Version:  2.0.40 2.2.26 2.4.37 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 4.0

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

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