Version:  2.0.40 2.2.26 2.4.37 3.3 3.4 3.5 3.6 3.7 3.8 3.9 3.10 3.11 3.12 3.13 3.14 3.15 3.16 3.17 3.18 3.19

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_on_dfl(cur->css.cgroup) && !is_cpuset_subset(trial, par))
477                 goto out;
478 
479         /*
480          * If either I or some sibling (!= me) is exclusive, we can't
481          * overlap
482          */
483         ret = -EINVAL;
484         cpuset_for_each_child(c, css, par) {
485                 if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
486                     c != cur &&
487                     cpumask_intersects(trial->cpus_allowed, c->cpus_allowed))
488                         goto out;
489                 if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) &&
490                     c != cur &&
491                     nodes_intersects(trial->mems_allowed, c->mems_allowed))
492                         goto out;
493         }
494 
495         /*
496          * Cpusets with tasks - existing or newly being attached - can't
497          * be changed to have empty cpus_allowed or mems_allowed.
498          */
499         ret = -ENOSPC;
500         if ((cgroup_has_tasks(cur->css.cgroup) || cur->attach_in_progress)) {
501                 if (!cpumask_empty(cur->cpus_allowed) &&
502                     cpumask_empty(trial->cpus_allowed))
503                         goto out;
504                 if (!nodes_empty(cur->mems_allowed) &&
505                     nodes_empty(trial->mems_allowed))
506                         goto out;
507         }
508 
509         /*
510          * We can't shrink if we won't have enough room for SCHED_DEADLINE
511          * tasks.
512          */
513         ret = -EBUSY;
514         if (is_cpu_exclusive(cur) &&
515             !cpuset_cpumask_can_shrink(cur->cpus_allowed,
516                                        trial->cpus_allowed))
517                 goto out;
518 
519         ret = 0;
520 out:
521         rcu_read_unlock();
522         return ret;
523 }
524 
525 #ifdef CONFIG_SMP
526 /*
527  * Helper routine for generate_sched_domains().
528  * Do cpusets a, b have overlapping effective cpus_allowed masks?
529  */
530 static int cpusets_overlap(struct cpuset *a, struct cpuset *b)
531 {
532         return cpumask_intersects(a->effective_cpus, b->effective_cpus);
533 }
534 
535 static void
536 update_domain_attr(struct sched_domain_attr *dattr, struct cpuset *c)
537 {
538         if (dattr->relax_domain_level < c->relax_domain_level)
539                 dattr->relax_domain_level = c->relax_domain_level;
540         return;
541 }
542 
543 static void update_domain_attr_tree(struct sched_domain_attr *dattr,
544                                     struct cpuset *root_cs)
545 {
546         struct cpuset *cp;
547         struct cgroup_subsys_state *pos_css;
548 
549         rcu_read_lock();
550         cpuset_for_each_descendant_pre(cp, pos_css, root_cs) {
551                 if (cp == root_cs)
552                         continue;
553 
554                 /* skip the whole subtree if @cp doesn't have any CPU */
555                 if (cpumask_empty(cp->cpus_allowed)) {
556                         pos_css = css_rightmost_descendant(pos_css);
557                         continue;
558                 }
559 
560                 if (is_sched_load_balance(cp))
561                         update_domain_attr(dattr, cp);
562         }
563         rcu_read_unlock();
564 }
565 
566 /*
567  * generate_sched_domains()
568  *
569  * This function builds a partial partition of the systems CPUs
570  * A 'partial partition' is a set of non-overlapping subsets whose
571  * union is a subset of that set.
572  * The output of this function needs to be passed to kernel/sched/core.c
573  * partition_sched_domains() routine, which will rebuild the scheduler's
574  * load balancing domains (sched domains) as specified by that partial
575  * partition.
576  *
577  * See "What is sched_load_balance" in Documentation/cgroups/cpusets.txt
578  * for a background explanation of this.
579  *
580  * Does not return errors, on the theory that the callers of this
581  * routine would rather not worry about failures to rebuild sched
582  * domains when operating in the severe memory shortage situations
583  * that could cause allocation failures below.
584  *
585  * Must be called with cpuset_mutex held.
586  *
587  * The three key local variables below are:
588  *    q  - a linked-list queue of cpuset pointers, used to implement a
589  *         top-down scan of all cpusets.  This scan loads a pointer
590  *         to each cpuset marked is_sched_load_balance into the
591  *         array 'csa'.  For our purposes, rebuilding the schedulers
592  *         sched domains, we can ignore !is_sched_load_balance cpusets.
593  *  csa  - (for CpuSet Array) Array of pointers to all the cpusets
594  *         that need to be load balanced, for convenient iterative
595  *         access by the subsequent code that finds the best partition,
596  *         i.e the set of domains (subsets) of CPUs such that the
597  *         cpus_allowed of every cpuset marked is_sched_load_balance
598  *         is a subset of one of these domains, while there are as
599  *         many such domains as possible, each as small as possible.
600  * doms  - Conversion of 'csa' to an array of cpumasks, for passing to
601  *         the kernel/sched/core.c routine partition_sched_domains() in a
602  *         convenient format, that can be easily compared to the prior
603  *         value to determine what partition elements (sched domains)
604  *         were changed (added or removed.)
605  *
606  * Finding the best partition (set of domains):
607  *      The triple nested loops below over i, j, k scan over the
608  *      load balanced cpusets (using the array of cpuset pointers in
609  *      csa[]) looking for pairs of cpusets that have overlapping
610  *      cpus_allowed, but which don't have the same 'pn' partition
611  *      number and gives them in the same partition number.  It keeps
612  *      looping on the 'restart' label until it can no longer find
613  *      any such pairs.
614  *
615  *      The union of the cpus_allowed masks from the set of
616  *      all cpusets having the same 'pn' value then form the one
617  *      element of the partition (one sched domain) to be passed to
618  *      partition_sched_domains().
619  */
620 static int generate_sched_domains(cpumask_var_t **domains,
621                         struct sched_domain_attr **attributes)
622 {
623         struct cpuset *cp;      /* scans q */
624         struct cpuset **csa;    /* array of all cpuset ptrs */
625         int csn;                /* how many cpuset ptrs in csa so far */
626         int i, j, k;            /* indices for partition finding loops */
627         cpumask_var_t *doms;    /* resulting partition; i.e. sched domains */
628         struct sched_domain_attr *dattr;  /* attributes for custom domains */
629         int ndoms = 0;          /* number of sched domains in result */
630         int nslot;              /* next empty doms[] struct cpumask slot */
631         struct cgroup_subsys_state *pos_css;
632 
633         doms = NULL;
634         dattr = NULL;
635         csa = NULL;
636 
637         /* Special case for the 99% of systems with one, full, sched domain */
638         if (is_sched_load_balance(&top_cpuset)) {
639                 ndoms = 1;
640                 doms = alloc_sched_domains(ndoms);
641                 if (!doms)
642                         goto done;
643 
644                 dattr = kmalloc(sizeof(struct sched_domain_attr), GFP_KERNEL);
645                 if (dattr) {
646                         *dattr = SD_ATTR_INIT;
647                         update_domain_attr_tree(dattr, &top_cpuset);
648                 }
649                 cpumask_copy(doms[0], top_cpuset.effective_cpus);
650 
651                 goto done;
652         }
653 
654         csa = kmalloc(nr_cpusets() * sizeof(cp), GFP_KERNEL);
655         if (!csa)
656                 goto done;
657         csn = 0;
658 
659         rcu_read_lock();
660         cpuset_for_each_descendant_pre(cp, pos_css, &top_cpuset) {
661                 if (cp == &top_cpuset)
662                         continue;
663                 /*
664                  * Continue traversing beyond @cp iff @cp has some CPUs and
665                  * isn't load balancing.  The former is obvious.  The
666                  * latter: All child cpusets contain a subset of the
667                  * parent's cpus, so just skip them, and then we call
668                  * update_domain_attr_tree() to calc relax_domain_level of
669                  * the corresponding sched domain.
670                  */
671                 if (!cpumask_empty(cp->cpus_allowed) &&
672                     !is_sched_load_balance(cp))
673                         continue;
674 
675                 if (is_sched_load_balance(cp))
676                         csa[csn++] = cp;
677 
678                 /* skip @cp's subtree */
679                 pos_css = css_rightmost_descendant(pos_css);
680         }
681         rcu_read_unlock();
682 
683         for (i = 0; i < csn; i++)
684                 csa[i]->pn = i;
685         ndoms = csn;
686 
687 restart:
688         /* Find the best partition (set of sched domains) */
689         for (i = 0; i < csn; i++) {
690                 struct cpuset *a = csa[i];
691                 int apn = a->pn;
692 
693                 for (j = 0; j < csn; j++) {
694                         struct cpuset *b = csa[j];
695                         int bpn = b->pn;
696 
697                         if (apn != bpn && cpusets_overlap(a, b)) {
698                                 for (k = 0; k < csn; k++) {
699                                         struct cpuset *c = csa[k];
700 
701                                         if (c->pn == bpn)
702                                                 c->pn = apn;
703                                 }
704                                 ndoms--;        /* one less element */
705                                 goto restart;
706                         }
707                 }
708         }
709 
710         /*
711          * Now we know how many domains to create.
712          * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
713          */
714         doms = alloc_sched_domains(ndoms);
715         if (!doms)
716                 goto done;
717 
718         /*
719          * The rest of the code, including the scheduler, can deal with
720          * dattr==NULL case. No need to abort if alloc fails.
721          */
722         dattr = kmalloc(ndoms * sizeof(struct sched_domain_attr), GFP_KERNEL);
723 
724         for (nslot = 0, i = 0; i < csn; i++) {
725                 struct cpuset *a = csa[i];
726                 struct cpumask *dp;
727                 int apn = a->pn;
728 
729                 if (apn < 0) {
730                         /* Skip completed partitions */
731                         continue;
732                 }
733 
734                 dp = doms[nslot];
735 
736                 if (nslot == ndoms) {
737                         static int warnings = 10;
738                         if (warnings) {
739                                 pr_warn("rebuild_sched_domains confused: nslot %d, ndoms %d, csn %d, i %d, apn %d\n",
740                                         nslot, ndoms, csn, i, apn);
741                                 warnings--;
742                         }
743                         continue;
744                 }
745 
746                 cpumask_clear(dp);
747                 if (dattr)
748                         *(dattr + nslot) = SD_ATTR_INIT;
749                 for (j = i; j < csn; j++) {
750                         struct cpuset *b = csa[j];
751 
752                         if (apn == b->pn) {
753                                 cpumask_or(dp, dp, b->effective_cpus);
754                                 if (dattr)
755                                         update_domain_attr_tree(dattr + nslot, b);
756 
757                                 /* Done with this partition */
758                                 b->pn = -1;
759                         }
760                 }
761                 nslot++;
762         }
763         BUG_ON(nslot != ndoms);
764 
765 done:
766         kfree(csa);
767 
768         /*
769          * Fallback to the default domain if kmalloc() failed.
770          * See comments in partition_sched_domains().
771          */
772         if (doms == NULL)
773                 ndoms = 1;
774 
775         *domains    = doms;
776         *attributes = dattr;
777         return ndoms;
778 }
779 
780 /*
781  * Rebuild scheduler domains.
782  *
783  * If the flag 'sched_load_balance' of any cpuset with non-empty
784  * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
785  * which has that flag enabled, or if any cpuset with a non-empty
786  * 'cpus' is removed, then call this routine to rebuild the
787  * scheduler's dynamic sched domains.
788  *
789  * Call with cpuset_mutex held.  Takes get_online_cpus().
790  */
791 static void rebuild_sched_domains_locked(void)
792 {
793         struct sched_domain_attr *attr;
794         cpumask_var_t *doms;
795         int ndoms;
796 
797         lockdep_assert_held(&cpuset_mutex);
798         get_online_cpus();
799 
800         /*
801          * We have raced with CPU hotplug. Don't do anything to avoid
802          * passing doms with offlined cpu to partition_sched_domains().
803          * Anyways, hotplug work item will rebuild sched domains.
804          */
805         if (!cpumask_equal(top_cpuset.effective_cpus, cpu_active_mask))
806                 goto out;
807 
808         /* Generate domain masks and attrs */
809         ndoms = generate_sched_domains(&doms, &attr);
810 
811         /* Have scheduler rebuild the domains */
812         partition_sched_domains(ndoms, doms, attr);
813 out:
814         put_online_cpus();
815 }
816 #else /* !CONFIG_SMP */
817 static void rebuild_sched_domains_locked(void)
818 {
819 }
820 #endif /* CONFIG_SMP */
821 
822 void rebuild_sched_domains(void)
823 {
824         mutex_lock(&cpuset_mutex);
825         rebuild_sched_domains_locked();
826         mutex_unlock(&cpuset_mutex);
827 }
828 
829 /**
830  * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
831  * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
832  *
833  * Iterate through each task of @cs updating its cpus_allowed to the
834  * effective cpuset's.  As this function is called with cpuset_mutex held,
835  * cpuset membership stays stable.
836  */
837 static void update_tasks_cpumask(struct cpuset *cs)
838 {
839         struct css_task_iter it;
840         struct task_struct *task;
841 
842         css_task_iter_start(&cs->css, &it);
843         while ((task = css_task_iter_next(&it)))
844                 set_cpus_allowed_ptr(task, cs->effective_cpus);
845         css_task_iter_end(&it);
846 }
847 
848 /*
849  * update_cpumasks_hier - Update effective cpumasks and tasks in the subtree
850  * @cs: the cpuset to consider
851  * @new_cpus: temp variable for calculating new effective_cpus
852  *
853  * When congifured cpumask is changed, the effective cpumasks of this cpuset
854  * and all its descendants need to be updated.
855  *
856  * On legacy hierachy, effective_cpus will be the same with cpu_allowed.
857  *
858  * Called with cpuset_mutex held
859  */
860 static void update_cpumasks_hier(struct cpuset *cs, struct cpumask *new_cpus)
861 {
862         struct cpuset *cp;
863         struct cgroup_subsys_state *pos_css;
864         bool need_rebuild_sched_domains = false;
865 
866         rcu_read_lock();
867         cpuset_for_each_descendant_pre(cp, pos_css, cs) {
868                 struct cpuset *parent = parent_cs(cp);
869 
870                 cpumask_and(new_cpus, cp->cpus_allowed, parent->effective_cpus);
871 
872                 /*
873                  * If it becomes empty, inherit the effective mask of the
874                  * parent, which is guaranteed to have some CPUs.
875                  */
876                 if (cpumask_empty(new_cpus))
877                         cpumask_copy(new_cpus, parent->effective_cpus);
878 
879                 /* Skip the whole subtree if the cpumask remains the same. */
880                 if (cpumask_equal(new_cpus, cp->effective_cpus)) {
881                         pos_css = css_rightmost_descendant(pos_css);
882                         continue;
883                 }
884 
885                 if (!css_tryget_online(&cp->css))
886                         continue;
887                 rcu_read_unlock();
888 
889                 spin_lock_irq(&callback_lock);
890                 cpumask_copy(cp->effective_cpus, new_cpus);
891                 spin_unlock_irq(&callback_lock);
892 
893                 WARN_ON(!cgroup_on_dfl(cp->css.cgroup) &&
894                         !cpumask_equal(cp->cpus_allowed, cp->effective_cpus));
895 
896                 update_tasks_cpumask(cp);
897 
898                 /*
899                  * If the effective cpumask of any non-empty cpuset is changed,
900                  * we need to rebuild sched domains.
901                  */
902                 if (!cpumask_empty(cp->cpus_allowed) &&
903                     is_sched_load_balance(cp))
904                         need_rebuild_sched_domains = true;
905 
906                 rcu_read_lock();
907                 css_put(&cp->css);
908         }
909         rcu_read_unlock();
910 
911         if (need_rebuild_sched_domains)
912                 rebuild_sched_domains_locked();
913 }
914 
915 /**
916  * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
917  * @cs: the cpuset to consider
918  * @trialcs: trial cpuset
919  * @buf: buffer of cpu numbers written to this cpuset
920  */
921 static int update_cpumask(struct cpuset *cs, struct cpuset *trialcs,
922                           const char *buf)
923 {
924         int retval;
925 
926         /* top_cpuset.cpus_allowed tracks cpu_online_mask; it's read-only */
927         if (cs == &top_cpuset)
928                 return -EACCES;
929 
930         /*
931          * An empty cpus_allowed is ok only if the cpuset has no tasks.
932          * Since cpulist_parse() fails on an empty mask, we special case
933          * that parsing.  The validate_change() call ensures that cpusets
934          * with tasks have cpus.
935          */
936         if (!*buf) {
937                 cpumask_clear(trialcs->cpus_allowed);
938         } else {
939                 retval = cpulist_parse(buf, trialcs->cpus_allowed);
940                 if (retval < 0)
941                         return retval;
942 
943                 if (!cpumask_subset(trialcs->cpus_allowed,
944                                     top_cpuset.cpus_allowed))
945                         return -EINVAL;
946         }
947 
948         /* Nothing to do if the cpus didn't change */
949         if (cpumask_equal(cs->cpus_allowed, trialcs->cpus_allowed))
950                 return 0;
951 
952         retval = validate_change(cs, trialcs);
953         if (retval < 0)
954                 return retval;
955 
956         spin_lock_irq(&callback_lock);
957         cpumask_copy(cs->cpus_allowed, trialcs->cpus_allowed);
958         spin_unlock_irq(&callback_lock);
959 
960         /* use trialcs->cpus_allowed as a temp variable */
961         update_cpumasks_hier(cs, trialcs->cpus_allowed);
962         return 0;
963 }
964 
965 /*
966  * cpuset_migrate_mm
967  *
968  *    Migrate memory region from one set of nodes to another.
969  *
970  *    Temporarilly set tasks mems_allowed to target nodes of migration,
971  *    so that the migration code can allocate pages on these nodes.
972  *
973  *    While the mm_struct we are migrating is typically from some
974  *    other task, the task_struct mems_allowed that we are hacking
975  *    is for our current task, which must allocate new pages for that
976  *    migrating memory region.
977  */
978 
979 static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
980                                                         const nodemask_t *to)
981 {
982         struct task_struct *tsk = current;
983 
984         tsk->mems_allowed = *to;
985 
986         do_migrate_pages(mm, from, to, MPOL_MF_MOVE_ALL);
987 
988         rcu_read_lock();
989         guarantee_online_mems(task_cs(tsk), &tsk->mems_allowed);
990         rcu_read_unlock();
991 }
992 
993 /*
994  * cpuset_change_task_nodemask - change task's mems_allowed and mempolicy
995  * @tsk: the task to change
996  * @newmems: new nodes that the task will be set
997  *
998  * In order to avoid seeing no nodes if the old and new nodes are disjoint,
999  * we structure updates as setting all new allowed nodes, then clearing newly
1000  * disallowed ones.
1001  */
1002 static void cpuset_change_task_nodemask(struct task_struct *tsk,
1003                                         nodemask_t *newmems)
1004 {
1005         bool need_loop;
1006 
1007         /*
1008          * Allow tasks that have access to memory reserves because they have
1009          * been OOM killed to get memory anywhere.
1010          */
1011         if (unlikely(test_thread_flag(TIF_MEMDIE)))
1012                 return;
1013         if (current->flags & PF_EXITING) /* Let dying task have memory */
1014                 return;
1015 
1016         task_lock(tsk);
1017         /*
1018          * Determine if a loop is necessary if another thread is doing
1019          * read_mems_allowed_begin().  If at least one node remains unchanged and
1020          * tsk does not have a mempolicy, then an empty nodemask will not be
1021          * possible when mems_allowed is larger than a word.
1022          */
1023         need_loop = task_has_mempolicy(tsk) ||
1024                         !nodes_intersects(*newmems, tsk->mems_allowed);
1025 
1026         if (need_loop) {
1027                 local_irq_disable();
1028                 write_seqcount_begin(&tsk->mems_allowed_seq);
1029         }
1030 
1031         nodes_or(tsk->mems_allowed, tsk->mems_allowed, *newmems);
1032         mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP1);
1033 
1034         mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP2);
1035         tsk->mems_allowed = *newmems;
1036 
1037         if (need_loop) {
1038                 write_seqcount_end(&tsk->mems_allowed_seq);
1039                 local_irq_enable();
1040         }
1041 
1042         task_unlock(tsk);
1043 }
1044 
1045 static void *cpuset_being_rebound;
1046 
1047 /**
1048  * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
1049  * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
1050  *
1051  * Iterate through each task of @cs updating its mems_allowed to the
1052  * effective cpuset's.  As this function is called with cpuset_mutex held,
1053  * cpuset membership stays stable.
1054  */
1055 static void update_tasks_nodemask(struct cpuset *cs)
1056 {
1057         static nodemask_t newmems;      /* protected by cpuset_mutex */
1058         struct css_task_iter it;
1059         struct task_struct *task;
1060 
1061         cpuset_being_rebound = cs;              /* causes mpol_dup() rebind */
1062 
1063         guarantee_online_mems(cs, &newmems);
1064 
1065         /*
1066          * The mpol_rebind_mm() call takes mmap_sem, which we couldn't
1067          * take while holding tasklist_lock.  Forks can happen - the
1068          * mpol_dup() cpuset_being_rebound check will catch such forks,
1069          * and rebind their vma mempolicies too.  Because we still hold
1070          * the global cpuset_mutex, we know that no other rebind effort
1071          * will be contending for the global variable cpuset_being_rebound.
1072          * It's ok if we rebind the same mm twice; mpol_rebind_mm()
1073          * is idempotent.  Also migrate pages in each mm to new nodes.
1074          */
1075         css_task_iter_start(&cs->css, &it);
1076         while ((task = css_task_iter_next(&it))) {
1077                 struct mm_struct *mm;
1078                 bool migrate;
1079 
1080                 cpuset_change_task_nodemask(task, &newmems);
1081 
1082                 mm = get_task_mm(task);
1083                 if (!mm)
1084                         continue;
1085 
1086                 migrate = is_memory_migrate(cs);
1087 
1088                 mpol_rebind_mm(mm, &cs->mems_allowed);
1089                 if (migrate)
1090                         cpuset_migrate_mm(mm, &cs->old_mems_allowed, &newmems);
1091                 mmput(mm);
1092         }
1093         css_task_iter_end(&it);
1094 
1095         /*
1096          * All the tasks' nodemasks have been updated, update
1097          * cs->old_mems_allowed.
1098          */
1099         cs->old_mems_allowed = newmems;
1100 
1101         /* We're done rebinding vmas to this cpuset's new mems_allowed. */
1102         cpuset_being_rebound = NULL;
1103 }
1104 
1105 /*
1106  * update_nodemasks_hier - Update effective nodemasks and tasks in the subtree
1107  * @cs: the cpuset to consider
1108  * @new_mems: a temp variable for calculating new effective_mems
1109  *
1110  * When configured nodemask is changed, the effective nodemasks of this cpuset
1111  * and all its descendants need to be updated.
1112  *
1113  * On legacy hiearchy, effective_mems will be the same with mems_allowed.
1114  *
1115  * Called with cpuset_mutex held
1116  */
1117 static void update_nodemasks_hier(struct cpuset *cs, nodemask_t *new_mems)
1118 {
1119         struct cpuset *cp;
1120         struct cgroup_subsys_state *pos_css;
1121 
1122         rcu_read_lock();
1123         cpuset_for_each_descendant_pre(cp, pos_css, cs) {
1124                 struct cpuset *parent = parent_cs(cp);
1125 
1126                 nodes_and(*new_mems, cp->mems_allowed, parent->effective_mems);
1127 
1128                 /*
1129                  * If it becomes empty, inherit the effective mask of the
1130                  * parent, which is guaranteed to have some MEMs.
1131                  */
1132                 if (nodes_empty(*new_mems))
1133                         *new_mems = parent->effective_mems;
1134 
1135                 /* Skip the whole subtree if the nodemask remains the same. */
1136                 if (nodes_equal(*new_mems, cp->effective_mems)) {
1137                         pos_css = css_rightmost_descendant(pos_css);
1138                         continue;
1139                 }
1140 
1141                 if (!css_tryget_online(&cp->css))
1142                         continue;
1143                 rcu_read_unlock();
1144 
1145                 spin_lock_irq(&callback_lock);
1146                 cp->effective_mems = *new_mems;
1147                 spin_unlock_irq(&callback_lock);
1148 
1149                 WARN_ON(!cgroup_on_dfl(cp->css.cgroup) &&
1150                         !nodes_equal(cp->mems_allowed, cp->effective_mems));
1151 
1152                 update_tasks_nodemask(cp);
1153 
1154                 rcu_read_lock();
1155                 css_put(&cp->css);
1156         }
1157         rcu_read_unlock();
1158 }
1159 
1160 /*
1161  * Handle user request to change the 'mems' memory placement
1162  * of a cpuset.  Needs to validate the request, update the
1163  * cpusets mems_allowed, and for each task in the cpuset,
1164  * update mems_allowed and rebind task's mempolicy and any vma
1165  * mempolicies and if the cpuset is marked 'memory_migrate',
1166  * migrate the tasks pages to the new memory.
1167  *
1168  * Call with cpuset_mutex held. May take callback_lock during call.
1169  * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
1170  * lock each such tasks mm->mmap_sem, scan its vma's and rebind
1171  * their mempolicies to the cpusets new mems_allowed.
1172  */
1173 static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
1174                            const char *buf)
1175 {
1176         int retval;
1177 
1178         /*
1179          * top_cpuset.mems_allowed tracks node_stats[N_MEMORY];
1180          * it's read-only
1181          */
1182         if (cs == &top_cpuset) {
1183                 retval = -EACCES;
1184                 goto done;
1185         }
1186 
1187         /*
1188          * An empty mems_allowed is ok iff there are no tasks in the cpuset.
1189          * Since nodelist_parse() fails on an empty mask, we special case
1190          * that parsing.  The validate_change() call ensures that cpusets
1191          * with tasks have memory.
1192          */
1193         if (!*buf) {
1194                 nodes_clear(trialcs->mems_allowed);
1195         } else {
1196                 retval = nodelist_parse(buf, trialcs->mems_allowed);
1197                 if (retval < 0)
1198                         goto done;
1199 
1200                 if (!nodes_subset(trialcs->mems_allowed,
1201                                   top_cpuset.mems_allowed)) {
1202                         retval = -EINVAL;
1203                         goto done;
1204                 }
1205         }
1206 
1207         if (nodes_equal(cs->mems_allowed, trialcs->mems_allowed)) {
1208                 retval = 0;             /* Too easy - nothing to do */
1209                 goto done;
1210         }
1211         retval = validate_change(cs, trialcs);
1212         if (retval < 0)
1213                 goto done;
1214 
1215         spin_lock_irq(&callback_lock);
1216         cs->mems_allowed = trialcs->mems_allowed;
1217         spin_unlock_irq(&callback_lock);
1218 
1219         /* use trialcs->mems_allowed as a temp variable */
1220         update_nodemasks_hier(cs, &cs->mems_allowed);
1221 done:
1222         return retval;
1223 }
1224 
1225 int current_cpuset_is_being_rebound(void)
1226 {
1227         int ret;
1228 
1229         rcu_read_lock();
1230         ret = task_cs(current) == cpuset_being_rebound;
1231         rcu_read_unlock();
1232 
1233         return ret;
1234 }
1235 
1236 static int update_relax_domain_level(struct cpuset *cs, s64 val)
1237 {
1238 #ifdef CONFIG_SMP
1239         if (val < -1 || val >= sched_domain_level_max)
1240                 return -EINVAL;
1241 #endif
1242 
1243         if (val != cs->relax_domain_level) {
1244                 cs->relax_domain_level = val;
1245                 if (!cpumask_empty(cs->cpus_allowed) &&
1246                     is_sched_load_balance(cs))
1247                         rebuild_sched_domains_locked();
1248         }
1249 
1250         return 0;
1251 }
1252 
1253 /**
1254  * update_tasks_flags - update the spread flags of tasks in the cpuset.
1255  * @cs: the cpuset in which each task's spread flags needs to be changed
1256  *
1257  * Iterate through each task of @cs updating its spread flags.  As this
1258  * function is called with cpuset_mutex held, cpuset membership stays
1259  * stable.
1260  */
1261 static void update_tasks_flags(struct cpuset *cs)
1262 {
1263         struct css_task_iter it;
1264         struct task_struct *task;
1265 
1266         css_task_iter_start(&cs->css, &it);
1267         while ((task = css_task_iter_next(&it)))
1268                 cpuset_update_task_spread_flag(cs, task);
1269         css_task_iter_end(&it);
1270 }
1271 
1272 /*
1273  * update_flag - read a 0 or a 1 in a file and update associated flag
1274  * bit:         the bit to update (see cpuset_flagbits_t)
1275  * cs:          the cpuset to update
1276  * turning_on:  whether the flag is being set or cleared
1277  *
1278  * Call with cpuset_mutex held.
1279  */
1280 
1281 static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs,
1282                        int turning_on)
1283 {
1284         struct cpuset *trialcs;
1285         int balance_flag_changed;
1286         int spread_flag_changed;
1287         int err;
1288 
1289         trialcs = alloc_trial_cpuset(cs);
1290         if (!trialcs)
1291                 return -ENOMEM;
1292 
1293         if (turning_on)
1294                 set_bit(bit, &trialcs->flags);
1295         else
1296                 clear_bit(bit, &trialcs->flags);
1297 
1298         err = validate_change(cs, trialcs);
1299         if (err < 0)
1300                 goto out;
1301 
1302         balance_flag_changed = (is_sched_load_balance(cs) !=
1303                                 is_sched_load_balance(trialcs));
1304 
1305         spread_flag_changed = ((is_spread_slab(cs) != is_spread_slab(trialcs))
1306                         || (is_spread_page(cs) != is_spread_page(trialcs)));
1307 
1308         spin_lock_irq(&callback_lock);
1309         cs->flags = trialcs->flags;
1310         spin_unlock_irq(&callback_lock);
1311 
1312         if (!cpumask_empty(trialcs->cpus_allowed) && balance_flag_changed)
1313                 rebuild_sched_domains_locked();
1314 
1315         if (spread_flag_changed)
1316                 update_tasks_flags(cs);
1317 out:
1318         free_trial_cpuset(trialcs);
1319         return err;
1320 }
1321 
1322 /*
1323  * Frequency meter - How fast is some event occurring?
1324  *
1325  * These routines manage a digitally filtered, constant time based,
1326  * event frequency meter.  There are four routines:
1327  *   fmeter_init() - initialize a frequency meter.
1328  *   fmeter_markevent() - called each time the event happens.
1329  *   fmeter_getrate() - returns the recent rate of such events.
1330  *   fmeter_update() - internal routine used to update fmeter.
1331  *
1332  * A common data structure is passed to each of these routines,
1333  * which is used to keep track of the state required to manage the
1334  * frequency meter and its digital filter.
1335  *
1336  * The filter works on the number of events marked per unit time.
1337  * The filter is single-pole low-pass recursive (IIR).  The time unit
1338  * is 1 second.  Arithmetic is done using 32-bit integers scaled to
1339  * simulate 3 decimal digits of precision (multiplied by 1000).
1340  *
1341  * With an FM_COEF of 933, and a time base of 1 second, the filter
1342  * has a half-life of 10 seconds, meaning that if the events quit
1343  * happening, then the rate returned from the fmeter_getrate()
1344  * will be cut in half each 10 seconds, until it converges to zero.
1345  *
1346  * It is not worth doing a real infinitely recursive filter.  If more
1347  * than FM_MAXTICKS ticks have elapsed since the last filter event,
1348  * just compute FM_MAXTICKS ticks worth, by which point the level
1349  * will be stable.
1350  *
1351  * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
1352  * arithmetic overflow in the fmeter_update() routine.
1353  *
1354  * Given the simple 32 bit integer arithmetic used, this meter works
1355  * best for reporting rates between one per millisecond (msec) and
1356  * one per 32 (approx) seconds.  At constant rates faster than one
1357  * per msec it maxes out at values just under 1,000,000.  At constant
1358  * rates between one per msec, and one per second it will stabilize
1359  * to a value N*1000, where N is the rate of events per second.
1360  * At constant rates between one per second and one per 32 seconds,
1361  * it will be choppy, moving up on the seconds that have an event,
1362  * and then decaying until the next event.  At rates slower than
1363  * about one in 32 seconds, it decays all the way back to zero between
1364  * each event.
1365  */
1366 
1367 #define FM_COEF 933             /* coefficient for half-life of 10 secs */
1368 #define FM_MAXTICKS ((time_t)99) /* useless computing more ticks than this */
1369 #define FM_MAXCNT 1000000       /* limit cnt to avoid overflow */
1370 #define FM_SCALE 1000           /* faux fixed point scale */
1371 
1372 /* Initialize a frequency meter */
1373 static void fmeter_init(struct fmeter *fmp)
1374 {
1375         fmp->cnt = 0;
1376         fmp->val = 0;
1377         fmp->time = 0;
1378         spin_lock_init(&fmp->lock);
1379 }
1380 
1381 /* Internal meter update - process cnt events and update value */
1382 static void fmeter_update(struct fmeter *fmp)
1383 {
1384         time_t now = get_seconds();
1385         time_t ticks = now - fmp->time;
1386 
1387         if (ticks == 0)
1388                 return;
1389 
1390         ticks = min(FM_MAXTICKS, ticks);
1391         while (ticks-- > 0)
1392                 fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
1393         fmp->time = now;
1394 
1395         fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
1396         fmp->cnt = 0;
1397 }
1398 
1399 /* Process any previous ticks, then bump cnt by one (times scale). */
1400 static void fmeter_markevent(struct fmeter *fmp)
1401 {
1402         spin_lock(&fmp->lock);
1403         fmeter_update(fmp);
1404         fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
1405         spin_unlock(&fmp->lock);
1406 }
1407 
1408 /* Process any previous ticks, then return current value. */
1409 static int fmeter_getrate(struct fmeter *fmp)
1410 {
1411         int val;
1412 
1413         spin_lock(&fmp->lock);
1414         fmeter_update(fmp);
1415         val = fmp->val;
1416         spin_unlock(&fmp->lock);
1417         return val;
1418 }
1419 
1420 static struct cpuset *cpuset_attach_old_cs;
1421 
1422 /* Called by cgroups to determine if a cpuset is usable; cpuset_mutex held */
1423 static int cpuset_can_attach(struct cgroup_subsys_state *css,
1424                              struct cgroup_taskset *tset)
1425 {
1426         struct cpuset *cs = css_cs(css);
1427         struct task_struct *task;
1428         int ret;
1429 
1430         /* used later by cpuset_attach() */
1431         cpuset_attach_old_cs = task_cs(cgroup_taskset_first(tset));
1432 
1433         mutex_lock(&cpuset_mutex);
1434 
1435         /* allow moving tasks into an empty cpuset if on default hierarchy */
1436         ret = -ENOSPC;
1437         if (!cgroup_on_dfl(css->cgroup) &&
1438             (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed)))
1439                 goto out_unlock;
1440 
1441         cgroup_taskset_for_each(task, tset) {
1442                 ret = task_can_attach(task, cs->cpus_allowed);
1443                 if (ret)
1444                         goto out_unlock;
1445                 ret = security_task_setscheduler(task);
1446                 if (ret)
1447                         goto out_unlock;
1448         }
1449 
1450         /*
1451          * Mark attach is in progress.  This makes validate_change() fail
1452          * changes which zero cpus/mems_allowed.
1453          */
1454         cs->attach_in_progress++;
1455         ret = 0;
1456 out_unlock:
1457         mutex_unlock(&cpuset_mutex);
1458         return ret;
1459 }
1460 
1461 static void cpuset_cancel_attach(struct cgroup_subsys_state *css,
1462                                  struct cgroup_taskset *tset)
1463 {
1464         mutex_lock(&cpuset_mutex);
1465         css_cs(css)->attach_in_progress--;
1466         mutex_unlock(&cpuset_mutex);
1467 }
1468 
1469 /*
1470  * Protected by cpuset_mutex.  cpus_attach is used only by cpuset_attach()
1471  * but we can't allocate it dynamically there.  Define it global and
1472  * allocate from cpuset_init().
1473  */
1474 static cpumask_var_t cpus_attach;
1475 
1476 static void cpuset_attach(struct cgroup_subsys_state *css,
1477                           struct cgroup_taskset *tset)
1478 {
1479         /* static buf protected by cpuset_mutex */
1480         static nodemask_t cpuset_attach_nodemask_to;
1481         struct mm_struct *mm;
1482         struct task_struct *task;
1483         struct task_struct *leader = cgroup_taskset_first(tset);
1484         struct cpuset *cs = css_cs(css);
1485         struct cpuset *oldcs = cpuset_attach_old_cs;
1486 
1487         mutex_lock(&cpuset_mutex);
1488 
1489         /* prepare for attach */
1490         if (cs == &top_cpuset)
1491                 cpumask_copy(cpus_attach, cpu_possible_mask);
1492         else
1493                 guarantee_online_cpus(cs, cpus_attach);
1494 
1495         guarantee_online_mems(cs, &cpuset_attach_nodemask_to);
1496 
1497         cgroup_taskset_for_each(task, tset) {
1498                 /*
1499                  * can_attach beforehand should guarantee that this doesn't
1500                  * fail.  TODO: have a better way to handle failure here
1501                  */
1502                 WARN_ON_ONCE(set_cpus_allowed_ptr(task, cpus_attach));
1503 
1504                 cpuset_change_task_nodemask(task, &cpuset_attach_nodemask_to);
1505                 cpuset_update_task_spread_flag(cs, task);
1506         }
1507 
1508         /*
1509          * Change mm, possibly for multiple threads in a threadgroup. This is
1510          * expensive and may sleep.
1511          */
1512         cpuset_attach_nodemask_to = cs->effective_mems;
1513         mm = get_task_mm(leader);
1514         if (mm) {
1515                 mpol_rebind_mm(mm, &cpuset_attach_nodemask_to);
1516 
1517                 /*
1518                  * old_mems_allowed is the same with mems_allowed here, except
1519                  * if this task is being moved automatically due to hotplug.
1520                  * In that case @mems_allowed has been updated and is empty,
1521                  * so @old_mems_allowed is the right nodesets that we migrate
1522                  * mm from.
1523                  */
1524                 if (is_memory_migrate(cs)) {
1525                         cpuset_migrate_mm(mm, &oldcs->old_mems_allowed,
1526                                           &cpuset_attach_nodemask_to);
1527                 }
1528                 mmput(mm);
1529         }
1530 
1531         cs->old_mems_allowed = cpuset_attach_nodemask_to;
1532 
1533         cs->attach_in_progress--;
1534         if (!cs->attach_in_progress)
1535                 wake_up(&cpuset_attach_wq);
1536 
1537         mutex_unlock(&cpuset_mutex);
1538 }
1539 
1540 /* The various types of files and directories in a cpuset file system */
1541 
1542 typedef enum {
1543         FILE_MEMORY_MIGRATE,
1544         FILE_CPULIST,
1545         FILE_MEMLIST,
1546         FILE_EFFECTIVE_CPULIST,
1547         FILE_EFFECTIVE_MEMLIST,
1548         FILE_CPU_EXCLUSIVE,
1549         FILE_MEM_EXCLUSIVE,
1550         FILE_MEM_HARDWALL,
1551         FILE_SCHED_LOAD_BALANCE,
1552         FILE_SCHED_RELAX_DOMAIN_LEVEL,
1553         FILE_MEMORY_PRESSURE_ENABLED,
1554         FILE_MEMORY_PRESSURE,
1555         FILE_SPREAD_PAGE,
1556         FILE_SPREAD_SLAB,
1557 } cpuset_filetype_t;
1558 
1559 static int cpuset_write_u64(struct cgroup_subsys_state *css, struct cftype *cft,
1560                             u64 val)
1561 {
1562         struct cpuset *cs = css_cs(css);
1563         cpuset_filetype_t type = cft->private;
1564         int retval = 0;
1565 
1566         mutex_lock(&cpuset_mutex);
1567         if (!is_cpuset_online(cs)) {
1568                 retval = -ENODEV;
1569                 goto out_unlock;
1570         }
1571 
1572         switch (type) {
1573         case FILE_CPU_EXCLUSIVE:
1574                 retval = update_flag(CS_CPU_EXCLUSIVE, cs, val);
1575                 break;
1576         case FILE_MEM_EXCLUSIVE:
1577                 retval = update_flag(CS_MEM_EXCLUSIVE, cs, val);
1578                 break;
1579         case FILE_MEM_HARDWALL:
1580                 retval = update_flag(CS_MEM_HARDWALL, cs, val);
1581                 break;
1582         case FILE_SCHED_LOAD_BALANCE:
1583                 retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, val);
1584                 break;
1585         case FILE_MEMORY_MIGRATE:
1586                 retval = update_flag(CS_MEMORY_MIGRATE, cs, val);
1587                 break;
1588         case FILE_MEMORY_PRESSURE_ENABLED:
1589                 cpuset_memory_pressure_enabled = !!val;
1590                 break;
1591         case FILE_MEMORY_PRESSURE:
1592                 retval = -EACCES;
1593                 break;
1594         case FILE_SPREAD_PAGE:
1595                 retval = update_flag(CS_SPREAD_PAGE, cs, val);
1596                 break;
1597         case FILE_SPREAD_SLAB:
1598                 retval = update_flag(CS_SPREAD_SLAB, cs, val);
1599                 break;
1600         default:
1601                 retval = -EINVAL;
1602                 break;
1603         }
1604 out_unlock:
1605         mutex_unlock(&cpuset_mutex);
1606         return retval;
1607 }
1608 
1609 static int cpuset_write_s64(struct cgroup_subsys_state *css, struct cftype *cft,
1610                             s64 val)
1611 {
1612         struct cpuset *cs = css_cs(css);
1613         cpuset_filetype_t type = cft->private;
1614         int retval = -ENODEV;
1615 
1616         mutex_lock(&cpuset_mutex);
1617         if (!is_cpuset_online(cs))
1618                 goto out_unlock;
1619 
1620         switch (type) {
1621         case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1622                 retval = update_relax_domain_level(cs, val);
1623                 break;
1624         default:
1625                 retval = -EINVAL;
1626                 break;
1627         }
1628 out_unlock:
1629         mutex_unlock(&cpuset_mutex);
1630         return retval;
1631 }
1632 
1633 /*
1634  * Common handling for a write to a "cpus" or "mems" file.
1635  */
1636 static ssize_t cpuset_write_resmask(struct kernfs_open_file *of,
1637                                     char *buf, size_t nbytes, loff_t off)
1638 {
1639         struct cpuset *cs = css_cs(of_css(of));
1640         struct cpuset *trialcs;
1641         int retval = -ENODEV;
1642 
1643         buf = strstrip(buf);
1644 
1645         /*
1646          * CPU or memory hotunplug may leave @cs w/o any execution
1647          * resources, in which case the hotplug code asynchronously updates
1648          * configuration and transfers all tasks to the nearest ancestor
1649          * which can execute.
1650          *
1651          * As writes to "cpus" or "mems" may restore @cs's execution
1652          * resources, wait for the previously scheduled operations before
1653          * proceeding, so that we don't end up keep removing tasks added
1654          * after execution capability is restored.
1655          *
1656          * cpuset_hotplug_work calls back into cgroup core via
1657          * cgroup_transfer_tasks() and waiting for it from a cgroupfs
1658          * operation like this one can lead to a deadlock through kernfs
1659          * active_ref protection.  Let's break the protection.  Losing the
1660          * protection is okay as we check whether @cs is online after
1661          * grabbing cpuset_mutex anyway.  This only happens on the legacy
1662          * hierarchies.
1663          */
1664         css_get(&cs->css);
1665         kernfs_break_active_protection(of->kn);
1666         flush_work(&cpuset_hotplug_work);
1667 
1668         mutex_lock(&cpuset_mutex);
1669         if (!is_cpuset_online(cs))
1670                 goto out_unlock;
1671 
1672         trialcs = alloc_trial_cpuset(cs);
1673         if (!trialcs) {
1674                 retval = -ENOMEM;
1675                 goto out_unlock;
1676         }
1677 
1678         switch (of_cft(of)->private) {
1679         case FILE_CPULIST:
1680                 retval = update_cpumask(cs, trialcs, buf);
1681                 break;
1682         case FILE_MEMLIST:
1683                 retval = update_nodemask(cs, trialcs, buf);
1684                 break;
1685         default:
1686                 retval = -EINVAL;
1687                 break;
1688         }
1689 
1690         free_trial_cpuset(trialcs);
1691 out_unlock:
1692         mutex_unlock(&cpuset_mutex);
1693         kernfs_unbreak_active_protection(of->kn);
1694         css_put(&cs->css);
1695         return retval ?: nbytes;
1696 }
1697 
1698 /*
1699  * These ascii lists should be read in a single call, by using a user
1700  * buffer large enough to hold the entire map.  If read in smaller
1701  * chunks, there is no guarantee of atomicity.  Since the display format
1702  * used, list of ranges of sequential numbers, is variable length,
1703  * and since these maps can change value dynamically, one could read
1704  * gibberish by doing partial reads while a list was changing.
1705  */
1706 static int cpuset_common_seq_show(struct seq_file *sf, void *v)
1707 {
1708         struct cpuset *cs = css_cs(seq_css(sf));
1709         cpuset_filetype_t type = seq_cft(sf)->private;
1710         ssize_t count;
1711         char *buf, *s;
1712         int ret = 0;
1713 
1714         count = seq_get_buf(sf, &buf);
1715         s = buf;
1716 
1717         spin_lock_irq(&callback_lock);
1718 
1719         switch (type) {
1720         case FILE_CPULIST:
1721                 s += cpulist_scnprintf(s, count, cs->cpus_allowed);
1722                 break;
1723         case FILE_MEMLIST:
1724                 s += nodelist_scnprintf(s, count, cs->mems_allowed);
1725                 break;
1726         case FILE_EFFECTIVE_CPULIST:
1727                 s += cpulist_scnprintf(s, count, cs->effective_cpus);
1728                 break;
1729         case FILE_EFFECTIVE_MEMLIST:
1730                 s += nodelist_scnprintf(s, count, cs->effective_mems);
1731                 break;
1732         default:
1733                 ret = -EINVAL;
1734                 goto out_unlock;
1735         }
1736 
1737         if (s < buf + count - 1) {
1738                 *s++ = '\n';
1739                 seq_commit(sf, s - buf);
1740         } else {
1741                 seq_commit(sf, -1);
1742         }
1743 out_unlock:
1744         spin_unlock_irq(&callback_lock);
1745         return ret;
1746 }
1747 
1748 static u64 cpuset_read_u64(struct cgroup_subsys_state *css, struct cftype *cft)
1749 {
1750         struct cpuset *cs = css_cs(css);
1751         cpuset_filetype_t type = cft->private;
1752         switch (type) {
1753         case FILE_CPU_EXCLUSIVE:
1754                 return is_cpu_exclusive(cs);
1755         case FILE_MEM_EXCLUSIVE:
1756                 return is_mem_exclusive(cs);
1757         case FILE_MEM_HARDWALL:
1758                 return is_mem_hardwall(cs);
1759         case FILE_SCHED_LOAD_BALANCE:
1760                 return is_sched_load_balance(cs);
1761         case FILE_MEMORY_MIGRATE:
1762                 return is_memory_migrate(cs);
1763         case FILE_MEMORY_PRESSURE_ENABLED:
1764                 return cpuset_memory_pressure_enabled;
1765         case FILE_MEMORY_PRESSURE:
1766                 return fmeter_getrate(&cs->fmeter);
1767         case FILE_SPREAD_PAGE:
1768                 return is_spread_page(cs);
1769         case FILE_SPREAD_SLAB:
1770                 return is_spread_slab(cs);
1771         default:
1772                 BUG();
1773         }
1774 
1775         /* Unreachable but makes gcc happy */
1776         return 0;
1777 }
1778 
1779 static s64 cpuset_read_s64(struct cgroup_subsys_state *css, struct cftype *cft)
1780 {
1781         struct cpuset *cs = css_cs(css);
1782         cpuset_filetype_t type = cft->private;
1783         switch (type) {
1784         case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1785                 return cs->relax_domain_level;
1786         default:
1787                 BUG();
1788         }
1789 
1790         /* Unrechable but makes gcc happy */
1791         return 0;
1792 }
1793 
1794 
1795 /*
1796  * for the common functions, 'private' gives the type of file
1797  */
1798 
1799 static struct cftype files[] = {
1800         {
1801                 .name = "cpus",
1802                 .seq_show = cpuset_common_seq_show,
1803                 .write = cpuset_write_resmask,
1804                 .max_write_len = (100U + 6 * NR_CPUS),
1805                 .private = FILE_CPULIST,
1806         },
1807 
1808         {
1809                 .name = "mems",
1810                 .seq_show = cpuset_common_seq_show,
1811                 .write = cpuset_write_resmask,
1812                 .max_write_len = (100U + 6 * MAX_NUMNODES),
1813                 .private = FILE_MEMLIST,
1814         },
1815 
1816         {
1817                 .name = "effective_cpus",
1818                 .seq_show = cpuset_common_seq_show,
1819                 .private = FILE_EFFECTIVE_CPULIST,
1820         },
1821 
1822         {
1823                 .name = "effective_mems",
1824                 .seq_show = cpuset_common_seq_show,
1825                 .private = FILE_EFFECTIVE_MEMLIST,
1826         },
1827 
1828         {
1829                 .name = "cpu_exclusive",
1830                 .read_u64 = cpuset_read_u64,
1831                 .write_u64 = cpuset_write_u64,
1832                 .private = FILE_CPU_EXCLUSIVE,
1833         },
1834 
1835         {
1836                 .name = "mem_exclusive",
1837                 .read_u64 = cpuset_read_u64,
1838                 .write_u64 = cpuset_write_u64,
1839                 .private = FILE_MEM_EXCLUSIVE,
1840         },
1841 
1842         {
1843                 .name = "mem_hardwall",
1844                 .read_u64 = cpuset_read_u64,
1845                 .write_u64 = cpuset_write_u64,
1846                 .private = FILE_MEM_HARDWALL,
1847         },
1848 
1849         {
1850                 .name = "sched_load_balance",
1851                 .read_u64 = cpuset_read_u64,
1852                 .write_u64 = cpuset_write_u64,
1853                 .private = FILE_SCHED_LOAD_BALANCE,
1854         },
1855 
1856         {
1857                 .name = "sched_relax_domain_level",
1858                 .read_s64 = cpuset_read_s64,
1859                 .write_s64 = cpuset_write_s64,
1860                 .private = FILE_SCHED_RELAX_DOMAIN_LEVEL,
1861         },
1862 
1863         {
1864                 .name = "memory_migrate",
1865                 .read_u64 = cpuset_read_u64,
1866                 .write_u64 = cpuset_write_u64,
1867                 .private = FILE_MEMORY_MIGRATE,
1868         },
1869 
1870         {
1871                 .name = "memory_pressure",
1872                 .read_u64 = cpuset_read_u64,
1873                 .write_u64 = cpuset_write_u64,
1874                 .private = FILE_MEMORY_PRESSURE,
1875                 .mode = S_IRUGO,
1876         },
1877 
1878         {
1879                 .name = "memory_spread_page",
1880                 .read_u64 = cpuset_read_u64,
1881                 .write_u64 = cpuset_write_u64,
1882                 .private = FILE_SPREAD_PAGE,
1883         },
1884 
1885         {
1886                 .name = "memory_spread_slab",
1887                 .read_u64 = cpuset_read_u64,
1888                 .write_u64 = cpuset_write_u64,
1889                 .private = FILE_SPREAD_SLAB,
1890         },
1891 
1892         {
1893                 .name = "memory_pressure_enabled",
1894                 .flags = CFTYPE_ONLY_ON_ROOT,
1895                 .read_u64 = cpuset_read_u64,
1896                 .write_u64 = cpuset_write_u64,
1897                 .private = FILE_MEMORY_PRESSURE_ENABLED,
1898         },
1899 
1900         { }     /* terminate */
1901 };
1902 
1903 /*
1904  *      cpuset_css_alloc - allocate a cpuset css
1905  *      cgrp:   control group that the new cpuset will be part of
1906  */
1907 
1908 static struct cgroup_subsys_state *
1909 cpuset_css_alloc(struct cgroup_subsys_state *parent_css)
1910 {
1911         struct cpuset *cs;
1912 
1913         if (!parent_css)
1914                 return &top_cpuset.css;
1915 
1916         cs = kzalloc(sizeof(*cs), GFP_KERNEL);
1917         if (!cs)
1918                 return ERR_PTR(-ENOMEM);
1919         if (!alloc_cpumask_var(&cs->cpus_allowed, GFP_KERNEL))
1920                 goto free_cs;
1921         if (!alloc_cpumask_var(&cs->effective_cpus, GFP_KERNEL))
1922                 goto free_cpus;
1923 
1924         set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
1925         cpumask_clear(cs->cpus_allowed);
1926         nodes_clear(cs->mems_allowed);
1927         cpumask_clear(cs->effective_cpus);
1928         nodes_clear(cs->effective_mems);
1929         fmeter_init(&cs->fmeter);
1930         cs->relax_domain_level = -1;
1931 
1932         return &cs->css;
1933 
1934 free_cpus:
1935         free_cpumask_var(cs->cpus_allowed);
1936 free_cs:
1937         kfree(cs);
1938         return ERR_PTR(-ENOMEM);
1939 }
1940 
1941 static int cpuset_css_online(struct cgroup_subsys_state *css)
1942 {
1943         struct cpuset *cs = css_cs(css);
1944         struct cpuset *parent = parent_cs(cs);
1945         struct cpuset *tmp_cs;
1946         struct cgroup_subsys_state *pos_css;
1947 
1948         if (!parent)
1949                 return 0;
1950 
1951         mutex_lock(&cpuset_mutex);
1952 
1953         set_bit(CS_ONLINE, &cs->flags);
1954         if (is_spread_page(parent))
1955                 set_bit(CS_SPREAD_PAGE, &cs->flags);
1956         if (is_spread_slab(parent))
1957                 set_bit(CS_SPREAD_SLAB, &cs->flags);
1958 
1959         cpuset_inc();
1960 
1961         spin_lock_irq(&callback_lock);
1962         if (cgroup_on_dfl(cs->css.cgroup)) {
1963                 cpumask_copy(cs->effective_cpus, parent->effective_cpus);
1964                 cs->effective_mems = parent->effective_mems;
1965         }
1966         spin_unlock_irq(&callback_lock);
1967 
1968         if (!test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags))
1969                 goto out_unlock;
1970 
1971         /*
1972          * Clone @parent's configuration if CGRP_CPUSET_CLONE_CHILDREN is
1973          * set.  This flag handling is implemented in cgroup core for
1974          * histrical reasons - the flag may be specified during mount.
1975          *
1976          * Currently, if any sibling cpusets have exclusive cpus or mem, we
1977          * refuse to clone the configuration - thereby refusing the task to
1978          * be entered, and as a result refusing the sys_unshare() or
1979          * clone() which initiated it.  If this becomes a problem for some
1980          * users who wish to allow that scenario, then this could be
1981          * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
1982          * (and likewise for mems) to the new cgroup.
1983          */
1984         rcu_read_lock();
1985         cpuset_for_each_child(tmp_cs, pos_css, parent) {
1986                 if (is_mem_exclusive(tmp_cs) || is_cpu_exclusive(tmp_cs)) {
1987                         rcu_read_unlock();
1988                         goto out_unlock;
1989                 }
1990         }
1991         rcu_read_unlock();
1992 
1993         spin_lock_irq(&callback_lock);
1994         cs->mems_allowed = parent->mems_allowed;
1995         cpumask_copy(cs->cpus_allowed, parent->cpus_allowed);
1996         spin_unlock_irq(&callback_lock);
1997 out_unlock:
1998         mutex_unlock(&cpuset_mutex);
1999         return 0;
2000 }
2001 
2002 /*
2003  * If the cpuset being removed has its flag 'sched_load_balance'
2004  * enabled, then simulate turning sched_load_balance off, which
2005  * will call rebuild_sched_domains_locked().
2006  */
2007 
2008 static void cpuset_css_offline(struct cgroup_subsys_state *css)
2009 {
2010         struct cpuset *cs = css_cs(css);
2011 
2012         mutex_lock(&cpuset_mutex);
2013 
2014         if (is_sched_load_balance(cs))
2015                 update_flag(CS_SCHED_LOAD_BALANCE, cs, 0);
2016 
2017         cpuset_dec();
2018         clear_bit(CS_ONLINE, &cs->flags);
2019 
2020         mutex_unlock(&cpuset_mutex);
2021 }
2022 
2023 static void cpuset_css_free(struct cgroup_subsys_state *css)
2024 {
2025         struct cpuset *cs = css_cs(css);
2026 
2027         free_cpumask_var(cs->effective_cpus);
2028         free_cpumask_var(cs->cpus_allowed);
2029         kfree(cs);
2030 }
2031 
2032 static void cpuset_bind(struct cgroup_subsys_state *root_css)
2033 {
2034         mutex_lock(&cpuset_mutex);
2035         spin_lock_irq(&callback_lock);
2036 
2037         if (cgroup_on_dfl(root_css->cgroup)) {
2038                 cpumask_copy(top_cpuset.cpus_allowed, cpu_possible_mask);
2039                 top_cpuset.mems_allowed = node_possible_map;
2040         } else {
2041                 cpumask_copy(top_cpuset.cpus_allowed,
2042                              top_cpuset.effective_cpus);
2043                 top_cpuset.mems_allowed = top_cpuset.effective_mems;
2044         }
2045 
2046         spin_unlock_irq(&callback_lock);
2047         mutex_unlock(&cpuset_mutex);
2048 }
2049 
2050 struct cgroup_subsys cpuset_cgrp_subsys = {
2051         .css_alloc      = cpuset_css_alloc,
2052         .css_online     = cpuset_css_online,
2053         .css_offline    = cpuset_css_offline,
2054         .css_free       = cpuset_css_free,
2055         .can_attach     = cpuset_can_attach,
2056         .cancel_attach  = cpuset_cancel_attach,
2057         .attach         = cpuset_attach,
2058         .bind           = cpuset_bind,
2059         .legacy_cftypes = files,
2060         .early_init     = 1,
2061 };
2062 
2063 /**
2064  * cpuset_init - initialize cpusets at system boot
2065  *
2066  * Description: Initialize top_cpuset and the cpuset internal file system,
2067  **/
2068 
2069 int __init cpuset_init(void)
2070 {
2071         int err = 0;
2072 
2073         if (!alloc_cpumask_var(&top_cpuset.cpus_allowed, GFP_KERNEL))
2074                 BUG();
2075         if (!alloc_cpumask_var(&top_cpuset.effective_cpus, GFP_KERNEL))
2076                 BUG();
2077 
2078         cpumask_setall(top_cpuset.cpus_allowed);
2079         nodes_setall(top_cpuset.mems_allowed);
2080         cpumask_setall(top_cpuset.effective_cpus);
2081         nodes_setall(top_cpuset.effective_mems);
2082 
2083         fmeter_init(&top_cpuset.fmeter);
2084         set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags);
2085         top_cpuset.relax_domain_level = -1;
2086 
2087         err = register_filesystem(&cpuset_fs_type);
2088         if (err < 0)
2089                 return err;
2090 
2091         if (!alloc_cpumask_var(&cpus_attach, GFP_KERNEL))
2092                 BUG();
2093 
2094         return 0;
2095 }
2096 
2097 /*
2098  * If CPU and/or memory hotplug handlers, below, unplug any CPUs
2099  * or memory nodes, we need to walk over the cpuset hierarchy,
2100  * removing that CPU or node from all cpusets.  If this removes the
2101  * last CPU or node from a cpuset, then move the tasks in the empty
2102  * cpuset to its next-highest non-empty parent.
2103  */
2104 static void remove_tasks_in_empty_cpuset(struct cpuset *cs)
2105 {
2106         struct cpuset *parent;
2107 
2108         /*
2109          * Find its next-highest non-empty parent, (top cpuset
2110          * has online cpus, so can't be empty).
2111          */
2112         parent = parent_cs(cs);
2113         while (cpumask_empty(parent->cpus_allowed) ||
2114                         nodes_empty(parent->mems_allowed))
2115                 parent = parent_cs(parent);
2116 
2117         if (cgroup_transfer_tasks(parent->css.cgroup, cs->css.cgroup)) {
2118                 pr_err("cpuset: failed to transfer tasks out of empty cpuset ");
2119                 pr_cont_cgroup_name(cs->css.cgroup);
2120                 pr_cont("\n");
2121         }
2122 }
2123 
2124 static void
2125 hotplug_update_tasks_legacy(struct cpuset *cs,
2126                             struct cpumask *new_cpus, nodemask_t *new_mems,
2127                             bool cpus_updated, bool mems_updated)
2128 {
2129         bool is_empty;
2130 
2131         spin_lock_irq(&callback_lock);
2132         cpumask_copy(cs->cpus_allowed, new_cpus);
2133         cpumask_copy(cs->effective_cpus, new_cpus);
2134         cs->mems_allowed = *new_mems;
2135         cs->effective_mems = *new_mems;
2136         spin_unlock_irq(&callback_lock);
2137 
2138         /*
2139          * Don't call update_tasks_cpumask() if the cpuset becomes empty,
2140          * as the tasks will be migratecd to an ancestor.
2141          */
2142         if (cpus_updated && !cpumask_empty(cs->cpus_allowed))
2143                 update_tasks_cpumask(cs);
2144         if (mems_updated && !nodes_empty(cs->mems_allowed))
2145                 update_tasks_nodemask(cs);
2146 
2147         is_empty = cpumask_empty(cs->cpus_allowed) ||
2148                    nodes_empty(cs->mems_allowed);
2149 
2150         mutex_unlock(&cpuset_mutex);
2151 
2152         /*
2153          * Move tasks to the nearest ancestor with execution resources,
2154          * This is full cgroup operation which will also call back into
2155          * cpuset. Should be done outside any lock.
2156          */
2157         if (is_empty)
2158                 remove_tasks_in_empty_cpuset(cs);
2159 
2160         mutex_lock(&cpuset_mutex);
2161 }
2162 
2163 static void
2164 hotplug_update_tasks(struct cpuset *cs,
2165                      struct cpumask *new_cpus, nodemask_t *new_mems,
2166                      bool cpus_updated, bool mems_updated)
2167 {
2168         if (cpumask_empty(new_cpus))
2169                 cpumask_copy(new_cpus, parent_cs(cs)->effective_cpus);
2170         if (nodes_empty(*new_mems))
2171                 *new_mems = parent_cs(cs)->effective_mems;
2172 
2173         spin_lock_irq(&callback_lock);
2174         cpumask_copy(cs->effective_cpus, new_cpus);
2175         cs->effective_mems = *new_mems;
2176         spin_unlock_irq(&callback_lock);
2177 
2178         if (cpus_updated)
2179                 update_tasks_cpumask(cs);
2180         if (mems_updated)
2181                 update_tasks_nodemask(cs);
2182 }
2183 
2184 /**
2185  * cpuset_hotplug_update_tasks - update tasks in a cpuset for hotunplug
2186  * @cs: cpuset in interest
2187  *
2188  * Compare @cs's cpu and mem masks against top_cpuset and if some have gone
2189  * offline, update @cs accordingly.  If @cs ends up with no CPU or memory,
2190  * all its tasks are moved to the nearest ancestor with both resources.
2191  */
2192 static void cpuset_hotplug_update_tasks(struct cpuset *cs)
2193 {
2194         static cpumask_t new_cpus;
2195         static nodemask_t new_mems;
2196         bool cpus_updated;
2197         bool mems_updated;
2198 retry:
2199         wait_event(cpuset_attach_wq, cs->attach_in_progress == 0);
2200 
2201         mutex_lock(&cpuset_mutex);
2202 
2203         /*
2204          * We have raced with task attaching. We wait until attaching
2205          * is finished, so we won't attach a task to an empty cpuset.
2206          */
2207         if (cs->attach_in_progress) {
2208                 mutex_unlock(&cpuset_mutex);
2209                 goto retry;
2210         }
2211 
2212         cpumask_and(&new_cpus, cs->cpus_allowed, parent_cs(cs)->effective_cpus);
2213         nodes_and(new_mems, cs->mems_allowed, parent_cs(cs)->effective_mems);
2214 
2215         cpus_updated = !cpumask_equal(&new_cpus, cs->effective_cpus);
2216         mems_updated = !nodes_equal(new_mems, cs->effective_mems);
2217 
2218         if (cgroup_on_dfl(cs->css.cgroup))
2219                 hotplug_update_tasks(cs, &new_cpus, &new_mems,
2220                                      cpus_updated, mems_updated);
2221         else
2222                 hotplug_update_tasks_legacy(cs, &new_cpus, &new_mems,
2223                                             cpus_updated, mems_updated);
2224 
2225         mutex_unlock(&cpuset_mutex);
2226 }
2227 
2228 /**
2229  * cpuset_hotplug_workfn - handle CPU/memory hotunplug for a cpuset
2230  *
2231  * This function is called after either CPU or memory configuration has
2232  * changed and updates cpuset accordingly.  The top_cpuset is always
2233  * synchronized to cpu_active_mask and N_MEMORY, which is necessary in
2234  * order to make cpusets transparent (of no affect) on systems that are
2235  * actively using CPU hotplug but making no active use of cpusets.
2236  *
2237  * Non-root cpusets are only affected by offlining.  If any CPUs or memory
2238  * nodes have been taken down, cpuset_hotplug_update_tasks() is invoked on
2239  * all descendants.
2240  *
2241  * Note that CPU offlining during suspend is ignored.  We don't modify
2242  * cpusets across suspend/resume cycles at all.
2243  */
2244 static void cpuset_hotplug_workfn(struct work_struct *work)
2245 {
2246         static cpumask_t new_cpus;
2247         static nodemask_t new_mems;
2248         bool cpus_updated, mems_updated;
2249         bool on_dfl = cgroup_on_dfl(top_cpuset.css.cgroup);
2250 
2251         mutex_lock(&cpuset_mutex);
2252 
2253         /* fetch the available cpus/mems and find out which changed how */
2254         cpumask_copy(&new_cpus, cpu_active_mask);
2255         new_mems = node_states[N_MEMORY];
2256 
2257         cpus_updated = !cpumask_equal(top_cpuset.effective_cpus, &new_cpus);
2258         mems_updated = !nodes_equal(top_cpuset.effective_mems, new_mems);
2259 
2260         /* synchronize cpus_allowed to cpu_active_mask */
2261         if (cpus_updated) {
2262                 spin_lock_irq(&callback_lock);
2263                 if (!on_dfl)
2264                         cpumask_copy(top_cpuset.cpus_allowed, &new_cpus);
2265                 cpumask_copy(top_cpuset.effective_cpus, &new_cpus);
2266                 spin_unlock_irq(&callback_lock);
2267                 /* we don't mess with cpumasks of tasks in top_cpuset */
2268         }
2269 
2270         /* synchronize mems_allowed to N_MEMORY */
2271         if (mems_updated) {
2272                 spin_lock_irq(&callback_lock);
2273                 if (!on_dfl)
2274                         top_cpuset.mems_allowed = new_mems;
2275                 top_cpuset.effective_mems = new_mems;
2276                 spin_unlock_irq(&callback_lock);
2277                 update_tasks_nodemask(&top_cpuset);
2278         }
2279 
2280         mutex_unlock(&cpuset_mutex);
2281 
2282         /* if cpus or mems changed, we need to propagate to descendants */
2283         if (cpus_updated || mems_updated) {
2284                 struct cpuset *cs;
2285                 struct cgroup_subsys_state *pos_css;
2286 
2287                 rcu_read_lock();
2288                 cpuset_for_each_descendant_pre(cs, pos_css, &top_cpuset) {
2289                         if (cs == &top_cpuset || !css_tryget_online(&cs->css))
2290                                 continue;
2291                         rcu_read_unlock();
2292 
2293                         cpuset_hotplug_update_tasks(cs);
2294 
2295                         rcu_read_lock();
2296                         css_put(&cs->css);
2297                 }
2298                 rcu_read_unlock();
2299         }
2300 
2301         /* rebuild sched domains if cpus_allowed has changed */
2302         if (cpus_updated)
2303                 rebuild_sched_domains();
2304 }
2305 
2306 void cpuset_update_active_cpus(bool cpu_online)
2307 {
2308         /*
2309          * We're inside cpu hotplug critical region which usually nests
2310          * inside cgroup synchronization.  Bounce actual hotplug processing
2311          * to a work item to avoid reverse locking order.
2312          *
2313          * We still need to do partition_sched_domains() synchronously;
2314          * otherwise, the scheduler will get confused and put tasks to the
2315          * dead CPU.  Fall back to the default single domain.
2316          * cpuset_hotplug_workfn() will rebuild it as necessary.
2317          */
2318         partition_sched_domains(1, NULL, NULL);
2319         schedule_work(&cpuset_hotplug_work);
2320 }
2321 
2322 /*
2323  * Keep top_cpuset.mems_allowed tracking node_states[N_MEMORY].
2324  * Call this routine anytime after node_states[N_MEMORY] changes.
2325  * See cpuset_update_active_cpus() for CPU hotplug handling.
2326  */
2327 static int cpuset_track_online_nodes(struct notifier_block *self,
2328                                 unsigned long action, void *arg)
2329 {
2330         schedule_work(&cpuset_hotplug_work);
2331         return NOTIFY_OK;
2332 }
2333 
2334 static struct notifier_block cpuset_track_online_nodes_nb = {
2335         .notifier_call = cpuset_track_online_nodes,
2336         .priority = 10,         /* ??! */
2337 };
2338 
2339 /**
2340  * cpuset_init_smp - initialize cpus_allowed
2341  *
2342  * Description: Finish top cpuset after cpu, node maps are initialized
2343  */
2344 void __init cpuset_init_smp(void)
2345 {
2346         cpumask_copy(top_cpuset.cpus_allowed, cpu_active_mask);
2347         top_cpuset.mems_allowed = node_states[N_MEMORY];
2348         top_cpuset.old_mems_allowed = top_cpuset.mems_allowed;
2349 
2350         cpumask_copy(top_cpuset.effective_cpus, cpu_active_mask);
2351         top_cpuset.effective_mems = node_states[N_MEMORY];
2352 
2353         register_hotmemory_notifier(&cpuset_track_online_nodes_nb);
2354 }
2355 
2356 /**
2357  * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
2358  * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
2359  * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
2360  *
2361  * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
2362  * attached to the specified @tsk.  Guaranteed to return some non-empty
2363  * subset of cpu_online_mask, even if this means going outside the
2364  * tasks cpuset.
2365  **/
2366 
2367 void cpuset_cpus_allowed(struct task_struct *tsk, struct cpumask *pmask)
2368 {
2369         unsigned long flags;
2370 
2371         spin_lock_irqsave(&callback_lock, flags);
2372         rcu_read_lock();
2373         guarantee_online_cpus(task_cs(tsk), pmask);
2374         rcu_read_unlock();
2375         spin_unlock_irqrestore(&callback_lock, flags);
2376 }
2377 
2378 void cpuset_cpus_allowed_fallback(struct task_struct *tsk)
2379 {
2380         rcu_read_lock();
2381         do_set_cpus_allowed(tsk, task_cs(tsk)->effective_cpus);
2382         rcu_read_unlock();
2383 
2384         /*
2385          * We own tsk->cpus_allowed, nobody can change it under us.
2386          *
2387          * But we used cs && cs->cpus_allowed lockless and thus can
2388          * race with cgroup_attach_task() or update_cpumask() and get
2389          * the wrong tsk->cpus_allowed. However, both cases imply the
2390          * subsequent cpuset_change_cpumask()->set_cpus_allowed_ptr()
2391          * which takes task_rq_lock().
2392          *
2393          * If we are called after it dropped the lock we must see all
2394          * changes in tsk_cs()->cpus_allowed. Otherwise we can temporary
2395          * set any mask even if it is not right from task_cs() pov,
2396          * the pending set_cpus_allowed_ptr() will fix things.
2397          *
2398          * select_fallback_rq() will fix things ups and set cpu_possible_mask
2399          * if required.
2400          */
2401 }
2402 
2403 void cpuset_init_current_mems_allowed(void)
2404 {
2405         nodes_setall(current->mems_allowed);
2406 }
2407 
2408 /**
2409  * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
2410  * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
2411  *
2412  * Description: Returns the nodemask_t mems_allowed of the cpuset
2413  * attached to the specified @tsk.  Guaranteed to return some non-empty
2414  * subset of node_states[N_MEMORY], even if this means going outside the
2415  * tasks cpuset.
2416  **/
2417 
2418 nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
2419 {
2420         nodemask_t mask;
2421         unsigned long flags;
2422 
2423         spin_lock_irqsave(&callback_lock, flags);
2424         rcu_read_lock();
2425         guarantee_online_mems(task_cs(tsk), &mask);
2426         rcu_read_unlock();
2427         spin_unlock_irqrestore(&callback_lock, flags);
2428 
2429         return mask;
2430 }
2431 
2432 /**
2433  * cpuset_nodemask_valid_mems_allowed - check nodemask vs. curremt mems_allowed
2434  * @nodemask: the nodemask to be checked
2435  *
2436  * Are any of the nodes in the nodemask allowed in current->mems_allowed?
2437  */
2438 int cpuset_nodemask_valid_mems_allowed(nodemask_t *nodemask)
2439 {
2440         return nodes_intersects(*nodemask, current->mems_allowed);
2441 }
2442 
2443 /*
2444  * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
2445  * mem_hardwall ancestor to the specified cpuset.  Call holding
2446  * callback_lock.  If no ancestor is mem_exclusive or mem_hardwall
2447  * (an unusual configuration), then returns the root cpuset.
2448  */
2449 static struct cpuset *nearest_hardwall_ancestor(struct cpuset *cs)
2450 {
2451         while (!(is_mem_exclusive(cs) || is_mem_hardwall(cs)) && parent_cs(cs))
2452                 cs = parent_cs(cs);
2453         return cs;
2454 }
2455 
2456 /**
2457  * cpuset_node_allowed - Can we allocate on a memory node?
2458  * @node: is this an allowed node?
2459  * @gfp_mask: memory allocation flags
2460  *
2461  * If we're in interrupt, yes, we can always allocate.  If __GFP_THISNODE is
2462  * set, yes, we can always allocate.  If node is in our task's mems_allowed,
2463  * yes.  If it's not a __GFP_HARDWALL request and this node is in the nearest
2464  * hardwalled cpuset ancestor to this task's cpuset, yes.  If the task has been
2465  * OOM killed and has access to memory reserves as specified by the TIF_MEMDIE
2466  * flag, yes.
2467  * Otherwise, no.
2468  *
2469  * The __GFP_THISNODE placement logic is really handled elsewhere,
2470  * by forcibly using a zonelist starting at a specified node, and by
2471  * (in get_page_from_freelist()) refusing to consider the zones for
2472  * any node on the zonelist except the first.  By the time any such
2473  * calls get to this routine, we should just shut up and say 'yes'.
2474  *
2475  * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
2476  * and do not allow allocations outside the current tasks cpuset
2477  * unless the task has been OOM killed as is marked TIF_MEMDIE.
2478  * GFP_KERNEL allocations are not so marked, so can escape to the
2479  * nearest enclosing hardwalled ancestor cpuset.
2480  *
2481  * Scanning up parent cpusets requires callback_lock.  The
2482  * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
2483  * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
2484  * current tasks mems_allowed came up empty on the first pass over
2485  * the zonelist.  So only GFP_KERNEL allocations, if all nodes in the
2486  * cpuset are short of memory, might require taking the callback_lock.
2487  *
2488  * The first call here from mm/page_alloc:get_page_from_freelist()
2489  * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
2490  * so no allocation on a node outside the cpuset is allowed (unless
2491  * in interrupt, of course).
2492  *
2493  * The second pass through get_page_from_freelist() doesn't even call
2494  * here for GFP_ATOMIC calls.  For those calls, the __alloc_pages()
2495  * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
2496  * in alloc_flags.  That logic and the checks below have the combined
2497  * affect that:
2498  *      in_interrupt - any node ok (current task context irrelevant)
2499  *      GFP_ATOMIC   - any node ok
2500  *      TIF_MEMDIE   - any node ok
2501  *      GFP_KERNEL   - any node in enclosing hardwalled cpuset ok
2502  *      GFP_USER     - only nodes in current tasks mems allowed ok.
2503  */
2504 int __cpuset_node_allowed(int node, gfp_t gfp_mask)
2505 {
2506         struct cpuset *cs;              /* current cpuset ancestors */
2507         int allowed;                    /* is allocation in zone z allowed? */
2508         unsigned long flags;
2509 
2510         if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
2511                 return 1;
2512         if (node_isset(node, current->mems_allowed))
2513                 return 1;
2514         /*
2515          * Allow tasks that have access to memory reserves because they have
2516          * been OOM killed to get memory anywhere.
2517          */
2518         if (unlikely(test_thread_flag(TIF_MEMDIE)))
2519                 return 1;
2520         if (gfp_mask & __GFP_HARDWALL)  /* If hardwall request, stop here */
2521                 return 0;
2522 
2523         if (current->flags & PF_EXITING) /* Let dying task have memory */
2524                 return 1;
2525 
2526         /* Not hardwall and node outside mems_allowed: scan up cpusets */
2527         spin_lock_irqsave(&callback_lock, flags);
2528 
2529         rcu_read_lock();
2530         cs = nearest_hardwall_ancestor(task_cs(current));
2531         allowed = node_isset(node, cs->mems_allowed);
2532         rcu_read_unlock();
2533 
2534         spin_unlock_irqrestore(&callback_lock, flags);
2535         return allowed;
2536 }
2537 
2538 /**
2539  * cpuset_mem_spread_node() - On which node to begin search for a file page
2540  * cpuset_slab_spread_node() - On which node to begin search for a slab page
2541  *
2542  * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
2543  * tasks in a cpuset with is_spread_page or is_spread_slab set),
2544  * and if the memory allocation used cpuset_mem_spread_node()
2545  * to determine on which node to start looking, as it will for
2546  * certain page cache or slab cache pages such as used for file
2547  * system buffers and inode caches, then instead of starting on the
2548  * local node to look for a free page, rather spread the starting
2549  * node around the tasks mems_allowed nodes.
2550  *
2551  * We don't have to worry about the returned node being offline
2552  * because "it can't happen", and even if it did, it would be ok.
2553  *
2554  * The routines calling guarantee_online_mems() are careful to
2555  * only set nodes in task->mems_allowed that are online.  So it
2556  * should not be possible for the following code to return an
2557  * offline node.  But if it did, that would be ok, as this routine
2558  * is not returning the node where the allocation must be, only
2559  * the node where the search should start.  The zonelist passed to
2560  * __alloc_pages() will include all nodes.  If the slab allocator
2561  * is passed an offline node, it will fall back to the local node.
2562  * See kmem_cache_alloc_node().
2563  */
2564 
2565 static int cpuset_spread_node(int *rotor)
2566 {
2567         int node;
2568 
2569         node = next_node(*rotor, current->mems_allowed);
2570         if (node == MAX_NUMNODES)
2571                 node = first_node(current->mems_allowed);
2572         *rotor = node;
2573         return node;
2574 }
2575 
2576 int cpuset_mem_spread_node(void)
2577 {
2578         if (current->cpuset_mem_spread_rotor == NUMA_NO_NODE)
2579                 current->cpuset_mem_spread_rotor =
2580                         node_random(&current->mems_allowed);
2581 
2582         return cpuset_spread_node(&current->cpuset_mem_spread_rotor);
2583 }
2584 
2585 int cpuset_slab_spread_node(void)
2586 {
2587         if (current->cpuset_slab_spread_rotor == NUMA_NO_NODE)
2588                 current->cpuset_slab_spread_rotor =
2589                         node_random(&current->mems_allowed);
2590 
2591         return cpuset_spread_node(&current->cpuset_slab_spread_rotor);
2592 }
2593 
2594 EXPORT_SYMBOL_GPL(cpuset_mem_spread_node);
2595 
2596 /**
2597  * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
2598  * @tsk1: pointer to task_struct of some task.
2599  * @tsk2: pointer to task_struct of some other task.
2600  *
2601  * Description: Return true if @tsk1's mems_allowed intersects the
2602  * mems_allowed of @tsk2.  Used by the OOM killer to determine if
2603  * one of the task's memory usage might impact the memory available
2604  * to the other.
2605  **/
2606 
2607 int cpuset_mems_allowed_intersects(const struct task_struct *tsk1,
2608                                    const struct task_struct *tsk2)
2609 {
2610         return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed);
2611 }
2612 
2613 #define CPUSET_NODELIST_LEN     (256)
2614 
2615 /**
2616  * cpuset_print_task_mems_allowed - prints task's cpuset and mems_allowed
2617  * @tsk: pointer to task_struct of some task.
2618  *
2619  * Description: Prints @task's name, cpuset name, and cached copy of its
2620  * mems_allowed to the kernel log.
2621  */
2622 void cpuset_print_task_mems_allowed(struct task_struct *tsk)
2623 {
2624          /* Statically allocated to prevent using excess stack. */
2625         static char cpuset_nodelist[CPUSET_NODELIST_LEN];
2626         static DEFINE_SPINLOCK(cpuset_buffer_lock);
2627         struct cgroup *cgrp;
2628 
2629         spin_lock(&cpuset_buffer_lock);
2630         rcu_read_lock();
2631 
2632         cgrp = task_cs(tsk)->css.cgroup;
2633         nodelist_scnprintf(cpuset_nodelist, CPUSET_NODELIST_LEN,
2634                            tsk->mems_allowed);
2635         pr_info("%s cpuset=", tsk->comm);
2636         pr_cont_cgroup_name(cgrp);
2637         pr_cont(" mems_allowed=%s\n", cpuset_nodelist);
2638 
2639         rcu_read_unlock();
2640         spin_unlock(&cpuset_buffer_lock);
2641 }
2642 
2643 /*
2644  * Collection of memory_pressure is suppressed unless
2645  * this flag is enabled by writing "1" to the special
2646  * cpuset file 'memory_pressure_enabled' in the root cpuset.
2647  */
2648 
2649 int cpuset_memory_pressure_enabled __read_mostly;
2650 
2651 /**
2652  * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
2653  *
2654  * Keep a running average of the rate of synchronous (direct)
2655  * page reclaim efforts initiated by tasks in each cpuset.
2656  *
2657  * This represents the rate at which some task in the cpuset
2658  * ran low on memory on all nodes it was allowed to use, and
2659  * had to enter the kernels page reclaim code in an effort to
2660  * create more free memory by tossing clean pages or swapping
2661  * or writing dirty pages.
2662  *
2663  * Display to user space in the per-cpuset read-only file
2664  * "memory_pressure".  Value displayed is an integer
2665  * representing the recent rate of entry into the synchronous
2666  * (direct) page reclaim by any task attached to the cpuset.
2667  **/
2668 
2669 void __cpuset_memory_pressure_bump(void)
2670 {
2671         rcu_read_lock();
2672         fmeter_markevent(&task_cs(current)->fmeter);
2673         rcu_read_unlock();
2674 }
2675 
2676 #ifdef CONFIG_PROC_PID_CPUSET
2677 /*
2678  * proc_cpuset_show()
2679  *  - Print tasks cpuset path into seq_file.
2680  *  - Used for /proc/<pid>/cpuset.
2681  *  - No need to task_lock(tsk) on this tsk->cpuset reference, as it
2682  *    doesn't really matter if tsk->cpuset changes after we read it,
2683  *    and we take cpuset_mutex, keeping cpuset_attach() from changing it
2684  *    anyway.
2685  */
2686 int proc_cpuset_show(struct seq_file *m, struct pid_namespace *ns,
2687                      struct pid *pid, struct task_struct *tsk)
2688 {
2689         char *buf, *p;
2690         struct cgroup_subsys_state *css;
2691         int retval;
2692 
2693         retval = -ENOMEM;
2694         buf = kmalloc(PATH_MAX, GFP_KERNEL);
2695         if (!buf)
2696                 goto out;
2697 
2698         retval = -ENAMETOOLONG;
2699         rcu_read_lock();
2700         css = task_css(tsk, cpuset_cgrp_id);
2701         p = cgroup_path(css->cgroup, buf, PATH_MAX);
2702         rcu_read_unlock();
2703         if (!p)
2704                 goto out_free;
2705         seq_puts(m, p);
2706         seq_putc(m, '\n');
2707         retval = 0;
2708 out_free:
2709         kfree(buf);
2710 out:
2711         return retval;
2712 }
2713 #endif /* CONFIG_PROC_PID_CPUSET */
2714 
2715 /* Display task mems_allowed in /proc/<pid>/status file. */
2716 void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task)
2717 {
2718         seq_puts(m, "Mems_allowed:\t");
2719         seq_nodemask(m, &task->mems_allowed);
2720         seq_puts(m, "\n");
2721         seq_puts(m, "Mems_allowed_list:\t");
2722         seq_nodemask_list(m, &task->mems_allowed);
2723         seq_puts(m, "\n");
2724 }
2725 

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