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

TOMOYO Linux Cross Reference
Linux/kernel/cpuset.c

Version: ~ [ linux-5.8 ] ~ [ linux-5.7.14 ] ~ [ linux-5.6.19 ] ~ [ linux-5.5.19 ] ~ [ linux-5.4.57 ] ~ [ linux-5.3.18 ] ~ [ linux-5.2.21 ] ~ [ linux-5.1.21 ] ~ [ linux-5.0.21 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.138 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.193 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.232 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.232 ] ~ [ linux-4.3.6 ] ~ [ linux-4.2.8 ] ~ [ linux-4.1.52 ] ~ [ linux-4.0.9 ] ~ [ linux-3.19.8 ] ~ [ linux-3.18.140 ] ~ [ linux-3.17.8 ] ~ [ linux-3.16.85 ] ~ [ linux-3.15.10 ] ~ [ linux-3.14.79 ] ~ [ linux-3.13.11 ] ~ [ linux-3.12.74 ] ~ [ linux-3.11.10 ] ~ [ linux-3.10.108 ] ~ [ linux-2.6.32.71 ] ~ [ linux-2.6.0 ] ~ [ linux-2.4.37.11 ] ~ [ unix-v6-master ] ~ [ ccs-tools-1.8.5 ] ~ [ policy-sample ] ~
Architecture: ~ [ i386 ] ~ [ alpha ] ~ [ m68k ] ~ [ mips ] ~ [ ppc ] ~ [ sparc ] ~ [ sparc64 ] ~

  1 /*
  2  *  kernel/cpuset.c
  3  *
  4  *  Processor and Memory placement constraints for sets of tasks.
  5  *
  6  *  Copyright (C) 2003 BULL SA.
  7  *  Copyright (C) 2004-2007 Silicon Graphics, Inc.
  8  *  Copyright (C) 2006 Google, Inc
  9  *
 10  *  Portions derived from Patrick Mochel's sysfs code.
 11  *  sysfs is Copyright (c) 2001-3 Patrick Mochel
 12  *
 13  *  2003-10-10 Written by Simon Derr.
 14  *  2003-10-22 Updates by Stephen Hemminger.
 15  *  2004 May-July Rework by Paul Jackson.
 16  *  2006 Rework by Paul Menage to use generic cgroups
 17  *  2008 Rework of the scheduler domains and CPU hotplug handling
 18  *       by Max Krasnyansky
 19  *
 20  *  This file is subject to the terms and conditions of the GNU General Public
 21  *  License.  See the file COPYING in the main directory of the Linux
 22  *  distribution for more details.
 23  */
 24 
 25 #include <linux/cpu.h>
 26 #include <linux/cpumask.h>
 27 #include <linux/cpuset.h>
 28 #include <linux/err.h>
 29 #include <linux/errno.h>
 30 #include <linux/file.h>
 31 #include <linux/fs.h>
 32 #include <linux/init.h>
 33 #include <linux/interrupt.h>
 34 #include <linux/kernel.h>
 35 #include <linux/kmod.h>
 36 #include <linux/list.h>
 37 #include <linux/mempolicy.h>
 38 #include <linux/mm.h>
 39 #include <linux/memory.h>
 40 #include <linux/export.h>
 41 #include <linux/mount.h>
 42 #include <linux/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 mutexes guarding cpuset structures - cpuset_mutex
252  * and callback_mutex.  The latter may nest inside the former.  We also
253  * require taking task_lock() when dereferencing a task's cpuset pointer.
254  * See "The task_lock() exception", at the end of this comment.
255  *
256  * A task must hold both mutexes 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_mutex 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_mutex to query cpusets.
263  * Once it is ready to make the changes, it takes callback_mutex, blocking
264  * everyone else.
265  *
266  * Calls to the kernel memory allocator can not be made while holding
267  * callback_mutex, as that would risk double tripping on callback_mutex
268  * from one of the callbacks into the cpuset code from within
269  * __alloc_pages().
270  *
271  * If a task is only holding callback_mutex, 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_mutex 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_MUTEX(callback_mutex);
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.
327  *
328  * One way or another, we guarantee to return some non-empty subset
329  * of cpu_online_mask.
330  *
331  * Call with callback_mutex held.
332  */
333 static void guarantee_online_cpus(struct cpuset *cs, struct cpumask *pmask)
334 {
335         while (!cpumask_intersects(cs->effective_cpus, cpu_online_mask)) {
336                 cs = parent_cs(cs);
337                 if (unlikely(!cs)) {
338                         /*
339                          * The top cpuset doesn't have any online cpu as a
340                          * consequence of a race between cpuset_hotplug_work
341                          * and cpu hotplug notifier.  But we know the top
342                          * cpuset's effective_cpus is on its way to to be
343                          * identical to cpu_online_mask.
344                          */
345                         cpumask_copy(pmask, cpu_online_mask);
346                         return;
347                 }
348         }
349         cpumask_and(pmask, cs->effective_cpus, cpu_online_mask);
350 }
351 
352 /*
353  * Return in *pmask the portion of a cpusets's mems_allowed that
354  * are online, with memory.  If none are online with memory, walk
355  * up the cpuset hierarchy until we find one that does have some
356  * online mems.  The top cpuset always has some mems online.
357  *
358  * One way or another, we guarantee to return some non-empty subset
359  * of node_states[N_MEMORY].
360  *
361  * Call with callback_mutex held.
362  */
363 static void guarantee_online_mems(struct cpuset *cs, nodemask_t *pmask)
364 {
365         while (!nodes_intersects(cs->effective_mems, node_states[N_MEMORY]))
366                 cs = parent_cs(cs);
367         nodes_and(*pmask, cs->effective_mems, node_states[N_MEMORY]);
368 }
369 
370 /*
371  * update task's spread flag if cpuset's page/slab spread flag is set
372  *
373  * Called with callback_mutex/cpuset_mutex held
374  */
375 static void cpuset_update_task_spread_flag(struct cpuset *cs,
376                                         struct task_struct *tsk)
377 {
378         if (is_spread_page(cs))
379                 task_set_spread_page(tsk);
380         else
381                 task_clear_spread_page(tsk);
382 
383         if (is_spread_slab(cs))
384                 task_set_spread_slab(tsk);
385         else
386                 task_clear_spread_slab(tsk);
387 }
388 
389 /*
390  * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
391  *
392  * One cpuset is a subset of another if all its allowed CPUs and
393  * Memory Nodes are a subset of the other, and its exclusive flags
394  * are only set if the other's are set.  Call holding cpuset_mutex.
395  */
396 
397 static int is_cpuset_subset(const struct cpuset *p, const struct cpuset *q)
398 {
399         return  cpumask_subset(p->cpus_allowed, q->cpus_allowed) &&
400                 nodes_subset(p->mems_allowed, q->mems_allowed) &&
401                 is_cpu_exclusive(p) <= is_cpu_exclusive(q) &&
402                 is_mem_exclusive(p) <= is_mem_exclusive(q);
403 }
404 
405 /**
406  * alloc_trial_cpuset - allocate a trial cpuset
407  * @cs: the cpuset that the trial cpuset duplicates
408  */
409 static struct cpuset *alloc_trial_cpuset(struct cpuset *cs)
410 {
411         struct cpuset *trial;
412 
413         trial = kmemdup(cs, sizeof(*cs), GFP_KERNEL);
414         if (!trial)
415                 return NULL;
416 
417         if (!alloc_cpumask_var(&trial->cpus_allowed, GFP_KERNEL))
418                 goto free_cs;
419         if (!alloc_cpumask_var(&trial->effective_cpus, GFP_KERNEL))
420                 goto free_cpus;
421 
422         cpumask_copy(trial->cpus_allowed, cs->cpus_allowed);
423         cpumask_copy(trial->effective_cpus, cs->effective_cpus);
424         return trial;
425 
426 free_cpus:
427         free_cpumask_var(trial->cpus_allowed);
428 free_cs:
429         kfree(trial);
430         return NULL;
431 }
432 
433 /**
434  * free_trial_cpuset - free the trial cpuset
435  * @trial: the trial cpuset to be freed
436  */
437 static void free_trial_cpuset(struct cpuset *trial)
438 {
439         free_cpumask_var(trial->effective_cpus);
440         free_cpumask_var(trial->cpus_allowed);
441         kfree(trial);
442 }
443 
444 /*
445  * validate_change() - Used to validate that any proposed cpuset change
446  *                     follows the structural rules for cpusets.
447  *
448  * If we replaced the flag and mask values of the current cpuset
449  * (cur) with those values in the trial cpuset (trial), would
450  * our various subset and exclusive rules still be valid?  Presumes
451  * cpuset_mutex held.
452  *
453  * 'cur' is the address of an actual, in-use cpuset.  Operations
454  * such as list traversal that depend on the actual address of the
455  * cpuset in the list must use cur below, not trial.
456  *
457  * 'trial' is the address of bulk structure copy of cur, with
458  * perhaps one or more of the fields cpus_allowed, mems_allowed,
459  * or flags changed to new, trial values.
460  *
461  * Return 0 if valid, -errno if not.
462  */
463 
464 static int validate_change(struct cpuset *cur, struct cpuset *trial)
465 {
466         struct cgroup_subsys_state *css;
467         struct cpuset *c, *par;
468         int ret;
469 
470         rcu_read_lock();
471 
472         /* Each of our child cpusets must be a subset of us */
473         ret = -EBUSY;
474         cpuset_for_each_child(c, css, cur)
475                 if (!is_cpuset_subset(c, trial))
476                         goto out;
477 
478         /* Remaining checks don't apply to root cpuset */
479         ret = 0;
480         if (cur == &top_cpuset)
481                 goto out;
482 
483         par = parent_cs(cur);
484 
485         /* On legacy hiearchy, we must be a subset of our parent cpuset. */
486         ret = -EACCES;
487         if (!cgroup_on_dfl(cur->css.cgroup) && !is_cpuset_subset(trial, par))
488                 goto out;
489 
490         /*
491          * If either I or some sibling (!= me) is exclusive, we can't
492          * overlap
493          */
494         ret = -EINVAL;
495         cpuset_for_each_child(c, css, par) {
496                 if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
497                     c != cur &&
498                     cpumask_intersects(trial->cpus_allowed, c->cpus_allowed))
499                         goto out;
500                 if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) &&
501                     c != cur &&
502                     nodes_intersects(trial->mems_allowed, c->mems_allowed))
503                         goto out;
504         }
505 
506         /*
507          * Cpusets with tasks - existing or newly being attached - can't
508          * be changed to have empty cpus_allowed or mems_allowed.
509          */
510         ret = -ENOSPC;
511         if ((cgroup_has_tasks(cur->css.cgroup) || cur->attach_in_progress)) {
512                 if (!cpumask_empty(cur->cpus_allowed) &&
513                     cpumask_empty(trial->cpus_allowed))
514                         goto out;
515                 if (!nodes_empty(cur->mems_allowed) &&
516                     nodes_empty(trial->mems_allowed))
517                         goto out;
518         }
519 
520         ret = 0;
521 out:
522         rcu_read_unlock();
523         return ret;
524 }
525 
526 #ifdef CONFIG_SMP
527 /*
528  * Helper routine for generate_sched_domains().
529  * Do cpusets a, b have overlapping effective cpus_allowed masks?
530  */
531 static int cpusets_overlap(struct cpuset *a, struct cpuset *b)
532 {
533         return cpumask_intersects(a->effective_cpus, b->effective_cpus);
534 }
535 
536 static void
537 update_domain_attr(struct sched_domain_attr *dattr, struct cpuset *c)
538 {
539         if (dattr->relax_domain_level < c->relax_domain_level)
540                 dattr->relax_domain_level = c->relax_domain_level;
541         return;
542 }
543 
544 static void update_domain_attr_tree(struct sched_domain_attr *dattr,
545                                     struct cpuset *root_cs)
546 {
547         struct cpuset *cp;
548         struct cgroup_subsys_state *pos_css;
549 
550         rcu_read_lock();
551         cpuset_for_each_descendant_pre(cp, pos_css, root_cs) {
552                 /* skip the whole subtree if @cp doesn't have any CPU */
553                 if (cpumask_empty(cp->cpus_allowed)) {
554                         pos_css = css_rightmost_descendant(pos_css);
555                         continue;
556                 }
557 
558                 if (is_sched_load_balance(cp))
559                         update_domain_attr(dattr, cp);
560         }
561         rcu_read_unlock();
562 }
563 
564 /*
565  * generate_sched_domains()
566  *
567  * This function builds a partial partition of the systems CPUs
568  * A 'partial partition' is a set of non-overlapping subsets whose
569  * union is a subset of that set.
570  * The output of this function needs to be passed to kernel/sched/core.c
571  * partition_sched_domains() routine, which will rebuild the scheduler's
572  * load balancing domains (sched domains) as specified by that partial
573  * partition.
574  *
575  * See "What is sched_load_balance" in Documentation/cgroups/cpusets.txt
576  * for a background explanation of this.
577  *
578  * Does not return errors, on the theory that the callers of this
579  * routine would rather not worry about failures to rebuild sched
580  * domains when operating in the severe memory shortage situations
581  * that could cause allocation failures below.
582  *
583  * Must be called with cpuset_mutex held.
584  *
585  * The three key local variables below are:
586  *    q  - a linked-list queue of cpuset pointers, used to implement a
587  *         top-down scan of all cpusets.  This scan loads a pointer
588  *         to each cpuset marked is_sched_load_balance into the
589  *         array 'csa'.  For our purposes, rebuilding the schedulers
590  *         sched domains, we can ignore !is_sched_load_balance cpusets.
591  *  csa  - (for CpuSet Array) Array of pointers to all the cpusets
592  *         that need to be load balanced, for convenient iterative
593  *         access by the subsequent code that finds the best partition,
594  *         i.e the set of domains (subsets) of CPUs such that the
595  *         cpus_allowed of every cpuset marked is_sched_load_balance
596  *         is a subset of one of these domains, while there are as
597  *         many such domains as possible, each as small as possible.
598  * doms  - Conversion of 'csa' to an array of cpumasks, for passing to
599  *         the kernel/sched/core.c routine partition_sched_domains() in a
600  *         convenient format, that can be easily compared to the prior
601  *         value to determine what partition elements (sched domains)
602  *         were changed (added or removed.)
603  *
604  * Finding the best partition (set of domains):
605  *      The triple nested loops below over i, j, k scan over the
606  *      load balanced cpusets (using the array of cpuset pointers in
607  *      csa[]) looking for pairs of cpusets that have overlapping
608  *      cpus_allowed, but which don't have the same 'pn' partition
609  *      number and gives them in the same partition number.  It keeps
610  *      looping on the 'restart' label until it can no longer find
611  *      any such pairs.
612  *
613  *      The union of the cpus_allowed masks from the set of
614  *      all cpusets having the same 'pn' value then form the one
615  *      element of the partition (one sched domain) to be passed to
616  *      partition_sched_domains().
617  */
618 static int generate_sched_domains(cpumask_var_t **domains,
619                         struct sched_domain_attr **attributes)
620 {
621         struct cpuset *cp;      /* scans q */
622         struct cpuset **csa;    /* array of all cpuset ptrs */
623         int csn;                /* how many cpuset ptrs in csa so far */
624         int i, j, k;            /* indices for partition finding loops */
625         cpumask_var_t *doms;    /* resulting partition; i.e. sched domains */
626         struct sched_domain_attr *dattr;  /* attributes for custom domains */
627         int ndoms = 0;          /* number of sched domains in result */
628         int nslot;              /* next empty doms[] struct cpumask slot */
629         struct cgroup_subsys_state *pos_css;
630 
631         doms = NULL;
632         dattr = NULL;
633         csa = NULL;
634 
635         /* Special case for the 99% of systems with one, full, sched domain */
636         if (is_sched_load_balance(&top_cpuset)) {
637                 ndoms = 1;
638                 doms = alloc_sched_domains(ndoms);
639                 if (!doms)
640                         goto done;
641 
642                 dattr = kmalloc(sizeof(struct sched_domain_attr), GFP_KERNEL);
643                 if (dattr) {
644                         *dattr = SD_ATTR_INIT;
645                         update_domain_attr_tree(dattr, &top_cpuset);
646                 }
647                 cpumask_copy(doms[0], top_cpuset.effective_cpus);
648 
649                 goto done;
650         }
651 
652         csa = kmalloc(nr_cpusets() * sizeof(cp), GFP_KERNEL);
653         if (!csa)
654                 goto done;
655         csn = 0;
656 
657         rcu_read_lock();
658         cpuset_for_each_descendant_pre(cp, pos_css, &top_cpuset) {
659                 if (cp == &top_cpuset)
660                         continue;
661                 /*
662                  * Continue traversing beyond @cp iff @cp has some CPUs and
663                  * isn't load balancing.  The former is obvious.  The
664                  * latter: All child cpusets contain a subset of the
665                  * parent's cpus, so just skip them, and then we call
666                  * update_domain_attr_tree() to calc relax_domain_level of
667                  * the corresponding sched domain.
668                  */
669                 if (!cpumask_empty(cp->cpus_allowed) &&
670                     !is_sched_load_balance(cp))
671                         continue;
672 
673                 if (is_sched_load_balance(cp))
674                         csa[csn++] = cp;
675 
676                 /* skip @cp's subtree */
677                 pos_css = css_rightmost_descendant(pos_css);
678         }
679         rcu_read_unlock();
680 
681         for (i = 0; i < csn; i++)
682                 csa[i]->pn = i;
683         ndoms = csn;
684 
685 restart:
686         /* Find the best partition (set of sched domains) */
687         for (i = 0; i < csn; i++) {
688                 struct cpuset *a = csa[i];
689                 int apn = a->pn;
690 
691                 for (j = 0; j < csn; j++) {
692                         struct cpuset *b = csa[j];
693                         int bpn = b->pn;
694 
695                         if (apn != bpn && cpusets_overlap(a, b)) {
696                                 for (k = 0; k < csn; k++) {
697                                         struct cpuset *c = csa[k];
698 
699                                         if (c->pn == bpn)
700                                                 c->pn = apn;
701                                 }
702                                 ndoms--;        /* one less element */
703                                 goto restart;
704                         }
705                 }
706         }
707 
708         /*
709          * Now we know how many domains to create.
710          * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
711          */
712         doms = alloc_sched_domains(ndoms);
713         if (!doms)
714                 goto done;
715 
716         /*
717          * The rest of the code, including the scheduler, can deal with
718          * dattr==NULL case. No need to abort if alloc fails.
719          */
720         dattr = kmalloc(ndoms * sizeof(struct sched_domain_attr), GFP_KERNEL);
721 
722         for (nslot = 0, i = 0; i < csn; i++) {
723                 struct cpuset *a = csa[i];
724                 struct cpumask *dp;
725                 int apn = a->pn;
726 
727                 if (apn < 0) {
728                         /* Skip completed partitions */
729                         continue;
730                 }
731 
732                 dp = doms[nslot];
733 
734                 if (nslot == ndoms) {
735                         static int warnings = 10;
736                         if (warnings) {
737                                 pr_warn("rebuild_sched_domains confused: nslot %d, ndoms %d, csn %d, i %d, apn %d\n",
738                                         nslot, ndoms, csn, i, apn);
739                                 warnings--;
740                         }
741                         continue;
742                 }
743 
744                 cpumask_clear(dp);
745                 if (dattr)
746                         *(dattr + nslot) = SD_ATTR_INIT;
747                 for (j = i; j < csn; j++) {
748                         struct cpuset *b = csa[j];
749 
750                         if (apn == b->pn) {
751                                 cpumask_or(dp, dp, b->effective_cpus);
752                                 if (dattr)
753                                         update_domain_attr_tree(dattr + nslot, b);
754 
755                                 /* Done with this partition */
756                                 b->pn = -1;
757                         }
758                 }
759                 nslot++;
760         }
761         BUG_ON(nslot != ndoms);
762 
763 done:
764         kfree(csa);
765 
766         /*
767          * Fallback to the default domain if kmalloc() failed.
768          * See comments in partition_sched_domains().
769          */
770         if (doms == NULL)
771                 ndoms = 1;
772 
773         *domains    = doms;
774         *attributes = dattr;
775         return ndoms;
776 }
777 
778 /*
779  * Rebuild scheduler domains.
780  *
781  * If the flag 'sched_load_balance' of any cpuset with non-empty
782  * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
783  * which has that flag enabled, or if any cpuset with a non-empty
784  * 'cpus' is removed, then call this routine to rebuild the
785  * scheduler's dynamic sched domains.
786  *
787  * Call with cpuset_mutex held.  Takes get_online_cpus().
788  */
789 static void rebuild_sched_domains_locked(void)
790 {
791         struct sched_domain_attr *attr;
792         cpumask_var_t *doms;
793         int ndoms;
794 
795         lockdep_assert_held(&cpuset_mutex);
796         get_online_cpus();
797 
798         /*
799          * We have raced with CPU hotplug. Don't do anything to avoid
800          * passing doms with offlined cpu to partition_sched_domains().
801          * Anyways, hotplug work item will rebuild sched domains.
802          */
803         if (!cpumask_equal(top_cpuset.effective_cpus, cpu_active_mask))
804                 goto out;
805 
806         /* Generate domain masks and attrs */
807         ndoms = generate_sched_domains(&doms, &attr);
808 
809         /* Have scheduler rebuild the domains */
810         partition_sched_domains(ndoms, doms, attr);
811 out:
812         put_online_cpus();
813 }
814 #else /* !CONFIG_SMP */
815 static void rebuild_sched_domains_locked(void)
816 {
817 }
818 #endif /* CONFIG_SMP */
819 
820 void rebuild_sched_domains(void)
821 {
822         mutex_lock(&cpuset_mutex);
823         rebuild_sched_domains_locked();
824         mutex_unlock(&cpuset_mutex);
825 }
826 
827 /**
828  * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
829  * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
830  *
831  * Iterate through each task of @cs updating its cpus_allowed to the
832  * effective cpuset's.  As this function is called with cpuset_mutex held,
833  * cpuset membership stays stable.
834  */
835 static void update_tasks_cpumask(struct cpuset *cs)
836 {
837         struct css_task_iter it;
838         struct task_struct *task;
839 
840         css_task_iter_start(&cs->css, &it);
841         while ((task = css_task_iter_next(&it)))
842                 set_cpus_allowed_ptr(task, cs->effective_cpus);
843         css_task_iter_end(&it);
844 }
845 
846 /*
847  * update_cpumasks_hier - Update effective cpumasks and tasks in the subtree
848  * @cs: the cpuset to consider
849  * @new_cpus: temp variable for calculating new effective_cpus
850  *
851  * When congifured cpumask is changed, the effective cpumasks of this cpuset
852  * and all its descendants need to be updated.
853  *
854  * On legacy hierachy, effective_cpus will be the same with cpu_allowed.
855  *
856  * Called with cpuset_mutex held
857  */
858 static void update_cpumasks_hier(struct cpuset *cs, struct cpumask *new_cpus)
859 {
860         struct cpuset *cp;
861         struct cgroup_subsys_state *pos_css;
862         bool need_rebuild_sched_domains = false;
863 
864         rcu_read_lock();
865         cpuset_for_each_descendant_pre(cp, pos_css, cs) {
866                 struct cpuset *parent = parent_cs(cp);
867 
868                 cpumask_and(new_cpus, cp->cpus_allowed, parent->effective_cpus);
869 
870                 /*
871                  * If it becomes empty, inherit the effective mask of the
872                  * parent, which is guaranteed to have some CPUs.
873                  */
874                 if (cgroup_on_dfl(cp->css.cgroup) && cpumask_empty(new_cpus))
875                         cpumask_copy(new_cpus, parent->effective_cpus);
876 
877                 /* Skip the whole subtree if the cpumask remains the same. */
878                 if (cpumask_equal(new_cpus, cp->effective_cpus)) {
879                         pos_css = css_rightmost_descendant(pos_css);
880                         continue;
881                 }
882 
883                 if (!css_tryget_online(&cp->css))
884                         continue;
885                 rcu_read_unlock();
886 
887                 mutex_lock(&callback_mutex);
888                 cpumask_copy(cp->effective_cpus, new_cpus);
889                 mutex_unlock(&callback_mutex);
890 
891                 WARN_ON(!cgroup_on_dfl(cp->css.cgroup) &&
892                         !cpumask_equal(cp->cpus_allowed, cp->effective_cpus));
893 
894                 update_tasks_cpumask(cp);
895 
896                 /*
897                  * If the effective cpumask of any non-empty cpuset is changed,
898                  * we need to rebuild sched domains.
899                  */
900                 if (!cpumask_empty(cp->cpus_allowed) &&
901                     is_sched_load_balance(cp))
902                         need_rebuild_sched_domains = true;
903 
904                 rcu_read_lock();
905                 css_put(&cp->css);
906         }
907         rcu_read_unlock();
908 
909         if (need_rebuild_sched_domains)
910                 rebuild_sched_domains_locked();
911 }
912 
913 /**
914  * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
915  * @cs: the cpuset to consider
916  * @trialcs: trial cpuset
917  * @buf: buffer of cpu numbers written to this cpuset
918  */
919 static int update_cpumask(struct cpuset *cs, struct cpuset *trialcs,
920                           const char *buf)
921 {
922         int retval;
923 
924         /* top_cpuset.cpus_allowed tracks cpu_online_mask; it's read-only */
925         if (cs == &top_cpuset)
926                 return -EACCES;
927 
928         /*
929          * An empty cpus_allowed is ok only if the cpuset has no tasks.
930          * Since cpulist_parse() fails on an empty mask, we special case
931          * that parsing.  The validate_change() call ensures that cpusets
932          * with tasks have cpus.
933          */
934         if (!*buf) {
935                 cpumask_clear(trialcs->cpus_allowed);
936         } else {
937                 retval = cpulist_parse(buf, trialcs->cpus_allowed);
938                 if (retval < 0)
939                         return retval;
940 
941                 if (!cpumask_subset(trialcs->cpus_allowed,
942                                     top_cpuset.cpus_allowed))
943                         return -EINVAL;
944         }
945 
946         /* Nothing to do if the cpus didn't change */
947         if (cpumask_equal(cs->cpus_allowed, trialcs->cpus_allowed))
948                 return 0;
949 
950         retval = validate_change(cs, trialcs);
951         if (retval < 0)
952                 return retval;
953 
954         mutex_lock(&callback_mutex);
955         cpumask_copy(cs->cpus_allowed, trialcs->cpus_allowed);
956         mutex_unlock(&callback_mutex);
957 
958         /* use trialcs->cpus_allowed as a temp variable */
959         update_cpumasks_hier(cs, trialcs->cpus_allowed);
960         return 0;
961 }
962 
963 /*
964  * cpuset_migrate_mm
965  *
966  *    Migrate memory region from one set of nodes to another.
967  *
968  *    Temporarilly set tasks mems_allowed to target nodes of migration,
969  *    so that the migration code can allocate pages on these nodes.
970  *
971  *    While the mm_struct we are migrating is typically from some
972  *    other task, the task_struct mems_allowed that we are hacking
973  *    is for our current task, which must allocate new pages for that
974  *    migrating memory region.
975  */
976 
977 static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
978                                                         const nodemask_t *to)
979 {
980         struct task_struct *tsk = current;
981 
982         tsk->mems_allowed = *to;
983 
984         do_migrate_pages(mm, from, to, MPOL_MF_MOVE_ALL);
985 
986         rcu_read_lock();
987         guarantee_online_mems(task_cs(tsk), &tsk->mems_allowed);
988         rcu_read_unlock();
989 }
990 
991 /*
992  * cpuset_change_task_nodemask - change task's mems_allowed and mempolicy
993  * @tsk: the task to change
994  * @newmems: new nodes that the task will be set
995  *
996  * In order to avoid seeing no nodes if the old and new nodes are disjoint,
997  * we structure updates as setting all new allowed nodes, then clearing newly
998  * disallowed ones.
999  */
1000 static void cpuset_change_task_nodemask(struct task_struct *tsk,
1001                                         nodemask_t *newmems)
1002 {
1003         bool need_loop;
1004 
1005         /*
1006          * Allow tasks that have access to memory reserves because they have
1007          * been OOM killed to get memory anywhere.
1008          */
1009         if (unlikely(test_thread_flag(TIF_MEMDIE)))
1010                 return;
1011         if (current->flags & PF_EXITING) /* Let dying task have memory */
1012                 return;
1013 
1014         task_lock(tsk);
1015         /*
1016          * Determine if a loop is necessary if another thread is doing
1017          * read_mems_allowed_begin().  If at least one node remains unchanged and
1018          * tsk does not have a mempolicy, then an empty nodemask will not be
1019          * possible when mems_allowed is larger than a word.
1020          */
1021         need_loop = task_has_mempolicy(tsk) ||
1022                         !nodes_intersects(*newmems, tsk->mems_allowed);
1023 
1024         if (need_loop) {
1025                 local_irq_disable();
1026                 write_seqcount_begin(&tsk->mems_allowed_seq);
1027         }
1028 
1029         nodes_or(tsk->mems_allowed, tsk->mems_allowed, *newmems);
1030         mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP1);
1031 
1032         mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP2);
1033         tsk->mems_allowed = *newmems;
1034 
1035         if (need_loop) {
1036                 write_seqcount_end(&tsk->mems_allowed_seq);
1037                 local_irq_enable();
1038         }
1039 
1040         task_unlock(tsk);
1041 }
1042 
1043 static void *cpuset_being_rebound;
1044 
1045 /**
1046  * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
1047  * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
1048  *
1049  * Iterate through each task of @cs updating its mems_allowed to the
1050  * effective cpuset's.  As this function is called with cpuset_mutex held,
1051  * cpuset membership stays stable.
1052  */
1053 static void update_tasks_nodemask(struct cpuset *cs)
1054 {
1055         static nodemask_t newmems;      /* protected by cpuset_mutex */
1056         struct css_task_iter it;
1057         struct task_struct *task;
1058 
1059         cpuset_being_rebound = cs;              /* causes mpol_dup() rebind */
1060 
1061         guarantee_online_mems(cs, &newmems);
1062 
1063         /*
1064          * The mpol_rebind_mm() call takes mmap_sem, which we couldn't
1065          * take while holding tasklist_lock.  Forks can happen - the
1066          * mpol_dup() cpuset_being_rebound check will catch such forks,
1067          * and rebind their vma mempolicies too.  Because we still hold
1068          * the global cpuset_mutex, we know that no other rebind effort
1069          * will be contending for the global variable cpuset_being_rebound.
1070          * It's ok if we rebind the same mm twice; mpol_rebind_mm()
1071          * is idempotent.  Also migrate pages in each mm to new nodes.
1072          */
1073         css_task_iter_start(&cs->css, &it);
1074         while ((task = css_task_iter_next(&it))) {
1075                 struct mm_struct *mm;
1076                 bool migrate;
1077 
1078                 cpuset_change_task_nodemask(task, &newmems);
1079 
1080                 mm = get_task_mm(task);
1081                 if (!mm)
1082                         continue;
1083 
1084                 migrate = is_memory_migrate(cs);
1085 
1086                 mpol_rebind_mm(mm, &cs->mems_allowed);
1087                 if (migrate)
1088                         cpuset_migrate_mm(mm, &cs->old_mems_allowed, &newmems);
1089                 mmput(mm);
1090         }
1091         css_task_iter_end(&it);
1092 
1093         /*
1094          * All the tasks' nodemasks have been updated, update
1095          * cs->old_mems_allowed.
1096          */
1097         cs->old_mems_allowed = newmems;
1098 
1099         /* We're done rebinding vmas to this cpuset's new mems_allowed. */
1100         cpuset_being_rebound = NULL;
1101 }
1102 
1103 /*
1104  * update_nodemasks_hier - Update effective nodemasks and tasks in the subtree
1105  * @cs: the cpuset to consider
1106  * @new_mems: a temp variable for calculating new effective_mems
1107  *
1108  * When configured nodemask is changed, the effective nodemasks of this cpuset
1109  * and all its descendants need to be updated.
1110  *
1111  * On legacy hiearchy, effective_mems will be the same with mems_allowed.
1112  *
1113  * Called with cpuset_mutex held
1114  */
1115 static void update_nodemasks_hier(struct cpuset *cs, nodemask_t *new_mems)
1116 {
1117         struct cpuset *cp;
1118         struct cgroup_subsys_state *pos_css;
1119 
1120         rcu_read_lock();
1121         cpuset_for_each_descendant_pre(cp, pos_css, cs) {
1122                 struct cpuset *parent = parent_cs(cp);
1123 
1124                 nodes_and(*new_mems, cp->mems_allowed, parent->effective_mems);
1125 
1126                 /*
1127                  * If it becomes empty, inherit the effective mask of the
1128                  * parent, which is guaranteed to have some MEMs.
1129                  */
1130                 if (cgroup_on_dfl(cp->css.cgroup) && nodes_empty(*new_mems))
1131                         *new_mems = parent->effective_mems;
1132 
1133                 /* Skip the whole subtree if the nodemask remains the same. */
1134                 if (nodes_equal(*new_mems, cp->effective_mems)) {
1135                         pos_css = css_rightmost_descendant(pos_css);
1136                         continue;
1137                 }
1138 
1139                 if (!css_tryget_online(&cp->css))
1140                         continue;
1141                 rcu_read_unlock();
1142 
1143                 mutex_lock(&callback_mutex);
1144                 cp->effective_mems = *new_mems;
1145                 mutex_unlock(&callback_mutex);
1146 
1147                 WARN_ON(!cgroup_on_dfl(cp->css.cgroup) &&
1148                         !nodes_equal(cp->mems_allowed, cp->effective_mems));
1149 
1150                 update_tasks_nodemask(cp);
1151 
1152                 rcu_read_lock();
1153                 css_put(&cp->css);
1154         }
1155         rcu_read_unlock();
1156 }
1157 
1158 /*
1159  * Handle user request to change the 'mems' memory placement
1160  * of a cpuset.  Needs to validate the request, update the
1161  * cpusets mems_allowed, and for each task in the cpuset,
1162  * update mems_allowed and rebind task's mempolicy and any vma
1163  * mempolicies and if the cpuset is marked 'memory_migrate',
1164  * migrate the tasks pages to the new memory.
1165  *
1166  * Call with cpuset_mutex held.  May take callback_mutex during call.
1167  * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
1168  * lock each such tasks mm->mmap_sem, scan its vma's and rebind
1169  * their mempolicies to the cpusets new mems_allowed.
1170  */
1171 static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
1172                            const char *buf)
1173 {
1174         int retval;
1175 
1176         /*
1177          * top_cpuset.mems_allowed tracks node_stats[N_MEMORY];
1178          * it's read-only
1179          */
1180         if (cs == &top_cpuset) {
1181                 retval = -EACCES;
1182                 goto done;
1183         }
1184 
1185         /*
1186          * An empty mems_allowed is ok iff there are no tasks in the cpuset.
1187          * Since nodelist_parse() fails on an empty mask, we special case
1188          * that parsing.  The validate_change() call ensures that cpusets
1189          * with tasks have memory.
1190          */
1191         if (!*buf) {
1192                 nodes_clear(trialcs->mems_allowed);
1193         } else {
1194                 retval = nodelist_parse(buf, trialcs->mems_allowed);
1195                 if (retval < 0)
1196                         goto done;
1197 
1198                 if (!nodes_subset(trialcs->mems_allowed,
1199                                   top_cpuset.mems_allowed)) {
1200                         retval = -EINVAL;
1201                         goto done;
1202                 }
1203         }
1204 
1205         if (nodes_equal(cs->mems_allowed, trialcs->mems_allowed)) {
1206                 retval = 0;             /* Too easy - nothing to do */
1207                 goto done;
1208         }
1209         retval = validate_change(cs, trialcs);
1210         if (retval < 0)
1211                 goto done;
1212 
1213         mutex_lock(&callback_mutex);
1214         cs->mems_allowed = trialcs->mems_allowed;
1215         mutex_unlock(&callback_mutex);
1216 
1217         /* use trialcs->mems_allowed as a temp variable */
1218         update_nodemasks_hier(cs, &trialcs->mems_allowed);
1219 done:
1220         return retval;
1221 }
1222 
1223 int current_cpuset_is_being_rebound(void)
1224 {
1225         int ret;
1226 
1227         rcu_read_lock();
1228         ret = task_cs(current) == cpuset_being_rebound;
1229         rcu_read_unlock();
1230 
1231         return ret;
1232 }
1233 
1234 static int update_relax_domain_level(struct cpuset *cs, s64 val)
1235 {
1236 #ifdef CONFIG_SMP
1237         if (val < -1 || val >= sched_domain_level_max)
1238                 return -EINVAL;
1239 #endif
1240 
1241         if (val != cs->relax_domain_level) {
1242                 cs->relax_domain_level = val;
1243                 if (!cpumask_empty(cs->cpus_allowed) &&
1244                     is_sched_load_balance(cs))
1245                         rebuild_sched_domains_locked();
1246         }
1247 
1248         return 0;
1249 }
1250 
1251 /**
1252  * update_tasks_flags - update the spread flags of tasks in the cpuset.
1253  * @cs: the cpuset in which each task's spread flags needs to be changed
1254  *
1255  * Iterate through each task of @cs updating its spread flags.  As this
1256  * function is called with cpuset_mutex held, cpuset membership stays
1257  * stable.
1258  */
1259 static void update_tasks_flags(struct cpuset *cs)
1260 {
1261         struct css_task_iter it;
1262         struct task_struct *task;
1263 
1264         css_task_iter_start(&cs->css, &it);
1265         while ((task = css_task_iter_next(&it)))
1266                 cpuset_update_task_spread_flag(cs, task);
1267         css_task_iter_end(&it);
1268 }
1269 
1270 /*
1271  * update_flag - read a 0 or a 1 in a file and update associated flag
1272  * bit:         the bit to update (see cpuset_flagbits_t)
1273  * cs:          the cpuset to update
1274  * turning_on:  whether the flag is being set or cleared
1275  *
1276  * Call with cpuset_mutex held.
1277  */
1278 
1279 static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs,
1280                        int turning_on)
1281 {
1282         struct cpuset *trialcs;
1283         int balance_flag_changed;
1284         int spread_flag_changed;
1285         int err;
1286 
1287         trialcs = alloc_trial_cpuset(cs);
1288         if (!trialcs)
1289                 return -ENOMEM;
1290 
1291         if (turning_on)
1292                 set_bit(bit, &trialcs->flags);
1293         else
1294                 clear_bit(bit, &trialcs->flags);
1295 
1296         err = validate_change(cs, trialcs);
1297         if (err < 0)
1298                 goto out;
1299 
1300         balance_flag_changed = (is_sched_load_balance(cs) !=
1301                                 is_sched_load_balance(trialcs));
1302 
1303         spread_flag_changed = ((is_spread_slab(cs) != is_spread_slab(trialcs))
1304                         || (is_spread_page(cs) != is_spread_page(trialcs)));
1305 
1306         mutex_lock(&callback_mutex);
1307         cs->flags = trialcs->flags;
1308         mutex_unlock(&callback_mutex);
1309 
1310         if (!cpumask_empty(trialcs->cpus_allowed) && balance_flag_changed)
1311                 rebuild_sched_domains_locked();
1312 
1313         if (spread_flag_changed)
1314                 update_tasks_flags(cs);
1315 out:
1316         free_trial_cpuset(trialcs);
1317         return err;
1318 }
1319 
1320 /*
1321  * Frequency meter - How fast is some event occurring?
1322  *
1323  * These routines manage a digitally filtered, constant time based,
1324  * event frequency meter.  There are four routines:
1325  *   fmeter_init() - initialize a frequency meter.
1326  *   fmeter_markevent() - called each time the event happens.
1327  *   fmeter_getrate() - returns the recent rate of such events.
1328  *   fmeter_update() - internal routine used to update fmeter.
1329  *
1330  * A common data structure is passed to each of these routines,
1331  * which is used to keep track of the state required to manage the
1332  * frequency meter and its digital filter.
1333  *
1334  * The filter works on the number of events marked per unit time.
1335  * The filter is single-pole low-pass recursive (IIR).  The time unit
1336  * is 1 second.  Arithmetic is done using 32-bit integers scaled to
1337  * simulate 3 decimal digits of precision (multiplied by 1000).
1338  *
1339  * With an FM_COEF of 933, and a time base of 1 second, the filter
1340  * has a half-life of 10 seconds, meaning that if the events quit
1341  * happening, then the rate returned from the fmeter_getrate()
1342  * will be cut in half each 10 seconds, until it converges to zero.
1343  *
1344  * It is not worth doing a real infinitely recursive filter.  If more
1345  * than FM_MAXTICKS ticks have elapsed since the last filter event,
1346  * just compute FM_MAXTICKS ticks worth, by which point the level
1347  * will be stable.
1348  *
1349  * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
1350  * arithmetic overflow in the fmeter_update() routine.
1351  *
1352  * Given the simple 32 bit integer arithmetic used, this meter works
1353  * best for reporting rates between one per millisecond (msec) and
1354  * one per 32 (approx) seconds.  At constant rates faster than one
1355  * per msec it maxes out at values just under 1,000,000.  At constant
1356  * rates between one per msec, and one per second it will stabilize
1357  * to a value N*1000, where N is the rate of events per second.
1358  * At constant rates between one per second and one per 32 seconds,
1359  * it will be choppy, moving up on the seconds that have an event,
1360  * and then decaying until the next event.  At rates slower than
1361  * about one in 32 seconds, it decays all the way back to zero between
1362  * each event.
1363  */
1364 
1365 #define FM_COEF 933             /* coefficient for half-life of 10 secs */
1366 #define FM_MAXTICKS ((time_t)99) /* useless computing more ticks than this */
1367 #define FM_MAXCNT 1000000       /* limit cnt to avoid overflow */
1368 #define FM_SCALE 1000           /* faux fixed point scale */
1369 
1370 /* Initialize a frequency meter */
1371 static void fmeter_init(struct fmeter *fmp)
1372 {
1373         fmp->cnt = 0;
1374         fmp->val = 0;
1375         fmp->time = 0;
1376         spin_lock_init(&fmp->lock);
1377 }
1378 
1379 /* Internal meter update - process cnt events and update value */
1380 static void fmeter_update(struct fmeter *fmp)
1381 {
1382         time_t now = get_seconds();
1383         time_t ticks = now - fmp->time;
1384 
1385         if (ticks == 0)
1386                 return;
1387 
1388         ticks = min(FM_MAXTICKS, ticks);
1389         while (ticks-- > 0)
1390                 fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
1391         fmp->time = now;
1392 
1393         fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
1394         fmp->cnt = 0;
1395 }
1396 
1397 /* Process any previous ticks, then bump cnt by one (times scale). */
1398 static void fmeter_markevent(struct fmeter *fmp)
1399 {
1400         spin_lock(&fmp->lock);
1401         fmeter_update(fmp);
1402         fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
1403         spin_unlock(&fmp->lock);
1404 }
1405 
1406 /* Process any previous ticks, then return current value. */
1407 static int fmeter_getrate(struct fmeter *fmp)
1408 {
1409         int val;
1410 
1411         spin_lock(&fmp->lock);
1412         fmeter_update(fmp);
1413         val = fmp->val;
1414         spin_unlock(&fmp->lock);
1415         return val;
1416 }
1417 
1418 static struct cpuset *cpuset_attach_old_cs;
1419 
1420 /* Called by cgroups to determine if a cpuset is usable; cpuset_mutex held */
1421 static int cpuset_can_attach(struct cgroup_subsys_state *css,
1422                              struct cgroup_taskset *tset)
1423 {
1424         struct cpuset *cs = css_cs(css);
1425         struct task_struct *task;
1426         int ret;
1427 
1428         /* used later by cpuset_attach() */
1429         cpuset_attach_old_cs = task_cs(cgroup_taskset_first(tset));
1430 
1431         mutex_lock(&cpuset_mutex);
1432 
1433         /* allow moving tasks into an empty cpuset if on default hierarchy */
1434         ret = -ENOSPC;
1435         if (!cgroup_on_dfl(css->cgroup) &&
1436             (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed)))
1437                 goto out_unlock;
1438 
1439         cgroup_taskset_for_each(task, tset) {
1440                 /*
1441                  * Kthreads which disallow setaffinity shouldn't be moved
1442                  * to a new cpuset; we don't want to change their cpu
1443                  * affinity and isolating such threads by their set of
1444                  * allowed nodes is unnecessary.  Thus, cpusets are not
1445                  * applicable for such threads.  This prevents checking for
1446                  * success of set_cpus_allowed_ptr() on all attached tasks
1447                  * before cpus_allowed may be changed.
1448                  */
1449                 ret = -EINVAL;
1450                 if (task->flags & PF_NO_SETAFFINITY)
1451                         goto out_unlock;
1452                 ret = security_task_setscheduler(task);
1453                 if (ret)
1454                         goto out_unlock;
1455         }
1456 
1457         /*
1458          * Mark attach is in progress.  This makes validate_change() fail
1459          * changes which zero cpus/mems_allowed.
1460          */
1461         cs->attach_in_progress++;
1462         ret = 0;
1463 out_unlock:
1464         mutex_unlock(&cpuset_mutex);
1465         return ret;
1466 }
1467 
1468 static void cpuset_cancel_attach(struct cgroup_subsys_state *css,
1469                                  struct cgroup_taskset *tset)
1470 {
1471         mutex_lock(&cpuset_mutex);
1472         css_cs(css)->attach_in_progress--;
1473         mutex_unlock(&cpuset_mutex);
1474 }
1475 
1476 /*
1477  * Protected by cpuset_mutex.  cpus_attach is used only by cpuset_attach()
1478  * but we can't allocate it dynamically there.  Define it global and
1479  * allocate from cpuset_init().
1480  */
1481 static cpumask_var_t cpus_attach;
1482 
1483 static void cpuset_attach(struct cgroup_subsys_state *css,
1484                           struct cgroup_taskset *tset)
1485 {
1486         /* static buf protected by cpuset_mutex */
1487         static nodemask_t cpuset_attach_nodemask_to;
1488         struct mm_struct *mm;
1489         struct task_struct *task;
1490         struct task_struct *leader = cgroup_taskset_first(tset);
1491         struct cpuset *cs = css_cs(css);
1492         struct cpuset *oldcs = cpuset_attach_old_cs;
1493 
1494         mutex_lock(&cpuset_mutex);
1495 
1496         /* prepare for attach */
1497         if (cs == &top_cpuset)
1498                 cpumask_copy(cpus_attach, cpu_possible_mask);
1499         else
1500                 guarantee_online_cpus(cs, cpus_attach);
1501 
1502         guarantee_online_mems(cs, &cpuset_attach_nodemask_to);
1503 
1504         cgroup_taskset_for_each(task, tset) {
1505                 /*
1506                  * can_attach beforehand should guarantee that this doesn't
1507                  * fail.  TODO: have a better way to handle failure here
1508                  */
1509                 WARN_ON_ONCE(set_cpus_allowed_ptr(task, cpus_attach));
1510 
1511                 cpuset_change_task_nodemask(task, &cpuset_attach_nodemask_to);
1512                 cpuset_update_task_spread_flag(cs, task);
1513         }
1514 
1515         /*
1516          * Change mm, possibly for multiple threads in a threadgroup. This is
1517          * expensive and may sleep.
1518          */
1519         cpuset_attach_nodemask_to = cs->effective_mems;
1520         mm = get_task_mm(leader);
1521         if (mm) {
1522                 mpol_rebind_mm(mm, &cpuset_attach_nodemask_to);
1523 
1524                 /*
1525                  * old_mems_allowed is the same with mems_allowed here, except
1526                  * if this task is being moved automatically due to hotplug.
1527                  * In that case @mems_allowed has been updated and is empty,
1528                  * so @old_mems_allowed is the right nodesets that we migrate
1529                  * mm from.
1530                  */
1531                 if (is_memory_migrate(cs)) {
1532                         cpuset_migrate_mm(mm, &oldcs->old_mems_allowed,
1533                                           &cpuset_attach_nodemask_to);
1534                 }
1535                 mmput(mm);
1536         }
1537 
1538         cs->old_mems_allowed = cpuset_attach_nodemask_to;
1539 
1540         cs->attach_in_progress--;
1541         if (!cs->attach_in_progress)
1542                 wake_up(&cpuset_attach_wq);
1543 
1544         mutex_unlock(&cpuset_mutex);
1545 }
1546 
1547 /* The various types of files and directories in a cpuset file system */
1548 
1549 typedef enum {
1550         FILE_MEMORY_MIGRATE,
1551         FILE_CPULIST,
1552         FILE_MEMLIST,
1553         FILE_EFFECTIVE_CPULIST,
1554         FILE_EFFECTIVE_MEMLIST,
1555         FILE_CPU_EXCLUSIVE,
1556         FILE_MEM_EXCLUSIVE,
1557         FILE_MEM_HARDWALL,
1558         FILE_SCHED_LOAD_BALANCE,
1559         FILE_SCHED_RELAX_DOMAIN_LEVEL,
1560         FILE_MEMORY_PRESSURE_ENABLED,
1561         FILE_MEMORY_PRESSURE,
1562         FILE_SPREAD_PAGE,
1563         FILE_SPREAD_SLAB,
1564 } cpuset_filetype_t;
1565 
1566 static int cpuset_write_u64(struct cgroup_subsys_state *css, struct cftype *cft,
1567                             u64 val)
1568 {
1569         struct cpuset *cs = css_cs(css);
1570         cpuset_filetype_t type = cft->private;
1571         int retval = 0;
1572 
1573         mutex_lock(&cpuset_mutex);
1574         if (!is_cpuset_online(cs)) {
1575                 retval = -ENODEV;
1576                 goto out_unlock;
1577         }
1578 
1579         switch (type) {
1580         case FILE_CPU_EXCLUSIVE:
1581                 retval = update_flag(CS_CPU_EXCLUSIVE, cs, val);
1582                 break;
1583         case FILE_MEM_EXCLUSIVE:
1584                 retval = update_flag(CS_MEM_EXCLUSIVE, cs, val);
1585                 break;
1586         case FILE_MEM_HARDWALL:
1587                 retval = update_flag(CS_MEM_HARDWALL, cs, val);
1588                 break;
1589         case FILE_SCHED_LOAD_BALANCE:
1590                 retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, val);
1591                 break;
1592         case FILE_MEMORY_MIGRATE:
1593                 retval = update_flag(CS_MEMORY_MIGRATE, cs, val);
1594                 break;
1595         case FILE_MEMORY_PRESSURE_ENABLED:
1596                 cpuset_memory_pressure_enabled = !!val;
1597                 break;
1598         case FILE_MEMORY_PRESSURE:
1599                 retval = -EACCES;
1600                 break;
1601         case FILE_SPREAD_PAGE:
1602                 retval = update_flag(CS_SPREAD_PAGE, cs, val);
1603                 break;
1604         case FILE_SPREAD_SLAB:
1605                 retval = update_flag(CS_SPREAD_SLAB, cs, val);
1606                 break;
1607         default:
1608                 retval = -EINVAL;
1609                 break;
1610         }
1611 out_unlock:
1612         mutex_unlock(&cpuset_mutex);
1613         return retval;
1614 }
1615 
1616 static int cpuset_write_s64(struct cgroup_subsys_state *css, struct cftype *cft,
1617                             s64 val)
1618 {
1619         struct cpuset *cs = css_cs(css);
1620         cpuset_filetype_t type = cft->private;
1621         int retval = -ENODEV;
1622 
1623         mutex_lock(&cpuset_mutex);
1624         if (!is_cpuset_online(cs))
1625                 goto out_unlock;
1626 
1627         switch (type) {
1628         case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1629                 retval = update_relax_domain_level(cs, val);
1630                 break;
1631         default:
1632                 retval = -EINVAL;
1633                 break;
1634         }
1635 out_unlock:
1636         mutex_unlock(&cpuset_mutex);
1637         return retval;
1638 }
1639 
1640 /*
1641  * Common handling for a write to a "cpus" or "mems" file.
1642  */
1643 static ssize_t cpuset_write_resmask(struct kernfs_open_file *of,
1644                                     char *buf, size_t nbytes, loff_t off)
1645 {
1646         struct cpuset *cs = css_cs(of_css(of));
1647         struct cpuset *trialcs;
1648         int retval = -ENODEV;
1649 
1650         buf = strstrip(buf);
1651 
1652         /*
1653          * CPU or memory hotunplug may leave @cs w/o any execution
1654          * resources, in which case the hotplug code asynchronously updates
1655          * configuration and transfers all tasks to the nearest ancestor
1656          * which can execute.
1657          *
1658          * As writes to "cpus" or "mems" may restore @cs's execution
1659          * resources, wait for the previously scheduled operations before
1660          * proceeding, so that we don't end up keep removing tasks added
1661          * after execution capability is restored.
1662          *
1663          * cpuset_hotplug_work calls back into cgroup core via
1664          * cgroup_transfer_tasks() and waiting for it from a cgroupfs
1665          * operation like this one can lead to a deadlock through kernfs
1666          * active_ref protection.  Let's break the protection.  Losing the
1667          * protection is okay as we check whether @cs is online after
1668          * grabbing cpuset_mutex anyway.  This only happens on the legacy
1669          * hierarchies.
1670          */
1671         css_get(&cs->css);
1672         kernfs_break_active_protection(of->kn);
1673         flush_work(&cpuset_hotplug_work);
1674 
1675         mutex_lock(&cpuset_mutex);
1676         if (!is_cpuset_online(cs))
1677                 goto out_unlock;
1678 
1679         trialcs = alloc_trial_cpuset(cs);
1680         if (!trialcs) {
1681                 retval = -ENOMEM;
1682                 goto out_unlock;
1683         }
1684 
1685         switch (of_cft(of)->private) {
1686         case FILE_CPULIST:
1687                 retval = update_cpumask(cs, trialcs, buf);
1688                 break;
1689         case FILE_MEMLIST:
1690                 retval = update_nodemask(cs, trialcs, buf);
1691                 break;
1692         default:
1693                 retval = -EINVAL;
1694                 break;
1695         }
1696 
1697         free_trial_cpuset(trialcs);
1698 out_unlock:
1699         mutex_unlock(&cpuset_mutex);
1700         kernfs_unbreak_active_protection(of->kn);
1701         css_put(&cs->css);
1702         return retval ?: nbytes;
1703 }
1704 
1705 /*
1706  * These ascii lists should be read in a single call, by using a user
1707  * buffer large enough to hold the entire map.  If read in smaller
1708  * chunks, there is no guarantee of atomicity.  Since the display format
1709  * used, list of ranges of sequential numbers, is variable length,
1710  * and since these maps can change value dynamically, one could read
1711  * gibberish by doing partial reads while a list was changing.
1712  */
1713 static int cpuset_common_seq_show(struct seq_file *sf, void *v)
1714 {
1715         struct cpuset *cs = css_cs(seq_css(sf));
1716         cpuset_filetype_t type = seq_cft(sf)->private;
1717         ssize_t count;
1718         char *buf, *s;
1719         int ret = 0;
1720 
1721         count = seq_get_buf(sf, &buf);
1722         s = buf;
1723 
1724         mutex_lock(&callback_mutex);
1725 
1726         switch (type) {
1727         case FILE_CPULIST:
1728                 s += cpulist_scnprintf(s, count, cs->cpus_allowed);
1729                 break;
1730         case FILE_MEMLIST:
1731                 s += nodelist_scnprintf(s, count, cs->mems_allowed);
1732                 break;
1733         case FILE_EFFECTIVE_CPULIST:
1734                 s += cpulist_scnprintf(s, count, cs->effective_cpus);
1735                 break;
1736         case FILE_EFFECTIVE_MEMLIST:
1737                 s += nodelist_scnprintf(s, count, cs->effective_mems);
1738                 break;
1739         default:
1740                 ret = -EINVAL;
1741                 goto out_unlock;
1742         }
1743 
1744         if (s < buf + count - 1) {
1745                 *s++ = '\n';
1746                 seq_commit(sf, s - buf);
1747         } else {
1748                 seq_commit(sf, -1);
1749         }
1750 out_unlock:
1751         mutex_unlock(&callback_mutex);
1752         return ret;
1753 }
1754 
1755 static u64 cpuset_read_u64(struct cgroup_subsys_state *css, struct cftype *cft)
1756 {
1757         struct cpuset *cs = css_cs(css);
1758         cpuset_filetype_t type = cft->private;
1759         switch (type) {
1760         case FILE_CPU_EXCLUSIVE:
1761                 return is_cpu_exclusive(cs);
1762         case FILE_MEM_EXCLUSIVE:
1763                 return is_mem_exclusive(cs);
1764         case FILE_MEM_HARDWALL:
1765                 return is_mem_hardwall(cs);
1766         case FILE_SCHED_LOAD_BALANCE:
1767                 return is_sched_load_balance(cs);
1768         case FILE_MEMORY_MIGRATE:
1769                 return is_memory_migrate(cs);
1770         case FILE_MEMORY_PRESSURE_ENABLED:
1771                 return cpuset_memory_pressure_enabled;
1772         case FILE_MEMORY_PRESSURE:
1773                 return fmeter_getrate(&cs->fmeter);
1774         case FILE_SPREAD_PAGE:
1775                 return is_spread_page(cs);
1776         case FILE_SPREAD_SLAB:
1777                 return is_spread_slab(cs);
1778         default:
1779                 BUG();
1780         }
1781 
1782         /* Unreachable but makes gcc happy */
1783         return 0;
1784 }
1785 
1786 static s64 cpuset_read_s64(struct cgroup_subsys_state *css, struct cftype *cft)
1787 {
1788         struct cpuset *cs = css_cs(css);
1789         cpuset_filetype_t type = cft->private;
1790         switch (type) {
1791         case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1792                 return cs->relax_domain_level;
1793         default:
1794                 BUG();
1795         }
1796 
1797         /* Unrechable but makes gcc happy */
1798         return 0;
1799 }
1800 
1801 
1802 /*
1803  * for the common functions, 'private' gives the type of file
1804  */
1805 
1806 static struct cftype files[] = {
1807         {
1808                 .name = "cpus",
1809                 .seq_show = cpuset_common_seq_show,
1810                 .write = cpuset_write_resmask,
1811                 .max_write_len = (100U + 6 * NR_CPUS),
1812                 .private = FILE_CPULIST,
1813         },
1814 
1815         {
1816                 .name = "mems",
1817                 .seq_show = cpuset_common_seq_show,
1818                 .write = cpuset_write_resmask,
1819                 .max_write_len = (100U + 6 * MAX_NUMNODES),
1820                 .private = FILE_MEMLIST,
1821         },
1822 
1823         {
1824                 .name = "effective_cpus",
1825                 .seq_show = cpuset_common_seq_show,
1826                 .private = FILE_EFFECTIVE_CPULIST,
1827         },
1828 
1829         {
1830                 .name = "effective_mems",
1831                 .seq_show = cpuset_common_seq_show,
1832                 .private = FILE_EFFECTIVE_MEMLIST,
1833         },
1834 
1835         {
1836                 .name = "cpu_exclusive",
1837                 .read_u64 = cpuset_read_u64,
1838                 .write_u64 = cpuset_write_u64,
1839                 .private = FILE_CPU_EXCLUSIVE,
1840         },
1841 
1842         {
1843                 .name = "mem_exclusive",
1844                 .read_u64 = cpuset_read_u64,
1845                 .write_u64 = cpuset_write_u64,
1846                 .private = FILE_MEM_EXCLUSIVE,
1847         },
1848 
1849         {
1850                 .name = "mem_hardwall",
1851                 .read_u64 = cpuset_read_u64,
1852                 .write_u64 = cpuset_write_u64,
1853                 .private = FILE_MEM_HARDWALL,
1854         },
1855 
1856         {
1857                 .name = "sched_load_balance",
1858                 .read_u64 = cpuset_read_u64,
1859                 .write_u64 = cpuset_write_u64,
1860                 .private = FILE_SCHED_LOAD_BALANCE,
1861         },
1862 
1863         {
1864                 .name = "sched_relax_domain_level",
1865                 .read_s64 = cpuset_read_s64,
1866                 .write_s64 = cpuset_write_s64,
1867                 .private = FILE_SCHED_RELAX_DOMAIN_LEVEL,
1868         },
1869 
1870         {
1871                 .name = "memory_migrate",
1872                 .read_u64 = cpuset_read_u64,
1873                 .write_u64 = cpuset_write_u64,
1874                 .private = FILE_MEMORY_MIGRATE,
1875         },
1876 
1877         {
1878                 .name = "memory_pressure",
1879                 .read_u64 = cpuset_read_u64,
1880                 .write_u64 = cpuset_write_u64,
1881                 .private = FILE_MEMORY_PRESSURE,
1882                 .mode = S_IRUGO,
1883         },
1884 
1885         {
1886                 .name = "memory_spread_page",
1887                 .read_u64 = cpuset_read_u64,
1888                 .write_u64 = cpuset_write_u64,
1889                 .private = FILE_SPREAD_PAGE,
1890         },
1891 
1892         {
1893                 .name = "memory_spread_slab",
1894                 .read_u64 = cpuset_read_u64,
1895                 .write_u64 = cpuset_write_u64,
1896                 .private = FILE_SPREAD_SLAB,
1897         },
1898 
1899         {
1900                 .name = "memory_pressure_enabled",
1901                 .flags = CFTYPE_ONLY_ON_ROOT,
1902                 .read_u64 = cpuset_read_u64,
1903                 .write_u64 = cpuset_write_u64,
1904                 .private = FILE_MEMORY_PRESSURE_ENABLED,
1905         },
1906 
1907         { }     /* terminate */
1908 };
1909 
1910 /*
1911  *      cpuset_css_alloc - allocate a cpuset css
1912  *      cgrp:   control group that the new cpuset will be part of
1913  */
1914 
1915 static struct cgroup_subsys_state *
1916 cpuset_css_alloc(struct cgroup_subsys_state *parent_css)
1917 {
1918         struct cpuset *cs;
1919 
1920         if (!parent_css)
1921                 return &top_cpuset.css;
1922 
1923         cs = kzalloc(sizeof(*cs), GFP_KERNEL);
1924         if (!cs)
1925                 return ERR_PTR(-ENOMEM);
1926         if (!alloc_cpumask_var(&cs->cpus_allowed, GFP_KERNEL))
1927                 goto free_cs;
1928         if (!alloc_cpumask_var(&cs->effective_cpus, GFP_KERNEL))
1929                 goto free_cpus;
1930 
1931         set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
1932         cpumask_clear(cs->cpus_allowed);
1933         nodes_clear(cs->mems_allowed);
1934         cpumask_clear(cs->effective_cpus);
1935         nodes_clear(cs->effective_mems);
1936         fmeter_init(&cs->fmeter);
1937         cs->relax_domain_level = -1;
1938 
1939         return &cs->css;
1940 
1941 free_cpus:
1942         free_cpumask_var(cs->cpus_allowed);
1943 free_cs:
1944         kfree(cs);
1945         return ERR_PTR(-ENOMEM);
1946 }
1947 
1948 static int cpuset_css_online(struct cgroup_subsys_state *css)
1949 {
1950         struct cpuset *cs = css_cs(css);
1951         struct cpuset *parent = parent_cs(cs);
1952         struct cpuset *tmp_cs;
1953         struct cgroup_subsys_state *pos_css;
1954 
1955         if (!parent)
1956                 return 0;
1957 
1958         mutex_lock(&cpuset_mutex);
1959 
1960         set_bit(CS_ONLINE, &cs->flags);
1961         if (is_spread_page(parent))
1962                 set_bit(CS_SPREAD_PAGE, &cs->flags);
1963         if (is_spread_slab(parent))
1964                 set_bit(CS_SPREAD_SLAB, &cs->flags);
1965 
1966         cpuset_inc();
1967 
1968         mutex_lock(&callback_mutex);
1969         if (cgroup_on_dfl(cs->css.cgroup)) {
1970                 cpumask_copy(cs->effective_cpus, parent->effective_cpus);
1971                 cs->effective_mems = parent->effective_mems;
1972         }
1973         mutex_unlock(&callback_mutex);
1974 
1975         if (!test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags))
1976                 goto out_unlock;
1977 
1978         /*
1979          * Clone @parent's configuration if CGRP_CPUSET_CLONE_CHILDREN is
1980          * set.  This flag handling is implemented in cgroup core for
1981          * histrical reasons - the flag may be specified during mount.
1982          *
1983          * Currently, if any sibling cpusets have exclusive cpus or mem, we
1984          * refuse to clone the configuration - thereby refusing the task to
1985          * be entered, and as a result refusing the sys_unshare() or
1986          * clone() which initiated it.  If this becomes a problem for some
1987          * users who wish to allow that scenario, then this could be
1988          * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
1989          * (and likewise for mems) to the new cgroup.
1990          */
1991         rcu_read_lock();
1992         cpuset_for_each_child(tmp_cs, pos_css, parent) {
1993                 if (is_mem_exclusive(tmp_cs) || is_cpu_exclusive(tmp_cs)) {
1994                         rcu_read_unlock();
1995                         goto out_unlock;
1996                 }
1997         }
1998         rcu_read_unlock();
1999 
2000         mutex_lock(&callback_mutex);
2001         cs->mems_allowed = parent->mems_allowed;
2002         cs->effective_mems = parent->mems_allowed;
2003         cpumask_copy(cs->cpus_allowed, parent->cpus_allowed);
2004         cpumask_copy(cs->effective_cpus, parent->cpus_allowed);
2005         mutex_unlock(&callback_mutex);
2006 out_unlock:
2007         mutex_unlock(&cpuset_mutex);
2008         return 0;
2009 }
2010 
2011 /*
2012  * If the cpuset being removed has its flag 'sched_load_balance'
2013  * enabled, then simulate turning sched_load_balance off, which
2014  * will call rebuild_sched_domains_locked().
2015  */
2016 
2017 static void cpuset_css_offline(struct cgroup_subsys_state *css)
2018 {
2019         struct cpuset *cs = css_cs(css);
2020 
2021         mutex_lock(&cpuset_mutex);
2022 
2023         if (is_sched_load_balance(cs))
2024                 update_flag(CS_SCHED_LOAD_BALANCE, cs, 0);
2025 
2026         cpuset_dec();
2027         clear_bit(CS_ONLINE, &cs->flags);
2028 
2029         mutex_unlock(&cpuset_mutex);
2030 }
2031 
2032 static void cpuset_css_free(struct cgroup_subsys_state *css)
2033 {
2034         struct cpuset *cs = css_cs(css);
2035 
2036         free_cpumask_var(cs->effective_cpus);
2037         free_cpumask_var(cs->cpus_allowed);
2038         kfree(cs);
2039 }
2040 
2041 static void cpuset_bind(struct cgroup_subsys_state *root_css)
2042 {
2043         mutex_lock(&cpuset_mutex);
2044         mutex_lock(&callback_mutex);
2045 
2046         if (cgroup_on_dfl(root_css->cgroup)) {
2047                 cpumask_copy(top_cpuset.cpus_allowed, cpu_possible_mask);
2048                 top_cpuset.mems_allowed = node_possible_map;
2049         } else {
2050                 cpumask_copy(top_cpuset.cpus_allowed,
2051                              top_cpuset.effective_cpus);
2052                 top_cpuset.mems_allowed = top_cpuset.effective_mems;
2053         }
2054 
2055         mutex_unlock(&callback_mutex);
2056         mutex_unlock(&cpuset_mutex);
2057 }
2058 
2059 /*
2060  * Make sure the new task conform to the current state of its parent,
2061  * which could have been changed by cpuset just after it inherits the
2062  * state from the parent and before it sits on the cgroup's task list.
2063  */
2064 void cpuset_fork(struct task_struct *task)
2065 {
2066         if (task_css_is_root(task, cpuset_cgrp_id))
2067                 return;
2068 
2069         set_cpus_allowed_ptr(task, &current->cpus_allowed);
2070         task->mems_allowed = current->mems_allowed;
2071 }
2072 
2073 struct cgroup_subsys cpuset_cgrp_subsys = {
2074         .css_alloc      = cpuset_css_alloc,
2075         .css_online     = cpuset_css_online,
2076         .css_offline    = cpuset_css_offline,
2077         .css_free       = cpuset_css_free,
2078         .can_attach     = cpuset_can_attach,
2079         .cancel_attach  = cpuset_cancel_attach,
2080         .attach         = cpuset_attach,
2081         .bind           = cpuset_bind,
2082         .fork           = cpuset_fork,
2083         .legacy_cftypes = files,
2084         .early_init     = 1,
2085 };
2086 
2087 /**
2088  * cpuset_init - initialize cpusets at system boot
2089  *
2090  * Description: Initialize top_cpuset and the cpuset internal file system,
2091  **/
2092 
2093 int __init cpuset_init(void)
2094 {
2095         int err = 0;
2096 
2097         if (!alloc_cpumask_var(&top_cpuset.cpus_allowed, GFP_KERNEL))
2098                 BUG();
2099         if (!alloc_cpumask_var(&top_cpuset.effective_cpus, GFP_KERNEL))
2100                 BUG();
2101 
2102         cpumask_setall(top_cpuset.cpus_allowed);
2103         nodes_setall(top_cpuset.mems_allowed);
2104         cpumask_setall(top_cpuset.effective_cpus);
2105         nodes_setall(top_cpuset.effective_mems);
2106 
2107         fmeter_init(&top_cpuset.fmeter);
2108         set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags);
2109         top_cpuset.relax_domain_level = -1;
2110 
2111         err = register_filesystem(&cpuset_fs_type);
2112         if (err < 0)
2113                 return err;
2114 
2115         if (!alloc_cpumask_var(&cpus_attach, GFP_KERNEL))
2116                 BUG();
2117 
2118         return 0;
2119 }
2120 
2121 /*
2122  * If CPU and/or memory hotplug handlers, below, unplug any CPUs
2123  * or memory nodes, we need to walk over the cpuset hierarchy,
2124  * removing that CPU or node from all cpusets.  If this removes the
2125  * last CPU or node from a cpuset, then move the tasks in the empty
2126  * cpuset to its next-highest non-empty parent.
2127  */
2128 static void remove_tasks_in_empty_cpuset(struct cpuset *cs)
2129 {
2130         struct cpuset *parent;
2131 
2132         /*
2133          * Find its next-highest non-empty parent, (top cpuset
2134          * has online cpus, so can't be empty).
2135          */
2136         parent = parent_cs(cs);
2137         while (cpumask_empty(parent->cpus_allowed) ||
2138                         nodes_empty(parent->mems_allowed))
2139                 parent = parent_cs(parent);
2140 
2141         if (cgroup_transfer_tasks(parent->css.cgroup, cs->css.cgroup)) {
2142                 pr_err("cpuset: failed to transfer tasks out of empty cpuset ");
2143                 pr_cont_cgroup_name(cs->css.cgroup);
2144                 pr_cont("\n");
2145         }
2146 }
2147 
2148 static void
2149 hotplug_update_tasks_legacy(struct cpuset *cs,
2150                             struct cpumask *new_cpus, nodemask_t *new_mems,
2151                             bool cpus_updated, bool mems_updated)
2152 {
2153         bool is_empty;
2154 
2155         mutex_lock(&callback_mutex);
2156         cpumask_copy(cs->cpus_allowed, new_cpus);
2157         cpumask_copy(cs->effective_cpus, new_cpus);
2158         cs->mems_allowed = *new_mems;
2159         cs->effective_mems = *new_mems;
2160         mutex_unlock(&callback_mutex);
2161 
2162         /*
2163          * Don't call update_tasks_cpumask() if the cpuset becomes empty,
2164          * as the tasks will be migratecd to an ancestor.
2165          */
2166         if (cpus_updated && !cpumask_empty(cs->cpus_allowed))
2167                 update_tasks_cpumask(cs);
2168         if (mems_updated && !nodes_empty(cs->mems_allowed))
2169                 update_tasks_nodemask(cs);
2170 
2171         is_empty = cpumask_empty(cs->cpus_allowed) ||
2172                    nodes_empty(cs->mems_allowed);
2173 
2174         mutex_unlock(&cpuset_mutex);
2175 
2176         /*
2177          * Move tasks to the nearest ancestor with execution resources,
2178          * This is full cgroup operation which will also call back into
2179          * cpuset. Should be done outside any lock.
2180          */
2181         if (is_empty)
2182                 remove_tasks_in_empty_cpuset(cs);
2183 
2184         mutex_lock(&cpuset_mutex);
2185 }
2186 
2187 static void
2188 hotplug_update_tasks(struct cpuset *cs,
2189                      struct cpumask *new_cpus, nodemask_t *new_mems,
2190                      bool cpus_updated, bool mems_updated)
2191 {
2192         if (cpumask_empty(new_cpus))
2193                 cpumask_copy(new_cpus, parent_cs(cs)->effective_cpus);
2194         if (nodes_empty(*new_mems))
2195                 *new_mems = parent_cs(cs)->effective_mems;
2196 
2197         mutex_lock(&callback_mutex);
2198         cpumask_copy(cs->effective_cpus, new_cpus);
2199         cs->effective_mems = *new_mems;
2200         mutex_unlock(&callback_mutex);
2201 
2202         if (cpus_updated)
2203                 update_tasks_cpumask(cs);
2204         if (mems_updated)
2205                 update_tasks_nodemask(cs);
2206 }
2207 
2208 /**
2209  * cpuset_hotplug_update_tasks - update tasks in a cpuset for hotunplug
2210  * @cs: cpuset in interest
2211  *
2212  * Compare @cs's cpu and mem masks against top_cpuset and if some have gone
2213  * offline, update @cs accordingly.  If @cs ends up with no CPU or memory,
2214  * all its tasks are moved to the nearest ancestor with both resources.
2215  */
2216 static void cpuset_hotplug_update_tasks(struct cpuset *cs)
2217 {
2218         static cpumask_t new_cpus;
2219         static nodemask_t new_mems;
2220         bool cpus_updated;
2221         bool mems_updated;
2222 retry:
2223         wait_event(cpuset_attach_wq, cs->attach_in_progress == 0);
2224 
2225         mutex_lock(&cpuset_mutex);
2226 
2227         /*
2228          * We have raced with task attaching. We wait until attaching
2229          * is finished, so we won't attach a task to an empty cpuset.
2230          */
2231         if (cs->attach_in_progress) {
2232                 mutex_unlock(&cpuset_mutex);
2233                 goto retry;
2234         }
2235 
2236         cpumask_and(&new_cpus, cs->cpus_allowed, parent_cs(cs)->effective_cpus);
2237         nodes_and(new_mems, cs->mems_allowed, parent_cs(cs)->effective_mems);
2238 
2239         cpus_updated = !cpumask_equal(&new_cpus, cs->effective_cpus);
2240         mems_updated = !nodes_equal(new_mems, cs->effective_mems);
2241 
2242         if (cgroup_on_dfl(cs->css.cgroup))
2243                 hotplug_update_tasks(cs, &new_cpus, &new_mems,
2244                                      cpus_updated, mems_updated);
2245         else
2246                 hotplug_update_tasks_legacy(cs, &new_cpus, &new_mems,
2247                                             cpus_updated, mems_updated);
2248 
2249         mutex_unlock(&cpuset_mutex);
2250 }
2251 
2252 /**
2253  * cpuset_hotplug_workfn - handle CPU/memory hotunplug for a cpuset
2254  *
2255  * This function is called after either CPU or memory configuration has
2256  * changed and updates cpuset accordingly.  The top_cpuset is always
2257  * synchronized to cpu_active_mask and N_MEMORY, which is necessary in
2258  * order to make cpusets transparent (of no affect) on systems that are
2259  * actively using CPU hotplug but making no active use of cpusets.
2260  *
2261  * Non-root cpusets are only affected by offlining.  If any CPUs or memory
2262  * nodes have been taken down, cpuset_hotplug_update_tasks() is invoked on
2263  * all descendants.
2264  *
2265  * Note that CPU offlining during suspend is ignored.  We don't modify
2266  * cpusets across suspend/resume cycles at all.
2267  */
2268 static void cpuset_hotplug_workfn(struct work_struct *work)
2269 {
2270         static cpumask_t new_cpus;
2271         static nodemask_t new_mems;
2272         bool cpus_updated, mems_updated;
2273         bool on_dfl = cgroup_on_dfl(top_cpuset.css.cgroup);
2274 
2275         mutex_lock(&cpuset_mutex);
2276 
2277         /* fetch the available cpus/mems and find out which changed how */
2278         cpumask_copy(&new_cpus, cpu_active_mask);
2279         new_mems = node_states[N_MEMORY];
2280 
2281         cpus_updated = !cpumask_equal(top_cpuset.effective_cpus, &new_cpus);
2282         mems_updated = !nodes_equal(top_cpuset.effective_mems, new_mems);
2283 
2284         /* synchronize cpus_allowed to cpu_active_mask */
2285         if (cpus_updated) {
2286                 mutex_lock(&callback_mutex);
2287                 if (!on_dfl)
2288                         cpumask_copy(top_cpuset.cpus_allowed, &new_cpus);
2289                 cpumask_copy(top_cpuset.effective_cpus, &new_cpus);
2290                 mutex_unlock(&callback_mutex);
2291                 /* we don't mess with cpumasks of tasks in top_cpuset */
2292         }
2293 
2294         /* synchronize mems_allowed to N_MEMORY */
2295         if (mems_updated) {
2296                 mutex_lock(&callback_mutex);
2297                 if (!on_dfl)
2298                         top_cpuset.mems_allowed = new_mems;
2299                 top_cpuset.effective_mems = new_mems;
2300                 mutex_unlock(&callback_mutex);
2301                 update_tasks_nodemask(&top_cpuset);
2302         }
2303 
2304         mutex_unlock(&cpuset_mutex);
2305 
2306         /* if cpus or mems changed, we need to propagate to descendants */
2307         if (cpus_updated || mems_updated) {
2308                 struct cpuset *cs;
2309                 struct cgroup_subsys_state *pos_css;
2310 
2311                 rcu_read_lock();
2312                 cpuset_for_each_descendant_pre(cs, pos_css, &top_cpuset) {
2313                         if (cs == &top_cpuset || !css_tryget_online(&cs->css))
2314                                 continue;
2315                         rcu_read_unlock();
2316 
2317                         cpuset_hotplug_update_tasks(cs);
2318 
2319                         rcu_read_lock();
2320                         css_put(&cs->css);
2321                 }
2322                 rcu_read_unlock();
2323         }
2324 
2325         /* rebuild sched domains if cpus_allowed has changed */
2326         if (cpus_updated)
2327                 rebuild_sched_domains();
2328 }
2329 
2330 void cpuset_update_active_cpus(bool cpu_online)
2331 {
2332         /*
2333          * We're inside cpu hotplug critical region which usually nests
2334          * inside cgroup synchronization.  Bounce actual hotplug processing
2335          * to a work item to avoid reverse locking order.
2336          *
2337          * We still need to do partition_sched_domains() synchronously;
2338          * otherwise, the scheduler will get confused and put tasks to the
2339          * dead CPU.  Fall back to the default single domain.
2340          * cpuset_hotplug_workfn() will rebuild it as necessary.
2341          */
2342         partition_sched_domains(1, NULL, NULL);
2343         schedule_work(&cpuset_hotplug_work);
2344 }
2345 
2346 /*
2347  * Keep top_cpuset.mems_allowed tracking node_states[N_MEMORY].
2348  * Call this routine anytime after node_states[N_MEMORY] changes.
2349  * See cpuset_update_active_cpus() for CPU hotplug handling.
2350  */
2351 static int cpuset_track_online_nodes(struct notifier_block *self,
2352                                 unsigned long action, void *arg)
2353 {
2354         schedule_work(&cpuset_hotplug_work);
2355         return NOTIFY_OK;
2356 }
2357 
2358 static struct notifier_block cpuset_track_online_nodes_nb = {
2359         .notifier_call = cpuset_track_online_nodes,
2360         .priority = 10,         /* ??! */
2361 };
2362 
2363 /**
2364  * cpuset_init_smp - initialize cpus_allowed
2365  *
2366  * Description: Finish top cpuset after cpu, node maps are initialized
2367  */
2368 void __init cpuset_init_smp(void)
2369 {
2370         cpumask_copy(top_cpuset.cpus_allowed, cpu_active_mask);
2371         top_cpuset.mems_allowed = node_states[N_MEMORY];
2372         top_cpuset.old_mems_allowed = top_cpuset.mems_allowed;
2373 
2374         cpumask_copy(top_cpuset.effective_cpus, cpu_active_mask);
2375         top_cpuset.effective_mems = node_states[N_MEMORY];
2376 
2377         register_hotmemory_notifier(&cpuset_track_online_nodes_nb);
2378 }
2379 
2380 /**
2381  * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
2382  * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
2383  * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
2384  *
2385  * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
2386  * attached to the specified @tsk.  Guaranteed to return some non-empty
2387  * subset of cpu_online_mask, even if this means going outside the
2388  * tasks cpuset.
2389  **/
2390 
2391 void cpuset_cpus_allowed(struct task_struct *tsk, struct cpumask *pmask)
2392 {
2393         mutex_lock(&callback_mutex);
2394         rcu_read_lock();
2395         guarantee_online_cpus(task_cs(tsk), pmask);
2396         rcu_read_unlock();
2397         mutex_unlock(&callback_mutex);
2398 }
2399 
2400 void cpuset_cpus_allowed_fallback(struct task_struct *tsk)
2401 {
2402         rcu_read_lock();
2403         do_set_cpus_allowed(tsk, task_cs(tsk)->effective_cpus);
2404         rcu_read_unlock();
2405 
2406         /*
2407          * We own tsk->cpus_allowed, nobody can change it under us.
2408          *
2409          * But we used cs && cs->cpus_allowed lockless and thus can
2410          * race with cgroup_attach_task() or update_cpumask() and get
2411          * the wrong tsk->cpus_allowed. However, both cases imply the
2412          * subsequent cpuset_change_cpumask()->set_cpus_allowed_ptr()
2413          * which takes task_rq_lock().
2414          *
2415          * If we are called after it dropped the lock we must see all
2416          * changes in tsk_cs()->cpus_allowed. Otherwise we can temporary
2417          * set any mask even if it is not right from task_cs() pov,
2418          * the pending set_cpus_allowed_ptr() will fix things.
2419          *
2420          * select_fallback_rq() will fix things ups and set cpu_possible_mask
2421          * if required.
2422          */
2423 }
2424 
2425 void cpuset_init_current_mems_allowed(void)
2426 {
2427         nodes_setall(current->mems_allowed);
2428 }
2429 
2430 /**
2431  * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
2432  * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
2433  *
2434  * Description: Returns the nodemask_t mems_allowed of the cpuset
2435  * attached to the specified @tsk.  Guaranteed to return some non-empty
2436  * subset of node_states[N_MEMORY], even if this means going outside the
2437  * tasks cpuset.
2438  **/
2439 
2440 nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
2441 {
2442         nodemask_t mask;
2443 
2444         mutex_lock(&callback_mutex);
2445         rcu_read_lock();
2446         guarantee_online_mems(task_cs(tsk), &mask);
2447         rcu_read_unlock();
2448         mutex_unlock(&callback_mutex);
2449 
2450         return mask;
2451 }
2452 
2453 /**
2454  * cpuset_nodemask_valid_mems_allowed - check nodemask vs. curremt mems_allowed
2455  * @nodemask: the nodemask to be checked
2456  *
2457  * Are any of the nodes in the nodemask allowed in current->mems_allowed?
2458  */
2459 int cpuset_nodemask_valid_mems_allowed(nodemask_t *nodemask)
2460 {
2461         return nodes_intersects(*nodemask, current->mems_allowed);
2462 }
2463 
2464 /*
2465  * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
2466  * mem_hardwall ancestor to the specified cpuset.  Call holding
2467  * callback_mutex.  If no ancestor is mem_exclusive or mem_hardwall
2468  * (an unusual configuration), then returns the root cpuset.
2469  */
2470 static struct cpuset *nearest_hardwall_ancestor(struct cpuset *cs)
2471 {
2472         while (!(is_mem_exclusive(cs) || is_mem_hardwall(cs)) && parent_cs(cs))
2473                 cs = parent_cs(cs);
2474         return cs;
2475 }
2476 
2477 /**
2478  * cpuset_node_allowed_softwall - Can we allocate on a memory node?
2479  * @node: is this an allowed node?
2480  * @gfp_mask: memory allocation flags
2481  *
2482  * If we're in interrupt, yes, we can always allocate.  If __GFP_THISNODE is
2483  * set, yes, we can always allocate.  If node is in our task's mems_allowed,
2484  * yes.  If it's not a __GFP_HARDWALL request and this node is in the nearest
2485  * hardwalled cpuset ancestor to this task's cpuset, yes.  If the task has been
2486  * OOM killed and has access to memory reserves as specified by the TIF_MEMDIE
2487  * flag, yes.
2488  * Otherwise, no.
2489  *
2490  * If __GFP_HARDWALL is set, cpuset_node_allowed_softwall() reduces to
2491  * cpuset_node_allowed_hardwall().  Otherwise, cpuset_node_allowed_softwall()
2492  * might sleep, and might allow a node from an enclosing cpuset.
2493  *
2494  * cpuset_node_allowed_hardwall() only handles the simpler case of hardwall
2495  * cpusets, and never sleeps.
2496  *
2497  * The __GFP_THISNODE placement logic is really handled elsewhere,
2498  * by forcibly using a zonelist starting at a specified node, and by
2499  * (in get_page_from_freelist()) refusing to consider the zones for
2500  * any node on the zonelist except the first.  By the time any such
2501  * calls get to this routine, we should just shut up and say 'yes'.
2502  *
2503  * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
2504  * and do not allow allocations outside the current tasks cpuset
2505  * unless the task has been OOM killed as is marked TIF_MEMDIE.
2506  * GFP_KERNEL allocations are not so marked, so can escape to the
2507  * nearest enclosing hardwalled ancestor cpuset.
2508  *
2509  * Scanning up parent cpusets requires callback_mutex.  The
2510  * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
2511  * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
2512  * current tasks mems_allowed came up empty on the first pass over
2513  * the zonelist.  So only GFP_KERNEL allocations, if all nodes in the
2514  * cpuset are short of memory, might require taking the callback_mutex
2515  * mutex.
2516  *
2517  * The first call here from mm/page_alloc:get_page_from_freelist()
2518  * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
2519  * so no allocation on a node outside the cpuset is allowed (unless
2520  * in interrupt, of course).
2521  *
2522  * The second pass through get_page_from_freelist() doesn't even call
2523  * here for GFP_ATOMIC calls.  For those calls, the __alloc_pages()
2524  * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
2525  * in alloc_flags.  That logic and the checks below have the combined
2526  * affect that:
2527  *      in_interrupt - any node ok (current task context irrelevant)
2528  *      GFP_ATOMIC   - any node ok
2529  *      TIF_MEMDIE   - any node ok
2530  *      GFP_KERNEL   - any node in enclosing hardwalled cpuset ok
2531  *      GFP_USER     - only nodes in current tasks mems allowed ok.
2532  *
2533  * Rule:
2534  *    Don't call cpuset_node_allowed_softwall if you can't sleep, unless you
2535  *    pass in the __GFP_HARDWALL flag set in gfp_flag, which disables
2536  *    the code that might scan up ancestor cpusets and sleep.
2537  */
2538 int __cpuset_node_allowed_softwall(int node, gfp_t gfp_mask)
2539 {
2540         struct cpuset *cs;              /* current cpuset ancestors */
2541         int allowed;                    /* is allocation in zone z allowed? */
2542 
2543         if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
2544                 return 1;
2545         might_sleep_if(!(gfp_mask & __GFP_HARDWALL));
2546         if (node_isset(node, current->mems_allowed))
2547                 return 1;
2548         /*
2549          * Allow tasks that have access to memory reserves because they have
2550          * been OOM killed to get memory anywhere.
2551          */
2552         if (unlikely(test_thread_flag(TIF_MEMDIE)))
2553                 return 1;
2554         if (gfp_mask & __GFP_HARDWALL)  /* If hardwall request, stop here */
2555                 return 0;
2556 
2557         if (current->flags & PF_EXITING) /* Let dying task have memory */
2558                 return 1;
2559 
2560         /* Not hardwall and node outside mems_allowed: scan up cpusets */
2561         mutex_lock(&callback_mutex);
2562 
2563         rcu_read_lock();
2564         cs = nearest_hardwall_ancestor(task_cs(current));
2565         allowed = node_isset(node, cs->mems_allowed);
2566         rcu_read_unlock();
2567 
2568         mutex_unlock(&callback_mutex);
2569         return allowed;
2570 }
2571 
2572 /*
2573  * cpuset_node_allowed_hardwall - Can we allocate on a memory node?
2574  * @node: is this an allowed node?
2575  * @gfp_mask: memory allocation flags
2576  *
2577  * If we're in interrupt, yes, we can always allocate.  If __GFP_THISNODE is
2578  * set, yes, we can always allocate.  If node is in our task's mems_allowed,
2579  * yes.  If the task has been OOM killed and has access to memory reserves as
2580  * specified by the TIF_MEMDIE flag, yes.
2581  * Otherwise, no.
2582  *
2583  * The __GFP_THISNODE placement logic is really handled elsewhere,
2584  * by forcibly using a zonelist starting at a specified node, and by
2585  * (in get_page_from_freelist()) refusing to consider the zones for
2586  * any node on the zonelist except the first.  By the time any such
2587  * calls get to this routine, we should just shut up and say 'yes'.
2588  *
2589  * Unlike the cpuset_node_allowed_softwall() variant, above,
2590  * this variant requires that the node be in the current task's
2591  * mems_allowed or that we're in interrupt.  It does not scan up the
2592  * cpuset hierarchy for the nearest enclosing mem_exclusive cpuset.
2593  * It never sleeps.
2594  */
2595 int __cpuset_node_allowed_hardwall(int node, gfp_t gfp_mask)
2596 {
2597         if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
2598                 return 1;
2599         if (node_isset(node, current->mems_allowed))
2600                 return 1;
2601         /*
2602          * Allow tasks that have access to memory reserves because they have
2603          * been OOM killed to get memory anywhere.
2604          */
2605         if (unlikely(test_thread_flag(TIF_MEMDIE)))
2606                 return 1;
2607         return 0;
2608 }
2609 
2610 /**
2611  * cpuset_mem_spread_node() - On which node to begin search for a file page
2612  * cpuset_slab_spread_node() - On which node to begin search for a slab page
2613  *
2614  * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
2615  * tasks in a cpuset with is_spread_page or is_spread_slab set),
2616  * and if the memory allocation used cpuset_mem_spread_node()
2617  * to determine on which node to start looking, as it will for
2618  * certain page cache or slab cache pages such as used for file
2619  * system buffers and inode caches, then instead of starting on the
2620  * local node to look for a free page, rather spread the starting
2621  * node around the tasks mems_allowed nodes.
2622  *
2623  * We don't have to worry about the returned node being offline
2624  * because "it can't happen", and even if it did, it would be ok.
2625  *
2626  * The routines calling guarantee_online_mems() are careful to
2627  * only set nodes in task->mems_allowed that are online.  So it
2628  * should not be possible for the following code to return an
2629  * offline node.  But if it did, that would be ok, as this routine
2630  * is not returning the node where the allocation must be, only
2631  * the node where the search should start.  The zonelist passed to
2632  * __alloc_pages() will include all nodes.  If the slab allocator
2633  * is passed an offline node, it will fall back to the local node.
2634  * See kmem_cache_alloc_node().
2635  */
2636 
2637 static int cpuset_spread_node(int *rotor)
2638 {
2639         int node;
2640 
2641         node = next_node(*rotor, current->mems_allowed);
2642         if (node == MAX_NUMNODES)
2643                 node = first_node(current->mems_allowed);
2644         *rotor = node;
2645         return node;
2646 }
2647 
2648 int cpuset_mem_spread_node(void)
2649 {
2650         if (current->cpuset_mem_spread_rotor == NUMA_NO_NODE)
2651                 current->cpuset_mem_spread_rotor =
2652                         node_random(&current->mems_allowed);
2653 
2654         return cpuset_spread_node(&current->cpuset_mem_spread_rotor);
2655 }
2656 
2657 int cpuset_slab_spread_node(void)
2658 {
2659         if (current->cpuset_slab_spread_rotor == NUMA_NO_NODE)
2660                 current->cpuset_slab_spread_rotor =
2661                         node_random(&current->mems_allowed);
2662 
2663         return cpuset_spread_node(&current->cpuset_slab_spread_rotor);
2664 }
2665 
2666 EXPORT_SYMBOL_GPL(cpuset_mem_spread_node);
2667 
2668 /**
2669  * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
2670  * @tsk1: pointer to task_struct of some task.
2671  * @tsk2: pointer to task_struct of some other task.
2672  *
2673  * Description: Return true if @tsk1's mems_allowed intersects the
2674  * mems_allowed of @tsk2.  Used by the OOM killer to determine if
2675  * one of the task's memory usage might impact the memory available
2676  * to the other.
2677  **/
2678 
2679 int cpuset_mems_allowed_intersects(const struct task_struct *tsk1,
2680                                    const struct task_struct *tsk2)
2681 {
2682         return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed);
2683 }
2684 
2685 #define CPUSET_NODELIST_LEN     (256)
2686 
2687 /**
2688  * cpuset_print_task_mems_allowed - prints task's cpuset and mems_allowed
2689  * @tsk: pointer to task_struct of some task.
2690  *
2691  * Description: Prints @task's name, cpuset name, and cached copy of its
2692  * mems_allowed to the kernel log.
2693  */
2694 void cpuset_print_task_mems_allowed(struct task_struct *tsk)
2695 {
2696          /* Statically allocated to prevent using excess stack. */
2697         static char cpuset_nodelist[CPUSET_NODELIST_LEN];
2698         static DEFINE_SPINLOCK(cpuset_buffer_lock);
2699         struct cgroup *cgrp;
2700 
2701         spin_lock(&cpuset_buffer_lock);
2702         rcu_read_lock();
2703 
2704         cgrp = task_cs(tsk)->css.cgroup;
2705         nodelist_scnprintf(cpuset_nodelist, CPUSET_NODELIST_LEN,
2706                            tsk->mems_allowed);
2707         pr_info("%s cpuset=", tsk->comm);
2708         pr_cont_cgroup_name(cgrp);
2709         pr_cont(" mems_allowed=%s\n", cpuset_nodelist);
2710 
2711         rcu_read_unlock();
2712         spin_unlock(&cpuset_buffer_lock);
2713 }
2714 
2715 /*
2716  * Collection of memory_pressure is suppressed unless
2717  * this flag is enabled by writing "1" to the special
2718  * cpuset file 'memory_pressure_enabled' in the root cpuset.
2719  */
2720 
2721 int cpuset_memory_pressure_enabled __read_mostly;
2722 
2723 /**
2724  * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
2725  *
2726  * Keep a running average of the rate of synchronous (direct)
2727  * page reclaim efforts initiated by tasks in each cpuset.
2728  *
2729  * This represents the rate at which some task in the cpuset
2730  * ran low on memory on all nodes it was allowed to use, and
2731  * had to enter the kernels page reclaim code in an effort to
2732  * create more free memory by tossing clean pages or swapping
2733  * or writing dirty pages.
2734  *
2735  * Display to user space in the per-cpuset read-only file
2736  * "memory_pressure".  Value displayed is an integer
2737  * representing the recent rate of entry into the synchronous
2738  * (direct) page reclaim by any task attached to the cpuset.
2739  **/
2740 
2741 void __cpuset_memory_pressure_bump(void)
2742 {
2743         rcu_read_lock();
2744         fmeter_markevent(&task_cs(current)->fmeter);
2745         rcu_read_unlock();
2746 }
2747 
2748 #ifdef CONFIG_PROC_PID_CPUSET
2749 /*
2750  * proc_cpuset_show()
2751  *  - Print tasks cpuset path into seq_file.
2752  *  - Used for /proc/<pid>/cpuset.
2753  *  - No need to task_lock(tsk) on this tsk->cpuset reference, as it
2754  *    doesn't really matter if tsk->cpuset changes after we read it,
2755  *    and we take cpuset_mutex, keeping cpuset_attach() from changing it
2756  *    anyway.
2757  */
2758 int proc_cpuset_show(struct seq_file *m, struct pid_namespace *ns,
2759                      struct pid *pid, struct task_struct *tsk)
2760 {
2761         char *buf, *p;
2762         struct cgroup_subsys_state *css;
2763         int retval;
2764 
2765         retval = -ENOMEM;
2766         buf = kmalloc(PATH_MAX, GFP_KERNEL);
2767         if (!buf)
2768                 goto out;
2769 
2770         retval = -ENAMETOOLONG;
2771         rcu_read_lock();
2772         css = task_css(tsk, cpuset_cgrp_id);
2773         p = cgroup_path(css->cgroup, buf, PATH_MAX);
2774         rcu_read_unlock();
2775         if (!p)
2776                 goto out_free;
2777         seq_puts(m, p);
2778         seq_putc(m, '\n');
2779         retval = 0;
2780 out_free:
2781         kfree(buf);
2782 out:
2783         return retval;
2784 }
2785 #endif /* CONFIG_PROC_PID_CPUSET */
2786 
2787 /* Display task mems_allowed in /proc/<pid>/status file. */
2788 void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task)
2789 {
2790         seq_puts(m, "Mems_allowed:\t");
2791         seq_nodemask(m, &task->mems_allowed);
2792         seq_puts(m, "\n");
2793         seq_puts(m, "Mems_allowed_list:\t");
2794         seq_nodemask_list(m, &task->mems_allowed);
2795         seq_puts(m, "\n");
2796 }
2797 

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

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

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

osdn.jp