Version:  2.0.40 2.2.26 2.4.37 3.8 3.9 3.10 3.11 3.12 3.13 3.14 3.15 3.16 3.17 3.18 3.19 4.0 4.1 4.2 4.3 4.4

Linux/kernel/cpuset.c

  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/namei.h>
 43 #include <linux/pagemap.h>
 44 #include <linux/proc_fs.h>
 45 #include <linux/rcupdate.h>
 46 #include <linux/sched.h>
 47 #include <linux/seq_file.h>
 48 #include <linux/security.h>
 49 #include <linux/slab.h>
 50 #include <linux/spinlock.h>
 51 #include <linux/stat.h>
 52 #include <linux/string.h>
 53 #include <linux/time.h>
 54 #include <linux/backing-dev.h>
 55 #include <linux/sort.h>
 56 
 57 #include <asm/uaccess.h>
 58 #include <linux/atomic.h>
 59 #include <linux/mutex.h>
 60 #include <linux/workqueue.h>
 61 #include <linux/cgroup.h>
 62 #include <linux/wait.h>
 63 
 64 struct static_key cpusets_enabled_key __read_mostly = STATIC_KEY_INIT_FALSE;
 65 
 66 /* See "Frequency meter" comments, below. */
 67 
 68 struct fmeter {
 69         int cnt;                /* unprocessed events count */
 70         int val;                /* most recent output value */
 71         time_t time;            /* clock (secs) when val computed */
 72         spinlock_t lock;        /* guards read or write of above */
 73 };
 74 
 75 struct cpuset {
 76         struct cgroup_subsys_state css;
 77 
 78         unsigned long flags;            /* "unsigned long" so bitops work */
 79 
 80         /*
 81          * On default hierarchy:
 82          *
 83          * The user-configured masks can only be changed by writing to
 84          * cpuset.cpus and cpuset.mems, and won't be limited by the
 85          * parent masks.
 86          *
 87          * The effective masks is the real masks that apply to the tasks
 88          * in the cpuset. They may be changed if the configured masks are
 89          * changed or hotplug happens.
 90          *
 91          * effective_mask == configured_mask & parent's effective_mask,
 92          * and if it ends up empty, it will inherit the parent's mask.
 93          *
 94          *
 95          * On legacy hierachy:
 96          *
 97          * The user-configured masks are always the same with effective masks.
 98          */
 99 
100         /* user-configured CPUs and Memory Nodes allow to tasks */
101         cpumask_var_t cpus_allowed;
102         nodemask_t mems_allowed;
103 
104         /* effective CPUs and Memory Nodes allow to tasks */
105         cpumask_var_t effective_cpus;
106         nodemask_t effective_mems;
107 
108         /*
109          * This is old Memory Nodes tasks took on.
110          *
111          * - top_cpuset.old_mems_allowed is initialized to mems_allowed.
112          * - A new cpuset's old_mems_allowed is initialized when some
113          *   task is moved into it.
114          * - old_mems_allowed is used in cpuset_migrate_mm() when we change
115          *   cpuset.mems_allowed and have tasks' nodemask updated, and
116          *   then old_mems_allowed is updated to mems_allowed.
117          */
118         nodemask_t old_mems_allowed;
119 
120         struct fmeter fmeter;           /* memory_pressure filter */
121 
122         /*
123          * Tasks are being attached to this cpuset.  Used to prevent
124          * zeroing cpus/mems_allowed between ->can_attach() and ->attach().
125          */
126         int attach_in_progress;
127 
128         /* partition number for rebuild_sched_domains() */
129         int pn;
130 
131         /* for custom sched domain */
132         int relax_domain_level;
133 };
134 
135 static inline struct cpuset *css_cs(struct cgroup_subsys_state *css)
136 {
137         return css ? container_of(css, struct cpuset, css) : NULL;
138 }
139 
140 /* Retrieve the cpuset for a task */
141 static inline struct cpuset *task_cs(struct task_struct *task)
142 {
143         return css_cs(task_css(task, cpuset_cgrp_id));
144 }
145 
146 static inline struct cpuset *parent_cs(struct cpuset *cs)
147 {
148         return css_cs(cs->css.parent);
149 }
150 
151 #ifdef CONFIG_NUMA
152 static inline bool task_has_mempolicy(struct task_struct *task)
153 {
154         return task->mempolicy;
155 }
156 #else
157 static inline bool task_has_mempolicy(struct task_struct *task)
158 {
159         return false;
160 }
161 #endif
162 
163 
164 /* bits in struct cpuset flags field */
165 typedef enum {
166         CS_ONLINE,
167         CS_CPU_EXCLUSIVE,
168         CS_MEM_EXCLUSIVE,
169         CS_MEM_HARDWALL,
170         CS_MEMORY_MIGRATE,
171         CS_SCHED_LOAD_BALANCE,
172         CS_SPREAD_PAGE,
173         CS_SPREAD_SLAB,
174 } cpuset_flagbits_t;
175 
176 /* convenient tests for these bits */
177 static inline bool is_cpuset_online(const struct cpuset *cs)
178 {
179         return test_bit(CS_ONLINE, &cs->flags);
180 }
181 
182 static inline int is_cpu_exclusive(const struct cpuset *cs)
183 {
184         return test_bit(CS_CPU_EXCLUSIVE, &cs->flags);
185 }
186 
187 static inline int is_mem_exclusive(const struct cpuset *cs)
188 {
189         return test_bit(CS_MEM_EXCLUSIVE, &cs->flags);
190 }
191 
192 static inline int is_mem_hardwall(const struct cpuset *cs)
193 {
194         return test_bit(CS_MEM_HARDWALL, &cs->flags);
195 }
196 
197 static inline int is_sched_load_balance(const struct cpuset *cs)
198 {
199         return test_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
200 }
201 
202 static inline int is_memory_migrate(const struct cpuset *cs)
203 {
204         return test_bit(CS_MEMORY_MIGRATE, &cs->flags);
205 }
206 
207 static inline int is_spread_page(const struct cpuset *cs)
208 {
209         return test_bit(CS_SPREAD_PAGE, &cs->flags);
210 }
211 
212 static inline int is_spread_slab(const struct cpuset *cs)
213 {
214         return test_bit(CS_SPREAD_SLAB, &cs->flags);
215 }
216 
217 static struct cpuset top_cpuset = {
218         .flags = ((1 << CS_ONLINE) | (1 << CS_CPU_EXCLUSIVE) |
219                   (1 << CS_MEM_EXCLUSIVE)),
220 };
221 
222 /**
223  * cpuset_for_each_child - traverse online children of a cpuset
224  * @child_cs: loop cursor pointing to the current child
225  * @pos_css: used for iteration
226  * @parent_cs: target cpuset to walk children of
227  *
228  * Walk @child_cs through the online children of @parent_cs.  Must be used
229  * with RCU read locked.
230  */
231 #define cpuset_for_each_child(child_cs, pos_css, parent_cs)             \
232         css_for_each_child((pos_css), &(parent_cs)->css)                \
233                 if (is_cpuset_online(((child_cs) = css_cs((pos_css)))))
234 
235 /**
236  * cpuset_for_each_descendant_pre - pre-order walk of a cpuset's descendants
237  * @des_cs: loop cursor pointing to the current descendant
238  * @pos_css: used for iteration
239  * @root_cs: target cpuset to walk ancestor of
240  *
241  * Walk @des_cs through the online descendants of @root_cs.  Must be used
242  * with RCU read locked.  The caller may modify @pos_css by calling
243  * css_rightmost_descendant() to skip subtree.  @root_cs is included in the
244  * iteration and the first node to be visited.
245  */
246 #define cpuset_for_each_descendant_pre(des_cs, pos_css, root_cs)        \
247         css_for_each_descendant_pre((pos_css), &(root_cs)->css)         \
248                 if (is_cpuset_online(((des_cs) = css_cs((pos_css)))))
249 
250 /*
251  * There are two global locks guarding cpuset structures - cpuset_mutex and
252  * callback_lock. We also require taking task_lock() when dereferencing a
253  * task's cpuset pointer. See "The task_lock() exception", at the end of this
254  * comment.
255  *
256  * A task must hold both locks to modify cpusets.  If a task holds
257  * cpuset_mutex, then it blocks others wanting that mutex, ensuring that it
258  * is the only task able to also acquire callback_lock and be able to
259  * modify cpusets.  It can perform various checks on the cpuset structure
260  * first, knowing nothing will change.  It can also allocate memory while
261  * just holding cpuset_mutex.  While it is performing these checks, various
262  * callback routines can briefly acquire callback_lock to query cpusets.
263  * Once it is ready to make the changes, it takes callback_lock, blocking
264  * everyone else.
265  *
266  * Calls to the kernel memory allocator can not be made while holding
267  * callback_lock, as that would risk double tripping on callback_lock
268  * from one of the callbacks into the cpuset code from within
269  * __alloc_pages().
270  *
271  * If a task is only holding callback_lock, then it has read-only
272  * access to cpusets.
273  *
274  * Now, the task_struct fields mems_allowed and mempolicy may be changed
275  * by other task, we use alloc_lock in the task_struct fields to protect
276  * them.
277  *
278  * The cpuset_common_file_read() handlers only hold callback_lock across
279  * small pieces of code, such as when reading out possibly multi-word
280  * cpumasks and nodemasks.
281  *
282  * Accessing a task's cpuset should be done in accordance with the
283  * guidelines for accessing subsystem state in kernel/cgroup.c
284  */
285 
286 static DEFINE_MUTEX(cpuset_mutex);
287 static DEFINE_SPINLOCK(callback_lock);
288 
289 /*
290  * CPU / memory hotplug is handled asynchronously.
291  */
292 static void cpuset_hotplug_workfn(struct work_struct *work);
293 static DECLARE_WORK(cpuset_hotplug_work, cpuset_hotplug_workfn);
294 
295 static DECLARE_WAIT_QUEUE_HEAD(cpuset_attach_wq);
296 
297 /*
298  * This is ugly, but preserves the userspace API for existing cpuset
299  * users. If someone tries to mount the "cpuset" filesystem, we
300  * silently switch it to mount "cgroup" instead
301  */
302 static struct dentry *cpuset_mount(struct file_system_type *fs_type,
303                          int flags, const char *unused_dev_name, void *data)
304 {
305         struct file_system_type *cgroup_fs = get_fs_type("cgroup");
306         struct dentry *ret = ERR_PTR(-ENODEV);
307         if (cgroup_fs) {
308                 char mountopts[] =
309                         "cpuset,noprefix,"
310                         "release_agent=/sbin/cpuset_release_agent";
311                 ret = cgroup_fs->mount(cgroup_fs, flags,
312                                            unused_dev_name, mountopts);
313                 put_filesystem(cgroup_fs);
314         }
315         return ret;
316 }
317 
318 static struct file_system_type cpuset_fs_type = {
319         .name = "cpuset",
320         .mount = cpuset_mount,
321 };
322 
323 /*
324  * Return in pmask the portion of a cpusets's cpus_allowed that
325  * are online.  If none are online, walk up the cpuset hierarchy
326  * until we find one that does have some online cpus.  The top
327  * cpuset always has some cpus online.
328  *
329  * One way or another, we guarantee to return some non-empty subset
330  * of cpu_online_mask.
331  *
332  * Call with callback_lock or cpuset_mutex held.
333  */
334 static void guarantee_online_cpus(struct cpuset *cs, struct cpumask *pmask)
335 {
336         while (!cpumask_intersects(cs->effective_cpus, cpu_online_mask))
337                 cs = parent_cs(cs);
338         cpumask_and(pmask, cs->effective_cpus, cpu_online_mask);
339 }
340 
341 /*
342  * Return in *pmask the portion of a cpusets's mems_allowed that
343  * are online, with memory.  If none are online with memory, walk
344  * up the cpuset hierarchy until we find one that does have some
345  * online mems.  The top cpuset always has some mems online.
346  *
347  * One way or another, we guarantee to return some non-empty subset
348  * of node_states[N_MEMORY].
349  *
350  * Call with callback_lock or cpuset_mutex held.
351  */
352 static void guarantee_online_mems(struct cpuset *cs, nodemask_t *pmask)
353 {
354         while (!nodes_intersects(cs->effective_mems, node_states[N_MEMORY]))
355                 cs = parent_cs(cs);
356         nodes_and(*pmask, cs->effective_mems, node_states[N_MEMORY]);
357 }
358 
359 /*
360  * update task's spread flag if cpuset's page/slab spread flag is set
361  *
362  * Call with callback_lock or cpuset_mutex held.
363  */
364 static void cpuset_update_task_spread_flag(struct cpuset *cs,
365                                         struct task_struct *tsk)
366 {
367         if (is_spread_page(cs))
368                 task_set_spread_page(tsk);
369         else
370                 task_clear_spread_page(tsk);
371 
372         if (is_spread_slab(cs))
373                 task_set_spread_slab(tsk);
374         else
375                 task_clear_spread_slab(tsk);
376 }
377 
378 /*
379  * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
380  *
381  * One cpuset is a subset of another if all its allowed CPUs and
382  * Memory Nodes are a subset of the other, and its exclusive flags
383  * are only set if the other's are set.  Call holding cpuset_mutex.
384  */
385 
386 static int is_cpuset_subset(const struct cpuset *p, const struct cpuset *q)
387 {
388         return  cpumask_subset(p->cpus_allowed, q->cpus_allowed) &&
389                 nodes_subset(p->mems_allowed, q->mems_allowed) &&
390                 is_cpu_exclusive(p) <= is_cpu_exclusive(q) &&
391                 is_mem_exclusive(p) <= is_mem_exclusive(q);
392 }
393 
394 /**
395  * alloc_trial_cpuset - allocate a trial cpuset
396  * @cs: the cpuset that the trial cpuset duplicates
397  */
398 static struct cpuset *alloc_trial_cpuset(struct cpuset *cs)
399 {
400         struct cpuset *trial;
401 
402         trial = kmemdup(cs, sizeof(*cs), GFP_KERNEL);
403         if (!trial)
404                 return NULL;
405 
406         if (!alloc_cpumask_var(&trial->cpus_allowed, GFP_KERNEL))
407                 goto free_cs;
408         if (!alloc_cpumask_var(&trial->effective_cpus, GFP_KERNEL))
409                 goto free_cpus;
410 
411         cpumask_copy(trial->cpus_allowed, cs->cpus_allowed);
412         cpumask_copy(trial->effective_cpus, cs->effective_cpus);
413         return trial;
414 
415 free_cpus:
416         free_cpumask_var(trial->cpus_allowed);
417 free_cs:
418         kfree(trial);
419         return NULL;
420 }
421 
422 /**
423  * free_trial_cpuset - free the trial cpuset
424  * @trial: the trial cpuset to be freed
425  */
426 static void free_trial_cpuset(struct cpuset *trial)
427 {
428         free_cpumask_var(trial->effective_cpus);
429         free_cpumask_var(trial->cpus_allowed);
430         kfree(trial);
431 }
432 
433 /*
434  * validate_change() - Used to validate that any proposed cpuset change
435  *                     follows the structural rules for cpusets.
436  *
437  * If we replaced the flag and mask values of the current cpuset
438  * (cur) with those values in the trial cpuset (trial), would
439  * our various subset and exclusive rules still be valid?  Presumes
440  * cpuset_mutex held.
441  *
442  * 'cur' is the address of an actual, in-use cpuset.  Operations
443  * such as list traversal that depend on the actual address of the
444  * cpuset in the list must use cur below, not trial.
445  *
446  * 'trial' is the address of bulk structure copy of cur, with
447  * perhaps one or more of the fields cpus_allowed, mems_allowed,
448  * or flags changed to new, trial values.
449  *
450  * Return 0 if valid, -errno if not.
451  */
452 
453 static int validate_change(struct cpuset *cur, struct cpuset *trial)
454 {
455         struct cgroup_subsys_state *css;
456         struct cpuset *c, *par;
457         int ret;
458 
459         rcu_read_lock();
460 
461         /* Each of our child cpusets must be a subset of us */
462         ret = -EBUSY;
463         cpuset_for_each_child(c, css, cur)
464                 if (!is_cpuset_subset(c, trial))
465                         goto out;
466 
467         /* Remaining checks don't apply to root cpuset */
468         ret = 0;
469         if (cur == &top_cpuset)
470                 goto out;
471 
472         par = parent_cs(cur);
473 
474         /* On legacy hiearchy, we must be a subset of our parent cpuset. */
475         ret = -EACCES;
476         if (!cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
477             !is_cpuset_subset(trial, par))
478                 goto out;
479 
480         /*
481          * If either I or some sibling (!= me) is exclusive, we can't
482          * overlap
483          */
484         ret = -EINVAL;
485         cpuset_for_each_child(c, css, par) {
486                 if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
487                     c != cur &&
488                     cpumask_intersects(trial->cpus_allowed, c->cpus_allowed))
489                         goto out;
490                 if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) &&
491                     c != cur &&
492                     nodes_intersects(trial->mems_allowed, c->mems_allowed))
493                         goto out;
494         }
495 
496         /*
497          * Cpusets with tasks - existing or newly being attached - can't
498          * be changed to have empty cpus_allowed or mems_allowed.
499          */
500         ret = -ENOSPC;
501         if ((cgroup_is_populated(cur->css.cgroup) || cur->attach_in_progress)) {
502                 if (!cpumask_empty(cur->cpus_allowed) &&
503                     cpumask_empty(trial->cpus_allowed))
504                         goto out;
505                 if (!nodes_empty(cur->mems_allowed) &&
506                     nodes_empty(trial->mems_allowed))
507                         goto out;
508         }
509 
510         /*
511          * We can't shrink if we won't have enough room for SCHED_DEADLINE
512          * tasks.
513          */
514         ret = -EBUSY;
515         if (is_cpu_exclusive(cur) &&
516             !cpuset_cpumask_can_shrink(cur->cpus_allowed,
517                                        trial->cpus_allowed))
518                 goto out;
519 
520         ret = 0;
521 out:
522         rcu_read_unlock();
523         return ret;
524 }
525 
526 #ifdef CONFIG_SMP
527 /*
528  * Helper routine for generate_sched_domains().
529  * Do cpusets a, b have overlapping effective cpus_allowed masks?
530  */
531 static int cpusets_overlap(struct cpuset *a, struct cpuset *b)
532 {
533         return cpumask_intersects(a->effective_cpus, b->effective_cpus);
534 }
535 
536 static void
537 update_domain_attr(struct sched_domain_attr *dattr, struct cpuset *c)
538 {
539         if (dattr->relax_domain_level < c->relax_domain_level)
540                 dattr->relax_domain_level = c->relax_domain_level;
541         return;
542 }
543 
544 static void update_domain_attr_tree(struct sched_domain_attr *dattr,
545                                     struct cpuset *root_cs)
546 {
547         struct cpuset *cp;
548         struct cgroup_subsys_state *pos_css;
549 
550         rcu_read_lock();
551         cpuset_for_each_descendant_pre(cp, pos_css, root_cs) {
552                 /* skip the whole subtree if @cp doesn't have any CPU */
553                 if (cpumask_empty(cp->cpus_allowed)) {
554                         pos_css = css_rightmost_descendant(pos_css);
555                         continue;
556                 }
557 
558                 if (is_sched_load_balance(cp))
559                         update_domain_attr(dattr, cp);
560         }
561         rcu_read_unlock();
562 }
563 
564 /*
565  * generate_sched_domains()
566  *
567  * This function builds a partial partition of the systems CPUs
568  * A 'partial partition' is a set of non-overlapping subsets whose
569  * union is a subset of that set.
570  * The output of this function needs to be passed to kernel/sched/core.c
571  * partition_sched_domains() routine, which will rebuild the scheduler's
572  * load balancing domains (sched domains) as specified by that partial
573  * partition.
574  *
575  * See "What is sched_load_balance" in Documentation/cgroups/cpusets.txt
576  * for a background explanation of this.
577  *
578  * Does not return errors, on the theory that the callers of this
579  * routine would rather not worry about failures to rebuild sched
580  * domains when operating in the severe memory shortage situations
581  * that could cause allocation failures below.
582  *
583  * Must be called with cpuset_mutex held.
584  *
585  * The three key local variables below are:
586  *    q  - a linked-list queue of cpuset pointers, used to implement a
587  *         top-down scan of all cpusets.  This scan loads a pointer
588  *         to each cpuset marked is_sched_load_balance into the
589  *         array 'csa'.  For our purposes, rebuilding the schedulers
590  *         sched domains, we can ignore !is_sched_load_balance cpusets.
591  *  csa  - (for CpuSet Array) Array of pointers to all the cpusets
592  *         that need to be load balanced, for convenient iterative
593  *         access by the subsequent code that finds the best partition,
594  *         i.e the set of domains (subsets) of CPUs such that the
595  *         cpus_allowed of every cpuset marked is_sched_load_balance
596  *         is a subset of one of these domains, while there are as
597  *         many such domains as possible, each as small as possible.
598  * doms  - Conversion of 'csa' to an array of cpumasks, for passing to
599  *         the kernel/sched/core.c routine partition_sched_domains() in a
600  *         convenient format, that can be easily compared to the prior
601  *         value to determine what partition elements (sched domains)
602  *         were changed (added or removed.)
603  *
604  * Finding the best partition (set of domains):
605  *      The triple nested loops below over i, j, k scan over the
606  *      load balanced cpusets (using the array of cpuset pointers in
607  *      csa[]) looking for pairs of cpusets that have overlapping
608  *      cpus_allowed, but which don't have the same 'pn' partition
609  *      number and gives them in the same partition number.  It keeps
610  *      looping on the 'restart' label until it can no longer find
611  *      any such pairs.
612  *
613  *      The union of the cpus_allowed masks from the set of
614  *      all cpusets having the same 'pn' value then form the one
615  *      element of the partition (one sched domain) to be passed to
616  *      partition_sched_domains().
617  */
618 static int generate_sched_domains(cpumask_var_t **domains,
619                         struct sched_domain_attr **attributes)
620 {
621         struct cpuset *cp;      /* scans q */
622         struct cpuset **csa;    /* array of all cpuset ptrs */
623         int csn;                /* how many cpuset ptrs in csa so far */
624         int i, j, k;            /* indices for partition finding loops */
625         cpumask_var_t *doms;    /* resulting partition; i.e. sched domains */
626         cpumask_var_t non_isolated_cpus;  /* load balanced CPUs */
627         struct sched_domain_attr *dattr;  /* attributes for custom domains */
628         int ndoms = 0;          /* number of sched domains in result */
629         int nslot;              /* next empty doms[] struct cpumask slot */
630         struct cgroup_subsys_state *pos_css;
631 
632         doms = NULL;
633         dattr = NULL;
634         csa = NULL;
635 
636         if (!alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL))
637                 goto done;
638         cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
639 
640         /* Special case for the 99% of systems with one, full, sched domain */
641         if (is_sched_load_balance(&top_cpuset)) {
642                 ndoms = 1;
643                 doms = alloc_sched_domains(ndoms);
644                 if (!doms)
645                         goto done;
646 
647                 dattr = kmalloc(sizeof(struct sched_domain_attr), GFP_KERNEL);
648                 if (dattr) {
649                         *dattr = SD_ATTR_INIT;
650                         update_domain_attr_tree(dattr, &top_cpuset);
651                 }
652                 cpumask_and(doms[0], top_cpuset.effective_cpus,
653                                      non_isolated_cpus);
654 
655                 goto done;
656         }
657 
658         csa = kmalloc(nr_cpusets() * sizeof(cp), GFP_KERNEL);
659         if (!csa)
660                 goto done;
661         csn = 0;
662 
663         rcu_read_lock();
664         cpuset_for_each_descendant_pre(cp, pos_css, &top_cpuset) {
665                 if (cp == &top_cpuset)
666                         continue;
667                 /*
668                  * Continue traversing beyond @cp iff @cp has some CPUs and
669                  * isn't load balancing.  The former is obvious.  The
670                  * latter: All child cpusets contain a subset of the
671                  * parent's cpus, so just skip them, and then we call
672                  * update_domain_attr_tree() to calc relax_domain_level of
673                  * the corresponding sched domain.
674                  */
675                 if (!cpumask_empty(cp->cpus_allowed) &&
676                     !(is_sched_load_balance(cp) &&
677                       cpumask_intersects(cp->cpus_allowed, non_isolated_cpus)))
678                         continue;
679 
680                 if (is_sched_load_balance(cp))
681                         csa[csn++] = cp;
682 
683                 /* skip @cp's subtree */
684                 pos_css = css_rightmost_descendant(pos_css);
685         }
686         rcu_read_unlock();
687 
688         for (i = 0; i < csn; i++)
689                 csa[i]->pn = i;
690         ndoms = csn;
691 
692 restart:
693         /* Find the best partition (set of sched domains) */
694         for (i = 0; i < csn; i++) {
695                 struct cpuset *a = csa[i];
696                 int apn = a->pn;
697 
698                 for (j = 0; j < csn; j++) {
699                         struct cpuset *b = csa[j];
700                         int bpn = b->pn;
701 
702                         if (apn != bpn && cpusets_overlap(a, b)) {
703                                 for (k = 0; k < csn; k++) {
704                                         struct cpuset *c = csa[k];
705 
706                                         if (c->pn == bpn)
707                                                 c->pn = apn;
708                                 }
709                                 ndoms--;        /* one less element */
710                                 goto restart;
711                         }
712                 }
713         }
714 
715         /*
716          * Now we know how many domains to create.
717          * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
718          */
719         doms = alloc_sched_domains(ndoms);
720         if (!doms)
721                 goto done;
722 
723         /*
724          * The rest of the code, including the scheduler, can deal with
725          * dattr==NULL case. No need to abort if alloc fails.
726          */
727         dattr = kmalloc(ndoms * sizeof(struct sched_domain_attr), GFP_KERNEL);
728 
729         for (nslot = 0, i = 0; i < csn; i++) {
730                 struct cpuset *a = csa[i];
731                 struct cpumask *dp;
732                 int apn = a->pn;
733 
734                 if (apn < 0) {
735                         /* Skip completed partitions */
736                         continue;
737                 }
738 
739                 dp = doms[nslot];
740 
741                 if (nslot == ndoms) {
742                         static int warnings = 10;
743                         if (warnings) {
744                                 pr_warn("rebuild_sched_domains confused: nslot %d, ndoms %d, csn %d, i %d, apn %d\n",
745                                         nslot, ndoms, csn, i, apn);
746                                 warnings--;
747                         }
748                         continue;
749                 }
750 
751                 cpumask_clear(dp);
752                 if (dattr)
753                         *(dattr + nslot) = SD_ATTR_INIT;
754                 for (j = i; j < csn; j++) {
755                         struct cpuset *b = csa[j];
756 
757                         if (apn == b->pn) {
758                                 cpumask_or(dp, dp, b->effective_cpus);
759                                 cpumask_and(dp, dp, non_isolated_cpus);
760                                 if (dattr)
761                                         update_domain_attr_tree(dattr + nslot, b);
762 
763                                 /* Done with this partition */
764                                 b->pn = -1;
765                         }
766                 }
767                 nslot++;
768         }
769         BUG_ON(nslot != ndoms);
770 
771 done:
772         free_cpumask_var(non_isolated_cpus);
773         kfree(csa);
774 
775         /*
776          * Fallback to the default domain if kmalloc() failed.
777          * See comments in partition_sched_domains().
778          */
779         if (doms == NULL)
780                 ndoms = 1;
781 
782         *domains    = doms;
783         *attributes = dattr;
784         return ndoms;
785 }
786 
787 /*
788  * Rebuild scheduler domains.
789  *
790  * If the flag 'sched_load_balance' of any cpuset with non-empty
791  * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
792  * which has that flag enabled, or if any cpuset with a non-empty
793  * 'cpus' is removed, then call this routine to rebuild the
794  * scheduler's dynamic sched domains.
795  *
796  * Call with cpuset_mutex held.  Takes get_online_cpus().
797  */
798 static void rebuild_sched_domains_locked(void)
799 {
800         struct sched_domain_attr *attr;
801         cpumask_var_t *doms;
802         int ndoms;
803 
804         lockdep_assert_held(&cpuset_mutex);
805         get_online_cpus();
806 
807         /*
808          * We have raced with CPU hotplug. Don't do anything to avoid
809          * passing doms with offlined cpu to partition_sched_domains().
810          * Anyways, hotplug work item will rebuild sched domains.
811          */
812         if (!cpumask_equal(top_cpuset.effective_cpus, cpu_active_mask))
813                 goto out;
814 
815         /* Generate domain masks and attrs */
816         ndoms = generate_sched_domains(&doms, &attr);
817 
818         /* Have scheduler rebuild the domains */
819         partition_sched_domains(ndoms, doms, attr);
820 out:
821         put_online_cpus();
822 }
823 #else /* !CONFIG_SMP */
824 static void rebuild_sched_domains_locked(void)
825 {
826 }
827 #endif /* CONFIG_SMP */
828 
829 void rebuild_sched_domains(void)
830 {
831         mutex_lock(&cpuset_mutex);
832         rebuild_sched_domains_locked();
833         mutex_unlock(&cpuset_mutex);
834 }
835 
836 /**
837  * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
838  * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
839  *
840  * Iterate through each task of @cs updating its cpus_allowed to the
841  * effective cpuset's.  As this function is called with cpuset_mutex held,
842  * cpuset membership stays stable.
843  */
844 static void update_tasks_cpumask(struct cpuset *cs)
845 {
846         struct css_task_iter it;
847         struct task_struct *task;
848 
849         css_task_iter_start(&cs->css, &it);
850         while ((task = css_task_iter_next(&it)))
851                 set_cpus_allowed_ptr(task, cs->effective_cpus);
852         css_task_iter_end(&it);
853 }
854 
855 /*
856  * update_cpumasks_hier - Update effective cpumasks and tasks in the subtree
857  * @cs: the cpuset to consider
858  * @new_cpus: temp variable for calculating new effective_cpus
859  *
860  * When congifured cpumask is changed, the effective cpumasks of this cpuset
861  * and all its descendants need to be updated.
862  *
863  * On legacy hierachy, effective_cpus will be the same with cpu_allowed.
864  *
865  * Called with cpuset_mutex held
866  */
867 static void update_cpumasks_hier(struct cpuset *cs, struct cpumask *new_cpus)
868 {
869         struct cpuset *cp;
870         struct cgroup_subsys_state *pos_css;
871         bool need_rebuild_sched_domains = false;
872 
873         rcu_read_lock();
874         cpuset_for_each_descendant_pre(cp, pos_css, cs) {
875                 struct cpuset *parent = parent_cs(cp);
876 
877                 cpumask_and(new_cpus, cp->cpus_allowed, parent->effective_cpus);
878 
879                 /*
880                  * If it becomes empty, inherit the effective mask of the
881                  * parent, which is guaranteed to have some CPUs.
882                  */
883                 if (cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
884                     cpumask_empty(new_cpus))
885                         cpumask_copy(new_cpus, parent->effective_cpus);
886 
887                 /* Skip the whole subtree if the cpumask remains the same. */
888                 if (cpumask_equal(new_cpus, cp->effective_cpus)) {
889                         pos_css = css_rightmost_descendant(pos_css);
890                         continue;
891                 }
892 
893                 if (!css_tryget_online(&cp->css))
894                         continue;
895                 rcu_read_unlock();
896 
897                 spin_lock_irq(&callback_lock);
898                 cpumask_copy(cp->effective_cpus, new_cpus);
899                 spin_unlock_irq(&callback_lock);
900 
901                 WARN_ON(!cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
902                         !cpumask_equal(cp->cpus_allowed, cp->effective_cpus));
903 
904                 update_tasks_cpumask(cp);
905 
906                 /*
907                  * If the effective cpumask of any non-empty cpuset is changed,
908                  * we need to rebuild sched domains.
909                  */
910                 if (!cpumask_empty(cp->cpus_allowed) &&
911                     is_sched_load_balance(cp))
912                         need_rebuild_sched_domains = true;
913 
914                 rcu_read_lock();
915                 css_put(&cp->css);
916         }
917         rcu_read_unlock();
918 
919         if (need_rebuild_sched_domains)
920                 rebuild_sched_domains_locked();
921 }
922 
923 /**
924  * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
925  * @cs: the cpuset to consider
926  * @trialcs: trial cpuset
927  * @buf: buffer of cpu numbers written to this cpuset
928  */
929 static int update_cpumask(struct cpuset *cs, struct cpuset *trialcs,
930                           const char *buf)
931 {
932         int retval;
933 
934         /* top_cpuset.cpus_allowed tracks cpu_online_mask; it's read-only */
935         if (cs == &top_cpuset)
936                 return -EACCES;
937 
938         /*
939          * An empty cpus_allowed is ok only if the cpuset has no tasks.
940          * Since cpulist_parse() fails on an empty mask, we special case
941          * that parsing.  The validate_change() call ensures that cpusets
942          * with tasks have cpus.
943          */
944         if (!*buf) {
945                 cpumask_clear(trialcs->cpus_allowed);
946         } else {
947                 retval = cpulist_parse(buf, trialcs->cpus_allowed);
948                 if (retval < 0)
949                         return retval;
950 
951                 if (!cpumask_subset(trialcs->cpus_allowed,
952                                     top_cpuset.cpus_allowed))
953                         return -EINVAL;
954         }
955 
956         /* Nothing to do if the cpus didn't change */
957         if (cpumask_equal(cs->cpus_allowed, trialcs->cpus_allowed))
958                 return 0;
959 
960         retval = validate_change(cs, trialcs);
961         if (retval < 0)
962                 return retval;
963 
964         spin_lock_irq(&callback_lock);
965         cpumask_copy(cs->cpus_allowed, trialcs->cpus_allowed);
966         spin_unlock_irq(&callback_lock);
967 
968         /* use trialcs->cpus_allowed as a temp variable */
969         update_cpumasks_hier(cs, trialcs->cpus_allowed);
970         return 0;
971 }
972 
973 /*
974  * cpuset_migrate_mm
975  *
976  *    Migrate memory region from one set of nodes to another.
977  *
978  *    Temporarilly set tasks mems_allowed to target nodes of migration,
979  *    so that the migration code can allocate pages on these nodes.
980  *
981  *    While the mm_struct we are migrating is typically from some
982  *    other task, the task_struct mems_allowed that we are hacking
983  *    is for our current task, which must allocate new pages for that
984  *    migrating memory region.
985  */
986 
987 static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
988                                                         const nodemask_t *to)
989 {
990         struct task_struct *tsk = current;
991 
992         tsk->mems_allowed = *to;
993 
994         do_migrate_pages(mm, from, to, MPOL_MF_MOVE_ALL);
995 
996         rcu_read_lock();
997         guarantee_online_mems(task_cs(tsk), &tsk->mems_allowed);
998         rcu_read_unlock();
999 }
1000 
1001 /*
1002  * cpuset_change_task_nodemask - change task's mems_allowed and mempolicy
1003  * @tsk: the task to change
1004  * @newmems: new nodes that the task will be set
1005  *
1006  * In order to avoid seeing no nodes if the old and new nodes are disjoint,
1007  * we structure updates as setting all new allowed nodes, then clearing newly
1008  * disallowed ones.
1009  */
1010 static void cpuset_change_task_nodemask(struct task_struct *tsk,
1011                                         nodemask_t *newmems)
1012 {
1013         bool need_loop;
1014 
1015         /*
1016          * Allow tasks that have access to memory reserves because they have
1017          * been OOM killed to get memory anywhere.
1018          */
1019         if (unlikely(test_thread_flag(TIF_MEMDIE)))
1020                 return;
1021         if (current->flags & PF_EXITING) /* Let dying task have memory */
1022                 return;
1023 
1024         task_lock(tsk);
1025         /*
1026          * Determine if a loop is necessary if another thread is doing
1027          * read_mems_allowed_begin().  If at least one node remains unchanged and
1028          * tsk does not have a mempolicy, then an empty nodemask will not be
1029          * possible when mems_allowed is larger than a word.
1030          */
1031         need_loop = task_has_mempolicy(tsk) ||
1032                         !nodes_intersects(*newmems, tsk->mems_allowed);
1033 
1034         if (need_loop) {
1035                 local_irq_disable();
1036                 write_seqcount_begin(&tsk->mems_allowed_seq);
1037         }
1038 
1039         nodes_or(tsk->mems_allowed, tsk->mems_allowed, *newmems);
1040         mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP1);
1041 
1042         mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP2);
1043         tsk->mems_allowed = *newmems;
1044 
1045         if (need_loop) {
1046                 write_seqcount_end(&tsk->mems_allowed_seq);
1047                 local_irq_enable();
1048         }
1049 
1050         task_unlock(tsk);
1051 }
1052 
1053 static void *cpuset_being_rebound;
1054 
1055 /**
1056  * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
1057  * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
1058  *
1059  * Iterate through each task of @cs updating its mems_allowed to the
1060  * effective cpuset's.  As this function is called with cpuset_mutex held,
1061  * cpuset membership stays stable.
1062  */
1063 static void update_tasks_nodemask(struct cpuset *cs)
1064 {
1065         static nodemask_t newmems;      /* protected by cpuset_mutex */
1066         struct css_task_iter it;
1067         struct task_struct *task;
1068 
1069         cpuset_being_rebound = cs;              /* causes mpol_dup() rebind */
1070 
1071         guarantee_online_mems(cs, &newmems);
1072 
1073         /*
1074          * The mpol_rebind_mm() call takes mmap_sem, which we couldn't
1075          * take while holding tasklist_lock.  Forks can happen - the
1076          * mpol_dup() cpuset_being_rebound check will catch such forks,
1077          * and rebind their vma mempolicies too.  Because we still hold
1078          * the global cpuset_mutex, we know that no other rebind effort
1079          * will be contending for the global variable cpuset_being_rebound.
1080          * It's ok if we rebind the same mm twice; mpol_rebind_mm()
1081          * is idempotent.  Also migrate pages in each mm to new nodes.
1082          */
1083         css_task_iter_start(&cs->css, &it);
1084         while ((task = css_task_iter_next(&it))) {
1085                 struct mm_struct *mm;
1086                 bool migrate;
1087 
1088                 cpuset_change_task_nodemask(task, &newmems);
1089 
1090                 mm = get_task_mm(task);
1091                 if (!mm)
1092                         continue;
1093 
1094                 migrate = is_memory_migrate(cs);
1095 
1096                 mpol_rebind_mm(mm, &cs->mems_allowed);
1097                 if (migrate)
1098                         cpuset_migrate_mm(mm, &cs->old_mems_allowed, &newmems);
1099                 mmput(mm);
1100         }
1101         css_task_iter_end(&it);
1102 
1103         /*
1104          * All the tasks' nodemasks have been updated, update
1105          * cs->old_mems_allowed.
1106          */
1107         cs->old_mems_allowed = newmems;
1108 
1109         /* We're done rebinding vmas to this cpuset's new mems_allowed. */
1110         cpuset_being_rebound = NULL;
1111 }
1112 
1113 /*
1114  * update_nodemasks_hier - Update effective nodemasks and tasks in the subtree
1115  * @cs: the cpuset to consider
1116  * @new_mems: a temp variable for calculating new effective_mems
1117  *
1118  * When configured nodemask is changed, the effective nodemasks of this cpuset
1119  * and all its descendants need to be updated.
1120  *
1121  * On legacy hiearchy, effective_mems will be the same with mems_allowed.
1122  *
1123  * Called with cpuset_mutex held
1124  */
1125 static void update_nodemasks_hier(struct cpuset *cs, nodemask_t *new_mems)
1126 {
1127         struct cpuset *cp;
1128         struct cgroup_subsys_state *pos_css;
1129 
1130         rcu_read_lock();
1131         cpuset_for_each_descendant_pre(cp, pos_css, cs) {
1132                 struct cpuset *parent = parent_cs(cp);
1133 
1134                 nodes_and(*new_mems, cp->mems_allowed, parent->effective_mems);
1135 
1136                 /*
1137                  * If it becomes empty, inherit the effective mask of the
1138                  * parent, which is guaranteed to have some MEMs.
1139                  */
1140                 if (cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
1141                     nodes_empty(*new_mems))
1142                         *new_mems = parent->effective_mems;
1143 
1144                 /* Skip the whole subtree if the nodemask remains the same. */
1145                 if (nodes_equal(*new_mems, cp->effective_mems)) {
1146                         pos_css = css_rightmost_descendant(pos_css);
1147                         continue;
1148                 }
1149 
1150                 if (!css_tryget_online(&cp->css))
1151                         continue;
1152                 rcu_read_unlock();
1153 
1154                 spin_lock_irq(&callback_lock);
1155                 cp->effective_mems = *new_mems;
1156                 spin_unlock_irq(&callback_lock);
1157 
1158                 WARN_ON(!cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
1159                         !nodes_equal(cp->mems_allowed, cp->effective_mems));
1160 
1161                 update_tasks_nodemask(cp);
1162 
1163                 rcu_read_lock();
1164                 css_put(&cp->css);
1165         }
1166         rcu_read_unlock();
1167 }
1168 
1169 /*
1170  * Handle user request to change the 'mems' memory placement
1171  * of a cpuset.  Needs to validate the request, update the
1172  * cpusets mems_allowed, and for each task in the cpuset,
1173  * update mems_allowed and rebind task's mempolicy and any vma
1174  * mempolicies and if the cpuset is marked 'memory_migrate',
1175  * migrate the tasks pages to the new memory.
1176  *
1177  * Call with cpuset_mutex held. May take callback_lock during call.
1178  * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
1179  * lock each such tasks mm->mmap_sem, scan its vma's and rebind
1180  * their mempolicies to the cpusets new mems_allowed.
1181  */
1182 static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
1183                            const char *buf)
1184 {
1185         int retval;
1186 
1187         /*
1188          * top_cpuset.mems_allowed tracks node_stats[N_MEMORY];
1189          * it's read-only
1190          */
1191         if (cs == &top_cpuset) {
1192                 retval = -EACCES;
1193                 goto done;
1194         }
1195 
1196         /*
1197          * An empty mems_allowed is ok iff there are no tasks in the cpuset.
1198          * Since nodelist_parse() fails on an empty mask, we special case
1199          * that parsing.  The validate_change() call ensures that cpusets
1200          * with tasks have memory.
1201          */
1202         if (!*buf) {
1203                 nodes_clear(trialcs->mems_allowed);
1204         } else {
1205                 retval = nodelist_parse(buf, trialcs->mems_allowed);
1206                 if (retval < 0)
1207                         goto done;
1208 
1209                 if (!nodes_subset(trialcs->mems_allowed,
1210                                   top_cpuset.mems_allowed)) {
1211                         retval = -EINVAL;
1212                         goto done;
1213                 }
1214         }
1215 
1216         if (nodes_equal(cs->mems_allowed, trialcs->mems_allowed)) {
1217                 retval = 0;             /* Too easy - nothing to do */
1218                 goto done;
1219         }
1220         retval = validate_change(cs, trialcs);
1221         if (retval < 0)
1222                 goto done;
1223 
1224         spin_lock_irq(&callback_lock);
1225         cs->mems_allowed = trialcs->mems_allowed;
1226         spin_unlock_irq(&callback_lock);
1227 
1228         /* use trialcs->mems_allowed as a temp variable */
1229         update_nodemasks_hier(cs, &trialcs->mems_allowed);
1230 done:
1231         return retval;
1232 }
1233 
1234 int current_cpuset_is_being_rebound(void)
1235 {
1236         int ret;
1237 
1238         rcu_read_lock();
1239         ret = task_cs(current) == cpuset_being_rebound;
1240         rcu_read_unlock();
1241 
1242         return ret;
1243 }
1244 
1245 static int update_relax_domain_level(struct cpuset *cs, s64 val)
1246 {
1247 #ifdef CONFIG_SMP
1248         if (val < -1 || val >= sched_domain_level_max)
1249                 return -EINVAL;
1250 #endif
1251 
1252         if (val != cs->relax_domain_level) {
1253                 cs->relax_domain_level = val;
1254                 if (!cpumask_empty(cs->cpus_allowed) &&
1255                     is_sched_load_balance(cs))
1256                         rebuild_sched_domains_locked();
1257         }
1258 
1259         return 0;
1260 }
1261 
1262 /**
1263  * update_tasks_flags - update the spread flags of tasks in the cpuset.
1264  * @cs: the cpuset in which each task's spread flags needs to be changed
1265  *
1266  * Iterate through each task of @cs updating its spread flags.  As this
1267  * function is called with cpuset_mutex held, cpuset membership stays
1268  * stable.
1269  */
1270 static void update_tasks_flags(struct cpuset *cs)
1271 {
1272         struct css_task_iter it;
1273         struct task_struct *task;
1274 
1275         css_task_iter_start(&cs->css, &it);
1276         while ((task = css_task_iter_next(&it)))
1277                 cpuset_update_task_spread_flag(cs, task);
1278         css_task_iter_end(&it);
1279 }
1280 
1281 /*
1282  * update_flag - read a 0 or a 1 in a file and update associated flag
1283  * bit:         the bit to update (see cpuset_flagbits_t)
1284  * cs:          the cpuset to update
1285  * turning_on:  whether the flag is being set or cleared
1286  *
1287  * Call with cpuset_mutex held.
1288  */
1289 
1290 static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs,
1291                        int turning_on)
1292 {
1293         struct cpuset *trialcs;
1294         int balance_flag_changed;
1295         int spread_flag_changed;
1296         int err;
1297 
1298         trialcs = alloc_trial_cpuset(cs);
1299         if (!trialcs)
1300                 return -ENOMEM;
1301 
1302         if (turning_on)
1303                 set_bit(bit, &trialcs->flags);
1304         else
1305                 clear_bit(bit, &trialcs->flags);
1306 
1307         err = validate_change(cs, trialcs);
1308         if (err < 0)
1309                 goto out;
1310 
1311         balance_flag_changed = (is_sched_load_balance(cs) !=
1312                                 is_sched_load_balance(trialcs));
1313 
1314         spread_flag_changed = ((is_spread_slab(cs) != is_spread_slab(trialcs))
1315                         || (is_spread_page(cs) != is_spread_page(trialcs)));
1316 
1317         spin_lock_irq(&callback_lock);
1318         cs->flags = trialcs->flags;
1319         spin_unlock_irq(&callback_lock);
1320 
1321         if (!cpumask_empty(trialcs->cpus_allowed) && balance_flag_changed)
1322                 rebuild_sched_domains_locked();
1323 
1324         if (spread_flag_changed)
1325                 update_tasks_flags(cs);
1326 out:
1327         free_trial_cpuset(trialcs);
1328         return err;
1329 }
1330 
1331 /*
1332  * Frequency meter - How fast is some event occurring?
1333  *
1334  * These routines manage a digitally filtered, constant time based,
1335  * event frequency meter.  There are four routines:
1336  *   fmeter_init() - initialize a frequency meter.
1337  *   fmeter_markevent() - called each time the event happens.
1338  *   fmeter_getrate() - returns the recent rate of such events.
1339  *   fmeter_update() - internal routine used to update fmeter.
1340  *
1341  * A common data structure is passed to each of these routines,
1342  * which is used to keep track of the state required to manage the
1343  * frequency meter and its digital filter.
1344  *
1345  * The filter works on the number of events marked per unit time.
1346  * The filter is single-pole low-pass recursive (IIR).  The time unit
1347  * is 1 second.  Arithmetic is done using 32-bit integers scaled to
1348  * simulate 3 decimal digits of precision (multiplied by 1000).
1349  *
1350  * With an FM_COEF of 933, and a time base of 1 second, the filter
1351  * has a half-life of 10 seconds, meaning that if the events quit
1352  * happening, then the rate returned from the fmeter_getrate()
1353  * will be cut in half each 10 seconds, until it converges to zero.
1354  *
1355  * It is not worth doing a real infinitely recursive filter.  If more
1356  * than FM_MAXTICKS ticks have elapsed since the last filter event,
1357  * just compute FM_MAXTICKS ticks worth, by which point the level
1358  * will be stable.
1359  *
1360  * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
1361  * arithmetic overflow in the fmeter_update() routine.
1362  *
1363  * Given the simple 32 bit integer arithmetic used, this meter works
1364  * best for reporting rates between one per millisecond (msec) and
1365  * one per 32 (approx) seconds.  At constant rates faster than one
1366  * per msec it maxes out at values just under 1,000,000.  At constant
1367  * rates between one per msec, and one per second it will stabilize
1368  * to a value N*1000, where N is the rate of events per second.
1369  * At constant rates between one per second and one per 32 seconds,
1370  * it will be choppy, moving up on the seconds that have an event,
1371  * and then decaying until the next event.  At rates slower than
1372  * about one in 32 seconds, it decays all the way back to zero between
1373  * each event.
1374  */
1375 
1376 #define FM_COEF 933             /* coefficient for half-life of 10 secs */
1377 #define FM_MAXTICKS ((time_t)99) /* useless computing more ticks than this */
1378 #define FM_MAXCNT 1000000       /* limit cnt to avoid overflow */
1379 #define FM_SCALE 1000           /* faux fixed point scale */
1380 
1381 /* Initialize a frequency meter */
1382 static void fmeter_init(struct fmeter *fmp)
1383 {
1384         fmp->cnt = 0;
1385         fmp->val = 0;
1386         fmp->time = 0;
1387         spin_lock_init(&fmp->lock);
1388 }
1389 
1390 /* Internal meter update - process cnt events and update value */
1391 static void fmeter_update(struct fmeter *fmp)
1392 {
1393         time_t now = get_seconds();
1394         time_t ticks = now - fmp->time;
1395 
1396         if (ticks == 0)
1397                 return;
1398 
1399         ticks = min(FM_MAXTICKS, ticks);
1400         while (ticks-- > 0)
1401                 fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
1402         fmp->time = now;
1403 
1404         fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
1405         fmp->cnt = 0;
1406 }
1407 
1408 /* Process any previous ticks, then bump cnt by one (times scale). */
1409 static void fmeter_markevent(struct fmeter *fmp)
1410 {
1411         spin_lock(&fmp->lock);
1412         fmeter_update(fmp);
1413         fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
1414         spin_unlock(&fmp->lock);
1415 }
1416 
1417 /* Process any previous ticks, then return current value. */
1418 static int fmeter_getrate(struct fmeter *fmp)
1419 {
1420         int val;
1421 
1422         spin_lock(&fmp->lock);
1423         fmeter_update(fmp);
1424         val = fmp->val;
1425         spin_unlock(&fmp->lock);
1426         return val;
1427 }
1428 
1429 static struct cpuset *cpuset_attach_old_cs;
1430 
1431 /* Called by cgroups to determine if a cpuset is usable; cpuset_mutex held */
1432 static int cpuset_can_attach(struct cgroup_taskset *tset)
1433 {
1434         struct cgroup_subsys_state *css;
1435         struct cpuset *cs;
1436         struct task_struct *task;
1437         int ret;
1438 
1439         /* used later by cpuset_attach() */
1440         cpuset_attach_old_cs = task_cs(cgroup_taskset_first(tset, &css));
1441         cs = css_cs(css);
1442 
1443         mutex_lock(&cpuset_mutex);
1444 
1445         /* allow moving tasks into an empty cpuset if on default hierarchy */
1446         ret = -ENOSPC;
1447         if (!cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
1448             (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed)))
1449                 goto out_unlock;
1450 
1451         cgroup_taskset_for_each(task, css, tset) {
1452                 ret = task_can_attach(task, cs->cpus_allowed);
1453                 if (ret)
1454                         goto out_unlock;
1455                 ret = security_task_setscheduler(task);
1456                 if (ret)
1457                         goto out_unlock;
1458         }
1459 
1460         /*
1461          * Mark attach is in progress.  This makes validate_change() fail
1462          * changes which zero cpus/mems_allowed.
1463          */
1464         cs->attach_in_progress++;
1465         ret = 0;
1466 out_unlock:
1467         mutex_unlock(&cpuset_mutex);
1468         return ret;
1469 }
1470 
1471 static void cpuset_cancel_attach(struct cgroup_taskset *tset)
1472 {
1473         struct cgroup_subsys_state *css;
1474         struct cpuset *cs;
1475 
1476         cgroup_taskset_first(tset, &css);
1477         cs = css_cs(css);
1478 
1479         mutex_lock(&cpuset_mutex);
1480         css_cs(css)->attach_in_progress--;
1481         mutex_unlock(&cpuset_mutex);
1482 }
1483 
1484 /*
1485  * Protected by cpuset_mutex.  cpus_attach is used only by cpuset_attach()
1486  * but we can't allocate it dynamically there.  Define it global and
1487  * allocate from cpuset_init().
1488  */
1489 static cpumask_var_t cpus_attach;
1490 
1491 static void cpuset_attach(struct cgroup_taskset *tset)
1492 {
1493         /* static buf protected by cpuset_mutex */
1494         static nodemask_t cpuset_attach_nodemask_to;
1495         struct task_struct *task;
1496         struct task_struct *leader;
1497         struct cgroup_subsys_state *css;
1498         struct cpuset *cs;
1499         struct cpuset *oldcs = cpuset_attach_old_cs;
1500 
1501         cgroup_taskset_first(tset, &css);
1502         cs = css_cs(css);
1503 
1504         mutex_lock(&cpuset_mutex);
1505 
1506         /* prepare for attach */
1507         if (cs == &top_cpuset)
1508                 cpumask_copy(cpus_attach, cpu_possible_mask);
1509         else
1510                 guarantee_online_cpus(cs, cpus_attach);
1511 
1512         guarantee_online_mems(cs, &cpuset_attach_nodemask_to);
1513 
1514         cgroup_taskset_for_each(task, css, tset) {
1515                 /*
1516                  * can_attach beforehand should guarantee that this doesn't
1517                  * fail.  TODO: have a better way to handle failure here
1518                  */
1519                 WARN_ON_ONCE(set_cpus_allowed_ptr(task, cpus_attach));
1520 
1521                 cpuset_change_task_nodemask(task, &cpuset_attach_nodemask_to);
1522                 cpuset_update_task_spread_flag(cs, task);
1523         }
1524 
1525         /*
1526          * Change mm for all threadgroup leaders. This is expensive and may
1527          * sleep and should be moved outside migration path proper.
1528          */
1529         cpuset_attach_nodemask_to = cs->effective_mems;
1530         cgroup_taskset_for_each_leader(leader, css, tset) {
1531                 struct mm_struct *mm = get_task_mm(leader);
1532 
1533                 if (mm) {
1534                         mpol_rebind_mm(mm, &cpuset_attach_nodemask_to);
1535 
1536                         /*
1537                          * old_mems_allowed is the same with mems_allowed
1538                          * here, except if this task is being moved
1539                          * automatically due to hotplug.  In that case
1540                          * @mems_allowed has been updated and is empty, so
1541                          * @old_mems_allowed is the right nodesets that we
1542                          * migrate mm from.
1543                          */
1544                         if (is_memory_migrate(cs)) {
1545                                 cpuset_migrate_mm(mm, &oldcs->old_mems_allowed,
1546                                                   &cpuset_attach_nodemask_to);
1547                         }
1548                         mmput(mm);
1549                 }
1550         }
1551 
1552         cs->old_mems_allowed = cpuset_attach_nodemask_to;
1553 
1554         cs->attach_in_progress--;
1555         if (!cs->attach_in_progress)
1556                 wake_up(&cpuset_attach_wq);
1557 
1558         mutex_unlock(&cpuset_mutex);
1559 }
1560 
1561 /* The various types of files and directories in a cpuset file system */
1562 
1563 typedef enum {
1564         FILE_MEMORY_MIGRATE,
1565         FILE_CPULIST,
1566         FILE_MEMLIST,
1567         FILE_EFFECTIVE_CPULIST,
1568         FILE_EFFECTIVE_MEMLIST,
1569         FILE_CPU_EXCLUSIVE,
1570         FILE_MEM_EXCLUSIVE,
1571         FILE_MEM_HARDWALL,
1572         FILE_SCHED_LOAD_BALANCE,
1573         FILE_SCHED_RELAX_DOMAIN_LEVEL,
1574         FILE_MEMORY_PRESSURE_ENABLED,
1575         FILE_MEMORY_PRESSURE,
1576         FILE_SPREAD_PAGE,
1577         FILE_SPREAD_SLAB,
1578 } cpuset_filetype_t;
1579 
1580 static int cpuset_write_u64(struct cgroup_subsys_state *css, struct cftype *cft,
1581                             u64 val)
1582 {
1583         struct cpuset *cs = css_cs(css);
1584         cpuset_filetype_t type = cft->private;
1585         int retval = 0;
1586 
1587         mutex_lock(&cpuset_mutex);
1588         if (!is_cpuset_online(cs)) {
1589                 retval = -ENODEV;
1590                 goto out_unlock;
1591         }
1592 
1593         switch (type) {
1594         case FILE_CPU_EXCLUSIVE:
1595                 retval = update_flag(CS_CPU_EXCLUSIVE, cs, val);
1596                 break;
1597         case FILE_MEM_EXCLUSIVE:
1598                 retval = update_flag(CS_MEM_EXCLUSIVE, cs, val);
1599                 break;
1600         case FILE_MEM_HARDWALL:
1601                 retval = update_flag(CS_MEM_HARDWALL, cs, val);
1602                 break;
1603         case FILE_SCHED_LOAD_BALANCE:
1604                 retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, val);
1605                 break;
1606         case FILE_MEMORY_MIGRATE:
1607                 retval = update_flag(CS_MEMORY_MIGRATE, cs, val);
1608                 break;
1609         case FILE_MEMORY_PRESSURE_ENABLED:
1610                 cpuset_memory_pressure_enabled = !!val;
1611                 break;
1612         case FILE_SPREAD_PAGE:
1613                 retval = update_flag(CS_SPREAD_PAGE, cs, val);
1614                 break;
1615         case FILE_SPREAD_SLAB:
1616                 retval = update_flag(CS_SPREAD_SLAB, cs, val);
1617                 break;
1618         default:
1619                 retval = -EINVAL;
1620                 break;
1621         }
1622 out_unlock:
1623         mutex_unlock(&cpuset_mutex);
1624         return retval;
1625 }
1626 
1627 static int cpuset_write_s64(struct cgroup_subsys_state *css, struct cftype *cft,
1628                             s64 val)
1629 {
1630         struct cpuset *cs = css_cs(css);
1631         cpuset_filetype_t type = cft->private;
1632         int retval = -ENODEV;
1633 
1634         mutex_lock(&cpuset_mutex);
1635         if (!is_cpuset_online(cs))
1636                 goto out_unlock;
1637 
1638         switch (type) {
1639         case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1640                 retval = update_relax_domain_level(cs, val);
1641                 break;
1642         default:
1643                 retval = -EINVAL;
1644                 break;
1645         }
1646 out_unlock:
1647         mutex_unlock(&cpuset_mutex);
1648         return retval;
1649 }
1650 
1651 /*
1652  * Common handling for a write to a "cpus" or "mems" file.
1653  */
1654 static ssize_t cpuset_write_resmask(struct kernfs_open_file *of,
1655                                     char *buf, size_t nbytes, loff_t off)
1656 {
1657         struct cpuset *cs = css_cs(of_css(of));
1658         struct cpuset *trialcs;
1659         int retval = -ENODEV;
1660 
1661         buf = strstrip(buf);
1662 
1663         /*
1664          * CPU or memory hotunplug may leave @cs w/o any execution
1665          * resources, in which case the hotplug code asynchronously updates
1666          * configuration and transfers all tasks to the nearest ancestor
1667          * which can execute.
1668          *
1669          * As writes to "cpus" or "mems" may restore @cs's execution
1670          * resources, wait for the previously scheduled operations before
1671          * proceeding, so that we don't end up keep removing tasks added
1672          * after execution capability is restored.
1673          *
1674          * cpuset_hotplug_work calls back into cgroup core via
1675          * cgroup_transfer_tasks() and waiting for it from a cgroupfs
1676          * operation like this one can lead to a deadlock through kernfs
1677          * active_ref protection.  Let's break the protection.  Losing the
1678          * protection is okay as we check whether @cs is online after
1679          * grabbing cpuset_mutex anyway.  This only happens on the legacy
1680          * hierarchies.
1681          */
1682         css_get(&cs->css);
1683         kernfs_break_active_protection(of->kn);
1684         flush_work(&cpuset_hotplug_work);
1685 
1686         mutex_lock(&cpuset_mutex);
1687         if (!is_cpuset_online(cs))
1688                 goto out_unlock;
1689 
1690         trialcs = alloc_trial_cpuset(cs);
1691         if (!trialcs) {
1692                 retval = -ENOMEM;
1693                 goto out_unlock;
1694         }
1695 
1696         switch (of_cft(of)->private) {
1697         case FILE_CPULIST:
1698                 retval = update_cpumask(cs, trialcs, buf);
1699                 break;
1700         case FILE_MEMLIST:
1701                 retval = update_nodemask(cs, trialcs, buf);
1702                 break;
1703         default:
1704                 retval = -EINVAL;
1705                 break;
1706         }
1707 
1708         free_trial_cpuset(trialcs);
1709 out_unlock:
1710         mutex_unlock(&cpuset_mutex);
1711         kernfs_unbreak_active_protection(of->kn);
1712         css_put(&cs->css);
1713         return retval ?: nbytes;
1714 }
1715 
1716 /*
1717  * These ascii lists should be read in a single call, by using a user
1718  * buffer large enough to hold the entire map.  If read in smaller
1719  * chunks, there is no guarantee of atomicity.  Since the display format
1720  * used, list of ranges of sequential numbers, is variable length,
1721  * and since these maps can change value dynamically, one could read
1722  * gibberish by doing partial reads while a list was changing.
1723  */
1724 static int cpuset_common_seq_show(struct seq_file *sf, void *v)
1725 {
1726         struct cpuset *cs = css_cs(seq_css(sf));
1727         cpuset_filetype_t type = seq_cft(sf)->private;
1728         int ret = 0;
1729 
1730         spin_lock_irq(&callback_lock);
1731 
1732         switch (type) {
1733         case FILE_CPULIST:
1734                 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->cpus_allowed));
1735                 break;
1736         case FILE_MEMLIST:
1737                 seq_printf(sf, "%*pbl\n", nodemask_pr_args(&cs->mems_allowed));
1738                 break;
1739         case FILE_EFFECTIVE_CPULIST:
1740                 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->effective_cpus));
1741                 break;
1742         case FILE_EFFECTIVE_MEMLIST:
1743                 seq_printf(sf, "%*pbl\n", nodemask_pr_args(&cs->effective_mems));
1744                 break;
1745         default:
1746                 ret = -EINVAL;
1747         }
1748 
1749         spin_unlock_irq(&callback_lock);
1750         return ret;
1751 }
1752 
1753 static u64 cpuset_read_u64(struct cgroup_subsys_state *css, struct cftype *cft)
1754 {
1755         struct cpuset *cs = css_cs(css);
1756         cpuset_filetype_t type = cft->private;
1757         switch (type) {
1758         case FILE_CPU_EXCLUSIVE:
1759                 return is_cpu_exclusive(cs);
1760         case FILE_MEM_EXCLUSIVE:
1761                 return is_mem_exclusive(cs);
1762         case FILE_MEM_HARDWALL:
1763                 return is_mem_hardwall(cs);
1764         case FILE_SCHED_LOAD_BALANCE:
1765                 return is_sched_load_balance(cs);
1766         case FILE_MEMORY_MIGRATE:
1767                 return is_memory_migrate(cs);
1768         case FILE_MEMORY_PRESSURE_ENABLED:
1769                 return cpuset_memory_pressure_enabled;
1770         case FILE_MEMORY_PRESSURE:
1771                 return fmeter_getrate(&cs->fmeter);
1772         case FILE_SPREAD_PAGE:
1773                 return is_spread_page(cs);
1774         case FILE_SPREAD_SLAB:
1775                 return is_spread_slab(cs);
1776         default:
1777                 BUG();
1778         }
1779 
1780         /* Unreachable but makes gcc happy */
1781         return 0;
1782 }
1783 
1784 static s64 cpuset_read_s64(struct cgroup_subsys_state *css, struct cftype *cft)
1785 {
1786         struct cpuset *cs = css_cs(css);
1787         cpuset_filetype_t type = cft->private;
1788         switch (type) {
1789         case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1790                 return cs->relax_domain_level;
1791         default:
1792                 BUG();
1793         }
1794 
1795         /* Unrechable but makes gcc happy */
1796         return 0;
1797 }
1798 
1799 
1800 /*
1801  * for the common functions, 'private' gives the type of file
1802  */
1803 
1804 static struct cftype files[] = {
1805         {
1806                 .name = "cpus",
1807                 .seq_show = cpuset_common_seq_show,
1808                 .write = cpuset_write_resmask,
1809                 .max_write_len = (100U + 6 * NR_CPUS),
1810                 .private = FILE_CPULIST,
1811         },
1812 
1813         {
1814                 .name = "mems",
1815                 .seq_show = cpuset_common_seq_show,
1816                 .write = cpuset_write_resmask,
1817                 .max_write_len = (100U + 6 * MAX_NUMNODES),
1818                 .private = FILE_MEMLIST,
1819         },
1820 
1821         {
1822                 .name = "effective_cpus",
1823                 .seq_show = cpuset_common_seq_show,
1824                 .private = FILE_EFFECTIVE_CPULIST,
1825         },
1826 
1827         {
1828                 .name = "effective_mems",
1829                 .seq_show = cpuset_common_seq_show,
1830                 .private = FILE_EFFECTIVE_MEMLIST,
1831         },
1832 
1833         {
1834                 .name = "cpu_exclusive",
1835                 .read_u64 = cpuset_read_u64,
1836                 .write_u64 = cpuset_write_u64,
1837                 .private = FILE_CPU_EXCLUSIVE,
1838         },
1839 
1840         {
1841                 .name = "mem_exclusive",
1842                 .read_u64 = cpuset_read_u64,
1843                 .write_u64 = cpuset_write_u64,
1844                 .private = FILE_MEM_EXCLUSIVE,
1845         },
1846 
1847         {
1848                 .name = "mem_hardwall",
1849                 .read_u64 = cpuset_read_u64,
1850                 .write_u64 = cpuset_write_u64,
1851                 .private = FILE_MEM_HARDWALL,
1852         },
1853 
1854         {
1855                 .name = "sched_load_balance",
1856                 .read_u64 = cpuset_read_u64,
1857                 .write_u64 = cpuset_write_u64,
1858                 .private = FILE_SCHED_LOAD_BALANCE,
1859         },
1860 
1861         {
1862                 .name = "sched_relax_domain_level",
1863                 .read_s64 = cpuset_read_s64,
1864                 .write_s64 = cpuset_write_s64,
1865                 .private = FILE_SCHED_RELAX_DOMAIN_LEVEL,
1866         },
1867 
1868         {
1869                 .name = "memory_migrate",
1870                 .read_u64 = cpuset_read_u64,
1871                 .write_u64 = cpuset_write_u64,
1872                 .private = FILE_MEMORY_MIGRATE,
1873         },
1874 
1875         {
1876                 .name = "memory_pressure",
1877                 .read_u64 = cpuset_read_u64,
1878         },
1879 
1880         {
1881                 .name = "memory_spread_page",
1882                 .read_u64 = cpuset_read_u64,
1883                 .write_u64 = cpuset_write_u64,
1884                 .private = FILE_SPREAD_PAGE,
1885         },
1886 
1887         {
1888                 .name = "memory_spread_slab",
1889                 .read_u64 = cpuset_read_u64,
1890                 .write_u64 = cpuset_write_u64,
1891                 .private = FILE_SPREAD_SLAB,
1892         },
1893 
1894         {
1895                 .name = "memory_pressure_enabled",
1896                 .flags = CFTYPE_ONLY_ON_ROOT,
1897                 .read_u64 = cpuset_read_u64,
1898                 .write_u64 = cpuset_write_u64,
1899                 .private = FILE_MEMORY_PRESSURE_ENABLED,
1900         },
1901 
1902         { }     /* terminate */
1903 };
1904 
1905 /*
1906  *      cpuset_css_alloc - allocate a cpuset css
1907  *      cgrp:   control group that the new cpuset will be part of
1908  */
1909 
1910 static struct cgroup_subsys_state *
1911 cpuset_css_alloc(struct cgroup_subsys_state *parent_css)
1912 {
1913         struct cpuset *cs;
1914 
1915         if (!parent_css)
1916                 return &top_cpuset.css;
1917 
1918         cs = kzalloc(sizeof(*cs), GFP_KERNEL);
1919         if (!cs)
1920                 return ERR_PTR(-ENOMEM);
1921         if (!alloc_cpumask_var(&cs->cpus_allowed, GFP_KERNEL))
1922                 goto free_cs;
1923         if (!alloc_cpumask_var(&cs->effective_cpus, GFP_KERNEL))
1924                 goto free_cpus;
1925 
1926         set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
1927         cpumask_clear(cs->cpus_allowed);
1928         nodes_clear(cs->mems_allowed);
1929         cpumask_clear(cs->effective_cpus);
1930         nodes_clear(cs->effective_mems);
1931         fmeter_init(&cs->fmeter);
1932         cs->relax_domain_level = -1;
1933 
1934         return &cs->css;
1935 
1936 free_cpus:
1937         free_cpumask_var(cs->cpus_allowed);
1938 free_cs:
1939         kfree(cs);
1940         return ERR_PTR(-ENOMEM);
1941 }
1942 
1943 static int cpuset_css_online(struct cgroup_subsys_state *css)
1944 {
1945         struct cpuset *cs = css_cs(css);
1946         struct cpuset *parent = parent_cs(cs);
1947         struct cpuset *tmp_cs;
1948         struct cgroup_subsys_state *pos_css;
1949 
1950         if (!parent)
1951                 return 0;
1952 
1953         mutex_lock(&cpuset_mutex);
1954 
1955         set_bit(CS_ONLINE, &cs->flags);
1956         if (is_spread_page(parent))
1957                 set_bit(CS_SPREAD_PAGE, &cs->flags);
1958         if (is_spread_slab(parent))
1959                 set_bit(CS_SPREAD_SLAB, &cs->flags);
1960 
1961         cpuset_inc();
1962 
1963         spin_lock_irq(&callback_lock);
1964         if (cgroup_subsys_on_dfl(cpuset_cgrp_subsys)) {
1965                 cpumask_copy(cs->effective_cpus, parent->effective_cpus);
1966                 cs->effective_mems = parent->effective_mems;
1967         }
1968         spin_unlock_irq(&callback_lock);
1969 
1970         if (!test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags))
1971                 goto out_unlock;
1972 
1973         /*
1974          * Clone @parent's configuration if CGRP_CPUSET_CLONE_CHILDREN is
1975          * set.  This flag handling is implemented in cgroup core for
1976          * histrical reasons - the flag may be specified during mount.
1977          *
1978          * Currently, if any sibling cpusets have exclusive cpus or mem, we
1979          * refuse to clone the configuration - thereby refusing the task to
1980          * be entered, and as a result refusing the sys_unshare() or
1981          * clone() which initiated it.  If this becomes a problem for some
1982          * users who wish to allow that scenario, then this could be
1983          * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
1984          * (and likewise for mems) to the new cgroup.
1985          */
1986         rcu_read_lock();
1987         cpuset_for_each_child(tmp_cs, pos_css, parent) {
1988                 if (is_mem_exclusive(tmp_cs) || is_cpu_exclusive(tmp_cs)) {
1989                         rcu_read_unlock();
1990                         goto out_unlock;
1991                 }
1992         }
1993         rcu_read_unlock();
1994 
1995         spin_lock_irq(&callback_lock);
1996         cs->mems_allowed = parent->mems_allowed;
1997         cs->effective_mems = parent->mems_allowed;
1998         cpumask_copy(cs->cpus_allowed, parent->cpus_allowed);
1999         cpumask_copy(cs->effective_cpus, parent->cpus_allowed);
2000         spin_unlock_irq(&callback_lock);
2001 out_unlock:
2002         mutex_unlock(&cpuset_mutex);
2003         return 0;
2004 }
2005 
2006 /*
2007  * If the cpuset being removed has its flag 'sched_load_balance'
2008  * enabled, then simulate turning sched_load_balance off, which
2009  * will call rebuild_sched_domains_locked().
2010  */
2011 
2012 static void cpuset_css_offline(struct cgroup_subsys_state *css)
2013 {
2014         struct cpuset *cs = css_cs(css);
2015 
2016         mutex_lock(&cpuset_mutex);
2017 
2018         if (is_sched_load_balance(cs))
2019                 update_flag(CS_SCHED_LOAD_BALANCE, cs, 0);
2020 
2021         cpuset_dec();
2022         clear_bit(CS_ONLINE, &cs->flags);
2023 
2024         mutex_unlock(&cpuset_mutex);
2025 }
2026 
2027 static void cpuset_css_free(struct cgroup_subsys_state *css)
2028 {
2029         struct cpuset *cs = css_cs(css);
2030 
2031         free_cpumask_var(cs->effective_cpus);
2032         free_cpumask_var(cs->cpus_allowed);
2033         kfree(cs);
2034 }
2035 
2036 static void cpuset_bind(struct cgroup_subsys_state *root_css)
2037 {
2038         mutex_lock(&cpuset_mutex);
2039         spin_lock_irq(&callback_lock);
2040 
2041         if (cgroup_subsys_on_dfl(cpuset_cgrp_subsys)) {
2042                 cpumask_copy(top_cpuset.cpus_allowed, cpu_possible_mask);
2043                 top_cpuset.mems_allowed = node_possible_map;
2044         } else {
2045                 cpumask_copy(top_cpuset.cpus_allowed,
2046                              top_cpuset.effective_cpus);
2047                 top_cpuset.mems_allowed = top_cpuset.effective_mems;
2048         }
2049 
2050         spin_unlock_irq(&callback_lock);
2051         mutex_unlock(&cpuset_mutex);
2052 }
2053 
2054 struct cgroup_subsys cpuset_cgrp_subsys = {
2055         .css_alloc      = cpuset_css_alloc,
2056         .css_online     = cpuset_css_online,
2057         .css_offline    = cpuset_css_offline,
2058         .css_free       = cpuset_css_free,
2059         .can_attach     = cpuset_can_attach,
2060         .cancel_attach  = cpuset_cancel_attach,
2061         .attach         = cpuset_attach,
2062         .bind           = cpuset_bind,
2063         .legacy_cftypes = files,
2064         .early_init     = 1,
2065 };
2066 
2067 /**
2068  * cpuset_init - initialize cpusets at system boot
2069  *
2070  * Description: Initialize top_cpuset and the cpuset internal file system,
2071  **/
2072 
2073 int __init cpuset_init(void)
2074 {
2075         int err = 0;
2076 
2077         if (!alloc_cpumask_var(&top_cpuset.cpus_allowed, GFP_KERNEL))
2078                 BUG();
2079         if (!alloc_cpumask_var(&top_cpuset.effective_cpus, GFP_KERNEL))
2080                 BUG();
2081 
2082         cpumask_setall(top_cpuset.cpus_allowed);
2083         nodes_setall(top_cpuset.mems_allowed);
2084         cpumask_setall(top_cpuset.effective_cpus);
2085         nodes_setall(top_cpuset.effective_mems);
2086 
2087         fmeter_init(&top_cpuset.fmeter);
2088         set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags);
2089         top_cpuset.relax_domain_level = -1;
2090 
2091         err = register_filesystem(&cpuset_fs_type);
2092         if (err < 0)
2093                 return err;
2094 
2095         if (!alloc_cpumask_var(&cpus_attach, GFP_KERNEL))
2096                 BUG();
2097 
2098         return 0;
2099 }
2100 
2101 /*
2102  * If CPU and/or memory hotplug handlers, below, unplug any CPUs
2103  * or memory nodes, we need to walk over the cpuset hierarchy,
2104  * removing that CPU or node from all cpusets.  If this removes the
2105  * last CPU or node from a cpuset, then move the tasks in the empty
2106  * cpuset to its next-highest non-empty parent.
2107  */
2108 static void remove_tasks_in_empty_cpuset(struct cpuset *cs)
2109 {
2110         struct cpuset *parent;
2111 
2112         /*
2113          * Find its next-highest non-empty parent, (top cpuset
2114          * has online cpus, so can't be empty).
2115          */
2116         parent = parent_cs(cs);
2117         while (cpumask_empty(parent->cpus_allowed) ||
2118                         nodes_empty(parent->mems_allowed))
2119                 parent = parent_cs(parent);
2120 
2121         if (cgroup_transfer_tasks(parent->css.cgroup, cs->css.cgroup)) {
2122                 pr_err("cpuset: failed to transfer tasks out of empty cpuset ");
2123                 pr_cont_cgroup_name(cs->css.cgroup);
2124                 pr_cont("\n");
2125         }
2126 }
2127 
2128 static void
2129 hotplug_update_tasks_legacy(struct cpuset *cs,
2130                             struct cpumask *new_cpus, nodemask_t *new_mems,
2131                             bool cpus_updated, bool mems_updated)
2132 {
2133         bool is_empty;
2134 
2135         spin_lock_irq(&callback_lock);
2136         cpumask_copy(cs->cpus_allowed, new_cpus);
2137         cpumask_copy(cs->effective_cpus, new_cpus);
2138         cs->mems_allowed = *new_mems;
2139         cs->effective_mems = *new_mems;
2140         spin_unlock_irq(&callback_lock);
2141 
2142         /*
2143          * Don't call update_tasks_cpumask() if the cpuset becomes empty,
2144          * as the tasks will be migratecd to an ancestor.
2145          */
2146         if (cpus_updated && !cpumask_empty(cs->cpus_allowed))
2147                 update_tasks_cpumask(cs);
2148         if (mems_updated && !nodes_empty(cs->mems_allowed))
2149                 update_tasks_nodemask(cs);
2150 
2151         is_empty = cpumask_empty(cs->cpus_allowed) ||
2152                    nodes_empty(cs->mems_allowed);
2153 
2154         mutex_unlock(&cpuset_mutex);
2155 
2156         /*
2157          * Move tasks to the nearest ancestor with execution resources,
2158          * This is full cgroup operation which will also call back into
2159          * cpuset. Should be done outside any lock.
2160          */
2161         if (is_empty)
2162                 remove_tasks_in_empty_cpuset(cs);
2163 
2164         mutex_lock(&cpuset_mutex);
2165 }
2166 
2167 static void
2168 hotplug_update_tasks(struct cpuset *cs,
2169                      struct cpumask *new_cpus, nodemask_t *new_mems,
2170                      bool cpus_updated, bool mems_updated)
2171 {
2172         if (cpumask_empty(new_cpus))
2173                 cpumask_copy(new_cpus, parent_cs(cs)->effective_cpus);
2174         if (nodes_empty(*new_mems))
2175                 *new_mems = parent_cs(cs)->effective_mems;
2176 
2177         spin_lock_irq(&callback_lock);
2178         cpumask_copy(cs->effective_cpus, new_cpus);
2179         cs->effective_mems = *new_mems;
2180         spin_unlock_irq(&callback_lock);
2181 
2182         if (cpus_updated)
2183                 update_tasks_cpumask(cs);
2184         if (mems_updated)
2185                 update_tasks_nodemask(cs);
2186 }
2187 
2188 /**
2189  * cpuset_hotplug_update_tasks - update tasks in a cpuset for hotunplug
2190  * @cs: cpuset in interest
2191  *
2192  * Compare @cs's cpu and mem masks against top_cpuset and if some have gone
2193  * offline, update @cs accordingly.  If @cs ends up with no CPU or memory,
2194  * all its tasks are moved to the nearest ancestor with both resources.
2195  */
2196 static void cpuset_hotplug_update_tasks(struct cpuset *cs)
2197 {
2198         static cpumask_t new_cpus;
2199         static nodemask_t new_mems;
2200         bool cpus_updated;
2201         bool mems_updated;
2202 retry:
2203         wait_event(cpuset_attach_wq, cs->attach_in_progress == 0);
2204 
2205         mutex_lock(&cpuset_mutex);
2206 
2207         /*
2208          * We have raced with task attaching. We wait until attaching
2209          * is finished, so we won't attach a task to an empty cpuset.
2210          */
2211         if (cs->attach_in_progress) {
2212                 mutex_unlock(&cpuset_mutex);
2213                 goto retry;
2214         }
2215 
2216         cpumask_and(&new_cpus, cs->cpus_allowed, parent_cs(cs)->effective_cpus);
2217         nodes_and(new_mems, cs->mems_allowed, parent_cs(cs)->effective_mems);
2218 
2219         cpus_updated = !cpumask_equal(&new_cpus, cs->effective_cpus);
2220         mems_updated = !nodes_equal(new_mems, cs->effective_mems);
2221 
2222         if (cgroup_subsys_on_dfl(cpuset_cgrp_subsys))
2223                 hotplug_update_tasks(cs, &new_cpus, &new_mems,
2224                                      cpus_updated, mems_updated);
2225         else
2226                 hotplug_update_tasks_legacy(cs, &new_cpus, &new_mems,
2227                                             cpus_updated, mems_updated);
2228 
2229         mutex_unlock(&cpuset_mutex);
2230 }
2231 
2232 /**
2233  * cpuset_hotplug_workfn - handle CPU/memory hotunplug for a cpuset
2234  *
2235  * This function is called after either CPU or memory configuration has
2236  * changed and updates cpuset accordingly.  The top_cpuset is always
2237  * synchronized to cpu_active_mask and N_MEMORY, which is necessary in
2238  * order to make cpusets transparent (of no affect) on systems that are
2239  * actively using CPU hotplug but making no active use of cpusets.
2240  *
2241  * Non-root cpusets are only affected by offlining.  If any CPUs or memory
2242  * nodes have been taken down, cpuset_hotplug_update_tasks() is invoked on
2243  * all descendants.
2244  *
2245  * Note that CPU offlining during suspend is ignored.  We don't modify
2246  * cpusets across suspend/resume cycles at all.
2247  */
2248 static void cpuset_hotplug_workfn(struct work_struct *work)
2249 {
2250         static cpumask_t new_cpus;
2251         static nodemask_t new_mems;
2252         bool cpus_updated, mems_updated;
2253         bool on_dfl = cgroup_subsys_on_dfl(cpuset_cgrp_subsys);
2254 
2255         mutex_lock(&cpuset_mutex);
2256 
2257         /* fetch the available cpus/mems and find out which changed how */
2258         cpumask_copy(&new_cpus, cpu_active_mask);
2259         new_mems = node_states[N_MEMORY];
2260 
2261         cpus_updated = !cpumask_equal(top_cpuset.effective_cpus, &new_cpus);
2262         mems_updated = !nodes_equal(top_cpuset.effective_mems, new_mems);
2263 
2264         /* synchronize cpus_allowed to cpu_active_mask */
2265         if (cpus_updated) {
2266                 spin_lock_irq(&callback_lock);
2267                 if (!on_dfl)
2268                         cpumask_copy(top_cpuset.cpus_allowed, &new_cpus);
2269                 cpumask_copy(top_cpuset.effective_cpus, &new_cpus);
2270                 spin_unlock_irq(&callback_lock);
2271                 /* we don't mess with cpumasks of tasks in top_cpuset */
2272         }
2273 
2274         /* synchronize mems_allowed to N_MEMORY */
2275         if (mems_updated) {
2276                 spin_lock_irq(&callback_lock);
2277                 if (!on_dfl)
2278                         top_cpuset.mems_allowed = new_mems;
2279                 top_cpuset.effective_mems = new_mems;
2280                 spin_unlock_irq(&callback_lock);
2281                 update_tasks_nodemask(&top_cpuset);
2282         }
2283 
2284         mutex_unlock(&cpuset_mutex);
2285 
2286         /* if cpus or mems changed, we need to propagate to descendants */
2287         if (cpus_updated || mems_updated) {
2288                 struct cpuset *cs;
2289                 struct cgroup_subsys_state *pos_css;
2290 
2291                 rcu_read_lock();
2292                 cpuset_for_each_descendant_pre(cs, pos_css, &top_cpuset) {
2293                         if (cs == &top_cpuset || !css_tryget_online(&cs->css))
2294                                 continue;
2295                         rcu_read_unlock();
2296 
2297                         cpuset_hotplug_update_tasks(cs);
2298 
2299                         rcu_read_lock();
2300                         css_put(&cs->css);
2301                 }
2302                 rcu_read_unlock();
2303         }
2304 
2305         /* rebuild sched domains if cpus_allowed has changed */
2306         if (cpus_updated)
2307                 rebuild_sched_domains();
2308 }
2309 
2310 void cpuset_update_active_cpus(bool cpu_online)
2311 {
2312         /*
2313          * We're inside cpu hotplug critical region which usually nests
2314          * inside cgroup synchronization.  Bounce actual hotplug processing
2315          * to a work item to avoid reverse locking order.
2316          *
2317          * We still need to do partition_sched_domains() synchronously;
2318          * otherwise, the scheduler will get confused and put tasks to the
2319          * dead CPU.  Fall back to the default single domain.
2320          * cpuset_hotplug_workfn() will rebuild it as necessary.
2321          */
2322         partition_sched_domains(1, NULL, NULL);
2323         schedule_work(&cpuset_hotplug_work);
2324 }
2325 
2326 /*
2327  * Keep top_cpuset.mems_allowed tracking node_states[N_MEMORY].
2328  * Call this routine anytime after node_states[N_MEMORY] changes.
2329  * See cpuset_update_active_cpus() for CPU hotplug handling.
2330  */
2331 static int cpuset_track_online_nodes(struct notifier_block *self,
2332                                 unsigned long action, void *arg)
2333 {
2334         schedule_work(&cpuset_hotplug_work);
2335         return NOTIFY_OK;
2336 }
2337 
2338 static struct notifier_block cpuset_track_online_nodes_nb = {
2339         .notifier_call = cpuset_track_online_nodes,
2340         .priority = 10,         /* ??! */
2341 };
2342 
2343 /**
2344  * cpuset_init_smp - initialize cpus_allowed
2345  *
2346  * Description: Finish top cpuset after cpu, node maps are initialized
2347  */
2348 void __init cpuset_init_smp(void)
2349 {
2350         cpumask_copy(top_cpuset.cpus_allowed, cpu_active_mask);
2351         top_cpuset.mems_allowed = node_states[N_MEMORY];
2352         top_cpuset.old_mems_allowed = top_cpuset.mems_allowed;
2353 
2354         cpumask_copy(top_cpuset.effective_cpus, cpu_active_mask);
2355         top_cpuset.effective_mems = node_states[N_MEMORY];
2356 
2357         register_hotmemory_notifier(&cpuset_track_online_nodes_nb);
2358 }
2359 
2360 /**
2361  * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
2362  * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
2363  * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
2364  *
2365  * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
2366  * attached to the specified @tsk.  Guaranteed to return some non-empty
2367  * subset of cpu_online_mask, even if this means going outside the
2368  * tasks cpuset.
2369  **/
2370 
2371 void cpuset_cpus_allowed(struct task_struct *tsk, struct cpumask *pmask)
2372 {
2373         unsigned long flags;
2374 
2375         spin_lock_irqsave(&callback_lock, flags);
2376         rcu_read_lock();
2377         guarantee_online_cpus(task_cs(tsk), pmask);
2378         rcu_read_unlock();
2379         spin_unlock_irqrestore(&callback_lock, flags);
2380 }
2381 
2382 void cpuset_cpus_allowed_fallback(struct task_struct *tsk)
2383 {
2384         rcu_read_lock();
2385         do_set_cpus_allowed(tsk, task_cs(tsk)->effective_cpus);
2386         rcu_read_unlock();
2387 
2388         /*
2389          * We own tsk->cpus_allowed, nobody can change it under us.
2390          *
2391          * But we used cs && cs->cpus_allowed lockless and thus can
2392          * race with cgroup_attach_task() or update_cpumask() and get
2393          * the wrong tsk->cpus_allowed. However, both cases imply the
2394          * subsequent cpuset_change_cpumask()->set_cpus_allowed_ptr()
2395          * which takes task_rq_lock().
2396          *
2397          * If we are called after it dropped the lock we must see all
2398          * changes in tsk_cs()->cpus_allowed. Otherwise we can temporary
2399          * set any mask even if it is not right from task_cs() pov,
2400          * the pending set_cpus_allowed_ptr() will fix things.
2401          *
2402          * select_fallback_rq() will fix things ups and set cpu_possible_mask
2403          * if required.
2404          */
2405 }
2406 
2407 void __init cpuset_init_current_mems_allowed(void)
2408 {
2409         nodes_setall(current->mems_allowed);
2410 }
2411 
2412 /**
2413  * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
2414  * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
2415  *
2416  * Description: Returns the nodemask_t mems_allowed of the cpuset
2417  * attached to the specified @tsk.  Guaranteed to return some non-empty
2418  * subset of node_states[N_MEMORY], even if this means going outside the
2419  * tasks cpuset.
2420  **/
2421 
2422 nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
2423 {
2424         nodemask_t mask;
2425         unsigned long flags;
2426 
2427         spin_lock_irqsave(&callback_lock, flags);
2428         rcu_read_lock();
2429         guarantee_online_mems(task_cs(tsk), &mask);
2430         rcu_read_unlock();
2431         spin_unlock_irqrestore(&callback_lock, flags);
2432 
2433         return mask;
2434 }
2435 
2436 /**
2437  * cpuset_nodemask_valid_mems_allowed - check nodemask vs. curremt mems_allowed
2438  * @nodemask: the nodemask to be checked
2439  *
2440  * Are any of the nodes in the nodemask allowed in current->mems_allowed?
2441  */
2442 int cpuset_nodemask_valid_mems_allowed(nodemask_t *nodemask)
2443 {
2444         return nodes_intersects(*nodemask, current->mems_allowed);
2445 }
2446 
2447 /*
2448  * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
2449  * mem_hardwall ancestor to the specified cpuset.  Call holding
2450  * callback_lock.  If no ancestor is mem_exclusive or mem_hardwall
2451  * (an unusual configuration), then returns the root cpuset.
2452  */
2453 static struct cpuset *nearest_hardwall_ancestor(struct cpuset *cs)
2454 {
2455         while (!(is_mem_exclusive(cs) || is_mem_hardwall(cs)) && parent_cs(cs))
2456                 cs = parent_cs(cs);
2457         return cs;
2458 }
2459 
2460 /**
2461  * cpuset_node_allowed - Can we allocate on a memory node?
2462  * @node: is this an allowed node?
2463  * @gfp_mask: memory allocation flags
2464  *
2465  * If we're in interrupt, yes, we can always allocate.  If @node is set in
2466  * current's mems_allowed, yes.  If it's not a __GFP_HARDWALL request and this
2467  * node is set in the nearest hardwalled cpuset ancestor to current's cpuset,
2468  * yes.  If current has access to memory reserves due to TIF_MEMDIE, yes.
2469  * Otherwise, no.
2470  *
2471  * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
2472  * and do not allow allocations outside the current tasks cpuset
2473  * unless the task has been OOM killed as is marked TIF_MEMDIE.
2474  * GFP_KERNEL allocations are not so marked, so can escape to the
2475  * nearest enclosing hardwalled ancestor cpuset.
2476  *
2477  * Scanning up parent cpusets requires callback_lock.  The
2478  * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
2479  * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
2480  * current tasks mems_allowed came up empty on the first pass over
2481  * the zonelist.  So only GFP_KERNEL allocations, if all nodes in the
2482  * cpuset are short of memory, might require taking the callback_lock.
2483  *
2484  * The first call here from mm/page_alloc:get_page_from_freelist()
2485  * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
2486  * so no allocation on a node outside the cpuset is allowed (unless
2487  * in interrupt, of course).
2488  *
2489  * The second pass through get_page_from_freelist() doesn't even call
2490  * here for GFP_ATOMIC calls.  For those calls, the __alloc_pages()
2491  * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
2492  * in alloc_flags.  That logic and the checks below have the combined
2493  * affect that:
2494  *      in_interrupt - any node ok (current task context irrelevant)
2495  *      GFP_ATOMIC   - any node ok
2496  *      TIF_MEMDIE   - any node ok
2497  *      GFP_KERNEL   - any node in enclosing hardwalled cpuset ok
2498  *      GFP_USER     - only nodes in current tasks mems allowed ok.
2499  */
2500 int __cpuset_node_allowed(int node, gfp_t gfp_mask)
2501 {
2502         struct cpuset *cs;              /* current cpuset ancestors */
2503         int allowed;                    /* is allocation in zone z allowed? */
2504         unsigned long flags;
2505 
2506         if (in_interrupt())
2507                 return 1;
2508         if (node_isset(node, current->mems_allowed))
2509                 return 1;
2510         /*
2511          * Allow tasks that have access to memory reserves because they have
2512          * been OOM killed to get memory anywhere.
2513          */
2514         if (unlikely(test_thread_flag(TIF_MEMDIE)))
2515                 return 1;
2516         if (gfp_mask & __GFP_HARDWALL)  /* If hardwall request, stop here */
2517                 return 0;
2518 
2519         if (current->flags & PF_EXITING) /* Let dying task have memory */
2520                 return 1;
2521 
2522         /* Not hardwall and node outside mems_allowed: scan up cpusets */
2523         spin_lock_irqsave(&callback_lock, flags);
2524 
2525         rcu_read_lock();
2526         cs = nearest_hardwall_ancestor(task_cs(current));
2527         allowed = node_isset(node, cs->mems_allowed);
2528         rcu_read_unlock();
2529 
2530         spin_unlock_irqrestore(&callback_lock, flags);
2531         return allowed;
2532 }
2533 
2534 /**
2535  * cpuset_mem_spread_node() - On which node to begin search for a file page
2536  * cpuset_slab_spread_node() - On which node to begin search for a slab page
2537  *
2538  * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
2539  * tasks in a cpuset with is_spread_page or is_spread_slab set),
2540  * and if the memory allocation used cpuset_mem_spread_node()
2541  * to determine on which node to start looking, as it will for
2542  * certain page cache or slab cache pages such as used for file
2543  * system buffers and inode caches, then instead of starting on the
2544  * local node to look for a free page, rather spread the starting
2545  * node around the tasks mems_allowed nodes.
2546  *
2547  * We don't have to worry about the returned node being offline
2548  * because "it can't happen", and even if it did, it would be ok.
2549  *
2550  * The routines calling guarantee_online_mems() are careful to
2551  * only set nodes in task->mems_allowed that are online.  So it
2552  * should not be possible for the following code to return an
2553  * offline node.  But if it did, that would be ok, as this routine
2554  * is not returning the node where the allocation must be, only
2555  * the node where the search should start.  The zonelist passed to
2556  * __alloc_pages() will include all nodes.  If the slab allocator
2557  * is passed an offline node, it will fall back to the local node.
2558  * See kmem_cache_alloc_node().
2559  */
2560 
2561 static int cpuset_spread_node(int *rotor)
2562 {
2563         int node;
2564 
2565         node = next_node(*rotor, current->mems_allowed);
2566         if (node == MAX_NUMNODES)
2567                 node = first_node(current->mems_allowed);
2568         *rotor = node;
2569         return node;
2570 }
2571 
2572 int cpuset_mem_spread_node(void)
2573 {
2574         if (current->cpuset_mem_spread_rotor == NUMA_NO_NODE)
2575                 current->cpuset_mem_spread_rotor =
2576                         node_random(&current->mems_allowed);
2577 
2578         return cpuset_spread_node(&current->cpuset_mem_spread_rotor);
2579 }
2580 
2581 int cpuset_slab_spread_node(void)
2582 {
2583         if (current->cpuset_slab_spread_rotor == NUMA_NO_NODE)
2584                 current->cpuset_slab_spread_rotor =
2585                         node_random(&current->mems_allowed);
2586 
2587         return cpuset_spread_node(&current->cpuset_slab_spread_rotor);
2588 }
2589 
2590 EXPORT_SYMBOL_GPL(cpuset_mem_spread_node);
2591 
2592 /**
2593  * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
2594  * @tsk1: pointer to task_struct of some task.
2595  * @tsk2: pointer to task_struct of some other task.
2596  *
2597  * Description: Return true if @tsk1's mems_allowed intersects the
2598  * mems_allowed of @tsk2.  Used by the OOM killer to determine if
2599  * one of the task's memory usage might impact the memory available
2600  * to the other.
2601  **/
2602 
2603 int cpuset_mems_allowed_intersects(const struct task_struct *tsk1,
2604                                    const struct task_struct *tsk2)
2605 {
2606         return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed);
2607 }
2608 
2609 /**
2610  * cpuset_print_current_mems_allowed - prints current's cpuset and mems_allowed
2611  *
2612  * Description: Prints current's name, cpuset name, and cached copy of its
2613  * mems_allowed to the kernel log.
2614  */
2615 void cpuset_print_current_mems_allowed(void)
2616 {
2617         struct cgroup *cgrp;
2618 
2619         rcu_read_lock();
2620 
2621         cgrp = task_cs(current)->css.cgroup;
2622         pr_info("%s cpuset=", current->comm);
2623         pr_cont_cgroup_name(cgrp);
2624         pr_cont(" mems_allowed=%*pbl\n",
2625                 nodemask_pr_args(&current->mems_allowed));
2626 
2627         rcu_read_unlock();
2628 }
2629 
2630 /*
2631  * Collection of memory_pressure is suppressed unless
2632  * this flag is enabled by writing "1" to the special
2633  * cpuset file 'memory_pressure_enabled' in the root cpuset.
2634  */
2635 
2636 int cpuset_memory_pressure_enabled __read_mostly;
2637 
2638 /**
2639  * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
2640  *
2641  * Keep a running average of the rate of synchronous (direct)
2642  * page reclaim efforts initiated by tasks in each cpuset.
2643  *
2644  * This represents the rate at which some task in the cpuset
2645  * ran low on memory on all nodes it was allowed to use, and
2646  * had to enter the kernels page reclaim code in an effort to
2647  * create more free memory by tossing clean pages or swapping
2648  * or writing dirty pages.
2649  *
2650  * Display to user space in the per-cpuset read-only file
2651  * "memory_pressure".  Value displayed is an integer
2652  * representing the recent rate of entry into the synchronous
2653  * (direct) page reclaim by any task attached to the cpuset.
2654  **/
2655 
2656 void __cpuset_memory_pressure_bump(void)
2657 {
2658         rcu_read_lock();
2659         fmeter_markevent(&task_cs(current)->fmeter);
2660         rcu_read_unlock();
2661 }
2662 
2663 #ifdef CONFIG_PROC_PID_CPUSET
2664 /*
2665  * proc_cpuset_show()
2666  *  - Print tasks cpuset path into seq_file.
2667  *  - Used for /proc/<pid>/cpuset.
2668  *  - No need to task_lock(tsk) on this tsk->cpuset reference, as it
2669  *    doesn't really matter if tsk->cpuset changes after we read it,
2670  *    and we take cpuset_mutex, keeping cpuset_attach() from changing it
2671  *    anyway.
2672  */
2673 int proc_cpuset_show(struct seq_file *m, struct pid_namespace *ns,
2674                      struct pid *pid, struct task_struct *tsk)
2675 {
2676         char *buf, *p;
2677         struct cgroup_subsys_state *css;
2678         int retval;
2679 
2680         retval = -ENOMEM;
2681         buf = kmalloc(PATH_MAX, GFP_KERNEL);
2682         if (!buf)
2683                 goto out;
2684 
2685         retval = -ENAMETOOLONG;
2686         rcu_read_lock();
2687         css = task_css(tsk, cpuset_cgrp_id);
2688         p = cgroup_path(css->cgroup, buf, PATH_MAX);
2689         rcu_read_unlock();
2690         if (!p)
2691                 goto out_free;
2692         seq_puts(m, p);
2693         seq_putc(m, '\n');
2694         retval = 0;
2695 out_free:
2696         kfree(buf);
2697 out:
2698         return retval;
2699 }
2700 #endif /* CONFIG_PROC_PID_CPUSET */
2701 
2702 /* Display task mems_allowed in /proc/<pid>/status file. */
2703 void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task)
2704 {
2705         seq_printf(m, "Mems_allowed:\t%*pb\n",
2706                    nodemask_pr_args(&task->mems_allowed));
2707         seq_printf(m, "Mems_allowed_list:\t%*pbl\n",
2708                    nodemask_pr_args(&task->mems_allowed));
2709 }
2710 

This page was automatically generated by LXR 0.3.1 (source).  •  Linux is a registered trademark of Linus Torvalds  •  Contact us