 kfree(agentbuf); 811 kfree(pathbuf); 812 } 813 814 /* 815 * cgroup_rename - Only allow simple rename of directories in place. 816 */ 817 static int cgroup1_rename(struct kernfs_node *kn, struct kernfs_node *new_parent, 818 const char *new_name_str) 819 { 820 struct cgroup *cgrp = kn->priv; 821 int ret; 822 823 if (kernfs_type(kn) != KERNFS_DIR) 824 return -ENOTDIR; 825 if (kn->parent != new_parent) 826 return -EIO; 827 828 /* 829 * We're gonna grab cgroup_mutex which nests outside kernfs 830 * active_ref. kernfs_rename() doesn't require active_ref 831 * protection. Break them before grabbing cgroup_mutex. 832 */ 833 kernfs_break_active_protection(new_parent); 834 kernfs_break_active_protection(kn); 835 836 mutex_lock(&cgroup_mutex); 837 838 ret = kernfs_rename(kn, new_parent, new_name_str); 839 if (!ret) 840 TRACE_CGROUP_PATH(rename, cgrp); 841 842 mutex_unlock(&cgroup_mutex); 843 844 kernfs_unbreak_active_protection(kn); 845 kernfs_unbreak_active_protection(new_parent); 846 return ret; 847 } 848 849 static int cgroup1_show_options(struct seq_file *seq, struct kernfs_root *kf_root) 850 { 851 struct cgroup_root *root = cgroup_root_from_kf(kf_root); 852 struct cgroup_subsys *ss; 853 int ssid; 854 855 for_each_subsys(ss, ssid) 856 if (root->subsys_mask & (1 << ssid)) 857 seq_show_option(seq, ss->legacy_name, NULL); 858 if (root->flags & CGRP_ROOT_NOPREFIX) 859 seq_puts(seq, ",noprefix"); 860 if (root->flags & CGRP_ROOT_XATTR) 861 seq_puts(seq, ",xattr"); 862 if (root->flags & CGRP_ROOT_CPUSET_V2_MODE) 863 seq_puts(seq, ",cpuset_v2_mode"); 864 865 spin_lock(&release_agent_path_lock); 866 if (strlen(root->release_agent_path)) 867 seq_show_option(seq, "release_agent", 868 root->release_agent_path); 869 spin_unlock(&release_agent_path_lock); 870 871 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags)) 872 seq_puts(seq, ",clone_children"); 873 if (strlen(root->name)) 874 seq_show_option(seq, "name", root->name); 875 return 0; 876 } 877 878 enum cgroup1_param { 879 Opt_all, 880 Opt_clone_children, 881 Opt_cpuset_v2_mode, 882 Opt_name, 883 Opt_none, 884 Opt_noprefix, 885 Opt_release_agent, 886 Opt_xattr, 887 }; 888 889 const struct fs_parameter_spec cgroup1_fs_parameters[] = { 890 fsparam_flag ("all", Opt_all), 891 fsparam_flag ("clone_children", Opt_clone_children), 892 fsparam_flag ("cpuset_v2_mode", Opt_cpuset_v2_mode), 893 fsparam_string("name", Opt_name), 894 fsparam_flag ("none", Opt_none), 895 fsparam_flag ("noprefix", Opt_noprefix), 896 fsparam_string("release_agent", Opt_release_agent), 897 fsparam_flag ("xattr", Opt_xattr), 898 {} 899 }; 900 901 int cgroup1_parse_param(struct fs_context *fc, struct fs_parameter *param) 902 { 903 struct cgroup_fs_context *ctx = cgroup_fc2context(fc); 904 struct cgroup_subsys *ss; 905 struct fs_parse_result result; 906 int opt, i; 907 908 opt = fs_parse(fc, cgroup1_fs_parameters, param, &result); 909 if (opt == -ENOPARAM) { 910 if (strcmp(param->key, "source") == 0) { 911 if (fc->source) 912 return invalf(fc, "Multiple sources not supported"); 913 fc->source = param->string; 914 param->string = NULL; 915 return 0; 916 } 917 for_each_subsys(ss, i) { 918 if (strcmp(param->key, ss->legacy_name)) 919 continue; 920 if (!cgroup_ssid_enabled(i) || cgroup1_ssid_disabled(i)) 921 return invalfc(fc, "Disabled controller '%s'", 922 param->key); 923 ctx->subsys_mask |= (1 << i); 924 return 0; 925 } 926 return invalfc(fc, "Unknown subsys name '%s'", param->key); 927 } 928 if (opt < 0) 929 return opt; 930 931 switch (opt) { 932 case Opt_none: 933 /* Explicitly have no subsystems */ 934 ctx->none = true; 935 break; 936 case Opt_all: 937 ctx->all_ss = true; 938 break; 939 case Opt_noprefix: 940 ctx->flags |= CGRP_ROOT_NOPREFIX; 941 break; 942 case Opt_clone_children: 943 ctx->cpuset_clone_children = true; 944 break; 945 case Opt_cpuset_v2_mode: 946 ctx->flags |= CGRP_ROOT_CPUSET_V2_MODE; 947 break; 948 case Opt_xattr: 949 ctx->flags |= CGRP_ROOT_XATTR; 950 break; 951 case Opt_release_agent: 952 /* Specifying two release agents is forbidden */ 953 if (ctx->release_agent) 954 return invalfc(fc, "release_agent respecified"); 955 ctx->release_agent = param->string; 956 param->string = NULL; 957 break; 958 case Opt_name: 959 /* blocked by boot param? */ 960 if (cgroup_no_v1_named) 961 return -ENOENT; 962 /* Can't specify an empty name */ 963 if (!param->size) 964 return invalfc(fc, "Empty name"); 965 if (param->size > MAX_CGROUP_ROOT_NAMELEN - 1) 966 return invalfc(fc, "Name too long"); 967 /* Must match [\w.-]+ */ 968 for (i = 0; i < param->size; i++) { 969 char c = param->string[i]; 970 if (isalnum(c)) 971 continue; 972 if ((c == '.') || (c == '-') || (c == '_')) 973 continue; 974 return invalfc(fc, "Invalid name"); 975 } 976 /* Specifying two names is forbidden */ 977 if (ctx->name) 978 return invalfc(fc, "name respecified"); 979 ctx->name = param->string; 980 param->string = NULL; 981 break; 982 } 983 return 0; 984 } 985 986 static int check_cgroupfs_options(struct fs_context *fc) 987 { 988 struct cgroup_fs_context *ctx = cgroup_fc2context(fc); 989 u16 mask = U16_MAX; 990 u16 enabled = 0; 991 struct cgroup_subsys *ss; 992 int i; 993 994 #ifdef CONFIG_CPUSETS 995 mask = ~((u16)1 << cpuset_cgrp_id); 996 #endif 997 for_each_subsys(ss, i) 998 if (cgroup_ssid_enabled(i) && !cgroup1_ssid_disabled(i)) 999 enabled |= 1 << i; 1000 1001 ctx->subsys_mask &= enabled; 1002 1003 /* 1004 * In absense of 'none', 'name=' or subsystem name options, 1005 * let's default to 'all'. 1006 */ 1007 if (!ctx->subsys_mask && !ctx->none && !ctx->name) 1008 ctx->all_ss = true; 1009 1010 if (ctx->all_ss) { 1011 /* Mutually exclusive option 'all' + subsystem name */ 1012 if (ctx->subsys_mask) 1013 return invalfc(fc, "subsys name conflicts with all"); 1014 /* 'all' => select all the subsystems */ 1015 ctx->subsys_mask = enabled; 1016 } 1017 1018 /* 1019 * We either have to specify by name or by subsystems. (So all 1020 * empty hierarchies must have a name). 1021 */ 1022 if (!ctx->subsys_mask && !ctx->name) 1023 return invalfc(fc, "Need name or subsystem set"); 1024 1025 /* 1026 * Option noprefix was introduced just for backward compatibility 1027 * with the old cpuset, so we allow noprefix only if mounting just 1028 * the cpuset subsystem. 1029 */ 1030 if ((ctx->flags & CGRP_ROOT_NOPREFIX) && (ctx->subsys_mask & mask)) 1031 return invalfc(fc, "noprefix used incorrectly"); 1032 1033 /* Can't specify "none" and some subsystems */ 1034 if (ctx->subsys_mask && ctx->none) 1035 return invalfc(fc, "none used incorrectly"); 1036 1037 return 0; 1038 } 1039 1040 int cgroup1_reconfigure(struct fs_context *fc) 1041 { 1042 struct cgroup_fs_context *ctx = cgroup_fc2context(fc); 1043 struct kernfs_root *kf_root = kernfs_root_from_sb(fc->root->d_sb); 1044 struct cgroup_root *root = cgroup_root_from_kf(kf_root); 1045 int ret = 0; 1046 u16 added_mask, removed_mask; 1047 1048 cgroup_lock_and_drain_offline(&cgrp_dfl_root.cgrp); 1049 1050 /* See what subsystems are wanted */ 1051 ret = check_cgroupfs_options(fc); 1052 if (ret) 1053 goto out_unlock; 1054 1055 if (ctx->subsys_mask != root->subsys_mask || ctx->release_agent) 1056 pr_warn("option changes via remount are deprecated (pid=%d comm=%s)\n", 1057 task_tgid_nr(current), current->comm); 1058 1059 added_mask = ctx->subsys_mask & ~root->subsys_mask; 1060 removed_mask = root->subsys_mask & ~ctx->subsys_mask; 1061 1062 /* Don't allow flags or name to change at remount */ 1063 if ((ctx->flags ^ root->flags) || 1064 (ctx->name && strcmp(ctx->name, root->name))) { 1065 errorfc(fc, "option or name mismatch, new: 0x%x \"%s\", old: 0x%x \"%s\"", 1066 ctx->flags, ctx->name ?: "", root->flags, root->name); 1067 ret = -EINVAL; 1068 goto out_unlock; 1069 } 1070 1071 /* remounting is not allowed for populated hierarchies */ 1072 if (!list_empty(&root->cgrp.self.children)) { 1073 ret = -EBUSY; 1074 goto out_unlock; 1075 } 1076 1077 ret = rebind_subsystems(root, added_mask); 1078 if (ret) 1079 goto out_unlock; 1080 1081 WARN_ON(rebind_subsystems(&cgrp_dfl_root, removed_mask)); 1082 1083 if (ctx->release_agent) { 1084 spin_lock(&release_agent_path_lock); 1085 strcpy(root->release_agent_path, ctx->release_agent); 1086 spin_unlock(&release_agent_path_lock); 1087 } 1088 1089 trace_cgroup_remount(root); 1090 1091 out_unlock: 1092 mutex_unlock(&cgroup_mutex); 1093 return ret; 1094 } 1095 1096 struct kernfs_syscall_ops cgroup1_kf_syscall_ops = { 1097 .rename = cgroup1_rename, 1098 .show_options = cgroup1_show_options, 1099 .mkdir = cgroup_mkdir, 1100 .rmdir = cgroup_rmdir, 1101 .show_path = cgroup_show_path, 1102 }; 1103 1104 /* 1105 * The guts of cgroup1 mount - find or create cgroup_root to use. 1106 * Called with cgroup_mutex held; returns 0 on success, -E... on 1107 * error and positive - in case when the candidate is busy dying. 1108 * On success it stashes a reference to cgroup_root into given 1109 * cgroup_fs_context; that reference is *NOT* counting towards the 1110 * cgroup_root refcount. 1111 */ 1112 static int cgroup1_root_to_use(struct fs_context *fc) 1113 { 1114 struct cgroup_fs_context *ctx = cgroup_fc2context(fc); 1115 struct cgroup_root *root; 1116 struct cgroup_subsys *ss; 1117 int i, ret; 1118 1119 /* First find the desired set of subsystems */ 1120 ret = check_cgroupfs_options(fc); 1121 if (ret) 1122 return ret; 1123 1124 /* 1125 * Destruction of cgroup root is asynchronous, so subsystems may 1126 * still be dying after the previous unmount. Let's drain the 1127 * dying subsystems. We just need to ensure that the ones 1128 * unmounted previously finish dying and don't care about new ones 1129 * starting. Testing ref liveliness is good enough. 1130 */ 1131 for_each_subsys(ss, i) { 1132 if (!(ctx->subsys_mask & (1 << i)) || 1133 ss->root == &cgrp_dfl_root) 1134 continue; 1135 1136 if (!percpu_ref_tryget_live(&ss->root->cgrp.self.refcnt)) 1137 return 1; /* restart */ 1138 cgroup_put(&ss->root->cgrp); 1139 } 1140 1141 for_each_root(root) { 1142 bool name_match = false; 1143 1144 if (root == &cgrp_dfl_root) 1145 continue; 1146 1147 /* 1148 * If we asked for a name then it must match. Also, if 1149 * name matches but sybsys_mask doesn't, we should fail. 1150 * Remember whether name matched. 1151 */ 1152 if (ctx->name) { 1153 if (strcmp(ctx->name, root->name)) 1154 continue; 1155 name_match = true; 1156 } 1157 1158 /* 1159 * If we asked for subsystems (or explicitly for no 1160 * subsystems) then they must match. 1161 */ 1162 if ((ctx->subsys_mask || ctx->none) && 1163 (ctx->subsys_mask != root->subsys_mask)) { 1164 if (!name_match) 1165 continue; 1166 return -EBUSY; 1167 } 1168 1169 if (root->flags ^ ctx->flags) 1170 pr_warn("new mount options do not match the existing superblock, will be ignored\n"); 1171 1172 ctx->root = root; 1173 return 0; 1174 } 1175 1176 /* 1177 * No such thing, create a new one. name= matching without subsys 1178 * specification is allowed for already existing hierarchies but we 1179 * can't create new one without subsys specification. 1180 */ 1181 if (!ctx->subsys_mask && !ctx->none) 1182 return invalfc(fc, "No subsys list or none specified"); 1183 1184 /* Hierarchies may only be created in the initial cgroup namespace. */ 1185 if (ctx->ns != &init_cgroup_ns) 1186 return -EPERM; 1187 1188 root = kzalloc(sizeof(*root), GFP_KERNEL); 1189 if (!root) 1190 return -ENOMEM; 1191 1192 ctx->root = root; 1193 init_cgroup_root(ctx); 1194 1195 ret = cgroup_setup_root(root, ctx->subsys_mask); 1196 if (ret) 1197 cgroup_free_root(root); 1198 return ret; 1199 } 1200 1201 int cgroup1_get_tree(struct fs_context *fc) 1202 { 1203 struct cgroup_fs_context *ctx = cgroup_fc2context(fc); 1204 int ret; 1205 1206 /* Check if the caller has permission to mount. */ 1207 if (!ns_capable(ctx->ns->user_ns, CAP_SYS_ADMIN)) 1208 return -EPERM; 1209 1210 cgroup_lock_and_drain_offline(&cgrp_dfl_root.cgrp); 1211 1212 ret = cgroup1_root_to_use(fc); 1213 if (!ret && !percpu_ref_tryget_live(&ctx->root->cgrp.self.refcnt)) 1214 ret = 1; /* restart */ 1215 1216 mutex_unlock(&cgroup_mutex); 1217 1218 if (!ret) 1219 ret = cgroup_do_get_tree(fc); 1220 1221 if (!ret && percpu_ref_is_dying(&ctx->root->cgrp.self.refcnt)) { 1222 struct super_block *sb = fc->root->d_sb; 1223 dput(fc->root); 1224 deactivate_locked_super(sb); 1225 ret = 1; 1226 } 1227 1228 if (unlikely(ret > 0)) { 1229 msleep(10); 1230 return restart_syscall(); 1231 } 1232 return ret; 1233 } 1234 1235 static int __init cgroup1_wq_init(void) 1236 { 1237 /* 1238 * Used to destroy pidlists and separate to serve as flush domain. 1239 * Cap @max_active to 1 too. 1240 */ 1241 cgroup_pidlist_destroy_wq = alloc_workqueue("cgroup_pidlist_destroy", 1242 0, 1); 1243 BUG_ON(!cgroup_pidlist_destroy_wq); 1244 return 0; 1245 } 1246 core_initcall(cgroup1_wq_init); 1247 1248 static int __init cgroup_no_v1(char *str) 1249 { 1250 struct cgroup_subsys *ss; 1251 char *token; 1252 int i; 1253 1254 while ((token = strsep(&str, ",")) != NULL) { 1255 if (!*token) 1256 continue; 1257 1258 if (!strcmp(token, "all")) { 1259 cgroup_no_v1_mask = U16_MAX; 1260 continue; 1261 } 1262 1263 if (!strcmp(token, "named")) { 1264 cgroup_no_v1_named = true; 1265 continue; 1266 } 1267 1268 for_each_subsys(ss, i) { 1269 if (strcmp(token, ss->name) && 1270 strcmp(token, ss->legacy_name)) 1271 continue; 1272 1273 cgroup_no_v1_mask |= 1 << i; 1274 } 1275 } 1276 return 1; 1277 } 1278 __setup("cgroup_no_v1=", cgroup_no_v1); 1279