Version:  2.0.40 2.2.26 2.4.37 3.1 3.2 3.3 3.4 3.5 3.6 3.7 3.8 3.9 3.10 3.11 3.12 3.13 3.14 3.15 3.16 3.17

Linux/kernel/cpuset.c

  1 /*
  2  *  kernel/cpuset.c
  3  *
  4  *  Processor and Memory placement constraints for sets of tasks.
  5  *
  6  *  Copyright (C) 2003 BULL SA.
  7  *  Copyright (C) 2004-2007 Silicon Graphics, Inc.
  8  *  Copyright (C) 2006 Google, Inc
  9  *
 10  *  Portions derived from Patrick Mochel's sysfs code.
 11  *  sysfs is Copyright (c) 2001-3 Patrick Mochel
 12  *
 13  *  2003-10-10 Written by Simon Derr.
 14  *  2003-10-22 Updates by Stephen Hemminger.
 15  *  2004 May-July Rework by Paul Jackson.
 16  *  2006 Rework by Paul Menage to use generic cgroups
 17  *  2008 Rework of the scheduler domains and CPU hotplug handling
 18  *       by Max Krasnyansky
 19  *
 20  *  This file is subject to the terms and conditions of the GNU General Public
 21  *  License.  See the file COPYING in the main directory of the Linux
 22  *  distribution for more details.
 23  */
 24 
 25 #include <linux/cpu.h>
 26 #include <linux/cpumask.h>
 27 #include <linux/cpuset.h>
 28 #include <linux/err.h>
 29 #include <linux/errno.h>
 30 #include <linux/file.h>
 31 #include <linux/fs.h>
 32 #include <linux/init.h>
 33 #include <linux/interrupt.h>
 34 #include <linux/kernel.h>
 35 #include <linux/kmod.h>
 36 #include <linux/list.h>
 37 #include <linux/mempolicy.h>
 38 #include <linux/mm.h>
 39 #include <linux/memory.h>
 40 #include <linux/export.h>
 41 #include <linux/mount.h>
 42 #include <linux/namei.h>
 43 #include <linux/pagemap.h>
 44 #include <linux/proc_fs.h>
 45 #include <linux/rcupdate.h>
 46 #include <linux/sched.h>
 47 #include <linux/seq_file.h>
 48 #include <linux/security.h>
 49 #include <linux/slab.h>
 50 #include <linux/spinlock.h>
 51 #include <linux/stat.h>
 52 #include <linux/string.h>
 53 #include <linux/time.h>
 54 #include <linux/backing-dev.h>
 55 #include <linux/sort.h>
 56 
 57 #include <asm/uaccess.h>
 58 #include <linux/atomic.h>
 59 #include <linux/mutex.h>
 60 #include <linux/workqueue.h>
 61 #include <linux/cgroup.h>
 62 #include <linux/wait.h>
 63 
 64 struct static_key cpusets_enabled_key __read_mostly = STATIC_KEY_INIT_FALSE;
 65 
 66 /* See "Frequency meter" comments, below. */
 67 
 68 struct fmeter {
 69         int cnt;                /* unprocessed events count */
 70         int val;                /* most recent output value */
 71         time_t time;            /* clock (secs) when val computed */
 72         spinlock_t lock;        /* guards read or write of above */
 73 };
 74 
 75 struct cpuset {
 76         struct cgroup_subsys_state css;
 77 
 78         unsigned long flags;            /* "unsigned long" so bitops work */
 79 
 80         /*
 81          * On default hierarchy:
 82          *
 83          * The user-configured masks can only be changed by writing to
 84          * cpuset.cpus and cpuset.mems, and won't be limited by the
 85          * parent masks.
 86          *
 87          * The effective masks is the real masks that apply to the tasks
 88          * in the cpuset. They may be changed if the configured masks are
 89          * changed or hotplug happens.
 90          *
 91          * effective_mask == configured_mask & parent's effective_mask,
 92          * and if it ends up empty, it will inherit the parent's mask.
 93          *
 94          *
 95          * On legacy hierachy:
 96          *
 97          * The user-configured masks are always the same with effective masks.
 98          */
 99 
100         /* user-configured CPUs and Memory Nodes allow to tasks */
101         cpumask_var_t cpus_allowed;
102         nodemask_t mems_allowed;
103 
104         /* effective CPUs and Memory Nodes allow to tasks */
105         cpumask_var_t effective_cpus;
106         nodemask_t effective_mems;
107 
108         /*
109          * This is old Memory Nodes tasks took on.
110          *
111          * - top_cpuset.old_mems_allowed is initialized to mems_allowed.
112          * - A new cpuset's old_mems_allowed is initialized when some
113          *   task is moved into it.
114          * - old_mems_allowed is used in cpuset_migrate_mm() when we change
115          *   cpuset.mems_allowed and have tasks' nodemask updated, and
116          *   then old_mems_allowed is updated to mems_allowed.
117          */
118         nodemask_t old_mems_allowed;
119 
120         struct fmeter fmeter;           /* memory_pressure filter */
121 
122         /*
123          * Tasks are being attached to this cpuset.  Used to prevent
124          * zeroing cpus/mems_allowed between ->can_attach() and ->attach().
125          */
126         int attach_in_progress;
127 
128         /* partition number for rebuild_sched_domains() */
129         int pn;
130 
131         /* for custom sched domain */
132         int relax_domain_level;
133 };
134 
135 static inline struct cpuset *css_cs(struct cgroup_subsys_state *css)
136 {
137         return css ? container_of(css, struct cpuset, css) : NULL;
138 }
139 
140 /* Retrieve the cpuset for a task */
141 static inline struct cpuset *task_cs(struct task_struct *task)
142 {
143         return css_cs(task_css(task, cpuset_cgrp_id));
144 }
145 
146 static inline struct cpuset *parent_cs(struct cpuset *cs)
147 {
148         return css_cs(cs->css.parent);
149 }
150 
151 #ifdef CONFIG_NUMA
152 static inline bool task_has_mempolicy(struct task_struct *task)
153 {
154         return task->mempolicy;
155 }
156 #else
157 static inline bool task_has_mempolicy(struct task_struct *task)
158 {
159         return false;
160 }
161 #endif
162 
163 
164 /* bits in struct cpuset flags field */
165 typedef enum {
166         CS_ONLINE,
167         CS_CPU_EXCLUSIVE,
168         CS_MEM_EXCLUSIVE,
169         CS_MEM_HARDWALL,
170         CS_MEMORY_MIGRATE,
171         CS_SCHED_LOAD_BALANCE,
172         CS_SPREAD_PAGE,
173         CS_SPREAD_SLAB,
174 } cpuset_flagbits_t;
175 
176 /* convenient tests for these bits */
177 static inline bool is_cpuset_online(const struct cpuset *cs)
178 {
179         return test_bit(CS_ONLINE, &cs->flags);
180 }
181 
182 static inline int is_cpu_exclusive(const struct cpuset *cs)
183 {
184         return test_bit(CS_CPU_EXCLUSIVE, &cs->flags);
185 }
186 
187 static inline int is_mem_exclusive(const struct cpuset *cs)
188 {
189         return test_bit(CS_MEM_EXCLUSIVE, &cs->flags);
190 }
191 
192 static inline int is_mem_hardwall(const struct cpuset *cs)
193 {
194         return test_bit(CS_MEM_HARDWALL, &cs->flags);
195 }
196 
197 static inline int is_sched_load_balance(const struct cpuset *cs)
198 {
199         return test_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
200 }
201 
202 static inline int is_memory_migrate(const struct cpuset *cs)
203 {
204         return test_bit(CS_MEMORY_MIGRATE, &cs->flags);
205 }
206 
207 static inline int is_spread_page(const struct cpuset *cs)
208 {
209         return test_bit(CS_SPREAD_PAGE, &cs->flags);
210 }
211 
212 static inline int is_spread_slab(const struct cpuset *cs)
213 {
214         return test_bit(CS_SPREAD_SLAB, &cs->flags);
215 }
216 
217 static struct cpuset top_cpuset = {
218         .flags = ((1 << CS_ONLINE) | (1 << CS_CPU_EXCLUSIVE) |
219                   (1 << CS_MEM_EXCLUSIVE)),
220 };
221 
222 /**
223  * cpuset_for_each_child - traverse online children of a cpuset
224  * @child_cs: loop cursor pointing to the current child
225  * @pos_css: used for iteration
226  * @parent_cs: target cpuset to walk children of
227  *
228  * Walk @child_cs through the online children of @parent_cs.  Must be used
229  * with RCU read locked.
230  */
231 #define cpuset_for_each_child(child_cs, pos_css, parent_cs)             \
232         css_for_each_child((pos_css), &(parent_cs)->css)                \
233                 if (is_cpuset_online(((child_cs) = css_cs((pos_css)))))
234 
235 /**
236  * cpuset_for_each_descendant_pre - pre-order walk of a cpuset's descendants
237  * @des_cs: loop cursor pointing to the current descendant
238  * @pos_css: used for iteration
239  * @root_cs: target cpuset to walk ancestor of
240  *
241  * Walk @des_cs through the online descendants of @root_cs.  Must be used
242  * with RCU read locked.  The caller may modify @pos_css by calling
243  * css_rightmost_descendant() to skip subtree.  @root_cs is included in the
244  * iteration and the first node to be visited.
245  */
246 #define cpuset_for_each_descendant_pre(des_cs, pos_css, root_cs)        \
247         css_for_each_descendant_pre((pos_css), &(root_cs)->css)         \
248                 if (is_cpuset_online(((des_cs) = css_cs((pos_css)))))
249 
250 /*
251  * There are two global 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.  The top
327  * cpuset always has some cpus online.
328  *
329  * One way or another, we guarantee to return some non-empty subset
330  * of cpu_online_mask.
331  *
332  * Call with callback_mutex held.
333  */
334 static void guarantee_online_cpus(struct cpuset *cs, struct cpumask *pmask)
335 {
336         while (!cpumask_intersects(cs->effective_cpus, cpu_online_mask))
337                 cs = parent_cs(cs);
338         cpumask_and(pmask, cs->effective_cpus, cpu_online_mask);
339 }
340 
341 /*
342  * Return in *pmask the portion of a cpusets's mems_allowed that
343  * are online, with memory.  If none are online with memory, walk
344  * up the cpuset hierarchy until we find one that does have some
345  * online mems.  The top cpuset always has some mems online.
346  *
347  * One way or another, we guarantee to return some non-empty subset
348  * of node_states[N_MEMORY].
349  *
350  * Call with callback_mutex held.
351  */
352 static void guarantee_online_mems(struct cpuset *cs, nodemask_t *pmask)
353 {
354         while (!nodes_intersects(cs->effective_mems, node_states[N_MEMORY]))
355                 cs = parent_cs(cs);
356         nodes_and(*pmask, cs->effective_mems, node_states[N_MEMORY]);
357 }
358 
359 /*
360  * update task's spread flag if cpuset's page/slab spread flag is set
361  *
362  * Called with callback_mutex/cpuset_mutex held
363  */
364 static void cpuset_update_task_spread_flag(struct cpuset *cs,
365                                         struct task_struct *tsk)
366 {
367         if (is_spread_page(cs))
368                 task_set_spread_page(tsk);
369         else
370                 task_clear_spread_page(tsk);
371 
372         if (is_spread_slab(cs))
373                 task_set_spread_slab(tsk);
374         else
375                 task_clear_spread_slab(tsk);
376 }
377 
378 /*
379  * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
380  *
381  * One cpuset is a subset of another if all its allowed CPUs and
382  * Memory Nodes are a subset of the other, and its exclusive flags
383  * are only set if the other's are set.  Call holding cpuset_mutex.
384  */
385 
386 static int is_cpuset_subset(const struct cpuset *p, const struct cpuset *q)
387 {
388         return  cpumask_subset(p->cpus_allowed, q->cpus_allowed) &&
389                 nodes_subset(p->mems_allowed, q->mems_allowed) &&
390                 is_cpu_exclusive(p) <= is_cpu_exclusive(q) &&
391                 is_mem_exclusive(p) <= is_mem_exclusive(q);
392 }
393 
394 /**
395  * alloc_trial_cpuset - allocate a trial cpuset
396  * @cs: the cpuset that the trial cpuset duplicates
397  */
398 static struct cpuset *alloc_trial_cpuset(struct cpuset *cs)
399 {
400         struct cpuset *trial;
401 
402         trial = kmemdup(cs, sizeof(*cs), GFP_KERNEL);
403         if (!trial)
404                 return NULL;
405 
406         if (!alloc_cpumask_var(&trial->cpus_allowed, GFP_KERNEL))
407                 goto free_cs;
408         if (!alloc_cpumask_var(&trial->effective_cpus, GFP_KERNEL))
409                 goto free_cpus;
410 
411         cpumask_copy(trial->cpus_allowed, cs->cpus_allowed);
412         cpumask_copy(trial->effective_cpus, cs->effective_cpus);
413         return trial;
414 
415 free_cpus:
416         free_cpumask_var(trial->cpus_allowed);
417 free_cs:
418         kfree(trial);
419         return NULL;
420 }
421 
422 /**
423  * free_trial_cpuset - free the trial cpuset
424  * @trial: the trial cpuset to be freed
425  */
426 static void free_trial_cpuset(struct cpuset *trial)
427 {
428         free_cpumask_var(trial->effective_cpus);
429         free_cpumask_var(trial->cpus_allowed);
430         kfree(trial);
431 }
432 
433 /*
434  * validate_change() - Used to validate that any proposed cpuset change
435  *                     follows the structural rules for cpusets.
436  *
437  * If we replaced the flag and mask values of the current cpuset
438  * (cur) with those values in the trial cpuset (trial), would
439  * our various subset and exclusive rules still be valid?  Presumes
440  * cpuset_mutex held.
441  *
442  * 'cur' is the address of an actual, in-use cpuset.  Operations
443  * such as list traversal that depend on the actual address of the
444  * cpuset in the list must use cur below, not trial.
445  *
446  * 'trial' is the address of bulk structure copy of cur, with
447  * perhaps one or more of the fields cpus_allowed, mems_allowed,
448  * or flags changed to new, trial values.
449  *
450  * Return 0 if valid, -errno if not.
451  */
452 
453 static int validate_change(struct cpuset *cur, struct cpuset *trial)
454 {
455         struct cgroup_subsys_state *css;
456         struct cpuset *c, *par;
457         int ret;
458 
459         rcu_read_lock();
460 
461         /* Each of our child cpusets must be a subset of us */
462         ret = -EBUSY;
463         cpuset_for_each_child(c, css, cur)
464                 if (!is_cpuset_subset(c, trial))
465                         goto out;
466 
467         /* Remaining checks don't apply to root cpuset */
468         ret = 0;
469         if (cur == &top_cpuset)
470                 goto out;
471 
472         par = parent_cs(cur);
473 
474         /* On legacy hiearchy, we must be a subset of our parent cpuset. */
475         ret = -EACCES;
476         if (!cgroup_on_dfl(cur->css.cgroup) && !is_cpuset_subset(trial, par))
477                 goto out;
478 
479         /*
480          * If either I or some sibling (!= me) is exclusive, we can't
481          * overlap
482          */
483         ret = -EINVAL;
484         cpuset_for_each_child(c, css, par) {
485                 if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
486                     c != cur &&
487                     cpumask_intersects(trial->cpus_allowed, c->cpus_allowed))
488                         goto out;
489                 if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) &&
490                     c != cur &&
491                     nodes_intersects(trial->mems_allowed, c->mems_allowed))
492                         goto out;
493         }
494 
495         /*
496          * Cpusets with tasks - existing or newly being attached - can't
497          * be changed to have empty cpus_allowed or mems_allowed.
498          */
499         ret = -ENOSPC;
500         if ((cgroup_has_tasks(cur->css.cgroup) || cur->attach_in_progress)) {
501                 if (!cpumask_empty(cur->cpus_allowed) &&
502                     cpumask_empty(trial->cpus_allowed))
503                         goto out;
504                 if (!nodes_empty(cur->mems_allowed) &&
505                     nodes_empty(trial->mems_allowed))
506                         goto out;
507         }
508 
509         ret = 0;
510 out:
511         rcu_read_unlock();
512         return ret;
513 }
514 
515 #ifdef CONFIG_SMP
516 /*
517  * Helper routine for generate_sched_domains().
518  * Do cpusets a, b have overlapping effective cpus_allowed masks?
519  */
520 static int cpusets_overlap(struct cpuset *a, struct cpuset *b)
521 {
522         return cpumask_intersects(a->effective_cpus, b->effective_cpus);
523 }
524 
525 static void
526 update_domain_attr(struct sched_domain_attr *dattr, struct cpuset *c)
527 {
528         if (dattr->relax_domain_level < c->relax_domain_level)
529                 dattr->relax_domain_level = c->relax_domain_level;
530         return;
531 }
532 
533 static void update_domain_attr_tree(struct sched_domain_attr *dattr,
534                                     struct cpuset *root_cs)
535 {
536         struct cpuset *cp;
537         struct cgroup_subsys_state *pos_css;
538 
539         rcu_read_lock();
540         cpuset_for_each_descendant_pre(cp, pos_css, root_cs) {
541                 if (cp == root_cs)
542                         continue;
543 
544                 /* skip the whole subtree if @cp doesn't have any CPU */
545                 if (cpumask_empty(cp->cpus_allowed)) {
546                         pos_css = css_rightmost_descendant(pos_css);
547                         continue;
548                 }
549 
550                 if (is_sched_load_balance(cp))
551                         update_domain_attr(dattr, cp);
552         }
553         rcu_read_unlock();
554 }
555 
556 /*
557  * generate_sched_domains()
558  *
559  * This function builds a partial partition of the systems CPUs
560  * A 'partial partition' is a set of non-overlapping subsets whose
561  * union is a subset of that set.
562  * The output of this function needs to be passed to kernel/sched/core.c
563  * partition_sched_domains() routine, which will rebuild the scheduler's
564  * load balancing domains (sched domains) as specified by that partial
565  * partition.
566  *
567  * See "What is sched_load_balance" in Documentation/cgroups/cpusets.txt
568  * for a background explanation of this.
569  *
570  * Does not return errors, on the theory that the callers of this
571  * routine would rather not worry about failures to rebuild sched
572  * domains when operating in the severe memory shortage situations
573  * that could cause allocation failures below.
574  *
575  * Must be called with cpuset_mutex held.
576  *
577  * The three key local variables below are:
578  *    q  - a linked-list queue of cpuset pointers, used to implement a
579  *         top-down scan of all cpusets.  This scan loads a pointer
580  *         to each cpuset marked is_sched_load_balance into the
581  *         array 'csa'.  For our purposes, rebuilding the schedulers
582  *         sched domains, we can ignore !is_sched_load_balance cpusets.
583  *  csa  - (for CpuSet Array) Array of pointers to all the cpusets
584  *         that need to be load balanced, for convenient iterative
585  *         access by the subsequent code that finds the best partition,
586  *         i.e the set of domains (subsets) of CPUs such that the
587  *         cpus_allowed of every cpuset marked is_sched_load_balance
588  *         is a subset of one of these domains, while there are as
589  *         many such domains as possible, each as small as possible.
590  * doms  - Conversion of 'csa' to an array of cpumasks, for passing to
591  *         the kernel/sched/core.c routine partition_sched_domains() in a
592  *         convenient format, that can be easily compared to the prior
593  *         value to determine what partition elements (sched domains)
594  *         were changed (added or removed.)
595  *
596  * Finding the best partition (set of domains):
597  *      The triple nested loops below over i, j, k scan over the
598  *      load balanced cpusets (using the array of cpuset pointers in
599  *      csa[]) looking for pairs of cpusets that have overlapping
600  *      cpus_allowed, but which don't have the same 'pn' partition
601  *      number and gives them in the same partition number.  It keeps
602  *      looping on the 'restart' label until it can no longer find
603  *      any such pairs.
604  *
605  *      The union of the cpus_allowed masks from the set of
606  *      all cpusets having the same 'pn' value then form the one
607  *      element of the partition (one sched domain) to be passed to
608  *      partition_sched_domains().
609  */
610 static int generate_sched_domains(cpumask_var_t **domains,
611                         struct sched_domain_attr **attributes)
612 {
613         struct cpuset *cp;      /* scans q */
614         struct cpuset **csa;    /* array of all cpuset ptrs */
615         int csn;                /* how many cpuset ptrs in csa so far */
616         int i, j, k;            /* indices for partition finding loops */
617         cpumask_var_t *doms;    /* resulting partition; i.e. sched domains */
618         struct sched_domain_attr *dattr;  /* attributes for custom domains */
619         int ndoms = 0;          /* number of sched domains in result */
620         int nslot;              /* next empty doms[] struct cpumask slot */
621         struct cgroup_subsys_state *pos_css;
622 
623         doms = NULL;
624         dattr = NULL;
625         csa = NULL;
626 
627         /* Special case for the 99% of systems with one, full, sched domain */
628         if (is_sched_load_balance(&top_cpuset)) {
629                 ndoms = 1;
630                 doms = alloc_sched_domains(ndoms);
631                 if (!doms)
632                         goto done;
633 
634                 dattr = kmalloc(sizeof(struct sched_domain_attr), GFP_KERNEL);
635                 if (dattr) {
636                         *dattr = SD_ATTR_INIT;
637                         update_domain_attr_tree(dattr, &top_cpuset);
638                 }
639                 cpumask_copy(doms[0], top_cpuset.effective_cpus);
640 
641                 goto done;
642         }
643 
644         csa = kmalloc(nr_cpusets() * sizeof(cp), GFP_KERNEL);
645         if (!csa)
646                 goto done;
647         csn = 0;
648 
649         rcu_read_lock();
650         cpuset_for_each_descendant_pre(cp, pos_css, &top_cpuset) {
651                 if (cp == &top_cpuset)
652                         continue;
653                 /*
654                  * Continue traversing beyond @cp iff @cp has some CPUs and
655                  * isn't load balancing.  The former is obvious.  The
656                  * latter: All child cpusets contain a subset of the
657                  * parent's cpus, so just skip them, and then we call
658                  * update_domain_attr_tree() to calc relax_domain_level of
659                  * the corresponding sched domain.
660                  */
661                 if (!cpumask_empty(cp->cpus_allowed) &&
662                     !is_sched_load_balance(cp))
663                         continue;
664 
665                 if (is_sched_load_balance(cp))
666                         csa[csn++] = cp;
667 
668                 /* skip @cp's subtree */
669                 pos_css = css_rightmost_descendant(pos_css);
670         }
671         rcu_read_unlock();
672 
673         for (i = 0; i < csn; i++)
674                 csa[i]->pn = i;
675         ndoms = csn;
676 
677 restart:
678         /* Find the best partition (set of sched domains) */
679         for (i = 0; i < csn; i++) {
680                 struct cpuset *a = csa[i];
681                 int apn = a->pn;
682 
683                 for (j = 0; j < csn; j++) {
684                         struct cpuset *b = csa[j];
685                         int bpn = b->pn;
686 
687                         if (apn != bpn && cpusets_overlap(a, b)) {
688                                 for (k = 0; k < csn; k++) {
689                                         struct cpuset *c = csa[k];
690 
691                                         if (c->pn == bpn)
692                                                 c->pn = apn;
693                                 }
694                                 ndoms--;        /* one less element */
695                                 goto restart;
696                         }
697                 }
698         }
699 
700         /*
701          * Now we know how many domains to create.
702          * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
703          */
704         doms = alloc_sched_domains(ndoms);
705         if (!doms)
706                 goto done;
707 
708         /*
709          * The rest of the code, including the scheduler, can deal with
710          * dattr==NULL case. No need to abort if alloc fails.
711          */
712         dattr = kmalloc(ndoms * sizeof(struct sched_domain_attr), GFP_KERNEL);
713 
714         for (nslot = 0, i = 0; i < csn; i++) {
715                 struct cpuset *a = csa[i];
716                 struct cpumask *dp;
717                 int apn = a->pn;
718 
719                 if (apn < 0) {
720                         /* Skip completed partitions */
721                         continue;
722                 }
723 
724                 dp = doms[nslot];
725 
726                 if (nslot == ndoms) {
727                         static int warnings = 10;
728                         if (warnings) {
729                                 pr_warn("rebuild_sched_domains confused: nslot %d, ndoms %d, csn %d, i %d, apn %d\n",
730                                         nslot, ndoms, csn, i, apn);
731                                 warnings--;
732                         }
733                         continue;
734                 }
735 
736                 cpumask_clear(dp);
737                 if (dattr)
738                         *(dattr + nslot) = SD_ATTR_INIT;
739                 for (j = i; j < csn; j++) {
740                         struct cpuset *b = csa[j];
741 
742                         if (apn == b->pn) {
743                                 cpumask_or(dp, dp, b->effective_cpus);
744                                 if (dattr)
745                                         update_domain_attr_tree(dattr + nslot, b);
746 
747                                 /* Done with this partition */
748                                 b->pn = -1;
749                         }
750                 }
751                 nslot++;
752         }
753         BUG_ON(nslot != ndoms);
754 
755 done:
756         kfree(csa);
757 
758         /*
759          * Fallback to the default domain if kmalloc() failed.
760          * See comments in partition_sched_domains().
761          */
762         if (doms == NULL)
763                 ndoms = 1;
764 
765         *domains    = doms;
766         *attributes = dattr;
767         return ndoms;
768 }
769 
770 /*
771  * Rebuild scheduler domains.
772  *
773  * If the flag 'sched_load_balance' of any cpuset with non-empty
774  * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
775  * which has that flag enabled, or if any cpuset with a non-empty
776  * 'cpus' is removed, then call this routine to rebuild the
777  * scheduler's dynamic sched domains.
778  *
779  * Call with cpuset_mutex held.  Takes get_online_cpus().
780  */
781 static void rebuild_sched_domains_locked(void)
782 {
783         struct sched_domain_attr *attr;
784         cpumask_var_t *doms;
785         int ndoms;
786 
787         lockdep_assert_held(&cpuset_mutex);
788         get_online_cpus();
789 
790         /*
791          * We have raced with CPU hotplug. Don't do anything to avoid
792          * passing doms with offlined cpu to partition_sched_domains().
793          * Anyways, hotplug work item will rebuild sched domains.
794          */
795         if (!cpumask_equal(top_cpuset.effective_cpus, cpu_active_mask))
796                 goto out;
797 
798         /* Generate domain masks and attrs */
799         ndoms = generate_sched_domains(&doms, &attr);
800 
801         /* Have scheduler rebuild the domains */
802         partition_sched_domains(ndoms, doms, attr);
803 out:
804         put_online_cpus();
805 }
806 #else /* !CONFIG_SMP */
807 static void rebuild_sched_domains_locked(void)
808 {
809 }
810 #endif /* CONFIG_SMP */
811 
812 void rebuild_sched_domains(void)
813 {
814         mutex_lock(&cpuset_mutex);
815         rebuild_sched_domains_locked();
816         mutex_unlock(&cpuset_mutex);
817 }
818 
819 /**
820  * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
821  * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
822  *
823  * Iterate through each task of @cs updating its cpus_allowed to the
824  * effective cpuset's.  As this function is called with cpuset_mutex held,
825  * cpuset membership stays stable.
826  */
827 static void update_tasks_cpumask(struct cpuset *cs)
828 {
829         struct css_task_iter it;
830         struct task_struct *task;
831 
832         css_task_iter_start(&cs->css, &it);
833         while ((task = css_task_iter_next(&it)))
834                 set_cpus_allowed_ptr(task, cs->effective_cpus);
835         css_task_iter_end(&it);
836 }
837 
838 /*
839  * update_cpumasks_hier - Update effective cpumasks and tasks in the subtree
840  * @cs: the cpuset to consider
841  * @new_cpus: temp variable for calculating new effective_cpus
842  *
843  * When congifured cpumask is changed, the effective cpumasks of this cpuset
844  * and all its descendants need to be updated.
845  *
846  * On legacy hierachy, effective_cpus will be the same with cpu_allowed.
847  *
848  * Called with cpuset_mutex held
849  */
850 static void update_cpumasks_hier(struct cpuset *cs, struct cpumask *new_cpus)
851 {
852         struct cpuset *cp;
853         struct cgroup_subsys_state *pos_css;
854         bool need_rebuild_sched_domains = false;
855 
856         rcu_read_lock();
857         cpuset_for_each_descendant_pre(cp, pos_css, cs) {
858                 struct cpuset *parent = parent_cs(cp);
859 
860                 cpumask_and(new_cpus, cp->cpus_allowed, parent->effective_cpus);
861 
862                 /*
863                  * If it becomes empty, inherit the effective mask of the
864                  * parent, which is guaranteed to have some CPUs.
865                  */
866                 if (cpumask_empty(new_cpus))
867                         cpumask_copy(new_cpus, parent->effective_cpus);
868 
869                 /* Skip the whole subtree if the cpumask remains the same. */
870                 if (cpumask_equal(new_cpus, cp->effective_cpus)) {
871                         pos_css = css_rightmost_descendant(pos_css);
872                         continue;
873                 }
874 
875                 if (!css_tryget_online(&cp->css))
876                         continue;
877                 rcu_read_unlock();
878 
879                 mutex_lock(&callback_mutex);
880                 cpumask_copy(cp->effective_cpus, new_cpus);
881                 mutex_unlock(&callback_mutex);
882 
883                 WARN_ON(!cgroup_on_dfl(cp->css.cgroup) &&
884                         !cpumask_equal(cp->cpus_allowed, cp->effective_cpus));
885 
886                 update_tasks_cpumask(cp);
887 
888                 /*
889                  * If the effective cpumask of any non-empty cpuset is changed,
890                  * we need to rebuild sched domains.
891                  */
892                 if (!cpumask_empty(cp->cpus_allowed) &&
893                     is_sched_load_balance(cp))
894                         need_rebuild_sched_domains = true;
895 
896                 rcu_read_lock();
897                 css_put(&cp->css);
898         }
899         rcu_read_unlock();
900 
901         if (need_rebuild_sched_domains)
902                 rebuild_sched_domains_locked();
903 }
904 
905 /**
906  * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
907  * @cs: the cpuset to consider
908  * @trialcs: trial cpuset
909  * @buf: buffer of cpu numbers written to this cpuset
910  */
911 static int update_cpumask(struct cpuset *cs, struct cpuset *trialcs,
912                           const char *buf)
913 {
914         int retval;
915 
916         /* top_cpuset.cpus_allowed tracks cpu_online_mask; it's read-only */
917         if (cs == &top_cpuset)
918                 return -EACCES;
919 
920         /*
921          * An empty cpus_allowed is ok only if the cpuset has no tasks.
922          * Since cpulist_parse() fails on an empty mask, we special case
923          * that parsing.  The validate_change() call ensures that cpusets
924          * with tasks have cpus.
925          */
926         if (!*buf) {
927                 cpumask_clear(trialcs->cpus_allowed);
928         } else {
929                 retval = cpulist_parse(buf, trialcs->cpus_allowed);
930                 if (retval < 0)
931                         return retval;
932 
933                 if (!cpumask_subset(trialcs->cpus_allowed,
934                                     top_cpuset.cpus_allowed))
935                         return -EINVAL;
936         }
937 
938         /* Nothing to do if the cpus didn't change */
939         if (cpumask_equal(cs->cpus_allowed, trialcs->cpus_allowed))
940                 return 0;
941 
942         retval = validate_change(cs, trialcs);
943         if (retval < 0)
944                 return retval;
945 
946         mutex_lock(&callback_mutex);
947         cpumask_copy(cs->cpus_allowed, trialcs->cpus_allowed);
948         mutex_unlock(&callback_mutex);
949 
950         /* use trialcs->cpus_allowed as a temp variable */
951         update_cpumasks_hier(cs, trialcs->cpus_allowed);
952         return 0;
953 }
954 
955 /*
956  * cpuset_migrate_mm
957  *
958  *    Migrate memory region from one set of nodes to another.
959  *
960  *    Temporarilly set tasks mems_allowed to target nodes of migration,
961  *    so that the migration code can allocate pages on these nodes.
962  *
963  *    While the mm_struct we are migrating is typically from some
964  *    other task, the task_struct mems_allowed that we are hacking
965  *    is for our current task, which must allocate new pages for that
966  *    migrating memory region.
967  */
968 
969 static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
970                                                         const nodemask_t *to)
971 {
972         struct task_struct *tsk = current;
973 
974         tsk->mems_allowed = *to;
975 
976         do_migrate_pages(mm, from, to, MPOL_MF_MOVE_ALL);
977 
978         rcu_read_lock();
979         guarantee_online_mems(task_cs(tsk), &tsk->mems_allowed);
980         rcu_read_unlock();
981 }
982 
983 /*
984  * cpuset_change_task_nodemask - change task's mems_allowed and mempolicy
985  * @tsk: the task to change
986  * @newmems: new nodes that the task will be set
987  *
988  * In order to avoid seeing no nodes if the old and new nodes are disjoint,
989  * we structure updates as setting all new allowed nodes, then clearing newly
990  * disallowed ones.
991  */
992 static void cpuset_change_task_nodemask(struct task_struct *tsk,
993                                         nodemask_t *newmems)
994 {
995         bool need_loop;
996 
997         /*
998          * Allow tasks that have access to memory reserves because they have
999          * been OOM killed to get memory anywhere.
1000          */
1001         if (unlikely(test_thread_flag(TIF_MEMDIE)))
1002                 return;
1003         if (current->flags & PF_EXITING) /* Let dying task have memory */
1004                 return;
1005 
1006         task_lock(tsk);
1007         /*
1008          * Determine if a loop is necessary if another thread is doing
1009          * read_mems_allowed_begin().  If at least one node remains unchanged and
1010          * tsk does not have a mempolicy, then an empty nodemask will not be
1011          * possible when mems_allowed is larger than a word.
1012          */
1013         need_loop = task_has_mempolicy(tsk) ||
1014                         !nodes_intersects(*newmems, tsk->mems_allowed);
1015 
1016         if (need_loop) {
1017                 local_irq_disable();
1018                 write_seqcount_begin(&tsk->mems_allowed_seq);
1019         }
1020 
1021         nodes_or(tsk->mems_allowed, tsk->mems_allowed, *newmems);
1022         mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP1);
1023 
1024         mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP2);
1025         tsk->mems_allowed = *newmems;
1026 
1027         if (need_loop) {
1028                 write_seqcount_end(&tsk->mems_allowed_seq);
1029                 local_irq_enable();
1030         }
1031 
1032         task_unlock(tsk);
1033 }
1034 
1035 static void *cpuset_being_rebound;
1036 
1037 /**
1038  * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
1039  * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
1040  *
1041  * Iterate through each task of @cs updating its mems_allowed to the
1042  * effective cpuset's.  As this function is called with cpuset_mutex held,
1043  * cpuset membership stays stable.
1044  */
1045 static void update_tasks_nodemask(struct cpuset *cs)
1046 {
1047         static nodemask_t newmems;      /* protected by cpuset_mutex */
1048         struct css_task_iter it;
1049         struct task_struct *task;
1050 
1051         cpuset_being_rebound = cs;              /* causes mpol_dup() rebind */
1052 
1053         guarantee_online_mems(cs, &newmems);
1054 
1055         /*
1056          * The mpol_rebind_mm() call takes mmap_sem, which we couldn't
1057          * take while holding tasklist_lock.  Forks can happen - the
1058          * mpol_dup() cpuset_being_rebound check will catch such forks,
1059          * and rebind their vma mempolicies too.  Because we still hold
1060          * the global cpuset_mutex, we know that no other rebind effort
1061          * will be contending for the global variable cpuset_being_rebound.
1062          * It's ok if we rebind the same mm twice; mpol_rebind_mm()
1063          * is idempotent.  Also migrate pages in each mm to new nodes.
1064          */
1065         css_task_iter_start(&cs->css, &it);
1066         while ((task = css_task_iter_next(&it))) {
1067                 struct mm_struct *mm;
1068                 bool migrate;
1069 
1070                 cpuset_change_task_nodemask(task, &newmems);
1071 
1072                 mm = get_task_mm(task);
1073                 if (!mm)
1074                         continue;
1075 
1076                 migrate = is_memory_migrate(cs);
1077 
1078                 mpol_rebind_mm(mm, &cs->mems_allowed);
1079                 if (migrate)
1080                         cpuset_migrate_mm(mm, &cs->old_mems_allowed, &newmems);
1081                 mmput(mm);
1082         }
1083         css_task_iter_end(&it);
1084 
1085         /*
1086          * All the tasks' nodemasks have been updated, update
1087          * cs->old_mems_allowed.
1088          */
1089         cs->old_mems_allowed = newmems;
1090 
1091         /* We're done rebinding vmas to this cpuset's new mems_allowed. */
1092         cpuset_being_rebound = NULL;
1093 }
1094 
1095 /*
1096  * update_nodemasks_hier - Update effective nodemasks and tasks in the subtree
1097  * @cs: the cpuset to consider
1098  * @new_mems: a temp variable for calculating new effective_mems
1099  *
1100  * When configured nodemask is changed, the effective nodemasks of this cpuset
1101  * and all its descendants need to be updated.
1102  *
1103  * On legacy hiearchy, effective_mems will be the same with mems_allowed.
1104  *
1105  * Called with cpuset_mutex held
1106  */
1107 static void update_nodemasks_hier(struct cpuset *cs, nodemask_t *new_mems)
1108 {
1109         struct cpuset *cp;
1110         struct cgroup_subsys_state *pos_css;
1111 
1112         rcu_read_lock();
1113         cpuset_for_each_descendant_pre(cp, pos_css, cs) {
1114                 struct cpuset *parent = parent_cs(cp);
1115 
1116                 nodes_and(*new_mems, cp->mems_allowed, parent->effective_mems);
1117 
1118                 /*
1119                  * If it becomes empty, inherit the effective mask of the
1120                  * parent, which is guaranteed to have some MEMs.
1121                  */
1122                 if (nodes_empty(*new_mems))
1123                         *new_mems = parent->effective_mems;
1124 
1125                 /* Skip the whole subtree if the nodemask remains the same. */
1126                 if (nodes_equal(*new_mems, cp->effective_mems)) {
1127                         pos_css = css_rightmost_descendant(pos_css);
1128                         continue;
1129                 }
1130 
1131                 if (!css_tryget_online(&cp->css))
1132                         continue;
1133                 rcu_read_unlock();
1134 
1135                 mutex_lock(&callback_mutex);
1136                 cp->effective_mems = *new_mems;
1137                 mutex_unlock(&callback_mutex);
1138 
1139                 WARN_ON(!cgroup_on_dfl(cp->css.cgroup) &&
1140                         !nodes_equal(cp->mems_allowed, cp->effective_mems));
1141 
1142                 update_tasks_nodemask(cp);
1143 
1144                 rcu_read_lock();
1145                 css_put(&cp->css);
1146         }
1147         rcu_read_unlock();
1148 }
1149 
1150 /*
1151  * Handle user request to change the 'mems' memory placement
1152  * of a cpuset.  Needs to validate the request, update the
1153  * cpusets mems_allowed, and for each task in the cpuset,
1154  * update mems_allowed and rebind task's mempolicy and any vma
1155  * mempolicies and if the cpuset is marked 'memory_migrate',
1156  * migrate the tasks pages to the new memory.
1157  *
1158  * Call with cpuset_mutex held.  May take callback_mutex during call.
1159  * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
1160  * lock each such tasks mm->mmap_sem, scan its vma's and rebind
1161  * their mempolicies to the cpusets new mems_allowed.
1162  */
1163 static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
1164                            const char *buf)
1165 {
1166         int retval;
1167 
1168         /*
1169          * top_cpuset.mems_allowed tracks node_stats[N_MEMORY];
1170          * it's read-only
1171          */
1172         if (cs == &top_cpuset) {
1173                 retval = -EACCES;
1174                 goto done;
1175         }
1176 
1177         /*
1178          * An empty mems_allowed is ok iff there are no tasks in the cpuset.
1179          * Since nodelist_parse() fails on an empty mask, we special case
1180          * that parsing.  The validate_change() call ensures that cpusets
1181          * with tasks have memory.
1182          */
1183         if (!*buf) {
1184                 nodes_clear(trialcs->mems_allowed);
1185         } else {
1186                 retval = nodelist_parse(buf, trialcs->mems_allowed);
1187                 if (retval < 0)
1188                         goto done;
1189 
1190                 if (!nodes_subset(trialcs->mems_allowed,
1191                                   top_cpuset.mems_allowed)) {
1192                         retval = -EINVAL;
1193                         goto done;
1194                 }
1195         }
1196 
1197         if (nodes_equal(cs->mems_allowed, trialcs->mems_allowed)) {
1198                 retval = 0;             /* Too easy - nothing to do */
1199                 goto done;
1200         }
1201         retval = validate_change(cs, trialcs);
1202         if (retval < 0)
1203                 goto done;
1204 
1205         mutex_lock(&callback_mutex);
1206         cs->mems_allowed = trialcs->mems_allowed;
1207         mutex_unlock(&callback_mutex);
1208 
1209         /* use trialcs->mems_allowed as a temp variable */
1210         update_nodemasks_hier(cs, &cs->mems_allowed);
1211 done:
1212         return retval;
1213 }
1214 
1215 int current_cpuset_is_being_rebound(void)
1216 {
1217         int ret;
1218 
1219         rcu_read_lock();
1220         ret = task_cs(current) == cpuset_being_rebound;
1221         rcu_read_unlock();
1222 
1223         return ret;
1224 }
1225 
1226 static int update_relax_domain_level(struct cpuset *cs, s64 val)
1227 {
1228 #ifdef CONFIG_SMP
1229         if (val < -1 || val >= sched_domain_level_max)
1230                 return -EINVAL;
1231 #endif
1232 
1233         if (val != cs->relax_domain_level) {
1234                 cs->relax_domain_level = val;
1235                 if (!cpumask_empty(cs->cpus_allowed) &&
1236                     is_sched_load_balance(cs))
1237                         rebuild_sched_domains_locked();
1238         }
1239 
1240         return 0;
1241 }
1242 
1243 /**
1244  * update_tasks_flags - update the spread flags of tasks in the cpuset.
1245  * @cs: the cpuset in which each task's spread flags needs to be changed
1246  *
1247  * Iterate through each task of @cs updating its spread flags.  As this
1248  * function is called with cpuset_mutex held, cpuset membership stays
1249  * stable.
1250  */
1251 static void update_tasks_flags(struct cpuset *cs)
1252 {
1253         struct css_task_iter it;
1254         struct task_struct *task;
1255 
1256         css_task_iter_start(&cs->css, &it);
1257         while ((task = css_task_iter_next(&it)))
1258                 cpuset_update_task_spread_flag(cs, task);
1259         css_task_iter_end(&it);
1260 }
1261 
1262 /*
1263  * update_flag - read a 0 or a 1 in a file and update associated flag
1264  * bit:         the bit to update (see cpuset_flagbits_t)
1265  * cs:          the cpuset to update
1266  * turning_on:  whether the flag is being set or cleared
1267  *
1268  * Call with cpuset_mutex held.
1269  */
1270 
1271 static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs,
1272                        int turning_on)
1273 {
1274         struct cpuset *trialcs;
1275         int balance_flag_changed;
1276         int spread_flag_changed;
1277         int err;
1278 
1279         trialcs = alloc_trial_cpuset(cs);
1280         if (!trialcs)
1281                 return -ENOMEM;
1282 
1283         if (turning_on)
1284                 set_bit(bit, &trialcs->flags);
1285         else
1286                 clear_bit(bit, &trialcs->flags);
1287 
1288         err = validate_change(cs, trialcs);
1289         if (err < 0)
1290                 goto out;
1291 
1292         balance_flag_changed = (is_sched_load_balance(cs) !=
1293                                 is_sched_load_balance(trialcs));
1294 
1295         spread_flag_changed = ((is_spread_slab(cs) != is_spread_slab(trialcs))
1296                         || (is_spread_page(cs) != is_spread_page(trialcs)));
1297 
1298         mutex_lock(&callback_mutex);
1299         cs->flags = trialcs->flags;
1300         mutex_unlock(&callback_mutex);
1301 
1302         if (!cpumask_empty(trialcs->cpus_allowed) && balance_flag_changed)
1303                 rebuild_sched_domains_locked();
1304 
1305         if (spread_flag_changed)
1306                 update_tasks_flags(cs);
1307 out:
1308         free_trial_cpuset(trialcs);
1309         return err;
1310 }
1311 
1312 /*
1313  * Frequency meter - How fast is some event occurring?
1314  *
1315  * These routines manage a digitally filtered, constant time based,
1316  * event frequency meter.  There are four routines:
1317  *   fmeter_init() - initialize a frequency meter.
1318  *   fmeter_markevent() - called each time the event happens.
1319  *   fmeter_getrate() - returns the recent rate of such events.
1320  *   fmeter_update() - internal routine used to update fmeter.
1321  *
1322  * A common data structure is passed to each of these routines,
1323  * which is used to keep track of the state required to manage the
1324  * frequency meter and its digital filter.
1325  *
1326  * The filter works on the number of events marked per unit time.
1327  * The filter is single-pole low-pass recursive (IIR).  The time unit
1328  * is 1 second.  Arithmetic is done using 32-bit integers scaled to
1329  * simulate 3 decimal digits of precision (multiplied by 1000).
1330  *
1331  * With an FM_COEF of 933, and a time base of 1 second, the filter
1332  * has a half-life of 10 seconds, meaning that if the events quit
1333  * happening, then the rate returned from the fmeter_getrate()
1334  * will be cut in half each 10 seconds, until it converges to zero.
1335  *
1336  * It is not worth doing a real infinitely recursive filter.  If more
1337  * than FM_MAXTICKS ticks have elapsed since the last filter event,
1338  * just compute FM_MAXTICKS ticks worth, by which point the level
1339  * will be stable.
1340  *
1341  * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
1342  * arithmetic overflow in the fmeter_update() routine.
1343  *
1344  * Given the simple 32 bit integer arithmetic used, this meter works
1345  * best for reporting rates between one per millisecond (msec) and
1346  * one per 32 (approx) seconds.  At constant rates faster than one
1347  * per msec it maxes out at values just under 1,000,000.  At constant
1348  * rates between one per msec, and one per second it will stabilize
1349  * to a value N*1000, where N is the rate of events per second.
1350  * At constant rates between one per second and one per 32 seconds,
1351  * it will be choppy, moving up on the seconds that have an event,
1352  * and then decaying until the next event.  At rates slower than
1353  * about one in 32 seconds, it decays all the way back to zero between
1354  * each event.
1355  */
1356 
1357 #define FM_COEF 933             /* coefficient for half-life of 10 secs */
1358 #define FM_MAXTICKS ((time_t)99) /* useless computing more ticks than this */
1359 #define FM_MAXCNT 1000000       /* limit cnt to avoid overflow */
1360 #define FM_SCALE 1000           /* faux fixed point scale */
1361 
1362 /* Initialize a frequency meter */
1363 static void fmeter_init(struct fmeter *fmp)
1364 {
1365         fmp->cnt = 0;
1366         fmp->val = 0;
1367         fmp->time = 0;
1368         spin_lock_init(&fmp->lock);
1369 }
1370 
1371 /* Internal meter update - process cnt events and update value */
1372 static void fmeter_update(struct fmeter *fmp)
1373 {
1374         time_t now = get_seconds();
1375         time_t ticks = now - fmp->time;
1376 
1377         if (ticks == 0)
1378                 return;
1379 
1380         ticks = min(FM_MAXTICKS, ticks);
1381         while (ticks-- > 0)
1382                 fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
1383         fmp->time = now;
1384 
1385         fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
1386         fmp->cnt = 0;
1387 }
1388 
1389 /* Process any previous ticks, then bump cnt by one (times scale). */
1390 static void fmeter_markevent(struct fmeter *fmp)
1391 {
1392         spin_lock(&fmp->lock);
1393         fmeter_update(fmp);
1394         fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
1395         spin_unlock(&fmp->lock);
1396 }
1397 
1398 /* Process any previous ticks, then return current value. */
1399 static int fmeter_getrate(struct fmeter *fmp)
1400 {
1401         int val;
1402 
1403         spin_lock(&fmp->lock);
1404         fmeter_update(fmp);
1405         val = fmp->val;
1406         spin_unlock(&fmp->lock);
1407         return val;
1408 }
1409 
1410 static struct cpuset *cpuset_attach_old_cs;
1411 
1412 /* Called by cgroups to determine if a cpuset is usable; cpuset_mutex held */
1413 static int cpuset_can_attach(struct cgroup_subsys_state *css,
1414                              struct cgroup_taskset *tset)
1415 {
1416         struct cpuset *cs = css_cs(css);
1417         struct task_struct *task;
1418         int ret;
1419 
1420         /* used later by cpuset_attach() */
1421         cpuset_attach_old_cs = task_cs(cgroup_taskset_first(tset));
1422 
1423         mutex_lock(&cpuset_mutex);
1424 
1425         /* allow moving tasks into an empty cpuset if on default hierarchy */
1426         ret = -ENOSPC;
1427         if (!cgroup_on_dfl(css->cgroup) &&
1428             (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed)))
1429                 goto out_unlock;
1430 
1431         cgroup_taskset_for_each(task, tset) {
1432                 /*
1433                  * Kthreads which disallow setaffinity shouldn't be moved
1434                  * to a new cpuset; we don't want to change their cpu
1435                  * affinity and isolating such threads by their set of
1436                  * allowed nodes is unnecessary.  Thus, cpusets are not
1437                  * applicable for such threads.  This prevents checking for
1438                  * success of set_cpus_allowed_ptr() on all attached tasks
1439                  * before cpus_allowed may be changed.
1440                  */
1441                 ret = -EINVAL;
1442                 if (task->flags & PF_NO_SETAFFINITY)
1443                         goto out_unlock;
1444                 ret = security_task_setscheduler(task);
1445                 if (ret)
1446                         goto out_unlock;
1447         }
1448 
1449         /*
1450          * Mark attach is in progress.  This makes validate_change() fail
1451          * changes which zero cpus/mems_allowed.
1452          */
1453         cs->attach_in_progress++;
1454         ret = 0;
1455 out_unlock:
1456         mutex_unlock(&cpuset_mutex);
1457         return ret;
1458 }
1459 
1460 static void cpuset_cancel_attach(struct cgroup_subsys_state *css,
1461                                  struct cgroup_taskset *tset)
1462 {
1463         mutex_lock(&cpuset_mutex);
1464         css_cs(css)->attach_in_progress--;
1465         mutex_unlock(&cpuset_mutex);
1466 }
1467 
1468 /*
1469  * Protected by cpuset_mutex.  cpus_attach is used only by cpuset_attach()
1470  * but we can't allocate it dynamically there.  Define it global and
1471  * allocate from cpuset_init().
1472  */
1473 static cpumask_var_t cpus_attach;
1474 
1475 static void cpuset_attach(struct cgroup_subsys_state *css,
1476                           struct cgroup_taskset *tset)
1477 {
1478         /* static buf protected by cpuset_mutex */
1479         static nodemask_t cpuset_attach_nodemask_to;
1480         struct mm_struct *mm;
1481         struct task_struct *task;
1482         struct task_struct *leader = cgroup_taskset_first(tset);
1483         struct cpuset *cs = css_cs(css);
1484         struct cpuset *oldcs = cpuset_attach_old_cs;
1485 
1486         mutex_lock(&cpuset_mutex);
1487 
1488         /* prepare for attach */
1489         if (cs == &top_cpuset)
1490                 cpumask_copy(cpus_attach, cpu_possible_mask);
1491         else
1492                 guarantee_online_cpus(cs, cpus_attach);
1493 
1494         guarantee_online_mems(cs, &cpuset_attach_nodemask_to);
1495 
1496         cgroup_taskset_for_each(task, tset) {
1497                 /*
1498                  * can_attach beforehand should guarantee that this doesn't
1499                  * fail.  TODO: have a better way to handle failure here
1500                  */
1501                 WARN_ON_ONCE(set_cpus_allowed_ptr(task, cpus_attach));
1502 
1503                 cpuset_change_task_nodemask(task, &cpuset_attach_nodemask_to);
1504                 cpuset_update_task_spread_flag(cs, task);
1505         }
1506 
1507         /*
1508          * Change mm, possibly for multiple threads in a threadgroup. This is
1509          * expensive and may sleep.
1510          */
1511         cpuset_attach_nodemask_to = cs->effective_mems;
1512         mm = get_task_mm(leader);
1513         if (mm) {
1514                 mpol_rebind_mm(mm, &cpuset_attach_nodemask_to);
1515 
1516                 /*
1517                  * old_mems_allowed is the same with mems_allowed here, except
1518                  * if this task is being moved automatically due to hotplug.
1519                  * In that case @mems_allowed has been updated and is empty,
1520                  * so @old_mems_allowed is the right nodesets that we migrate
1521                  * mm from.
1522                  */
1523                 if (is_memory_migrate(cs)) {
1524                         cpuset_migrate_mm(mm, &oldcs->old_mems_allowed,
1525                                           &cpuset_attach_nodemask_to);
1526                 }
1527                 mmput(mm);
1528         }
1529 
1530         cs->old_mems_allowed = cpuset_attach_nodemask_to;
1531 
1532         cs->attach_in_progress--;
1533         if (!cs->attach_in_progress)
1534                 wake_up(&cpuset_attach_wq);
1535 
1536         mutex_unlock(&cpuset_mutex);
1537 }
1538 
1539 /* The various types of files and directories in a cpuset file system */
1540 
1541 typedef enum {
1542         FILE_MEMORY_MIGRATE,
1543         FILE_CPULIST,
1544         FILE_MEMLIST,
1545         FILE_EFFECTIVE_CPULIST,
1546         FILE_EFFECTIVE_MEMLIST,
1547         FILE_CPU_EXCLUSIVE,
1548         FILE_MEM_EXCLUSIVE,
1549         FILE_MEM_HARDWALL,
1550         FILE_SCHED_LOAD_BALANCE,
1551         FILE_SCHED_RELAX_DOMAIN_LEVEL,
1552         FILE_MEMORY_PRESSURE_ENABLED,
1553         FILE_MEMORY_PRESSURE,
1554         FILE_SPREAD_PAGE,
1555         FILE_SPREAD_SLAB,
1556 } cpuset_filetype_t;
1557 
1558 static int cpuset_write_u64(struct cgroup_subsys_state *css, struct cftype *cft,
1559                             u64 val)
1560 {
1561         struct cpuset *cs = css_cs(css);
1562         cpuset_filetype_t type = cft->private;
1563         int retval = 0;
1564 
1565         mutex_lock(&cpuset_mutex);
1566         if (!is_cpuset_online(cs)) {
1567                 retval = -ENODEV;
1568                 goto out_unlock;
1569         }
1570 
1571         switch (type) {
1572         case FILE_CPU_EXCLUSIVE:
1573                 retval = update_flag(CS_CPU_EXCLUSIVE, cs, val);
1574                 break;
1575         case FILE_MEM_EXCLUSIVE:
1576                 retval = update_flag(CS_MEM_EXCLUSIVE, cs, val);
1577                 break;
1578         case FILE_MEM_HARDWALL:
1579                 retval = update_flag(CS_MEM_HARDWALL, cs, val);
1580                 break;
1581         case FILE_SCHED_LOAD_BALANCE:
1582                 retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, val);
1583                 break;
1584         case FILE_MEMORY_MIGRATE:
1585                 retval = update_flag(CS_MEMORY_MIGRATE, cs, val);
1586                 break;
1587         case FILE_MEMORY_PRESSURE_ENABLED:
1588                 cpuset_memory_pressure_enabled = !!val;
1589                 break;
1590         case FILE_MEMORY_PRESSURE:
1591                 retval = -EACCES;
1592                 break;
1593         case FILE_SPREAD_PAGE:
1594                 retval = update_flag(CS_SPREAD_PAGE, cs, val);
1595                 break;
1596         case FILE_SPREAD_SLAB:
1597                 retval = update_flag(CS_SPREAD_SLAB, cs, val);
1598                 break;
1599         default:
1600                 retval = -EINVAL;
1601                 break;
1602         }
1603 out_unlock:
1604         mutex_unlock(&cpuset_mutex);
1605         return retval;
1606 }
1607 
1608 static int cpuset_write_s64(struct cgroup_subsys_state *css, struct cftype *cft,
1609                             s64 val)
1610 {
1611         struct cpuset *cs = css_cs(css);
1612         cpuset_filetype_t type = cft->private;
1613         int retval = -ENODEV;
1614 
1615         mutex_lock(&cpuset_mutex);
1616         if (!is_cpuset_online(cs))
1617                 goto out_unlock;
1618 
1619         switch (type) {
1620         case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1621                 retval = update_relax_domain_level(cs, val);
1622                 break;
1623         default:
1624                 retval = -EINVAL;
1625                 break;
1626         }
1627 out_unlock:
1628         mutex_unlock(&cpuset_mutex);
1629         return retval;
1630 }
1631 
1632 /*
1633  * Common handling for a write to a "cpus" or "mems" file.
1634  */
1635 static ssize_t cpuset_write_resmask(struct kernfs_open_file *of,
1636                                     char *buf, size_t nbytes, loff_t off)
1637 {
1638         struct cpuset *cs = css_cs(of_css(of));
1639         struct cpuset *trialcs;
1640         int retval = -ENODEV;
1641 
1642         buf = strstrip(buf);
1643 
1644         /*
1645          * CPU or memory hotunplug may leave @cs w/o any execution
1646          * resources, in which case the hotplug code asynchronously updates
1647          * configuration and transfers all tasks to the nearest ancestor
1648          * which can execute.
1649          *
1650          * As writes to "cpus" or "mems" may restore @cs's execution
1651          * resources, wait for the previously scheduled operations before
1652          * proceeding, so that we don't end up keep removing tasks added
1653          * after execution capability is restored.
1654          *
1655          * cpuset_hotplug_work calls back into cgroup core via
1656          * cgroup_transfer_tasks() and waiting for it from a cgroupfs
1657          * operation like this one can lead to a deadlock through kernfs
1658          * active_ref protection.  Let's break the protection.  Losing the
1659          * protection is okay as we check whether @cs is online after
1660          * grabbing cpuset_mutex anyway.  This only happens on the legacy
1661          * hierarchies.
1662          */
1663         css_get(&cs->css);
1664         kernfs_break_active_protection(of->kn);
1665         flush_work(&cpuset_hotplug_work);
1666 
1667         mutex_lock(&cpuset_mutex);
1668         if (!is_cpuset_online(cs))
1669                 goto out_unlock;
1670 
1671         trialcs = alloc_trial_cpuset(cs);
1672         if (!trialcs) {
1673                 retval = -ENOMEM;
1674                 goto out_unlock;
1675         }
1676 
1677         switch (of_cft(of)->private) {
1678         case FILE_CPULIST:
1679                 retval = update_cpumask(cs, trialcs, buf);
1680                 break;
1681         case FILE_MEMLIST:
1682                 retval = update_nodemask(cs, trialcs, buf);
1683                 break;
1684         default:
1685                 retval = -EINVAL;
1686                 break;
1687         }
1688 
1689         free_trial_cpuset(trialcs);
1690 out_unlock:
1691         mutex_unlock(&cpuset_mutex);
1692         kernfs_unbreak_active_protection(of->kn);
1693         css_put(&cs->css);
1694         return retval ?: nbytes;
1695 }
1696 
1697 /*
1698  * These ascii lists should be read in a single call, by using a user
1699  * buffer large enough to hold the entire map.  If read in smaller
1700  * chunks, there is no guarantee of atomicity.  Since the display format
1701  * used, list of ranges of sequential numbers, is variable length,
1702  * and since these maps can change value dynamically, one could read
1703  * gibberish by doing partial reads while a list was changing.
1704  */
1705 static int cpuset_common_seq_show(struct seq_file *sf, void *v)
1706 {
1707         struct cpuset *cs = css_cs(seq_css(sf));
1708         cpuset_filetype_t type = seq_cft(sf)->private;
1709         ssize_t count;
1710         char *buf, *s;
1711         int ret = 0;
1712 
1713         count = seq_get_buf(sf, &buf);
1714         s = buf;
1715 
1716         mutex_lock(&callback_mutex);
1717 
1718         switch (type) {
1719         case FILE_CPULIST:
1720                 s += cpulist_scnprintf(s, count, cs->cpus_allowed);
1721                 break;
1722         case FILE_MEMLIST:
1723                 s += nodelist_scnprintf(s, count, cs->mems_allowed);
1724                 break;
1725         case FILE_EFFECTIVE_CPULIST:
1726                 s += cpulist_scnprintf(s, count, cs->effective_cpus);
1727                 break;
1728         case FILE_EFFECTIVE_MEMLIST:
1729                 s += nodelist_scnprintf(s, count, cs->effective_mems);
1730                 break;
1731         default:
1732                 ret = -EINVAL;
1733                 goto out_unlock;
1734         }
1735 
1736         if (s < buf + count - 1) {
1737                 *s++ = '\n';
1738                 seq_commit(sf, s - buf);
1739         } else {
1740                 seq_commit(sf, -1);
1741         }
1742 out_unlock:
1743         mutex_unlock(&callback_mutex);
1744         return ret;
1745 }
1746 
1747 static u64 cpuset_read_u64(struct cgroup_subsys_state *css, struct cftype *cft)
1748 {
1749         struct cpuset *cs = css_cs(css);
1750         cpuset_filetype_t type = cft->private;
1751         switch (type) {
1752         case FILE_CPU_EXCLUSIVE:
1753                 return is_cpu_exclusive(cs);
1754         case FILE_MEM_EXCLUSIVE:
1755                 return is_mem_exclusive(cs);
1756         case FILE_MEM_HARDWALL:
1757                 return is_mem_hardwall(cs);
1758         case FILE_SCHED_LOAD_BALANCE:
1759                 return is_sched_load_balance(cs);
1760         case FILE_MEMORY_MIGRATE:
1761                 return is_memory_migrate(cs);
1762         case FILE_MEMORY_PRESSURE_ENABLED:
1763                 return cpuset_memory_pressure_enabled;
1764         case FILE_MEMORY_PRESSURE:
1765                 return fmeter_getrate(&cs->fmeter);
1766         case FILE_SPREAD_PAGE:
1767                 return is_spread_page(cs);
1768         case FILE_SPREAD_SLAB:
1769                 return is_spread_slab(cs);
1770         default:
1771                 BUG();
1772         }
1773 
1774         /* Unreachable but makes gcc happy */
1775         return 0;
1776 }
1777 
1778 static s64 cpuset_read_s64(struct cgroup_subsys_state *css, struct cftype *cft)
1779 {
1780         struct cpuset *cs = css_cs(css);
1781         cpuset_filetype_t type = cft->private;
1782         switch (type) {
1783         case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1784                 return cs->relax_domain_level;
1785         default:
1786                 BUG();
1787         }
1788 
1789         /* Unrechable but makes gcc happy */
1790         return 0;
1791 }
1792 
1793 
1794 /*
1795  * for the common functions, 'private' gives the type of file
1796  */
1797 
1798 static struct cftype files[] = {
1799         {
1800                 .name = "cpus",
1801                 .seq_show = cpuset_common_seq_show,
1802                 .write = cpuset_write_resmask,
1803                 .max_write_len = (100U + 6 * NR_CPUS),
1804                 .private = FILE_CPULIST,
1805         },
1806 
1807         {
1808                 .name = "mems",
1809                 .seq_show = cpuset_common_seq_show,
1810                 .write = cpuset_write_resmask,
1811                 .max_write_len = (100U + 6 * MAX_NUMNODES),
1812                 .private = FILE_MEMLIST,
1813         },
1814 
1815         {
1816                 .name = "effective_cpus",
1817                 .seq_show = cpuset_common_seq_show,
1818                 .private = FILE_EFFECTIVE_CPULIST,
1819         },
1820 
1821         {
1822                 .name = "effective_mems",
1823                 .seq_show = cpuset_common_seq_show,
1824                 .private = FILE_EFFECTIVE_MEMLIST,
1825         },
1826 
1827         {
1828                 .name = "cpu_exclusive",
1829                 .read_u64 = cpuset_read_u64,
1830                 .write_u64 = cpuset_write_u64,
1831                 .private = FILE_CPU_EXCLUSIVE,
1832         },
1833 
1834         {
1835                 .name = "mem_exclusive",
1836                 .read_u64 = cpuset_read_u64,
1837                 .write_u64 = cpuset_write_u64,
1838                 .private = FILE_MEM_EXCLUSIVE,
1839         },
1840 
1841         {
1842                 .name = "mem_hardwall",
1843                 .read_u64 = cpuset_read_u64,
1844                 .write_u64 = cpuset_write_u64,
1845                 .private = FILE_MEM_HARDWALL,
1846         },
1847 
1848         {
1849                 .name = "sched_load_balance",
1850                 .read_u64 = cpuset_read_u64,
1851                 .write_u64 = cpuset_write_u64,
1852                 .private = FILE_SCHED_LOAD_BALANCE,
1853         },
1854 
1855         {
1856                 .name = "sched_relax_domain_level",
1857                 .read_s64 = cpuset_read_s64,
1858                 .write_s64 = cpuset_write_s64,
1859                 .private = FILE_SCHED_RELAX_DOMAIN_LEVEL,
1860         },
1861 
1862         {
1863                 .name = "memory_migrate",
1864                 .read_u64 = cpuset_read_u64,
1865                 .write_u64 = cpuset_write_u64,
1866                 .private = FILE_MEMORY_MIGRATE,
1867         },
1868 
1869         {
1870                 .name = "memory_pressure",
1871                 .read_u64 = cpuset_read_u64,
1872                 .write_u64 = cpuset_write_u64,
1873                 .private = FILE_MEMORY_PRESSURE,
1874                 .mode = S_IRUGO,
1875         },
1876 
1877         {
1878                 .name = "memory_spread_page",
1879                 .read_u64 = cpuset_read_u64,
1880                 .write_u64 = cpuset_write_u64,
1881                 .private = FILE_SPREAD_PAGE,
1882         },
1883 
1884         {
1885                 .name = "memory_spread_slab",
1886                 .read_u64 = cpuset_read_u64,
1887                 .write_u64 = cpuset_write_u64,
1888                 .private = FILE_SPREAD_SLAB,
1889         },
1890 
1891         {
1892                 .name = "memory_pressure_enabled",
1893                 .flags = CFTYPE_ONLY_ON_ROOT,
1894                 .read_u64 = cpuset_read_u64,
1895                 .write_u64 = cpuset_write_u64,
1896                 .private = FILE_MEMORY_PRESSURE_ENABLED,
1897         },
1898 
1899         { }     /* terminate */
1900 };
1901 
1902 /*
1903  *      cpuset_css_alloc - allocate a cpuset css
1904  *      cgrp:   control group that the new cpuset will be part of
1905  */
1906 
1907 static struct cgroup_subsys_state *
1908 cpuset_css_alloc(struct cgroup_subsys_state *parent_css)
1909 {
1910         struct cpuset *cs;
1911 
1912         if (!parent_css)
1913                 return &top_cpuset.css;
1914 
1915         cs = kzalloc(sizeof(*cs), GFP_KERNEL);
1916         if (!cs)
1917                 return ERR_PTR(-ENOMEM);
1918         if (!alloc_cpumask_var(&cs->cpus_allowed, GFP_KERNEL))
1919                 goto free_cs;
1920         if (!alloc_cpumask_var(&cs->effective_cpus, GFP_KERNEL))
1921                 goto free_cpus;
1922 
1923         set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
1924         cpumask_clear(cs->cpus_allowed);
1925         nodes_clear(cs->mems_allowed);
1926         cpumask_clear(cs->effective_cpus);
1927         nodes_clear(cs->effective_mems);
1928         fmeter_init(&cs->fmeter);
1929         cs->relax_domain_level = -1;
1930 
1931         return &cs->css;
1932 
1933 free_cpus:
1934         free_cpumask_var(cs->cpus_allowed);
1935 free_cs:
1936         kfree(cs);
1937         return ERR_PTR(-ENOMEM);
1938 }
1939 
1940 static int cpuset_css_online(struct cgroup_subsys_state *css)
1941 {
1942         struct cpuset *cs = css_cs(css);
1943         struct cpuset *parent = parent_cs(cs);
1944         struct cpuset *tmp_cs;
1945         struct cgroup_subsys_state *pos_css;
1946 
1947         if (!parent)
1948                 return 0;
1949 
1950         mutex_lock(&cpuset_mutex);
1951 
1952         set_bit(CS_ONLINE, &cs->flags);
1953         if (is_spread_page(parent))
1954                 set_bit(CS_SPREAD_PAGE, &cs->flags);
1955         if (is_spread_slab(parent))
1956                 set_bit(CS_SPREAD_SLAB, &cs->flags);
1957 
1958         cpuset_inc();
1959 
1960         mutex_lock(&callback_mutex);
1961         if (cgroup_on_dfl(cs->css.cgroup)) {
1962                 cpumask_copy(cs->effective_cpus, parent->effective_cpus);
1963                 cs->effective_mems = parent->effective_mems;
1964         }
1965         mutex_unlock(&callback_mutex);
1966 
1967         if (!test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags))
1968                 goto out_unlock;
1969 
1970         /*
1971          * Clone @parent's configuration if CGRP_CPUSET_CLONE_CHILDREN is
1972          * set.  This flag handling is implemented in cgroup core for
1973          * histrical reasons - the flag may be specified during mount.
1974          *
1975          * Currently, if any sibling cpusets have exclusive cpus or mem, we
1976          * refuse to clone the configuration - thereby refusing the task to
1977          * be entered, and as a result refusing the sys_unshare() or
1978          * clone() which initiated it.  If this becomes a problem for some
1979          * users who wish to allow that scenario, then this could be
1980          * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
1981          * (and likewise for mems) to the new cgroup.
1982          */
1983         rcu_read_lock();
1984         cpuset_for_each_child(tmp_cs, pos_css, parent) {
1985                 if (is_mem_exclusive(tmp_cs) || is_cpu_exclusive(tmp_cs)) {
1986                         rcu_read_unlock();
1987                         goto out_unlock;
1988                 }
1989         }
1990         rcu_read_unlock();
1991 
1992         mutex_lock(&callback_mutex);
1993         cs->mems_allowed = parent->mems_allowed;
1994         cpumask_copy(cs->cpus_allowed, parent->cpus_allowed);
1995         mutex_unlock(&callback_mutex);
1996 out_unlock:
1997         mutex_unlock(&cpuset_mutex);
1998         return 0;
1999 }
2000 
2001 /*
2002  * If the cpuset being removed has its flag 'sched_load_balance'
2003  * enabled, then simulate turning sched_load_balance off, which
2004  * will call rebuild_sched_domains_locked().
2005  */
2006 
2007 static void cpuset_css_offline(struct cgroup_subsys_state *css)
2008 {
2009         struct cpuset *cs = css_cs(css);
2010 
2011         mutex_lock(&cpuset_mutex);
2012 
2013         if (is_sched_load_balance(cs))
2014                 update_flag(CS_SCHED_LOAD_BALANCE, cs, 0);
2015 
2016         cpuset_dec();
2017         clear_bit(CS_ONLINE, &cs->flags);
2018 
2019         mutex_unlock(&cpuset_mutex);
2020 }
2021 
2022 static void cpuset_css_free(struct cgroup_subsys_state *css)
2023 {
2024         struct cpuset *cs = css_cs(css);
2025 
2026         free_cpumask_var(cs->effective_cpus);
2027         free_cpumask_var(cs->cpus_allowed);
2028         kfree(cs);
2029 }
2030 
2031 static void cpuset_bind(struct cgroup_subsys_state *root_css)
2032 {
2033         mutex_lock(&cpuset_mutex);
2034         mutex_lock(&callback_mutex);
2035 
2036         if (cgroup_on_dfl(root_css->cgroup)) {
2037                 cpumask_copy(top_cpuset.cpus_allowed, cpu_possible_mask);
2038                 top_cpuset.mems_allowed = node_possible_map;
2039         } else {
2040                 cpumask_copy(top_cpuset.cpus_allowed,
2041                              top_cpuset.effective_cpus);
2042                 top_cpuset.mems_allowed = top_cpuset.effective_mems;
2043         }
2044 
2045         mutex_unlock(&callback_mutex);
2046         mutex_unlock(&cpuset_mutex);
2047 }
2048 
2049 struct cgroup_subsys cpuset_cgrp_subsys = {
2050         .css_alloc      = cpuset_css_alloc,
2051         .css_online     = cpuset_css_online,
2052         .css_offline    = cpuset_css_offline,
2053         .css_free       = cpuset_css_free,
2054         .can_attach     = cpuset_can_attach,
2055         .cancel_attach  = cpuset_cancel_attach,
2056         .attach         = cpuset_attach,
2057         .bind           = cpuset_bind,
2058         .legacy_cftypes = files,
2059         .early_init     = 1,
2060 };
2061 
2062 /**
2063  * cpuset_init - initialize cpusets at system boot
2064  *
2065  * Description: Initialize top_cpuset and the cpuset internal file system,
2066  **/
2067 
2068 int __init cpuset_init(void)
2069 {
2070         int err = 0;
2071 
2072         if (!alloc_cpumask_var(&top_cpuset.cpus_allowed, GFP_KERNEL))
2073                 BUG();
2074         if (!alloc_cpumask_var(&top_cpuset.effective_cpus, GFP_KERNEL))
2075                 BUG();
2076 
2077         cpumask_setall(top_cpuset.cpus_allowed);
2078         nodes_setall(top_cpuset.mems_allowed);
2079         cpumask_setall(top_cpuset.effective_cpus);
2080         nodes_setall(top_cpuset.effective_mems);
2081 
2082         fmeter_init(&top_cpuset.fmeter);
2083         set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags);
2084         top_cpuset.relax_domain_level = -1;
2085 
2086         err = register_filesystem(&cpuset_fs_type);
2087         if (err < 0)
2088                 return err;
2089 
2090         if (!alloc_cpumask_var(&cpus_attach, GFP_KERNEL))
2091                 BUG();
2092 
2093         return 0;
2094 }
2095 
2096 /*
2097  * If CPU and/or memory hotplug handlers, below, unplug any CPUs
2098  * or memory nodes, we need to walk over the cpuset hierarchy,
2099  * removing that CPU or node from all cpusets.  If this removes the
2100  * last CPU or node from a cpuset, then move the tasks in the empty
2101  * cpuset to its next-highest non-empty parent.
2102  */
2103 static void remove_tasks_in_empty_cpuset(struct cpuset *cs)
2104 {
2105         struct cpuset *parent;
2106 
2107         /*
2108          * Find its next-highest non-empty parent, (top cpuset
2109          * has online cpus, so can't be empty).
2110          */
2111         parent = parent_cs(cs);
2112         while (cpumask_empty(parent->cpus_allowed) ||
2113                         nodes_empty(parent->mems_allowed))
2114                 parent = parent_cs(parent);
2115 
2116         if (cgroup_transfer_tasks(parent->css.cgroup, cs->css.cgroup)) {
2117                 pr_err("cpuset: failed to transfer tasks out of empty cpuset ");
2118                 pr_cont_cgroup_name(cs->css.cgroup);
2119                 pr_cont("\n");
2120         }
2121 }
2122 
2123 static void
2124 hotplug_update_tasks_legacy(struct cpuset *cs,
2125                             struct cpumask *new_cpus, nodemask_t *new_mems,
2126                             bool cpus_updated, bool mems_updated)
2127 {
2128         bool is_empty;
2129 
2130         mutex_lock(&callback_mutex);
2131         cpumask_copy(cs->cpus_allowed, new_cpus);
2132         cpumask_copy(cs->effective_cpus, new_cpus);
2133         cs->mems_allowed = *new_mems;
2134         cs->effective_mems = *new_mems;
2135         mutex_unlock(&callback_mutex);
2136 
2137         /*
2138          * Don't call update_tasks_cpumask() if the cpuset becomes empty,
2139          * as the tasks will be migratecd to an ancestor.
2140          */
2141         if (cpus_updated && !cpumask_empty(cs->cpus_allowed))
2142                 update_tasks_cpumask(cs);
2143         if (mems_updated && !nodes_empty(cs->mems_allowed))
2144                 update_tasks_nodemask(cs);
2145 
2146         is_empty = cpumask_empty(cs->cpus_allowed) ||
2147                    nodes_empty(cs->mems_allowed);
2148 
2149         mutex_unlock(&cpuset_mutex);
2150 
2151         /*
2152          * Move tasks to the nearest ancestor with execution resources,
2153          * This is full cgroup operation which will also call back into
2154          * cpuset. Should be done outside any lock.
2155          */
2156         if (is_empty)
2157                 remove_tasks_in_empty_cpuset(cs);
2158 
2159         mutex_lock(&cpuset_mutex);
2160 }
2161 
2162 static void
2163 hotplug_update_tasks(struct cpuset *cs,
2164                      struct cpumask *new_cpus, nodemask_t *new_mems,
2165                      bool cpus_updated, bool mems_updated)
2166 {
2167         if (cpumask_empty(new_cpus))
2168                 cpumask_copy(new_cpus, parent_cs(cs)->effective_cpus);
2169         if (nodes_empty(*new_mems))
2170                 *new_mems = parent_cs(cs)->effective_mems;
2171 
2172         mutex_lock(&callback_mutex);
2173         cpumask_copy(cs->effective_cpus, new_cpus);
2174         cs->effective_mems = *new_mems;
2175         mutex_unlock(&callback_mutex);
2176 
2177         if (cpus_updated)
2178                 update_tasks_cpumask(cs);
2179         if (mems_updated)
2180                 update_tasks_nodemask(cs);
2181 }
2182 
2183 /**
2184  * cpuset_hotplug_update_tasks - update tasks in a cpuset for hotunplug
2185  * @cs: cpuset in interest
2186  *
2187  * Compare @cs's cpu and mem masks against top_cpuset and if some have gone
2188  * offline, update @cs accordingly.  If @cs ends up with no CPU or memory,
2189  * all its tasks are moved to the nearest ancestor with both resources.
2190  */
2191 static void cpuset_hotplug_update_tasks(struct cpuset *cs)
2192 {
2193         static cpumask_t new_cpus;
2194         static nodemask_t new_mems;
2195         bool cpus_updated;
2196         bool mems_updated;
2197 retry:
2198         wait_event(cpuset_attach_wq, cs->attach_in_progress == 0);
2199 
2200         mutex_lock(&cpuset_mutex);
2201 
2202         /*
2203          * We have raced with task attaching. We wait until attaching
2204          * is finished, so we won't attach a task to an empty cpuset.
2205          */
2206         if (cs->attach_in_progress) {
2207                 mutex_unlock(&cpuset_mutex);
2208                 goto retry;
2209         }
2210 
2211         cpumask_and(&new_cpus, cs->cpus_allowed, parent_cs(cs)->effective_cpus);
2212         nodes_and(new_mems, cs->mems_allowed, parent_cs(cs)->effective_mems);
2213 
2214         cpus_updated = !cpumask_equal(&new_cpus, cs->effective_cpus);
2215         mems_updated = !nodes_equal(new_mems, cs->effective_mems);
2216 
2217         if (cgroup_on_dfl(cs->css.cgroup))
2218                 hotplug_update_tasks(cs, &new_cpus, &new_mems,
2219                                      cpus_updated, mems_updated);
2220         else
2221                 hotplug_update_tasks_legacy(cs, &new_cpus, &new_mems,
2222                                             cpus_updated, mems_updated);
2223 
2224         mutex_unlock(&cpuset_mutex);
2225 }
2226 
2227 /**
2228  * cpuset_hotplug_workfn - handle CPU/memory hotunplug for a cpuset
2229  *
2230  * This function is called after either CPU or memory configuration has
2231  * changed and updates cpuset accordingly.  The top_cpuset is always
2232  * synchronized to cpu_active_mask and N_MEMORY, which is necessary in
2233  * order to make cpusets transparent (of no affect) on systems that are
2234  * actively using CPU hotplug but making no active use of cpusets.
2235  *
2236  * Non-root cpusets are only affected by offlining.  If any CPUs or memory
2237  * nodes have been taken down, cpuset_hotplug_update_tasks() is invoked on
2238  * all descendants.
2239  *
2240  * Note that CPU offlining during suspend is ignored.  We don't modify
2241  * cpusets across suspend/resume cycles at all.
2242  */
2243 static void cpuset_hotplug_workfn(struct work_struct *work)
2244 {
2245         static cpumask_t new_cpus;
2246         static nodemask_t new_mems;
2247         bool cpus_updated, mems_updated;
2248         bool on_dfl = cgroup_on_dfl(top_cpuset.css.cgroup);
2249 
2250         mutex_lock(&cpuset_mutex);
2251 
2252         /* fetch the available cpus/mems and find out which changed how */
2253         cpumask_copy(&new_cpus, cpu_active_mask);
2254         new_mems = node_states[N_MEMORY];
2255 
2256         cpus_updated = !cpumask_equal(top_cpuset.effective_cpus, &new_cpus);
2257         mems_updated = !nodes_equal(top_cpuset.effective_mems, new_mems);
2258 
2259         /* synchronize cpus_allowed to cpu_active_mask */
2260         if (cpus_updated) {
2261                 mutex_lock(&callback_mutex);
2262                 if (!on_dfl)
2263                         cpumask_copy(top_cpuset.cpus_allowed, &new_cpus);
2264                 cpumask_copy(top_cpuset.effective_cpus, &new_cpus);
2265                 mutex_unlock(&callback_mutex);
2266                 /* we don't mess with cpumasks of tasks in top_cpuset */
2267         }
2268 
2269         /* synchronize mems_allowed to N_MEMORY */
2270         if (mems_updated) {
2271                 mutex_lock(&callback_mutex);
2272                 if (!on_dfl)
2273                         top_cpuset.mems_allowed = new_mems;
2274                 top_cpuset.effective_mems = new_mems;
2275                 mutex_unlock(&callback_mutex);
2276                 update_tasks_nodemask(&top_cpuset);
2277         }
2278 
2279         mutex_unlock(&cpuset_mutex);
2280 
2281         /* if cpus or mems changed, we need to propagate to descendants */
2282         if (cpus_updated || mems_updated) {
2283                 struct cpuset *cs;
2284                 struct cgroup_subsys_state *pos_css;
2285 
2286                 rcu_read_lock();
2287                 cpuset_for_each_descendant_pre(cs, pos_css, &top_cpuset) {
2288                         if (cs == &top_cpuset || !css_tryget_online(&cs->css))
2289                                 continue;
2290                         rcu_read_unlock();
2291 
2292                         cpuset_hotplug_update_tasks(cs);
2293 
2294                         rcu_read_lock();
2295                         css_put(&cs->css);
2296                 }
2297                 rcu_read_unlock();
2298         }
2299 
2300         /* rebuild sched domains if cpus_allowed has changed */
2301         if (cpus_updated)
2302                 rebuild_sched_domains();
2303 }
2304 
2305 void cpuset_update_active_cpus(bool cpu_online)
2306 {
2307         /*
2308          * We're inside cpu hotplug critical region which usually nests
2309          * inside cgroup synchronization.  Bounce actual hotplug processing
2310          * to a work item to avoid reverse locking order.
2311          *
2312          * We still need to do partition_sched_domains() synchronously;
2313          * otherwise, the scheduler will get confused and put tasks to the
2314          * dead CPU.  Fall back to the default single domain.
2315          * cpuset_hotplug_workfn() will rebuild it as necessary.
2316          */
2317         partition_sched_domains(1, NULL, NULL);
2318         schedule_work(&cpuset_hotplug_work);
2319 }
2320 
2321 /*
2322  * Keep top_cpuset.mems_allowed tracking node_states[N_MEMORY].
2323  * Call this routine anytime after node_states[N_MEMORY] changes.
2324  * See cpuset_update_active_cpus() for CPU hotplug handling.
2325  */
2326 static int cpuset_track_online_nodes(struct notifier_block *self,
2327                                 unsigned long action, void *arg)
2328 {
2329         schedule_work(&cpuset_hotplug_work);
2330         return NOTIFY_OK;
2331 }
2332 
2333 static struct notifier_block cpuset_track_online_nodes_nb = {
2334         .notifier_call = cpuset_track_online_nodes,
2335         .priority = 10,         /* ??! */
2336 };
2337 
2338 /**
2339  * cpuset_init_smp - initialize cpus_allowed
2340  *
2341  * Description: Finish top cpuset after cpu, node maps are initialized
2342  */
2343 void __init cpuset_init_smp(void)
2344 {
2345         cpumask_copy(top_cpuset.cpus_allowed, cpu_active_mask);
2346         top_cpuset.mems_allowed = node_states[N_MEMORY];
2347         top_cpuset.old_mems_allowed = top_cpuset.mems_allowed;
2348 
2349         cpumask_copy(top_cpuset.effective_cpus, cpu_active_mask);
2350         top_cpuset.effective_mems = node_states[N_MEMORY];
2351 
2352         register_hotmemory_notifier(&cpuset_track_online_nodes_nb);
2353 }
2354 
2355 /**
2356  * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
2357  * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
2358  * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
2359  *
2360  * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
2361  * attached to the specified @tsk.  Guaranteed to return some non-empty
2362  * subset of cpu_online_mask, even if this means going outside the
2363  * tasks cpuset.
2364  **/
2365 
2366 void cpuset_cpus_allowed(struct task_struct *tsk, struct cpumask *pmask)
2367 {
2368         mutex_lock(&callback_mutex);
2369         rcu_read_lock();
2370         guarantee_online_cpus(task_cs(tsk), pmask);
2371         rcu_read_unlock();
2372         mutex_unlock(&callback_mutex);
2373 }
2374 
2375 void cpuset_cpus_allowed_fallback(struct task_struct *tsk)
2376 {
2377         rcu_read_lock();
2378         do_set_cpus_allowed(tsk, task_cs(tsk)->effective_cpus);
2379         rcu_read_unlock();
2380 
2381         /*
2382          * We own tsk->cpus_allowed, nobody can change it under us.
2383          *
2384          * But we used cs && cs->cpus_allowed lockless and thus can
2385          * race with cgroup_attach_task() or update_cpumask() and get
2386          * the wrong tsk->cpus_allowed. However, both cases imply the
2387          * subsequent cpuset_change_cpumask()->set_cpus_allowed_ptr()
2388          * which takes task_rq_lock().
2389          *
2390          * If we are called after it dropped the lock we must see all
2391          * changes in tsk_cs()->cpus_allowed. Otherwise we can temporary
2392          * set any mask even if it is not right from task_cs() pov,
2393          * the pending set_cpus_allowed_ptr() will fix things.
2394          *
2395          * select_fallback_rq() will fix things ups and set cpu_possible_mask
2396          * if required.
2397          */
2398 }
2399 
2400 void cpuset_init_current_mems_allowed(void)
2401 {
2402         nodes_setall(current->mems_allowed);
2403 }
2404 
2405 /**
2406  * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
2407  * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
2408  *
2409  * Description: Returns the nodemask_t mems_allowed of the cpuset
2410  * attached to the specified @tsk.  Guaranteed to return some non-empty
2411  * subset of node_states[N_MEMORY], even if this means going outside the
2412  * tasks cpuset.
2413  **/
2414 
2415 nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
2416 {
2417         nodemask_t mask;
2418 
2419         mutex_lock(&callback_mutex);
2420         rcu_read_lock();
2421         guarantee_online_mems(task_cs(tsk), &mask);
2422         rcu_read_unlock();
2423         mutex_unlock(&callback_mutex);
2424 
2425         return mask;
2426 }
2427 
2428 /**
2429  * cpuset_nodemask_valid_mems_allowed - check nodemask vs. curremt mems_allowed
2430  * @nodemask: the nodemask to be checked
2431  *
2432  * Are any of the nodes in the nodemask allowed in current->mems_allowed?
2433  */
2434 int cpuset_nodemask_valid_mems_allowed(nodemask_t *nodemask)
2435 {
2436         return nodes_intersects(*nodemask, current->mems_allowed);
2437 }
2438 
2439 /*
2440  * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
2441  * mem_hardwall ancestor to the specified cpuset.  Call holding
2442  * callback_mutex.  If no ancestor is mem_exclusive or mem_hardwall
2443  * (an unusual configuration), then returns the root cpuset.
2444  */
2445 static struct cpuset *nearest_hardwall_ancestor(struct cpuset *cs)
2446 {
2447         while (!(is_mem_exclusive(cs) || is_mem_hardwall(cs)) && parent_cs(cs))
2448                 cs = parent_cs(cs);
2449         return cs;
2450 }
2451 
2452 /**
2453  * cpuset_node_allowed_softwall - Can we allocate on a memory node?
2454  * @node: is this an allowed node?
2455  * @gfp_mask: memory allocation flags
2456  *
2457  * If we're in interrupt, yes, we can always allocate.  If __GFP_THISNODE is
2458  * set, yes, we can always allocate.  If node is in our task's mems_allowed,
2459  * yes.  If it's not a __GFP_HARDWALL request and this node is in the nearest
2460  * hardwalled cpuset ancestor to this task's cpuset, yes.  If the task has been
2461  * OOM killed and has access to memory reserves as specified by the TIF_MEMDIE
2462  * flag, yes.
2463  * Otherwise, no.
2464  *
2465  * If __GFP_HARDWALL is set, cpuset_node_allowed_softwall() reduces to
2466  * cpuset_node_allowed_hardwall().  Otherwise, cpuset_node_allowed_softwall()
2467  * might sleep, and might allow a node from an enclosing cpuset.
2468  *
2469  * cpuset_node_allowed_hardwall() only handles the simpler case of hardwall
2470  * cpusets, and never sleeps.
2471  *
2472  * The __GFP_THISNODE placement logic is really handled elsewhere,
2473  * by forcibly using a zonelist starting at a specified node, and by
2474  * (in get_page_from_freelist()) refusing to consider the zones for
2475  * any node on the zonelist except the first.  By the time any such
2476  * calls get to this routine, we should just shut up and say 'yes'.
2477  *
2478  * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
2479  * and do not allow allocations outside the current tasks cpuset
2480  * unless the task has been OOM killed as is marked TIF_MEMDIE.
2481  * GFP_KERNEL allocations are not so marked, so can escape to the
2482  * nearest enclosing hardwalled ancestor cpuset.
2483  *
2484  * Scanning up parent cpusets requires callback_mutex.  The
2485  * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
2486  * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
2487  * current tasks mems_allowed came up empty on the first pass over
2488  * the zonelist.  So only GFP_KERNEL allocations, if all nodes in the
2489  * cpuset are short of memory, might require taking the callback_mutex
2490  * mutex.
2491  *
2492  * The first call here from mm/page_alloc:get_page_from_freelist()
2493  * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
2494  * so no allocation on a node outside the cpuset is allowed (unless
2495  * in interrupt, of course).
2496  *
2497  * The second pass through get_page_from_freelist() doesn't even call
2498  * here for GFP_ATOMIC calls.  For those calls, the __alloc_pages()
2499  * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
2500  * in alloc_flags.  That logic and the checks below have the combined
2501  * affect that:
2502  *      in_interrupt - any node ok (current task context irrelevant)
2503  *      GFP_ATOMIC   - any node ok
2504  *      TIF_MEMDIE   - any node ok
2505  *      GFP_KERNEL   - any node in enclosing hardwalled cpuset ok
2506  *      GFP_USER     - only nodes in current tasks mems allowed ok.
2507  *
2508  * Rule:
2509  *    Don't call cpuset_node_allowed_softwall if you can't sleep, unless you
2510  *    pass in the __GFP_HARDWALL flag set in gfp_flag, which disables
2511  *    the code that might scan up ancestor cpusets and sleep.
2512  */
2513 int __cpuset_node_allowed_softwall(int node, gfp_t gfp_mask)
2514 {
2515         struct cpuset *cs;              /* current cpuset ancestors */
2516         int allowed;                    /* is allocation in zone z allowed? */
2517 
2518         if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
2519                 return 1;
2520         might_sleep_if(!(gfp_mask & __GFP_HARDWALL));
2521         if (node_isset(node, current->mems_allowed))
2522                 return 1;
2523         /*
2524          * Allow tasks that have access to memory reserves because they have
2525          * been OOM killed to get memory anywhere.
2526          */
2527         if (unlikely(test_thread_flag(TIF_MEMDIE)))
2528                 return 1;
2529         if (gfp_mask & __GFP_HARDWALL)  /* If hardwall request, stop here */
2530                 return 0;
2531 
2532         if (current->flags & PF_EXITING) /* Let dying task have memory */
2533                 return 1;
2534 
2535         /* Not hardwall and node outside mems_allowed: scan up cpusets */
2536         mutex_lock(&callback_mutex);
2537 
2538         rcu_read_lock();
2539         cs = nearest_hardwall_ancestor(task_cs(current));
2540         allowed = node_isset(node, cs->mems_allowed);
2541         rcu_read_unlock();
2542 
2543         mutex_unlock(&callback_mutex);
2544         return allowed;
2545 }
2546 
2547 /*
2548  * cpuset_node_allowed_hardwall - Can we allocate on a memory node?
2549  * @node: is this an allowed node?
2550  * @gfp_mask: memory allocation flags
2551  *
2552  * If we're in interrupt, yes, we can always allocate.  If __GFP_THISNODE is
2553  * set, yes, we can always allocate.  If node is in our task's mems_allowed,
2554  * yes.  If the task has been OOM killed and has access to memory reserves as
2555  * specified by the TIF_MEMDIE flag, yes.
2556  * Otherwise, no.
2557  *
2558  * The __GFP_THISNODE placement logic is really handled elsewhere,
2559  * by forcibly using a zonelist starting at a specified node, and by
2560  * (in get_page_from_freelist()) refusing to consider the zones for
2561  * any node on the zonelist except the first.  By the time any such
2562  * calls get to this routine, we should just shut up and say 'yes'.
2563  *
2564  * Unlike the cpuset_node_allowed_softwall() variant, above,
2565  * this variant requires that the node be in the current task's
2566  * mems_allowed or that we're in interrupt.  It does not scan up the
2567  * cpuset hierarchy for the nearest enclosing mem_exclusive cpuset.
2568  * It never sleeps.
2569  */
2570 int __cpuset_node_allowed_hardwall(int node, gfp_t gfp_mask)
2571 {
2572         if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
2573                 return 1;
2574         if (node_isset(node, current->mems_allowed))
2575                 return 1;
2576         /*
2577          * Allow tasks that have access to memory reserves because they have
2578          * been OOM killed to get memory anywhere.
2579          */
2580         if (unlikely(test_thread_flag(TIF_MEMDIE)))
2581                 return 1;
2582         return 0;
2583 }
2584 
2585 /**
2586  * cpuset_mem_spread_node() - On which node to begin search for a file page
2587  * cpuset_slab_spread_node() - On which node to begin search for a slab page
2588  *
2589  * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
2590  * tasks in a cpuset with is_spread_page or is_spread_slab set),
2591  * and if the memory allocation used cpuset_mem_spread_node()
2592  * to determine on which node to start looking, as it will for
2593  * certain page cache or slab cache pages such as used for file
2594  * system buffers and inode caches, then instead of starting on the
2595  * local node to look for a free page, rather spread the starting
2596  * node around the tasks mems_allowed nodes.
2597  *
2598  * We don't have to worry about the returned node being offline
2599  * because "it can't happen", and even if it did, it would be ok.
2600  *
2601  * The routines calling guarantee_online_mems() are careful to
2602  * only set nodes in task->mems_allowed that are online.  So it
2603  * should not be possible for the following code to return an
2604  * offline node.  But if it did, that would be ok, as this routine
2605  * is not returning the node where the allocation must be, only
2606  * the node where the search should start.  The zonelist passed to
2607  * __alloc_pages() will include all nodes.  If the slab allocator
2608  * is passed an offline node, it will fall back to the local node.
2609  * See kmem_cache_alloc_node().
2610  */
2611 
2612 static int cpuset_spread_node(int *rotor)
2613 {
2614         int node;
2615 
2616         node = next_node(*rotor, current->mems_allowed);
2617         if (node == MAX_NUMNODES)
2618                 node = first_node(current->mems_allowed);
2619         *rotor = node;
2620         return node;
2621 }
2622 
2623 int cpuset_mem_spread_node(void)
2624 {
2625         if (current->cpuset_mem_spread_rotor == NUMA_NO_NODE)
2626                 current->cpuset_mem_spread_rotor =
2627                         node_random(&current->mems_allowed);
2628 
2629         return cpuset_spread_node(&current->cpuset_mem_spread_rotor);
2630 }
2631 
2632 int cpuset_slab_spread_node(void)
2633 {
2634         if (current->cpuset_slab_spread_rotor == NUMA_NO_NODE)
2635                 current->cpuset_slab_spread_rotor =
2636                         node_random(&current->mems_allowed);
2637 
2638         return cpuset_spread_node(&current->cpuset_slab_spread_rotor);
2639 }
2640 
2641 EXPORT_SYMBOL_GPL(cpuset_mem_spread_node);
2642 
2643 /**
2644  * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
2645  * @tsk1: pointer to task_struct of some task.
2646  * @tsk2: pointer to task_struct of some other task.
2647  *
2648  * Description: Return true if @tsk1's mems_allowed intersects the
2649  * mems_allowed of @tsk2.  Used by the OOM killer to determine if
2650  * one of the task's memory usage might impact the memory available
2651  * to the other.
2652  **/
2653 
2654 int cpuset_mems_allowed_intersects(const struct task_struct *tsk1,
2655                                    const struct task_struct *tsk2)
2656 {
2657         return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed);
2658 }
2659 
2660 #define CPUSET_NODELIST_LEN     (256)
2661 
2662 /**
2663  * cpuset_print_task_mems_allowed - prints task's cpuset and mems_allowed
2664  * @tsk: pointer to task_struct of some task.
2665  *
2666  * Description: Prints @task's name, cpuset name, and cached copy of its
2667  * mems_allowed to the kernel log.
2668  */
2669 void cpuset_print_task_mems_allowed(struct task_struct *tsk)
2670 {
2671          /* Statically allocated to prevent using excess stack. */
2672         static char cpuset_nodelist[CPUSET_NODELIST_LEN];
2673         static DEFINE_SPINLOCK(cpuset_buffer_lock);
2674         struct cgroup *cgrp;
2675 
2676         spin_lock(&cpuset_buffer_lock);
2677         rcu_read_lock();
2678 
2679         cgrp = task_cs(tsk)->css.cgroup;
2680         nodelist_scnprintf(cpuset_nodelist, CPUSET_NODELIST_LEN,
2681                            tsk->mems_allowed);
2682         pr_info("%s cpuset=", tsk->comm);
2683         pr_cont_cgroup_name(cgrp);
2684         pr_cont(" mems_allowed=%s\n", cpuset_nodelist);
2685 
2686         rcu_read_unlock();
2687         spin_unlock(&cpuset_buffer_lock);
2688 }
2689 
2690 /*
2691  * Collection of memory_pressure is suppressed unless
2692  * this flag is enabled by writing "1" to the special
2693  * cpuset file 'memory_pressure_enabled' in the root cpuset.
2694  */
2695 
2696 int cpuset_memory_pressure_enabled __read_mostly;
2697 
2698 /**
2699  * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
2700  *
2701  * Keep a running average of the rate of synchronous (direct)
2702  * page reclaim efforts initiated by tasks in each cpuset.
2703  *
2704  * This represents the rate at which some task in the cpuset
2705  * ran low on memory on all nodes it was allowed to use, and
2706  * had to enter the kernels page reclaim code in an effort to
2707  * create more free memory by tossing clean pages or swapping
2708  * or writing dirty pages.
2709  *
2710  * Display to user space in the per-cpuset read-only file
2711  * "memory_pressure".  Value displayed is an integer
2712  * representing the recent rate of entry into the synchronous
2713  * (direct) page reclaim by any task attached to the cpuset.
2714  **/
2715 
2716 void __cpuset_memory_pressure_bump(void)
2717 {
2718         rcu_read_lock();
2719         fmeter_markevent(&task_cs(current)->fmeter);
2720         rcu_read_unlock();
2721 }
2722 
2723 #ifdef CONFIG_PROC_PID_CPUSET
2724 /*
2725  * proc_cpuset_show()
2726  *  - Print tasks cpuset path into seq_file.
2727  *  - Used for /proc/<pid>/cpuset.
2728  *  - No need to task_lock(tsk) on this tsk->cpuset reference, as it
2729  *    doesn't really matter if tsk->cpuset changes after we read it,
2730  *    and we take cpuset_mutex, keeping cpuset_attach() from changing it
2731  *    anyway.
2732  */
2733 int proc_cpuset_show(struct seq_file *m, void *unused_v)
2734 {
2735         struct pid *pid;
2736         struct task_struct *tsk;
2737         char *buf, *p;
2738         struct cgroup_subsys_state *css;
2739         int retval;
2740 
2741         retval = -ENOMEM;
2742         buf = kmalloc(PATH_MAX, GFP_KERNEL);
2743         if (!buf)
2744                 goto out;
2745 
2746         retval = -ESRCH;
2747         pid = m->private;
2748         tsk = get_pid_task(pid, PIDTYPE_PID);
2749         if (!tsk)
2750                 goto out_free;
2751 
2752         retval = -ENAMETOOLONG;
2753         rcu_read_lock();
2754         css = task_css(tsk, cpuset_cgrp_id);
2755         p = cgroup_path(css->cgroup, buf, PATH_MAX);
2756         rcu_read_unlock();
2757         if (!p)
2758                 goto out_put_task;
2759         seq_puts(m, p);
2760         seq_putc(m, '\n');
2761         retval = 0;
2762 out_put_task:
2763         put_task_struct(tsk);
2764 out_free:
2765         kfree(buf);
2766 out:
2767         return retval;
2768 }
2769 #endif /* CONFIG_PROC_PID_CPUSET */
2770 
2771 /* Display task mems_allowed in /proc/<pid>/status file. */
2772 void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task)
2773 {
2774         seq_puts(m, "Mems_allowed:\t");
2775         seq_nodemask(m, &task->mems_allowed);
2776         seq_puts(m, "\n");
2777         seq_puts(m, "Mems_allowed_list:\t");
2778         seq_nodemask_list(m, &task->mems_allowed);
2779         seq_puts(m, "\n");
2780 }
2781 

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