Version:  2.0.40 2.2.26 2.4.37 3.8 3.9 3.10 3.11 3.12 3.13 3.14 3.15 3.16 3.17 3.18 3.19 4.0 4.1 4.2 4.3 4.4 4.5

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/time64.h>
 55 #include <linux/backing-dev.h>
 56 #include <linux/sort.h>
 57 
 58 #include <asm/uaccess.h>
 59 #include <linux/atomic.h>
 60 #include <linux/mutex.h>
 61 #include <linux/workqueue.h>
 62 #include <linux/cgroup.h>
 63 #include <linux/wait.h>
 64 
 65 struct static_key cpusets_enabled_key __read_mostly = STATIC_KEY_INIT_FALSE;
 66 
 67 /* See "Frequency meter" comments, below. */
 68 
 69 struct fmeter {
 70         int cnt;                /* unprocessed events count */
 71         int val;                /* most recent output value */
 72         time64_t time;          /* clock (secs) when val computed */
 73         spinlock_t lock;        /* guards read or write of above */
 74 };
 75 
 76 struct cpuset {
 77         struct cgroup_subsys_state css;
 78 
 79         unsigned long flags;            /* "unsigned long" so bitops work */
 80 
 81         /*
 82          * On default hierarchy:
 83          *
 84          * The user-configured masks can only be changed by writing to
 85          * cpuset.cpus and cpuset.mems, and won't be limited by the
 86          * parent masks.
 87          *
 88          * The effective masks is the real masks that apply to the tasks
 89          * in the cpuset. They may be changed if the configured masks are
 90          * changed or hotplug happens.
 91          *
 92          * effective_mask == configured_mask & parent's effective_mask,
 93          * and if it ends up empty, it will inherit the parent's mask.
 94          *
 95          *
 96          * On legacy hierachy:
 97          *
 98          * The user-configured masks are always the same with effective masks.
 99          */
100 
101         /* user-configured CPUs and Memory Nodes allow to tasks */
102         cpumask_var_t cpus_allowed;
103         nodemask_t mems_allowed;
104 
105         /* effective CPUs and Memory Nodes allow to tasks */
106         cpumask_var_t effective_cpus;
107         nodemask_t effective_mems;
108 
109         /*
110          * This is old Memory Nodes tasks took on.
111          *
112          * - top_cpuset.old_mems_allowed is initialized to mems_allowed.
113          * - A new cpuset's old_mems_allowed is initialized when some
114          *   task is moved into it.
115          * - old_mems_allowed is used in cpuset_migrate_mm() when we change
116          *   cpuset.mems_allowed and have tasks' nodemask updated, and
117          *   then old_mems_allowed is updated to mems_allowed.
118          */
119         nodemask_t old_mems_allowed;
120 
121         struct fmeter fmeter;           /* memory_pressure filter */
122 
123         /*
124          * Tasks are being attached to this cpuset.  Used to prevent
125          * zeroing cpus/mems_allowed between ->can_attach() and ->attach().
126          */
127         int attach_in_progress;
128 
129         /* partition number for rebuild_sched_domains() */
130         int pn;
131 
132         /* for custom sched domain */
133         int relax_domain_level;
134 };
135 
136 static inline struct cpuset *css_cs(struct cgroup_subsys_state *css)
137 {
138         return css ? container_of(css, struct cpuset, css) : NULL;
139 }
140 
141 /* Retrieve the cpuset for a task */
142 static inline struct cpuset *task_cs(struct task_struct *task)
143 {
144         return css_cs(task_css(task, cpuset_cgrp_id));
145 }
146 
147 static inline struct cpuset *parent_cs(struct cpuset *cs)
148 {
149         return css_cs(cs->css.parent);
150 }
151 
152 #ifdef CONFIG_NUMA
153 static inline bool task_has_mempolicy(struct task_struct *task)
154 {
155         return task->mempolicy;
156 }
157 #else
158 static inline bool task_has_mempolicy(struct task_struct *task)
159 {
160         return false;
161 }
162 #endif
163 
164 
165 /* bits in struct cpuset flags field */
166 typedef enum {
167         CS_ONLINE,
168         CS_CPU_EXCLUSIVE,
169         CS_MEM_EXCLUSIVE,
170         CS_MEM_HARDWALL,
171         CS_MEMORY_MIGRATE,
172         CS_SCHED_LOAD_BALANCE,
173         CS_SPREAD_PAGE,
174         CS_SPREAD_SLAB,
175 } cpuset_flagbits_t;
176 
177 /* convenient tests for these bits */
178 static inline bool is_cpuset_online(const struct cpuset *cs)
179 {
180         return test_bit(CS_ONLINE, &cs->flags);
181 }
182 
183 static inline int is_cpu_exclusive(const struct cpuset *cs)
184 {
185         return test_bit(CS_CPU_EXCLUSIVE, &cs->flags);
186 }
187 
188 static inline int is_mem_exclusive(const struct cpuset *cs)
189 {
190         return test_bit(CS_MEM_EXCLUSIVE, &cs->flags);
191 }
192 
193 static inline int is_mem_hardwall(const struct cpuset *cs)
194 {
195         return test_bit(CS_MEM_HARDWALL, &cs->flags);
196 }
197 
198 static inline int is_sched_load_balance(const struct cpuset *cs)
199 {
200         return test_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
201 }
202 
203 static inline int is_memory_migrate(const struct cpuset *cs)
204 {
205         return test_bit(CS_MEMORY_MIGRATE, &cs->flags);
206 }
207 
208 static inline int is_spread_page(const struct cpuset *cs)
209 {
210         return test_bit(CS_SPREAD_PAGE, &cs->flags);
211 }
212 
213 static inline int is_spread_slab(const struct cpuset *cs)
214 {
215         return test_bit(CS_SPREAD_SLAB, &cs->flags);
216 }
217 
218 static struct cpuset top_cpuset = {
219         .flags = ((1 << CS_ONLINE) | (1 << CS_CPU_EXCLUSIVE) |
220                   (1 << CS_MEM_EXCLUSIVE)),
221 };
222 
223 /**
224  * cpuset_for_each_child - traverse online children of a cpuset
225  * @child_cs: loop cursor pointing to the current child
226  * @pos_css: used for iteration
227  * @parent_cs: target cpuset to walk children of
228  *
229  * Walk @child_cs through the online children of @parent_cs.  Must be used
230  * with RCU read locked.
231  */
232 #define cpuset_for_each_child(child_cs, pos_css, parent_cs)             \
233         css_for_each_child((pos_css), &(parent_cs)->css)                \
234                 if (is_cpuset_online(((child_cs) = css_cs((pos_css)))))
235 
236 /**
237  * cpuset_for_each_descendant_pre - pre-order walk of a cpuset's descendants
238  * @des_cs: loop cursor pointing to the current descendant
239  * @pos_css: used for iteration
240  * @root_cs: target cpuset to walk ancestor of
241  *
242  * Walk @des_cs through the online descendants of @root_cs.  Must be used
243  * with RCU read locked.  The caller may modify @pos_css by calling
244  * css_rightmost_descendant() to skip subtree.  @root_cs is included in the
245  * iteration and the first node to be visited.
246  */
247 #define cpuset_for_each_descendant_pre(des_cs, pos_css, root_cs)        \
248         css_for_each_descendant_pre((pos_css), &(root_cs)->css)         \
249                 if (is_cpuset_online(((des_cs) = css_cs((pos_css)))))
250 
251 /*
252  * There are two global locks guarding cpuset structures - cpuset_mutex and
253  * callback_lock. We also require taking task_lock() when dereferencing a
254  * task's cpuset pointer. See "The task_lock() exception", at the end of this
255  * comment.
256  *
257  * A task must hold both locks to modify cpusets.  If a task holds
258  * cpuset_mutex, then it blocks others wanting that mutex, ensuring that it
259  * is the only task able to also acquire callback_lock and be able to
260  * modify cpusets.  It can perform various checks on the cpuset structure
261  * first, knowing nothing will change.  It can also allocate memory while
262  * just holding cpuset_mutex.  While it is performing these checks, various
263  * callback routines can briefly acquire callback_lock to query cpusets.
264  * Once it is ready to make the changes, it takes callback_lock, blocking
265  * everyone else.
266  *
267  * Calls to the kernel memory allocator can not be made while holding
268  * callback_lock, as that would risk double tripping on callback_lock
269  * from one of the callbacks into the cpuset code from within
270  * __alloc_pages().
271  *
272  * If a task is only holding callback_lock, then it has read-only
273  * access to cpusets.
274  *
275  * Now, the task_struct fields mems_allowed and mempolicy may be changed
276  * by other task, we use alloc_lock in the task_struct fields to protect
277  * them.
278  *
279  * The cpuset_common_file_read() handlers only hold callback_lock across
280  * small pieces of code, such as when reading out possibly multi-word
281  * cpumasks and nodemasks.
282  *
283  * Accessing a task's cpuset should be done in accordance with the
284  * guidelines for accessing subsystem state in kernel/cgroup.c
285  */
286 
287 static DEFINE_MUTEX(cpuset_mutex);
288 static DEFINE_SPINLOCK(callback_lock);
289 
290 static struct workqueue_struct *cpuset_migrate_mm_wq;
291 
292 /*
293  * CPU / memory hotplug is handled asynchronously.
294  */
295 static void cpuset_hotplug_workfn(struct work_struct *work);
296 static DECLARE_WORK(cpuset_hotplug_work, cpuset_hotplug_workfn);
297 
298 static DECLARE_WAIT_QUEUE_HEAD(cpuset_attach_wq);
299 
300 /*
301  * This is ugly, but preserves the userspace API for existing cpuset
302  * users. If someone tries to mount the "cpuset" filesystem, we
303  * silently switch it to mount "cgroup" instead
304  */
305 static struct dentry *cpuset_mount(struct file_system_type *fs_type,
306                          int flags, const char *unused_dev_name, void *data)
307 {
308         struct file_system_type *cgroup_fs = get_fs_type("cgroup");
309         struct dentry *ret = ERR_PTR(-ENODEV);
310         if (cgroup_fs) {
311                 char mountopts[] =
312                         "cpuset,noprefix,"
313                         "release_agent=/sbin/cpuset_release_agent";
314                 ret = cgroup_fs->mount(cgroup_fs, flags,
315                                            unused_dev_name, mountopts);
316                 put_filesystem(cgroup_fs);
317         }
318         return ret;
319 }
320 
321 static struct file_system_type cpuset_fs_type = {
322         .name = "cpuset",
323         .mount = cpuset_mount,
324 };
325 
326 /*
327  * Return in pmask the portion of a cpusets's cpus_allowed that
328  * are online.  If none are online, walk up the cpuset hierarchy
329  * until we find one that does have some online cpus.  The top
330  * cpuset always has some cpus online.
331  *
332  * One way or another, we guarantee to return some non-empty subset
333  * of cpu_online_mask.
334  *
335  * Call with callback_lock or cpuset_mutex held.
336  */
337 static void guarantee_online_cpus(struct cpuset *cs, struct cpumask *pmask)
338 {
339         while (!cpumask_intersects(cs->effective_cpus, cpu_online_mask))
340                 cs = parent_cs(cs);
341         cpumask_and(pmask, cs->effective_cpus, cpu_online_mask);
342 }
343 
344 /*
345  * Return in *pmask the portion of a cpusets's mems_allowed that
346  * are online, with memory.  If none are online with memory, walk
347  * up the cpuset hierarchy until we find one that does have some
348  * online mems.  The top cpuset always has some mems online.
349  *
350  * One way or another, we guarantee to return some non-empty subset
351  * of node_states[N_MEMORY].
352  *
353  * Call with callback_lock or cpuset_mutex held.
354  */
355 static void guarantee_online_mems(struct cpuset *cs, nodemask_t *pmask)
356 {
357         while (!nodes_intersects(cs->effective_mems, node_states[N_MEMORY]))
358                 cs = parent_cs(cs);
359         nodes_and(*pmask, cs->effective_mems, node_states[N_MEMORY]);
360 }
361 
362 /*
363  * update task's spread flag if cpuset's page/slab spread flag is set
364  *
365  * Call with callback_lock or cpuset_mutex held.
366  */
367 static void cpuset_update_task_spread_flag(struct cpuset *cs,
368                                         struct task_struct *tsk)
369 {
370         if (is_spread_page(cs))
371                 task_set_spread_page(tsk);
372         else
373                 task_clear_spread_page(tsk);
374 
375         if (is_spread_slab(cs))
376                 task_set_spread_slab(tsk);
377         else
378                 task_clear_spread_slab(tsk);
379 }
380 
381 /*
382  * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
383  *
384  * One cpuset is a subset of another if all its allowed CPUs and
385  * Memory Nodes are a subset of the other, and its exclusive flags
386  * are only set if the other's are set.  Call holding cpuset_mutex.
387  */
388 
389 static int is_cpuset_subset(const struct cpuset *p, const struct cpuset *q)
390 {
391         return  cpumask_subset(p->cpus_allowed, q->cpus_allowed) &&
392                 nodes_subset(p->mems_allowed, q->mems_allowed) &&
393                 is_cpu_exclusive(p) <= is_cpu_exclusive(q) &&
394                 is_mem_exclusive(p) <= is_mem_exclusive(q);
395 }
396 
397 /**
398  * alloc_trial_cpuset - allocate a trial cpuset
399  * @cs: the cpuset that the trial cpuset duplicates
400  */
401 static struct cpuset *alloc_trial_cpuset(struct cpuset *cs)
402 {
403         struct cpuset *trial;
404 
405         trial = kmemdup(cs, sizeof(*cs), GFP_KERNEL);
406         if (!trial)
407                 return NULL;
408 
409         if (!alloc_cpumask_var(&trial->cpus_allowed, GFP_KERNEL))
410                 goto free_cs;
411         if (!alloc_cpumask_var(&trial->effective_cpus, GFP_KERNEL))
412                 goto free_cpus;
413 
414         cpumask_copy(trial->cpus_allowed, cs->cpus_allowed);
415         cpumask_copy(trial->effective_cpus, cs->effective_cpus);
416         return trial;
417 
418 free_cpus:
419         free_cpumask_var(trial->cpus_allowed);
420 free_cs:
421         kfree(trial);
422         return NULL;
423 }
424 
425 /**
426  * free_trial_cpuset - free the trial cpuset
427  * @trial: the trial cpuset to be freed
428  */
429 static void free_trial_cpuset(struct cpuset *trial)
430 {
431         free_cpumask_var(trial->effective_cpus);
432         free_cpumask_var(trial->cpus_allowed);
433         kfree(trial);
434 }
435 
436 /*
437  * validate_change() - Used to validate that any proposed cpuset change
438  *                     follows the structural rules for cpusets.
439  *
440  * If we replaced the flag and mask values of the current cpuset
441  * (cur) with those values in the trial cpuset (trial), would
442  * our various subset and exclusive rules still be valid?  Presumes
443  * cpuset_mutex held.
444  *
445  * 'cur' is the address of an actual, in-use cpuset.  Operations
446  * such as list traversal that depend on the actual address of the
447  * cpuset in the list must use cur below, not trial.
448  *
449  * 'trial' is the address of bulk structure copy of cur, with
450  * perhaps one or more of the fields cpus_allowed, mems_allowed,
451  * or flags changed to new, trial values.
452  *
453  * Return 0 if valid, -errno if not.
454  */
455 
456 static int validate_change(struct cpuset *cur, struct cpuset *trial)
457 {
458         struct cgroup_subsys_state *css;
459         struct cpuset *c, *par;
460         int ret;
461 
462         rcu_read_lock();
463 
464         /* Each of our child cpusets must be a subset of us */
465         ret = -EBUSY;
466         cpuset_for_each_child(c, css, cur)
467                 if (!is_cpuset_subset(c, trial))
468                         goto out;
469 
470         /* Remaining checks don't apply to root cpuset */
471         ret = 0;
472         if (cur == &top_cpuset)
473                 goto out;
474 
475         par = parent_cs(cur);
476 
477         /* On legacy hiearchy, we must be a subset of our parent cpuset. */
478         ret = -EACCES;
479         if (!cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
480             !is_cpuset_subset(trial, par))
481                 goto out;
482 
483         /*
484          * If either I or some sibling (!= me) is exclusive, we can't
485          * overlap
486          */
487         ret = -EINVAL;
488         cpuset_for_each_child(c, css, par) {
489                 if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
490                     c != cur &&
491                     cpumask_intersects(trial->cpus_allowed, c->cpus_allowed))
492                         goto out;
493                 if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) &&
494                     c != cur &&
495                     nodes_intersects(trial->mems_allowed, c->mems_allowed))
496                         goto out;
497         }
498 
499         /*
500          * Cpusets with tasks - existing or newly being attached - can't
501          * be changed to have empty cpus_allowed or mems_allowed.
502          */
503         ret = -ENOSPC;
504         if ((cgroup_is_populated(cur->css.cgroup) || cur->attach_in_progress)) {
505                 if (!cpumask_empty(cur->cpus_allowed) &&
506                     cpumask_empty(trial->cpus_allowed))
507                         goto out;
508                 if (!nodes_empty(cur->mems_allowed) &&
509                     nodes_empty(trial->mems_allowed))
510                         goto out;
511         }
512 
513         /*
514          * We can't shrink if we won't have enough room for SCHED_DEADLINE
515          * tasks.
516          */
517         ret = -EBUSY;
518         if (is_cpu_exclusive(cur) &&
519             !cpuset_cpumask_can_shrink(cur->cpus_allowed,
520                                        trial->cpus_allowed))
521                 goto out;
522 
523         ret = 0;
524 out:
525         rcu_read_unlock();
526         return ret;
527 }
528 
529 #ifdef CONFIG_SMP
530 /*
531  * Helper routine for generate_sched_domains().
532  * Do cpusets a, b have overlapping effective cpus_allowed masks?
533  */
534 static int cpusets_overlap(struct cpuset *a, struct cpuset *b)
535 {
536         return cpumask_intersects(a->effective_cpus, b->effective_cpus);
537 }
538 
539 static void
540 update_domain_attr(struct sched_domain_attr *dattr, struct cpuset *c)
541 {
542         if (dattr->relax_domain_level < c->relax_domain_level)
543                 dattr->relax_domain_level = c->relax_domain_level;
544         return;
545 }
546 
547 static void update_domain_attr_tree(struct sched_domain_attr *dattr,
548                                     struct cpuset *root_cs)
549 {
550         struct cpuset *cp;
551         struct cgroup_subsys_state *pos_css;
552 
553         rcu_read_lock();
554         cpuset_for_each_descendant_pre(cp, pos_css, root_cs) {
555                 /* skip the whole subtree if @cp doesn't have any CPU */
556                 if (cpumask_empty(cp->cpus_allowed)) {
557                         pos_css = css_rightmost_descendant(pos_css);
558                         continue;
559                 }
560 
561                 if (is_sched_load_balance(cp))
562                         update_domain_attr(dattr, cp);
563         }
564         rcu_read_unlock();
565 }
566 
567 /*
568  * generate_sched_domains()
569  *
570  * This function builds a partial partition of the systems CPUs
571  * A 'partial partition' is a set of non-overlapping subsets whose
572  * union is a subset of that set.
573  * The output of this function needs to be passed to kernel/sched/core.c
574  * partition_sched_domains() routine, which will rebuild the scheduler's
575  * load balancing domains (sched domains) as specified by that partial
576  * partition.
577  *
578  * See "What is sched_load_balance" in Documentation/cgroups/cpusets.txt
579  * for a background explanation of this.
580  *
581  * Does not return errors, on the theory that the callers of this
582  * routine would rather not worry about failures to rebuild sched
583  * domains when operating in the severe memory shortage situations
584  * that could cause allocation failures below.
585  *
586  * Must be called with cpuset_mutex held.
587  *
588  * The three key local variables below are:
589  *    q  - a linked-list queue of cpuset pointers, used to implement a
590  *         top-down scan of all cpusets.  This scan loads a pointer
591  *         to each cpuset marked is_sched_load_balance into the
592  *         array 'csa'.  For our purposes, rebuilding the schedulers
593  *         sched domains, we can ignore !is_sched_load_balance cpusets.
594  *  csa  - (for CpuSet Array) Array of pointers to all the cpusets
595  *         that need to be load balanced, for convenient iterative
596  *         access by the subsequent code that finds the best partition,
597  *         i.e the set of domains (subsets) of CPUs such that the
598  *         cpus_allowed of every cpuset marked is_sched_load_balance
599  *         is a subset of one of these domains, while there are as
600  *         many such domains as possible, each as small as possible.
601  * doms  - Conversion of 'csa' to an array of cpumasks, for passing to
602  *         the kernel/sched/core.c routine partition_sched_domains() in a
603  *         convenient format, that can be easily compared to the prior
604  *         value to determine what partition elements (sched domains)
605  *         were changed (added or removed.)
606  *
607  * Finding the best partition (set of domains):
608  *      The triple nested loops below over i, j, k scan over the
609  *      load balanced cpusets (using the array of cpuset pointers in
610  *      csa[]) looking for pairs of cpusets that have overlapping
611  *      cpus_allowed, but which don't have the same 'pn' partition
612  *      number and gives them in the same partition number.  It keeps
613  *      looping on the 'restart' label until it can no longer find
614  *      any such pairs.
615  *
616  *      The union of the cpus_allowed masks from the set of
617  *      all cpusets having the same 'pn' value then form the one
618  *      element of the partition (one sched domain) to be passed to
619  *      partition_sched_domains().
620  */
621 static int generate_sched_domains(cpumask_var_t **domains,
622                         struct sched_domain_attr **attributes)
623 {
624         struct cpuset *cp;      /* scans q */
625         struct cpuset **csa;    /* array of all cpuset ptrs */
626         int csn;                /* how many cpuset ptrs in csa so far */
627         int i, j, k;            /* indices for partition finding loops */
628         cpumask_var_t *doms;    /* resulting partition; i.e. sched domains */
629         cpumask_var_t non_isolated_cpus;  /* load balanced CPUs */
630         struct sched_domain_attr *dattr;  /* attributes for custom domains */
631         int ndoms = 0;          /* number of sched domains in result */
632         int nslot;              /* next empty doms[] struct cpumask slot */
633         struct cgroup_subsys_state *pos_css;
634 
635         doms = NULL;
636         dattr = NULL;
637         csa = NULL;
638 
639         if (!alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL))
640                 goto done;
641         cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
642 
643         /* Special case for the 99% of systems with one, full, sched domain */
644         if (is_sched_load_balance(&top_cpuset)) {
645                 ndoms = 1;
646                 doms = alloc_sched_domains(ndoms);
647                 if (!doms)
648                         goto done;
649 
650                 dattr = kmalloc(sizeof(struct sched_domain_attr), GFP_KERNEL);
651                 if (dattr) {
652                         *dattr = SD_ATTR_INIT;
653                         update_domain_attr_tree(dattr, &top_cpuset);
654                 }
655                 cpumask_and(doms[0], top_cpuset.effective_cpus,
656                                      non_isolated_cpus);
657 
658                 goto done;
659         }
660 
661         csa = kmalloc(nr_cpusets() * sizeof(cp), GFP_KERNEL);
662         if (!csa)
663                 goto done;
664         csn = 0;
665 
666         rcu_read_lock();
667         cpuset_for_each_descendant_pre(cp, pos_css, &top_cpuset) {
668                 if (cp == &top_cpuset)
669                         continue;
670                 /*
671                  * Continue traversing beyond @cp iff @cp has some CPUs and
672                  * isn't load balancing.  The former is obvious.  The
673                  * latter: All child cpusets contain a subset of the
674                  * parent's cpus, so just skip them, and then we call
675                  * update_domain_attr_tree() to calc relax_domain_level of
676                  * the corresponding sched domain.
677                  */
678                 if (!cpumask_empty(cp->cpus_allowed) &&
679                     !(is_sched_load_balance(cp) &&
680                       cpumask_intersects(cp->cpus_allowed, non_isolated_cpus)))
681                         continue;
682 
683                 if (is_sched_load_balance(cp))
684                         csa[csn++] = cp;
685 
686                 /* skip @cp's subtree */
687                 pos_css = css_rightmost_descendant(pos_css);
688         }
689         rcu_read_unlock();
690 
691         for (i = 0; i < csn; i++)
692                 csa[i]->pn = i;
693         ndoms = csn;
694 
695 restart:
696         /* Find the best partition (set of sched domains) */
697         for (i = 0; i < csn; i++) {
698                 struct cpuset *a = csa[i];
699                 int apn = a->pn;
700 
701                 for (j = 0; j < csn; j++) {
702                         struct cpuset *b = csa[j];
703                         int bpn = b->pn;
704 
705                         if (apn != bpn && cpusets_overlap(a, b)) {
706                                 for (k = 0; k < csn; k++) {
707                                         struct cpuset *c = csa[k];
708 
709                                         if (c->pn == bpn)
710                                                 c->pn = apn;
711                                 }
712                                 ndoms--;        /* one less element */
713                                 goto restart;
714                         }
715                 }
716         }
717 
718         /*
719          * Now we know how many domains to create.
720          * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
721          */
722         doms = alloc_sched_domains(ndoms);
723         if (!doms)
724                 goto done;
725 
726         /*
727          * The rest of the code, including the scheduler, can deal with
728          * dattr==NULL case. No need to abort if alloc fails.
729          */
730         dattr = kmalloc(ndoms * sizeof(struct sched_domain_attr), GFP_KERNEL);
731 
732         for (nslot = 0, i = 0; i < csn; i++) {
733                 struct cpuset *a = csa[i];
734                 struct cpumask *dp;
735                 int apn = a->pn;
736 
737                 if (apn < 0) {
738                         /* Skip completed partitions */
739                         continue;
740                 }
741 
742                 dp = doms[nslot];
743 
744                 if (nslot == ndoms) {
745                         static int warnings = 10;
746                         if (warnings) {
747                                 pr_warn("rebuild_sched_domains confused: nslot %d, ndoms %d, csn %d, i %d, apn %d\n",
748                                         nslot, ndoms, csn, i, apn);
749                                 warnings--;
750                         }
751                         continue;
752                 }
753 
754                 cpumask_clear(dp);
755                 if (dattr)
756                         *(dattr + nslot) = SD_ATTR_INIT;
757                 for (j = i; j < csn; j++) {
758                         struct cpuset *b = csa[j];
759 
760                         if (apn == b->pn) {
761                                 cpumask_or(dp, dp, b->effective_cpus);
762                                 cpumask_and(dp, dp, non_isolated_cpus);
763                                 if (dattr)
764                                         update_domain_attr_tree(dattr + nslot, b);
765 
766                                 /* Done with this partition */
767                                 b->pn = -1;
768                         }
769                 }
770                 nslot++;
771         }
772         BUG_ON(nslot != ndoms);
773 
774 done:
775         free_cpumask_var(non_isolated_cpus);
776         kfree(csa);
777 
778         /*
779          * Fallback to the default domain if kmalloc() failed.
780          * See comments in partition_sched_domains().
781          */
782         if (doms == NULL)
783                 ndoms = 1;
784 
785         *domains    = doms;
786         *attributes = dattr;
787         return ndoms;
788 }
789 
790 /*
791  * Rebuild scheduler domains.
792  *
793  * If the flag 'sched_load_balance' of any cpuset with non-empty
794  * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
795  * which has that flag enabled, or if any cpuset with a non-empty
796  * 'cpus' is removed, then call this routine to rebuild the
797  * scheduler's dynamic sched domains.
798  *
799  * Call with cpuset_mutex held.  Takes get_online_cpus().
800  */
801 static void rebuild_sched_domains_locked(void)
802 {
803         struct sched_domain_attr *attr;
804         cpumask_var_t *doms;
805         int ndoms;
806 
807         lockdep_assert_held(&cpuset_mutex);
808         get_online_cpus();
809 
810         /*
811          * We have raced with CPU hotplug. Don't do anything to avoid
812          * passing doms with offlined cpu to partition_sched_domains().
813          * Anyways, hotplug work item will rebuild sched domains.
814          */
815         if (!cpumask_equal(top_cpuset.effective_cpus, cpu_active_mask))
816                 goto out;
817 
818         /* Generate domain masks and attrs */
819         ndoms = generate_sched_domains(&doms, &attr);
820 
821         /* Have scheduler rebuild the domains */
822         partition_sched_domains(ndoms, doms, attr);
823 out:
824         put_online_cpus();
825 }
826 #else /* !CONFIG_SMP */
827 static void rebuild_sched_domains_locked(void)
828 {
829 }
830 #endif /* CONFIG_SMP */
831 
832 void rebuild_sched_domains(void)
833 {
834         mutex_lock(&cpuset_mutex);
835         rebuild_sched_domains_locked();
836         mutex_unlock(&cpuset_mutex);
837 }
838 
839 /**
840  * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
841  * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
842  *
843  * Iterate through each task of @cs updating its cpus_allowed to the
844  * effective cpuset's.  As this function is called with cpuset_mutex held,
845  * cpuset membership stays stable.
846  */
847 static void update_tasks_cpumask(struct cpuset *cs)
848 {
849         struct css_task_iter it;
850         struct task_struct *task;
851 
852         css_task_iter_start(&cs->css, &it);
853         while ((task = css_task_iter_next(&it)))
854                 set_cpus_allowed_ptr(task, cs->effective_cpus);
855         css_task_iter_end(&it);
856 }
857 
858 /*
859  * update_cpumasks_hier - Update effective cpumasks and tasks in the subtree
860  * @cs: the cpuset to consider
861  * @new_cpus: temp variable for calculating new effective_cpus
862  *
863  * When congifured cpumask is changed, the effective cpumasks of this cpuset
864  * and all its descendants need to be updated.
865  *
866  * On legacy hierachy, effective_cpus will be the same with cpu_allowed.
867  *
868  * Called with cpuset_mutex held
869  */
870 static void update_cpumasks_hier(struct cpuset *cs, struct cpumask *new_cpus)
871 {
872         struct cpuset *cp;
873         struct cgroup_subsys_state *pos_css;
874         bool need_rebuild_sched_domains = false;
875 
876         rcu_read_lock();
877         cpuset_for_each_descendant_pre(cp, pos_css, cs) {
878                 struct cpuset *parent = parent_cs(cp);
879 
880                 cpumask_and(new_cpus, cp->cpus_allowed, parent->effective_cpus);
881 
882                 /*
883                  * If it becomes empty, inherit the effective mask of the
884                  * parent, which is guaranteed to have some CPUs.
885                  */
886                 if (cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
887                     cpumask_empty(new_cpus))
888                         cpumask_copy(new_cpus, parent->effective_cpus);
889 
890                 /* Skip the whole subtree if the cpumask remains the same. */
891                 if (cpumask_equal(new_cpus, cp->effective_cpus)) {
892                         pos_css = css_rightmost_descendant(pos_css);
893                         continue;
894                 }
895 
896                 if (!css_tryget_online(&cp->css))
897                         continue;
898                 rcu_read_unlock();
899 
900                 spin_lock_irq(&callback_lock);
901                 cpumask_copy(cp->effective_cpus, new_cpus);
902                 spin_unlock_irq(&callback_lock);
903 
904                 WARN_ON(!cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
905                         !cpumask_equal(cp->cpus_allowed, cp->effective_cpus));
906 
907                 update_tasks_cpumask(cp);
908 
909                 /*
910                  * If the effective cpumask of any non-empty cpuset is changed,
911                  * we need to rebuild sched domains.
912                  */
913                 if (!cpumask_empty(cp->cpus_allowed) &&
914                     is_sched_load_balance(cp))
915                         need_rebuild_sched_domains = true;
916 
917                 rcu_read_lock();
918                 css_put(&cp->css);
919         }
920         rcu_read_unlock();
921 
922         if (need_rebuild_sched_domains)
923                 rebuild_sched_domains_locked();
924 }
925 
926 /**
927  * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
928  * @cs: the cpuset to consider
929  * @trialcs: trial cpuset
930  * @buf: buffer of cpu numbers written to this cpuset
931  */
932 static int update_cpumask(struct cpuset *cs, struct cpuset *trialcs,
933                           const char *buf)
934 {
935         int retval;
936 
937         /* top_cpuset.cpus_allowed tracks cpu_online_mask; it's read-only */
938         if (cs == &top_cpuset)
939                 return -EACCES;
940 
941         /*
942          * An empty cpus_allowed is ok only if the cpuset has no tasks.
943          * Since cpulist_parse() fails on an empty mask, we special case
944          * that parsing.  The validate_change() call ensures that cpusets
945          * with tasks have cpus.
946          */
947         if (!*buf) {
948                 cpumask_clear(trialcs->cpus_allowed);
949         } else {
950                 retval = cpulist_parse(buf, trialcs->cpus_allowed);
951                 if (retval < 0)
952                         return retval;
953 
954                 if (!cpumask_subset(trialcs->cpus_allowed,
955                                     top_cpuset.cpus_allowed))
956                         return -EINVAL;
957         }
958 
959         /* Nothing to do if the cpus didn't change */
960         if (cpumask_equal(cs->cpus_allowed, trialcs->cpus_allowed))
961                 return 0;
962 
963         retval = validate_change(cs, trialcs);
964         if (retval < 0)
965                 return retval;
966 
967         spin_lock_irq(&callback_lock);
968         cpumask_copy(cs->cpus_allowed, trialcs->cpus_allowed);
969         spin_unlock_irq(&callback_lock);
970 
971         /* use trialcs->cpus_allowed as a temp variable */
972         update_cpumasks_hier(cs, trialcs->cpus_allowed);
973         return 0;
974 }
975 
976 /*
977  * Migrate memory region from one set of nodes to another.  This is
978  * performed asynchronously as it can be called from process migration path
979  * holding locks involved in process management.  All mm migrations are
980  * performed in the queued order and can be waited for by flushing
981  * cpuset_migrate_mm_wq.
982  */
983 
984 struct cpuset_migrate_mm_work {
985         struct work_struct      work;
986         struct mm_struct        *mm;
987         nodemask_t              from;
988         nodemask_t              to;
989 };
990 
991 static void cpuset_migrate_mm_workfn(struct work_struct *work)
992 {
993         struct cpuset_migrate_mm_work *mwork =
994                 container_of(work, struct cpuset_migrate_mm_work, work);
995 
996         /* on a wq worker, no need to worry about %current's mems_allowed */
997         do_migrate_pages(mwork->mm, &mwork->from, &mwork->to, MPOL_MF_MOVE_ALL);
998         mmput(mwork->mm);
999         kfree(mwork);
1000 }
1001 
1002 static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
1003                                                         const nodemask_t *to)
1004 {
1005         struct cpuset_migrate_mm_work *mwork;
1006 
1007         mwork = kzalloc(sizeof(*mwork), GFP_KERNEL);
1008         if (mwork) {
1009                 mwork->mm = mm;
1010                 mwork->from = *from;
1011                 mwork->to = *to;
1012                 INIT_WORK(&mwork->work, cpuset_migrate_mm_workfn);
1013                 queue_work(cpuset_migrate_mm_wq, &mwork->work);
1014         } else {
1015                 mmput(mm);
1016         }
1017 }
1018 
1019 void cpuset_post_attach_flush(void)
1020 {
1021         flush_workqueue(cpuset_migrate_mm_wq);
1022 }
1023 
1024 /*
1025  * cpuset_change_task_nodemask - change task's mems_allowed and mempolicy
1026  * @tsk: the task to change
1027  * @newmems: new nodes that the task will be set
1028  *
1029  * In order to avoid seeing no nodes if the old and new nodes are disjoint,
1030  * we structure updates as setting all new allowed nodes, then clearing newly
1031  * disallowed ones.
1032  */
1033 static void cpuset_change_task_nodemask(struct task_struct *tsk,
1034                                         nodemask_t *newmems)
1035 {
1036         bool need_loop;
1037 
1038         /*
1039          * Allow tasks that have access to memory reserves because they have
1040          * been OOM killed to get memory anywhere.
1041          */
1042         if (unlikely(test_thread_flag(TIF_MEMDIE)))
1043                 return;
1044         if (current->flags & PF_EXITING) /* Let dying task have memory */
1045                 return;
1046 
1047         task_lock(tsk);
1048         /*
1049          * Determine if a loop is necessary if another thread is doing
1050          * read_mems_allowed_begin().  If at least one node remains unchanged and
1051          * tsk does not have a mempolicy, then an empty nodemask will not be
1052          * possible when mems_allowed is larger than a word.
1053          */
1054         need_loop = task_has_mempolicy(tsk) ||
1055                         !nodes_intersects(*newmems, tsk->mems_allowed);
1056 
1057         if (need_loop) {
1058                 local_irq_disable();
1059                 write_seqcount_begin(&tsk->mems_allowed_seq);
1060         }
1061 
1062         nodes_or(tsk->mems_allowed, tsk->mems_allowed, *newmems);
1063         mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP1);
1064 
1065         mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP2);
1066         tsk->mems_allowed = *newmems;
1067 
1068         if (need_loop) {
1069                 write_seqcount_end(&tsk->mems_allowed_seq);
1070                 local_irq_enable();
1071         }
1072 
1073         task_unlock(tsk);
1074 }
1075 
1076 static void *cpuset_being_rebound;
1077 
1078 /**
1079  * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
1080  * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
1081  *
1082  * Iterate through each task of @cs updating its mems_allowed to the
1083  * effective cpuset's.  As this function is called with cpuset_mutex held,
1084  * cpuset membership stays stable.
1085  */
1086 static void update_tasks_nodemask(struct cpuset *cs)
1087 {
1088         static nodemask_t newmems;      /* protected by cpuset_mutex */
1089         struct css_task_iter it;
1090         struct task_struct *task;
1091 
1092         cpuset_being_rebound = cs;              /* causes mpol_dup() rebind */
1093 
1094         guarantee_online_mems(cs, &newmems);
1095 
1096         /*
1097          * The mpol_rebind_mm() call takes mmap_sem, which we couldn't
1098          * take while holding tasklist_lock.  Forks can happen - the
1099          * mpol_dup() cpuset_being_rebound check will catch such forks,
1100          * and rebind their vma mempolicies too.  Because we still hold
1101          * the global cpuset_mutex, we know that no other rebind effort
1102          * will be contending for the global variable cpuset_being_rebound.
1103          * It's ok if we rebind the same mm twice; mpol_rebind_mm()
1104          * is idempotent.  Also migrate pages in each mm to new nodes.
1105          */
1106         css_task_iter_start(&cs->css, &it);
1107         while ((task = css_task_iter_next(&it))) {
1108                 struct mm_struct *mm;
1109                 bool migrate;
1110 
1111                 cpuset_change_task_nodemask(task, &newmems);
1112 
1113                 mm = get_task_mm(task);
1114                 if (!mm)
1115                         continue;
1116 
1117                 migrate = is_memory_migrate(cs);
1118 
1119                 mpol_rebind_mm(mm, &cs->mems_allowed);
1120                 if (migrate)
1121                         cpuset_migrate_mm(mm, &cs->old_mems_allowed, &newmems);
1122                 else
1123                         mmput(mm);
1124         }
1125         css_task_iter_end(&it);
1126 
1127         /*
1128          * All the tasks' nodemasks have been updated, update
1129          * cs->old_mems_allowed.
1130          */
1131         cs->old_mems_allowed = newmems;
1132 
1133         /* We're done rebinding vmas to this cpuset's new mems_allowed. */
1134         cpuset_being_rebound = NULL;
1135 }
1136 
1137 /*
1138  * update_nodemasks_hier - Update effective nodemasks and tasks in the subtree
1139  * @cs: the cpuset to consider
1140  * @new_mems: a temp variable for calculating new effective_mems
1141  *
1142  * When configured nodemask is changed, the effective nodemasks of this cpuset
1143  * and all its descendants need to be updated.
1144  *
1145  * On legacy hiearchy, effective_mems will be the same with mems_allowed.
1146  *
1147  * Called with cpuset_mutex held
1148  */
1149 static void update_nodemasks_hier(struct cpuset *cs, nodemask_t *new_mems)
1150 {
1151         struct cpuset *cp;
1152         struct cgroup_subsys_state *pos_css;
1153 
1154         rcu_read_lock();
1155         cpuset_for_each_descendant_pre(cp, pos_css, cs) {
1156                 struct cpuset *parent = parent_cs(cp);
1157 
1158                 nodes_and(*new_mems, cp->mems_allowed, parent->effective_mems);
1159 
1160                 /*
1161                  * If it becomes empty, inherit the effective mask of the
1162                  * parent, which is guaranteed to have some MEMs.
1163                  */
1164                 if (cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
1165                     nodes_empty(*new_mems))
1166                         *new_mems = parent->effective_mems;
1167 
1168                 /* Skip the whole subtree if the nodemask remains the same. */
1169                 if (nodes_equal(*new_mems, cp->effective_mems)) {
1170                         pos_css = css_rightmost_descendant(pos_css);
1171                         continue;
1172                 }
1173 
1174                 if (!css_tryget_online(&cp->css))
1175                         continue;
1176                 rcu_read_unlock();
1177 
1178                 spin_lock_irq(&callback_lock);
1179                 cp->effective_mems = *new_mems;
1180                 spin_unlock_irq(&callback_lock);
1181 
1182                 WARN_ON(!cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
1183                         !nodes_equal(cp->mems_allowed, cp->effective_mems));
1184 
1185                 update_tasks_nodemask(cp);
1186 
1187                 rcu_read_lock();
1188                 css_put(&cp->css);
1189         }
1190         rcu_read_unlock();
1191 }
1192 
1193 /*
1194  * Handle user request to change the 'mems' memory placement
1195  * of a cpuset.  Needs to validate the request, update the
1196  * cpusets mems_allowed, and for each task in the cpuset,
1197  * update mems_allowed and rebind task's mempolicy and any vma
1198  * mempolicies and if the cpuset is marked 'memory_migrate',
1199  * migrate the tasks pages to the new memory.
1200  *
1201  * Call with cpuset_mutex held. May take callback_lock during call.
1202  * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
1203  * lock each such tasks mm->mmap_sem, scan its vma's and rebind
1204  * their mempolicies to the cpusets new mems_allowed.
1205  */
1206 static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
1207                            const char *buf)
1208 {
1209         int retval;
1210 
1211         /*
1212          * top_cpuset.mems_allowed tracks node_stats[N_MEMORY];
1213          * it's read-only
1214          */
1215         if (cs == &top_cpuset) {
1216                 retval = -EACCES;
1217                 goto done;
1218         }
1219 
1220         /*
1221          * An empty mems_allowed is ok iff there are no tasks in the cpuset.
1222          * Since nodelist_parse() fails on an empty mask, we special case
1223          * that parsing.  The validate_change() call ensures that cpusets
1224          * with tasks have memory.
1225          */
1226         if (!*buf) {
1227                 nodes_clear(trialcs->mems_allowed);
1228         } else {
1229                 retval = nodelist_parse(buf, trialcs->mems_allowed);
1230                 if (retval < 0)
1231                         goto done;
1232 
1233                 if (!nodes_subset(trialcs->mems_allowed,
1234                                   top_cpuset.mems_allowed)) {
1235                         retval = -EINVAL;
1236                         goto done;
1237                 }
1238         }
1239 
1240         if (nodes_equal(cs->mems_allowed, trialcs->mems_allowed)) {
1241                 retval = 0;             /* Too easy - nothing to do */
1242                 goto done;
1243         }
1244         retval = validate_change(cs, trialcs);
1245         if (retval < 0)
1246                 goto done;
1247 
1248         spin_lock_irq(&callback_lock);
1249         cs->mems_allowed = trialcs->mems_allowed;
1250         spin_unlock_irq(&callback_lock);
1251 
1252         /* use trialcs->mems_allowed as a temp variable */
1253         update_nodemasks_hier(cs, &trialcs->mems_allowed);
1254 done:
1255         return retval;
1256 }
1257 
1258 int current_cpuset_is_being_rebound(void)
1259 {
1260         int ret;
1261 
1262         rcu_read_lock();
1263         ret = task_cs(current) == cpuset_being_rebound;
1264         rcu_read_unlock();
1265 
1266         return ret;
1267 }
1268 
1269 static int update_relax_domain_level(struct cpuset *cs, s64 val)
1270 {
1271 #ifdef CONFIG_SMP
1272         if (val < -1 || val >= sched_domain_level_max)
1273                 return -EINVAL;
1274 #endif
1275 
1276         if (val != cs->relax_domain_level) {
1277                 cs->relax_domain_level = val;
1278                 if (!cpumask_empty(cs->cpus_allowed) &&
1279                     is_sched_load_balance(cs))
1280                         rebuild_sched_domains_locked();
1281         }
1282 
1283         return 0;
1284 }
1285 
1286 /**
1287  * update_tasks_flags - update the spread flags of tasks in the cpuset.
1288  * @cs: the cpuset in which each task's spread flags needs to be changed
1289  *
1290  * Iterate through each task of @cs updating its spread flags.  As this
1291  * function is called with cpuset_mutex held, cpuset membership stays
1292  * stable.
1293  */
1294 static void update_tasks_flags(struct cpuset *cs)
1295 {
1296         struct css_task_iter it;
1297         struct task_struct *task;
1298 
1299         css_task_iter_start(&cs->css, &it);
1300         while ((task = css_task_iter_next(&it)))
1301                 cpuset_update_task_spread_flag(cs, task);
1302         css_task_iter_end(&it);
1303 }
1304 
1305 /*
1306  * update_flag - read a 0 or a 1 in a file and update associated flag
1307  * bit:         the bit to update (see cpuset_flagbits_t)
1308  * cs:          the cpuset to update
1309  * turning_on:  whether the flag is being set or cleared
1310  *
1311  * Call with cpuset_mutex held.
1312  */
1313 
1314 static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs,
1315                        int turning_on)
1316 {
1317         struct cpuset *trialcs;
1318         int balance_flag_changed;
1319         int spread_flag_changed;
1320         int err;
1321 
1322         trialcs = alloc_trial_cpuset(cs);
1323         if (!trialcs)
1324                 return -ENOMEM;
1325 
1326         if (turning_on)
1327                 set_bit(bit, &trialcs->flags);
1328         else
1329                 clear_bit(bit, &trialcs->flags);
1330 
1331         err = validate_change(cs, trialcs);
1332         if (err < 0)
1333                 goto out;
1334 
1335         balance_flag_changed = (is_sched_load_balance(cs) !=
1336                                 is_sched_load_balance(trialcs));
1337 
1338         spread_flag_changed = ((is_spread_slab(cs) != is_spread_slab(trialcs))
1339                         || (is_spread_page(cs) != is_spread_page(trialcs)));
1340 
1341         spin_lock_irq(&callback_lock);
1342         cs->flags = trialcs->flags;
1343         spin_unlock_irq(&callback_lock);
1344 
1345         if (!cpumask_empty(trialcs->cpus_allowed) && balance_flag_changed)
1346                 rebuild_sched_domains_locked();
1347 
1348         if (spread_flag_changed)
1349                 update_tasks_flags(cs);
1350 out:
1351         free_trial_cpuset(trialcs);
1352         return err;
1353 }
1354 
1355 /*
1356  * Frequency meter - How fast is some event occurring?
1357  *
1358  * These routines manage a digitally filtered, constant time based,
1359  * event frequency meter.  There are four routines:
1360  *   fmeter_init() - initialize a frequency meter.
1361  *   fmeter_markevent() - called each time the event happens.
1362  *   fmeter_getrate() - returns the recent rate of such events.
1363  *   fmeter_update() - internal routine used to update fmeter.
1364  *
1365  * A common data structure is passed to each of these routines,
1366  * which is used to keep track of the state required to manage the
1367  * frequency meter and its digital filter.
1368  *
1369  * The filter works on the number of events marked per unit time.
1370  * The filter is single-pole low-pass recursive (IIR).  The time unit
1371  * is 1 second.  Arithmetic is done using 32-bit integers scaled to
1372  * simulate 3 decimal digits of precision (multiplied by 1000).
1373  *
1374  * With an FM_COEF of 933, and a time base of 1 second, the filter
1375  * has a half-life of 10 seconds, meaning that if the events quit
1376  * happening, then the rate returned from the fmeter_getrate()
1377  * will be cut in half each 10 seconds, until it converges to zero.
1378  *
1379  * It is not worth doing a real infinitely recursive filter.  If more
1380  * than FM_MAXTICKS ticks have elapsed since the last filter event,
1381  * just compute FM_MAXTICKS ticks worth, by which point the level
1382  * will be stable.
1383  *
1384  * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
1385  * arithmetic overflow in the fmeter_update() routine.
1386  *
1387  * Given the simple 32 bit integer arithmetic used, this meter works
1388  * best for reporting rates between one per millisecond (msec) and
1389  * one per 32 (approx) seconds.  At constant rates faster than one
1390  * per msec it maxes out at values just under 1,000,000.  At constant
1391  * rates between one per msec, and one per second it will stabilize
1392  * to a value N*1000, where N is the rate of events per second.
1393  * At constant rates between one per second and one per 32 seconds,
1394  * it will be choppy, moving up on the seconds that have an event,
1395  * and then decaying until the next event.  At rates slower than
1396  * about one in 32 seconds, it decays all the way back to zero between
1397  * each event.
1398  */
1399 
1400 #define FM_COEF 933             /* coefficient for half-life of 10 secs */
1401 #define FM_MAXTICKS ((u32)99)   /* useless computing more ticks than this */
1402 #define FM_MAXCNT 1000000       /* limit cnt to avoid overflow */
1403 #define FM_SCALE 1000           /* faux fixed point scale */
1404 
1405 /* Initialize a frequency meter */
1406 static void fmeter_init(struct fmeter *fmp)
1407 {
1408         fmp->cnt = 0;
1409         fmp->val = 0;
1410         fmp->time = 0;
1411         spin_lock_init(&fmp->lock);
1412 }
1413 
1414 /* Internal meter update - process cnt events and update value */
1415 static void fmeter_update(struct fmeter *fmp)
1416 {
1417         time64_t now;
1418         u32 ticks;
1419 
1420         now = ktime_get_seconds();
1421         ticks = now - fmp->time;
1422 
1423         if (ticks == 0)
1424                 return;
1425 
1426         ticks = min(FM_MAXTICKS, ticks);
1427         while (ticks-- > 0)
1428                 fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
1429         fmp->time = now;
1430 
1431         fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
1432         fmp->cnt = 0;
1433 }
1434 
1435 /* Process any previous ticks, then bump cnt by one (times scale). */
1436 static void fmeter_markevent(struct fmeter *fmp)
1437 {
1438         spin_lock(&fmp->lock);
1439         fmeter_update(fmp);
1440         fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
1441         spin_unlock(&fmp->lock);
1442 }
1443 
1444 /* Process any previous ticks, then return current value. */
1445 static int fmeter_getrate(struct fmeter *fmp)
1446 {
1447         int val;
1448 
1449         spin_lock(&fmp->lock);
1450         fmeter_update(fmp);
1451         val = fmp->val;
1452         spin_unlock(&fmp->lock);
1453         return val;
1454 }
1455 
1456 static struct cpuset *cpuset_attach_old_cs;
1457 
1458 /* Called by cgroups to determine if a cpuset is usable; cpuset_mutex held */
1459 static int cpuset_can_attach(struct cgroup_taskset *tset)
1460 {
1461         struct cgroup_subsys_state *css;
1462         struct cpuset *cs;
1463         struct task_struct *task;
1464         int ret;
1465 
1466         /* used later by cpuset_attach() */
1467         cpuset_attach_old_cs = task_cs(cgroup_taskset_first(tset, &css));
1468         cs = css_cs(css);
1469 
1470         mutex_lock(&cpuset_mutex);
1471 
1472         /* allow moving tasks into an empty cpuset if on default hierarchy */
1473         ret = -ENOSPC;
1474         if (!cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
1475             (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed)))
1476                 goto out_unlock;
1477 
1478         cgroup_taskset_for_each(task, css, tset) {
1479                 ret = task_can_attach(task, cs->cpus_allowed);
1480                 if (ret)
1481                         goto out_unlock;
1482                 ret = security_task_setscheduler(task);
1483                 if (ret)
1484                         goto out_unlock;
1485         }
1486 
1487         /*
1488          * Mark attach is in progress.  This makes validate_change() fail
1489          * changes which zero cpus/mems_allowed.
1490          */
1491         cs->attach_in_progress++;
1492         ret = 0;
1493 out_unlock:
1494         mutex_unlock(&cpuset_mutex);
1495         return ret;
1496 }
1497 
1498 static void cpuset_cancel_attach(struct cgroup_taskset *tset)
1499 {
1500         struct cgroup_subsys_state *css;
1501         struct cpuset *cs;
1502 
1503         cgroup_taskset_first(tset, &css);
1504         cs = css_cs(css);
1505 
1506         mutex_lock(&cpuset_mutex);
1507         css_cs(css)->attach_in_progress--;
1508         mutex_unlock(&cpuset_mutex);
1509 }
1510 
1511 /*
1512  * Protected by cpuset_mutex.  cpus_attach is used only by cpuset_attach()
1513  * but we can't allocate it dynamically there.  Define it global and
1514  * allocate from cpuset_init().
1515  */
1516 static cpumask_var_t cpus_attach;
1517 
1518 static void cpuset_attach(struct cgroup_taskset *tset)
1519 {
1520         /* static buf protected by cpuset_mutex */
1521         static nodemask_t cpuset_attach_nodemask_to;
1522         struct task_struct *task;
1523         struct task_struct *leader;
1524         struct cgroup_subsys_state *css;
1525         struct cpuset *cs;
1526         struct cpuset *oldcs = cpuset_attach_old_cs;
1527 
1528         cgroup_taskset_first(tset, &css);
1529         cs = css_cs(css);
1530 
1531         mutex_lock(&cpuset_mutex);
1532 
1533         /* prepare for attach */
1534         if (cs == &top_cpuset)
1535                 cpumask_copy(cpus_attach, cpu_possible_mask);
1536         else
1537                 guarantee_online_cpus(cs, cpus_attach);
1538 
1539         guarantee_online_mems(cs, &cpuset_attach_nodemask_to);
1540 
1541         cgroup_taskset_for_each(task, css, tset) {
1542                 /*
1543                  * can_attach beforehand should guarantee that this doesn't
1544                  * fail.  TODO: have a better way to handle failure here
1545                  */
1546                 WARN_ON_ONCE(set_cpus_allowed_ptr(task, cpus_attach));
1547 
1548                 cpuset_change_task_nodemask(task, &cpuset_attach_nodemask_to);
1549                 cpuset_update_task_spread_flag(cs, task);
1550         }
1551 
1552         /*
1553          * Change mm for all threadgroup leaders. This is expensive and may
1554          * sleep and should be moved outside migration path proper.
1555          */
1556         cpuset_attach_nodemask_to = cs->effective_mems;
1557         cgroup_taskset_for_each_leader(leader, css, tset) {
1558                 struct mm_struct *mm = get_task_mm(leader);
1559 
1560                 if (mm) {
1561                         mpol_rebind_mm(mm, &cpuset_attach_nodemask_to);
1562 
1563                         /*
1564                          * old_mems_allowed is the same with mems_allowed
1565                          * here, except if this task is being moved
1566                          * automatically due to hotplug.  In that case
1567                          * @mems_allowed has been updated and is empty, so
1568                          * @old_mems_allowed is the right nodesets that we
1569                          * migrate mm from.
1570                          */
1571                         if (is_memory_migrate(cs))
1572                                 cpuset_migrate_mm(mm, &oldcs->old_mems_allowed,
1573                                                   &cpuset_attach_nodemask_to);
1574                         else
1575                                 mmput(mm);
1576                 }
1577         }
1578 
1579         cs->old_mems_allowed = cpuset_attach_nodemask_to;
1580 
1581         cs->attach_in_progress--;
1582         if (!cs->attach_in_progress)
1583                 wake_up(&cpuset_attach_wq);
1584 
1585         mutex_unlock(&cpuset_mutex);
1586 }
1587 
1588 /* The various types of files and directories in a cpuset file system */
1589 
1590 typedef enum {
1591         FILE_MEMORY_MIGRATE,
1592         FILE_CPULIST,
1593         FILE_MEMLIST,
1594         FILE_EFFECTIVE_CPULIST,
1595         FILE_EFFECTIVE_MEMLIST,
1596         FILE_CPU_EXCLUSIVE,
1597         FILE_MEM_EXCLUSIVE,
1598         FILE_MEM_HARDWALL,
1599         FILE_SCHED_LOAD_BALANCE,
1600         FILE_SCHED_RELAX_DOMAIN_LEVEL,
1601         FILE_MEMORY_PRESSURE_ENABLED,
1602         FILE_MEMORY_PRESSURE,
1603         FILE_SPREAD_PAGE,
1604         FILE_SPREAD_SLAB,
1605 } cpuset_filetype_t;
1606 
1607 static int cpuset_write_u64(struct cgroup_subsys_state *css, struct cftype *cft,
1608                             u64 val)
1609 {
1610         struct cpuset *cs = css_cs(css);
1611         cpuset_filetype_t type = cft->private;
1612         int retval = 0;
1613 
1614         mutex_lock(&cpuset_mutex);
1615         if (!is_cpuset_online(cs)) {
1616                 retval = -ENODEV;
1617                 goto out_unlock;
1618         }
1619 
1620         switch (type) {
1621         case FILE_CPU_EXCLUSIVE:
1622                 retval = update_flag(CS_CPU_EXCLUSIVE, cs, val);
1623                 break;
1624         case FILE_MEM_EXCLUSIVE:
1625                 retval = update_flag(CS_MEM_EXCLUSIVE, cs, val);
1626                 break;
1627         case FILE_MEM_HARDWALL:
1628                 retval = update_flag(CS_MEM_HARDWALL, cs, val);
1629                 break;
1630         case FILE_SCHED_LOAD_BALANCE:
1631                 retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, val);
1632                 break;
1633         case FILE_MEMORY_MIGRATE:
1634                 retval = update_flag(CS_MEMORY_MIGRATE, cs, val);
1635                 break;
1636         case FILE_MEMORY_PRESSURE_ENABLED:
1637                 cpuset_memory_pressure_enabled = !!val;
1638                 break;
1639         case FILE_SPREAD_PAGE:
1640                 retval = update_flag(CS_SPREAD_PAGE, cs, val);
1641                 break;
1642         case FILE_SPREAD_SLAB:
1643                 retval = update_flag(CS_SPREAD_SLAB, cs, val);
1644                 break;
1645         default:
1646                 retval = -EINVAL;
1647                 break;
1648         }
1649 out_unlock:
1650         mutex_unlock(&cpuset_mutex);
1651         return retval;
1652 }
1653 
1654 static int cpuset_write_s64(struct cgroup_subsys_state *css, struct cftype *cft,
1655                             s64 val)
1656 {
1657         struct cpuset *cs = css_cs(css);
1658         cpuset_filetype_t type = cft->private;
1659         int retval = -ENODEV;
1660 
1661         mutex_lock(&cpuset_mutex);
1662         if (!is_cpuset_online(cs))
1663                 goto out_unlock;
1664 
1665         switch (type) {
1666         case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1667                 retval = update_relax_domain_level(cs, val);
1668                 break;
1669         default:
1670                 retval = -EINVAL;
1671                 break;
1672         }
1673 out_unlock:
1674         mutex_unlock(&cpuset_mutex);
1675         return retval;
1676 }
1677 
1678 /*
1679  * Common handling for a write to a "cpus" or "mems" file.
1680  */
1681 static ssize_t cpuset_write_resmask(struct kernfs_open_file *of,
1682                                     char *buf, size_t nbytes, loff_t off)
1683 {
1684         struct cpuset *cs = css_cs(of_css(of));
1685         struct cpuset *trialcs;
1686         int retval = -ENODEV;
1687 
1688         buf = strstrip(buf);
1689 
1690         /*
1691          * CPU or memory hotunplug may leave @cs w/o any execution
1692          * resources, in which case the hotplug code asynchronously updates
1693          * configuration and transfers all tasks to the nearest ancestor
1694          * which can execute.
1695          *
1696          * As writes to "cpus" or "mems" may restore @cs's execution
1697          * resources, wait for the previously scheduled operations before
1698          * proceeding, so that we don't end up keep removing tasks added
1699          * after execution capability is restored.
1700          *
1701          * cpuset_hotplug_work calls back into cgroup core via
1702          * cgroup_transfer_tasks() and waiting for it from a cgroupfs
1703          * operation like this one can lead to a deadlock through kernfs
1704          * active_ref protection.  Let's break the protection.  Losing the
1705          * protection is okay as we check whether @cs is online after
1706          * grabbing cpuset_mutex anyway.  This only happens on the legacy
1707          * hierarchies.
1708          */
1709         css_get(&cs->css);
1710         kernfs_break_active_protection(of->kn);
1711         flush_work(&cpuset_hotplug_work);
1712 
1713         mutex_lock(&cpuset_mutex);
1714         if (!is_cpuset_online(cs))
1715                 goto out_unlock;
1716 
1717         trialcs = alloc_trial_cpuset(cs);
1718         if (!trialcs) {
1719                 retval = -ENOMEM;
1720                 goto out_unlock;
1721         }
1722 
1723         switch (of_cft(of)->private) {
1724         case FILE_CPULIST:
1725                 retval = update_cpumask(cs, trialcs, buf);
1726                 break;
1727         case FILE_MEMLIST:
1728                 retval = update_nodemask(cs, trialcs, buf);
1729                 break;
1730         default:
1731                 retval = -EINVAL;
1732                 break;
1733         }
1734 
1735         free_trial_cpuset(trialcs);
1736 out_unlock:
1737         mutex_unlock(&cpuset_mutex);
1738         kernfs_unbreak_active_protection(of->kn);
1739         css_put(&cs->css);
1740         flush_workqueue(cpuset_migrate_mm_wq);
1741         return retval ?: nbytes;
1742 }
1743 
1744 /*
1745  * These ascii lists should be read in a single call, by using a user
1746  * buffer large enough to hold the entire map.  If read in smaller
1747  * chunks, there is no guarantee of atomicity.  Since the display format
1748  * used, list of ranges of sequential numbers, is variable length,
1749  * and since these maps can change value dynamically, one could read
1750  * gibberish by doing partial reads while a list was changing.
1751  */
1752 static int cpuset_common_seq_show(struct seq_file *sf, void *v)
1753 {
1754         struct cpuset *cs = css_cs(seq_css(sf));
1755         cpuset_filetype_t type = seq_cft(sf)->private;
1756         int ret = 0;
1757 
1758         spin_lock_irq(&callback_lock);
1759 
1760         switch (type) {
1761         case FILE_CPULIST:
1762                 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->cpus_allowed));
1763                 break;
1764         case FILE_MEMLIST:
1765                 seq_printf(sf, "%*pbl\n", nodemask_pr_args(&cs->mems_allowed));
1766                 break;
1767         case FILE_EFFECTIVE_CPULIST:
1768                 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->effective_cpus));
1769                 break;
1770         case FILE_EFFECTIVE_MEMLIST:
1771                 seq_printf(sf, "%*pbl\n", nodemask_pr_args(&cs->effective_mems));
1772                 break;
1773         default:
1774                 ret = -EINVAL;
1775         }
1776 
1777         spin_unlock_irq(&callback_lock);
1778         return ret;
1779 }
1780 
1781 static u64 cpuset_read_u64(struct cgroup_subsys_state *css, struct cftype *cft)
1782 {
1783         struct cpuset *cs = css_cs(css);
1784         cpuset_filetype_t type = cft->private;
1785         switch (type) {
1786         case FILE_CPU_EXCLUSIVE:
1787                 return is_cpu_exclusive(cs);
1788         case FILE_MEM_EXCLUSIVE:
1789                 return is_mem_exclusive(cs);
1790         case FILE_MEM_HARDWALL:
1791                 return is_mem_hardwall(cs);
1792         case FILE_SCHED_LOAD_BALANCE:
1793                 return is_sched_load_balance(cs);
1794         case FILE_MEMORY_MIGRATE:
1795                 return is_memory_migrate(cs);
1796         case FILE_MEMORY_PRESSURE_ENABLED:
1797                 return cpuset_memory_pressure_enabled;
1798         case FILE_MEMORY_PRESSURE:
1799                 return fmeter_getrate(&cs->fmeter);
1800         case FILE_SPREAD_PAGE:
1801                 return is_spread_page(cs);
1802         case FILE_SPREAD_SLAB:
1803                 return is_spread_slab(cs);
1804         default:
1805                 BUG();
1806         }
1807 
1808         /* Unreachable but makes gcc happy */
1809         return 0;
1810 }
1811 
1812 static s64 cpuset_read_s64(struct cgroup_subsys_state *css, struct cftype *cft)
1813 {
1814         struct cpuset *cs = css_cs(css);
1815         cpuset_filetype_t type = cft->private;
1816         switch (type) {
1817         case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1818                 return cs->relax_domain_level;
1819         default:
1820                 BUG();
1821         }
1822 
1823         /* Unrechable but makes gcc happy */
1824         return 0;
1825 }
1826 
1827 
1828 /*
1829  * for the common functions, 'private' gives the type of file
1830  */
1831 
1832 static struct cftype files[] = {
1833         {
1834                 .name = "cpus",
1835                 .seq_show = cpuset_common_seq_show,
1836                 .write = cpuset_write_resmask,
1837                 .max_write_len = (100U + 6 * NR_CPUS),
1838                 .private = FILE_CPULIST,
1839         },
1840 
1841         {
1842                 .name = "mems",
1843                 .seq_show = cpuset_common_seq_show,
1844                 .write = cpuset_write_resmask,
1845                 .max_write_len = (100U + 6 * MAX_NUMNODES),
1846                 .private = FILE_MEMLIST,
1847         },
1848 
1849         {
1850                 .name = "effective_cpus",
1851                 .seq_show = cpuset_common_seq_show,
1852                 .private = FILE_EFFECTIVE_CPULIST,
1853         },
1854 
1855         {
1856                 .name = "effective_mems",
1857                 .seq_show = cpuset_common_seq_show,
1858                 .private = FILE_EFFECTIVE_MEMLIST,
1859         },
1860 
1861         {
1862                 .name = "cpu_exclusive",
1863                 .read_u64 = cpuset_read_u64,
1864                 .write_u64 = cpuset_write_u64,
1865                 .private = FILE_CPU_EXCLUSIVE,
1866         },
1867 
1868         {
1869                 .name = "mem_exclusive",
1870                 .read_u64 = cpuset_read_u64,
1871                 .write_u64 = cpuset_write_u64,
1872                 .private = FILE_MEM_EXCLUSIVE,
1873         },
1874 
1875         {
1876                 .name = "mem_hardwall",
1877                 .read_u64 = cpuset_read_u64,
1878                 .write_u64 = cpuset_write_u64,
1879                 .private = FILE_MEM_HARDWALL,
1880         },
1881 
1882         {
1883                 .name = "sched_load_balance",
1884                 .read_u64 = cpuset_read_u64,
1885                 .write_u64 = cpuset_write_u64,
1886                 .private = FILE_SCHED_LOAD_BALANCE,
1887         },
1888 
1889         {
1890                 .name = "sched_relax_domain_level",
1891                 .read_s64 = cpuset_read_s64,
1892                 .write_s64 = cpuset_write_s64,
1893                 .private = FILE_SCHED_RELAX_DOMAIN_LEVEL,
1894         },
1895 
1896         {
1897                 .name = "memory_migrate",
1898                 .read_u64 = cpuset_read_u64,
1899                 .write_u64 = cpuset_write_u64,
1900                 .private = FILE_MEMORY_MIGRATE,
1901         },
1902 
1903         {
1904                 .name = "memory_pressure",
1905                 .read_u64 = cpuset_read_u64,
1906         },
1907 
1908         {
1909                 .name = "memory_spread_page",
1910                 .read_u64 = cpuset_read_u64,
1911                 .write_u64 = cpuset_write_u64,
1912                 .private = FILE_SPREAD_PAGE,
1913         },
1914 
1915         {
1916                 .name = "memory_spread_slab",
1917                 .read_u64 = cpuset_read_u64,
1918                 .write_u64 = cpuset_write_u64,
1919                 .private = FILE_SPREAD_SLAB,
1920         },
1921 
1922         {
1923                 .name = "memory_pressure_enabled",
1924                 .flags = CFTYPE_ONLY_ON_ROOT,
1925                 .read_u64 = cpuset_read_u64,
1926                 .write_u64 = cpuset_write_u64,
1927                 .private = FILE_MEMORY_PRESSURE_ENABLED,
1928         },
1929 
1930         { }     /* terminate */
1931 };
1932 
1933 /*
1934  *      cpuset_css_alloc - allocate a cpuset css
1935  *      cgrp:   control group that the new cpuset will be part of
1936  */
1937 
1938 static struct cgroup_subsys_state *
1939 cpuset_css_alloc(struct cgroup_subsys_state *parent_css)
1940 {
1941         struct cpuset *cs;
1942 
1943         if (!parent_css)
1944                 return &top_cpuset.css;
1945 
1946         cs = kzalloc(sizeof(*cs), GFP_KERNEL);
1947         if (!cs)
1948                 return ERR_PTR(-ENOMEM);
1949         if (!alloc_cpumask_var(&cs->cpus_allowed, GFP_KERNEL))
1950                 goto free_cs;
1951         if (!alloc_cpumask_var(&cs->effective_cpus, GFP_KERNEL))
1952                 goto free_cpus;
1953 
1954         set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
1955         cpumask_clear(cs->cpus_allowed);
1956         nodes_clear(cs->mems_allowed);
1957         cpumask_clear(cs->effective_cpus);
1958         nodes_clear(cs->effective_mems);
1959         fmeter_init(&cs->fmeter);
1960         cs->relax_domain_level = -1;
1961 
1962         return &cs->css;
1963 
1964 free_cpus:
1965         free_cpumask_var(cs->cpus_allowed);
1966 free_cs:
1967         kfree(cs);
1968         return ERR_PTR(-ENOMEM);
1969 }
1970 
1971 static int cpuset_css_online(struct cgroup_subsys_state *css)
1972 {
1973         struct cpuset *cs = css_cs(css);
1974         struct cpuset *parent = parent_cs(cs);
1975         struct cpuset *tmp_cs;
1976         struct cgroup_subsys_state *pos_css;
1977 
1978         if (!parent)
1979                 return 0;
1980 
1981         mutex_lock(&cpuset_mutex);
1982 
1983         set_bit(CS_ONLINE, &cs->flags);
1984         if (is_spread_page(parent))
1985                 set_bit(CS_SPREAD_PAGE, &cs->flags);
1986         if (is_spread_slab(parent))
1987                 set_bit(CS_SPREAD_SLAB, &cs->flags);
1988 
1989         cpuset_inc();
1990 
1991         spin_lock_irq(&callback_lock);
1992         if (cgroup_subsys_on_dfl(cpuset_cgrp_subsys)) {
1993                 cpumask_copy(cs->effective_cpus, parent->effective_cpus);
1994                 cs->effective_mems = parent->effective_mems;
1995         }
1996         spin_unlock_irq(&callback_lock);
1997 
1998         if (!test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags))
1999                 goto out_unlock;
2000 
2001         /*
2002          * Clone @parent's configuration if CGRP_CPUSET_CLONE_CHILDREN is
2003          * set.  This flag handling is implemented in cgroup core for
2004          * histrical reasons - the flag may be specified during mount.
2005          *
2006          * Currently, if any sibling cpusets have exclusive cpus or mem, we
2007          * refuse to clone the configuration - thereby refusing the task to
2008          * be entered, and as a result refusing the sys_unshare() or
2009          * clone() which initiated it.  If this becomes a problem for some
2010          * users who wish to allow that scenario, then this could be
2011          * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
2012          * (and likewise for mems) to the new cgroup.
2013          */
2014         rcu_read_lock();
2015         cpuset_for_each_child(tmp_cs, pos_css, parent) {
2016                 if (is_mem_exclusive(tmp_cs) || is_cpu_exclusive(tmp_cs)) {
2017                         rcu_read_unlock();
2018                         goto out_unlock;
2019                 }
2020         }
2021         rcu_read_unlock();
2022 
2023         spin_lock_irq(&callback_lock);
2024         cs->mems_allowed = parent->mems_allowed;
2025         cs->effective_mems = parent->mems_allowed;
2026         cpumask_copy(cs->cpus_allowed, parent->cpus_allowed);
2027         cpumask_copy(cs->effective_cpus, parent->cpus_allowed);
2028         spin_unlock_irq(&callback_lock);
2029 out_unlock:
2030         mutex_unlock(&cpuset_mutex);
2031         return 0;
2032 }
2033 
2034 /*
2035  * If the cpuset being removed has its flag 'sched_load_balance'
2036  * enabled, then simulate turning sched_load_balance off, which
2037  * will call rebuild_sched_domains_locked().
2038  */
2039 
2040 static void cpuset_css_offline(struct cgroup_subsys_state *css)
2041 {
2042         struct cpuset *cs = css_cs(css);
2043 
2044         mutex_lock(&cpuset_mutex);
2045 
2046         if (is_sched_load_balance(cs))
2047                 update_flag(CS_SCHED_LOAD_BALANCE, cs, 0);
2048 
2049         cpuset_dec();
2050         clear_bit(CS_ONLINE, &cs->flags);
2051 
2052         mutex_unlock(&cpuset_mutex);
2053 }
2054 
2055 static void cpuset_css_free(struct cgroup_subsys_state *css)
2056 {
2057         struct cpuset *cs = css_cs(css);
2058 
2059         free_cpumask_var(cs->effective_cpus);
2060         free_cpumask_var(cs->cpus_allowed);
2061         kfree(cs);
2062 }
2063 
2064 static void cpuset_bind(struct cgroup_subsys_state *root_css)
2065 {
2066         mutex_lock(&cpuset_mutex);
2067         spin_lock_irq(&callback_lock);
2068 
2069         if (cgroup_subsys_on_dfl(cpuset_cgrp_subsys)) {
2070                 cpumask_copy(top_cpuset.cpus_allowed, cpu_possible_mask);
2071                 top_cpuset.mems_allowed = node_possible_map;
2072         } else {
2073                 cpumask_copy(top_cpuset.cpus_allowed,
2074                              top_cpuset.effective_cpus);
2075                 top_cpuset.mems_allowed = top_cpuset.effective_mems;
2076         }
2077 
2078         spin_unlock_irq(&callback_lock);
2079         mutex_unlock(&cpuset_mutex);
2080 }
2081 
2082 struct cgroup_subsys cpuset_cgrp_subsys = {
2083         .css_alloc      = cpuset_css_alloc,
2084         .css_online     = cpuset_css_online,
2085         .css_offline    = cpuset_css_offline,
2086         .css_free       = cpuset_css_free,
2087         .can_attach     = cpuset_can_attach,
2088         .cancel_attach  = cpuset_cancel_attach,
2089         .attach         = cpuset_attach,
2090         .bind           = cpuset_bind,
2091         .legacy_cftypes = files,
2092         .early_init     = 1,
2093 };
2094 
2095 /**
2096  * cpuset_init - initialize cpusets at system boot
2097  *
2098  * Description: Initialize top_cpuset and the cpuset internal file system,
2099  **/
2100 
2101 int __init cpuset_init(void)
2102 {
2103         int err = 0;
2104 
2105         if (!alloc_cpumask_var(&top_cpuset.cpus_allowed, GFP_KERNEL))
2106                 BUG();
2107         if (!alloc_cpumask_var(&top_cpuset.effective_cpus, GFP_KERNEL))
2108                 BUG();
2109 
2110         cpumask_setall(top_cpuset.cpus_allowed);
2111         nodes_setall(top_cpuset.mems_allowed);
2112         cpumask_setall(top_cpuset.effective_cpus);
2113         nodes_setall(top_cpuset.effective_mems);
2114 
2115         fmeter_init(&top_cpuset.fmeter);
2116         set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags);
2117         top_cpuset.relax_domain_level = -1;
2118 
2119         err = register_filesystem(&cpuset_fs_type);
2120         if (err < 0)
2121                 return err;
2122 
2123         if (!alloc_cpumask_var(&cpus_attach, GFP_KERNEL))
2124                 BUG();
2125 
2126         return 0;
2127 }
2128 
2129 /*
2130  * If CPU and/or memory hotplug handlers, below, unplug any CPUs
2131  * or memory nodes, we need to walk over the cpuset hierarchy,
2132  * removing that CPU or node from all cpusets.  If this removes the
2133  * last CPU or node from a cpuset, then move the tasks in the empty
2134  * cpuset to its next-highest non-empty parent.
2135  */
2136 static void remove_tasks_in_empty_cpuset(struct cpuset *cs)
2137 {
2138         struct cpuset *parent;
2139 
2140         /*
2141          * Find its next-highest non-empty parent, (top cpuset
2142          * has online cpus, so can't be empty).
2143          */
2144         parent = parent_cs(cs);
2145         while (cpumask_empty(parent->cpus_allowed) ||
2146                         nodes_empty(parent->mems_allowed))
2147                 parent = parent_cs(parent);
2148 
2149         if (cgroup_transfer_tasks(parent->css.cgroup, cs->css.cgroup)) {
2150                 pr_err("cpuset: failed to transfer tasks out of empty cpuset ");
2151                 pr_cont_cgroup_name(cs->css.cgroup);
2152                 pr_cont("\n");
2153         }
2154 }
2155 
2156 static void
2157 hotplug_update_tasks_legacy(struct cpuset *cs,
2158                             struct cpumask *new_cpus, nodemask_t *new_mems,
2159                             bool cpus_updated, bool mems_updated)
2160 {
2161         bool is_empty;
2162 
2163         spin_lock_irq(&callback_lock);
2164         cpumask_copy(cs->cpus_allowed, new_cpus);
2165         cpumask_copy(cs->effective_cpus, new_cpus);
2166         cs->mems_allowed = *new_mems;
2167         cs->effective_mems = *new_mems;
2168         spin_unlock_irq(&callback_lock);
2169 
2170         /*
2171          * Don't call update_tasks_cpumask() if the cpuset becomes empty,
2172          * as the tasks will be migratecd to an ancestor.
2173          */
2174         if (cpus_updated && !cpumask_empty(cs->cpus_allowed))
2175                 update_tasks_cpumask(cs);
2176         if (mems_updated && !nodes_empty(cs->mems_allowed))
2177                 update_tasks_nodemask(cs);
2178 
2179         is_empty = cpumask_empty(cs->cpus_allowed) ||
2180                    nodes_empty(cs->mems_allowed);
2181 
2182         mutex_unlock(&cpuset_mutex);
2183 
2184         /*
2185          * Move tasks to the nearest ancestor with execution resources,
2186          * This is full cgroup operation which will also call back into
2187          * cpuset. Should be done outside any lock.
2188          */
2189         if (is_empty)
2190                 remove_tasks_in_empty_cpuset(cs);
2191 
2192         mutex_lock(&cpuset_mutex);
2193 }
2194 
2195 static void
2196 hotplug_update_tasks(struct cpuset *cs,
2197                      struct cpumask *new_cpus, nodemask_t *new_mems,
2198                      bool cpus_updated, bool mems_updated)
2199 {
2200         if (cpumask_empty(new_cpus))
2201                 cpumask_copy(new_cpus, parent_cs(cs)->effective_cpus);
2202         if (nodes_empty(*new_mems))
2203                 *new_mems = parent_cs(cs)->effective_mems;
2204 
2205         spin_lock_irq(&callback_lock);
2206         cpumask_copy(cs->effective_cpus, new_cpus);
2207         cs->effective_mems = *new_mems;
2208         spin_unlock_irq(&callback_lock);
2209 
2210         if (cpus_updated)
2211                 update_tasks_cpumask(cs);
2212         if (mems_updated)
2213                 update_tasks_nodemask(cs);
2214 }
2215 
2216 /**
2217  * cpuset_hotplug_update_tasks - update tasks in a cpuset for hotunplug
2218  * @cs: cpuset in interest
2219  *
2220  * Compare @cs's cpu and mem masks against top_cpuset and if some have gone
2221  * offline, update @cs accordingly.  If @cs ends up with no CPU or memory,
2222  * all its tasks are moved to the nearest ancestor with both resources.
2223  */
2224 static void cpuset_hotplug_update_tasks(struct cpuset *cs)
2225 {
2226         static cpumask_t new_cpus;
2227         static nodemask_t new_mems;
2228         bool cpus_updated;
2229         bool mems_updated;
2230 retry:
2231         wait_event(cpuset_attach_wq, cs->attach_in_progress == 0);
2232 
2233         mutex_lock(&cpuset_mutex);
2234 
2235         /*
2236          * We have raced with task attaching. We wait until attaching
2237          * is finished, so we won't attach a task to an empty cpuset.
2238          */
2239         if (cs->attach_in_progress) {
2240                 mutex_unlock(&cpuset_mutex);
2241                 goto retry;
2242         }
2243 
2244         cpumask_and(&new_cpus, cs->cpus_allowed, parent_cs(cs)->effective_cpus);
2245         nodes_and(new_mems, cs->mems_allowed, parent_cs(cs)->effective_mems);
2246 
2247         cpus_updated = !cpumask_equal(&new_cpus, cs->effective_cpus);
2248         mems_updated = !nodes_equal(new_mems, cs->effective_mems);
2249 
2250         if (cgroup_subsys_on_dfl(cpuset_cgrp_subsys))
2251                 hotplug_update_tasks(cs, &new_cpus, &new_mems,
2252                                      cpus_updated, mems_updated);
2253         else
2254                 hotplug_update_tasks_legacy(cs, &new_cpus, &new_mems,
2255                                             cpus_updated, mems_updated);
2256 
2257         mutex_unlock(&cpuset_mutex);
2258 }
2259 
2260 /**
2261  * cpuset_hotplug_workfn - handle CPU/memory hotunplug for a cpuset
2262  *
2263  * This function is called after either CPU or memory configuration has
2264  * changed and updates cpuset accordingly.  The top_cpuset is always
2265  * synchronized to cpu_active_mask and N_MEMORY, which is necessary in
2266  * order to make cpusets transparent (of no affect) on systems that are
2267  * actively using CPU hotplug but making no active use of cpusets.
2268  *
2269  * Non-root cpusets are only affected by offlining.  If any CPUs or memory
2270  * nodes have been taken down, cpuset_hotplug_update_tasks() is invoked on
2271  * all descendants.
2272  *
2273  * Note that CPU offlining during suspend is ignored.  We don't modify
2274  * cpusets across suspend/resume cycles at all.
2275  */
2276 static void cpuset_hotplug_workfn(struct work_struct *work)
2277 {
2278         static cpumask_t new_cpus;
2279         static nodemask_t new_mems;
2280         bool cpus_updated, mems_updated;
2281         bool on_dfl = cgroup_subsys_on_dfl(cpuset_cgrp_subsys);
2282 
2283         mutex_lock(&cpuset_mutex);
2284 
2285         /* fetch the available cpus/mems and find out which changed how */
2286         cpumask_copy(&new_cpus, cpu_active_mask);
2287         new_mems = node_states[N_MEMORY];
2288 
2289         cpus_updated = !cpumask_equal(top_cpuset.effective_cpus, &new_cpus);
2290         mems_updated = !nodes_equal(top_cpuset.effective_mems, new_mems);
2291 
2292         /* synchronize cpus_allowed to cpu_active_mask */
2293         if (cpus_updated) {
2294                 spin_lock_irq(&callback_lock);
2295                 if (!on_dfl)
2296                         cpumask_copy(top_cpuset.cpus_allowed, &new_cpus);
2297                 cpumask_copy(top_cpuset.effective_cpus, &new_cpus);
2298                 spin_unlock_irq(&callback_lock);
2299                 /* we don't mess with cpumasks of tasks in top_cpuset */
2300         }
2301 
2302         /* synchronize mems_allowed to N_MEMORY */
2303         if (mems_updated) {
2304                 spin_lock_irq(&callback_lock);
2305                 if (!on_dfl)
2306                         top_cpuset.mems_allowed = new_mems;
2307                 top_cpuset.effective_mems = new_mems;
2308                 spin_unlock_irq(&callback_lock);
2309                 update_tasks_nodemask(&top_cpuset);
2310         }
2311 
2312         mutex_unlock(&cpuset_mutex);
2313 
2314         /* if cpus or mems changed, we need to propagate to descendants */
2315         if (cpus_updated || mems_updated) {
2316                 struct cpuset *cs;
2317                 struct cgroup_subsys_state *pos_css;
2318 
2319                 rcu_read_lock();
2320                 cpuset_for_each_descendant_pre(cs, pos_css, &top_cpuset) {
2321                         if (cs == &top_cpuset || !css_tryget_online(&cs->css))
2322                                 continue;
2323                         rcu_read_unlock();
2324 
2325                         cpuset_hotplug_update_tasks(cs);
2326 
2327                         rcu_read_lock();
2328                         css_put(&cs->css);
2329                 }
2330                 rcu_read_unlock();
2331         }
2332 
2333         /* rebuild sched domains if cpus_allowed has changed */
2334         if (cpus_updated)
2335                 rebuild_sched_domains();
2336 }
2337 
2338 void cpuset_update_active_cpus(bool cpu_online)
2339 {
2340         /*
2341          * We're inside cpu hotplug critical region which usually nests
2342          * inside cgroup synchronization.  Bounce actual hotplug processing
2343          * to a work item to avoid reverse locking order.
2344          *
2345          * We still need to do partition_sched_domains() synchronously;
2346          * otherwise, the scheduler will get confused and put tasks to the
2347          * dead CPU.  Fall back to the default single domain.
2348          * cpuset_hotplug_workfn() will rebuild it as necessary.
2349          */
2350         partition_sched_domains(1, NULL, NULL);
2351         schedule_work(&cpuset_hotplug_work);
2352 }
2353 
2354 /*
2355  * Keep top_cpuset.mems_allowed tracking node_states[N_MEMORY].
2356  * Call this routine anytime after node_states[N_MEMORY] changes.
2357  * See cpuset_update_active_cpus() for CPU hotplug handling.
2358  */
2359 static int cpuset_track_online_nodes(struct notifier_block *self,
2360                                 unsigned long action, void *arg)
2361 {
2362         schedule_work(&cpuset_hotplug_work);
2363         return NOTIFY_OK;
2364 }
2365 
2366 static struct notifier_block cpuset_track_online_nodes_nb = {
2367         .notifier_call = cpuset_track_online_nodes,
2368         .priority = 10,         /* ??! */
2369 };
2370 
2371 /**
2372  * cpuset_init_smp - initialize cpus_allowed
2373  *
2374  * Description: Finish top cpuset after cpu, node maps are initialized
2375  */
2376 void __init cpuset_init_smp(void)
2377 {
2378         cpumask_copy(top_cpuset.cpus_allowed, cpu_active_mask);
2379         top_cpuset.mems_allowed = node_states[N_MEMORY];
2380         top_cpuset.old_mems_allowed = top_cpuset.mems_allowed;
2381 
2382         cpumask_copy(top_cpuset.effective_cpus, cpu_active_mask);
2383         top_cpuset.effective_mems = node_states[N_MEMORY];
2384 
2385         register_hotmemory_notifier(&cpuset_track_online_nodes_nb);
2386 
2387         cpuset_migrate_mm_wq = alloc_ordered_workqueue("cpuset_migrate_mm", 0);
2388         BUG_ON(!cpuset_migrate_mm_wq);
2389 }
2390 
2391 /**
2392  * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
2393  * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
2394  * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
2395  *
2396  * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
2397  * attached to the specified @tsk.  Guaranteed to return some non-empty
2398  * subset of cpu_online_mask, even if this means going outside the
2399  * tasks cpuset.
2400  **/
2401 
2402 void cpuset_cpus_allowed(struct task_struct *tsk, struct cpumask *pmask)
2403 {
2404         unsigned long flags;
2405 
2406         spin_lock_irqsave(&callback_lock, flags);
2407         rcu_read_lock();
2408         guarantee_online_cpus(task_cs(tsk), pmask);
2409         rcu_read_unlock();
2410         spin_unlock_irqrestore(&callback_lock, flags);
2411 }
2412 
2413 void cpuset_cpus_allowed_fallback(struct task_struct *tsk)
2414 {
2415         rcu_read_lock();
2416         do_set_cpus_allowed(tsk, task_cs(tsk)->effective_cpus);
2417         rcu_read_unlock();
2418 
2419         /*
2420          * We own tsk->cpus_allowed, nobody can change it under us.
2421          *
2422          * But we used cs && cs->cpus_allowed lockless and thus can
2423          * race with cgroup_attach_task() or update_cpumask() and get
2424          * the wrong tsk->cpus_allowed. However, both cases imply the
2425          * subsequent cpuset_change_cpumask()->set_cpus_allowed_ptr()
2426          * which takes task_rq_lock().
2427          *
2428          * If we are called after it dropped the lock we must see all
2429          * changes in tsk_cs()->cpus_allowed. Otherwise we can temporary
2430          * set any mask even if it is not right from task_cs() pov,
2431          * the pending set_cpus_allowed_ptr() will fix things.
2432          *
2433          * select_fallback_rq() will fix things ups and set cpu_possible_mask
2434          * if required.
2435          */
2436 }
2437 
2438 void __init cpuset_init_current_mems_allowed(void)
2439 {
2440         nodes_setall(current->mems_allowed);
2441 }
2442 
2443 /**
2444  * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
2445  * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
2446  *
2447  * Description: Returns the nodemask_t mems_allowed of the cpuset
2448  * attached to the specified @tsk.  Guaranteed to return some non-empty
2449  * subset of node_states[N_MEMORY], even if this means going outside the
2450  * tasks cpuset.
2451  **/
2452 
2453 nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
2454 {
2455         nodemask_t mask;
2456         unsigned long flags;
2457 
2458         spin_lock_irqsave(&callback_lock, flags);
2459         rcu_read_lock();
2460         guarantee_online_mems(task_cs(tsk), &mask);
2461         rcu_read_unlock();
2462         spin_unlock_irqrestore(&callback_lock, flags);
2463 
2464         return mask;
2465 }
2466 
2467 /**
2468  * cpuset_nodemask_valid_mems_allowed - check nodemask vs. curremt mems_allowed
2469  * @nodemask: the nodemask to be checked
2470  *
2471  * Are any of the nodes in the nodemask allowed in current->mems_allowed?
2472  */
2473 int cpuset_nodemask_valid_mems_allowed(nodemask_t *nodemask)
2474 {
2475         return nodes_intersects(*nodemask, current->mems_allowed);
2476 }
2477 
2478 /*
2479  * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
2480  * mem_hardwall ancestor to the specified cpuset.  Call holding
2481  * callback_lock.  If no ancestor is mem_exclusive or mem_hardwall
2482  * (an unusual configuration), then returns the root cpuset.
2483  */
2484 static struct cpuset *nearest_hardwall_ancestor(struct cpuset *cs)
2485 {
2486         while (!(is_mem_exclusive(cs) || is_mem_hardwall(cs)) && parent_cs(cs))
2487                 cs = parent_cs(cs);
2488         return cs;
2489 }
2490 
2491 /**
2492  * cpuset_node_allowed - Can we allocate on a memory node?
2493  * @node: is this an allowed node?
2494  * @gfp_mask: memory allocation flags
2495  *
2496  * If we're in interrupt, yes, we can always allocate.  If @node is set in
2497  * current's mems_allowed, yes.  If it's not a __GFP_HARDWALL request and this
2498  * node is set in the nearest hardwalled cpuset ancestor to current's cpuset,
2499  * yes.  If current has access to memory reserves due to TIF_MEMDIE, yes.
2500  * Otherwise, no.
2501  *
2502  * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
2503  * and do not allow allocations outside the current tasks cpuset
2504  * unless the task has been OOM killed as is marked TIF_MEMDIE.
2505  * GFP_KERNEL allocations are not so marked, so can escape to the
2506  * nearest enclosing hardwalled ancestor cpuset.
2507  *
2508  * Scanning up parent cpusets requires callback_lock.  The
2509  * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
2510  * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
2511  * current tasks mems_allowed came up empty on the first pass over
2512  * the zonelist.  So only GFP_KERNEL allocations, if all nodes in the
2513  * cpuset are short of memory, might require taking the callback_lock.
2514  *
2515  * The first call here from mm/page_alloc:get_page_from_freelist()
2516  * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
2517  * so no allocation on a node outside the cpuset is allowed (unless
2518  * in interrupt, of course).
2519  *
2520  * The second pass through get_page_from_freelist() doesn't even call
2521  * here for GFP_ATOMIC calls.  For those calls, the __alloc_pages()
2522  * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
2523  * in alloc_flags.  That logic and the checks below have the combined
2524  * affect that:
2525  *      in_interrupt - any node ok (current task context irrelevant)
2526  *      GFP_ATOMIC   - any node ok
2527  *      TIF_MEMDIE   - any node ok
2528  *      GFP_KERNEL   - any node in enclosing hardwalled cpuset ok
2529  *      GFP_USER     - only nodes in current tasks mems allowed ok.
2530  */
2531 int __cpuset_node_allowed(int node, gfp_t gfp_mask)
2532 {
2533         struct cpuset *cs;              /* current cpuset ancestors */
2534         int allowed;                    /* is allocation in zone z allowed? */
2535         unsigned long flags;
2536 
2537         if (in_interrupt())
2538                 return 1;
2539         if (node_isset(node, current->mems_allowed))
2540                 return 1;
2541         /*
2542          * Allow tasks that have access to memory reserves because they have
2543          * been OOM killed to get memory anywhere.
2544          */
2545         if (unlikely(test_thread_flag(TIF_MEMDIE)))
2546                 return 1;
2547         if (gfp_mask & __GFP_HARDWALL)  /* If hardwall request, stop here */
2548                 return 0;
2549 
2550         if (current->flags & PF_EXITING) /* Let dying task have memory */
2551                 return 1;
2552 
2553         /* Not hardwall and node outside mems_allowed: scan up cpusets */
2554         spin_lock_irqsave(&callback_lock, flags);
2555 
2556         rcu_read_lock();
2557         cs = nearest_hardwall_ancestor(task_cs(current));
2558         allowed = node_isset(node, cs->mems_allowed);
2559         rcu_read_unlock();
2560 
2561         spin_unlock_irqrestore(&callback_lock, flags);
2562         return allowed;
2563 }
2564 
2565 /**
2566  * cpuset_mem_spread_node() - On which node to begin search for a file page
2567  * cpuset_slab_spread_node() - On which node to begin search for a slab page
2568  *
2569  * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
2570  * tasks in a cpuset with is_spread_page or is_spread_slab set),
2571  * and if the memory allocation used cpuset_mem_spread_node()
2572  * to determine on which node to start looking, as it will for
2573  * certain page cache or slab cache pages such as used for file
2574  * system buffers and inode caches, then instead of starting on the
2575  * local node to look for a free page, rather spread the starting
2576  * node around the tasks mems_allowed nodes.
2577  *
2578  * We don't have to worry about the returned node being offline
2579  * because "it can't happen", and even if it did, it would be ok.
2580  *
2581  * The routines calling guarantee_online_mems() are careful to
2582  * only set nodes in task->mems_allowed that are online.  So it
2583  * should not be possible for the following code to return an
2584  * offline node.  But if it did, that would be ok, as this routine
2585  * is not returning the node where the allocation must be, only
2586  * the node where the search should start.  The zonelist passed to
2587  * __alloc_pages() will include all nodes.  If the slab allocator
2588  * is passed an offline node, it will fall back to the local node.
2589  * See kmem_cache_alloc_node().
2590  */
2591 
2592 static int cpuset_spread_node(int *rotor)
2593 {
2594         int node;
2595 
2596         node = next_node(*rotor, current->mems_allowed);
2597         if (node == MAX_NUMNODES)
2598                 node = first_node(current->mems_allowed);
2599         *rotor = node;
2600         return node;
2601 }
2602 
2603 int cpuset_mem_spread_node(void)
2604 {
2605         if (current->cpuset_mem_spread_rotor == NUMA_NO_NODE)
2606                 current->cpuset_mem_spread_rotor =
2607                         node_random(&current->mems_allowed);
2608 
2609         return cpuset_spread_node(&current->cpuset_mem_spread_rotor);
2610 }
2611 
2612 int cpuset_slab_spread_node(void)
2613 {
2614         if (current->cpuset_slab_spread_rotor == NUMA_NO_NODE)
2615                 current->cpuset_slab_spread_rotor =
2616                         node_random(&current->mems_allowed);
2617 
2618         return cpuset_spread_node(&current->cpuset_slab_spread_rotor);
2619 }
2620 
2621 EXPORT_SYMBOL_GPL(cpuset_mem_spread_node);
2622 
2623 /**
2624  * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
2625  * @tsk1: pointer to task_struct of some task.
2626  * @tsk2: pointer to task_struct of some other task.
2627  *
2628  * Description: Return true if @tsk1's mems_allowed intersects the
2629  * mems_allowed of @tsk2.  Used by the OOM killer to determine if
2630  * one of the task's memory usage might impact the memory available
2631  * to the other.
2632  **/
2633 
2634 int cpuset_mems_allowed_intersects(const struct task_struct *tsk1,
2635                                    const struct task_struct *tsk2)
2636 {
2637         return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed);
2638 }
2639 
2640 /**
2641  * cpuset_print_current_mems_allowed - prints current's cpuset and mems_allowed
2642  *
2643  * Description: Prints current's name, cpuset name, and cached copy of its
2644  * mems_allowed to the kernel log.
2645  */
2646 void cpuset_print_current_mems_allowed(void)
2647 {
2648         struct cgroup *cgrp;
2649 
2650         rcu_read_lock();
2651 
2652         cgrp = task_cs(current)->css.cgroup;
2653         pr_info("%s cpuset=", current->comm);
2654         pr_cont_cgroup_name(cgrp);
2655         pr_cont(" mems_allowed=%*pbl\n",
2656                 nodemask_pr_args(&current->mems_allowed));
2657 
2658         rcu_read_unlock();
2659 }
2660 
2661 /*
2662  * Collection of memory_pressure is suppressed unless
2663  * this flag is enabled by writing "1" to the special
2664  * cpuset file 'memory_pressure_enabled' in the root cpuset.
2665  */
2666 
2667 int cpuset_memory_pressure_enabled __read_mostly;
2668 
2669 /**
2670  * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
2671  *
2672  * Keep a running average of the rate of synchronous (direct)
2673  * page reclaim efforts initiated by tasks in each cpuset.
2674  *
2675  * This represents the rate at which some task in the cpuset
2676  * ran low on memory on all nodes it was allowed to use, and
2677  * had to enter the kernels page reclaim code in an effort to
2678  * create more free memory by tossing clean pages or swapping
2679  * or writing dirty pages.
2680  *
2681  * Display to user space in the per-cpuset read-only file
2682  * "memory_pressure".  Value displayed is an integer
2683  * representing the recent rate of entry into the synchronous
2684  * (direct) page reclaim by any task attached to the cpuset.
2685  **/
2686 
2687 void __cpuset_memory_pressure_bump(void)
2688 {
2689         rcu_read_lock();
2690         fmeter_markevent(&task_cs(current)->fmeter);
2691         rcu_read_unlock();
2692 }
2693 
2694 #ifdef CONFIG_PROC_PID_CPUSET
2695 /*
2696  * proc_cpuset_show()
2697  *  - Print tasks cpuset path into seq_file.
2698  *  - Used for /proc/<pid>/cpuset.
2699  *  - No need to task_lock(tsk) on this tsk->cpuset reference, as it
2700  *    doesn't really matter if tsk->cpuset changes after we read it,
2701  *    and we take cpuset_mutex, keeping cpuset_attach() from changing it
2702  *    anyway.
2703  */
2704 int proc_cpuset_show(struct seq_file *m, struct pid_namespace *ns,
2705                      struct pid *pid, struct task_struct *tsk)
2706 {
2707         char *buf, *p;
2708         struct cgroup_subsys_state *css;
2709         int retval;
2710 
2711         retval = -ENOMEM;
2712         buf = kmalloc(PATH_MAX, GFP_KERNEL);
2713         if (!buf)
2714                 goto out;
2715 
2716         retval = -ENAMETOOLONG;
2717         rcu_read_lock();
2718         css = task_css(tsk, cpuset_cgrp_id);
2719         p = cgroup_path(css->cgroup, buf, PATH_MAX);
2720         rcu_read_unlock();
2721         if (!p)
2722                 goto out_free;
2723         seq_puts(m, p);
2724         seq_putc(m, '\n');
2725         retval = 0;
2726 out_free:
2727         kfree(buf);
2728 out:
2729         return retval;
2730 }
2731 #endif /* CONFIG_PROC_PID_CPUSET */
2732 
2733 /* Display task mems_allowed in /proc/<pid>/status file. */
2734 void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task)
2735 {
2736         seq_printf(m, "Mems_allowed:\t%*pb\n",
2737                    nodemask_pr_args(&task->mems_allowed));
2738         seq_printf(m, "Mems_allowed_list:\t%*pbl\n",
2739                    nodemask_pr_args(&task->mems_allowed));
2740 }
2741 

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