~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

TOMOYO Linux Cross Reference
Linux/kernel/cgroup/cpuset.c

Version: ~ [ linux-5.13-rc5 ] ~ [ linux-5.12.9 ] ~ [ linux-5.11.22 ] ~ [ linux-5.10.42 ] ~ [ linux-5.9.16 ] ~ [ linux-5.8.18 ] ~ [ linux-5.7.19 ] ~ [ linux-5.6.19 ] ~ [ linux-5.5.19 ] ~ [ linux-5.4.124 ] ~ [ linux-5.3.18 ] ~ [ linux-5.2.21 ] ~ [ linux-5.1.21 ] ~ [ linux-5.0.21 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.193 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.235 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.271 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.271 ] ~ [ linux-4.3.6 ] ~ [ linux-4.2.8 ] ~ [ linux-4.1.52 ] ~ [ linux-4.0.9 ] ~ [ linux-3.18.140 ] ~ [ linux-3.16.85 ] ~ [ linux-3.14.79 ] ~ [ linux-3.12.74 ] ~ [ linux-3.10.108 ] ~ [ linux-2.6.32.71 ] ~ [ linux-2.6.0 ] ~ [ linux-2.4.37.11 ] ~ [ unix-v6-master ] ~ [ ccs-tools-1.8.5 ] ~ [ policy-sample ] ~
Architecture: ~ [ i386 ] ~ [ alpha ] ~ [ m68k ] ~ [ mips ] ~ [ ppc ] ~ [ sparc ] ~ [ sparc64 ] ~

  1 /*
  2  *  kernel/cpuset.c
  3  *
  4  *  Processor and Memory placement constraints for sets of tasks.
  5  *
  6  *  Copyright (C) 2003 BULL SA.
  7  *  Copyright (C) 2004-2007 Silicon Graphics, Inc.
  8  *  Copyright (C) 2006 Google, Inc
  9  *
 10  *  Portions derived from Patrick Mochel's sysfs code.
 11  *  sysfs is Copyright (c) 2001-3 Patrick Mochel
 12  *
 13  *  2003-10-10 Written by Simon Derr.
 14  *  2003-10-22 Updates by Stephen Hemminger.
 15  *  2004 May-July Rework by Paul Jackson.
 16  *  2006 Rework by Paul Menage to use generic cgroups
 17  *  2008 Rework of the scheduler domains and CPU hotplug handling
 18  *       by Max Krasnyansky
 19  *
 20  *  This file is subject to the terms and conditions of the GNU General Public
 21  *  License.  See the file COPYING in the main directory of the Linux
 22  *  distribution for more details.
 23  */
 24 
 25 #include <linux/cpu.h>
 26 #include <linux/cpumask.h>
 27 #include <linux/cpuset.h>
 28 #include <linux/err.h>
 29 #include <linux/errno.h>
 30 #include <linux/file.h>
 31 #include <linux/fs.h>
 32 #include <linux/init.h>
 33 #include <linux/interrupt.h>
 34 #include <linux/kernel.h>
 35 #include <linux/kmod.h>
 36 #include <linux/list.h>
 37 #include <linux/mempolicy.h>
 38 #include <linux/mm.h>
 39 #include <linux/memory.h>
 40 #include <linux/export.h>
 41 #include <linux/mount.h>
 42 #include <linux/fs_context.h>
 43 #include <linux/namei.h>
 44 #include <linux/pagemap.h>
 45 #include <linux/proc_fs.h>
 46 #include <linux/rcupdate.h>
 47 #include <linux/sched.h>
 48 #include <linux/sched/deadline.h>
 49 #include <linux/sched/mm.h>
 50 #include <linux/sched/task.h>
 51 #include <linux/seq_file.h>
 52 #include <linux/security.h>
 53 #include <linux/slab.h>
 54 #include <linux/spinlock.h>
 55 #include <linux/stat.h>
 56 #include <linux/string.h>
 57 #include <linux/time.h>
 58 #include <linux/time64.h>
 59 #include <linux/backing-dev.h>
 60 #include <linux/sort.h>
 61 #include <linux/oom.h>
 62 #include <linux/sched/isolation.h>
 63 #include <linux/uaccess.h>
 64 #include <linux/atomic.h>
 65 #include <linux/mutex.h>
 66 #include <linux/cgroup.h>
 67 #include <linux/wait.h>
 68 
 69 DEFINE_STATIC_KEY_FALSE(cpusets_pre_enable_key);
 70 DEFINE_STATIC_KEY_FALSE(cpusets_enabled_key);
 71 
 72 /* See "Frequency meter" comments, below. */
 73 
 74 struct fmeter {
 75         int cnt;                /* unprocessed events count */
 76         int val;                /* most recent output value */
 77         time64_t time;          /* clock (secs) when val computed */
 78         spinlock_t lock;        /* guards read or write of above */
 79 };
 80 
 81 struct cpuset {
 82         struct cgroup_subsys_state css;
 83 
 84         unsigned long flags;            /* "unsigned long" so bitops work */
 85 
 86         /*
 87          * On default hierarchy:
 88          *
 89          * The user-configured masks can only be changed by writing to
 90          * cpuset.cpus and cpuset.mems, and won't be limited by the
 91          * parent masks.
 92          *
 93          * The effective masks is the real masks that apply to the tasks
 94          * in the cpuset. They may be changed if the configured masks are
 95          * changed or hotplug happens.
 96          *
 97          * effective_mask == configured_mask & parent's effective_mask,
 98          * and if it ends up empty, it will inherit the parent's mask.
 99          *
100          *
101          * On legacy hierarchy:
102          *
103          * The user-configured masks are always the same with effective masks.
104          */
105 
106         /* user-configured CPUs and Memory Nodes allow to tasks */
107         cpumask_var_t cpus_allowed;
108         nodemask_t mems_allowed;
109 
110         /* effective CPUs and Memory Nodes allow to tasks */
111         cpumask_var_t effective_cpus;
112         nodemask_t effective_mems;
113 
114         /*
115          * CPUs allocated to child sub-partitions (default hierarchy only)
116          * - CPUs granted by the parent = effective_cpus U subparts_cpus
117          * - effective_cpus and subparts_cpus are mutually exclusive.
118          *
119          * effective_cpus contains only onlined CPUs, but subparts_cpus
120          * may have offlined ones.
121          */
122         cpumask_var_t subparts_cpus;
123 
124         /*
125          * This is old Memory Nodes tasks took on.
126          *
127          * - top_cpuset.old_mems_allowed is initialized to mems_allowed.
128          * - A new cpuset's old_mems_allowed is initialized when some
129          *   task is moved into it.
130          * - old_mems_allowed is used in cpuset_migrate_mm() when we change
131          *   cpuset.mems_allowed and have tasks' nodemask updated, and
132          *   then old_mems_allowed is updated to mems_allowed.
133          */
134         nodemask_t old_mems_allowed;
135 
136         struct fmeter fmeter;           /* memory_pressure filter */
137 
138         /*
139          * Tasks are being attached to this cpuset.  Used to prevent
140          * zeroing cpus/mems_allowed between ->can_attach() and ->attach().
141          */
142         int attach_in_progress;
143 
144         /* partition number for rebuild_sched_domains() */
145         int pn;
146 
147         /* for custom sched domain */
148         int relax_domain_level;
149 
150         /* number of CPUs in subparts_cpus */
151         int nr_subparts_cpus;
152 
153         /* partition root state */
154         int partition_root_state;
155 
156         /*
157          * Default hierarchy only:
158          * use_parent_ecpus - set if using parent's effective_cpus
159          * child_ecpus_count - # of children with use_parent_ecpus set
160          */
161         int use_parent_ecpus;
162         int child_ecpus_count;
163 };
164 
165 /*
166  * Partition root states:
167  *
168  *   0 - not a partition root
169  *
170  *   1 - partition root
171  *
172  *  -1 - invalid partition root
173  *       None of the cpus in cpus_allowed can be put into the parent's
174  *       subparts_cpus. In this case, the cpuset is not a real partition
175  *       root anymore.  However, the CPU_EXCLUSIVE bit will still be set
176  *       and the cpuset can be restored back to a partition root if the
177  *       parent cpuset can give more CPUs back to this child cpuset.
178  */
179 #define PRS_DISABLED            0
180 #define PRS_ENABLED             1
181 #define PRS_ERROR               -1
182 
183 /*
184  * Temporary cpumasks for working with partitions that are passed among
185  * functions to avoid memory allocation in inner functions.
186  */
187 struct tmpmasks {
188         cpumask_var_t addmask, delmask; /* For partition root */
189         cpumask_var_t new_cpus;         /* For update_cpumasks_hier() */
190 };
191 
192 static inline struct cpuset *css_cs(struct cgroup_subsys_state *css)
193 {
194         return css ? container_of(css, struct cpuset, css) : NULL;
195 }
196 
197 /* Retrieve the cpuset for a task */
198 static inline struct cpuset *task_cs(struct task_struct *task)
199 {
200         return css_cs(task_css(task, cpuset_cgrp_id));
201 }
202 
203 static inline struct cpuset *parent_cs(struct cpuset *cs)
204 {
205         return css_cs(cs->css.parent);
206 }
207 
208 /* bits in struct cpuset flags field */
209 typedef enum {
210         CS_ONLINE,
211         CS_CPU_EXCLUSIVE,
212         CS_MEM_EXCLUSIVE,
213         CS_MEM_HARDWALL,
214         CS_MEMORY_MIGRATE,
215         CS_SCHED_LOAD_BALANCE,
216         CS_SPREAD_PAGE,
217         CS_SPREAD_SLAB,
218 } cpuset_flagbits_t;
219 
220 /* convenient tests for these bits */
221 static inline bool is_cpuset_online(struct cpuset *cs)
222 {
223         return test_bit(CS_ONLINE, &cs->flags) && !css_is_dying(&cs->css);
224 }
225 
226 static inline int is_cpu_exclusive(const struct cpuset *cs)
227 {
228         return test_bit(CS_CPU_EXCLUSIVE, &cs->flags);
229 }
230 
231 static inline int is_mem_exclusive(const struct cpuset *cs)
232 {
233         return test_bit(CS_MEM_EXCLUSIVE, &cs->flags);
234 }
235 
236 static inline int is_mem_hardwall(const struct cpuset *cs)
237 {
238         return test_bit(CS_MEM_HARDWALL, &cs->flags);
239 }
240 
241 static inline int is_sched_load_balance(const struct cpuset *cs)
242 {
243         return test_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
244 }
245 
246 static inline int is_memory_migrate(const struct cpuset *cs)
247 {
248         return test_bit(CS_MEMORY_MIGRATE, &cs->flags);
249 }
250 
251 static inline int is_spread_page(const struct cpuset *cs)
252 {
253         return test_bit(CS_SPREAD_PAGE, &cs->flags);
254 }
255 
256 static inline int is_spread_slab(const struct cpuset *cs)
257 {
258         return test_bit(CS_SPREAD_SLAB, &cs->flags);
259 }
260 
261 static inline int is_partition_root(const struct cpuset *cs)
262 {
263         return cs->partition_root_state > 0;
264 }
265 
266 static struct cpuset top_cpuset = {
267         .flags = ((1 << CS_ONLINE) | (1 << CS_CPU_EXCLUSIVE) |
268                   (1 << CS_MEM_EXCLUSIVE)),
269         .partition_root_state = PRS_ENABLED,
270 };
271 
272 /**
273  * cpuset_for_each_child - traverse online children of a cpuset
274  * @child_cs: loop cursor pointing to the current child
275  * @pos_css: used for iteration
276  * @parent_cs: target cpuset to walk children of
277  *
278  * Walk @child_cs through the online children of @parent_cs.  Must be used
279  * with RCU read locked.
280  */
281 #define cpuset_for_each_child(child_cs, pos_css, parent_cs)             \
282         css_for_each_child((pos_css), &(parent_cs)->css)                \
283                 if (is_cpuset_online(((child_cs) = css_cs((pos_css)))))
284 
285 /**
286  * cpuset_for_each_descendant_pre - pre-order walk of a cpuset's descendants
287  * @des_cs: loop cursor pointing to the current descendant
288  * @pos_css: used for iteration
289  * @root_cs: target cpuset to walk ancestor of
290  *
291  * Walk @des_cs through the online descendants of @root_cs.  Must be used
292  * with RCU read locked.  The caller may modify @pos_css by calling
293  * css_rightmost_descendant() to skip subtree.  @root_cs is included in the
294  * iteration and the first node to be visited.
295  */
296 #define cpuset_for_each_descendant_pre(des_cs, pos_css, root_cs)        \
297         css_for_each_descendant_pre((pos_css), &(root_cs)->css)         \
298                 if (is_cpuset_online(((des_cs) = css_cs((pos_css)))))
299 
300 /*
301  * There are two global locks guarding cpuset structures - cpuset_mutex and
302  * callback_lock. We also require taking task_lock() when dereferencing a
303  * task's cpuset pointer. See "The task_lock() exception", at the end of this
304  * comment.
305  *
306  * A task must hold both locks to modify cpusets.  If a task holds
307  * cpuset_mutex, then it blocks others wanting that mutex, ensuring that it
308  * is the only task able to also acquire callback_lock and be able to
309  * modify cpusets.  It can perform various checks on the cpuset structure
310  * first, knowing nothing will change.  It can also allocate memory while
311  * just holding cpuset_mutex.  While it is performing these checks, various
312  * callback routines can briefly acquire callback_lock to query cpusets.
313  * Once it is ready to make the changes, it takes callback_lock, blocking
314  * everyone else.
315  *
316  * Calls to the kernel memory allocator can not be made while holding
317  * callback_lock, as that would risk double tripping on callback_lock
318  * from one of the callbacks into the cpuset code from within
319  * __alloc_pages().
320  *
321  * If a task is only holding callback_lock, then it has read-only
322  * access to cpusets.
323  *
324  * Now, the task_struct fields mems_allowed and mempolicy may be changed
325  * by other task, we use alloc_lock in the task_struct fields to protect
326  * them.
327  *
328  * The cpuset_common_file_read() handlers only hold callback_lock across
329  * small pieces of code, such as when reading out possibly multi-word
330  * cpumasks and nodemasks.
331  *
332  * Accessing a task's cpuset should be done in accordance with the
333  * guidelines for accessing subsystem state in kernel/cgroup.c
334  */
335 
336 DEFINE_STATIC_PERCPU_RWSEM(cpuset_rwsem);
337 
338 void cpuset_read_lock(void)
339 {
340         percpu_down_read(&cpuset_rwsem);
341 }
342 
343 void cpuset_read_unlock(void)
344 {
345         percpu_up_read(&cpuset_rwsem);
346 }
347 
348 static DEFINE_SPINLOCK(callback_lock);
349 
350 static struct workqueue_struct *cpuset_migrate_mm_wq;
351 
352 /*
353  * CPU / memory hotplug is handled asynchronously.
354  */
355 static void cpuset_hotplug_workfn(struct work_struct *work);
356 static DECLARE_WORK(cpuset_hotplug_work, cpuset_hotplug_workfn);
357 
358 static DECLARE_WAIT_QUEUE_HEAD(cpuset_attach_wq);
359 
360 /*
361  * Cgroup v2 behavior is used on the "cpus" and "mems" control files when
362  * on default hierarchy or when the cpuset_v2_mode flag is set by mounting
363  * the v1 cpuset cgroup filesystem with the "cpuset_v2_mode" mount option.
364  * With v2 behavior, "cpus" and "mems" are always what the users have
365  * requested and won't be changed by hotplug events. Only the effective
366  * cpus or mems will be affected.
367  */
368 static inline bool is_in_v2_mode(void)
369 {
370         return cgroup_subsys_on_dfl(cpuset_cgrp_subsys) ||
371               (cpuset_cgrp_subsys.root->flags & CGRP_ROOT_CPUSET_V2_MODE);
372 }
373 
374 /*
375  * Return in pmask the portion of a cpusets's cpus_allowed that
376  * are online.  If none are online, walk up the cpuset hierarchy
377  * until we find one that does have some online cpus.
378  *
379  * One way or another, we guarantee to return some non-empty subset
380  * of cpu_online_mask.
381  *
382  * Call with callback_lock or cpuset_mutex held.
383  */
384 static void guarantee_online_cpus(struct cpuset *cs, struct cpumask *pmask)
385 {
386         while (!cpumask_intersects(cs->effective_cpus, cpu_online_mask)) {
387                 cs = parent_cs(cs);
388                 if (unlikely(!cs)) {
389                         /*
390                          * The top cpuset doesn't have any online cpu as a
391                          * consequence of a race between cpuset_hotplug_work
392                          * and cpu hotplug notifier.  But we know the top
393                          * cpuset's effective_cpus is on its way to be
394                          * identical to cpu_online_mask.
395                          */
396                         cpumask_copy(pmask, cpu_online_mask);
397                         return;
398                 }
399         }
400         cpumask_and(pmask, cs->effective_cpus, cpu_online_mask);
401 }
402 
403 /*
404  * Return in *pmask the portion of a cpusets's mems_allowed that
405  * are online, with memory.  If none are online with memory, walk
406  * up the cpuset hierarchy until we find one that does have some
407  * online mems.  The top cpuset always has some mems online.
408  *
409  * One way or another, we guarantee to return some non-empty subset
410  * of node_states[N_MEMORY].
411  *
412  * Call with callback_lock or cpuset_mutex held.
413  */
414 static void guarantee_online_mems(struct cpuset *cs, nodemask_t *pmask)
415 {
416         while (!nodes_intersects(cs->effective_mems, node_states[N_MEMORY]))
417                 cs = parent_cs(cs);
418         nodes_and(*pmask, cs->effective_mems, node_states[N_MEMORY]);
419 }
420 
421 /*
422  * update task's spread flag if cpuset's page/slab spread flag is set
423  *
424  * Call with callback_lock or cpuset_mutex held.
425  */
426 static void cpuset_update_task_spread_flag(struct cpuset *cs,
427                                         struct task_struct *tsk)
428 {
429         if (is_spread_page(cs))
430                 task_set_spread_page(tsk);
431         else
432                 task_clear_spread_page(tsk);
433 
434         if (is_spread_slab(cs))
435                 task_set_spread_slab(tsk);
436         else
437                 task_clear_spread_slab(tsk);
438 }
439 
440 /*
441  * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
442  *
443  * One cpuset is a subset of another if all its allowed CPUs and
444  * Memory Nodes are a subset of the other, and its exclusive flags
445  * are only set if the other's are set.  Call holding cpuset_mutex.
446  */
447 
448 static int is_cpuset_subset(const struct cpuset *p, const struct cpuset *q)
449 {
450         return  cpumask_subset(p->cpus_allowed, q->cpus_allowed) &&
451                 nodes_subset(p->mems_allowed, q->mems_allowed) &&
452                 is_cpu_exclusive(p) <= is_cpu_exclusive(q) &&
453                 is_mem_exclusive(p) <= is_mem_exclusive(q);
454 }
455 
456 /**
457  * alloc_cpumasks - allocate three cpumasks for cpuset
458  * @cs:  the cpuset that have cpumasks to be allocated.
459  * @tmp: the tmpmasks structure pointer
460  * Return: 0 if successful, -ENOMEM otherwise.
461  *
462  * Only one of the two input arguments should be non-NULL.
463  */
464 static inline int alloc_cpumasks(struct cpuset *cs, struct tmpmasks *tmp)
465 {
466         cpumask_var_t *pmask1, *pmask2, *pmask3;
467 
468         if (cs) {
469                 pmask1 = &cs->cpus_allowed;
470                 pmask2 = &cs->effective_cpus;
471                 pmask3 = &cs->subparts_cpus;
472         } else {
473                 pmask1 = &tmp->new_cpus;
474                 pmask2 = &tmp->addmask;
475                 pmask3 = &tmp->delmask;
476         }
477 
478         if (!zalloc_cpumask_var(pmask1, GFP_KERNEL))
479                 return -ENOMEM;
480 
481         if (!zalloc_cpumask_var(pmask2, GFP_KERNEL))
482                 goto free_one;
483 
484         if (!zalloc_cpumask_var(pmask3, GFP_KERNEL))
485                 goto free_two;
486 
487         return 0;
488 
489 free_two:
490         free_cpumask_var(*pmask2);
491 free_one:
492         free_cpumask_var(*pmask1);
493         return -ENOMEM;
494 }
495 
496 /**
497  * free_cpumasks - free cpumasks in a tmpmasks structure
498  * @cs:  the cpuset that have cpumasks to be free.
499  * @tmp: the tmpmasks structure pointer
500  */
501 static inline void free_cpumasks(struct cpuset *cs, struct tmpmasks *tmp)
502 {
503         if (cs) {
504                 free_cpumask_var(cs->cpus_allowed);
505                 free_cpumask_var(cs->effective_cpus);
506                 free_cpumask_var(cs->subparts_cpus);
507         }
508         if (tmp) {
509                 free_cpumask_var(tmp->new_cpus);
510                 free_cpumask_var(tmp->addmask);
511                 free_cpumask_var(tmp->delmask);
512         }
513 }
514 
515 /**
516  * alloc_trial_cpuset - allocate a trial cpuset
517  * @cs: the cpuset that the trial cpuset duplicates
518  */
519 static struct cpuset *alloc_trial_cpuset(struct cpuset *cs)
520 {
521         struct cpuset *trial;
522 
523         trial = kmemdup(cs, sizeof(*cs), GFP_KERNEL);
524         if (!trial)
525                 return NULL;
526 
527         if (alloc_cpumasks(trial, NULL)) {
528                 kfree(trial);
529                 return NULL;
530         }
531 
532         cpumask_copy(trial->cpus_allowed, cs->cpus_allowed);
533         cpumask_copy(trial->effective_cpus, cs->effective_cpus);
534         return trial;
535 }
536 
537 /**
538  * free_cpuset - free the cpuset
539  * @cs: the cpuset to be freed
540  */
541 static inline void free_cpuset(struct cpuset *cs)
542 {
543         free_cpumasks(cs, NULL);
544         kfree(cs);
545 }
546 
547 /*
548  * validate_change() - Used to validate that any proposed cpuset change
549  *                     follows the structural rules for cpusets.
550  *
551  * If we replaced the flag and mask values of the current cpuset
552  * (cur) with those values in the trial cpuset (trial), would
553  * our various subset and exclusive rules still be valid?  Presumes
554  * cpuset_mutex held.
555  *
556  * 'cur' is the address of an actual, in-use cpuset.  Operations
557  * such as list traversal that depend on the actual address of the
558  * cpuset in the list must use cur below, not trial.
559  *
560  * 'trial' is the address of bulk structure copy of cur, with
561  * perhaps one or more of the fields cpus_allowed, mems_allowed,
562  * or flags changed to new, trial values.
563  *
564  * Return 0 if valid, -errno if not.
565  */
566 
567 static int validate_change(struct cpuset *cur, struct cpuset *trial)
568 {
569         struct cgroup_subsys_state *css;
570         struct cpuset *c, *par;
571         int ret;
572 
573         rcu_read_lock();
574 
575         /* Each of our child cpusets must be a subset of us */
576         ret = -EBUSY;
577         cpuset_for_each_child(c, css, cur)
578                 if (!is_cpuset_subset(c, trial))
579                         goto out;
580 
581         /* Remaining checks don't apply to root cpuset */
582         ret = 0;
583         if (cur == &top_cpuset)
584                 goto out;
585 
586         par = parent_cs(cur);
587 
588         /* On legacy hierarchy, we must be a subset of our parent cpuset. */
589         ret = -EACCES;
590         if (!is_in_v2_mode() && !is_cpuset_subset(trial, par))
591                 goto out;
592 
593         /*
594          * If either I or some sibling (!= me) is exclusive, we can't
595          * overlap
596          */
597         ret = -EINVAL;
598         cpuset_for_each_child(c, css, par) {
599                 if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
600                     c != cur &&
601                     cpumask_intersects(trial->cpus_allowed, c->cpus_allowed))
602                         goto out;
603                 if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) &&
604                     c != cur &&
605                     nodes_intersects(trial->mems_allowed, c->mems_allowed))
606                         goto out;
607         }
608 
609         /*
610          * Cpusets with tasks - existing or newly being attached - can't
611          * be changed to have empty cpus_allowed or mems_allowed.
612          */
613         ret = -ENOSPC;
614         if ((cgroup_is_populated(cur->css.cgroup) || cur->attach_in_progress)) {
615                 if (!cpumask_empty(cur->cpus_allowed) &&
616                     cpumask_empty(trial->cpus_allowed))
617                         goto out;
618                 if (!nodes_empty(cur->mems_allowed) &&
619                     nodes_empty(trial->mems_allowed))
620                         goto out;
621         }
622 
623         /*
624          * We can't shrink if we won't have enough room for SCHED_DEADLINE
625          * tasks.
626          */
627         ret = -EBUSY;
628         if (is_cpu_exclusive(cur) &&
629             !cpuset_cpumask_can_shrink(cur->cpus_allowed,
630                                        trial->cpus_allowed))
631                 goto out;
632 
633         ret = 0;
634 out:
635         rcu_read_unlock();
636         return ret;
637 }
638 
639 #ifdef CONFIG_SMP
640 /*
641  * Helper routine for generate_sched_domains().
642  * Do cpusets a, b have overlapping effective cpus_allowed masks?
643  */
644 static int cpusets_overlap(struct cpuset *a, struct cpuset *b)
645 {
646         return cpumask_intersects(a->effective_cpus, b->effective_cpus);
647 }
648 
649 static void
650 update_domain_attr(struct sched_domain_attr *dattr, struct cpuset *c)
651 {
652         if (dattr->relax_domain_level < c->relax_domain_level)
653                 dattr->relax_domain_level = c->relax_domain_level;
654         return;
655 }
656 
657 static void update_domain_attr_tree(struct sched_domain_attr *dattr,
658                                     struct cpuset *root_cs)
659 {
660         struct cpuset *cp;
661         struct cgroup_subsys_state *pos_css;
662 
663         rcu_read_lock();
664         cpuset_for_each_descendant_pre(cp, pos_css, root_cs) {
665                 /* skip the whole subtree if @cp doesn't have any CPU */
666                 if (cpumask_empty(cp->cpus_allowed)) {
667                         pos_css = css_rightmost_descendant(pos_css);
668                         continue;
669                 }
670 
671                 if (is_sched_load_balance(cp))
672                         update_domain_attr(dattr, cp);
673         }
674         rcu_read_unlock();
675 }
676 
677 /* Must be called with cpuset_mutex held.  */
678 static inline int nr_cpusets(void)
679 {
680         /* jump label reference count + the top-level cpuset */
681         return static_key_count(&cpusets_enabled_key.key) + 1;
682 }
683 
684 /*
685  * generate_sched_domains()
686  *
687  * This function builds a partial partition of the systems CPUs
688  * A 'partial partition' is a set of non-overlapping subsets whose
689  * union is a subset of that set.
690  * The output of this function needs to be passed to kernel/sched/core.c
691  * partition_sched_domains() routine, which will rebuild the scheduler's
692  * load balancing domains (sched domains) as specified by that partial
693  * partition.
694  *
695  * See "What is sched_load_balance" in Documentation/admin-guide/cgroup-v1/cpusets.rst
696  * for a background explanation of this.
697  *
698  * Does not return errors, on the theory that the callers of this
699  * routine would rather not worry about failures to rebuild sched
700  * domains when operating in the severe memory shortage situations
701  * that could cause allocation failures below.
702  *
703  * Must be called with cpuset_mutex held.
704  *
705  * The three key local variables below are:
706  *    cp - cpuset pointer, used (together with pos_css) to perform a
707  *         top-down scan of all cpusets. For our purposes, rebuilding
708  *         the schedulers sched domains, we can ignore !is_sched_load_
709  *         balance cpusets.
710  *  csa  - (for CpuSet Array) Array of pointers to all the cpusets
711  *         that need to be load balanced, for convenient iterative
712  *         access by the subsequent code that finds the best partition,
713  *         i.e the set of domains (subsets) of CPUs such that the
714  *         cpus_allowed of every cpuset marked is_sched_load_balance
715  *         is a subset of one of these domains, while there are as
716  *         many such domains as possible, each as small as possible.
717  * doms  - Conversion of 'csa' to an array of cpumasks, for passing to
718  *         the kernel/sched/core.c routine partition_sched_domains() in a
719  *         convenient format, that can be easily compared to the prior
720  *         value to determine what partition elements (sched domains)
721  *         were changed (added or removed.)
722  *
723  * Finding the best partition (set of domains):
724  *      The triple nested loops below over i, j, k scan over the
725  *      load balanced cpusets (using the array of cpuset pointers in
726  *      csa[]) looking for pairs of cpusets that have overlapping
727  *      cpus_allowed, but which don't have the same 'pn' partition
728  *      number and gives them in the same partition number.  It keeps
729  *      looping on the 'restart' label until it can no longer find
730  *      any such pairs.
731  *
732  *      The union of the cpus_allowed masks from the set of
733  *      all cpusets having the same 'pn' value then form the one
734  *      element of the partition (one sched domain) to be passed to
735  *      partition_sched_domains().
736  */
737 static int generate_sched_domains(cpumask_var_t **domains,
738                         struct sched_domain_attr **attributes)
739 {
740         struct cpuset *cp;      /* top-down scan of cpusets */
741         struct cpuset **csa;    /* array of all cpuset ptrs */
742         int csn;                /* how many cpuset ptrs in csa so far */
743         int i, j, k;            /* indices for partition finding loops */
744         cpumask_var_t *doms;    /* resulting partition; i.e. sched domains */
745         struct sched_domain_attr *dattr;  /* attributes for custom domains */
746         int ndoms = 0;          /* number of sched domains in result */
747         int nslot;              /* next empty doms[] struct cpumask slot */
748         struct cgroup_subsys_state *pos_css;
749         bool root_load_balance = is_sched_load_balance(&top_cpuset);
750 
751         doms = NULL;
752         dattr = NULL;
753         csa = NULL;
754 
755         /* Special case for the 99% of systems with one, full, sched domain */
756         if (root_load_balance && !top_cpuset.nr_subparts_cpus) {
757                 ndoms = 1;
758                 doms = alloc_sched_domains(ndoms);
759                 if (!doms)
760                         goto done;
761 
762                 dattr = kmalloc(sizeof(struct sched_domain_attr), GFP_KERNEL);
763                 if (dattr) {
764                         *dattr = SD_ATTR_INIT;
765                         update_domain_attr_tree(dattr, &top_cpuset);
766                 }
767                 cpumask_and(doms[0], top_cpuset.effective_cpus,
768                             housekeeping_cpumask(HK_FLAG_DOMAIN));
769 
770                 goto done;
771         }
772 
773         csa = kmalloc_array(nr_cpusets(), sizeof(cp), GFP_KERNEL);
774         if (!csa)
775                 goto done;
776         csn = 0;
777 
778         rcu_read_lock();
779         if (root_load_balance)
780                 csa[csn++] = &top_cpuset;
781         cpuset_for_each_descendant_pre(cp, pos_css, &top_cpuset) {
782                 if (cp == &top_cpuset)
783                         continue;
784                 /*
785                  * Continue traversing beyond @cp iff @cp has some CPUs and
786                  * isn't load balancing.  The former is obvious.  The
787                  * latter: All child cpusets contain a subset of the
788                  * parent's cpus, so just skip them, and then we call
789                  * update_domain_attr_tree() to calc relax_domain_level of
790                  * the corresponding sched domain.
791                  *
792                  * If root is load-balancing, we can skip @cp if it
793                  * is a subset of the root's effective_cpus.
794                  */
795                 if (!cpumask_empty(cp->cpus_allowed) &&
796                     !(is_sched_load_balance(cp) &&
797                       cpumask_intersects(cp->cpus_allowed,
798                                          housekeeping_cpumask(HK_FLAG_DOMAIN))))
799                         continue;
800 
801                 if (root_load_balance &&
802                     cpumask_subset(cp->cpus_allowed, top_cpuset.effective_cpus))
803                         continue;
804 
805                 if (is_sched_load_balance(cp) &&
806                     !cpumask_empty(cp->effective_cpus))
807                         csa[csn++] = cp;
808 
809                 /* skip @cp's subtree if not a partition root */
810                 if (!is_partition_root(cp))
811                         pos_css = css_rightmost_descendant(pos_css);
812         }
813         rcu_read_unlock();
814 
815         for (i = 0; i < csn; i++)
816                 csa[i]->pn = i;
817         ndoms = csn;
818 
819 restart:
820         /* Find the best partition (set of sched domains) */
821         for (i = 0; i < csn; i++) {
822                 struct cpuset *a = csa[i];
823                 int apn = a->pn;
824 
825                 for (j = 0; j < csn; j++) {
826                         struct cpuset *b = csa[j];
827                         int bpn = b->pn;
828 
829                         if (apn != bpn && cpusets_overlap(a, b)) {
830                                 for (k = 0; k < csn; k++) {
831                                         struct cpuset *c = csa[k];
832 
833                                         if (c->pn == bpn)
834                                                 c->pn = apn;
835                                 }
836                                 ndoms--;        /* one less element */
837                                 goto restart;
838                         }
839                 }
840         }
841 
842         /*
843          * Now we know how many domains to create.
844          * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
845          */
846         doms = alloc_sched_domains(ndoms);
847         if (!doms)
848                 goto done;
849 
850         /*
851          * The rest of the code, including the scheduler, can deal with
852          * dattr==NULL case. No need to abort if alloc fails.
853          */
854         dattr = kmalloc_array(ndoms, sizeof(struct sched_domain_attr),
855                               GFP_KERNEL);
856 
857         for (nslot = 0, i = 0; i < csn; i++) {
858                 struct cpuset *a = csa[i];
859                 struct cpumask *dp;
860                 int apn = a->pn;
861 
862                 if (apn < 0) {
863                         /* Skip completed partitions */
864                         continue;
865                 }
866 
867                 dp = doms[nslot];
868 
869                 if (nslot == ndoms) {
870                         static int warnings = 10;
871                         if (warnings) {
872                                 pr_warn("rebuild_sched_domains confused: nslot %d, ndoms %d, csn %d, i %d, apn %d\n",
873                                         nslot, ndoms, csn, i, apn);
874                                 warnings--;
875                         }
876                         continue;
877                 }
878 
879                 cpumask_clear(dp);
880                 if (dattr)
881                         *(dattr + nslot) = SD_ATTR_INIT;
882                 for (j = i; j < csn; j++) {
883                         struct cpuset *b = csa[j];
884 
885                         if (apn == b->pn) {
886                                 cpumask_or(dp, dp, b->effective_cpus);
887                                 cpumask_and(dp, dp, housekeeping_cpumask(HK_FLAG_DOMAIN));
888                                 if (dattr)
889                                         update_domain_attr_tree(dattr + nslot, b);
890 
891                                 /* Done with this partition */
892                                 b->pn = -1;
893                         }
894                 }
895                 nslot++;
896         }
897         BUG_ON(nslot != ndoms);
898 
899 done:
900         kfree(csa);
901 
902         /*
903          * Fallback to the default domain if kmalloc() failed.
904          * See comments in partition_sched_domains().
905          */
906         if (doms == NULL)
907                 ndoms = 1;
908 
909         *domains    = doms;
910         *attributes = dattr;
911         return ndoms;
912 }
913 
914 static void update_tasks_root_domain(struct cpuset *cs)
915 {
916         struct css_task_iter it;
917         struct task_struct *task;
918 
919         css_task_iter_start(&cs->css, 0, &it);
920 
921         while ((task = css_task_iter_next(&it)))
922                 dl_add_task_root_domain(task);
923 
924         css_task_iter_end(&it);
925 }
926 
927 static void rebuild_root_domains(void)
928 {
929         struct cpuset *cs = NULL;
930         struct cgroup_subsys_state *pos_css;
931 
932         percpu_rwsem_assert_held(&cpuset_rwsem);
933         lockdep_assert_cpus_held();
934         lockdep_assert_held(&sched_domains_mutex);
935 
936         rcu_read_lock();
937 
938         /*
939          * Clear default root domain DL accounting, it will be computed again
940          * if a task belongs to it.
941          */
942         dl_clear_root_domain(&def_root_domain);
943 
944         cpuset_for_each_descendant_pre(cs, pos_css, &top_cpuset) {
945 
946                 if (cpumask_empty(cs->effective_cpus)) {
947                         pos_css = css_rightmost_descendant(pos_css);
948                         continue;
949                 }
950 
951                 css_get(&cs->css);
952 
953                 rcu_read_unlock();
954 
955                 update_tasks_root_domain(cs);
956 
957                 rcu_read_lock();
958                 css_put(&cs->css);
959         }
960         rcu_read_unlock();
961 }
962 
963 static void
964 partition_and_rebuild_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
965                                     struct sched_domain_attr *dattr_new)
966 {
967         mutex_lock(&sched_domains_mutex);
968         partition_sched_domains_locked(ndoms_new, doms_new, dattr_new);
969         rebuild_root_domains();
970         mutex_unlock(&sched_domains_mutex);
971 }
972 
973 /*
974  * Rebuild scheduler domains.
975  *
976  * If the flag 'sched_load_balance' of any cpuset with non-empty
977  * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
978  * which has that flag enabled, or if any cpuset with a non-empty
979  * 'cpus' is removed, then call this routine to rebuild the
980  * scheduler's dynamic sched domains.
981  *
982  * Call with cpuset_mutex held.  Takes get_online_cpus().
983  */
984 static void rebuild_sched_domains_locked(void)
985 {
986         struct cgroup_subsys_state *pos_css;
987         struct sched_domain_attr *attr;
988         cpumask_var_t *doms;
989         struct cpuset *cs;
990         int ndoms;
991 
992         lockdep_assert_cpus_held();
993         percpu_rwsem_assert_held(&cpuset_rwsem);
994 
995         /*
996          * If we have raced with CPU hotplug, return early to avoid
997          * passing doms with offlined cpu to partition_sched_domains().
998          * Anyways, cpuset_hotplug_workfn() will rebuild sched domains.
999          *
1000          * With no CPUs in any subpartitions, top_cpuset's effective CPUs
1001          * should be the same as the active CPUs, so checking only top_cpuset
1002          * is enough to detect racing CPU offlines.
1003          */
1004         if (!top_cpuset.nr_subparts_cpus &&
1005             !cpumask_equal(top_cpuset.effective_cpus, cpu_active_mask))
1006                 return;
1007 
1008         /*
1009          * With subpartition CPUs, however, the effective CPUs of a partition
1010          * root should be only a subset of the active CPUs.  Since a CPU in any
1011          * partition root could be offlined, all must be checked.
1012          */
1013         if (top_cpuset.nr_subparts_cpus) {
1014                 rcu_read_lock();
1015                 cpuset_for_each_descendant_pre(cs, pos_css, &top_cpuset) {
1016                         if (!is_partition_root(cs)) {
1017                                 pos_css = css_rightmost_descendant(pos_css);
1018                                 continue;
1019                         }
1020                         if (!cpumask_subset(cs->effective_cpus,
1021                                             cpu_active_mask)) {
1022                                 rcu_read_unlock();
1023                                 return;
1024                         }
1025                 }
1026                 rcu_read_unlock();
1027         }
1028 
1029         /* Generate domain masks and attrs */
1030         ndoms = generate_sched_domains(&doms, &attr);
1031 
1032         /* Have scheduler rebuild the domains */
1033         partition_and_rebuild_sched_domains(ndoms, doms, attr);
1034 }
1035 #else /* !CONFIG_SMP */
1036 static void rebuild_sched_domains_locked(void)
1037 {
1038 }
1039 #endif /* CONFIG_SMP */
1040 
1041 void rebuild_sched_domains(void)
1042 {
1043         get_online_cpus();
1044         percpu_down_write(&cpuset_rwsem);
1045         rebuild_sched_domains_locked();
1046         percpu_up_write(&cpuset_rwsem);
1047         put_online_cpus();
1048 }
1049 
1050 /**
1051  * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
1052  * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
1053  *
1054  * Iterate through each task of @cs updating its cpus_allowed to the
1055  * effective cpuset's.  As this function is called with cpuset_mutex held,
1056  * cpuset membership stays stable.
1057  */
1058 static void update_tasks_cpumask(struct cpuset *cs)
1059 {
1060         struct css_task_iter it;
1061         struct task_struct *task;
1062 
1063         css_task_iter_start(&cs->css, 0, &it);
1064         while ((task = css_task_iter_next(&it)))
1065                 set_cpus_allowed_ptr(task, cs->effective_cpus);
1066         css_task_iter_end(&it);
1067 }
1068 
1069 /**
1070  * compute_effective_cpumask - Compute the effective cpumask of the cpuset
1071  * @new_cpus: the temp variable for the new effective_cpus mask
1072  * @cs: the cpuset the need to recompute the new effective_cpus mask
1073  * @parent: the parent cpuset
1074  *
1075  * If the parent has subpartition CPUs, include them in the list of
1076  * allowable CPUs in computing the new effective_cpus mask. Since offlined
1077  * CPUs are not removed from subparts_cpus, we have to use cpu_active_mask
1078  * to mask those out.
1079  */
1080 static void compute_effective_cpumask(struct cpumask *new_cpus,
1081                                       struct cpuset *cs, struct cpuset *parent)
1082 {
1083         if (parent->nr_subparts_cpus) {
1084                 cpumask_or(new_cpus, parent->effective_cpus,
1085                            parent->subparts_cpus);
1086                 cpumask_and(new_cpus, new_cpus, cs->cpus_allowed);
1087                 cpumask_and(new_cpus, new_cpus, cpu_active_mask);
1088         } else {
1089                 cpumask_and(new_cpus, cs->cpus_allowed, parent->effective_cpus);
1090         }
1091 }
1092 
1093 /*
1094  * Commands for update_parent_subparts_cpumask
1095  */
1096 enum subparts_cmd {
1097         partcmd_enable,         /* Enable partition root         */
1098         partcmd_disable,        /* Disable partition root        */
1099         partcmd_update,         /* Update parent's subparts_cpus */
1100 };
1101 
1102 /**
1103  * update_parent_subparts_cpumask - update subparts_cpus mask of parent cpuset
1104  * @cpuset:  The cpuset that requests change in partition root state
1105  * @cmd:     Partition root state change command
1106  * @newmask: Optional new cpumask for partcmd_update
1107  * @tmp:     Temporary addmask and delmask
1108  * Return:   0, 1 or an error code
1109  *
1110  * For partcmd_enable, the cpuset is being transformed from a non-partition
1111  * root to a partition root. The cpus_allowed mask of the given cpuset will
1112  * be put into parent's subparts_cpus and taken away from parent's
1113  * effective_cpus. The function will return 0 if all the CPUs listed in
1114  * cpus_allowed can be granted or an error code will be returned.
1115  *
1116  * For partcmd_disable, the cpuset is being transofrmed from a partition
1117  * root back to a non-partition root. any CPUs in cpus_allowed that are in
1118  * parent's subparts_cpus will be taken away from that cpumask and put back
1119  * into parent's effective_cpus. 0 should always be returned.
1120  *
1121  * For partcmd_update, if the optional newmask is specified, the cpu
1122  * list is to be changed from cpus_allowed to newmask. Otherwise,
1123  * cpus_allowed is assumed to remain the same. The cpuset should either
1124  * be a partition root or an invalid partition root. The partition root
1125  * state may change if newmask is NULL and none of the requested CPUs can
1126  * be granted by the parent. The function will return 1 if changes to
1127  * parent's subparts_cpus and effective_cpus happen or 0 otherwise.
1128  * Error code should only be returned when newmask is non-NULL.
1129  *
1130  * The partcmd_enable and partcmd_disable commands are used by
1131  * update_prstate(). The partcmd_update command is used by
1132  * update_cpumasks_hier() with newmask NULL and update_cpumask() with
1133  * newmask set.
1134  *
1135  * The checking is more strict when enabling partition root than the
1136  * other two commands.
1137  *
1138  * Because of the implicit cpu exclusive nature of a partition root,
1139  * cpumask changes that violates the cpu exclusivity rule will not be
1140  * permitted when checked by validate_change(). The validate_change()
1141  * function will also prevent any changes to the cpu list if it is not
1142  * a superset of children's cpu lists.
1143  */
1144 static int update_parent_subparts_cpumask(struct cpuset *cpuset, int cmd,
1145                                           struct cpumask *newmask,
1146                                           struct tmpmasks *tmp)
1147 {
1148         struct cpuset *parent = parent_cs(cpuset);
1149         int adding;     /* Moving cpus from effective_cpus to subparts_cpus */
1150         int deleting;   /* Moving cpus from subparts_cpus to effective_cpus */
1151         bool part_error = false;        /* Partition error? */
1152 
1153         percpu_rwsem_assert_held(&cpuset_rwsem);
1154 
1155         /*
1156          * The parent must be a partition root.
1157          * The new cpumask, if present, or the current cpus_allowed must
1158          * not be empty.
1159          */
1160         if (!is_partition_root(parent) ||
1161            (newmask && cpumask_empty(newmask)) ||
1162            (!newmask && cpumask_empty(cpuset->cpus_allowed)))
1163                 return -EINVAL;
1164 
1165         /*
1166          * Enabling/disabling partition root is not allowed if there are
1167          * online children.
1168          */
1169         if ((cmd != partcmd_update) && css_has_online_children(&cpuset->css))
1170                 return -EBUSY;
1171 
1172         /*
1173          * Enabling partition root is not allowed if not all the CPUs
1174          * can be granted from parent's effective_cpus or at least one
1175          * CPU will be left after that.
1176          */
1177         if ((cmd == partcmd_enable) &&
1178            (!cpumask_subset(cpuset->cpus_allowed, parent->effective_cpus) ||
1179              cpumask_equal(cpuset->cpus_allowed, parent->effective_cpus)))
1180                 return -EINVAL;
1181 
1182         /*
1183          * A cpumask update cannot make parent's effective_cpus become empty.
1184          */
1185         adding = deleting = false;
1186         if (cmd == partcmd_enable) {
1187                 cpumask_copy(tmp->addmask, cpuset->cpus_allowed);
1188                 adding = true;
1189         } else if (cmd == partcmd_disable) {
1190                 deleting = cpumask_and(tmp->delmask, cpuset->cpus_allowed,
1191                                        parent->subparts_cpus);
1192         } else if (newmask) {
1193                 /*
1194                  * partcmd_update with newmask:
1195                  *
1196                  * delmask = cpus_allowed & ~newmask & parent->subparts_cpus
1197                  * addmask = newmask & parent->effective_cpus
1198                  *                   & ~parent->subparts_cpus
1199                  */
1200                 cpumask_andnot(tmp->delmask, cpuset->cpus_allowed, newmask);
1201                 deleting = cpumask_and(tmp->delmask, tmp->delmask,
1202                                        parent->subparts_cpus);
1203 
1204                 cpumask_and(tmp->addmask, newmask, parent->effective_cpus);
1205                 adding = cpumask_andnot(tmp->addmask, tmp->addmask,
1206                                         parent->subparts_cpus);
1207                 /*
1208                  * Return error if the new effective_cpus could become empty.
1209                  */
1210                 if (adding &&
1211                     cpumask_equal(parent->effective_cpus, tmp->addmask)) {
1212                         if (!deleting)
1213                                 return -EINVAL;
1214                         /*
1215                          * As some of the CPUs in subparts_cpus might have
1216                          * been offlined, we need to compute the real delmask
1217                          * to confirm that.
1218                          */
1219                         if (!cpumask_and(tmp->addmask, tmp->delmask,
1220                                          cpu_active_mask))
1221                                 return -EINVAL;
1222                         cpumask_copy(tmp->addmask, parent->effective_cpus);
1223                 }
1224         } else {
1225                 /*
1226                  * partcmd_update w/o newmask:
1227                  *
1228                  * addmask = cpus_allowed & parent->effectiveb_cpus
1229                  *
1230                  * Note that parent's subparts_cpus may have been
1231                  * pre-shrunk in case there is a change in the cpu list.
1232                  * So no deletion is needed.
1233                  */
1234                 adding = cpumask_and(tmp->addmask, cpuset->cpus_allowed,
1235                                      parent->effective_cpus);
1236                 part_error = cpumask_equal(tmp->addmask,
1237                                            parent->effective_cpus);
1238         }
1239 
1240         if (cmd == partcmd_update) {
1241                 int prev_prs = cpuset->partition_root_state;
1242 
1243                 /*
1244                  * Check for possible transition between PRS_ENABLED
1245                  * and PRS_ERROR.
1246                  */
1247                 switch (cpuset->partition_root_state) {
1248                 case PRS_ENABLED:
1249                         if (part_error)
1250                                 cpuset->partition_root_state = PRS_ERROR;
1251                         break;
1252                 case PRS_ERROR:
1253                         if (!part_error)
1254                                 cpuset->partition_root_state = PRS_ENABLED;
1255                         break;
1256                 }
1257                 /*
1258                  * Set part_error if previously in invalid state.
1259                  */
1260                 part_error = (prev_prs == PRS_ERROR);
1261         }
1262 
1263         if (!part_error && (cpuset->partition_root_state == PRS_ERROR))
1264                 return 0;       /* Nothing need to be done */
1265 
1266         if (cpuset->partition_root_state == PRS_ERROR) {
1267                 /*
1268                  * Remove all its cpus from parent's subparts_cpus.
1269                  */
1270                 adding = false;
1271                 deleting = cpumask_and(tmp->delmask, cpuset->cpus_allowed,
1272                                        parent->subparts_cpus);
1273         }
1274 
1275         if (!adding && !deleting)
1276                 return 0;
1277 
1278         /*
1279          * Change the parent's subparts_cpus.
1280          * Newly added CPUs will be removed from effective_cpus and
1281          * newly deleted ones will be added back to effective_cpus.
1282          */
1283         spin_lock_irq(&callback_lock);
1284         if (adding) {
1285                 cpumask_or(parent->subparts_cpus,
1286                            parent->subparts_cpus, tmp->addmask);
1287                 cpumask_andnot(parent->effective_cpus,
1288                                parent->effective_cpus, tmp->addmask);
1289         }
1290         if (deleting) {
1291                 cpumask_andnot(parent->subparts_cpus,
1292                                parent->subparts_cpus, tmp->delmask);
1293                 /*
1294                  * Some of the CPUs in subparts_cpus might have been offlined.
1295                  */
1296                 cpumask_and(tmp->delmask, tmp->delmask, cpu_active_mask);
1297                 cpumask_or(parent->effective_cpus,
1298                            parent->effective_cpus, tmp->delmask);
1299         }
1300 
1301         parent->nr_subparts_cpus = cpumask_weight(parent->subparts_cpus);
1302         spin_unlock_irq(&callback_lock);
1303 
1304         return cmd == partcmd_update;
1305 }
1306 
1307 /*
1308  * update_cpumasks_hier - Update effective cpumasks and tasks in the subtree
1309  * @cs:  the cpuset to consider
1310  * @tmp: temp variables for calculating effective_cpus & partition setup
1311  *
1312  * When configured cpumask is changed, the effective cpumasks of this cpuset
1313  * and all its descendants need to be updated.
1314  *
1315  * On legacy hierarchy, effective_cpus will be the same with cpu_allowed.
1316  *
1317  * Called with cpuset_mutex held
1318  */
1319 static void update_cpumasks_hier(struct cpuset *cs, struct tmpmasks *tmp)
1320 {
1321         struct cpuset *cp;
1322         struct cgroup_subsys_state *pos_css;
1323         bool need_rebuild_sched_domains = false;
1324 
1325         rcu_read_lock();
1326         cpuset_for_each_descendant_pre(cp, pos_css, cs) {
1327                 struct cpuset *parent = parent_cs(cp);
1328 
1329                 compute_effective_cpumask(tmp->new_cpus, cp, parent);
1330 
1331                 /*
1332                  * If it becomes empty, inherit the effective mask of the
1333                  * parent, which is guaranteed to have some CPUs.
1334                  */
1335                 if (is_in_v2_mode() && cpumask_empty(tmp->new_cpus)) {
1336                         cpumask_copy(tmp->new_cpus, parent->effective_cpus);
1337                         if (!cp->use_parent_ecpus) {
1338                                 cp->use_parent_ecpus = true;
1339                                 parent->child_ecpus_count++;
1340                         }
1341                 } else if (cp->use_parent_ecpus) {
1342                         cp->use_parent_ecpus = false;
1343                         WARN_ON_ONCE(!parent->child_ecpus_count);
1344                         parent->child_ecpus_count--;
1345                 }
1346 
1347                 /*
1348                  * Skip the whole subtree if the cpumask remains the same
1349                  * and has no partition root state.
1350                  */
1351                 if (!cp->partition_root_state &&
1352                     cpumask_equal(tmp->new_cpus, cp->effective_cpus)) {
1353                         pos_css = css_rightmost_descendant(pos_css);
1354                         continue;
1355                 }
1356 
1357                 /*
1358                  * update_parent_subparts_cpumask() should have been called
1359                  * for cs already in update_cpumask(). We should also call
1360                  * update_tasks_cpumask() again for tasks in the parent
1361                  * cpuset if the parent's subparts_cpus changes.
1362                  */
1363                 if ((cp != cs) && cp->partition_root_state) {
1364                         switch (parent->partition_root_state) {
1365                         case PRS_DISABLED:
1366                                 /*
1367                                  * If parent is not a partition root or an
1368                                  * invalid partition root, clear the state
1369                                  * state and the CS_CPU_EXCLUSIVE flag.
1370                                  */
1371                                 WARN_ON_ONCE(cp->partition_root_state
1372                                              != PRS_ERROR);
1373                                 cp->partition_root_state = 0;
1374 
1375                                 /*
1376                                  * clear_bit() is an atomic operation and
1377                                  * readers aren't interested in the state
1378                                  * of CS_CPU_EXCLUSIVE anyway. So we can
1379                                  * just update the flag without holding
1380                                  * the callback_lock.
1381                                  */
1382                                 clear_bit(CS_CPU_EXCLUSIVE, &cp->flags);
1383                                 break;
1384 
1385                         case PRS_ENABLED:
1386                                 if (update_parent_subparts_cpumask(cp, partcmd_update, NULL, tmp))
1387                                         update_tasks_cpumask(parent);
1388                                 break;
1389 
1390                         case PRS_ERROR:
1391                                 /*
1392                                  * When parent is invalid, it has to be too.
1393                                  */
1394                                 cp->partition_root_state = PRS_ERROR;
1395                                 if (cp->nr_subparts_cpus) {
1396                                         cp->nr_subparts_cpus = 0;
1397                                         cpumask_clear(cp->subparts_cpus);
1398                                 }
1399                                 break;
1400                         }
1401                 }
1402 
1403                 if (!css_tryget_online(&cp->css))
1404                         continue;
1405                 rcu_read_unlock();
1406 
1407                 spin_lock_irq(&callback_lock);
1408 
1409                 cpumask_copy(cp->effective_cpus, tmp->new_cpus);
1410                 if (cp->nr_subparts_cpus &&
1411                    (cp->partition_root_state != PRS_ENABLED)) {
1412                         cp->nr_subparts_cpus = 0;
1413                         cpumask_clear(cp->subparts_cpus);
1414                 } else if (cp->nr_subparts_cpus) {
1415                         /*
1416                          * Make sure that effective_cpus & subparts_cpus
1417                          * are mutually exclusive.
1418                          *
1419                          * In the unlikely event that effective_cpus
1420                          * becomes empty. we clear cp->nr_subparts_cpus and
1421                          * let its child partition roots to compete for
1422                          * CPUs again.
1423                          */
1424                         cpumask_andnot(cp->effective_cpus, cp->effective_cpus,
1425                                        cp->subparts_cpus);
1426                         if (cpumask_empty(cp->effective_cpus)) {
1427                                 cpumask_copy(cp->effective_cpus, tmp->new_cpus);
1428                                 cpumask_clear(cp->subparts_cpus);
1429                                 cp->nr_subparts_cpus = 0;
1430                         } else if (!cpumask_subset(cp->subparts_cpus,
1431                                                    tmp->new_cpus)) {
1432                                 cpumask_andnot(cp->subparts_cpus,
1433                                         cp->subparts_cpus, tmp->new_cpus);
1434                                 cp->nr_subparts_cpus
1435                                         = cpumask_weight(cp->subparts_cpus);
1436                         }
1437                 }
1438                 spin_unlock_irq(&callback_lock);
1439 
1440                 WARN_ON(!is_in_v2_mode() &&
1441                         !cpumask_equal(cp->cpus_allowed, cp->effective_cpus));
1442 
1443                 update_tasks_cpumask(cp);
1444 
1445                 /*
1446                  * On legacy hierarchy, if the effective cpumask of any non-
1447                  * empty cpuset is changed, we need to rebuild sched domains.
1448                  * On default hierarchy, the cpuset needs to be a partition
1449                  * root as well.
1450                  */
1451                 if (!cpumask_empty(cp->cpus_allowed) &&
1452                     is_sched_load_balance(cp) &&
1453                    (!cgroup_subsys_on_dfl(cpuset_cgrp_subsys) ||
1454                     is_partition_root(cp)))
1455                         need_rebuild_sched_domains = true;
1456 
1457                 rcu_read_lock();
1458                 css_put(&cp->css);
1459         }
1460         rcu_read_unlock();
1461 
1462         if (need_rebuild_sched_domains)
1463                 rebuild_sched_domains_locked();
1464 }
1465 
1466 /**
1467  * update_sibling_cpumasks - Update siblings cpumasks
1468  * @parent:  Parent cpuset
1469  * @cs:      Current cpuset
1470  * @tmp:     Temp variables
1471  */
1472 static void update_sibling_cpumasks(struct cpuset *parent, struct cpuset *cs,
1473                                     struct tmpmasks *tmp)
1474 {
1475         struct cpuset *sibling;
1476         struct cgroup_subsys_state *pos_css;
1477 
1478         /*
1479          * Check all its siblings and call update_cpumasks_hier()
1480          * if their use_parent_ecpus flag is set in order for them
1481          * to use the right effective_cpus value.
1482          */
1483         rcu_read_lock();
1484         cpuset_for_each_child(sibling, pos_css, parent) {
1485                 if (sibling == cs)
1486                         continue;
1487                 if (!sibling->use_parent_ecpus)
1488                         continue;
1489 
1490                 update_cpumasks_hier(sibling, tmp);
1491         }
1492         rcu_read_unlock();
1493 }
1494 
1495 /**
1496  * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
1497  * @cs: the cpuset to consider
1498  * @trialcs: trial cpuset
1499  * @buf: buffer of cpu numbers written to this cpuset
1500  */
1501 static int update_cpumask(struct cpuset *cs, struct cpuset *trialcs,
1502                           const char *buf)
1503 {
1504         int retval;
1505         struct tmpmasks tmp;
1506 
1507         /* top_cpuset.cpus_allowed tracks cpu_online_mask; it's read-only */
1508         if (cs == &top_cpuset)
1509                 return -EACCES;
1510 
1511         /*
1512          * An empty cpus_allowed is ok only if the cpuset has no tasks.
1513          * Since cpulist_parse() fails on an empty mask, we special case
1514          * that parsing.  The validate_change() call ensures that cpusets
1515          * with tasks have cpus.
1516          */
1517         if (!*buf) {
1518                 cpumask_clear(trialcs->cpus_allowed);
1519         } else {
1520                 retval = cpulist_parse(buf, trialcs->cpus_allowed);
1521                 if (retval < 0)
1522                         return retval;
1523 
1524                 if (!cpumask_subset(trialcs->cpus_allowed,
1525                                     top_cpuset.cpus_allowed))
1526                         return -EINVAL;
1527         }
1528 
1529         /* Nothing to do if the cpus didn't change */
1530         if (cpumask_equal(cs->cpus_allowed, trialcs->cpus_allowed))
1531                 return 0;
1532 
1533         retval = validate_change(cs, trialcs);
1534         if (retval < 0)
1535                 return retval;
1536 
1537 #ifdef CONFIG_CPUMASK_OFFSTACK
1538         /*
1539          * Use the cpumasks in trialcs for tmpmasks when they are pointers
1540          * to allocated cpumasks.
1541          */
1542         tmp.addmask  = trialcs->subparts_cpus;
1543         tmp.delmask  = trialcs->effective_cpus;
1544         tmp.new_cpus = trialcs->cpus_allowed;
1545 #endif
1546 
1547         if (cs->partition_root_state) {
1548                 /* Cpumask of a partition root cannot be empty */
1549                 if (cpumask_empty(trialcs->cpus_allowed))
1550                         return -EINVAL;
1551                 if (update_parent_subparts_cpumask(cs, partcmd_update,
1552                                         trialcs->cpus_allowed, &tmp) < 0)
1553                         return -EINVAL;
1554         }
1555 
1556         spin_lock_irq(&callback_lock);
1557         cpumask_copy(cs->cpus_allowed, trialcs->cpus_allowed);
1558 
1559         /*
1560          * Make sure that subparts_cpus is a subset of cpus_allowed.
1561          */
1562         if (cs->nr_subparts_cpus) {
1563                 cpumask_andnot(cs->subparts_cpus, cs->subparts_cpus,
1564                                cs->cpus_allowed);
1565                 cs->nr_subparts_cpus = cpumask_weight(cs->subparts_cpus);
1566         }
1567         spin_unlock_irq(&callback_lock);
1568 
1569         update_cpumasks_hier(cs, &tmp);
1570 
1571         if (cs->partition_root_state) {
1572                 struct cpuset *parent = parent_cs(cs);
1573 
1574                 /*
1575                  * For partition root, update the cpumasks of sibling
1576                  * cpusets if they use parent's effective_cpus.
1577                  */
1578                 if (parent->child_ecpus_count)
1579                         update_sibling_cpumasks(parent, cs, &tmp);
1580         }
1581         return 0;
1582 }
1583 
1584 /*
1585  * Migrate memory region from one set of nodes to another.  This is
1586  * performed asynchronously as it can be called from process migration path
1587  * holding locks involved in process management.  All mm migrations are
1588  * performed in the queued order and can be waited for by flushing
1589  * cpuset_migrate_mm_wq.
1590  */
1591 
1592 struct cpuset_migrate_mm_work {
1593         struct work_struct      work;
1594         struct mm_struct        *mm;
1595         nodemask_t              from;
1596         nodemask_t              to;
1597 };
1598 
1599 static void cpuset_migrate_mm_workfn(struct work_struct *work)
1600 {
1601         struct cpuset_migrate_mm_work *mwork =
1602                 container_of(work, struct cpuset_migrate_mm_work, work);
1603 
1604         /* on a wq worker, no need to worry about %current's mems_allowed */
1605         do_migrate_pages(mwork->mm, &mwork->from, &mwork->to, MPOL_MF_MOVE_ALL);
1606         mmput(mwork->mm);
1607         kfree(mwork);
1608 }
1609 
1610 static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
1611                                                         const nodemask_t *to)
1612 {
1613         struct cpuset_migrate_mm_work *mwork;
1614 
1615         mwork = kzalloc(sizeof(*mwork), GFP_KERNEL);
1616         if (mwork) {
1617                 mwork->mm = mm;
1618                 mwork->from = *from;
1619                 mwork->to = *to;
1620                 INIT_WORK(&mwork->work, cpuset_migrate_mm_workfn);
1621                 queue_work(cpuset_migrate_mm_wq, &mwork->work);
1622         } else {
1623                 mmput(mm);
1624         }
1625 }
1626 
1627 static void cpuset_post_attach(void)
1628 {
1629         flush_workqueue(cpuset_migrate_mm_wq);
1630 }
1631 
1632 /*
1633  * cpuset_change_task_nodemask - change task's mems_allowed and mempolicy
1634  * @tsk: the task to change
1635  * @newmems: new nodes that the task will be set
1636  *
1637  * We use the mems_allowed_seq seqlock to safely update both tsk->mems_allowed
1638  * and rebind an eventual tasks' mempolicy. If the task is allocating in
1639  * parallel, it might temporarily see an empty intersection, which results in
1640  * a seqlock check and retry before OOM or allocation failure.
1641  */
1642 static void cpuset_change_task_nodemask(struct task_struct *tsk,
1643                                         nodemask_t *newmems)
1644 {
1645         task_lock(tsk);
1646 
1647         local_irq_disable();
1648         write_seqcount_begin(&tsk->mems_allowed_seq);
1649 
1650         nodes_or(tsk->mems_allowed, tsk->mems_allowed, *newmems);
1651         mpol_rebind_task(tsk, newmems);
1652         tsk->mems_allowed = *newmems;
1653 
1654         write_seqcount_end(&tsk->mems_allowed_seq);
1655         local_irq_enable();
1656 
1657         task_unlock(tsk);
1658 }
1659 
1660 static void *cpuset_being_rebound;
1661 
1662 /**
1663  * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
1664  * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
1665  *
1666  * Iterate through each task of @cs updating its mems_allowed to the
1667  * effective cpuset's.  As this function is called with cpuset_mutex held,
1668  * cpuset membership stays stable.
1669  */
1670 static void update_tasks_nodemask(struct cpuset *cs)
1671 {
1672         static nodemask_t newmems;      /* protected by cpuset_mutex */
1673         struct css_task_iter it;
1674         struct task_struct *task;
1675 
1676         cpuset_being_rebound = cs;              /* causes mpol_dup() rebind */
1677 
1678         guarantee_online_mems(cs, &newmems);
1679 
1680         /*
1681          * The mpol_rebind_mm() call takes mmap_lock, which we couldn't
1682          * take while holding tasklist_lock.  Forks can happen - the
1683          * mpol_dup() cpuset_being_rebound check will catch such forks,
1684          * and rebind their vma mempolicies too.  Because we still hold
1685          * the global cpuset_mutex, we know that no other rebind effort
1686          * will be contending for the global variable cpuset_being_rebound.
1687          * It's ok if we rebind the same mm twice; mpol_rebind_mm()
1688          * is idempotent.  Also migrate pages in each mm to new nodes.
1689          */
1690         css_task_iter_start(&cs->css, 0, &it);
1691         while ((task = css_task_iter_next(&it))) {
1692                 struct mm_struct *mm;
1693                 bool migrate;
1694 
1695                 cpuset_change_task_nodemask(task, &newmems);
1696 
1697                 mm = get_task_mm(task);
1698                 if (!mm)
1699                         continue;
1700 
1701                 migrate = is_memory_migrate(cs);
1702 
1703                 mpol_rebind_mm(mm, &cs->mems_allowed);
1704                 if (migrate)
1705                         cpuset_migrate_mm(mm, &cs->old_mems_allowed, &newmems);
1706                 else
1707                         mmput(mm);
1708         }
1709         css_task_iter_end(&it);
1710 
1711         /*
1712          * All the tasks' nodemasks have been updated, update
1713          * cs->old_mems_allowed.
1714          */
1715         cs->old_mems_allowed = newmems;
1716 
1717         /* We're done rebinding vmas to this cpuset's new mems_allowed. */
1718         cpuset_being_rebound = NULL;
1719 }
1720 
1721 /*
1722  * update_nodemasks_hier - Update effective nodemasks and tasks in the subtree
1723  * @cs: the cpuset to consider
1724  * @new_mems: a temp variable for calculating new effective_mems
1725  *
1726  * When configured nodemask is changed, the effective nodemasks of this cpuset
1727  * and all its descendants need to be updated.
1728  *
1729  * On legacy hierarchy, effective_mems will be the same with mems_allowed.
1730  *
1731  * Called with cpuset_mutex held
1732  */
1733 static void update_nodemasks_hier(struct cpuset *cs, nodemask_t *new_mems)
1734 {
1735         struct cpuset *cp;
1736         struct cgroup_subsys_state *pos_css;
1737 
1738         rcu_read_lock();
1739         cpuset_for_each_descendant_pre(cp, pos_css, cs) {
1740                 struct cpuset *parent = parent_cs(cp);
1741 
1742                 nodes_and(*new_mems, cp->mems_allowed, parent->effective_mems);
1743 
1744                 /*
1745                  * If it becomes empty, inherit the effective mask of the
1746                  * parent, which is guaranteed to have some MEMs.
1747                  */
1748                 if (is_in_v2_mode() && nodes_empty(*new_mems))
1749                         *new_mems = parent->effective_mems;
1750 
1751                 /* Skip the whole subtree if the nodemask remains the same. */
1752                 if (nodes_equal(*new_mems, cp->effective_mems)) {
1753                         pos_css = css_rightmost_descendant(pos_css);
1754                         continue;
1755                 }
1756 
1757                 if (!css_tryget_online(&cp->css))
1758                         continue;
1759                 rcu_read_unlock();
1760 
1761                 spin_lock_irq(&callback_lock);
1762                 cp->effective_mems = *new_mems;
1763                 spin_unlock_irq(&callback_lock);
1764 
1765                 WARN_ON(!is_in_v2_mode() &&
1766                         !nodes_equal(cp->mems_allowed, cp->effective_mems));
1767 
1768                 update_tasks_nodemask(cp);
1769 
1770                 rcu_read_lock();
1771                 css_put(&cp->css);
1772         }
1773         rcu_read_unlock();
1774 }
1775 
1776 /*
1777  * Handle user request to change the 'mems' memory placement
1778  * of a cpuset.  Needs to validate the request, update the
1779  * cpusets mems_allowed, and for each task in the cpuset,
1780  * update mems_allowed and rebind task's mempolicy and any vma
1781  * mempolicies and if the cpuset is marked 'memory_migrate',
1782  * migrate the tasks pages to the new memory.
1783  *
1784  * Call with cpuset_mutex held. May take callback_lock during call.
1785  * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
1786  * lock each such tasks mm->mmap_lock, scan its vma's and rebind
1787  * their mempolicies to the cpusets new mems_allowed.
1788  */
1789 static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
1790                            const char *buf)
1791 {
1792         int retval;
1793 
1794         /*
1795          * top_cpuset.mems_allowed tracks node_stats[N_MEMORY];
1796          * it's read-only
1797          */
1798         if (cs == &top_cpuset) {
1799                 retval = -EACCES;
1800                 goto done;
1801         }
1802 
1803         /*
1804          * An empty mems_allowed is ok iff there are no tasks in the cpuset.
1805          * Since nodelist_parse() fails on an empty mask, we special case
1806          * that parsing.  The validate_change() call ensures that cpusets
1807          * with tasks have memory.
1808          */
1809         if (!*buf) {
1810                 nodes_clear(trialcs->mems_allowed);
1811         } else {
1812                 retval = nodelist_parse(buf, trialcs->mems_allowed);
1813                 if (retval < 0)
1814                         goto done;
1815 
1816                 if (!nodes_subset(trialcs->mems_allowed,
1817                                   top_cpuset.mems_allowed)) {
1818                         retval = -EINVAL;
1819                         goto done;
1820                 }
1821         }
1822 
1823         if (nodes_equal(cs->mems_allowed, trialcs->mems_allowed)) {
1824                 retval = 0;             /* Too easy - nothing to do */
1825                 goto done;
1826         }
1827         retval = validate_change(cs, trialcs);
1828         if (retval < 0)
1829                 goto done;
1830 
1831         spin_lock_irq(&callback_lock);
1832         cs->mems_allowed = trialcs->mems_allowed;
1833         spin_unlock_irq(&callback_lock);
1834 
1835         /* use trialcs->mems_allowed as a temp variable */
1836         update_nodemasks_hier(cs, &trialcs->mems_allowed);
1837 done:
1838         return retval;
1839 }
1840 
1841 bool current_cpuset_is_being_rebound(void)
1842 {
1843         bool ret;
1844 
1845         rcu_read_lock();
1846         ret = task_cs(current) == cpuset_being_rebound;
1847         rcu_read_unlock();
1848 
1849         return ret;
1850 }
1851 
1852 static int update_relax_domain_level(struct cpuset *cs, s64 val)
1853 {
1854 #ifdef CONFIG_SMP
1855         if (val < -1 || val >= sched_domain_level_max)
1856                 return -EINVAL;
1857 #endif
1858 
1859         if (val != cs->relax_domain_level) {
1860                 cs->relax_domain_level = val;
1861                 if (!cpumask_empty(cs->cpus_allowed) &&
1862                     is_sched_load_balance(cs))
1863                         rebuild_sched_domains_locked();
1864         }
1865 
1866         return 0;
1867 }
1868 
1869 /**
1870  * update_tasks_flags - update the spread flags of tasks in the cpuset.
1871  * @cs: the cpuset in which each task's spread flags needs to be changed
1872  *
1873  * Iterate through each task of @cs updating its spread flags.  As this
1874  * function is called with cpuset_mutex held, cpuset membership stays
1875  * stable.
1876  */
1877 static void update_tasks_flags(struct cpuset *cs)
1878 {
1879         struct css_task_iter it;
1880         struct task_struct *task;
1881 
1882         css_task_iter_start(&cs->css, 0, &it);
1883         while ((task = css_task_iter_next(&it)))
1884                 cpuset_update_task_spread_flag(cs, task);
1885         css_task_iter_end(&it);
1886 }
1887 
1888 /*
1889  * update_flag - read a 0 or a 1 in a file and update associated flag
1890  * bit:         the bit to update (see cpuset_flagbits_t)
1891  * cs:          the cpuset to update
1892  * turning_on:  whether the flag is being set or cleared
1893  *
1894  * Call with cpuset_mutex held.
1895  */
1896 
1897 static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs,
1898                        int turning_on)
1899 {
1900         struct cpuset *trialcs;
1901         int balance_flag_changed;
1902         int spread_flag_changed;
1903         int err;
1904 
1905         trialcs = alloc_trial_cpuset(cs);
1906         if (!trialcs)
1907                 return -ENOMEM;
1908 
1909         if (turning_on)
1910                 set_bit(bit, &trialcs->flags);
1911         else
1912                 clear_bit(bit, &trialcs->flags);
1913 
1914         err = validate_change(cs, trialcs);
1915         if (err < 0)
1916                 goto out;
1917 
1918         balance_flag_changed = (is_sched_load_balance(cs) !=
1919                                 is_sched_load_balance(trialcs));
1920 
1921         spread_flag_changed = ((is_spread_slab(cs) != is_spread_slab(trialcs))
1922                         || (is_spread_page(cs) != is_spread_page(trialcs)));
1923 
1924         spin_lock_irq(&callback_lock);
1925         cs->flags = trialcs->flags;
1926         spin_unlock_irq(&callback_lock);
1927 
1928         if (!cpumask_empty(trialcs->cpus_allowed) && balance_flag_changed)
1929                 rebuild_sched_domains_locked();
1930 
1931         if (spread_flag_changed)
1932                 update_tasks_flags(cs);
1933 out:
1934         free_cpuset(trialcs);
1935         return err;
1936 }
1937 
1938 /*
1939  * update_prstate - update partititon_root_state
1940  * cs:  the cpuset to update
1941  * val: 0 - disabled, 1 - enabled
1942  *
1943  * Call with cpuset_mutex held.
1944  */
1945 static int update_prstate(struct cpuset *cs, int val)
1946 {
1947         int err;
1948         struct cpuset *parent = parent_cs(cs);
1949         struct tmpmasks tmp;
1950 
1951         if ((val != 0) && (val != 1))
1952                 return -EINVAL;
1953         if (val == cs->partition_root_state)
1954                 return 0;
1955 
1956         /*
1957          * Cannot force a partial or invalid partition root to a full
1958          * partition root.
1959          */
1960         if (val && cs->partition_root_state)
1961                 return -EINVAL;
1962 
1963         if (alloc_cpumasks(NULL, &tmp))
1964                 return -ENOMEM;
1965 
1966         err = -EINVAL;
1967         if (!cs->partition_root_state) {
1968                 /*
1969                  * Turning on partition root requires setting the
1970                  * CS_CPU_EXCLUSIVE bit implicitly as well and cpus_allowed
1971                  * cannot be NULL.
1972                  */
1973                 if (cpumask_empty(cs->cpus_allowed))
1974                         goto out;
1975 
1976                 err = update_flag(CS_CPU_EXCLUSIVE, cs, 1);
1977                 if (err)
1978                         goto out;
1979 
1980                 err = update_parent_subparts_cpumask(cs, partcmd_enable,
1981                                                      NULL, &tmp);
1982                 if (err) {
1983                         update_flag(CS_CPU_EXCLUSIVE, cs, 0);
1984                         goto out;
1985                 }
1986                 cs->partition_root_state = PRS_ENABLED;
1987         } else {
1988                 /*
1989                  * Turning off partition root will clear the
1990                  * CS_CPU_EXCLUSIVE bit.
1991                  */
1992                 if (cs->partition_root_state == PRS_ERROR) {
1993                         cs->partition_root_state = 0;
1994                         update_flag(CS_CPU_EXCLUSIVE, cs, 0);
1995                         err = 0;
1996                         goto out;
1997                 }
1998 
1999                 err = update_parent_subparts_cpumask(cs, partcmd_disable,
2000                                                      NULL, &tmp);
2001                 if (err)
2002                         goto out;
2003 
2004                 cs->partition_root_state = 0;
2005 
2006                 /* Turning off CS_CPU_EXCLUSIVE will not return error */
2007                 update_flag(CS_CPU_EXCLUSIVE, cs, 0);
2008         }
2009 
2010         /*
2011          * Update cpumask of parent's tasks except when it is the top
2012          * cpuset as some system daemons cannot be mapped to other CPUs.
2013          */
2014         if (parent != &top_cpuset)
2015                 update_tasks_cpumask(parent);
2016 
2017         if (parent->child_ecpus_count)
2018                 update_sibling_cpumasks(parent, cs, &tmp);
2019 
2020         rebuild_sched_domains_locked();
2021 out:
2022         free_cpumasks(NULL, &tmp);
2023         return err;
2024 }
2025 
2026 /*
2027  * Frequency meter - How fast is some event occurring?
2028  *
2029  * These routines manage a digitally filtered, constant time based,
2030  * event frequency meter.  There are four routines:
2031  *   fmeter_init() - initialize a frequency meter.
2032  *   fmeter_markevent() - called each time the event happens.
2033  *   fmeter_getrate() - returns the recent rate of such events.
2034  *   fmeter_update() - internal routine used to update fmeter.
2035  *
2036  * A common data structure is passed to each of these routines,
2037  * which is used to keep track of the state required to manage the
2038  * frequency meter and its digital filter.
2039  *
2040  * The filter works on the number of events marked per unit time.
2041  * The filter is single-pole low-pass recursive (IIR).  The time unit
2042  * is 1 second.  Arithmetic is done using 32-bit integers scaled to
2043  * simulate 3 decimal digits of precision (multiplied by 1000).
2044  *
2045  * With an FM_COEF of 933, and a time base of 1 second, the filter
2046  * has a half-life of 10 seconds, meaning that if the events quit
2047  * happening, then the rate returned from the fmeter_getrate()
2048  * will be cut in half each 10 seconds, until it converges to zero.
2049  *
2050  * It is not worth doing a real infinitely recursive filter.  If more
2051  * than FM_MAXTICKS ticks have elapsed since the last filter event,
2052  * just compute FM_MAXTICKS ticks worth, by which point the level
2053  * will be stable.
2054  *
2055  * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
2056  * arithmetic overflow in the fmeter_update() routine.
2057  *
2058  * Given the simple 32 bit integer arithmetic used, this meter works
2059  * best for reporting rates between one per millisecond (msec) and
2060  * one per 32 (approx) seconds.  At constant rates faster than one
2061  * per msec it maxes out at values just under 1,000,000.  At constant
2062  * rates between one per msec, and one per second it will stabilize
2063  * to a value N*1000, where N is the rate of events per second.
2064  * At constant rates between one per second and one per 32 seconds,
2065  * it will be choppy, moving up on the seconds that have an event,
2066  * and then decaying until the next event.  At rates slower than
2067  * about one in 32 seconds, it decays all the way back to zero between
2068  * each event.
2069  */
2070 
2071 #define FM_COEF 933             /* coefficient for half-life of 10 secs */
2072 #define FM_MAXTICKS ((u32)99)   /* useless computing more ticks than this */
2073 #define FM_MAXCNT 1000000       /* limit cnt to avoid overflow */
2074 #define FM_SCALE 1000           /* faux fixed point scale */
2075 
2076 /* Initialize a frequency meter */
2077 static void fmeter_init(struct fmeter *fmp)
2078 {
2079         fmp->cnt = 0;
2080         fmp->val = 0;
2081         fmp->time = 0;
2082         spin_lock_init(&fmp->lock);
2083 }
2084 
2085 /* Internal meter update - process cnt events and update value */
2086 static void fmeter_update(struct fmeter *fmp)
2087 {
2088         time64_t now;
2089         u32 ticks;
2090 
2091         now = ktime_get_seconds();
2092         ticks = now - fmp->time;
2093 
2094         if (ticks == 0)
2095                 return;
2096 
2097         ticks = min(FM_MAXTICKS, ticks);
2098         while (ticks-- > 0)
2099                 fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
2100         fmp->time = now;
2101 
2102         fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
2103         fmp->cnt = 0;
2104 }
2105 
2106 /* Process any previous ticks, then bump cnt by one (times scale). */
2107 static void fmeter_markevent(struct fmeter *fmp)
2108 {
2109         spin_lock(&fmp->lock);
2110         fmeter_update(fmp);
2111         fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
2112         spin_unlock(&fmp->lock);
2113 }
2114 
2115 /* Process any previous ticks, then return current value. */
2116 static int fmeter_getrate(struct fmeter *fmp)
2117 {
2118         int val;
2119 
2120         spin_lock(&fmp->lock);
2121         fmeter_update(fmp);
2122         val = fmp->val;
2123         spin_unlock(&fmp->lock);
2124         return val;
2125 }
2126 
2127 static struct cpuset *cpuset_attach_old_cs;
2128 
2129 /* Called by cgroups to determine if a cpuset is usable; cpuset_mutex held */
2130 static int cpuset_can_attach(struct cgroup_taskset *tset)
2131 {
2132         struct cgroup_subsys_state *css;
2133         struct cpuset *cs;
2134         struct task_struct *task;
2135         int ret;
2136 
2137         /* used later by cpuset_attach() */
2138         cpuset_attach_old_cs = task_cs(cgroup_taskset_first(tset, &css));
2139         cs = css_cs(css);
2140 
2141         percpu_down_write(&cpuset_rwsem);
2142 
2143         /* allow moving tasks into an empty cpuset if on default hierarchy */
2144         ret = -ENOSPC;
2145         if (!is_in_v2_mode() &&
2146             (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed)))
2147                 goto out_unlock;
2148 
2149         cgroup_taskset_for_each(task, css, tset) {
2150                 ret = task_can_attach(task, cs->cpus_allowed);
2151                 if (ret)
2152                         goto out_unlock;
2153                 ret = security_task_setscheduler(task);
2154                 if (ret)
2155                         goto out_unlock;
2156         }
2157 
2158         /*
2159          * Mark attach is in progress.  This makes validate_change() fail
2160          * changes which zero cpus/mems_allowed.
2161          */
2162         cs->attach_in_progress++;
2163         ret = 0;
2164 out_unlock:
2165         percpu_up_write(&cpuset_rwsem);
2166         return ret;
2167 }
2168 
2169 static void cpuset_cancel_attach(struct cgroup_taskset *tset)
2170 {
2171         struct cgroup_subsys_state *css;
2172 
2173         cgroup_taskset_first(tset, &css);
2174 
2175         percpu_down_write(&cpuset_rwsem);
2176         css_cs(css)->attach_in_progress--;
2177         percpu_up_write(&cpuset_rwsem);
2178 }
2179 
2180 /*
2181  * Protected by cpuset_mutex.  cpus_attach is used only by cpuset_attach()
2182  * but we can't allocate it dynamically there.  Define it global and
2183  * allocate from cpuset_init().
2184  */
2185 static cpumask_var_t cpus_attach;
2186 
2187 static void cpuset_attach(struct cgroup_taskset *tset)
2188 {
2189         /* static buf protected by cpuset_mutex */
2190         static nodemask_t cpuset_attach_nodemask_to;
2191         struct task_struct *task;
2192         struct task_struct *leader;
2193         struct cgroup_subsys_state *css;
2194         struct cpuset *cs;
2195         struct cpuset *oldcs = cpuset_attach_old_cs;
2196 
2197         cgroup_taskset_first(tset, &css);
2198         cs = css_cs(css);
2199 
2200         percpu_down_write(&cpuset_rwsem);
2201 
2202         /* prepare for attach */
2203         if (cs == &top_cpuset)
2204                 cpumask_copy(cpus_attach, cpu_possible_mask);
2205         else
2206                 guarantee_online_cpus(cs, cpus_attach);
2207 
2208         guarantee_online_mems(cs, &cpuset_attach_nodemask_to);
2209 
2210         cgroup_taskset_for_each(task, css, tset) {
2211                 /*
2212                  * can_attach beforehand should guarantee that this doesn't
2213                  * fail.  TODO: have a better way to handle failure here
2214                  */
2215                 WARN_ON_ONCE(set_cpus_allowed_ptr(task, cpus_attach));
2216 
2217                 cpuset_change_task_nodemask(task, &cpuset_attach_nodemask_to);
2218                 cpuset_update_task_spread_flag(cs, task);
2219         }
2220 
2221         /*
2222          * Change mm for all threadgroup leaders. This is expensive and may
2223          * sleep and should be moved outside migration path proper.
2224          */
2225         cpuset_attach_nodemask_to = cs->effective_mems;
2226         cgroup_taskset_for_each_leader(leader, css, tset) {
2227                 struct mm_struct *mm = get_task_mm(leader);
2228 
2229                 if (mm) {
2230                         mpol_rebind_mm(mm, &cpuset_attach_nodemask_to);
2231 
2232                         /*
2233                          * old_mems_allowed is the same with mems_allowed
2234                          * here, except if this task is being moved
2235                          * automatically due to hotplug.  In that case
2236                          * @mems_allowed has been updated and is empty, so
2237                          * @old_mems_allowed is the right nodesets that we
2238                          * migrate mm from.
2239                          */
2240                         if (is_memory_migrate(cs))
2241                                 cpuset_migrate_mm(mm, &oldcs->old_mems_allowed,
2242                                                   &cpuset_attach_nodemask_to);
2243                         else
2244                                 mmput(mm);
2245                 }
2246         }
2247 
2248         cs->old_mems_allowed = cpuset_attach_nodemask_to;
2249 
2250         cs->attach_in_progress--;
2251         if (!cs->attach_in_progress)
2252                 wake_up(&cpuset_attach_wq);
2253 
2254         percpu_up_write(&cpuset_rwsem);
2255 }
2256 
2257 /* The various types of files and directories in a cpuset file system */
2258 
2259 typedef enum {
2260         FILE_MEMORY_MIGRATE,
2261         FILE_CPULIST,
2262         FILE_MEMLIST,
2263         FILE_EFFECTIVE_CPULIST,
2264         FILE_EFFECTIVE_MEMLIST,
2265         FILE_SUBPARTS_CPULIST,
2266         FILE_CPU_EXCLUSIVE,
2267         FILE_MEM_EXCLUSIVE,
2268         FILE_MEM_HARDWALL,
2269         FILE_SCHED_LOAD_BALANCE,
2270         FILE_PARTITION_ROOT,
2271         FILE_SCHED_RELAX_DOMAIN_LEVEL,
2272         FILE_MEMORY_PRESSURE_ENABLED,
2273         FILE_MEMORY_PRESSURE,
2274         FILE_SPREAD_PAGE,
2275         FILE_SPREAD_SLAB,
2276 } cpuset_filetype_t;
2277 
2278 static int cpuset_write_u64(struct cgroup_subsys_state *css, struct cftype *cft,
2279                             u64 val)
2280 {
2281         struct cpuset *cs = css_cs(css);
2282         cpuset_filetype_t type = cft->private;
2283         int retval = 0;
2284 
2285         get_online_cpus();
2286         percpu_down_write(&cpuset_rwsem);
2287         if (!is_cpuset_online(cs)) {
2288                 retval = -ENODEV;
2289                 goto out_unlock;
2290         }
2291 
2292         switch (type) {
2293         case FILE_CPU_EXCLUSIVE:
2294                 retval = update_flag(CS_CPU_EXCLUSIVE, cs, val);
2295                 break;
2296         case FILE_MEM_EXCLUSIVE:
2297                 retval = update_flag(CS_MEM_EXCLUSIVE, cs, val);
2298                 break;
2299         case FILE_MEM_HARDWALL:
2300                 retval = update_flag(CS_MEM_HARDWALL, cs, val);
2301                 break;
2302         case FILE_SCHED_LOAD_BALANCE:
2303                 retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, val);
2304                 break;
2305         case FILE_MEMORY_MIGRATE:
2306                 retval = update_flag(CS_MEMORY_MIGRATE, cs, val);
2307                 break;
2308         case FILE_MEMORY_PRESSURE_ENABLED:
2309                 cpuset_memory_pressure_enabled = !!val;
2310                 break;
2311         case FILE_SPREAD_PAGE:
2312                 retval = update_flag(CS_SPREAD_PAGE, cs, val);
2313                 break;
2314         case FILE_SPREAD_SLAB:
2315                 retval = update_flag(CS_SPREAD_SLAB, cs, val);
2316                 break;
2317         default:
2318                 retval = -EINVAL;
2319                 break;
2320         }
2321 out_unlock:
2322         percpu_up_write(&cpuset_rwsem);
2323         put_online_cpus();
2324         return retval;
2325 }
2326 
2327 static int cpuset_write_s64(struct cgroup_subsys_state *css, struct cftype *cft,
2328                             s64 val)
2329 {
2330         struct cpuset *cs = css_cs(css);
2331         cpuset_filetype_t type = cft->private;
2332         int retval = -ENODEV;
2333 
2334         get_online_cpus();
2335         percpu_down_write(&cpuset_rwsem);
2336         if (!is_cpuset_online(cs))
2337                 goto out_unlock;
2338 
2339         switch (type) {
2340         case FILE_SCHED_RELAX_DOMAIN_LEVEL:
2341                 retval = update_relax_domain_level(cs, val);
2342                 break;
2343         default:
2344                 retval = -EINVAL;
2345                 break;
2346         }
2347 out_unlock:
2348         percpu_up_write(&cpuset_rwsem);
2349         put_online_cpus();
2350         return retval;
2351 }
2352 
2353 /*
2354  * Common handling for a write to a "cpus" or "mems" file.
2355  */
2356 static ssize_t cpuset_write_resmask(struct kernfs_open_file *of,
2357                                     char *buf, size_t nbytes, loff_t off)
2358 {
2359         struct cpuset *cs = css_cs(of_css(of));
2360         struct cpuset *trialcs;
2361         int retval = -ENODEV;
2362 
2363         buf = strstrip(buf);
2364 
2365         /*
2366          * CPU or memory hotunplug may leave @cs w/o any execution
2367          * resources, in which case the hotplug code asynchronously updates
2368          * configuration and transfers all tasks to the nearest ancestor
2369          * which can execute.
2370          *
2371          * As writes to "cpus" or "mems" may restore @cs's execution
2372          * resources, wait for the previously scheduled operations before
2373          * proceeding, so that we don't end up keep removing tasks added
2374          * after execution capability is restored.
2375          *
2376          * cpuset_hotplug_work calls back into cgroup core via
2377          * cgroup_transfer_tasks() and waiting for it from a cgroupfs
2378          * operation like this one can lead to a deadlock through kernfs
2379          * active_ref protection.  Let's break the protection.  Losing the
2380          * protection is okay as we check whether @cs is online after
2381          * grabbing cpuset_mutex anyway.  This only happens on the legacy
2382          * hierarchies.
2383          */
2384         css_get(&cs->css);
2385         kernfs_break_active_protection(of->kn);
2386         flush_work(&cpuset_hotplug_work);
2387 
2388         get_online_cpus();
2389         percpu_down_write(&cpuset_rwsem);
2390         if (!is_cpuset_online(cs))
2391                 goto out_unlock;
2392 
2393         trialcs = alloc_trial_cpuset(cs);
2394         if (!trialcs) {
2395                 retval = -ENOMEM;
2396                 goto out_unlock;
2397         }
2398 
2399         switch (of_cft(of)->private) {
2400         case FILE_CPULIST:
2401                 retval = update_cpumask(cs, trialcs, buf);
2402                 break;
2403         case FILE_MEMLIST:
2404                 retval = update_nodemask(cs, trialcs, buf);
2405                 break;
2406         default:
2407                 retval = -EINVAL;
2408                 break;
2409         }
2410 
2411         free_cpuset(trialcs);
2412 out_unlock:
2413         percpu_up_write(&cpuset_rwsem);
2414         put_online_cpus();
2415         kernfs_unbreak_active_protection(of->kn);
2416         css_put(&cs->css);
2417         flush_workqueue(cpuset_migrate_mm_wq);
2418         return retval ?: nbytes;
2419 }
2420 
2421 /*
2422  * These ascii lists should be read in a single call, by using a user
2423  * buffer large enough to hold the entire map.  If read in smaller
2424  * chunks, there is no guarantee of atomicity.  Since the display format
2425  * used, list of ranges of sequential numbers, is variable length,
2426  * and since these maps can change value dynamically, one could read
2427  * gibberish by doing partial reads while a list was changing.
2428  */
2429 static int cpuset_common_seq_show(struct seq_file *sf, void *v)
2430 {
2431         struct cpuset *cs = css_cs(seq_css(sf));
2432         cpuset_filetype_t type = seq_cft(sf)->private;
2433         int ret = 0;
2434 
2435         spin_lock_irq(&callback_lock);
2436 
2437         switch (type) {
2438         case FILE_CPULIST:
2439                 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->cpus_allowed));
2440                 break;
2441         case FILE_MEMLIST:
2442                 seq_printf(sf, "%*pbl\n", nodemask_pr_args(&cs->mems_allowed));
2443                 break;
2444         case FILE_EFFECTIVE_CPULIST:
2445                 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->effective_cpus));
2446                 break;
2447         case FILE_EFFECTIVE_MEMLIST:
2448                 seq_printf(sf, "%*pbl\n", nodemask_pr_args(&cs->effective_mems));
2449                 break;
2450         case FILE_SUBPARTS_CPULIST:
2451                 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->subparts_cpus));
2452                 break;
2453         default:
2454                 ret = -EINVAL;
2455         }
2456 
2457         spin_unlock_irq(&callback_lock);
2458         return ret;
2459 }
2460 
2461 static u64 cpuset_read_u64(struct cgroup_subsys_state *css, struct cftype *cft)
2462 {
2463         struct cpuset *cs = css_cs(css);
2464         cpuset_filetype_t type = cft->private;
2465         switch (type) {
2466         case FILE_CPU_EXCLUSIVE:
2467                 return is_cpu_exclusive(cs);
2468         case FILE_MEM_EXCLUSIVE:
2469                 return is_mem_exclusive(cs);
2470         case FILE_MEM_HARDWALL:
2471                 return is_mem_hardwall(cs);
2472         case FILE_SCHED_LOAD_BALANCE:
2473                 return is_sched_load_balance(cs);
2474         case FILE_MEMORY_MIGRATE:
2475                 return is_memory_migrate(cs);
2476         case FILE_MEMORY_PRESSURE_ENABLED:
2477                 return cpuset_memory_pressure_enabled;
2478         case FILE_MEMORY_PRESSURE:
2479                 return fmeter_getrate(&cs->fmeter);
2480         case FILE_SPREAD_PAGE:
2481                 return is_spread_page(cs);
2482         case FILE_SPREAD_SLAB:
2483                 return is_spread_slab(cs);
2484         default:
2485                 BUG();
2486         }
2487 
2488         /* Unreachable but makes gcc happy */
2489         return 0;
2490 }
2491 
2492 static s64 cpuset_read_s64(struct cgroup_subsys_state *css, struct cftype *cft)
2493 {
2494         struct cpuset *cs = css_cs(css);
2495         cpuset_filetype_t type = cft->private;
2496         switch (type) {
2497         case FILE_SCHED_RELAX_DOMAIN_LEVEL:
2498                 return cs->relax_domain_level;
2499         default:
2500                 BUG();
2501         }
2502 
2503         /* Unreachable but makes gcc happy */
2504         return 0;
2505 }
2506 
2507 static int sched_partition_show(struct seq_file *seq, void *v)
2508 {
2509         struct cpuset *cs = css_cs(seq_css(seq));
2510 
2511         switch (cs->partition_root_state) {
2512         case PRS_ENABLED:
2513                 seq_puts(seq, "root\n");
2514                 break;
2515         case PRS_DISABLED:
2516                 seq_puts(seq, "member\n");
2517                 break;
2518         case PRS_ERROR:
2519                 seq_puts(seq, "root invalid\n");
2520                 break;
2521         }
2522         return 0;
2523 }
2524 
2525 static ssize_t sched_partition_write(struct kernfs_open_file *of, char *buf,
2526                                      size_t nbytes, loff_t off)
2527 {
2528         struct cpuset *cs = css_cs(of_css(of));
2529         int val;
2530         int retval = -ENODEV;
2531 
2532         buf = strstrip(buf);
2533 
2534         /*
2535          * Convert "root" to ENABLED, and convert "member" to DISABLED.
2536          */
2537         if (!strcmp(buf, "root"))
2538                 val = PRS_ENABLED;
2539         else if (!strcmp(buf, "member"))
2540                 val = PRS_DISABLED;
2541         else
2542                 return -EINVAL;
2543 
2544         css_get(&cs->css);
2545         get_online_cpus();
2546         percpu_down_write(&cpuset_rwsem);
2547         if (!is_cpuset_online(cs))
2548                 goto out_unlock;
2549 
2550         retval = update_prstate(cs, val);
2551 out_unlock:
2552         percpu_up_write(&cpuset_rwsem);
2553         put_online_cpus();
2554         css_put(&cs->css);
2555         return retval ?: nbytes;
2556 }
2557 
2558 /*
2559  * for the common functions, 'private' gives the type of file
2560  */
2561 
2562 static struct cftype legacy_files[] = {
2563         {
2564                 .name = "cpus",
2565                 .seq_show = cpuset_common_seq_show,
2566                 .write = cpuset_write_resmask,
2567                 .max_write_len = (100U + 6 * NR_CPUS),
2568                 .private = FILE_CPULIST,
2569         },
2570 
2571         {
2572                 .name = "mems",
2573                 .seq_show = cpuset_common_seq_show,
2574                 .write = cpuset_write_resmask,
2575                 .max_write_len = (100U + 6 * MAX_NUMNODES),
2576                 .private = FILE_MEMLIST,
2577         },
2578 
2579         {
2580                 .name = "effective_cpus",
2581                 .seq_show = cpuset_common_seq_show,
2582                 .private = FILE_EFFECTIVE_CPULIST,
2583         },
2584 
2585         {
2586                 .name = "effective_mems",
2587                 .seq_show = cpuset_common_seq_show,
2588                 .private = FILE_EFFECTIVE_MEMLIST,
2589         },
2590 
2591         {
2592                 .name = "cpu_exclusive",
2593                 .read_u64 = cpuset_read_u64,
2594                 .write_u64 = cpuset_write_u64,
2595                 .private = FILE_CPU_EXCLUSIVE,
2596         },
2597 
2598         {
2599                 .name = "mem_exclusive",
2600                 .read_u64 = cpuset_read_u64,
2601                 .write_u64 = cpuset_write_u64,
2602                 .private = FILE_MEM_EXCLUSIVE,
2603         },
2604 
2605         {
2606                 .name = "mem_hardwall",
2607                 .read_u64 = cpuset_read_u64,
2608                 .write_u64 = cpuset_write_u64,
2609                 .private = FILE_MEM_HARDWALL,
2610         },
2611 
2612         {
2613                 .name = "sched_load_balance",
2614                 .read_u64 = cpuset_read_u64,
2615                 .write_u64 = cpuset_write_u64,
2616                 .private = FILE_SCHED_LOAD_BALANCE,
2617         },
2618 
2619         {
2620                 .name = "sched_relax_domain_level",
2621                 .read_s64 = cpuset_read_s64,
2622                 .write_s64 = cpuset_write_s64,
2623                 .private = FILE_SCHED_RELAX_DOMAIN_LEVEL,
2624         },
2625 
2626         {
2627                 .name = "memory_migrate",
2628                 .read_u64 = cpuset_read_u64,
2629                 .write_u64 = cpuset_write_u64,
2630                 .private = FILE_MEMORY_MIGRATE,
2631         },
2632 
2633         {
2634                 .name = "memory_pressure",
2635                 .read_u64 = cpuset_read_u64,
2636                 .private = FILE_MEMORY_PRESSURE,
2637         },
2638 
2639         {
2640                 .name = "memory_spread_page",
2641                 .read_u64 = cpuset_read_u64,
2642                 .write_u64 = cpuset_write_u64,
2643                 .private = FILE_SPREAD_PAGE,
2644         },
2645 
2646         {
2647                 .name = "memory_spread_slab",
2648                 .read_u64 = cpuset_read_u64,
2649                 .write_u64 = cpuset_write_u64,
2650                 .private = FILE_SPREAD_SLAB,
2651         },
2652 
2653         {
2654                 .name = "memory_pressure_enabled",
2655                 .flags = CFTYPE_ONLY_ON_ROOT,
2656                 .read_u64 = cpuset_read_u64,
2657                 .write_u64 = cpuset_write_u64,
2658                 .private = FILE_MEMORY_PRESSURE_ENABLED,
2659         },
2660 
2661         { }     /* terminate */
2662 };
2663 
2664 /*
2665  * This is currently a minimal set for the default hierarchy. It can be
2666  * expanded later on by migrating more features and control files from v1.
2667  */
2668 static struct cftype dfl_files[] = {
2669         {
2670                 .name = "cpus",
2671                 .seq_show = cpuset_common_seq_show,
2672                 .write = cpuset_write_resmask,
2673                 .max_write_len = (100U + 6 * NR_CPUS),
2674                 .private = FILE_CPULIST,
2675                 .flags = CFTYPE_NOT_ON_ROOT,
2676         },
2677 
2678         {
2679                 .name = "mems",
2680                 .seq_show = cpuset_common_seq_show,
2681                 .write = cpuset_write_resmask,
2682                 .max_write_len = (100U + 6 * MAX_NUMNODES),
2683                 .private = FILE_MEMLIST,
2684                 .flags = CFTYPE_NOT_ON_ROOT,
2685         },
2686 
2687         {
2688                 .name = "cpus.effective",
2689                 .seq_show = cpuset_common_seq_show,
2690                 .private = FILE_EFFECTIVE_CPULIST,
2691         },
2692 
2693         {
2694                 .name = "mems.effective",
2695                 .seq_show = cpuset_common_seq_show,
2696                 .private = FILE_EFFECTIVE_MEMLIST,
2697         },
2698 
2699         {
2700                 .name = "cpus.partition",
2701                 .seq_show = sched_partition_show,
2702                 .write = sched_partition_write,
2703                 .private = FILE_PARTITION_ROOT,
2704                 .flags = CFTYPE_NOT_ON_ROOT,
2705         },
2706 
2707         {
2708                 .name = "cpus.subpartitions",
2709                 .seq_show = cpuset_common_seq_show,
2710                 .private = FILE_SUBPARTS_CPULIST,
2711                 .flags = CFTYPE_DEBUG,
2712         },
2713 
2714         { }     /* terminate */
2715 };
2716 
2717 
2718 /*
2719  *      cpuset_css_alloc - allocate a cpuset css
2720  *      cgrp:   control group that the new cpuset will be part of
2721  */
2722 
2723 static struct cgroup_subsys_state *
2724 cpuset_css_alloc(struct cgroup_subsys_state *parent_css)
2725 {
2726         struct cpuset *cs;
2727 
2728         if (!parent_css)
2729                 return &top_cpuset.css;
2730 
2731         cs = kzalloc(sizeof(*cs), GFP_KERNEL);
2732         if (!cs)
2733                 return ERR_PTR(-ENOMEM);
2734 
2735         if (alloc_cpumasks(cs, NULL)) {
2736                 kfree(cs);
2737                 return ERR_PTR(-ENOMEM);
2738         }
2739 
2740         set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
2741         nodes_clear(cs->mems_allowed);
2742         nodes_clear(cs->effective_mems);
2743         fmeter_init(&cs->fmeter);
2744         cs->relax_domain_level = -1;
2745 
2746         return &cs->css;
2747 }
2748 
2749 static int cpuset_css_online(struct cgroup_subsys_state *css)
2750 {
2751         struct cpuset *cs = css_cs(css);
2752         struct cpuset *parent = parent_cs(cs);
2753         struct cpuset *tmp_cs;
2754         struct cgroup_subsys_state *pos_css;
2755 
2756         if (!parent)
2757                 return 0;
2758 
2759         get_online_cpus();
2760         percpu_down_write(&cpuset_rwsem);
2761 
2762         set_bit(CS_ONLINE, &cs->flags);
2763         if (is_spread_page(parent))
2764                 set_bit(CS_SPREAD_PAGE, &cs->flags);
2765         if (is_spread_slab(parent))
2766                 set_bit(CS_SPREAD_SLAB, &cs->flags);
2767 
2768         cpuset_inc();
2769 
2770         spin_lock_irq(&callback_lock);
2771         if (is_in_v2_mode()) {
2772                 cpumask_copy(cs->effective_cpus, parent->effective_cpus);
2773                 cs->effective_mems = parent->effective_mems;
2774                 cs->use_parent_ecpus = true;
2775                 parent->child_ecpus_count++;
2776         }
2777         spin_unlock_irq(&callback_lock);
2778 
2779         if (!test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags))
2780                 goto out_unlock;
2781 
2782         /*
2783          * Clone @parent's configuration if CGRP_CPUSET_CLONE_CHILDREN is
2784          * set.  This flag handling is implemented in cgroup core for
2785          * histrical reasons - the flag may be specified during mount.
2786          *
2787          * Currently, if any sibling cpusets have exclusive cpus or mem, we
2788          * refuse to clone the configuration - thereby refusing the task to
2789          * be entered, and as a result refusing the sys_unshare() or
2790          * clone() which initiated it.  If this becomes a problem for some
2791          * users who wish to allow that scenario, then this could be
2792          * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
2793          * (and likewise for mems) to the new cgroup.
2794          */
2795         rcu_read_lock();
2796         cpuset_for_each_child(tmp_cs, pos_css, parent) {
2797                 if (is_mem_exclusive(tmp_cs) || is_cpu_exclusive(tmp_cs)) {
2798                         rcu_read_unlock();
2799                         goto out_unlock;
2800                 }
2801         }
2802         rcu_read_unlock();
2803 
2804         spin_lock_irq(&callback_lock);
2805         cs->mems_allowed = parent->mems_allowed;
2806         cs->effective_mems = parent->mems_allowed;
2807         cpumask_copy(cs->cpus_allowed, parent->cpus_allowed);
2808         cpumask_copy(cs->effective_cpus, parent->cpus_allowed);
2809         spin_unlock_irq(&callback_lock);
2810 out_unlock:
2811         percpu_up_write(&cpuset_rwsem);
2812         put_online_cpus();
2813         return 0;
2814 }
2815 
2816 /*
2817  * If the cpuset being removed has its flag 'sched_load_balance'
2818  * enabled, then simulate turning sched_load_balance off, which
2819  * will call rebuild_sched_domains_locked(). That is not needed
2820  * in the default hierarchy where only changes in partition
2821  * will cause repartitioning.
2822  *
2823  * If the cpuset has the 'sched.partition' flag enabled, simulate
2824  * turning 'sched.partition" off.
2825  */
2826 
2827 static void cpuset_css_offline(struct cgroup_subsys_state *css)
2828 {
2829         struct cpuset *cs = css_cs(css);
2830 
2831         get_online_cpus();
2832         percpu_down_write(&cpuset_rwsem);
2833 
2834         if (is_partition_root(cs))
2835                 update_prstate(cs, 0);
2836 
2837         if (!cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
2838             is_sched_load_balance(cs))
2839                 update_flag(CS_SCHED_LOAD_BALANCE, cs, 0);
2840 
2841         if (cs->use_parent_ecpus) {
2842                 struct cpuset *parent = parent_cs(cs);
2843 
2844                 cs->use_parent_ecpus = false;
2845                 parent->child_ecpus_count--;
2846         }
2847 
2848         cpuset_dec();
2849         clear_bit(CS_ONLINE, &cs->flags);
2850 
2851         percpu_up_write(&cpuset_rwsem);
2852         put_online_cpus();
2853 }
2854 
2855 static void cpuset_css_free(struct cgroup_subsys_state *css)
2856 {
2857         struct cpuset *cs = css_cs(css);
2858 
2859         free_cpuset(cs);
2860 }
2861 
2862 static void cpuset_bind(struct cgroup_subsys_state *root_css)
2863 {
2864         percpu_down_write(&cpuset_rwsem);
2865         spin_lock_irq(&callback_lock);
2866 
2867         if (is_in_v2_mode()) {
2868                 cpumask_copy(top_cpuset.cpus_allowed, cpu_possible_mask);
2869                 top_cpuset.mems_allowed = node_possible_map;
2870         } else {
2871                 cpumask_copy(top_cpuset.cpus_allowed,
2872                              top_cpuset.effective_cpus);
2873                 top_cpuset.mems_allowed = top_cpuset.effective_mems;
2874         }
2875 
2876         spin_unlock_irq(&callback_lock);
2877         percpu_up_write(&cpuset_rwsem);
2878 }
2879 
2880 /*
2881  * Make sure the new task conform to the current state of its parent,
2882  * which could have been changed by cpuset just after it inherits the
2883  * state from the parent and before it sits on the cgroup's task list.
2884  */
2885 static void cpuset_fork(struct task_struct *task)
2886 {
2887         if (task_css_is_root(task, cpuset_cgrp_id))
2888                 return;
2889 
2890         set_cpus_allowed_ptr(task, current->cpus_ptr);
2891         task->mems_allowed = current->mems_allowed;
2892 }
2893 
2894 struct cgroup_subsys cpuset_cgrp_subsys = {
2895         .css_alloc      = cpuset_css_alloc,
2896         .css_online     = cpuset_css_online,
2897         .css_offline    = cpuset_css_offline,
2898         .css_free       = cpuset_css_free,
2899         .can_attach     = cpuset_can_attach,
2900         .cancel_attach  = cpuset_cancel_attach,
2901         .attach         = cpuset_attach,
2902         .post_attach    = cpuset_post_attach,
2903         .bind           = cpuset_bind,
2904         .fork           = cpuset_fork,
2905         .legacy_cftypes = legacy_files,
2906         .dfl_cftypes    = dfl_files,
2907         .early_init     = true,
2908         .threaded       = true,
2909 };
2910 
2911 /**
2912  * cpuset_init - initialize cpusets at system boot
2913  *
2914  * Description: Initialize top_cpuset
2915  **/
2916 
2917 int __init cpuset_init(void)
2918 {
2919         BUG_ON(percpu_init_rwsem(&cpuset_rwsem));
2920 
2921         BUG_ON(!alloc_cpumask_var(&top_cpuset.cpus_allowed, GFP_KERNEL));
2922         BUG_ON(!alloc_cpumask_var(&top_cpuset.effective_cpus, GFP_KERNEL));
2923         BUG_ON(!zalloc_cpumask_var(&top_cpuset.subparts_cpus, GFP_KERNEL));
2924 
2925         cpumask_setall(top_cpuset.cpus_allowed);
2926         nodes_setall(top_cpuset.mems_allowed);
2927         cpumask_setall(top_cpuset.effective_cpus);
2928         nodes_setall(top_cpuset.effective_mems);
2929 
2930         fmeter_init(&top_cpuset.fmeter);
2931         set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags);
2932         top_cpuset.relax_domain_level = -1;
2933 
2934         BUG_ON(!alloc_cpumask_var(&cpus_attach, GFP_KERNEL));
2935 
2936         return 0;
2937 }
2938 
2939 /*
2940  * If CPU and/or memory hotplug handlers, below, unplug any CPUs
2941  * or memory nodes, we need to walk over the cpuset hierarchy,
2942  * removing that CPU or node from all cpusets.  If this removes the
2943  * last CPU or node from a cpuset, then move the tasks in the empty
2944  * cpuset to its next-highest non-empty parent.
2945  */
2946 static void remove_tasks_in_empty_cpuset(struct cpuset *cs)
2947 {
2948         struct cpuset *parent;
2949 
2950         /*
2951          * Find its next-highest non-empty parent, (top cpuset
2952          * has online cpus, so can't be empty).
2953          */
2954         parent = parent_cs(cs);
2955         while (cpumask_empty(parent->cpus_allowed) ||
2956                         nodes_empty(parent->mems_allowed))
2957                 parent = parent_cs(parent);
2958 
2959         if (cgroup_transfer_tasks(parent->css.cgroup, cs->css.cgroup)) {
2960                 pr_err("cpuset: failed to transfer tasks out of empty cpuset ");
2961                 pr_cont_cgroup_name(cs->css.cgroup);
2962                 pr_cont("\n");
2963         }
2964 }
2965 
2966 static void
2967 hotplug_update_tasks_legacy(struct cpuset *cs,
2968                             struct cpumask *new_cpus, nodemask_t *new_mems,
2969                             bool cpus_updated, bool mems_updated)
2970 {
2971         bool is_empty;
2972 
2973         spin_lock_irq(&callback_lock);
2974         cpumask_copy(cs->cpus_allowed, new_cpus);
2975         cpumask_copy(cs->effective_cpus, new_cpus);
2976         cs->mems_allowed = *new_mems;
2977         cs->effective_mems = *new_mems;
2978         spin_unlock_irq(&callback_lock);
2979 
2980         /*
2981          * Don't call update_tasks_cpumask() if the cpuset becomes empty,
2982          * as the tasks will be migratecd to an ancestor.
2983          */
2984         if (cpus_updated && !cpumask_empty(cs->cpus_allowed))
2985                 update_tasks_cpumask(cs);
2986         if (mems_updated && !nodes_empty(cs->mems_allowed))
2987                 update_tasks_nodemask(cs);
2988 
2989         is_empty = cpumask_empty(cs->cpus_allowed) ||
2990                    nodes_empty(cs->mems_allowed);
2991 
2992         percpu_up_write(&cpuset_rwsem);
2993 
2994         /*
2995          * Move tasks to the nearest ancestor with execution resources,
2996          * This is full cgroup operation which will also call back into
2997          * cpuset. Should be done outside any lock.
2998          */
2999         if (is_empty)
3000                 remove_tasks_in_empty_cpuset(cs);
3001 
3002         percpu_down_write(&cpuset_rwsem);
3003 }
3004 
3005 static void
3006 hotplug_update_tasks(struct cpuset *cs,
3007                      struct cpumask *new_cpus, nodemask_t *new_mems,
3008                      bool cpus_updated, bool mems_updated)
3009 {
3010         if (cpumask_empty(new_cpus))
3011                 cpumask_copy(new_cpus, parent_cs(cs)->effective_cpus);
3012         if (nodes_empty(*new_mems))
3013                 *new_mems = parent_cs(cs)->effective_mems;
3014 
3015         spin_lock_irq(&callback_lock);
3016         cpumask_copy(cs->effective_cpus, new_cpus);
3017         cs->effective_mems = *new_mems;
3018         spin_unlock_irq(&callback_lock);
3019 
3020         if (cpus_updated)
3021                 update_tasks_cpumask(cs);
3022         if (mems_updated)
3023                 update_tasks_nodemask(cs);
3024 }
3025 
3026 static bool force_rebuild;
3027 
3028 void cpuset_force_rebuild(void)
3029 {
3030         force_rebuild = true;
3031 }
3032 
3033 /**
3034  * cpuset_hotplug_update_tasks - update tasks in a cpuset for hotunplug
3035  * @cs: cpuset in interest
3036  * @tmp: the tmpmasks structure pointer
3037  *
3038  * Compare @cs's cpu and mem masks against top_cpuset and if some have gone
3039  * offline, update @cs accordingly.  If @cs ends up with no CPU or memory,
3040  * all its tasks are moved to the nearest ancestor with both resources.
3041  */
3042 static void cpuset_hotplug_update_tasks(struct cpuset *cs, struct tmpmasks *tmp)
3043 {
3044         static cpumask_t new_cpus;
3045         static nodemask_t new_mems;
3046         bool cpus_updated;
3047         bool mems_updated;
3048         struct cpuset *parent;
3049 retry:
3050         wait_event(cpuset_attach_wq, cs->attach_in_progress == 0);
3051 
3052         percpu_down_write(&cpuset_rwsem);
3053 
3054         /*
3055          * We have raced with task attaching. We wait until attaching
3056          * is finished, so we won't attach a task to an empty cpuset.
3057          */
3058         if (cs->attach_in_progress) {
3059                 percpu_up_write(&cpuset_rwsem);
3060                 goto retry;
3061         }
3062 
3063         parent =  parent_cs(cs);
3064         compute_effective_cpumask(&new_cpus, cs, parent);
3065         nodes_and(new_mems, cs->mems_allowed, parent->effective_mems);
3066 
3067         if (cs->nr_subparts_cpus)
3068                 /*
3069                  * Make sure that CPUs allocated to child partitions
3070                  * do not show up in effective_cpus.
3071                  */
3072                 cpumask_andnot(&new_cpus, &new_cpus, cs->subparts_cpus);
3073 
3074         if (!tmp || !cs->partition_root_state)
3075                 goto update_tasks;
3076 
3077         /*
3078          * In the unlikely event that a partition root has empty
3079          * effective_cpus or its parent becomes erroneous, we have to
3080          * transition it to the erroneous state.
3081          */
3082         if (is_partition_root(cs) && (cpumask_empty(&new_cpus) ||
3083            (parent->partition_root_state == PRS_ERROR))) {
3084                 if (cs->nr_subparts_cpus) {
3085                         cs->nr_subparts_cpus = 0;
3086                         cpumask_clear(cs->subparts_cpus);
3087                         compute_effective_cpumask(&new_cpus, cs, parent);
3088                 }
3089 
3090                 /*
3091                  * If the effective_cpus is empty because the child
3092                  * partitions take away all the CPUs, we can keep
3093                  * the current partition and let the child partitions
3094                  * fight for available CPUs.
3095                  */
3096                 if ((parent->partition_root_state == PRS_ERROR) ||
3097                      cpumask_empty(&new_cpus)) {
3098                         update_parent_subparts_cpumask(cs, partcmd_disable,
3099                                                        NULL, tmp);
3100                         cs->partition_root_state = PRS_ERROR;
3101                 }
3102                 cpuset_force_rebuild();
3103         }
3104 
3105         /*
3106          * On the other hand, an erroneous partition root may be transitioned
3107          * back to a regular one or a partition root with no CPU allocated
3108          * from the parent may change to erroneous.
3109          */
3110         if (is_partition_root(parent) &&
3111            ((cs->partition_root_state == PRS_ERROR) ||
3112             !cpumask_intersects(&new_cpus, parent->subparts_cpus)) &&
3113              update_parent_subparts_cpumask(cs, partcmd_update, NULL, tmp))
3114                 cpuset_force_rebuild();
3115 
3116 update_tasks:
3117         cpus_updated = !cpumask_equal(&new_cpus, cs->effective_cpus);
3118         mems_updated = !nodes_equal(new_mems, cs->effective_mems);
3119 
3120         if (is_in_v2_mode())
3121                 hotplug_update_tasks(cs, &new_cpus, &new_mems,
3122                                      cpus_updated, mems_updated);
3123         else
3124                 hotplug_update_tasks_legacy(cs, &new_cpus, &new_mems,
3125                                             cpus_updated, mems_updated);
3126 
3127         percpu_up_write(&cpuset_rwsem);
3128 }
3129 
3130 /**
3131  * cpuset_hotplug_workfn - handle CPU/memory hotunplug for a cpuset
3132  *
3133  * This function is called after either CPU or memory configuration has
3134  * changed and updates cpuset accordingly.  The top_cpuset is always
3135  * synchronized to cpu_active_mask and N_MEMORY, which is necessary in
3136  * order to make cpusets transparent (of no affect) on systems that are
3137  * actively using CPU hotplug but making no active use of cpusets.
3138  *
3139  * Non-root cpusets are only affected by offlining.  If any CPUs or memory
3140  * nodes have been taken down, cpuset_hotplug_update_tasks() is invoked on
3141  * all descendants.
3142  *
3143  * Note that CPU offlining during suspend is ignored.  We don't modify
3144  * cpusets across suspend/resume cycles at all.
3145  */
3146 static void cpuset_hotplug_workfn(struct work_struct *work)
3147 {
3148         static cpumask_t new_cpus;
3149         static nodemask_t new_mems;
3150         bool cpus_updated, mems_updated;
3151         bool on_dfl = is_in_v2_mode();
3152         struct tmpmasks tmp, *ptmp = NULL;
3153 
3154         if (on_dfl && !alloc_cpumasks(NULL, &tmp))
3155                 ptmp = &tmp;
3156 
3157         percpu_down_write(&cpuset_rwsem);
3158 
3159         /* fetch the available cpus/mems and find out which changed how */
3160         cpumask_copy(&new_cpus, cpu_active_mask);
3161         new_mems = node_states[N_MEMORY];
3162 
3163         /*
3164          * If subparts_cpus is populated, it is likely that the check below
3165          * will produce a false positive on cpus_updated when the cpu list
3166          * isn't changed. It is extra work, but it is better to be safe.
3167          */
3168         cpus_updated = !cpumask_equal(top_cpuset.effective_cpus, &new_cpus);
3169         mems_updated = !nodes_equal(top_cpuset.effective_mems, new_mems);
3170 
3171         /* synchronize cpus_allowed to cpu_active_mask */
3172         if (cpus_updated) {
3173                 spin_lock_irq(&callback_lock);
3174                 if (!on_dfl)
3175                         cpumask_copy(top_cpuset.cpus_allowed, &new_cpus);
3176                 /*
3177                  * Make sure that CPUs allocated to child partitions
3178                  * do not show up in effective_cpus. If no CPU is left,
3179                  * we clear the subparts_cpus & let the child partitions
3180                  * fight for the CPUs again.
3181                  */
3182                 if (top_cpuset.nr_subparts_cpus) {
3183                         if (cpumask_subset(&new_cpus,
3184                                            top_cpuset.subparts_cpus)) {
3185                                 top_cpuset.nr_subparts_cpus = 0;
3186                                 cpumask_clear(top_cpuset.subparts_cpus);
3187                         } else {
3188                                 cpumask_andnot(&new_cpus, &new_cpus,
3189                                                top_cpuset.subparts_cpus);
3190                         }
3191                 }
3192                 cpumask_copy(top_cpuset.effective_cpus, &new_cpus);
3193                 spin_unlock_irq(&callback_lock);
3194                 /* we don't mess with cpumasks of tasks in top_cpuset */
3195         }
3196 
3197         /* synchronize mems_allowed to N_MEMORY */
3198         if (mems_updated) {
3199                 spin_lock_irq(&callback_lock);
3200                 if (!on_dfl)
3201                         top_cpuset.mems_allowed = new_mems;
3202                 top_cpuset.effective_mems = new_mems;
3203                 spin_unlock_irq(&callback_lock);
3204                 update_tasks_nodemask(&top_cpuset);
3205         }
3206 
3207         percpu_up_write(&cpuset_rwsem);
3208 
3209         /* if cpus or mems changed, we need to propagate to descendants */
3210         if (cpus_updated || mems_updated) {
3211                 struct cpuset *cs;
3212                 struct cgroup_subsys_state *pos_css;
3213 
3214                 rcu_read_lock();
3215                 cpuset_for_each_descendant_pre(cs, pos_css, &top_cpuset) {
3216                         if (cs == &top_cpuset || !css_tryget_online(&cs->css))
3217                                 continue;
3218                         rcu_read_unlock();
3219 
3220                         cpuset_hotplug_update_tasks(cs, ptmp);
3221 
3222                         rcu_read_lock();
3223                         css_put(&cs->css);
3224                 }
3225                 rcu_read_unlock();
3226         }
3227 
3228         /* rebuild sched domains if cpus_allowed has changed */
3229         if (cpus_updated || force_rebuild) {
3230                 force_rebuild = false;
3231                 rebuild_sched_domains();
3232         }
3233 
3234         free_cpumasks(NULL, ptmp);
3235 }
3236 
3237 void cpuset_update_active_cpus(void)
3238 {
3239         /*
3240          * We're inside cpu hotplug critical region which usually nests
3241          * inside cgroup synchronization.  Bounce actual hotplug processing
3242          * to a work item to avoid reverse locking order.
3243          */
3244         schedule_work(&cpuset_hotplug_work);
3245 }
3246 
3247 void cpuset_wait_for_hotplug(void)
3248 {
3249         flush_work(&cpuset_hotplug_work);
3250 }
3251 
3252 /*
3253  * Keep top_cpuset.mems_allowed tracking node_states[N_MEMORY].
3254  * Call this routine anytime after node_states[N_MEMORY] changes.
3255  * See cpuset_update_active_cpus() for CPU hotplug handling.
3256  */
3257 static int cpuset_track_online_nodes(struct notifier_block *self,
3258                                 unsigned long action, void *arg)
3259 {
3260         schedule_work(&cpuset_hotplug_work);
3261         return NOTIFY_OK;
3262 }
3263 
3264 static struct notifier_block cpuset_track_online_nodes_nb = {
3265         .notifier_call = cpuset_track_online_nodes,
3266         .priority = 10,         /* ??! */
3267 };
3268 
3269 /**
3270  * cpuset_init_smp - initialize cpus_allowed
3271  *
3272  * Description: Finish top cpuset after cpu, node maps are initialized
3273  */
3274 void __init cpuset_init_smp(void)
3275 {
3276         cpumask_copy(top_cpuset.cpus_allowed, cpu_active_mask);
3277         top_cpuset.mems_allowed = node_states[N_MEMORY];
3278         top_cpuset.old_mems_allowed = top_cpuset.mems_allowed;
3279 
3280         cpumask_copy(top_cpuset.effective_cpus, cpu_active_mask);
3281         top_cpuset.effective_mems = node_states[N_MEMORY];
3282 
3283         register_hotmemory_notifier(&cpuset_track_online_nodes_nb);
3284 
3285         cpuset_migrate_mm_wq = alloc_ordered_workqueue("cpuset_migrate_mm", 0);
3286         BUG_ON(!cpuset_migrate_mm_wq);
3287 }
3288 
3289 /**
3290  * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
3291  * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
3292  * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
3293  *
3294  * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
3295  * attached to the specified @tsk.  Guaranteed to return some non-empty
3296  * subset of cpu_online_mask, even if this means going outside the
3297  * tasks cpuset.
3298  **/
3299 
3300 void cpuset_cpus_allowed(struct task_struct *tsk, struct cpumask *pmask)
3301 {
3302         unsigned long flags;
3303 
3304         spin_lock_irqsave(&callback_lock, flags);
3305         rcu_read_lock();
3306         guarantee_online_cpus(task_cs(tsk), pmask);
3307         rcu_read_unlock();
3308         spin_unlock_irqrestore(&callback_lock, flags);
3309 }
3310 
3311 /**
3312  * cpuset_cpus_allowed_fallback - final fallback before complete catastrophe.
3313  * @tsk: pointer to task_struct with which the scheduler is struggling
3314  *
3315  * Description: In the case that the scheduler cannot find an allowed cpu in
3316  * tsk->cpus_allowed, we fall back to task_cs(tsk)->cpus_allowed. In legacy
3317  * mode however, this value is the same as task_cs(tsk)->effective_cpus,
3318  * which will not contain a sane cpumask during cases such as cpu hotplugging.
3319  * This is the absolute last resort for the scheduler and it is only used if
3320  * _every_ other avenue has been traveled.
3321  **/
3322 
3323 void cpuset_cpus_allowed_fallback(struct task_struct *tsk)
3324 {
3325         rcu_read_lock();
3326         do_set_cpus_allowed(tsk, is_in_v2_mode() ?
3327                 task_cs(tsk)->cpus_allowed : cpu_possible_mask);
3328         rcu_read_unlock();
3329 
3330         /*
3331          * We own tsk->cpus_allowed, nobody can change it under us.
3332          *
3333          * But we used cs && cs->cpus_allowed lockless and thus can
3334          * race with cgroup_attach_task() or update_cpumask() and get
3335          * the wrong tsk->cpus_allowed. However, both cases imply the
3336          * subsequent cpuset_change_cpumask()->set_cpus_allowed_ptr()
3337          * which takes task_rq_lock().
3338          *
3339          * If we are called after it dropped the lock we must see all
3340          * changes in tsk_cs()->cpus_allowed. Otherwise we can temporary
3341          * set any mask even if it is not right from task_cs() pov,
3342          * the pending set_cpus_allowed_ptr() will fix things.
3343          *
3344          * select_fallback_rq() will fix things ups and set cpu_possible_mask
3345          * if required.
3346          */
3347 }
3348 
3349 void __init cpuset_init_current_mems_allowed(void)
3350 {
3351         nodes_setall(current->mems_allowed);
3352 }
3353 
3354 /**
3355  * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
3356  * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
3357  *
3358  * Description: Returns the nodemask_t mems_allowed of the cpuset
3359  * attached to the specified @tsk.  Guaranteed to return some non-empty
3360  * subset of node_states[N_MEMORY], even if this means going outside the
3361  * tasks cpuset.
3362  **/
3363 
3364 nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
3365 {
3366         nodemask_t mask;
3367         unsigned long flags;
3368 
3369         spin_lock_irqsave(&callback_lock, flags);
3370         rcu_read_lock();
3371         guarantee_online_mems(task_cs(tsk), &mask);
3372         rcu_read_unlock();
3373         spin_unlock_irqrestore(&callback_lock, flags);
3374 
3375         return mask;
3376 }
3377 
3378 /**
3379  * cpuset_nodemask_valid_mems_allowed - check nodemask vs. current mems_allowed
3380  * @nodemask: the nodemask to be checked
3381  *
3382  * Are any of the nodes in the nodemask allowed in current->mems_allowed?
3383  */
3384 int cpuset_nodemask_valid_mems_allowed(nodemask_t *nodemask)
3385 {
3386         return nodes_intersects(*nodemask, current->mems_allowed);
3387 }
3388 
3389 /*
3390  * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
3391  * mem_hardwall ancestor to the specified cpuset.  Call holding
3392  * callback_lock.  If no ancestor is mem_exclusive or mem_hardwall
3393  * (an unusual configuration), then returns the root cpuset.
3394  */
3395 static struct cpuset *nearest_hardwall_ancestor(struct cpuset *cs)
3396 {
3397         while (!(is_mem_exclusive(cs) || is_mem_hardwall(cs)) && parent_cs(cs))
3398                 cs = parent_cs(cs);
3399         return cs;
3400 }
3401 
3402 /**
3403  * cpuset_node_allowed - Can we allocate on a memory node?
3404  * @node: is this an allowed node?
3405  * @gfp_mask: memory allocation flags
3406  *
3407  * If we're in interrupt, yes, we can always allocate.  If @node is set in
3408  * current's mems_allowed, yes.  If it's not a __GFP_HARDWALL request and this
3409  * node is set in the nearest hardwalled cpuset ancestor to current's cpuset,
3410  * yes.  If current has access to memory reserves as an oom victim, yes.
3411  * Otherwise, no.
3412  *
3413  * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
3414  * and do not allow allocations outside the current tasks cpuset
3415  * unless the task has been OOM killed.
3416  * GFP_KERNEL allocations are not so marked, so can escape to the
3417  * nearest enclosing hardwalled ancestor cpuset.
3418  *
3419  * Scanning up parent cpusets requires callback_lock.  The
3420  * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
3421  * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
3422  * current tasks mems_allowed came up empty on the first pass over
3423  * the zonelist.  So only GFP_KERNEL allocations, if all nodes in the
3424  * cpuset are short of memory, might require taking the callback_lock.
3425  *
3426  * The first call here from mm/page_alloc:get_page_from_freelist()
3427  * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
3428  * so no allocation on a node outside the cpuset is allowed (unless
3429  * in interrupt, of course).
3430  *
3431  * The second pass through get_page_from_freelist() doesn't even call
3432  * here for GFP_ATOMIC calls.  For those calls, the __alloc_pages()
3433  * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
3434  * in alloc_flags.  That logic and the checks below have the combined
3435  * affect that:
3436  *      in_interrupt - any node ok (current task context irrelevant)
3437  *      GFP_ATOMIC   - any node ok
3438  *      tsk_is_oom_victim   - any node ok
3439  *      GFP_KERNEL   - any node in enclosing hardwalled cpuset ok
3440  *      GFP_USER     - only nodes in current tasks mems allowed ok.
3441  */
3442 bool __cpuset_node_allowed(int node, gfp_t gfp_mask)
3443 {
3444         struct cpuset *cs;              /* current cpuset ancestors */
3445         int allowed;                    /* is allocation in zone z allowed? */
3446         unsigned long flags;
3447 
3448         if (in_interrupt())
3449                 return true;
3450         if (node_isset(node, current->mems_allowed))
3451                 return true;
3452         /*
3453          * Allow tasks that have access to memory reserves because they have
3454          * been OOM killed to get memory anywhere.
3455          */
3456         if (unlikely(tsk_is_oom_victim(current)))
3457                 return true;
3458         if (gfp_mask & __GFP_HARDWALL)  /* If hardwall request, stop here */
3459                 return false;
3460 
3461         if (current->flags & PF_EXITING) /* Let dying task have memory */
3462                 return true;
3463 
3464         /* Not hardwall and node outside mems_allowed: scan up cpusets */
3465         spin_lock_irqsave(&callback_lock, flags);
3466 
3467         rcu_read_lock();
3468         cs = nearest_hardwall_ancestor(task_cs(current));
3469         allowed = node_isset(node, cs->mems_allowed);
3470         rcu_read_unlock();
3471 
3472         spin_unlock_irqrestore(&callback_lock, flags);
3473         return allowed;
3474 }
3475 
3476 /**
3477  * cpuset_mem_spread_node() - On which node to begin search for a file page
3478  * cpuset_slab_spread_node() - On which node to begin search for a slab page
3479  *
3480  * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
3481  * tasks in a cpuset with is_spread_page or is_spread_slab set),
3482  * and if the memory allocation used cpuset_mem_spread_node()
3483  * to determine on which node to start looking, as it will for
3484  * certain page cache or slab cache pages such as used for file
3485  * system buffers and inode caches, then instead of starting on the
3486  * local node to look for a free page, rather spread the starting
3487  * node around the tasks mems_allowed nodes.
3488  *
3489  * We don't have to worry about the returned node being offline
3490  * because "it can't happen", and even if it did, it would be ok.
3491  *
3492  * The routines calling guarantee_online_mems() are careful to
3493  * only set nodes in task->mems_allowed that are online.  So it
3494  * should not be possible for the following code to return an
3495  * offline node.  But if it did, that would be ok, as this routine
3496  * is not returning the node where the allocation must be, only
3497  * the node where the search should start.  The zonelist passed to
3498  * __alloc_pages() will include all nodes.  If the slab allocator
3499  * is passed an offline node, it will fall back to the local node.
3500  * See kmem_cache_alloc_node().
3501  */
3502 
3503 static int cpuset_spread_node(int *rotor)
3504 {
3505         return *rotor = next_node_in(*rotor, current->mems_allowed);
3506 }
3507 
3508 int cpuset_mem_spread_node(void)
3509 {
3510         if (current->cpuset_mem_spread_rotor == NUMA_NO_NODE)
3511                 current->cpuset_mem_spread_rotor =
3512                         node_random(&current->mems_allowed);
3513 
3514         return cpuset_spread_node(&current->cpuset_mem_spread_rotor);
3515 }
3516 
3517 int cpuset_slab_spread_node(void)
3518 {
3519         if (current->cpuset_slab_spread_rotor == NUMA_NO_NODE)
3520                 current->cpuset_slab_spread_rotor =
3521                         node_random(&current->mems_allowed);
3522 
3523         return cpuset_spread_node(&current->cpuset_slab_spread_rotor);
3524 }
3525 
3526 EXPORT_SYMBOL_GPL(cpuset_mem_spread_node);
3527 
3528 /**
3529  * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
3530  * @tsk1: pointer to task_struct of some task.
3531  * @tsk2: pointer to task_struct of some other task.
3532  *
3533  * Description: Return true if @tsk1's mems_allowed intersects the
3534  * mems_allowed of @tsk2.  Used by the OOM killer to determine if
3535  * one of the task's memory usage might impact the memory available
3536  * to the other.
3537  **/
3538 
3539 int cpuset_mems_allowed_intersects(const struct task_struct *tsk1,
3540                                    const struct task_struct *tsk2)
3541 {
3542         return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed);
3543 }
3544 
3545 /**
3546  * cpuset_print_current_mems_allowed - prints current's cpuset and mems_allowed
3547  *
3548  * Description: Prints current's name, cpuset name, and cached copy of its
3549  * mems_allowed to the kernel log.
3550  */
3551 void cpuset_print_current_mems_allowed(void)
3552 {
3553         struct cgroup *cgrp;
3554 
3555         rcu_read_lock();
3556 
3557         cgrp = task_cs(current)->css.cgroup;
3558         pr_cont(",cpuset=");
3559         pr_cont_cgroup_name(cgrp);
3560         pr_cont(",mems_allowed=%*pbl",
3561                 nodemask_pr_args(&current->mems_allowed));
3562 
3563         rcu_read_unlock();
3564 }
3565 
3566 /*
3567  * Collection of memory_pressure is suppressed unless
3568  * this flag is enabled by writing "1" to the special
3569  * cpuset file 'memory_pressure_enabled' in the root cpuset.
3570  */
3571 
3572 int cpuset_memory_pressure_enabled __read_mostly;
3573 
3574 /**
3575  * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
3576  *
3577  * Keep a running average of the rate of synchronous (direct)
3578  * page reclaim efforts initiated by tasks in each cpuset.
3579  *
3580  * This represents the rate at which some task in the cpuset
3581  * ran low on memory on all nodes it was allowed to use, and
3582  * had to enter the kernels page reclaim code in an effort to
3583  * create more free memory by tossing clean pages or swapping
3584  * or writing dirty pages.
3585  *
3586  * Display to user space in the per-cpuset read-only file
3587  * "memory_pressure".  Value displayed is an integer
3588  * representing the recent rate of entry into the synchronous
3589  * (direct) page reclaim by any task attached to the cpuset.
3590  **/
3591 
3592 void __cpuset_memory_pressure_bump(void)
3593 {
3594         rcu_read_lock();
3595         fmeter_markevent(&task_cs(current)->fmeter);
3596         rcu_read_unlock();
3597 }
3598 
3599 #ifdef CONFIG_PROC_PID_CPUSET
3600 /*
3601  * proc_cpuset_show()
3602  *  - Print tasks cpuset path into seq_file.
3603  *  - Used for /proc/<pid>/cpuset.
3604  *  - No need to task_lock(tsk) on this tsk->cpuset reference, as it
3605  *    doesn't really matter if tsk->cpuset changes after we read it,
3606  *    and we take cpuset_mutex, keeping cpuset_attach() from changing it
3607  *    anyway.
3608  */
3609 int proc_cpuset_show(struct seq_file *m, struct pid_namespace *ns,
3610                      struct pid *pid, struct task_struct *tsk)
3611 {
3612         char *buf;
3613         struct cgroup_subsys_state *css;
3614         int retval;
3615 
3616         retval = -ENOMEM;
3617         buf = kmalloc(PATH_MAX, GFP_KERNEL);
3618         if (!buf)
3619                 goto out;
3620 
3621         css = task_get_css(tsk, cpuset_cgrp_id);
3622         retval = cgroup_path_ns(css->cgroup, buf, PATH_MAX,
3623                                 current->nsproxy->cgroup_ns);
3624         css_put(css);
3625         if (retval >= PATH_MAX)
3626                 retval = -ENAMETOOLONG;
3627         if (retval < 0)
3628                 goto out_free;
3629         seq_puts(m, buf);
3630         seq_putc(m, '\n');
3631         retval = 0;
3632 out_free:
3633         kfree(buf);
3634 out:
3635         return retval;
3636 }
3637 #endif /* CONFIG_PROC_PID_CPUSET */
3638 
3639 /* Display task mems_allowed in /proc/<pid>/status file. */
3640 void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task)
3641 {
3642         seq_printf(m, "Mems_allowed:\t%*pb\n",
3643                    nodemask_pr_args(&task->mems_allowed));
3644         seq_printf(m, "Mems_allowed_list:\t%*pbl\n",
3645                    nodemask_pr_args(&task->mems_allowed));
3646 }
3647 

~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

kernel.org | git.kernel.org | LWN.net | Project Home | Wiki (Japanese) | Wiki (English) | SVN repository | Mail admin

Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.

osdn.jp