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

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

Version: ~ [ linux-5.13-rc1 ] ~ [ linux-5.12.2 ] ~ [ linux-5.11.19 ] ~ [ linux-5.10.35 ] ~ [ linux-5.9.16 ] ~ [ linux-5.8.18 ] ~ [ linux-5.7.19 ] ~ [ linux-5.6.19 ] ~ [ linux-5.5.19 ] ~ [ linux-5.4.117 ] ~ [ linux-5.3.18 ] ~ [ linux-5.2.21 ] ~ [ linux-5.1.21 ] ~ [ linux-5.0.21 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.190 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.232 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.268 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.268 ] ~ [ 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 hierachy:
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 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 hiearchy, 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 sched_domain_attr *attr;
987         cpumask_var_t *doms;
988         int ndoms;
989 
990         lockdep_assert_cpus_held();
991         percpu_rwsem_assert_held(&cpuset_rwsem);
992 
993         /*
994          * We have raced with CPU hotplug. Don't do anything to avoid
995          * passing doms with offlined cpu to partition_sched_domains().
996          * Anyways, hotplug work item will rebuild sched domains.
997          */
998         if (!top_cpuset.nr_subparts_cpus &&
999             !cpumask_equal(top_cpuset.effective_cpus, cpu_active_mask))
1000                 return;
1001 
1002         if (top_cpuset.nr_subparts_cpus &&
1003            !cpumask_subset(top_cpuset.effective_cpus, cpu_active_mask))
1004                 return;
1005 
1006         /* Generate domain masks and attrs */
1007         ndoms = generate_sched_domains(&doms, &attr);
1008 
1009         /* Have scheduler rebuild the domains */
1010         partition_and_rebuild_sched_domains(ndoms, doms, attr);
1011 }
1012 #else /* !CONFIG_SMP */
1013 static void rebuild_sched_domains_locked(void)
1014 {
1015 }
1016 #endif /* CONFIG_SMP */
1017 
1018 void rebuild_sched_domains(void)
1019 {
1020         get_online_cpus();
1021         percpu_down_write(&cpuset_rwsem);
1022         rebuild_sched_domains_locked();
1023         percpu_up_write(&cpuset_rwsem);
1024         put_online_cpus();
1025 }
1026 
1027 /**
1028  * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
1029  * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
1030  *
1031  * Iterate through each task of @cs updating its cpus_allowed to the
1032  * effective cpuset's.  As this function is called with cpuset_mutex held,
1033  * cpuset membership stays stable.
1034  */
1035 static void update_tasks_cpumask(struct cpuset *cs)
1036 {
1037         struct css_task_iter it;
1038         struct task_struct *task;
1039 
1040         css_task_iter_start(&cs->css, 0, &it);
1041         while ((task = css_task_iter_next(&it)))
1042                 set_cpus_allowed_ptr(task, cs->effective_cpus);
1043         css_task_iter_end(&it);
1044 }
1045 
1046 /**
1047  * compute_effective_cpumask - Compute the effective cpumask of the cpuset
1048  * @new_cpus: the temp variable for the new effective_cpus mask
1049  * @cs: the cpuset the need to recompute the new effective_cpus mask
1050  * @parent: the parent cpuset
1051  *
1052  * If the parent has subpartition CPUs, include them in the list of
1053  * allowable CPUs in computing the new effective_cpus mask. Since offlined
1054  * CPUs are not removed from subparts_cpus, we have to use cpu_active_mask
1055  * to mask those out.
1056  */
1057 static void compute_effective_cpumask(struct cpumask *new_cpus,
1058                                       struct cpuset *cs, struct cpuset *parent)
1059 {
1060         if (parent->nr_subparts_cpus) {
1061                 cpumask_or(new_cpus, parent->effective_cpus,
1062                            parent->subparts_cpus);
1063                 cpumask_and(new_cpus, new_cpus, cs->cpus_allowed);
1064                 cpumask_and(new_cpus, new_cpus, cpu_active_mask);
1065         } else {
1066                 cpumask_and(new_cpus, cs->cpus_allowed, parent->effective_cpus);
1067         }
1068 }
1069 
1070 /*
1071  * Commands for update_parent_subparts_cpumask
1072  */
1073 enum subparts_cmd {
1074         partcmd_enable,         /* Enable partition root         */
1075         partcmd_disable,        /* Disable partition root        */
1076         partcmd_update,         /* Update parent's subparts_cpus */
1077 };
1078 
1079 /**
1080  * update_parent_subparts_cpumask - update subparts_cpus mask of parent cpuset
1081  * @cpuset:  The cpuset that requests change in partition root state
1082  * @cmd:     Partition root state change command
1083  * @newmask: Optional new cpumask for partcmd_update
1084  * @tmp:     Temporary addmask and delmask
1085  * Return:   0, 1 or an error code
1086  *
1087  * For partcmd_enable, the cpuset is being transformed from a non-partition
1088  * root to a partition root. The cpus_allowed mask of the given cpuset will
1089  * be put into parent's subparts_cpus and taken away from parent's
1090  * effective_cpus. The function will return 0 if all the CPUs listed in
1091  * cpus_allowed can be granted or an error code will be returned.
1092  *
1093  * For partcmd_disable, the cpuset is being transofrmed from a partition
1094  * root back to a non-partition root. any CPUs in cpus_allowed that are in
1095  * parent's subparts_cpus will be taken away from that cpumask and put back
1096  * into parent's effective_cpus. 0 should always be returned.
1097  *
1098  * For partcmd_update, if the optional newmask is specified, the cpu
1099  * list is to be changed from cpus_allowed to newmask. Otherwise,
1100  * cpus_allowed is assumed to remain the same. The cpuset should either
1101  * be a partition root or an invalid partition root. The partition root
1102  * state may change if newmask is NULL and none of the requested CPUs can
1103  * be granted by the parent. The function will return 1 if changes to
1104  * parent's subparts_cpus and effective_cpus happen or 0 otherwise.
1105  * Error code should only be returned when newmask is non-NULL.
1106  *
1107  * The partcmd_enable and partcmd_disable commands are used by
1108  * update_prstate(). The partcmd_update command is used by
1109  * update_cpumasks_hier() with newmask NULL and update_cpumask() with
1110  * newmask set.
1111  *
1112  * The checking is more strict when enabling partition root than the
1113  * other two commands.
1114  *
1115  * Because of the implicit cpu exclusive nature of a partition root,
1116  * cpumask changes that violates the cpu exclusivity rule will not be
1117  * permitted when checked by validate_change(). The validate_change()
1118  * function will also prevent any changes to the cpu list if it is not
1119  * a superset of children's cpu lists.
1120  */
1121 static int update_parent_subparts_cpumask(struct cpuset *cpuset, int cmd,
1122                                           struct cpumask *newmask,
1123                                           struct tmpmasks *tmp)
1124 {
1125         struct cpuset *parent = parent_cs(cpuset);
1126         int adding;     /* Moving cpus from effective_cpus to subparts_cpus */
1127         int deleting;   /* Moving cpus from subparts_cpus to effective_cpus */
1128         bool part_error = false;        /* Partition error? */
1129 
1130         percpu_rwsem_assert_held(&cpuset_rwsem);
1131 
1132         /*
1133          * The parent must be a partition root.
1134          * The new cpumask, if present, or the current cpus_allowed must
1135          * not be empty.
1136          */
1137         if (!is_partition_root(parent) ||
1138            (newmask && cpumask_empty(newmask)) ||
1139            (!newmask && cpumask_empty(cpuset->cpus_allowed)))
1140                 return -EINVAL;
1141 
1142         /*
1143          * Enabling/disabling partition root is not allowed if there are
1144          * online children.
1145          */
1146         if ((cmd != partcmd_update) && css_has_online_children(&cpuset->css))
1147                 return -EBUSY;
1148 
1149         /*
1150          * Enabling partition root is not allowed if not all the CPUs
1151          * can be granted from parent's effective_cpus or at least one
1152          * CPU will be left after that.
1153          */
1154         if ((cmd == partcmd_enable) &&
1155            (!cpumask_subset(cpuset->cpus_allowed, parent->effective_cpus) ||
1156              cpumask_equal(cpuset->cpus_allowed, parent->effective_cpus)))
1157                 return -EINVAL;
1158 
1159         /*
1160          * A cpumask update cannot make parent's effective_cpus become empty.
1161          */
1162         adding = deleting = false;
1163         if (cmd == partcmd_enable) {
1164                 cpumask_copy(tmp->addmask, cpuset->cpus_allowed);
1165                 adding = true;
1166         } else if (cmd == partcmd_disable) {
1167                 deleting = cpumask_and(tmp->delmask, cpuset->cpus_allowed,
1168                                        parent->subparts_cpus);
1169         } else if (newmask) {
1170                 /*
1171                  * partcmd_update with newmask:
1172                  *
1173                  * delmask = cpus_allowed & ~newmask & parent->subparts_cpus
1174                  * addmask = newmask & parent->effective_cpus
1175                  *                   & ~parent->subparts_cpus
1176                  */
1177                 cpumask_andnot(tmp->delmask, cpuset->cpus_allowed, newmask);
1178                 deleting = cpumask_and(tmp->delmask, tmp->delmask,
1179                                        parent->subparts_cpus);
1180 
1181                 cpumask_and(tmp->addmask, newmask, parent->effective_cpus);
1182                 adding = cpumask_andnot(tmp->addmask, tmp->addmask,
1183                                         parent->subparts_cpus);
1184                 /*
1185                  * Return error if the new effective_cpus could become empty.
1186                  */
1187                 if (adding &&
1188                     cpumask_equal(parent->effective_cpus, tmp->addmask)) {
1189                         if (!deleting)
1190                                 return -EINVAL;
1191                         /*
1192                          * As some of the CPUs in subparts_cpus might have
1193                          * been offlined, we need to compute the real delmask
1194                          * to confirm that.
1195                          */
1196                         if (!cpumask_and(tmp->addmask, tmp->delmask,
1197                                          cpu_active_mask))
1198                                 return -EINVAL;
1199                         cpumask_copy(tmp->addmask, parent->effective_cpus);
1200                 }
1201         } else {
1202                 /*
1203                  * partcmd_update w/o newmask:
1204                  *
1205                  * addmask = cpus_allowed & parent->effectiveb_cpus
1206                  *
1207                  * Note that parent's subparts_cpus may have been
1208                  * pre-shrunk in case there is a change in the cpu list.
1209                  * So no deletion is needed.
1210                  */
1211                 adding = cpumask_and(tmp->addmask, cpuset->cpus_allowed,
1212                                      parent->effective_cpus);
1213                 part_error = cpumask_equal(tmp->addmask,
1214                                            parent->effective_cpus);
1215         }
1216 
1217         if (cmd == partcmd_update) {
1218                 int prev_prs = cpuset->partition_root_state;
1219 
1220                 /*
1221                  * Check for possible transition between PRS_ENABLED
1222                  * and PRS_ERROR.
1223                  */
1224                 switch (cpuset->partition_root_state) {
1225                 case PRS_ENABLED:
1226                         if (part_error)
1227                                 cpuset->partition_root_state = PRS_ERROR;
1228                         break;
1229                 case PRS_ERROR:
1230                         if (!part_error)
1231                                 cpuset->partition_root_state = PRS_ENABLED;
1232                         break;
1233                 }
1234                 /*
1235                  * Set part_error if previously in invalid state.
1236                  */
1237                 part_error = (prev_prs == PRS_ERROR);
1238         }
1239 
1240         if (!part_error && (cpuset->partition_root_state == PRS_ERROR))
1241                 return 0;       /* Nothing need to be done */
1242 
1243         if (cpuset->partition_root_state == PRS_ERROR) {
1244                 /*
1245                  * Remove all its cpus from parent's subparts_cpus.
1246                  */
1247                 adding = false;
1248                 deleting = cpumask_and(tmp->delmask, cpuset->cpus_allowed,
1249                                        parent->subparts_cpus);
1250         }
1251 
1252         if (!adding && !deleting)
1253                 return 0;
1254 
1255         /*
1256          * Change the parent's subparts_cpus.
1257          * Newly added CPUs will be removed from effective_cpus and
1258          * newly deleted ones will be added back to effective_cpus.
1259          */
1260         spin_lock_irq(&callback_lock);
1261         if (adding) {
1262                 cpumask_or(parent->subparts_cpus,
1263                            parent->subparts_cpus, tmp->addmask);
1264                 cpumask_andnot(parent->effective_cpus,
1265                                parent->effective_cpus, tmp->addmask);
1266         }
1267         if (deleting) {
1268                 cpumask_andnot(parent->subparts_cpus,
1269                                parent->subparts_cpus, tmp->delmask);
1270                 /*
1271                  * Some of the CPUs in subparts_cpus might have been offlined.
1272                  */
1273                 cpumask_and(tmp->delmask, tmp->delmask, cpu_active_mask);
1274                 cpumask_or(parent->effective_cpus,
1275                            parent->effective_cpus, tmp->delmask);
1276         }
1277 
1278         parent->nr_subparts_cpus = cpumask_weight(parent->subparts_cpus);
1279         spin_unlock_irq(&callback_lock);
1280 
1281         return cmd == partcmd_update;
1282 }
1283 
1284 /*
1285  * update_cpumasks_hier - Update effective cpumasks and tasks in the subtree
1286  * @cs:  the cpuset to consider
1287  * @tmp: temp variables for calculating effective_cpus & partition setup
1288  *
1289  * When congifured cpumask is changed, the effective cpumasks of this cpuset
1290  * and all its descendants need to be updated.
1291  *
1292  * On legacy hierachy, effective_cpus will be the same with cpu_allowed.
1293  *
1294  * Called with cpuset_mutex held
1295  */
1296 static void update_cpumasks_hier(struct cpuset *cs, struct tmpmasks *tmp)
1297 {
1298         struct cpuset *cp;
1299         struct cgroup_subsys_state *pos_css;
1300         bool need_rebuild_sched_domains = false;
1301 
1302         rcu_read_lock();
1303         cpuset_for_each_descendant_pre(cp, pos_css, cs) {
1304                 struct cpuset *parent = parent_cs(cp);
1305 
1306                 compute_effective_cpumask(tmp->new_cpus, cp, parent);
1307 
1308                 /*
1309                  * If it becomes empty, inherit the effective mask of the
1310                  * parent, which is guaranteed to have some CPUs.
1311                  */
1312                 if (is_in_v2_mode() && cpumask_empty(tmp->new_cpus)) {
1313                         cpumask_copy(tmp->new_cpus, parent->effective_cpus);
1314                         if (!cp->use_parent_ecpus) {
1315                                 cp->use_parent_ecpus = true;
1316                                 parent->child_ecpus_count++;
1317                         }
1318                 } else if (cp->use_parent_ecpus) {
1319                         cp->use_parent_ecpus = false;
1320                         WARN_ON_ONCE(!parent->child_ecpus_count);
1321                         parent->child_ecpus_count--;
1322                 }
1323 
1324                 /*
1325                  * Skip the whole subtree if the cpumask remains the same
1326                  * and has no partition root state.
1327                  */
1328                 if (!cp->partition_root_state &&
1329                     cpumask_equal(tmp->new_cpus, cp->effective_cpus)) {
1330                         pos_css = css_rightmost_descendant(pos_css);
1331                         continue;
1332                 }
1333 
1334                 /*
1335                  * update_parent_subparts_cpumask() should have been called
1336                  * for cs already in update_cpumask(). We should also call
1337                  * update_tasks_cpumask() again for tasks in the parent
1338                  * cpuset if the parent's subparts_cpus changes.
1339                  */
1340                 if ((cp != cs) && cp->partition_root_state) {
1341                         switch (parent->partition_root_state) {
1342                         case PRS_DISABLED:
1343                                 /*
1344                                  * If parent is not a partition root or an
1345                                  * invalid partition root, clear the state
1346                                  * state and the CS_CPU_EXCLUSIVE flag.
1347                                  */
1348                                 WARN_ON_ONCE(cp->partition_root_state
1349                                              != PRS_ERROR);
1350                                 cp->partition_root_state = 0;
1351 
1352                                 /*
1353                                  * clear_bit() is an atomic operation and
1354                                  * readers aren't interested in the state
1355                                  * of CS_CPU_EXCLUSIVE anyway. So we can
1356                                  * just update the flag without holding
1357                                  * the callback_lock.
1358                                  */
1359                                 clear_bit(CS_CPU_EXCLUSIVE, &cp->flags);
1360                                 break;
1361 
1362                         case PRS_ENABLED:
1363                                 if (update_parent_subparts_cpumask(cp, partcmd_update, NULL, tmp))
1364                                         update_tasks_cpumask(parent);
1365                                 break;
1366 
1367                         case PRS_ERROR:
1368                                 /*
1369                                  * When parent is invalid, it has to be too.
1370                                  */
1371                                 cp->partition_root_state = PRS_ERROR;
1372                                 if (cp->nr_subparts_cpus) {
1373                                         cp->nr_subparts_cpus = 0;
1374                                         cpumask_clear(cp->subparts_cpus);
1375                                 }
1376                                 break;
1377                         }
1378                 }
1379 
1380                 if (!css_tryget_online(&cp->css))
1381                         continue;
1382                 rcu_read_unlock();
1383 
1384                 spin_lock_irq(&callback_lock);
1385 
1386                 cpumask_copy(cp->effective_cpus, tmp->new_cpus);
1387                 if (cp->nr_subparts_cpus &&
1388                    (cp->partition_root_state != PRS_ENABLED)) {
1389                         cp->nr_subparts_cpus = 0;
1390                         cpumask_clear(cp->subparts_cpus);
1391                 } else if (cp->nr_subparts_cpus) {
1392                         /*
1393                          * Make sure that effective_cpus & subparts_cpus
1394                          * are mutually exclusive.
1395                          *
1396                          * In the unlikely event that effective_cpus
1397                          * becomes empty. we clear cp->nr_subparts_cpus and
1398                          * let its child partition roots to compete for
1399                          * CPUs again.
1400                          */
1401                         cpumask_andnot(cp->effective_cpus, cp->effective_cpus,
1402                                        cp->subparts_cpus);
1403                         if (cpumask_empty(cp->effective_cpus)) {
1404                                 cpumask_copy(cp->effective_cpus, tmp->new_cpus);
1405                                 cpumask_clear(cp->subparts_cpus);
1406                                 cp->nr_subparts_cpus = 0;
1407                         } else if (!cpumask_subset(cp->subparts_cpus,
1408                                                    tmp->new_cpus)) {
1409                                 cpumask_andnot(cp->subparts_cpus,
1410                                         cp->subparts_cpus, tmp->new_cpus);
1411                                 cp->nr_subparts_cpus
1412                                         = cpumask_weight(cp->subparts_cpus);
1413                         }
1414                 }
1415                 spin_unlock_irq(&callback_lock);
1416 
1417                 WARN_ON(!is_in_v2_mode() &&
1418                         !cpumask_equal(cp->cpus_allowed, cp->effective_cpus));
1419 
1420                 update_tasks_cpumask(cp);
1421 
1422                 /*
1423                  * On legacy hierarchy, if the effective cpumask of any non-
1424                  * empty cpuset is changed, we need to rebuild sched domains.
1425                  * On default hierarchy, the cpuset needs to be a partition
1426                  * root as well.
1427                  */
1428                 if (!cpumask_empty(cp->cpus_allowed) &&
1429                     is_sched_load_balance(cp) &&
1430                    (!cgroup_subsys_on_dfl(cpuset_cgrp_subsys) ||
1431                     is_partition_root(cp)))
1432                         need_rebuild_sched_domains = true;
1433 
1434                 rcu_read_lock();
1435                 css_put(&cp->css);
1436         }
1437         rcu_read_unlock();
1438 
1439         if (need_rebuild_sched_domains)
1440                 rebuild_sched_domains_locked();
1441 }
1442 
1443 /**
1444  * update_sibling_cpumasks - Update siblings cpumasks
1445  * @parent:  Parent cpuset
1446  * @cs:      Current cpuset
1447  * @tmp:     Temp variables
1448  */
1449 static void update_sibling_cpumasks(struct cpuset *parent, struct cpuset *cs,
1450                                     struct tmpmasks *tmp)
1451 {
1452         struct cpuset *sibling;
1453         struct cgroup_subsys_state *pos_css;
1454 
1455         /*
1456          * Check all its siblings and call update_cpumasks_hier()
1457          * if their use_parent_ecpus flag is set in order for them
1458          * to use the right effective_cpus value.
1459          */
1460         rcu_read_lock();
1461         cpuset_for_each_child(sibling, pos_css, parent) {
1462                 if (sibling == cs)
1463                         continue;
1464                 if (!sibling->use_parent_ecpus)
1465                         continue;
1466 
1467                 update_cpumasks_hier(sibling, tmp);
1468         }
1469         rcu_read_unlock();
1470 }
1471 
1472 /**
1473  * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
1474  * @cs: the cpuset to consider
1475  * @trialcs: trial cpuset
1476  * @buf: buffer of cpu numbers written to this cpuset
1477  */
1478 static int update_cpumask(struct cpuset *cs, struct cpuset *trialcs,
1479                           const char *buf)
1480 {
1481         int retval;
1482         struct tmpmasks tmp;
1483 
1484         /* top_cpuset.cpus_allowed tracks cpu_online_mask; it's read-only */
1485         if (cs == &top_cpuset)
1486                 return -EACCES;
1487 
1488         /*
1489          * An empty cpus_allowed is ok only if the cpuset has no tasks.
1490          * Since cpulist_parse() fails on an empty mask, we special case
1491          * that parsing.  The validate_change() call ensures that cpusets
1492          * with tasks have cpus.
1493          */
1494         if (!*buf) {
1495                 cpumask_clear(trialcs->cpus_allowed);
1496         } else {
1497                 retval = cpulist_parse(buf, trialcs->cpus_allowed);
1498                 if (retval < 0)
1499                         return retval;
1500 
1501                 if (!cpumask_subset(trialcs->cpus_allowed,
1502                                     top_cpuset.cpus_allowed))
1503                         return -EINVAL;
1504         }
1505 
1506         /* Nothing to do if the cpus didn't change */
1507         if (cpumask_equal(cs->cpus_allowed, trialcs->cpus_allowed))
1508                 return 0;
1509 
1510         retval = validate_change(cs, trialcs);
1511         if (retval < 0)
1512                 return retval;
1513 
1514 #ifdef CONFIG_CPUMASK_OFFSTACK
1515         /*
1516          * Use the cpumasks in trialcs for tmpmasks when they are pointers
1517          * to allocated cpumasks.
1518          */
1519         tmp.addmask  = trialcs->subparts_cpus;
1520         tmp.delmask  = trialcs->effective_cpus;
1521         tmp.new_cpus = trialcs->cpus_allowed;
1522 #endif
1523 
1524         if (cs->partition_root_state) {
1525                 /* Cpumask of a partition root cannot be empty */
1526                 if (cpumask_empty(trialcs->cpus_allowed))
1527                         return -EINVAL;
1528                 if (update_parent_subparts_cpumask(cs, partcmd_update,
1529                                         trialcs->cpus_allowed, &tmp) < 0)
1530                         return -EINVAL;
1531         }
1532 
1533         spin_lock_irq(&callback_lock);
1534         cpumask_copy(cs->cpus_allowed, trialcs->cpus_allowed);
1535 
1536         /*
1537          * Make sure that subparts_cpus is a subset of cpus_allowed.
1538          */
1539         if (cs->nr_subparts_cpus) {
1540                 cpumask_andnot(cs->subparts_cpus, cs->subparts_cpus,
1541                                cs->cpus_allowed);
1542                 cs->nr_subparts_cpus = cpumask_weight(cs->subparts_cpus);
1543         }
1544         spin_unlock_irq(&callback_lock);
1545 
1546         update_cpumasks_hier(cs, &tmp);
1547 
1548         if (cs->partition_root_state) {
1549                 struct cpuset *parent = parent_cs(cs);
1550 
1551                 /*
1552                  * For partition root, update the cpumasks of sibling
1553                  * cpusets if they use parent's effective_cpus.
1554                  */
1555                 if (parent->child_ecpus_count)
1556                         update_sibling_cpumasks(parent, cs, &tmp);
1557         }
1558         return 0;
1559 }
1560 
1561 /*
1562  * Migrate memory region from one set of nodes to another.  This is
1563  * performed asynchronously as it can be called from process migration path
1564  * holding locks involved in process management.  All mm migrations are
1565  * performed in the queued order and can be waited for by flushing
1566  * cpuset_migrate_mm_wq.
1567  */
1568 
1569 struct cpuset_migrate_mm_work {
1570         struct work_struct      work;
1571         struct mm_struct        *mm;
1572         nodemask_t              from;
1573         nodemask_t              to;
1574 };
1575 
1576 static void cpuset_migrate_mm_workfn(struct work_struct *work)
1577 {
1578         struct cpuset_migrate_mm_work *mwork =
1579                 container_of(work, struct cpuset_migrate_mm_work, work);
1580 
1581         /* on a wq worker, no need to worry about %current's mems_allowed */
1582         do_migrate_pages(mwork->mm, &mwork->from, &mwork->to, MPOL_MF_MOVE_ALL);
1583         mmput(mwork->mm);
1584         kfree(mwork);
1585 }
1586 
1587 static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
1588                                                         const nodemask_t *to)
1589 {
1590         struct cpuset_migrate_mm_work *mwork;
1591 
1592         mwork = kzalloc(sizeof(*mwork), GFP_KERNEL);
1593         if (mwork) {
1594                 mwork->mm = mm;
1595                 mwork->from = *from;
1596                 mwork->to = *to;
1597                 INIT_WORK(&mwork->work, cpuset_migrate_mm_workfn);
1598                 queue_work(cpuset_migrate_mm_wq, &mwork->work);
1599         } else {
1600                 mmput(mm);
1601         }
1602 }
1603 
1604 static void cpuset_post_attach(void)
1605 {
1606         flush_workqueue(cpuset_migrate_mm_wq);
1607 }
1608 
1609 /*
1610  * cpuset_change_task_nodemask - change task's mems_allowed and mempolicy
1611  * @tsk: the task to change
1612  * @newmems: new nodes that the task will be set
1613  *
1614  * We use the mems_allowed_seq seqlock to safely update both tsk->mems_allowed
1615  * and rebind an eventual tasks' mempolicy. If the task is allocating in
1616  * parallel, it might temporarily see an empty intersection, which results in
1617  * a seqlock check and retry before OOM or allocation failure.
1618  */
1619 static void cpuset_change_task_nodemask(struct task_struct *tsk,
1620                                         nodemask_t *newmems)
1621 {
1622         task_lock(tsk);
1623 
1624         local_irq_disable();
1625         write_seqcount_begin(&tsk->mems_allowed_seq);
1626 
1627         nodes_or(tsk->mems_allowed, tsk->mems_allowed, *newmems);
1628         mpol_rebind_task(tsk, newmems);
1629         tsk->mems_allowed = *newmems;
1630 
1631         write_seqcount_end(&tsk->mems_allowed_seq);
1632         local_irq_enable();
1633 
1634         task_unlock(tsk);
1635 }
1636 
1637 static void *cpuset_being_rebound;
1638 
1639 /**
1640  * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
1641  * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
1642  *
1643  * Iterate through each task of @cs updating its mems_allowed to the
1644  * effective cpuset's.  As this function is called with cpuset_mutex held,
1645  * cpuset membership stays stable.
1646  */
1647 static void update_tasks_nodemask(struct cpuset *cs)
1648 {
1649         static nodemask_t newmems;      /* protected by cpuset_mutex */
1650         struct css_task_iter it;
1651         struct task_struct *task;
1652 
1653         cpuset_being_rebound = cs;              /* causes mpol_dup() rebind */
1654 
1655         guarantee_online_mems(cs, &newmems);
1656 
1657         /*
1658          * The mpol_rebind_mm() call takes mmap_sem, which we couldn't
1659          * take while holding tasklist_lock.  Forks can happen - the
1660          * mpol_dup() cpuset_being_rebound check will catch such forks,
1661          * and rebind their vma mempolicies too.  Because we still hold
1662          * the global cpuset_mutex, we know that no other rebind effort
1663          * will be contending for the global variable cpuset_being_rebound.
1664          * It's ok if we rebind the same mm twice; mpol_rebind_mm()
1665          * is idempotent.  Also migrate pages in each mm to new nodes.
1666          */
1667         css_task_iter_start(&cs->css, 0, &it);
1668         while ((task = css_task_iter_next(&it))) {
1669                 struct mm_struct *mm;
1670                 bool migrate;
1671 
1672                 cpuset_change_task_nodemask(task, &newmems);
1673 
1674                 mm = get_task_mm(task);
1675                 if (!mm)
1676                         continue;
1677 
1678                 migrate = is_memory_migrate(cs);
1679 
1680                 mpol_rebind_mm(mm, &cs->mems_allowed);
1681                 if (migrate)
1682                         cpuset_migrate_mm(mm, &cs->old_mems_allowed, &newmems);
1683                 else
1684                         mmput(mm);
1685         }
1686         css_task_iter_end(&it);
1687 
1688         /*
1689          * All the tasks' nodemasks have been updated, update
1690          * cs->old_mems_allowed.
1691          */
1692         cs->old_mems_allowed = newmems;
1693 
1694         /* We're done rebinding vmas to this cpuset's new mems_allowed. */
1695         cpuset_being_rebound = NULL;
1696 }
1697 
1698 /*
1699  * update_nodemasks_hier - Update effective nodemasks and tasks in the subtree
1700  * @cs: the cpuset to consider
1701  * @new_mems: a temp variable for calculating new effective_mems
1702  *
1703  * When configured nodemask is changed, the effective nodemasks of this cpuset
1704  * and all its descendants need to be updated.
1705  *
1706  * On legacy hiearchy, effective_mems will be the same with mems_allowed.
1707  *
1708  * Called with cpuset_mutex held
1709  */
1710 static void update_nodemasks_hier(struct cpuset *cs, nodemask_t *new_mems)
1711 {
1712         struct cpuset *cp;
1713         struct cgroup_subsys_state *pos_css;
1714 
1715         rcu_read_lock();
1716         cpuset_for_each_descendant_pre(cp, pos_css, cs) {
1717                 struct cpuset *parent = parent_cs(cp);
1718 
1719                 nodes_and(*new_mems, cp->mems_allowed, parent->effective_mems);
1720 
1721                 /*
1722                  * If it becomes empty, inherit the effective mask of the
1723                  * parent, which is guaranteed to have some MEMs.
1724                  */
1725                 if (is_in_v2_mode() && nodes_empty(*new_mems))
1726                         *new_mems = parent->effective_mems;
1727 
1728                 /* Skip the whole subtree if the nodemask remains the same. */
1729                 if (nodes_equal(*new_mems, cp->effective_mems)) {
1730                         pos_css = css_rightmost_descendant(pos_css);
1731                         continue;
1732                 }
1733 
1734                 if (!css_tryget_online(&cp->css))
1735                         continue;
1736                 rcu_read_unlock();
1737 
1738                 spin_lock_irq(&callback_lock);
1739                 cp->effective_mems = *new_mems;
1740                 spin_unlock_irq(&callback_lock);
1741 
1742                 WARN_ON(!is_in_v2_mode() &&
1743                         !nodes_equal(cp->mems_allowed, cp->effective_mems));
1744 
1745                 update_tasks_nodemask(cp);
1746 
1747                 rcu_read_lock();
1748                 css_put(&cp->css);
1749         }
1750         rcu_read_unlock();
1751 }
1752 
1753 /*
1754  * Handle user request to change the 'mems' memory placement
1755  * of a cpuset.  Needs to validate the request, update the
1756  * cpusets mems_allowed, and for each task in the cpuset,
1757  * update mems_allowed and rebind task's mempolicy and any vma
1758  * mempolicies and if the cpuset is marked 'memory_migrate',
1759  * migrate the tasks pages to the new memory.
1760  *
1761  * Call with cpuset_mutex held. May take callback_lock during call.
1762  * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
1763  * lock each such tasks mm->mmap_sem, scan its vma's and rebind
1764  * their mempolicies to the cpusets new mems_allowed.
1765  */
1766 static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
1767                            const char *buf)
1768 {
1769         int retval;
1770 
1771         /*
1772          * top_cpuset.mems_allowed tracks node_stats[N_MEMORY];
1773          * it's read-only
1774          */
1775         if (cs == &top_cpuset) {
1776                 retval = -EACCES;
1777                 goto done;
1778         }
1779 
1780         /*
1781          * An empty mems_allowed is ok iff there are no tasks in the cpuset.
1782          * Since nodelist_parse() fails on an empty mask, we special case
1783          * that parsing.  The validate_change() call ensures that cpusets
1784          * with tasks have memory.
1785          */
1786         if (!*buf) {
1787                 nodes_clear(trialcs->mems_allowed);
1788         } else {
1789                 retval = nodelist_parse(buf, trialcs->mems_allowed);
1790                 if (retval < 0)
1791                         goto done;
1792 
1793                 if (!nodes_subset(trialcs->mems_allowed,
1794                                   top_cpuset.mems_allowed)) {
1795                         retval = -EINVAL;
1796                         goto done;
1797                 }
1798         }
1799 
1800         if (nodes_equal(cs->mems_allowed, trialcs->mems_allowed)) {
1801                 retval = 0;             /* Too easy - nothing to do */
1802                 goto done;
1803         }
1804         retval = validate_change(cs, trialcs);
1805         if (retval < 0)
1806                 goto done;
1807 
1808         spin_lock_irq(&callback_lock);
1809         cs->mems_allowed = trialcs->mems_allowed;
1810         spin_unlock_irq(&callback_lock);
1811 
1812         /* use trialcs->mems_allowed as a temp variable */
1813         update_nodemasks_hier(cs, &trialcs->mems_allowed);
1814 done:
1815         return retval;
1816 }
1817 
1818 bool current_cpuset_is_being_rebound(void)
1819 {
1820         bool ret;
1821 
1822         rcu_read_lock();
1823         ret = task_cs(current) == cpuset_being_rebound;
1824         rcu_read_unlock();
1825 
1826         return ret;
1827 }
1828 
1829 static int update_relax_domain_level(struct cpuset *cs, s64 val)
1830 {
1831 #ifdef CONFIG_SMP
1832         if (val < -1 || val >= sched_domain_level_max)
1833                 return -EINVAL;
1834 #endif
1835 
1836         if (val != cs->relax_domain_level) {
1837                 cs->relax_domain_level = val;
1838                 if (!cpumask_empty(cs->cpus_allowed) &&
1839                     is_sched_load_balance(cs))
1840                         rebuild_sched_domains_locked();
1841         }
1842 
1843         return 0;
1844 }
1845 
1846 /**
1847  * update_tasks_flags - update the spread flags of tasks in the cpuset.
1848  * @cs: the cpuset in which each task's spread flags needs to be changed
1849  *
1850  * Iterate through each task of @cs updating its spread flags.  As this
1851  * function is called with cpuset_mutex held, cpuset membership stays
1852  * stable.
1853  */
1854 static void update_tasks_flags(struct cpuset *cs)
1855 {
1856         struct css_task_iter it;
1857         struct task_struct *task;
1858 
1859         css_task_iter_start(&cs->css, 0, &it);
1860         while ((task = css_task_iter_next(&it)))
1861                 cpuset_update_task_spread_flag(cs, task);
1862         css_task_iter_end(&it);
1863 }
1864 
1865 /*
1866  * update_flag - read a 0 or a 1 in a file and update associated flag
1867  * bit:         the bit to update (see cpuset_flagbits_t)
1868  * cs:          the cpuset to update
1869  * turning_on:  whether the flag is being set or cleared
1870  *
1871  * Call with cpuset_mutex held.
1872  */
1873 
1874 static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs,
1875                        int turning_on)
1876 {
1877         struct cpuset *trialcs;
1878         int balance_flag_changed;
1879         int spread_flag_changed;
1880         int err;
1881 
1882         trialcs = alloc_trial_cpuset(cs);
1883         if (!trialcs)
1884                 return -ENOMEM;
1885 
1886         if (turning_on)
1887                 set_bit(bit, &trialcs->flags);
1888         else
1889                 clear_bit(bit, &trialcs->flags);
1890 
1891         err = validate_change(cs, trialcs);
1892         if (err < 0)
1893                 goto out;
1894 
1895         balance_flag_changed = (is_sched_load_balance(cs) !=
1896                                 is_sched_load_balance(trialcs));
1897 
1898         spread_flag_changed = ((is_spread_slab(cs) != is_spread_slab(trialcs))
1899                         || (is_spread_page(cs) != is_spread_page(trialcs)));
1900 
1901         spin_lock_irq(&callback_lock);
1902         cs->flags = trialcs->flags;
1903         spin_unlock_irq(&callback_lock);
1904 
1905         if (!cpumask_empty(trialcs->cpus_allowed) && balance_flag_changed)
1906                 rebuild_sched_domains_locked();
1907 
1908         if (spread_flag_changed)
1909                 update_tasks_flags(cs);
1910 out:
1911         free_cpuset(trialcs);
1912         return err;
1913 }
1914 
1915 /*
1916  * update_prstate - update partititon_root_state
1917  * cs:  the cpuset to update
1918  * val: 0 - disabled, 1 - enabled
1919  *
1920  * Call with cpuset_mutex held.
1921  */
1922 static int update_prstate(struct cpuset *cs, int val)
1923 {
1924         int err;
1925         struct cpuset *parent = parent_cs(cs);
1926         struct tmpmasks tmp;
1927 
1928         if ((val != 0) && (val != 1))
1929                 return -EINVAL;
1930         if (val == cs->partition_root_state)
1931                 return 0;
1932 
1933         /*
1934          * Cannot force a partial or invalid partition root to a full
1935          * partition root.
1936          */
1937         if (val && cs->partition_root_state)
1938                 return -EINVAL;
1939 
1940         if (alloc_cpumasks(NULL, &tmp))
1941                 return -ENOMEM;
1942 
1943         err = -EINVAL;
1944         if (!cs->partition_root_state) {
1945                 /*
1946                  * Turning on partition root requires setting the
1947                  * CS_CPU_EXCLUSIVE bit implicitly as well and cpus_allowed
1948                  * cannot be NULL.
1949                  */
1950                 if (cpumask_empty(cs->cpus_allowed))
1951                         goto out;
1952 
1953                 err = update_flag(CS_CPU_EXCLUSIVE, cs, 1);
1954                 if (err)
1955                         goto out;
1956 
1957                 err = update_parent_subparts_cpumask(cs, partcmd_enable,
1958                                                      NULL, &tmp);
1959                 if (err) {
1960                         update_flag(CS_CPU_EXCLUSIVE, cs, 0);
1961                         goto out;
1962                 }
1963                 cs->partition_root_state = PRS_ENABLED;
1964         } else {
1965                 /*
1966                  * Turning off partition root will clear the
1967                  * CS_CPU_EXCLUSIVE bit.
1968                  */
1969                 if (cs->partition_root_state == PRS_ERROR) {
1970                         cs->partition_root_state = 0;
1971                         update_flag(CS_CPU_EXCLUSIVE, cs, 0);
1972                         err = 0;
1973                         goto out;
1974                 }
1975 
1976                 err = update_parent_subparts_cpumask(cs, partcmd_disable,
1977                                                      NULL, &tmp);
1978                 if (err)
1979                         goto out;
1980 
1981                 cs->partition_root_state = 0;
1982 
1983                 /* Turning off CS_CPU_EXCLUSIVE will not return error */
1984                 update_flag(CS_CPU_EXCLUSIVE, cs, 0);
1985         }
1986 
1987         /*
1988          * Update cpumask of parent's tasks except when it is the top
1989          * cpuset as some system daemons cannot be mapped to other CPUs.
1990          */
1991         if (parent != &top_cpuset)
1992                 update_tasks_cpumask(parent);
1993 
1994         if (parent->child_ecpus_count)
1995                 update_sibling_cpumasks(parent, cs, &tmp);
1996 
1997         rebuild_sched_domains_locked();
1998 out:
1999         free_cpumasks(NULL, &tmp);
2000         return err;
2001 }
2002 
2003 /*
2004  * Frequency meter - How fast is some event occurring?
2005  *
2006  * These routines manage a digitally filtered, constant time based,
2007  * event frequency meter.  There are four routines:
2008  *   fmeter_init() - initialize a frequency meter.
2009  *   fmeter_markevent() - called each time the event happens.
2010  *   fmeter_getrate() - returns the recent rate of such events.
2011  *   fmeter_update() - internal routine used to update fmeter.
2012  *
2013  * A common data structure is passed to each of these routines,
2014  * which is used to keep track of the state required to manage the
2015  * frequency meter and its digital filter.
2016  *
2017  * The filter works on the number of events marked per unit time.
2018  * The filter is single-pole low-pass recursive (IIR).  The time unit
2019  * is 1 second.  Arithmetic is done using 32-bit integers scaled to
2020  * simulate 3 decimal digits of precision (multiplied by 1000).
2021  *
2022  * With an FM_COEF of 933, and a time base of 1 second, the filter
2023  * has a half-life of 10 seconds, meaning that if the events quit
2024  * happening, then the rate returned from the fmeter_getrate()
2025  * will be cut in half each 10 seconds, until it converges to zero.
2026  *
2027  * It is not worth doing a real infinitely recursive filter.  If more
2028  * than FM_MAXTICKS ticks have elapsed since the last filter event,
2029  * just compute FM_MAXTICKS ticks worth, by which point the level
2030  * will be stable.
2031  *
2032  * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
2033  * arithmetic overflow in the fmeter_update() routine.
2034  *
2035  * Given the simple 32 bit integer arithmetic used, this meter works
2036  * best for reporting rates between one per millisecond (msec) and
2037  * one per 32 (approx) seconds.  At constant rates faster than one
2038  * per msec it maxes out at values just under 1,000,000.  At constant
2039  * rates between one per msec, and one per second it will stabilize
2040  * to a value N*1000, where N is the rate of events per second.
2041  * At constant rates between one per second and one per 32 seconds,
2042  * it will be choppy, moving up on the seconds that have an event,
2043  * and then decaying until the next event.  At rates slower than
2044  * about one in 32 seconds, it decays all the way back to zero between
2045  * each event.
2046  */
2047 
2048 #define FM_COEF 933             /* coefficient for half-life of 10 secs */
2049 #define FM_MAXTICKS ((u32)99)   /* useless computing more ticks than this */
2050 #define FM_MAXCNT 1000000       /* limit cnt to avoid overflow */
2051 #define FM_SCALE 1000           /* faux fixed point scale */
2052 
2053 /* Initialize a frequency meter */
2054 static void fmeter_init(struct fmeter *fmp)
2055 {
2056         fmp->cnt = 0;
2057         fmp->val = 0;
2058         fmp->time = 0;
2059         spin_lock_init(&fmp->lock);
2060 }
2061 
2062 /* Internal meter update - process cnt events and update value */
2063 static void fmeter_update(struct fmeter *fmp)
2064 {
2065         time64_t now;
2066         u32 ticks;
2067 
2068         now = ktime_get_seconds();
2069         ticks = now - fmp->time;
2070 
2071         if (ticks == 0)
2072                 return;
2073 
2074         ticks = min(FM_MAXTICKS, ticks);
2075         while (ticks-- > 0)
2076                 fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
2077         fmp->time = now;
2078 
2079         fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
2080         fmp->cnt = 0;
2081 }
2082 
2083 /* Process any previous ticks, then bump cnt by one (times scale). */
2084 static void fmeter_markevent(struct fmeter *fmp)
2085 {
2086         spin_lock(&fmp->lock);
2087         fmeter_update(fmp);
2088         fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
2089         spin_unlock(&fmp->lock);
2090 }
2091 
2092 /* Process any previous ticks, then return current value. */
2093 static int fmeter_getrate(struct fmeter *fmp)
2094 {
2095         int val;
2096 
2097         spin_lock(&fmp->lock);
2098         fmeter_update(fmp);
2099         val = fmp->val;
2100         spin_unlock(&fmp->lock);
2101         return val;
2102 }
2103 
2104 static struct cpuset *cpuset_attach_old_cs;
2105 
2106 /* Called by cgroups to determine if a cpuset is usable; cpuset_mutex held */
2107 static int cpuset_can_attach(struct cgroup_taskset *tset)
2108 {
2109         struct cgroup_subsys_state *css;
2110         struct cpuset *cs;
2111         struct task_struct *task;
2112         int ret;
2113 
2114         /* used later by cpuset_attach() */
2115         cpuset_attach_old_cs = task_cs(cgroup_taskset_first(tset, &css));
2116         cs = css_cs(css);
2117 
2118         percpu_down_write(&cpuset_rwsem);
2119 
2120         /* allow moving tasks into an empty cpuset if on default hierarchy */
2121         ret = -ENOSPC;
2122         if (!is_in_v2_mode() &&
2123             (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed)))
2124                 goto out_unlock;
2125 
2126         cgroup_taskset_for_each(task, css, tset) {
2127                 ret = task_can_attach(task, cs->cpus_allowed);
2128                 if (ret)
2129                         goto out_unlock;
2130                 ret = security_task_setscheduler(task);
2131                 if (ret)
2132                         goto out_unlock;
2133         }
2134 
2135         /*
2136          * Mark attach is in progress.  This makes validate_change() fail
2137          * changes which zero cpus/mems_allowed.
2138          */
2139         cs->attach_in_progress++;
2140         ret = 0;
2141 out_unlock:
2142         percpu_up_write(&cpuset_rwsem);
2143         return ret;
2144 }
2145 
2146 static void cpuset_cancel_attach(struct cgroup_taskset *tset)
2147 {
2148         struct cgroup_subsys_state *css;
2149 
2150         cgroup_taskset_first(tset, &css);
2151 
2152         percpu_down_write(&cpuset_rwsem);
2153         css_cs(css)->attach_in_progress--;
2154         percpu_up_write(&cpuset_rwsem);
2155 }
2156 
2157 /*
2158  * Protected by cpuset_mutex.  cpus_attach is used only by cpuset_attach()
2159  * but we can't allocate it dynamically there.  Define it global and
2160  * allocate from cpuset_init().
2161  */
2162 static cpumask_var_t cpus_attach;
2163 
2164 static void cpuset_attach(struct cgroup_taskset *tset)
2165 {
2166         /* static buf protected by cpuset_mutex */
2167         static nodemask_t cpuset_attach_nodemask_to;
2168         struct task_struct *task;
2169         struct task_struct *leader;
2170         struct cgroup_subsys_state *css;
2171         struct cpuset *cs;
2172         struct cpuset *oldcs = cpuset_attach_old_cs;
2173 
2174         cgroup_taskset_first(tset, &css);
2175         cs = css_cs(css);
2176 
2177         percpu_down_write(&cpuset_rwsem);
2178 
2179         /* prepare for attach */
2180         if (cs == &top_cpuset)
2181                 cpumask_copy(cpus_attach, cpu_possible_mask);
2182         else
2183                 guarantee_online_cpus(cs, cpus_attach);
2184 
2185         guarantee_online_mems(cs, &cpuset_attach_nodemask_to);
2186 
2187         cgroup_taskset_for_each(task, css, tset) {
2188                 /*
2189                  * can_attach beforehand should guarantee that this doesn't
2190                  * fail.  TODO: have a better way to handle failure here
2191                  */
2192                 WARN_ON_ONCE(set_cpus_allowed_ptr(task, cpus_attach));
2193 
2194                 cpuset_change_task_nodemask(task, &cpuset_attach_nodemask_to);
2195                 cpuset_update_task_spread_flag(cs, task);
2196         }
2197 
2198         /*
2199          * Change mm for all threadgroup leaders. This is expensive and may
2200          * sleep and should be moved outside migration path proper.
2201          */
2202         cpuset_attach_nodemask_to = cs->effective_mems;
2203         cgroup_taskset_for_each_leader(leader, css, tset) {
2204                 struct mm_struct *mm = get_task_mm(leader);
2205 
2206                 if (mm) {
2207                         mpol_rebind_mm(mm, &cpuset_attach_nodemask_to);
2208 
2209                         /*
2210                          * old_mems_allowed is the same with mems_allowed
2211                          * here, except if this task is being moved
2212                          * automatically due to hotplug.  In that case
2213                          * @mems_allowed has been updated and is empty, so
2214                          * @old_mems_allowed is the right nodesets that we
2215                          * migrate mm from.
2216                          */
2217                         if (is_memory_migrate(cs))
2218                                 cpuset_migrate_mm(mm, &oldcs->old_mems_allowed,
2219                                                   &cpuset_attach_nodemask_to);
2220                         else
2221                                 mmput(mm);
2222                 }
2223         }
2224 
2225         cs->old_mems_allowed = cpuset_attach_nodemask_to;
2226 
2227         cs->attach_in_progress--;
2228         if (!cs->attach_in_progress)
2229                 wake_up(&cpuset_attach_wq);
2230 
2231         percpu_up_write(&cpuset_rwsem);
2232 }
2233 
2234 /* The various types of files and directories in a cpuset file system */
2235 
2236 typedef enum {
2237         FILE_MEMORY_MIGRATE,
2238         FILE_CPULIST,
2239         FILE_MEMLIST,
2240         FILE_EFFECTIVE_CPULIST,
2241         FILE_EFFECTIVE_MEMLIST,
2242         FILE_SUBPARTS_CPULIST,
2243         FILE_CPU_EXCLUSIVE,
2244         FILE_MEM_EXCLUSIVE,
2245         FILE_MEM_HARDWALL,
2246         FILE_SCHED_LOAD_BALANCE,
2247         FILE_PARTITION_ROOT,
2248         FILE_SCHED_RELAX_DOMAIN_LEVEL,
2249         FILE_MEMORY_PRESSURE_ENABLED,
2250         FILE_MEMORY_PRESSURE,
2251         FILE_SPREAD_PAGE,
2252         FILE_SPREAD_SLAB,
2253 } cpuset_filetype_t;
2254 
2255 static int cpuset_write_u64(struct cgroup_subsys_state *css, struct cftype *cft,
2256                             u64 val)
2257 {
2258         struct cpuset *cs = css_cs(css);
2259         cpuset_filetype_t type = cft->private;
2260         int retval = 0;
2261 
2262         get_online_cpus();
2263         percpu_down_write(&cpuset_rwsem);
2264         if (!is_cpuset_online(cs)) {
2265                 retval = -ENODEV;
2266                 goto out_unlock;
2267         }
2268 
2269         switch (type) {
2270         case FILE_CPU_EXCLUSIVE:
2271                 retval = update_flag(CS_CPU_EXCLUSIVE, cs, val);
2272                 break;
2273         case FILE_MEM_EXCLUSIVE:
2274                 retval = update_flag(CS_MEM_EXCLUSIVE, cs, val);
2275                 break;
2276         case FILE_MEM_HARDWALL:
2277                 retval = update_flag(CS_MEM_HARDWALL, cs, val);
2278                 break;
2279         case FILE_SCHED_LOAD_BALANCE:
2280                 retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, val);
2281                 break;
2282         case FILE_MEMORY_MIGRATE:
2283                 retval = update_flag(CS_MEMORY_MIGRATE, cs, val);
2284                 break;
2285         case FILE_MEMORY_PRESSURE_ENABLED:
2286                 cpuset_memory_pressure_enabled = !!val;
2287                 break;
2288         case FILE_SPREAD_PAGE:
2289                 retval = update_flag(CS_SPREAD_PAGE, cs, val);
2290                 break;
2291         case FILE_SPREAD_SLAB:
2292                 retval = update_flag(CS_SPREAD_SLAB, cs, val);
2293                 break;
2294         default:
2295                 retval = -EINVAL;
2296                 break;
2297         }
2298 out_unlock:
2299         percpu_up_write(&cpuset_rwsem);
2300         put_online_cpus();
2301         return retval;
2302 }
2303 
2304 static int cpuset_write_s64(struct cgroup_subsys_state *css, struct cftype *cft,
2305                             s64 val)
2306 {
2307         struct cpuset *cs = css_cs(css);
2308         cpuset_filetype_t type = cft->private;
2309         int retval = -ENODEV;
2310 
2311         get_online_cpus();
2312         percpu_down_write(&cpuset_rwsem);
2313         if (!is_cpuset_online(cs))
2314                 goto out_unlock;
2315 
2316         switch (type) {
2317         case FILE_SCHED_RELAX_DOMAIN_LEVEL:
2318                 retval = update_relax_domain_level(cs, val);
2319                 break;
2320         default:
2321                 retval = -EINVAL;
2322                 break;
2323         }
2324 out_unlock:
2325         percpu_up_write(&cpuset_rwsem);
2326         put_online_cpus();
2327         return retval;
2328 }
2329 
2330 /*
2331  * Common handling for a write to a "cpus" or "mems" file.
2332  */
2333 static ssize_t cpuset_write_resmask(struct kernfs_open_file *of,
2334                                     char *buf, size_t nbytes, loff_t off)
2335 {
2336         struct cpuset *cs = css_cs(of_css(of));
2337         struct cpuset *trialcs;
2338         int retval = -ENODEV;
2339 
2340         buf = strstrip(buf);
2341 
2342         /*
2343          * CPU or memory hotunplug may leave @cs w/o any execution
2344          * resources, in which case the hotplug code asynchronously updates
2345          * configuration and transfers all tasks to the nearest ancestor
2346          * which can execute.
2347          *
2348          * As writes to "cpus" or "mems" may restore @cs's execution
2349          * resources, wait for the previously scheduled operations before
2350          * proceeding, so that we don't end up keep removing tasks added
2351          * after execution capability is restored.
2352          *
2353          * cpuset_hotplug_work calls back into cgroup core via
2354          * cgroup_transfer_tasks() and waiting for it from a cgroupfs
2355          * operation like this one can lead to a deadlock through kernfs
2356          * active_ref protection.  Let's break the protection.  Losing the
2357          * protection is okay as we check whether @cs is online after
2358          * grabbing cpuset_mutex anyway.  This only happens on the legacy
2359          * hierarchies.
2360          */
2361         css_get(&cs->css);
2362         kernfs_break_active_protection(of->kn);
2363         flush_work(&cpuset_hotplug_work);
2364 
2365         get_online_cpus();
2366         percpu_down_write(&cpuset_rwsem);
2367         if (!is_cpuset_online(cs))
2368                 goto out_unlock;
2369 
2370         trialcs = alloc_trial_cpuset(cs);
2371         if (!trialcs) {
2372                 retval = -ENOMEM;
2373                 goto out_unlock;
2374         }
2375 
2376         switch (of_cft(of)->private) {
2377         case FILE_CPULIST:
2378                 retval = update_cpumask(cs, trialcs, buf);
2379                 break;
2380         case FILE_MEMLIST:
2381                 retval = update_nodemask(cs, trialcs, buf);
2382                 break;
2383         default:
2384                 retval = -EINVAL;
2385                 break;
2386         }
2387 
2388         free_cpuset(trialcs);
2389 out_unlock:
2390         percpu_up_write(&cpuset_rwsem);
2391         put_online_cpus();
2392         kernfs_unbreak_active_protection(of->kn);
2393         css_put(&cs->css);
2394         flush_workqueue(cpuset_migrate_mm_wq);
2395         return retval ?: nbytes;
2396 }
2397 
2398 /*
2399  * These ascii lists should be read in a single call, by using a user
2400  * buffer large enough to hold the entire map.  If read in smaller
2401  * chunks, there is no guarantee of atomicity.  Since the display format
2402  * used, list of ranges of sequential numbers, is variable length,
2403  * and since these maps can change value dynamically, one could read
2404  * gibberish by doing partial reads while a list was changing.
2405  */
2406 static int cpuset_common_seq_show(struct seq_file *sf, void *v)
2407 {
2408         struct cpuset *cs = css_cs(seq_css(sf));
2409         cpuset_filetype_t type = seq_cft(sf)->private;
2410         int ret = 0;
2411 
2412         spin_lock_irq(&callback_lock);
2413 
2414         switch (type) {
2415         case FILE_CPULIST:
2416                 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->cpus_allowed));
2417                 break;
2418         case FILE_MEMLIST:
2419                 seq_printf(sf, "%*pbl\n", nodemask_pr_args(&cs->mems_allowed));
2420                 break;
2421         case FILE_EFFECTIVE_CPULIST:
2422                 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->effective_cpus));
2423                 break;
2424         case FILE_EFFECTIVE_MEMLIST:
2425                 seq_printf(sf, "%*pbl\n", nodemask_pr_args(&cs->effective_mems));
2426                 break;
2427         case FILE_SUBPARTS_CPULIST:
2428                 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->subparts_cpus));
2429                 break;
2430         default:
2431                 ret = -EINVAL;
2432         }
2433 
2434         spin_unlock_irq(&callback_lock);
2435         return ret;
2436 }
2437 
2438 static u64 cpuset_read_u64(struct cgroup_subsys_state *css, struct cftype *cft)
2439 {
2440         struct cpuset *cs = css_cs(css);
2441         cpuset_filetype_t type = cft->private;
2442         switch (type) {
2443         case FILE_CPU_EXCLUSIVE:
2444                 return is_cpu_exclusive(cs);
2445         case FILE_MEM_EXCLUSIVE:
2446                 return is_mem_exclusive(cs);
2447         case FILE_MEM_HARDWALL:
2448                 return is_mem_hardwall(cs);
2449         case FILE_SCHED_LOAD_BALANCE:
2450                 return is_sched_load_balance(cs);
2451         case FILE_MEMORY_MIGRATE:
2452                 return is_memory_migrate(cs);
2453         case FILE_MEMORY_PRESSURE_ENABLED:
2454                 return cpuset_memory_pressure_enabled;
2455         case FILE_MEMORY_PRESSURE:
2456                 return fmeter_getrate(&cs->fmeter);
2457         case FILE_SPREAD_PAGE:
2458                 return is_spread_page(cs);
2459         case FILE_SPREAD_SLAB:
2460                 return is_spread_slab(cs);
2461         default:
2462                 BUG();
2463         }
2464 
2465         /* Unreachable but makes gcc happy */
2466         return 0;
2467 }
2468 
2469 static s64 cpuset_read_s64(struct cgroup_subsys_state *css, struct cftype *cft)
2470 {
2471         struct cpuset *cs = css_cs(css);
2472         cpuset_filetype_t type = cft->private;
2473         switch (type) {
2474         case FILE_SCHED_RELAX_DOMAIN_LEVEL:
2475                 return cs->relax_domain_level;
2476         default:
2477                 BUG();
2478         }
2479 
2480         /* Unrechable but makes gcc happy */
2481         return 0;
2482 }
2483 
2484 static int sched_partition_show(struct seq_file *seq, void *v)
2485 {
2486         struct cpuset *cs = css_cs(seq_css(seq));
2487 
2488         switch (cs->partition_root_state) {
2489         case PRS_ENABLED:
2490                 seq_puts(seq, "root\n");
2491                 break;
2492         case PRS_DISABLED:
2493                 seq_puts(seq, "member\n");
2494                 break;
2495         case PRS_ERROR:
2496                 seq_puts(seq, "root invalid\n");
2497                 break;
2498         }
2499         return 0;
2500 }
2501 
2502 static ssize_t sched_partition_write(struct kernfs_open_file *of, char *buf,
2503                                      size_t nbytes, loff_t off)
2504 {
2505         struct cpuset *cs = css_cs(of_css(of));
2506         int val;
2507         int retval = -ENODEV;
2508 
2509         buf = strstrip(buf);
2510 
2511         /*
2512          * Convert "root" to ENABLED, and convert "member" to DISABLED.
2513          */
2514         if (!strcmp(buf, "root"))
2515                 val = PRS_ENABLED;
2516         else if (!strcmp(buf, "member"))
2517                 val = PRS_DISABLED;
2518         else
2519                 return -EINVAL;
2520 
2521         css_get(&cs->css);
2522         get_online_cpus();
2523         percpu_down_write(&cpuset_rwsem);
2524         if (!is_cpuset_online(cs))
2525                 goto out_unlock;
2526 
2527         retval = update_prstate(cs, val);
2528 out_unlock:
2529         percpu_up_write(&cpuset_rwsem);
2530         put_online_cpus();
2531         css_put(&cs->css);
2532         return retval ?: nbytes;
2533 }
2534 
2535 /*
2536  * for the common functions, 'private' gives the type of file
2537  */
2538 
2539 static struct cftype legacy_files[] = {
2540         {
2541                 .name = "cpus",
2542                 .seq_show = cpuset_common_seq_show,
2543                 .write = cpuset_write_resmask,
2544                 .max_write_len = (100U + 6 * NR_CPUS),
2545                 .private = FILE_CPULIST,
2546         },
2547 
2548         {
2549                 .name = "mems",
2550                 .seq_show = cpuset_common_seq_show,
2551                 .write = cpuset_write_resmask,
2552                 .max_write_len = (100U + 6 * MAX_NUMNODES),
2553                 .private = FILE_MEMLIST,
2554         },
2555 
2556         {
2557                 .name = "effective_cpus",
2558                 .seq_show = cpuset_common_seq_show,
2559                 .private = FILE_EFFECTIVE_CPULIST,
2560         },
2561 
2562         {
2563                 .name = "effective_mems",
2564                 .seq_show = cpuset_common_seq_show,
2565                 .private = FILE_EFFECTIVE_MEMLIST,
2566         },
2567 
2568         {
2569                 .name = "cpu_exclusive",
2570                 .read_u64 = cpuset_read_u64,
2571                 .write_u64 = cpuset_write_u64,
2572                 .private = FILE_CPU_EXCLUSIVE,
2573         },
2574 
2575         {
2576                 .name = "mem_exclusive",
2577                 .read_u64 = cpuset_read_u64,
2578                 .write_u64 = cpuset_write_u64,
2579                 .private = FILE_MEM_EXCLUSIVE,
2580         },
2581 
2582         {
2583                 .name = "mem_hardwall",
2584                 .read_u64 = cpuset_read_u64,
2585                 .write_u64 = cpuset_write_u64,
2586                 .private = FILE_MEM_HARDWALL,
2587         },
2588 
2589         {
2590                 .name = "sched_load_balance",
2591                 .read_u64 = cpuset_read_u64,
2592                 .write_u64 = cpuset_write_u64,
2593                 .private = FILE_SCHED_LOAD_BALANCE,
2594         },
2595 
2596         {
2597                 .name = "sched_relax_domain_level",
2598                 .read_s64 = cpuset_read_s64,
2599                 .write_s64 = cpuset_write_s64,
2600                 .private = FILE_SCHED_RELAX_DOMAIN_LEVEL,
2601         },
2602 
2603         {
2604                 .name = "memory_migrate",
2605                 .read_u64 = cpuset_read_u64,
2606                 .write_u64 = cpuset_write_u64,
2607                 .private = FILE_MEMORY_MIGRATE,
2608         },
2609 
2610         {
2611                 .name = "memory_pressure",
2612                 .read_u64 = cpuset_read_u64,
2613                 .private = FILE_MEMORY_PRESSURE,
2614         },
2615 
2616         {
2617                 .name = "memory_spread_page",
2618                 .read_u64 = cpuset_read_u64,
2619                 .write_u64 = cpuset_write_u64,
2620                 .private = FILE_SPREAD_PAGE,
2621         },
2622 
2623         {
2624                 .name = "memory_spread_slab",
2625                 .read_u64 = cpuset_read_u64,
2626                 .write_u64 = cpuset_write_u64,
2627                 .private = FILE_SPREAD_SLAB,
2628         },
2629 
2630         {
2631                 .name = "memory_pressure_enabled",
2632                 .flags = CFTYPE_ONLY_ON_ROOT,
2633                 .read_u64 = cpuset_read_u64,
2634                 .write_u64 = cpuset_write_u64,
2635                 .private = FILE_MEMORY_PRESSURE_ENABLED,
2636         },
2637 
2638         { }     /* terminate */
2639 };
2640 
2641 /*
2642  * This is currently a minimal set for the default hierarchy. It can be
2643  * expanded later on by migrating more features and control files from v1.
2644  */
2645 static struct cftype dfl_files[] = {
2646         {
2647                 .name = "cpus",
2648                 .seq_show = cpuset_common_seq_show,
2649                 .write = cpuset_write_resmask,
2650                 .max_write_len = (100U + 6 * NR_CPUS),
2651                 .private = FILE_CPULIST,
2652                 .flags = CFTYPE_NOT_ON_ROOT,
2653         },
2654 
2655         {
2656                 .name = "mems",
2657                 .seq_show = cpuset_common_seq_show,
2658                 .write = cpuset_write_resmask,
2659                 .max_write_len = (100U + 6 * MAX_NUMNODES),
2660                 .private = FILE_MEMLIST,
2661                 .flags = CFTYPE_NOT_ON_ROOT,
2662         },
2663 
2664         {
2665                 .name = "cpus.effective",
2666                 .seq_show = cpuset_common_seq_show,
2667                 .private = FILE_EFFECTIVE_CPULIST,
2668         },
2669 
2670         {
2671                 .name = "mems.effective",
2672                 .seq_show = cpuset_common_seq_show,
2673                 .private = FILE_EFFECTIVE_MEMLIST,
2674         },
2675 
2676         {
2677                 .name = "cpus.partition",
2678                 .seq_show = sched_partition_show,
2679                 .write = sched_partition_write,
2680                 .private = FILE_PARTITION_ROOT,
2681                 .flags = CFTYPE_NOT_ON_ROOT,
2682         },
2683 
2684         {
2685                 .name = "cpus.subpartitions",
2686                 .seq_show = cpuset_common_seq_show,
2687                 .private = FILE_SUBPARTS_CPULIST,
2688                 .flags = CFTYPE_DEBUG,
2689         },
2690 
2691         { }     /* terminate */
2692 };
2693 
2694 
2695 /*
2696  *      cpuset_css_alloc - allocate a cpuset css
2697  *      cgrp:   control group that the new cpuset will be part of
2698  */
2699 
2700 static struct cgroup_subsys_state *
2701 cpuset_css_alloc(struct cgroup_subsys_state *parent_css)
2702 {
2703         struct cpuset *cs;
2704 
2705         if (!parent_css)
2706                 return &top_cpuset.css;
2707 
2708         cs = kzalloc(sizeof(*cs), GFP_KERNEL);
2709         if (!cs)
2710                 return ERR_PTR(-ENOMEM);
2711 
2712         if (alloc_cpumasks(cs, NULL)) {
2713                 kfree(cs);
2714                 return ERR_PTR(-ENOMEM);
2715         }
2716 
2717         set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
2718         nodes_clear(cs->mems_allowed);
2719         nodes_clear(cs->effective_mems);
2720         fmeter_init(&cs->fmeter);
2721         cs->relax_domain_level = -1;
2722 
2723         return &cs->css;
2724 }
2725 
2726 static int cpuset_css_online(struct cgroup_subsys_state *css)
2727 {
2728         struct cpuset *cs = css_cs(css);
2729         struct cpuset *parent = parent_cs(cs);
2730         struct cpuset *tmp_cs;
2731         struct cgroup_subsys_state *pos_css;
2732 
2733         if (!parent)
2734                 return 0;
2735 
2736         get_online_cpus();
2737         percpu_down_write(&cpuset_rwsem);
2738 
2739         set_bit(CS_ONLINE, &cs->flags);
2740         if (is_spread_page(parent))
2741                 set_bit(CS_SPREAD_PAGE, &cs->flags);
2742         if (is_spread_slab(parent))
2743                 set_bit(CS_SPREAD_SLAB, &cs->flags);
2744 
2745         cpuset_inc();
2746 
2747         spin_lock_irq(&callback_lock);
2748         if (is_in_v2_mode()) {
2749                 cpumask_copy(cs->effective_cpus, parent->effective_cpus);
2750                 cs->effective_mems = parent->effective_mems;
2751                 cs->use_parent_ecpus = true;
2752                 parent->child_ecpus_count++;
2753         }
2754         spin_unlock_irq(&callback_lock);
2755 
2756         if (!test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags))
2757                 goto out_unlock;
2758 
2759         /*
2760          * Clone @parent's configuration if CGRP_CPUSET_CLONE_CHILDREN is
2761          * set.  This flag handling is implemented in cgroup core for
2762          * histrical reasons - the flag may be specified during mount.
2763          *
2764          * Currently, if any sibling cpusets have exclusive cpus or mem, we
2765          * refuse to clone the configuration - thereby refusing the task to
2766          * be entered, and as a result refusing the sys_unshare() or
2767          * clone() which initiated it.  If this becomes a problem for some
2768          * users who wish to allow that scenario, then this could be
2769          * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
2770          * (and likewise for mems) to the new cgroup.
2771          */
2772         rcu_read_lock();
2773         cpuset_for_each_child(tmp_cs, pos_css, parent) {
2774                 if (is_mem_exclusive(tmp_cs) || is_cpu_exclusive(tmp_cs)) {
2775                         rcu_read_unlock();
2776                         goto out_unlock;
2777                 }
2778         }
2779         rcu_read_unlock();
2780 
2781         spin_lock_irq(&callback_lock);
2782         cs->mems_allowed = parent->mems_allowed;
2783         cs->effective_mems = parent->mems_allowed;
2784         cpumask_copy(cs->cpus_allowed, parent->cpus_allowed);
2785         cpumask_copy(cs->effective_cpus, parent->cpus_allowed);
2786         spin_unlock_irq(&callback_lock);
2787 out_unlock:
2788         percpu_up_write(&cpuset_rwsem);
2789         put_online_cpus();
2790         return 0;
2791 }
2792 
2793 /*
2794  * If the cpuset being removed has its flag 'sched_load_balance'
2795  * enabled, then simulate turning sched_load_balance off, which
2796  * will call rebuild_sched_domains_locked(). That is not needed
2797  * in the default hierarchy where only changes in partition
2798  * will cause repartitioning.
2799  *
2800  * If the cpuset has the 'sched.partition' flag enabled, simulate
2801  * turning 'sched.partition" off.
2802  */
2803 
2804 static void cpuset_css_offline(struct cgroup_subsys_state *css)
2805 {
2806         struct cpuset *cs = css_cs(css);
2807 
2808         get_online_cpus();
2809         percpu_down_write(&cpuset_rwsem);
2810 
2811         if (is_partition_root(cs))
2812                 update_prstate(cs, 0);
2813 
2814         if (!cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
2815             is_sched_load_balance(cs))
2816                 update_flag(CS_SCHED_LOAD_BALANCE, cs, 0);
2817 
2818         if (cs->use_parent_ecpus) {
2819                 struct cpuset *parent = parent_cs(cs);
2820 
2821                 cs->use_parent_ecpus = false;
2822                 parent->child_ecpus_count--;
2823         }
2824 
2825         cpuset_dec();
2826         clear_bit(CS_ONLINE, &cs->flags);
2827 
2828         percpu_up_write(&cpuset_rwsem);
2829         put_online_cpus();
2830 }
2831 
2832 static void cpuset_css_free(struct cgroup_subsys_state *css)
2833 {
2834         struct cpuset *cs = css_cs(css);
2835 
2836         free_cpuset(cs);
2837 }
2838 
2839 static void cpuset_bind(struct cgroup_subsys_state *root_css)
2840 {
2841         percpu_down_write(&cpuset_rwsem);
2842         spin_lock_irq(&callback_lock);
2843 
2844         if (is_in_v2_mode()) {
2845                 cpumask_copy(top_cpuset.cpus_allowed, cpu_possible_mask);
2846                 top_cpuset.mems_allowed = node_possible_map;
2847         } else {
2848                 cpumask_copy(top_cpuset.cpus_allowed,
2849                              top_cpuset.effective_cpus);
2850                 top_cpuset.mems_allowed = top_cpuset.effective_mems;
2851         }
2852 
2853         spin_unlock_irq(&callback_lock);
2854         percpu_up_write(&cpuset_rwsem);
2855 }
2856 
2857 /*
2858  * Make sure the new task conform to the current state of its parent,
2859  * which could have been changed by cpuset just after it inherits the
2860  * state from the parent and before it sits on the cgroup's task list.
2861  */
2862 static void cpuset_fork(struct task_struct *task)
2863 {
2864         if (task_css_is_root(task, cpuset_cgrp_id))
2865                 return;
2866 
2867         set_cpus_allowed_ptr(task, current->cpus_ptr);
2868         task->mems_allowed = current->mems_allowed;
2869 }
2870 
2871 struct cgroup_subsys cpuset_cgrp_subsys = {
2872         .css_alloc      = cpuset_css_alloc,
2873         .css_online     = cpuset_css_online,
2874         .css_offline    = cpuset_css_offline,
2875         .css_free       = cpuset_css_free,
2876         .can_attach     = cpuset_can_attach,
2877         .cancel_attach  = cpuset_cancel_attach,
2878         .attach         = cpuset_attach,
2879         .post_attach    = cpuset_post_attach,
2880         .bind           = cpuset_bind,
2881         .fork           = cpuset_fork,
2882         .legacy_cftypes = legacy_files,
2883         .dfl_cftypes    = dfl_files,
2884         .early_init     = true,
2885         .threaded       = true,
2886 };
2887 
2888 /**
2889  * cpuset_init - initialize cpusets at system boot
2890  *
2891  * Description: Initialize top_cpuset
2892  **/
2893 
2894 int __init cpuset_init(void)
2895 {
2896         BUG_ON(percpu_init_rwsem(&cpuset_rwsem));
2897 
2898         BUG_ON(!alloc_cpumask_var(&top_cpuset.cpus_allowed, GFP_KERNEL));
2899         BUG_ON(!alloc_cpumask_var(&top_cpuset.effective_cpus, GFP_KERNEL));
2900         BUG_ON(!zalloc_cpumask_var(&top_cpuset.subparts_cpus, GFP_KERNEL));
2901 
2902         cpumask_setall(top_cpuset.cpus_allowed);
2903         nodes_setall(top_cpuset.mems_allowed);
2904         cpumask_setall(top_cpuset.effective_cpus);
2905         nodes_setall(top_cpuset.effective_mems);
2906 
2907         fmeter_init(&top_cpuset.fmeter);
2908         set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags);
2909         top_cpuset.relax_domain_level = -1;
2910 
2911         BUG_ON(!alloc_cpumask_var(&cpus_attach, GFP_KERNEL));
2912 
2913         return 0;
2914 }
2915 
2916 /*
2917  * If CPU and/or memory hotplug handlers, below, unplug any CPUs
2918  * or memory nodes, we need to walk over the cpuset hierarchy,
2919  * removing that CPU or node from all cpusets.  If this removes the
2920  * last CPU or node from a cpuset, then move the tasks in the empty
2921  * cpuset to its next-highest non-empty parent.
2922  */
2923 static void remove_tasks_in_empty_cpuset(struct cpuset *cs)
2924 {
2925         struct cpuset *parent;
2926 
2927         /*
2928          * Find its next-highest non-empty parent, (top cpuset
2929          * has online cpus, so can't be empty).
2930          */
2931         parent = parent_cs(cs);
2932         while (cpumask_empty(parent->cpus_allowed) ||
2933                         nodes_empty(parent->mems_allowed))
2934                 parent = parent_cs(parent);
2935 
2936         if (cgroup_transfer_tasks(parent->css.cgroup, cs->css.cgroup)) {
2937                 pr_err("cpuset: failed to transfer tasks out of empty cpuset ");
2938                 pr_cont_cgroup_name(cs->css.cgroup);
2939                 pr_cont("\n");
2940         }
2941 }
2942 
2943 static void
2944 hotplug_update_tasks_legacy(struct cpuset *cs,
2945                             struct cpumask *new_cpus, nodemask_t *new_mems,
2946                             bool cpus_updated, bool mems_updated)
2947 {
2948         bool is_empty;
2949 
2950         spin_lock_irq(&callback_lock);
2951         cpumask_copy(cs->cpus_allowed, new_cpus);
2952         cpumask_copy(cs->effective_cpus, new_cpus);
2953         cs->mems_allowed = *new_mems;
2954         cs->effective_mems = *new_mems;
2955         spin_unlock_irq(&callback_lock);
2956 
2957         /*
2958          * Don't call update_tasks_cpumask() if the cpuset becomes empty,
2959          * as the tasks will be migratecd to an ancestor.
2960          */
2961         if (cpus_updated && !cpumask_empty(cs->cpus_allowed))
2962                 update_tasks_cpumask(cs);
2963         if (mems_updated && !nodes_empty(cs->mems_allowed))
2964                 update_tasks_nodemask(cs);
2965 
2966         is_empty = cpumask_empty(cs->cpus_allowed) ||
2967                    nodes_empty(cs->mems_allowed);
2968 
2969         percpu_up_write(&cpuset_rwsem);
2970 
2971         /*
2972          * Move tasks to the nearest ancestor with execution resources,
2973          * This is full cgroup operation which will also call back into
2974          * cpuset. Should be done outside any lock.
2975          */
2976         if (is_empty)
2977                 remove_tasks_in_empty_cpuset(cs);
2978 
2979         percpu_down_write(&cpuset_rwsem);
2980 }
2981 
2982 static void
2983 hotplug_update_tasks(struct cpuset *cs,
2984                      struct cpumask *new_cpus, nodemask_t *new_mems,
2985                      bool cpus_updated, bool mems_updated)
2986 {
2987         if (cpumask_empty(new_cpus))
2988                 cpumask_copy(new_cpus, parent_cs(cs)->effective_cpus);
2989         if (nodes_empty(*new_mems))
2990                 *new_mems = parent_cs(cs)->effective_mems;
2991 
2992         spin_lock_irq(&callback_lock);
2993         cpumask_copy(cs->effective_cpus, new_cpus);
2994         cs->effective_mems = *new_mems;
2995         spin_unlock_irq(&callback_lock);
2996 
2997         if (cpus_updated)
2998                 update_tasks_cpumask(cs);
2999         if (mems_updated)
3000                 update_tasks_nodemask(cs);
3001 }
3002 
3003 static bool force_rebuild;
3004 
3005 void cpuset_force_rebuild(void)
3006 {
3007         force_rebuild = true;
3008 }
3009 
3010 /**
3011  * cpuset_hotplug_update_tasks - update tasks in a cpuset for hotunplug
3012  * @cs: cpuset in interest
3013  * @tmp: the tmpmasks structure pointer
3014  *
3015  * Compare @cs's cpu and mem masks against top_cpuset and if some have gone
3016  * offline, update @cs accordingly.  If @cs ends up with no CPU or memory,
3017  * all its tasks are moved to the nearest ancestor with both resources.
3018  */
3019 static void cpuset_hotplug_update_tasks(struct cpuset *cs, struct tmpmasks *tmp)
3020 {
3021         static cpumask_t new_cpus;
3022         static nodemask_t new_mems;
3023         bool cpus_updated;
3024         bool mems_updated;
3025         struct cpuset *parent;
3026 retry:
3027         wait_event(cpuset_attach_wq, cs->attach_in_progress == 0);
3028 
3029         percpu_down_write(&cpuset_rwsem);
3030 
3031         /*
3032          * We have raced with task attaching. We wait until attaching
3033          * is finished, so we won't attach a task to an empty cpuset.
3034          */
3035         if (cs->attach_in_progress) {
3036                 percpu_up_write(&cpuset_rwsem);
3037                 goto retry;
3038         }
3039 
3040         parent =  parent_cs(cs);
3041         compute_effective_cpumask(&new_cpus, cs, parent);
3042         nodes_and(new_mems, cs->mems_allowed, parent->effective_mems);
3043 
3044         if (cs->nr_subparts_cpus)
3045                 /*
3046                  * Make sure that CPUs allocated to child partitions
3047                  * do not show up in effective_cpus.
3048                  */
3049                 cpumask_andnot(&new_cpus, &new_cpus, cs->subparts_cpus);
3050 
3051         if (!tmp || !cs->partition_root_state)
3052                 goto update_tasks;
3053 
3054         /*
3055          * In the unlikely event that a partition root has empty
3056          * effective_cpus or its parent becomes erroneous, we have to
3057          * transition it to the erroneous state.
3058          */
3059         if (is_partition_root(cs) && (cpumask_empty(&new_cpus) ||
3060            (parent->partition_root_state == PRS_ERROR))) {
3061                 if (cs->nr_subparts_cpus) {
3062                         cs->nr_subparts_cpus = 0;
3063                         cpumask_clear(cs->subparts_cpus);
3064                         compute_effective_cpumask(&new_cpus, cs, parent);
3065                 }
3066 
3067                 /*
3068                  * If the effective_cpus is empty because the child
3069                  * partitions take away all the CPUs, we can keep
3070                  * the current partition and let the child partitions
3071                  * fight for available CPUs.
3072                  */
3073                 if ((parent->partition_root_state == PRS_ERROR) ||
3074                      cpumask_empty(&new_cpus)) {
3075                         update_parent_subparts_cpumask(cs, partcmd_disable,
3076                                                        NULL, tmp);
3077                         cs->partition_root_state = PRS_ERROR;
3078                 }
3079                 cpuset_force_rebuild();
3080         }
3081 
3082         /*
3083          * On the other hand, an erroneous partition root may be transitioned
3084          * back to a regular one or a partition root with no CPU allocated
3085          * from the parent may change to erroneous.
3086          */
3087         if (is_partition_root(parent) &&
3088            ((cs->partition_root_state == PRS_ERROR) ||
3089             !cpumask_intersects(&new_cpus, parent->subparts_cpus)) &&
3090              update_parent_subparts_cpumask(cs, partcmd_update, NULL, tmp))
3091                 cpuset_force_rebuild();
3092 
3093 update_tasks:
3094         cpus_updated = !cpumask_equal(&new_cpus, cs->effective_cpus);
3095         mems_updated = !nodes_equal(new_mems, cs->effective_mems);
3096 
3097         if (is_in_v2_mode())
3098                 hotplug_update_tasks(cs, &new_cpus, &new_mems,
3099                                      cpus_updated, mems_updated);
3100         else
3101                 hotplug_update_tasks_legacy(cs, &new_cpus, &new_mems,
3102                                             cpus_updated, mems_updated);
3103 
3104         percpu_up_write(&cpuset_rwsem);
3105 }
3106 
3107 /**
3108  * cpuset_hotplug_workfn - handle CPU/memory hotunplug for a cpuset
3109  *
3110  * This function is called after either CPU or memory configuration has
3111  * changed and updates cpuset accordingly.  The top_cpuset is always
3112  * synchronized to cpu_active_mask and N_MEMORY, which is necessary in
3113  * order to make cpusets transparent (of no affect) on systems that are
3114  * actively using CPU hotplug but making no active use of cpusets.
3115  *
3116  * Non-root cpusets are only affected by offlining.  If any CPUs or memory
3117  * nodes have been taken down, cpuset_hotplug_update_tasks() is invoked on
3118  * all descendants.
3119  *
3120  * Note that CPU offlining during suspend is ignored.  We don't modify
3121  * cpusets across suspend/resume cycles at all.
3122  */
3123 static void cpuset_hotplug_workfn(struct work_struct *work)
3124 {
3125         static cpumask_t new_cpus;
3126         static nodemask_t new_mems;
3127         bool cpus_updated, mems_updated;
3128         bool on_dfl = is_in_v2_mode();
3129         struct tmpmasks tmp, *ptmp = NULL;
3130 
3131         if (on_dfl && !alloc_cpumasks(NULL, &tmp))
3132                 ptmp = &tmp;
3133 
3134         percpu_down_write(&cpuset_rwsem);
3135 
3136         /* fetch the available cpus/mems and find out which changed how */
3137         cpumask_copy(&new_cpus, cpu_active_mask);
3138         new_mems = node_states[N_MEMORY];
3139 
3140         /*
3141          * If subparts_cpus is populated, it is likely that the check below
3142          * will produce a false positive on cpus_updated when the cpu list
3143          * isn't changed. It is extra work, but it is better to be safe.
3144          */
3145         cpus_updated = !cpumask_equal(top_cpuset.effective_cpus, &new_cpus);
3146         mems_updated = !nodes_equal(top_cpuset.effective_mems, new_mems);
3147 
3148         /* synchronize cpus_allowed to cpu_active_mask */
3149         if (cpus_updated) {
3150                 spin_lock_irq(&callback_lock);
3151                 if (!on_dfl)
3152                         cpumask_copy(top_cpuset.cpus_allowed, &new_cpus);
3153                 /*
3154                  * Make sure that CPUs allocated to child partitions
3155                  * do not show up in effective_cpus. If no CPU is left,
3156                  * we clear the subparts_cpus & let the child partitions
3157                  * fight for the CPUs again.
3158                  */
3159                 if (top_cpuset.nr_subparts_cpus) {
3160                         if (cpumask_subset(&new_cpus,
3161                                            top_cpuset.subparts_cpus)) {
3162                                 top_cpuset.nr_subparts_cpus = 0;
3163                                 cpumask_clear(top_cpuset.subparts_cpus);
3164                         } else {
3165                                 cpumask_andnot(&new_cpus, &new_cpus,
3166                                                top_cpuset.subparts_cpus);
3167                         }
3168                 }
3169                 cpumask_copy(top_cpuset.effective_cpus, &new_cpus);
3170                 spin_unlock_irq(&callback_lock);
3171                 /* we don't mess with cpumasks of tasks in top_cpuset */
3172         }
3173 
3174         /* synchronize mems_allowed to N_MEMORY */
3175         if (mems_updated) {
3176                 spin_lock_irq(&callback_lock);
3177                 if (!on_dfl)
3178                         top_cpuset.mems_allowed = new_mems;
3179                 top_cpuset.effective_mems = new_mems;
3180                 spin_unlock_irq(&callback_lock);
3181                 update_tasks_nodemask(&top_cpuset);
3182         }
3183 
3184         percpu_up_write(&cpuset_rwsem);
3185 
3186         /* if cpus or mems changed, we need to propagate to descendants */
3187         if (cpus_updated || mems_updated) {
3188                 struct cpuset *cs;
3189                 struct cgroup_subsys_state *pos_css;
3190 
3191                 rcu_read_lock();
3192                 cpuset_for_each_descendant_pre(cs, pos_css, &top_cpuset) {
3193                         if (cs == &top_cpuset || !css_tryget_online(&cs->css))
3194                                 continue;
3195                         rcu_read_unlock();
3196 
3197                         cpuset_hotplug_update_tasks(cs, ptmp);
3198 
3199                         rcu_read_lock();
3200                         css_put(&cs->css);
3201                 }
3202                 rcu_read_unlock();
3203         }
3204 
3205         /* rebuild sched domains if cpus_allowed has changed */
3206         if (cpus_updated || force_rebuild) {
3207                 force_rebuild = false;
3208                 rebuild_sched_domains();
3209         }
3210 
3211         free_cpumasks(NULL, ptmp);
3212 }
3213 
3214 void cpuset_update_active_cpus(void)
3215 {
3216         /*
3217          * We're inside cpu hotplug critical region which usually nests
3218          * inside cgroup synchronization.  Bounce actual hotplug processing
3219          * to a work item to avoid reverse locking order.
3220          */
3221         schedule_work(&cpuset_hotplug_work);
3222 }
3223 
3224 void cpuset_wait_for_hotplug(void)
3225 {
3226         flush_work(&cpuset_hotplug_work);
3227 }
3228 
3229 /*
3230  * Keep top_cpuset.mems_allowed tracking node_states[N_MEMORY].
3231  * Call this routine anytime after node_states[N_MEMORY] changes.
3232  * See cpuset_update_active_cpus() for CPU hotplug handling.
3233  */
3234 static int cpuset_track_online_nodes(struct notifier_block *self,
3235                                 unsigned long action, void *arg)
3236 {
3237         schedule_work(&cpuset_hotplug_work);
3238         return NOTIFY_OK;
3239 }
3240 
3241 static struct notifier_block cpuset_track_online_nodes_nb = {
3242         .notifier_call = cpuset_track_online_nodes,
3243         .priority = 10,         /* ??! */
3244 };
3245 
3246 /**
3247  * cpuset_init_smp - initialize cpus_allowed
3248  *
3249  * Description: Finish top cpuset after cpu, node maps are initialized
3250  */
3251 void __init cpuset_init_smp(void)
3252 {
3253         cpumask_copy(top_cpuset.cpus_allowed, cpu_active_mask);
3254         top_cpuset.mems_allowed = node_states[N_MEMORY];
3255         top_cpuset.old_mems_allowed = top_cpuset.mems_allowed;
3256 
3257         cpumask_copy(top_cpuset.effective_cpus, cpu_active_mask);
3258         top_cpuset.effective_mems = node_states[N_MEMORY];
3259 
3260         register_hotmemory_notifier(&cpuset_track_online_nodes_nb);
3261 
3262         cpuset_migrate_mm_wq = alloc_ordered_workqueue("cpuset_migrate_mm", 0);
3263         BUG_ON(!cpuset_migrate_mm_wq);
3264 }
3265 
3266 /**
3267  * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
3268  * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
3269  * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
3270  *
3271  * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
3272  * attached to the specified @tsk.  Guaranteed to return some non-empty
3273  * subset of cpu_online_mask, even if this means going outside the
3274  * tasks cpuset.
3275  **/
3276 
3277 void cpuset_cpus_allowed(struct task_struct *tsk, struct cpumask *pmask)
3278 {
3279         unsigned long flags;
3280 
3281         spin_lock_irqsave(&callback_lock, flags);
3282         rcu_read_lock();
3283         guarantee_online_cpus(task_cs(tsk), pmask);
3284         rcu_read_unlock();
3285         spin_unlock_irqrestore(&callback_lock, flags);
3286 }
3287 
3288 /**
3289  * cpuset_cpus_allowed_fallback - final fallback before complete catastrophe.
3290  * @tsk: pointer to task_struct with which the scheduler is struggling
3291  *
3292  * Description: In the case that the scheduler cannot find an allowed cpu in
3293  * tsk->cpus_allowed, we fall back to task_cs(tsk)->cpus_allowed. In legacy
3294  * mode however, this value is the same as task_cs(tsk)->effective_cpus,
3295  * which will not contain a sane cpumask during cases such as cpu hotplugging.
3296  * This is the absolute last resort for the scheduler and it is only used if
3297  * _every_ other avenue has been traveled.
3298  **/
3299 
3300 void cpuset_cpus_allowed_fallback(struct task_struct *tsk)
3301 {
3302         rcu_read_lock();
3303         do_set_cpus_allowed(tsk, is_in_v2_mode() ?
3304                 task_cs(tsk)->cpus_allowed : cpu_possible_mask);
3305         rcu_read_unlock();
3306 
3307         /*
3308          * We own tsk->cpus_allowed, nobody can change it under us.
3309          *
3310          * But we used cs && cs->cpus_allowed lockless and thus can
3311          * race with cgroup_attach_task() or update_cpumask() and get
3312          * the wrong tsk->cpus_allowed. However, both cases imply the
3313          * subsequent cpuset_change_cpumask()->set_cpus_allowed_ptr()
3314          * which takes task_rq_lock().
3315          *
3316          * If we are called after it dropped the lock we must see all
3317          * changes in tsk_cs()->cpus_allowed. Otherwise we can temporary
3318          * set any mask even if it is not right from task_cs() pov,
3319          * the pending set_cpus_allowed_ptr() will fix things.
3320          *
3321          * select_fallback_rq() will fix things ups and set cpu_possible_mask
3322          * if required.
3323          */
3324 }
3325 
3326 void __init cpuset_init_current_mems_allowed(void)
3327 {
3328         nodes_setall(current->mems_allowed);
3329 }
3330 
3331 /**
3332  * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
3333  * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
3334  *
3335  * Description: Returns the nodemask_t mems_allowed of the cpuset
3336  * attached to the specified @tsk.  Guaranteed to return some non-empty
3337  * subset of node_states[N_MEMORY], even if this means going outside the
3338  * tasks cpuset.
3339  **/
3340 
3341 nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
3342 {
3343         nodemask_t mask;
3344         unsigned long flags;
3345 
3346         spin_lock_irqsave(&callback_lock, flags);
3347         rcu_read_lock();
3348         guarantee_online_mems(task_cs(tsk), &mask);
3349         rcu_read_unlock();
3350         spin_unlock_irqrestore(&callback_lock, flags);
3351 
3352         return mask;
3353 }
3354 
3355 /**
3356  * cpuset_nodemask_valid_mems_allowed - check nodemask vs. curremt mems_allowed
3357  * @nodemask: the nodemask to be checked
3358  *
3359  * Are any of the nodes in the nodemask allowed in current->mems_allowed?
3360  */
3361 int cpuset_nodemask_valid_mems_allowed(nodemask_t *nodemask)
3362 {
3363         return nodes_intersects(*nodemask, current->mems_allowed);
3364 }
3365 
3366 /*
3367  * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
3368  * mem_hardwall ancestor to the specified cpuset.  Call holding
3369  * callback_lock.  If no ancestor is mem_exclusive or mem_hardwall
3370  * (an unusual configuration), then returns the root cpuset.
3371  */
3372 static struct cpuset *nearest_hardwall_ancestor(struct cpuset *cs)
3373 {
3374         while (!(is_mem_exclusive(cs) || is_mem_hardwall(cs)) && parent_cs(cs))
3375                 cs = parent_cs(cs);
3376         return cs;
3377 }
3378 
3379 /**
3380  * cpuset_node_allowed - Can we allocate on a memory node?
3381  * @node: is this an allowed node?
3382  * @gfp_mask: memory allocation flags
3383  *
3384  * If we're in interrupt, yes, we can always allocate.  If @node is set in
3385  * current's mems_allowed, yes.  If it's not a __GFP_HARDWALL request and this
3386  * node is set in the nearest hardwalled cpuset ancestor to current's cpuset,
3387  * yes.  If current has access to memory reserves as an oom victim, yes.
3388  * Otherwise, no.
3389  *
3390  * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
3391  * and do not allow allocations outside the current tasks cpuset
3392  * unless the task has been OOM killed.
3393  * GFP_KERNEL allocations are not so marked, so can escape to the
3394  * nearest enclosing hardwalled ancestor cpuset.
3395  *
3396  * Scanning up parent cpusets requires callback_lock.  The
3397  * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
3398  * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
3399  * current tasks mems_allowed came up empty on the first pass over
3400  * the zonelist.  So only GFP_KERNEL allocations, if all nodes in the
3401  * cpuset are short of memory, might require taking the callback_lock.
3402  *
3403  * The first call here from mm/page_alloc:get_page_from_freelist()
3404  * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
3405  * so no allocation on a node outside the cpuset is allowed (unless
3406  * in interrupt, of course).
3407  *
3408  * The second pass through get_page_from_freelist() doesn't even call
3409  * here for GFP_ATOMIC calls.  For those calls, the __alloc_pages()
3410  * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
3411  * in alloc_flags.  That logic and the checks below have the combined
3412  * affect that:
3413  *      in_interrupt - any node ok (current task context irrelevant)
3414  *      GFP_ATOMIC   - any node ok
3415  *      tsk_is_oom_victim   - any node ok
3416  *      GFP_KERNEL   - any node in enclosing hardwalled cpuset ok
3417  *      GFP_USER     - only nodes in current tasks mems allowed ok.
3418  */
3419 bool __cpuset_node_allowed(int node, gfp_t gfp_mask)
3420 {
3421         struct cpuset *cs;              /* current cpuset ancestors */
3422         int allowed;                    /* is allocation in zone z allowed? */
3423         unsigned long flags;
3424 
3425         if (in_interrupt())
3426                 return true;
3427         if (node_isset(node, current->mems_allowed))
3428                 return true;
3429         /*
3430          * Allow tasks that have access to memory reserves because they have
3431          * been OOM killed to get memory anywhere.
3432          */
3433         if (unlikely(tsk_is_oom_victim(current)))
3434                 return true;
3435         if (gfp_mask & __GFP_HARDWALL)  /* If hardwall request, stop here */
3436                 return false;
3437 
3438         if (current->flags & PF_EXITING) /* Let dying task have memory */
3439                 return true;
3440 
3441         /* Not hardwall and node outside mems_allowed: scan up cpusets */
3442         spin_lock_irqsave(&callback_lock, flags);
3443 
3444         rcu_read_lock();
3445         cs = nearest_hardwall_ancestor(task_cs(current));
3446         allowed = node_isset(node, cs->mems_allowed);
3447         rcu_read_unlock();
3448 
3449         spin_unlock_irqrestore(&callback_lock, flags);
3450         return allowed;
3451 }
3452 
3453 /**
3454  * cpuset_mem_spread_node() - On which node to begin search for a file page
3455  * cpuset_slab_spread_node() - On which node to begin search for a slab page
3456  *
3457  * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
3458  * tasks in a cpuset with is_spread_page or is_spread_slab set),
3459  * and if the memory allocation used cpuset_mem_spread_node()
3460  * to determine on which node to start looking, as it will for
3461  * certain page cache or slab cache pages such as used for file
3462  * system buffers and inode caches, then instead of starting on the
3463  * local node to look for a free page, rather spread the starting
3464  * node around the tasks mems_allowed nodes.
3465  *
3466  * We don't have to worry about the returned node being offline
3467  * because "it can't happen", and even if it did, it would be ok.
3468  *
3469  * The routines calling guarantee_online_mems() are careful to
3470  * only set nodes in task->mems_allowed that are online.  So it
3471  * should not be possible for the following code to return an
3472  * offline node.  But if it did, that would be ok, as this routine
3473  * is not returning the node where the allocation must be, only
3474  * the node where the search should start.  The zonelist passed to
3475  * __alloc_pages() will include all nodes.  If the slab allocator
3476  * is passed an offline node, it will fall back to the local node.
3477  * See kmem_cache_alloc_node().
3478  */
3479 
3480 static int cpuset_spread_node(int *rotor)
3481 {
3482         return *rotor = next_node_in(*rotor, current->mems_allowed);
3483 }
3484 
3485 int cpuset_mem_spread_node(void)
3486 {
3487         if (current->cpuset_mem_spread_rotor == NUMA_NO_NODE)
3488                 current->cpuset_mem_spread_rotor =
3489                         node_random(&current->mems_allowed);
3490 
3491         return cpuset_spread_node(&current->cpuset_mem_spread_rotor);
3492 }
3493 
3494 int cpuset_slab_spread_node(void)
3495 {
3496         if (current->cpuset_slab_spread_rotor == NUMA_NO_NODE)
3497                 current->cpuset_slab_spread_rotor =
3498                         node_random(&current->mems_allowed);
3499 
3500         return cpuset_spread_node(&current->cpuset_slab_spread_rotor);
3501 }
3502 
3503 EXPORT_SYMBOL_GPL(cpuset_mem_spread_node);
3504 
3505 /**
3506  * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
3507  * @tsk1: pointer to task_struct of some task.
3508  * @tsk2: pointer to task_struct of some other task.
3509  *
3510  * Description: Return true if @tsk1's mems_allowed intersects the
3511  * mems_allowed of @tsk2.  Used by the OOM killer to determine if
3512  * one of the task's memory usage might impact the memory available
3513  * to the other.
3514  **/
3515 
3516 int cpuset_mems_allowed_intersects(const struct task_struct *tsk1,
3517                                    const struct task_struct *tsk2)
3518 {
3519         return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed);
3520 }
3521 
3522 /**
3523  * cpuset_print_current_mems_allowed - prints current's cpuset and mems_allowed
3524  *
3525  * Description: Prints current's name, cpuset name, and cached copy of its
3526  * mems_allowed to the kernel log.
3527  */
3528 void cpuset_print_current_mems_allowed(void)
3529 {
3530         struct cgroup *cgrp;
3531 
3532         rcu_read_lock();
3533 
3534         cgrp = task_cs(current)->css.cgroup;
3535         pr_cont(",cpuset=");
3536         pr_cont_cgroup_name(cgrp);
3537         pr_cont(",mems_allowed=%*pbl",
3538                 nodemask_pr_args(&current->mems_allowed));
3539 
3540         rcu_read_unlock();
3541 }
3542 
3543 /*
3544  * Collection of memory_pressure is suppressed unless
3545  * this flag is enabled by writing "1" to the special
3546  * cpuset file 'memory_pressure_enabled' in the root cpuset.
3547  */
3548 
3549 int cpuset_memory_pressure_enabled __read_mostly;
3550 
3551 /**
3552  * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
3553  *
3554  * Keep a running average of the rate of synchronous (direct)
3555  * page reclaim efforts initiated by tasks in each cpuset.
3556  *
3557  * This represents the rate at which some task in the cpuset
3558  * ran low on memory on all nodes it was allowed to use, and
3559  * had to enter the kernels page reclaim code in an effort to
3560  * create more free memory by tossing clean pages or swapping
3561  * or writing dirty pages.
3562  *
3563  * Display to user space in the per-cpuset read-only file
3564  * "memory_pressure".  Value displayed is an integer
3565  * representing the recent rate of entry into the synchronous
3566  * (direct) page reclaim by any task attached to the cpuset.
3567  **/
3568 
3569 void __cpuset_memory_pressure_bump(void)
3570 {
3571         rcu_read_lock();
3572         fmeter_markevent(&task_cs(current)->fmeter);
3573         rcu_read_unlock();
3574 }
3575 
3576 #ifdef CONFIG_PROC_PID_CPUSET
3577 /*
3578  * proc_cpuset_show()
3579  *  - Print tasks cpuset path into seq_file.
3580  *  - Used for /proc/<pid>/cpuset.
3581  *  - No need to task_lock(tsk) on this tsk->cpuset reference, as it
3582  *    doesn't really matter if tsk->cpuset changes after we read it,
3583  *    and we take cpuset_mutex, keeping cpuset_attach() from changing it
3584  *    anyway.
3585  */
3586 int proc_cpuset_show(struct seq_file *m, struct pid_namespace *ns,
3587                      struct pid *pid, struct task_struct *tsk)
3588 {
3589         char *buf;
3590         struct cgroup_subsys_state *css;
3591         int retval;
3592 
3593         retval = -ENOMEM;
3594         buf = kmalloc(PATH_MAX, GFP_KERNEL);
3595         if (!buf)
3596                 goto out;
3597 
3598         css = task_get_css(tsk, cpuset_cgrp_id);
3599         retval = cgroup_path_ns(css->cgroup, buf, PATH_MAX,
3600                                 current->nsproxy->cgroup_ns);
3601         css_put(css);
3602         if (retval >= PATH_MAX)
3603                 retval = -ENAMETOOLONG;
3604         if (retval < 0)
3605                 goto out_free;
3606         seq_puts(m, buf);
3607         seq_putc(m, '\n');
3608         retval = 0;
3609 out_free:
3610         kfree(buf);
3611 out:
3612         return retval;
3613 }
3614 #endif /* CONFIG_PROC_PID_CPUSET */
3615 
3616 /* Display task mems_allowed in /proc/<pid>/status file. */
3617 void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task)
3618 {
3619         seq_printf(m, "Mems_allowed:\t%*pb\n",
3620                    nodemask_pr_args(&task->mems_allowed));
3621         seq_printf(m, "Mems_allowed_list:\t%*pbl\n",
3622                    nodemask_pr_args(&task->mems_allowed));
3623 }
3624 

~ [ 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