Linux Kernel  3.7.1
 All Data Structures Namespaces Files Functions Variables Typedefs Enumerations Enumerator Macros Groups Pages
cpuset.c
Go to the documentation of this file.
1 /*
2  * kernel/cpuset.c
3  *
4  * Processor and Memory placement constraints for sets of tasks.
5  *
6  * Copyright (C) 2003 BULL SA.
7  * Copyright (C) 2004-2007 Silicon Graphics, Inc.
8  * Copyright (C) 2006 Google, Inc
9  *
10  * Portions derived from Patrick Mochel's sysfs code.
11  * sysfs is Copyright (c) 2001-3 Patrick Mochel
12  *
13  * 2003-10-10 Written by Simon Derr.
14  * 2003-10-22 Updates by Stephen Hemminger.
15  * 2004 May-July Rework by Paul Jackson.
16  * 2006 Rework by Paul Menage to use generic cgroups
17  * 2008 Rework of the scheduler domains and CPU hotplug handling
18  * by Max Krasnyansky
19  *
20  * This file is subject to the terms and conditions of the GNU General Public
21  * License. See the file COPYING in the main directory of the Linux
22  * distribution for more details.
23  */
24 
25 #include <linux/cpu.h>
26 #include <linux/cpumask.h>
27 #include <linux/cpuset.h>
28 #include <linux/err.h>
29 #include <linux/errno.h>
30 #include <linux/file.h>
31 #include <linux/fs.h>
32 #include <linux/init.h>
33 #include <linux/interrupt.h>
34 #include <linux/kernel.h>
35 #include <linux/kmod.h>
36 #include <linux/list.h>
37 #include <linux/mempolicy.h>
38 #include <linux/mm.h>
39 #include <linux/memory.h>
40 #include <linux/export.h>
41 #include <linux/mount.h>
42 #include <linux/namei.h>
43 #include <linux/pagemap.h>
44 #include <linux/proc_fs.h>
45 #include <linux/rcupdate.h>
46 #include <linux/sched.h>
47 #include <linux/seq_file.h>
48 #include <linux/security.h>
49 #include <linux/slab.h>
50 #include <linux/spinlock.h>
51 #include <linux/stat.h>
52 #include <linux/string.h>
53 #include <linux/time.h>
54 #include <linux/backing-dev.h>
55 #include <linux/sort.h>
56 
57 #include <asm/uaccess.h>
58 #include <linux/atomic.h>
59 #include <linux/mutex.h>
60 #include <linux/workqueue.h>
61 #include <linux/cgroup.h>
62 
63 /*
64  * Workqueue for cpuset related tasks.
65  *
66  * Using kevent workqueue may cause deadlock when memory_migrate
67  * is set. So we create a separate workqueue thread for cpuset.
68  */
69 static struct workqueue_struct *cpuset_wq;
70 
71 /*
72  * Tracks how many cpusets are currently defined in system.
73  * When there is only one cpuset (the root cpuset) we can
74  * short circuit some hooks.
75  */
76 int number_of_cpusets __read_mostly;
77 
78 /* Forward declare cgroup structures */
79 struct cgroup_subsys cpuset_subsys;
80 struct cpuset;
81 
82 /* See "Frequency meter" comments, below. */
83 
84 struct fmeter {
85  int cnt; /* unprocessed events count */
86  int val; /* most recent output value */
87  time_t time; /* clock (secs) when val computed */
88  spinlock_t lock; /* guards read or write of above */
89 };
90 
91 struct cpuset {
92  struct cgroup_subsys_state css;
93 
94  unsigned long flags; /* "unsigned long" so bitops work */
95  cpumask_var_t cpus_allowed; /* CPUs allowed to tasks in cpuset */
96  nodemask_t mems_allowed; /* Memory Nodes allowed to tasks */
97 
98  struct cpuset *parent; /* my parent */
99 
100  struct fmeter fmeter; /* memory_pressure filter */
101 
102  /* partition number for rebuild_sched_domains() */
103  int pn;
104 
105  /* for custom sched domain */
107 
108  /* used for walking a cpuset hierarchy */
110 };
111 
112 /* Retrieve the cpuset for a cgroup */
113 static inline struct cpuset *cgroup_cs(struct cgroup *cont)
114 {
115  return container_of(cgroup_subsys_state(cont, cpuset_subsys_id),
116  struct cpuset, css);
117 }
118 
119 /* Retrieve the cpuset for a task */
120 static inline struct cpuset *task_cs(struct task_struct *task)
121 {
122  return container_of(task_subsys_state(task, cpuset_subsys_id),
123  struct cpuset, css);
124 }
125 
126 #ifdef CONFIG_NUMA
127 static inline bool task_has_mempolicy(struct task_struct *task)
128 {
129  return task->mempolicy;
130 }
131 #else
132 static inline bool task_has_mempolicy(struct task_struct *task)
133 {
134  return false;
135 }
136 #endif
137 
138 
139 /* bits in struct cpuset flags field */
140 typedef enum {
149 
150 /* the type of hotplug event */
154 };
155 
156 /* convenient tests for these bits */
157 static inline int is_cpu_exclusive(const struct cpuset *cs)
158 {
159  return test_bit(CS_CPU_EXCLUSIVE, &cs->flags);
160 }
161 
162 static inline int is_mem_exclusive(const struct cpuset *cs)
163 {
164  return test_bit(CS_MEM_EXCLUSIVE, &cs->flags);
165 }
166 
167 static inline int is_mem_hardwall(const struct cpuset *cs)
168 {
169  return test_bit(CS_MEM_HARDWALL, &cs->flags);
170 }
171 
172 static inline int is_sched_load_balance(const struct cpuset *cs)
173 {
174  return test_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
175 }
176 
177 static inline int is_memory_migrate(const struct cpuset *cs)
178 {
179  return test_bit(CS_MEMORY_MIGRATE, &cs->flags);
180 }
181 
182 static inline int is_spread_page(const struct cpuset *cs)
183 {
184  return test_bit(CS_SPREAD_PAGE, &cs->flags);
185 }
186 
187 static inline int is_spread_slab(const struct cpuset *cs)
188 {
189  return test_bit(CS_SPREAD_SLAB, &cs->flags);
190 }
191 
192 static struct cpuset top_cpuset = {
193  .flags = ((1 << CS_CPU_EXCLUSIVE) | (1 << CS_MEM_EXCLUSIVE)),
194 };
195 
196 /*
197  * There are two global mutexes guarding cpuset structures. The first
198  * is the main control groups cgroup_mutex, accessed via
199  * cgroup_lock()/cgroup_unlock(). The second is the cpuset-specific
200  * callback_mutex, below. They can nest. It is ok to first take
201  * cgroup_mutex, then nest callback_mutex. We also require taking
202  * task_lock() when dereferencing a task's cpuset pointer. See "The
203  * task_lock() exception", at the end of this comment.
204  *
205  * A task must hold both mutexes to modify cpusets. If a task
206  * holds cgroup_mutex, then it blocks others wanting that mutex,
207  * ensuring that it is the only task able to also acquire callback_mutex
208  * and be able to modify cpusets. It can perform various checks on
209  * the cpuset structure first, knowing nothing will change. It can
210  * also allocate memory while just holding cgroup_mutex. While it is
211  * performing these checks, various callback routines can briefly
212  * acquire callback_mutex to query cpusets. Once it is ready to make
213  * the changes, it takes callback_mutex, blocking everyone else.
214  *
215  * Calls to the kernel memory allocator can not be made while holding
216  * callback_mutex, as that would risk double tripping on callback_mutex
217  * from one of the callbacks into the cpuset code from within
218  * __alloc_pages().
219  *
220  * If a task is only holding callback_mutex, then it has read-only
221  * access to cpusets.
222  *
223  * Now, the task_struct fields mems_allowed and mempolicy may be changed
224  * by other task, we use alloc_lock in the task_struct fields to protect
225  * them.
226  *
227  * The cpuset_common_file_read() handlers only hold callback_mutex across
228  * small pieces of code, such as when reading out possibly multi-word
229  * cpumasks and nodemasks.
230  *
231  * Accessing a task's cpuset should be done in accordance with the
232  * guidelines for accessing subsystem state in kernel/cgroup.c
233  */
234 
235 static DEFINE_MUTEX(callback_mutex);
236 
237 /*
238  * cpuset_buffer_lock protects both the cpuset_name and cpuset_nodelist
239  * buffers. They are statically allocated to prevent using excess stack
240  * when calling cpuset_print_task_mems_allowed().
241  */
242 #define CPUSET_NAME_LEN (128)
243 #define CPUSET_NODELIST_LEN (256)
244 static char cpuset_name[CPUSET_NAME_LEN];
245 static char cpuset_nodelist[CPUSET_NODELIST_LEN];
246 static DEFINE_SPINLOCK(cpuset_buffer_lock);
247 
248 /*
249  * This is ugly, but preserves the userspace API for existing cpuset
250  * users. If someone tries to mount the "cpuset" filesystem, we
251  * silently switch it to mount "cgroup" instead
252  */
253 static struct dentry *cpuset_mount(struct file_system_type *fs_type,
254  int flags, const char *unused_dev_name, void *data)
255 {
256  struct file_system_type *cgroup_fs = get_fs_type("cgroup");
257  struct dentry *ret = ERR_PTR(-ENODEV);
258  if (cgroup_fs) {
259  char mountopts[] =
260  "cpuset,noprefix,"
261  "release_agent=/sbin/cpuset_release_agent";
262  ret = cgroup_fs->mount(cgroup_fs, flags,
263  unused_dev_name, mountopts);
264  put_filesystem(cgroup_fs);
265  }
266  return ret;
267 }
268 
269 static struct file_system_type cpuset_fs_type = {
270  .name = "cpuset",
271  .mount = cpuset_mount,
272 };
273 
274 /*
275  * Return in pmask the portion of a cpusets's cpus_allowed that
276  * are online. If none are online, walk up the cpuset hierarchy
277  * until we find one that does have some online cpus. If we get
278  * all the way to the top and still haven't found any online cpus,
279  * return cpu_online_mask. Or if passed a NULL cs from an exit'ing
280  * task, return cpu_online_mask.
281  *
282  * One way or another, we guarantee to return some non-empty subset
283  * of cpu_online_mask.
284  *
285  * Call with callback_mutex held.
286  */
287 
288 static void guarantee_online_cpus(const struct cpuset *cs,
289  struct cpumask *pmask)
290 {
291  while (cs && !cpumask_intersects(cs->cpus_allowed, cpu_online_mask))
292  cs = cs->parent;
293  if (cs)
294  cpumask_and(pmask, cs->cpus_allowed, cpu_online_mask);
295  else
296  cpumask_copy(pmask, cpu_online_mask);
297  BUG_ON(!cpumask_intersects(pmask, cpu_online_mask));
298 }
299 
300 /*
301  * Return in *pmask the portion of a cpusets's mems_allowed that
302  * are online, with memory. If none are online with memory, walk
303  * up the cpuset hierarchy until we find one that does have some
304  * online mems. If we get all the way to the top and still haven't
305  * found any online mems, return node_states[N_HIGH_MEMORY].
306  *
307  * One way or another, we guarantee to return some non-empty subset
308  * of node_states[N_HIGH_MEMORY].
309  *
310  * Call with callback_mutex held.
311  */
312 
313 static void guarantee_online_mems(const struct cpuset *cs, nodemask_t *pmask)
314 {
315  while (cs && !nodes_intersects(cs->mems_allowed,
317  cs = cs->parent;
318  if (cs)
319  nodes_and(*pmask, cs->mems_allowed,
320  node_states[N_HIGH_MEMORY]);
321  else
322  *pmask = node_states[N_HIGH_MEMORY];
323  BUG_ON(!nodes_intersects(*pmask, node_states[N_HIGH_MEMORY]));
324 }
325 
326 /*
327  * update task's spread flag if cpuset's page/slab spread flag is set
328  *
329  * Called with callback_mutex/cgroup_mutex held
330  */
331 static void cpuset_update_task_spread_flag(struct cpuset *cs,
332  struct task_struct *tsk)
333 {
334  if (is_spread_page(cs))
335  tsk->flags |= PF_SPREAD_PAGE;
336  else
337  tsk->flags &= ~PF_SPREAD_PAGE;
338  if (is_spread_slab(cs))
339  tsk->flags |= PF_SPREAD_SLAB;
340  else
341  tsk->flags &= ~PF_SPREAD_SLAB;
342 }
343 
344 /*
345  * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
346  *
347  * One cpuset is a subset of another if all its allowed CPUs and
348  * Memory Nodes are a subset of the other, and its exclusive flags
349  * are only set if the other's are set. Call holding cgroup_mutex.
350  */
351 
352 static int is_cpuset_subset(const struct cpuset *p, const struct cpuset *q)
353 {
354  return cpumask_subset(p->cpus_allowed, q->cpus_allowed) &&
356  is_cpu_exclusive(p) <= is_cpu_exclusive(q) &&
357  is_mem_exclusive(p) <= is_mem_exclusive(q);
358 }
359 
364 static struct cpuset *alloc_trial_cpuset(const struct cpuset *cs)
365 {
366  struct cpuset *trial;
367 
368  trial = kmemdup(cs, sizeof(*cs), GFP_KERNEL);
369  if (!trial)
370  return NULL;
371 
372  if (!alloc_cpumask_var(&trial->cpus_allowed, GFP_KERNEL)) {
373  kfree(trial);
374  return NULL;
375  }
376  cpumask_copy(trial->cpus_allowed, cs->cpus_allowed);
377 
378  return trial;
379 }
380 
385 static void free_trial_cpuset(struct cpuset *trial)
386 {
387  free_cpumask_var(trial->cpus_allowed);
388  kfree(trial);
389 }
390 
391 /*
392  * validate_change() - Used to validate that any proposed cpuset change
393  * follows the structural rules for cpusets.
394  *
395  * If we replaced the flag and mask values of the current cpuset
396  * (cur) with those values in the trial cpuset (trial), would
397  * our various subset and exclusive rules still be valid? Presumes
398  * cgroup_mutex held.
399  *
400  * 'cur' is the address of an actual, in-use cpuset. Operations
401  * such as list traversal that depend on the actual address of the
402  * cpuset in the list must use cur below, not trial.
403  *
404  * 'trial' is the address of bulk structure copy of cur, with
405  * perhaps one or more of the fields cpus_allowed, mems_allowed,
406  * or flags changed to new, trial values.
407  *
408  * Return 0 if valid, -errno if not.
409  */
410 
411 static int validate_change(const struct cpuset *cur, const struct cpuset *trial)
412 {
413  struct cgroup *cont;
414  struct cpuset *c, *par;
415 
416  /* Each of our child cpusets must be a subset of us */
417  list_for_each_entry(cont, &cur->css.cgroup->children, sibling) {
418  if (!is_cpuset_subset(cgroup_cs(cont), trial))
419  return -EBUSY;
420  }
421 
422  /* Remaining checks don't apply to root cpuset */
423  if (cur == &top_cpuset)
424  return 0;
425 
426  par = cur->parent;
427 
428  /* We must be a subset of our parent cpuset */
429  if (!is_cpuset_subset(trial, par))
430  return -EACCES;
431 
432  /*
433  * If either I or some sibling (!= me) is exclusive, we can't
434  * overlap
435  */
436  list_for_each_entry(cont, &par->css.cgroup->children, sibling) {
437  c = cgroup_cs(cont);
438  if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
439  c != cur &&
440  cpumask_intersects(trial->cpus_allowed, c->cpus_allowed))
441  return -EINVAL;
442  if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) &&
443  c != cur &&
445  return -EINVAL;
446  }
447 
448  /* Cpusets with tasks can't have empty cpus_allowed or mems_allowed */
449  if (cgroup_task_count(cur->css.cgroup)) {
450  if (cpumask_empty(trial->cpus_allowed) ||
451  nodes_empty(trial->mems_allowed)) {
452  return -ENOSPC;
453  }
454  }
455 
456  return 0;
457 }
458 
459 #ifdef CONFIG_SMP
460 /*
461  * Helper routine for generate_sched_domains().
462  * Do cpusets a, b have overlapping cpus_allowed masks?
463  */
464 static int cpusets_overlap(struct cpuset *a, struct cpuset *b)
465 {
466  return cpumask_intersects(a->cpus_allowed, b->cpus_allowed);
467 }
468 
469 static void
470 update_domain_attr(struct sched_domain_attr *dattr, struct cpuset *c)
471 {
472  if (dattr->relax_domain_level < c->relax_domain_level)
473  dattr->relax_domain_level = c->relax_domain_level;
474  return;
475 }
476 
477 static void
478 update_domain_attr_tree(struct sched_domain_attr *dattr, struct cpuset *c)
479 {
480  LIST_HEAD(q);
481 
482  list_add(&c->stack_list, &q);
483  while (!list_empty(&q)) {
484  struct cpuset *cp;
485  struct cgroup *cont;
486  struct cpuset *child;
487 
488  cp = list_first_entry(&q, struct cpuset, stack_list);
489  list_del(q.next);
490 
491  if (cpumask_empty(cp->cpus_allowed))
492  continue;
493 
494  if (is_sched_load_balance(cp))
495  update_domain_attr(dattr, cp);
496 
497  list_for_each_entry(cont, &cp->css.cgroup->children, sibling) {
498  child = cgroup_cs(cont);
499  list_add_tail(&child->stack_list, &q);
500  }
501  }
502 }
503 
504 /*
505  * generate_sched_domains()
506  *
507  * This function builds a partial partition of the systems CPUs
508  * A 'partial partition' is a set of non-overlapping subsets whose
509  * union is a subset of that set.
510  * The output of this function needs to be passed to kernel/sched.c
511  * partition_sched_domains() routine, which will rebuild the scheduler's
512  * load balancing domains (sched domains) as specified by that partial
513  * partition.
514  *
515  * See "What is sched_load_balance" in Documentation/cgroups/cpusets.txt
516  * for a background explanation of this.
517  *
518  * Does not return errors, on the theory that the callers of this
519  * routine would rather not worry about failures to rebuild sched
520  * domains when operating in the severe memory shortage situations
521  * that could cause allocation failures below.
522  *
523  * Must be called with cgroup_lock held.
524  *
525  * The three key local variables below are:
526  * q - a linked-list queue of cpuset pointers, used to implement a
527  * top-down scan of all cpusets. This scan loads a pointer
528  * to each cpuset marked is_sched_load_balance into the
529  * array 'csa'. For our purposes, rebuilding the schedulers
530  * sched domains, we can ignore !is_sched_load_balance cpusets.
531  * csa - (for CpuSet Array) Array of pointers to all the cpusets
532  * that need to be load balanced, for convenient iterative
533  * access by the subsequent code that finds the best partition,
534  * i.e the set of domains (subsets) of CPUs such that the
535  * cpus_allowed of every cpuset marked is_sched_load_balance
536  * is a subset of one of these domains, while there are as
537  * many such domains as possible, each as small as possible.
538  * doms - Conversion of 'csa' to an array of cpumasks, for passing to
539  * the kernel/sched.c routine partition_sched_domains() in a
540  * convenient format, that can be easily compared to the prior
541  * value to determine what partition elements (sched domains)
542  * were changed (added or removed.)
543  *
544  * Finding the best partition (set of domains):
545  * The triple nested loops below over i, j, k scan over the
546  * load balanced cpusets (using the array of cpuset pointers in
547  * csa[]) looking for pairs of cpusets that have overlapping
548  * cpus_allowed, but which don't have the same 'pn' partition
549  * number and gives them in the same partition number. It keeps
550  * looping on the 'restart' label until it can no longer find
551  * any such pairs.
552  *
553  * The union of the cpus_allowed masks from the set of
554  * all cpusets having the same 'pn' value then form the one
555  * element of the partition (one sched domain) to be passed to
556  * partition_sched_domains().
557  */
558 static int generate_sched_domains(cpumask_var_t **domains,
559  struct sched_domain_attr **attributes)
560 {
561  LIST_HEAD(q); /* queue of cpusets to be scanned */
562  struct cpuset *cp; /* scans q */
563  struct cpuset **csa; /* array of all cpuset ptrs */
564  int csn; /* how many cpuset ptrs in csa so far */
565  int i, j, k; /* indices for partition finding loops */
566  cpumask_var_t *doms; /* resulting partition; i.e. sched domains */
567  struct sched_domain_attr *dattr; /* attributes for custom domains */
568  int ndoms = 0; /* number of sched domains in result */
569  int nslot; /* next empty doms[] struct cpumask slot */
570 
571  doms = NULL;
572  dattr = NULL;
573  csa = NULL;
574 
575  /* Special case for the 99% of systems with one, full, sched domain */
576  if (is_sched_load_balance(&top_cpuset)) {
577  ndoms = 1;
578  doms = alloc_sched_domains(ndoms);
579  if (!doms)
580  goto done;
581 
582  dattr = kmalloc(sizeof(struct sched_domain_attr), GFP_KERNEL);
583  if (dattr) {
584  *dattr = SD_ATTR_INIT;
585  update_domain_attr_tree(dattr, &top_cpuset);
586  }
587  cpumask_copy(doms[0], top_cpuset.cpus_allowed);
588 
589  goto done;
590  }
591 
592  csa = kmalloc(number_of_cpusets * sizeof(cp), GFP_KERNEL);
593  if (!csa)
594  goto done;
595  csn = 0;
596 
597  list_add(&top_cpuset.stack_list, &q);
598  while (!list_empty(&q)) {
599  struct cgroup *cont;
600  struct cpuset *child; /* scans child cpusets of cp */
601 
602  cp = list_first_entry(&q, struct cpuset, stack_list);
603  list_del(q.next);
604 
605  if (cpumask_empty(cp->cpus_allowed))
606  continue;
607 
608  /*
609  * All child cpusets contain a subset of the parent's cpus, so
610  * just skip them, and then we call update_domain_attr_tree()
611  * to calc relax_domain_level of the corresponding sched
612  * domain.
613  */
614  if (is_sched_load_balance(cp)) {
615  csa[csn++] = cp;
616  continue;
617  }
618 
619  list_for_each_entry(cont, &cp->css.cgroup->children, sibling) {
620  child = cgroup_cs(cont);
621  list_add_tail(&child->stack_list, &q);
622  }
623  }
624 
625  for (i = 0; i < csn; i++)
626  csa[i]->pn = i;
627  ndoms = csn;
628 
629 restart:
630  /* Find the best partition (set of sched domains) */
631  for (i = 0; i < csn; i++) {
632  struct cpuset *a = csa[i];
633  int apn = a->pn;
634 
635  for (j = 0; j < csn; j++) {
636  struct cpuset *b = csa[j];
637  int bpn = b->pn;
638 
639  if (apn != bpn && cpusets_overlap(a, b)) {
640  for (k = 0; k < csn; k++) {
641  struct cpuset *c = csa[k];
642 
643  if (c->pn == bpn)
644  c->pn = apn;
645  }
646  ndoms--; /* one less element */
647  goto restart;
648  }
649  }
650  }
651 
652  /*
653  * Now we know how many domains to create.
654  * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
655  */
656  doms = alloc_sched_domains(ndoms);
657  if (!doms)
658  goto done;
659 
660  /*
661  * The rest of the code, including the scheduler, can deal with
662  * dattr==NULL case. No need to abort if alloc fails.
663  */
664  dattr = kmalloc(ndoms * sizeof(struct sched_domain_attr), GFP_KERNEL);
665 
666  for (nslot = 0, i = 0; i < csn; i++) {
667  struct cpuset *a = csa[i];
668  struct cpumask *dp;
669  int apn = a->pn;
670 
671  if (apn < 0) {
672  /* Skip completed partitions */
673  continue;
674  }
675 
676  dp = doms[nslot];
677 
678  if (nslot == ndoms) {
679  static int warnings = 10;
680  if (warnings) {
682  "rebuild_sched_domains confused:"
683  " nslot %d, ndoms %d, csn %d, i %d,"
684  " apn %d\n",
685  nslot, ndoms, csn, i, apn);
686  warnings--;
687  }
688  continue;
689  }
690 
691  cpumask_clear(dp);
692  if (dattr)
693  *(dattr + nslot) = SD_ATTR_INIT;
694  for (j = i; j < csn; j++) {
695  struct cpuset *b = csa[j];
696 
697  if (apn == b->pn) {
698  cpumask_or(dp, dp, b->cpus_allowed);
699  if (dattr)
700  update_domain_attr_tree(dattr + nslot, b);
701 
702  /* Done with this partition */
703  b->pn = -1;
704  }
705  }
706  nslot++;
707  }
708  BUG_ON(nslot != ndoms);
709 
710 done:
711  kfree(csa);
712 
713  /*
714  * Fallback to the default domain if kmalloc() failed.
715  * See comments in partition_sched_domains().
716  */
717  if (doms == NULL)
718  ndoms = 1;
719 
720  *domains = doms;
721  *attributes = dattr;
722  return ndoms;
723 }
724 
725 /*
726  * Rebuild scheduler domains.
727  *
728  * Call with neither cgroup_mutex held nor within get_online_cpus().
729  * Takes both cgroup_mutex and get_online_cpus().
730  *
731  * Cannot be directly called from cpuset code handling changes
732  * to the cpuset pseudo-filesystem, because it cannot be called
733  * from code that already holds cgroup_mutex.
734  */
735 static void do_rebuild_sched_domains(struct work_struct *unused)
736 {
737  struct sched_domain_attr *attr;
738  cpumask_var_t *doms;
739  int ndoms;
740 
741  get_online_cpus();
742 
743  /* Generate domain masks and attrs */
744  cgroup_lock();
745  ndoms = generate_sched_domains(&doms, &attr);
746  cgroup_unlock();
747 
748  /* Have scheduler rebuild the domains */
749  partition_sched_domains(ndoms, doms, attr);
750 
751  put_online_cpus();
752 }
753 #else /* !CONFIG_SMP */
754 static void do_rebuild_sched_domains(struct work_struct *unused)
755 {
756 }
757 
758 static int generate_sched_domains(cpumask_var_t **domains,
759  struct sched_domain_attr **attributes)
760 {
761  *domains = NULL;
762  return 1;
763 }
764 #endif /* CONFIG_SMP */
765 
766 static DECLARE_WORK(rebuild_sched_domains_work, do_rebuild_sched_domains);
767 
768 /*
769  * Rebuild scheduler domains, asynchronously via workqueue.
770  *
771  * If the flag 'sched_load_balance' of any cpuset with non-empty
772  * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
773  * which has that flag enabled, or if any cpuset with a non-empty
774  * 'cpus' is removed, then call this routine to rebuild the
775  * scheduler's dynamic sched domains.
776  *
777  * The rebuild_sched_domains() and partition_sched_domains()
778  * routines must nest cgroup_lock() inside get_online_cpus(),
779  * but such cpuset changes as these must nest that locking the
780  * other way, holding cgroup_lock() for much of the code.
781  *
782  * So in order to avoid an ABBA deadlock, the cpuset code handling
783  * these user changes delegates the actual sched domain rebuilding
784  * to a separate workqueue thread, which ends up processing the
785  * above do_rebuild_sched_domains() function.
786  */
787 static void async_rebuild_sched_domains(void)
788 {
789  queue_work(cpuset_wq, &rebuild_sched_domains_work);
790 }
791 
792 /*
793  * Accomplishes the same scheduler domain rebuild as the above
794  * async_rebuild_sched_domains(), however it directly calls the
795  * rebuild routine synchronously rather than calling it via an
796  * asynchronous work thread.
797  *
798  * This can only be called from code that is not holding
799  * cgroup_mutex (not nested in a cgroup_lock() call.)
800  */
802 {
803  do_rebuild_sched_domains(NULL);
804 }
805 
816 static int cpuset_test_cpumask(struct task_struct *tsk,
817  struct cgroup_scanner *scan)
818 {
819  return !cpumask_equal(&tsk->cpus_allowed,
820  (cgroup_cs(scan->cg))->cpus_allowed);
821 }
822 
834 static void cpuset_change_cpumask(struct task_struct *tsk,
835  struct cgroup_scanner *scan)
836 {
837  set_cpus_allowed_ptr(tsk, ((cgroup_cs(scan->cg))->cpus_allowed));
838 }
839 
853 static void update_tasks_cpumask(struct cpuset *cs, struct ptr_heap *heap)
854 {
855  struct cgroup_scanner scan;
856 
857  scan.cg = cs->css.cgroup;
858  scan.test_task = cpuset_test_cpumask;
859  scan.process_task = cpuset_change_cpumask;
860  scan.heap = heap;
861  cgroup_scan_tasks(&scan);
862 }
863 
869 static int update_cpumask(struct cpuset *cs, struct cpuset *trialcs,
870  const char *buf)
871 {
872  struct ptr_heap heap;
873  int retval;
874  int is_load_balanced;
875 
876  /* top_cpuset.cpus_allowed tracks cpu_online_mask; it's read-only */
877  if (cs == &top_cpuset)
878  return -EACCES;
879 
880  /*
881  * An empty cpus_allowed is ok only if the cpuset has no tasks.
882  * Since cpulist_parse() fails on an empty mask, we special case
883  * that parsing. The validate_change() call ensures that cpusets
884  * with tasks have cpus.
885  */
886  if (!*buf) {
887  cpumask_clear(trialcs->cpus_allowed);
888  } else {
889  retval = cpulist_parse(buf, trialcs->cpus_allowed);
890  if (retval < 0)
891  return retval;
892 
893  if (!cpumask_subset(trialcs->cpus_allowed, cpu_active_mask))
894  return -EINVAL;
895  }
896  retval = validate_change(cs, trialcs);
897  if (retval < 0)
898  return retval;
899 
900  /* Nothing to do if the cpus didn't change */
901  if (cpumask_equal(cs->cpus_allowed, trialcs->cpus_allowed))
902  return 0;
903 
904  retval = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
905  if (retval)
906  return retval;
907 
908  is_load_balanced = is_sched_load_balance(trialcs);
909 
910  mutex_lock(&callback_mutex);
911  cpumask_copy(cs->cpus_allowed, trialcs->cpus_allowed);
912  mutex_unlock(&callback_mutex);
913 
914  /*
915  * Scan tasks in the cpuset, and update the cpumasks of any
916  * that need an update.
917  */
918  update_tasks_cpumask(cs, &heap);
919 
920  heap_free(&heap);
921 
922  if (is_load_balanced)
923  async_rebuild_sched_domains();
924  return 0;
925 }
926 
927 /*
928  * cpuset_migrate_mm
929  *
930  * Migrate memory region from one set of nodes to another.
931  *
932  * Temporarilly set tasks mems_allowed to target nodes of migration,
933  * so that the migration code can allocate pages on these nodes.
934  *
935  * Call holding cgroup_mutex, so current's cpuset won't change
936  * during this call, as manage_mutex holds off any cpuset_attach()
937  * calls. Therefore we don't need to take task_lock around the
938  * call to guarantee_online_mems(), as we know no one is changing
939  * our task's cpuset.
940  *
941  * While the mm_struct we are migrating is typically from some
942  * other task, the task_struct mems_allowed that we are hacking
943  * is for our current task, which must allocate new pages for that
944  * migrating memory region.
945  */
946 
947 static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
948  const nodemask_t *to)
949 {
950  struct task_struct *tsk = current;
951 
952  tsk->mems_allowed = *to;
953 
954  do_migrate_pages(mm, from, to, MPOL_MF_MOVE_ALL);
955 
956  guarantee_online_mems(task_cs(tsk),&tsk->mems_allowed);
957 }
958 
959 /*
960  * cpuset_change_task_nodemask - change task's mems_allowed and mempolicy
961  * @tsk: the task to change
962  * @newmems: new nodes that the task will be set
963  *
964  * In order to avoid seeing no nodes if the old and new nodes are disjoint,
965  * we structure updates as setting all new allowed nodes, then clearing newly
966  * disallowed ones.
967  */
968 static void cpuset_change_task_nodemask(struct task_struct *tsk,
969  nodemask_t *newmems)
970 {
971  bool need_loop;
972 
973  /*
974  * Allow tasks that have access to memory reserves because they have
975  * been OOM killed to get memory anywhere.
976  */
977  if (unlikely(test_thread_flag(TIF_MEMDIE)))
978  return;
979  if (current->flags & PF_EXITING) /* Let dying task have memory */
980  return;
981 
982  task_lock(tsk);
983  /*
984  * Determine if a loop is necessary if another thread is doing
985  * get_mems_allowed(). If at least one node remains unchanged and
986  * tsk does not have a mempolicy, then an empty nodemask will not be
987  * possible when mems_allowed is larger than a word.
988  */
989  need_loop = task_has_mempolicy(tsk) ||
990  !nodes_intersects(*newmems, tsk->mems_allowed);
991 
992  if (need_loop)
993  write_seqcount_begin(&tsk->mems_allowed_seq);
994 
995  nodes_or(tsk->mems_allowed, tsk->mems_allowed, *newmems);
996  mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP1);
997 
998  mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP2);
999  tsk->mems_allowed = *newmems;
1000 
1001  if (need_loop)
1002  write_seqcount_end(&tsk->mems_allowed_seq);
1003 
1004  task_unlock(tsk);
1005 }
1006 
1007 /*
1008  * Update task's mems_allowed and rebind its mempolicy and vmas' mempolicy
1009  * of it to cpuset's new mems_allowed, and migrate pages to new nodes if
1010  * memory_migrate flag is set. Called with cgroup_mutex held.
1011  */
1012 static void cpuset_change_nodemask(struct task_struct *p,
1013  struct cgroup_scanner *scan)
1014 {
1015  struct mm_struct *mm;
1016  struct cpuset *cs;
1017  int migrate;
1018  const nodemask_t *oldmem = scan->data;
1019  static nodemask_t newmems; /* protected by cgroup_mutex */
1020 
1021  cs = cgroup_cs(scan->cg);
1022  guarantee_online_mems(cs, &newmems);
1023 
1024  cpuset_change_task_nodemask(p, &newmems);
1025 
1026  mm = get_task_mm(p);
1027  if (!mm)
1028  return;
1029 
1030  migrate = is_memory_migrate(cs);
1031 
1032  mpol_rebind_mm(mm, &cs->mems_allowed);
1033  if (migrate)
1034  cpuset_migrate_mm(mm, oldmem, &cs->mems_allowed);
1035  mmput(mm);
1036 }
1037 
1038 static void *cpuset_being_rebound;
1039 
1050 static void update_tasks_nodemask(struct cpuset *cs, const nodemask_t *oldmem,
1051  struct ptr_heap *heap)
1052 {
1053  struct cgroup_scanner scan;
1054 
1055  cpuset_being_rebound = cs; /* causes mpol_dup() rebind */
1056 
1057  scan.cg = cs->css.cgroup;
1058  scan.test_task = NULL;
1059  scan.process_task = cpuset_change_nodemask;
1060  scan.heap = heap;
1061  scan.data = (nodemask_t *)oldmem;
1062 
1063  /*
1064  * The mpol_rebind_mm() call takes mmap_sem, which we couldn't
1065  * take while holding tasklist_lock. Forks can happen - the
1066  * mpol_dup() cpuset_being_rebound check will catch such forks,
1067  * and rebind their vma mempolicies too. Because we still hold
1068  * the global cgroup_mutex, we know that no other rebind effort
1069  * will be contending for the global variable cpuset_being_rebound.
1070  * It's ok if we rebind the same mm twice; mpol_rebind_mm()
1071  * is idempotent. Also migrate pages in each mm to new nodes.
1072  */
1073  cgroup_scan_tasks(&scan);
1074 
1075  /* We're done rebinding vmas to this cpuset's new mems_allowed. */
1076  cpuset_being_rebound = NULL;
1077 }
1078 
1079 /*
1080  * Handle user request to change the 'mems' memory placement
1081  * of a cpuset. Needs to validate the request, update the
1082  * cpusets mems_allowed, and for each task in the cpuset,
1083  * update mems_allowed and rebind task's mempolicy and any vma
1084  * mempolicies and if the cpuset is marked 'memory_migrate',
1085  * migrate the tasks pages to the new memory.
1086  *
1087  * Call with cgroup_mutex held. May take callback_mutex during call.
1088  * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
1089  * lock each such tasks mm->mmap_sem, scan its vma's and rebind
1090  * their mempolicies to the cpusets new mems_allowed.
1091  */
1092 static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
1093  const char *buf)
1094 {
1096  int retval;
1097  struct ptr_heap heap;
1098 
1099  if (!oldmem)
1100  return -ENOMEM;
1101 
1102  /*
1103  * top_cpuset.mems_allowed tracks node_stats[N_HIGH_MEMORY];
1104  * it's read-only
1105  */
1106  if (cs == &top_cpuset) {
1107  retval = -EACCES;
1108  goto done;
1109  }
1110 
1111  /*
1112  * An empty mems_allowed is ok iff there are no tasks in the cpuset.
1113  * Since nodelist_parse() fails on an empty mask, we special case
1114  * that parsing. The validate_change() call ensures that cpusets
1115  * with tasks have memory.
1116  */
1117  if (!*buf) {
1118  nodes_clear(trialcs->mems_allowed);
1119  } else {
1120  retval = nodelist_parse(buf, trialcs->mems_allowed);
1121  if (retval < 0)
1122  goto done;
1123 
1124  if (!nodes_subset(trialcs->mems_allowed,
1125  node_states[N_HIGH_MEMORY])) {
1126  retval = -EINVAL;
1127  goto done;
1128  }
1129  }
1130  *oldmem = cs->mems_allowed;
1131  if (nodes_equal(*oldmem, trialcs->mems_allowed)) {
1132  retval = 0; /* Too easy - nothing to do */
1133  goto done;
1134  }
1135  retval = validate_change(cs, trialcs);
1136  if (retval < 0)
1137  goto done;
1138 
1139  retval = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
1140  if (retval < 0)
1141  goto done;
1142 
1143  mutex_lock(&callback_mutex);
1144  cs->mems_allowed = trialcs->mems_allowed;
1145  mutex_unlock(&callback_mutex);
1146 
1147  update_tasks_nodemask(cs, oldmem, &heap);
1148 
1149  heap_free(&heap);
1150 done:
1151  NODEMASK_FREE(oldmem);
1152  return retval;
1153 }
1154 
1156 {
1157  return task_cs(current) == cpuset_being_rebound;
1158 }
1159 
1160 static int update_relax_domain_level(struct cpuset *cs, s64 val)
1161 {
1162 #ifdef CONFIG_SMP
1163  if (val < -1 || val >= sched_domain_level_max)
1164  return -EINVAL;
1165 #endif
1166 
1167  if (val != cs->relax_domain_level) {
1168  cs->relax_domain_level = val;
1169  if (!cpumask_empty(cs->cpus_allowed) &&
1170  is_sched_load_balance(cs))
1171  async_rebuild_sched_domains();
1172  }
1173 
1174  return 0;
1175 }
1176 
1177 /*
1178  * cpuset_change_flag - make a task's spread flags the same as its cpuset's
1179  * @tsk: task to be updated
1180  * @scan: struct cgroup_scanner containing the cgroup of the task
1181  *
1182  * Called by cgroup_scan_tasks() for each task in a cgroup.
1183  *
1184  * We don't need to re-check for the cgroup/cpuset membership, since we're
1185  * holding cgroup_lock() at this point.
1186  */
1187 static void cpuset_change_flag(struct task_struct *tsk,
1188  struct cgroup_scanner *scan)
1189 {
1190  cpuset_update_task_spread_flag(cgroup_cs(scan->cg), tsk);
1191 }
1192 
1193 /*
1194  * update_tasks_flags - update the spread flags of tasks in the cpuset.
1195  * @cs: the cpuset in which each task's spread flags needs to be changed
1196  * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
1197  *
1198  * Called with cgroup_mutex held
1199  *
1200  * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
1201  * calling callback functions for each.
1202  *
1203  * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
1204  * if @heap != NULL.
1205  */
1206 static void update_tasks_flags(struct cpuset *cs, struct ptr_heap *heap)
1207 {
1208  struct cgroup_scanner scan;
1209 
1210  scan.cg = cs->css.cgroup;
1211  scan.test_task = NULL;
1212  scan.process_task = cpuset_change_flag;
1213  scan.heap = heap;
1214  cgroup_scan_tasks(&scan);
1215 }
1216 
1217 /*
1218  * update_flag - read a 0 or a 1 in a file and update associated flag
1219  * bit: the bit to update (see cpuset_flagbits_t)
1220  * cs: the cpuset to update
1221  * turning_on: whether the flag is being set or cleared
1222  *
1223  * Call with cgroup_mutex held.
1224  */
1225 
1226 static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs,
1227  int turning_on)
1228 {
1229  struct cpuset *trialcs;
1230  int balance_flag_changed;
1231  int spread_flag_changed;
1232  struct ptr_heap heap;
1233  int err;
1234 
1235  trialcs = alloc_trial_cpuset(cs);
1236  if (!trialcs)
1237  return -ENOMEM;
1238 
1239  if (turning_on)
1240  set_bit(bit, &trialcs->flags);
1241  else
1242  clear_bit(bit, &trialcs->flags);
1243 
1244  err = validate_change(cs, trialcs);
1245  if (err < 0)
1246  goto out;
1247 
1248  err = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
1249  if (err < 0)
1250  goto out;
1251 
1252  balance_flag_changed = (is_sched_load_balance(cs) !=
1253  is_sched_load_balance(trialcs));
1254 
1255  spread_flag_changed = ((is_spread_slab(cs) != is_spread_slab(trialcs))
1256  || (is_spread_page(cs) != is_spread_page(trialcs)));
1257 
1258  mutex_lock(&callback_mutex);
1259  cs->flags = trialcs->flags;
1260  mutex_unlock(&callback_mutex);
1261 
1262  if (!cpumask_empty(trialcs->cpus_allowed) && balance_flag_changed)
1263  async_rebuild_sched_domains();
1264 
1265  if (spread_flag_changed)
1266  update_tasks_flags(cs, &heap);
1267  heap_free(&heap);
1268 out:
1269  free_trial_cpuset(trialcs);
1270  return err;
1271 }
1272 
1273 /*
1274  * Frequency meter - How fast is some event occurring?
1275  *
1276  * These routines manage a digitally filtered, constant time based,
1277  * event frequency meter. There are four routines:
1278  * fmeter_init() - initialize a frequency meter.
1279  * fmeter_markevent() - called each time the event happens.
1280  * fmeter_getrate() - returns the recent rate of such events.
1281  * fmeter_update() - internal routine used to update fmeter.
1282  *
1283  * A common data structure is passed to each of these routines,
1284  * which is used to keep track of the state required to manage the
1285  * frequency meter and its digital filter.
1286  *
1287  * The filter works on the number of events marked per unit time.
1288  * The filter is single-pole low-pass recursive (IIR). The time unit
1289  * is 1 second. Arithmetic is done using 32-bit integers scaled to
1290  * simulate 3 decimal digits of precision (multiplied by 1000).
1291  *
1292  * With an FM_COEF of 933, and a time base of 1 second, the filter
1293  * has a half-life of 10 seconds, meaning that if the events quit
1294  * happening, then the rate returned from the fmeter_getrate()
1295  * will be cut in half each 10 seconds, until it converges to zero.
1296  *
1297  * It is not worth doing a real infinitely recursive filter. If more
1298  * than FM_MAXTICKS ticks have elapsed since the last filter event,
1299  * just compute FM_MAXTICKS ticks worth, by which point the level
1300  * will be stable.
1301  *
1302  * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
1303  * arithmetic overflow in the fmeter_update() routine.
1304  *
1305  * Given the simple 32 bit integer arithmetic used, this meter works
1306  * best for reporting rates between one per millisecond (msec) and
1307  * one per 32 (approx) seconds. At constant rates faster than one
1308  * per msec it maxes out at values just under 1,000,000. At constant
1309  * rates between one per msec, and one per second it will stabilize
1310  * to a value N*1000, where N is the rate of events per second.
1311  * At constant rates between one per second and one per 32 seconds,
1312  * it will be choppy, moving up on the seconds that have an event,
1313  * and then decaying until the next event. At rates slower than
1314  * about one in 32 seconds, it decays all the way back to zero between
1315  * each event.
1316  */
1317 
1318 #define FM_COEF 933 /* coefficient for half-life of 10 secs */
1319 #define FM_MAXTICKS ((time_t)99) /* useless computing more ticks than this */
1320 #define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
1321 #define FM_SCALE 1000 /* faux fixed point scale */
1322 
1323 /* Initialize a frequency meter */
1324 static void fmeter_init(struct fmeter *fmp)
1325 {
1326  fmp->cnt = 0;
1327  fmp->val = 0;
1328  fmp->time = 0;
1329  spin_lock_init(&fmp->lock);
1330 }
1331 
1332 /* Internal meter update - process cnt events and update value */
1333 static void fmeter_update(struct fmeter *fmp)
1334 {
1335  time_t now = get_seconds();
1336  time_t ticks = now - fmp->time;
1337 
1338  if (ticks == 0)
1339  return;
1340 
1341  ticks = min(FM_MAXTICKS, ticks);
1342  while (ticks-- > 0)
1343  fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
1344  fmp->time = now;
1345 
1346  fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
1347  fmp->cnt = 0;
1348 }
1349 
1350 /* Process any previous ticks, then bump cnt by one (times scale). */
1351 static void fmeter_markevent(struct fmeter *fmp)
1352 {
1353  spin_lock(&fmp->lock);
1354  fmeter_update(fmp);
1355  fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
1356  spin_unlock(&fmp->lock);
1357 }
1358 
1359 /* Process any previous ticks, then return current value. */
1360 static int fmeter_getrate(struct fmeter *fmp)
1361 {
1362  int val;
1363 
1364  spin_lock(&fmp->lock);
1365  fmeter_update(fmp);
1366  val = fmp->val;
1367  spin_unlock(&fmp->lock);
1368  return val;
1369 }
1370 
1371 /*
1372  * Protected by cgroup_lock. The nodemasks must be stored globally because
1373  * dynamically allocating them is not allowed in can_attach, and they must
1374  * persist until attach.
1375  */
1376 static cpumask_var_t cpus_attach;
1377 static nodemask_t cpuset_attach_nodemask_from;
1378 static nodemask_t cpuset_attach_nodemask_to;
1379 
1380 /* Called by cgroups to determine if a cpuset is usable; cgroup_mutex held */
1381 static int cpuset_can_attach(struct cgroup *cgrp, struct cgroup_taskset *tset)
1382 {
1383  struct cpuset *cs = cgroup_cs(cgrp);
1384  struct task_struct *task;
1385  int ret;
1386 
1387  if (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed))
1388  return -ENOSPC;
1389 
1390  cgroup_taskset_for_each(task, cgrp, tset) {
1391  /*
1392  * Kthreads bound to specific cpus cannot be moved to a new
1393  * cpuset; we cannot change their cpu affinity and
1394  * isolating such threads by their set of allowed nodes is
1395  * unnecessary. Thus, cpusets are not applicable for such
1396  * threads. This prevents checking for success of
1397  * set_cpus_allowed_ptr() on all attached tasks before
1398  * cpus_allowed may be changed.
1399  */
1400  if (task->flags & PF_THREAD_BOUND)
1401  return -EINVAL;
1402  if ((ret = security_task_setscheduler(task)))
1403  return ret;
1404  }
1405 
1406  /* prepare for attach */
1407  if (cs == &top_cpuset)
1408  cpumask_copy(cpus_attach, cpu_possible_mask);
1409  else
1410  guarantee_online_cpus(cs, cpus_attach);
1411 
1412  guarantee_online_mems(cs, &cpuset_attach_nodemask_to);
1413 
1414  return 0;
1415 }
1416 
1417 static void cpuset_attach(struct cgroup *cgrp, struct cgroup_taskset *tset)
1418 {
1419  struct mm_struct *mm;
1420  struct task_struct *task;
1421  struct task_struct *leader = cgroup_taskset_first(tset);
1422  struct cgroup *oldcgrp = cgroup_taskset_cur_cgroup(tset);
1423  struct cpuset *cs = cgroup_cs(cgrp);
1424  struct cpuset *oldcs = cgroup_cs(oldcgrp);
1425 
1426  cgroup_taskset_for_each(task, cgrp, tset) {
1427  /*
1428  * can_attach beforehand should guarantee that this doesn't
1429  * fail. TODO: have a better way to handle failure here
1430  */
1431  WARN_ON_ONCE(set_cpus_allowed_ptr(task, cpus_attach));
1432 
1433  cpuset_change_task_nodemask(task, &cpuset_attach_nodemask_to);
1434  cpuset_update_task_spread_flag(cs, task);
1435  }
1436 
1437  /*
1438  * Change mm, possibly for multiple threads in a threadgroup. This is
1439  * expensive and may sleep.
1440  */
1441  cpuset_attach_nodemask_from = oldcs->mems_allowed;
1442  cpuset_attach_nodemask_to = cs->mems_allowed;
1443  mm = get_task_mm(leader);
1444  if (mm) {
1445  mpol_rebind_mm(mm, &cpuset_attach_nodemask_to);
1446  if (is_memory_migrate(cs))
1447  cpuset_migrate_mm(mm, &cpuset_attach_nodemask_from,
1448  &cpuset_attach_nodemask_to);
1449  mmput(mm);
1450  }
1451 }
1452 
1453 /* The various types of files and directories in a cpuset file system */
1454 
1455 typedef enum {
1469 
1470 static int cpuset_write_u64(struct cgroup *cgrp, struct cftype *cft, u64 val)
1471 {
1472  int retval = 0;
1473  struct cpuset *cs = cgroup_cs(cgrp);
1474  cpuset_filetype_t type = cft->private;
1475 
1476  if (!cgroup_lock_live_group(cgrp))
1477  return -ENODEV;
1478 
1479  switch (type) {
1480  case FILE_CPU_EXCLUSIVE:
1481  retval = update_flag(CS_CPU_EXCLUSIVE, cs, val);
1482  break;
1483  case FILE_MEM_EXCLUSIVE:
1484  retval = update_flag(CS_MEM_EXCLUSIVE, cs, val);
1485  break;
1486  case FILE_MEM_HARDWALL:
1487  retval = update_flag(CS_MEM_HARDWALL, cs, val);
1488  break;
1490  retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, val);
1491  break;
1492  case FILE_MEMORY_MIGRATE:
1493  retval = update_flag(CS_MEMORY_MIGRATE, cs, val);
1494  break;
1496  cpuset_memory_pressure_enabled = !!val;
1497  break;
1498  case FILE_MEMORY_PRESSURE:
1499  retval = -EACCES;
1500  break;
1501  case FILE_SPREAD_PAGE:
1502  retval = update_flag(CS_SPREAD_PAGE, cs, val);
1503  break;
1504  case FILE_SPREAD_SLAB:
1505  retval = update_flag(CS_SPREAD_SLAB, cs, val);
1506  break;
1507  default:
1508  retval = -EINVAL;
1509  break;
1510  }
1511  cgroup_unlock();
1512  return retval;
1513 }
1514 
1515 static int cpuset_write_s64(struct cgroup *cgrp, struct cftype *cft, s64 val)
1516 {
1517  int retval = 0;
1518  struct cpuset *cs = cgroup_cs(cgrp);
1519  cpuset_filetype_t type = cft->private;
1520 
1521  if (!cgroup_lock_live_group(cgrp))
1522  return -ENODEV;
1523 
1524  switch (type) {
1526  retval = update_relax_domain_level(cs, val);
1527  break;
1528  default:
1529  retval = -EINVAL;
1530  break;
1531  }
1532  cgroup_unlock();
1533  return retval;
1534 }
1535 
1536 /*
1537  * Common handling for a write to a "cpus" or "mems" file.
1538  */
1539 static int cpuset_write_resmask(struct cgroup *cgrp, struct cftype *cft,
1540  const char *buf)
1541 {
1542  int retval = 0;
1543  struct cpuset *cs = cgroup_cs(cgrp);
1544  struct cpuset *trialcs;
1545 
1546  if (!cgroup_lock_live_group(cgrp))
1547  return -ENODEV;
1548 
1549  trialcs = alloc_trial_cpuset(cs);
1550  if (!trialcs) {
1551  retval = -ENOMEM;
1552  goto out;
1553  }
1554 
1555  switch (cft->private) {
1556  case FILE_CPULIST:
1557  retval = update_cpumask(cs, trialcs, buf);
1558  break;
1559  case FILE_MEMLIST:
1560  retval = update_nodemask(cs, trialcs, buf);
1561  break;
1562  default:
1563  retval = -EINVAL;
1564  break;
1565  }
1566 
1567  free_trial_cpuset(trialcs);
1568 out:
1569  cgroup_unlock();
1570  return retval;
1571 }
1572 
1573 /*
1574  * These ascii lists should be read in a single call, by using a user
1575  * buffer large enough to hold the entire map. If read in smaller
1576  * chunks, there is no guarantee of atomicity. Since the display format
1577  * used, list of ranges of sequential numbers, is variable length,
1578  * and since these maps can change value dynamically, one could read
1579  * gibberish by doing partial reads while a list was changing.
1580  * A single large read to a buffer that crosses a page boundary is
1581  * ok, because the result being copied to user land is not recomputed
1582  * across a page fault.
1583  */
1584 
1585 static size_t cpuset_sprintf_cpulist(char *page, struct cpuset *cs)
1586 {
1587  size_t count;
1588 
1589  mutex_lock(&callback_mutex);
1590  count = cpulist_scnprintf(page, PAGE_SIZE, cs->cpus_allowed);
1591  mutex_unlock(&callback_mutex);
1592 
1593  return count;
1594 }
1595 
1596 static size_t cpuset_sprintf_memlist(char *page, struct cpuset *cs)
1597 {
1598  size_t count;
1599 
1600  mutex_lock(&callback_mutex);
1601  count = nodelist_scnprintf(page, PAGE_SIZE, cs->mems_allowed);
1602  mutex_unlock(&callback_mutex);
1603 
1604  return count;
1605 }
1606 
1607 static ssize_t cpuset_common_file_read(struct cgroup *cont,
1608  struct cftype *cft,
1609  struct file *file,
1610  char __user *buf,
1611  size_t nbytes, loff_t *ppos)
1612 {
1613  struct cpuset *cs = cgroup_cs(cont);
1614  cpuset_filetype_t type = cft->private;
1615  char *page;
1616  ssize_t retval = 0;
1617  char *s;
1618 
1619  if (!(page = (char *)__get_free_page(GFP_TEMPORARY)))
1620  return -ENOMEM;
1621 
1622  s = page;
1623 
1624  switch (type) {
1625  case FILE_CPULIST:
1626  s += cpuset_sprintf_cpulist(s, cs);
1627  break;
1628  case FILE_MEMLIST:
1629  s += cpuset_sprintf_memlist(s, cs);
1630  break;
1631  default:
1632  retval = -EINVAL;
1633  goto out;
1634  }
1635  *s++ = '\n';
1636 
1637  retval = simple_read_from_buffer(buf, nbytes, ppos, page, s - page);
1638 out:
1639  free_page((unsigned long)page);
1640  return retval;
1641 }
1642 
1643 static u64 cpuset_read_u64(struct cgroup *cont, struct cftype *cft)
1644 {
1645  struct cpuset *cs = cgroup_cs(cont);
1646  cpuset_filetype_t type = cft->private;
1647  switch (type) {
1648  case FILE_CPU_EXCLUSIVE:
1649  return is_cpu_exclusive(cs);
1650  case FILE_MEM_EXCLUSIVE:
1651  return is_mem_exclusive(cs);
1652  case FILE_MEM_HARDWALL:
1653  return is_mem_hardwall(cs);
1655  return is_sched_load_balance(cs);
1656  case FILE_MEMORY_MIGRATE:
1657  return is_memory_migrate(cs);
1659  return cpuset_memory_pressure_enabled;
1660  case FILE_MEMORY_PRESSURE:
1661  return fmeter_getrate(&cs->fmeter);
1662  case FILE_SPREAD_PAGE:
1663  return is_spread_page(cs);
1664  case FILE_SPREAD_SLAB:
1665  return is_spread_slab(cs);
1666  default:
1667  BUG();
1668  }
1669 
1670  /* Unreachable but makes gcc happy */
1671  return 0;
1672 }
1673 
1674 static s64 cpuset_read_s64(struct cgroup *cont, struct cftype *cft)
1675 {
1676  struct cpuset *cs = cgroup_cs(cont);
1677  cpuset_filetype_t type = cft->private;
1678  switch (type) {
1680  return cs->relax_domain_level;
1681  default:
1682  BUG();
1683  }
1684 
1685  /* Unrechable but makes gcc happy */
1686  return 0;
1687 }
1688 
1689 
1690 /*
1691  * for the common functions, 'private' gives the type of file
1692  */
1693 
1694 static struct cftype files[] = {
1695  {
1696  .name = "cpus",
1697  .read = cpuset_common_file_read,
1698  .write_string = cpuset_write_resmask,
1699  .max_write_len = (100U + 6 * NR_CPUS),
1700  .private = FILE_CPULIST,
1701  },
1702 
1703  {
1704  .name = "mems",
1705  .read = cpuset_common_file_read,
1706  .write_string = cpuset_write_resmask,
1707  .max_write_len = (100U + 6 * MAX_NUMNODES),
1708  .private = FILE_MEMLIST,
1709  },
1710 
1711  {
1712  .name = "cpu_exclusive",
1713  .read_u64 = cpuset_read_u64,
1714  .write_u64 = cpuset_write_u64,
1715  .private = FILE_CPU_EXCLUSIVE,
1716  },
1717 
1718  {
1719  .name = "mem_exclusive",
1720  .read_u64 = cpuset_read_u64,
1721  .write_u64 = cpuset_write_u64,
1722  .private = FILE_MEM_EXCLUSIVE,
1723  },
1724 
1725  {
1726  .name = "mem_hardwall",
1727  .read_u64 = cpuset_read_u64,
1728  .write_u64 = cpuset_write_u64,
1729  .private = FILE_MEM_HARDWALL,
1730  },
1731 
1732  {
1733  .name = "sched_load_balance",
1734  .read_u64 = cpuset_read_u64,
1735  .write_u64 = cpuset_write_u64,
1736  .private = FILE_SCHED_LOAD_BALANCE,
1737  },
1738 
1739  {
1740  .name = "sched_relax_domain_level",
1741  .read_s64 = cpuset_read_s64,
1742  .write_s64 = cpuset_write_s64,
1743  .private = FILE_SCHED_RELAX_DOMAIN_LEVEL,
1744  },
1745 
1746  {
1747  .name = "memory_migrate",
1748  .read_u64 = cpuset_read_u64,
1749  .write_u64 = cpuset_write_u64,
1750  .private = FILE_MEMORY_MIGRATE,
1751  },
1752 
1753  {
1754  .name = "memory_pressure",
1755  .read_u64 = cpuset_read_u64,
1756  .write_u64 = cpuset_write_u64,
1757  .private = FILE_MEMORY_PRESSURE,
1758  .mode = S_IRUGO,
1759  },
1760 
1761  {
1762  .name = "memory_spread_page",
1763  .read_u64 = cpuset_read_u64,
1764  .write_u64 = cpuset_write_u64,
1765  .private = FILE_SPREAD_PAGE,
1766  },
1767 
1768  {
1769  .name = "memory_spread_slab",
1770  .read_u64 = cpuset_read_u64,
1771  .write_u64 = cpuset_write_u64,
1772  .private = FILE_SPREAD_SLAB,
1773  },
1774 
1775  {
1776  .name = "memory_pressure_enabled",
1777  .flags = CFTYPE_ONLY_ON_ROOT,
1778  .read_u64 = cpuset_read_u64,
1779  .write_u64 = cpuset_write_u64,
1780  .private = FILE_MEMORY_PRESSURE_ENABLED,
1781  },
1782 
1783  { } /* terminate */
1784 };
1785 
1786 /*
1787  * post_clone() is called during cgroup_create() when the
1788  * clone_children mount argument was specified. The cgroup
1789  * can not yet have any tasks.
1790  *
1791  * Currently we refuse to set up the cgroup - thereby
1792  * refusing the task to be entered, and as a result refusing
1793  * the sys_unshare() or clone() which initiated it - if any
1794  * sibling cpusets have exclusive cpus or mem.
1795  *
1796  * If this becomes a problem for some users who wish to
1797  * allow that scenario, then cpuset_post_clone() could be
1798  * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
1799  * (and likewise for mems) to the new cgroup. Called with cgroup_mutex
1800  * held.
1801  */
1802 static void cpuset_post_clone(struct cgroup *cgroup)
1803 {
1804  struct cgroup *parent, *child;
1805  struct cpuset *cs, *parent_cs;
1806 
1807  parent = cgroup->parent;
1808  list_for_each_entry(child, &parent->children, sibling) {
1809  cs = cgroup_cs(child);
1810  if (is_mem_exclusive(cs) || is_cpu_exclusive(cs))
1811  return;
1812  }
1813  cs = cgroup_cs(cgroup);
1814  parent_cs = cgroup_cs(parent);
1815 
1816  mutex_lock(&callback_mutex);
1817  cs->mems_allowed = parent_cs->mems_allowed;
1818  cpumask_copy(cs->cpus_allowed, parent_cs->cpus_allowed);
1819  mutex_unlock(&callback_mutex);
1820  return;
1821 }
1822 
1823 /*
1824  * cpuset_create - create a cpuset
1825  * cont: control group that the new cpuset will be part of
1826  */
1827 
1828 static struct cgroup_subsys_state *cpuset_create(struct cgroup *cont)
1829 {
1830  struct cpuset *cs;
1831  struct cpuset *parent;
1832 
1833  if (!cont->parent) {
1834  return &top_cpuset.css;
1835  }
1836  parent = cgroup_cs(cont->parent);
1837  cs = kmalloc(sizeof(*cs), GFP_KERNEL);
1838  if (!cs)
1839  return ERR_PTR(-ENOMEM);
1840  if (!alloc_cpumask_var(&cs->cpus_allowed, GFP_KERNEL)) {
1841  kfree(cs);
1842  return ERR_PTR(-ENOMEM);
1843  }
1844 
1845  cs->flags = 0;
1846  if (is_spread_page(parent))
1847  set_bit(CS_SPREAD_PAGE, &cs->flags);
1848  if (is_spread_slab(parent))
1849  set_bit(CS_SPREAD_SLAB, &cs->flags);
1851  cpumask_clear(cs->cpus_allowed);
1853  fmeter_init(&cs->fmeter);
1854  cs->relax_domain_level = -1;
1855 
1856  cs->parent = parent;
1857  number_of_cpusets++;
1858  return &cs->css ;
1859 }
1860 
1861 /*
1862  * If the cpuset being removed has its flag 'sched_load_balance'
1863  * enabled, then simulate turning sched_load_balance off, which
1864  * will call async_rebuild_sched_domains().
1865  */
1866 
1867 static void cpuset_destroy(struct cgroup *cont)
1868 {
1869  struct cpuset *cs = cgroup_cs(cont);
1870 
1871  if (is_sched_load_balance(cs))
1872  update_flag(CS_SCHED_LOAD_BALANCE, cs, 0);
1873 
1874  number_of_cpusets--;
1875  free_cpumask_var(cs->cpus_allowed);
1876  kfree(cs);
1877 }
1878 
1879 struct cgroup_subsys cpuset_subsys = {
1880  .name = "cpuset",
1881  .create = cpuset_create,
1882  .destroy = cpuset_destroy,
1883  .can_attach = cpuset_can_attach,
1884  .attach = cpuset_attach,
1885  .post_clone = cpuset_post_clone,
1886  .subsys_id = cpuset_subsys_id,
1887  .base_cftypes = files,
1888  .early_init = 1,
1889 };
1890 
1898 {
1899  int err = 0;
1900 
1901  if (!alloc_cpumask_var(&top_cpuset.cpus_allowed, GFP_KERNEL))
1902  BUG();
1903 
1904  cpumask_setall(top_cpuset.cpus_allowed);
1905  nodes_setall(top_cpuset.mems_allowed);
1906 
1907  fmeter_init(&top_cpuset.fmeter);
1908  set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags);
1909  top_cpuset.relax_domain_level = -1;
1910 
1911  err = register_filesystem(&cpuset_fs_type);
1912  if (err < 0)
1913  return err;
1914 
1915  if (!alloc_cpumask_var(&cpus_attach, GFP_KERNEL))
1916  BUG();
1917 
1918  number_of_cpusets = 1;
1919  return 0;
1920 }
1921 
1930 static void cpuset_do_move_task(struct task_struct *tsk,
1931  struct cgroup_scanner *scan)
1932 {
1933  struct cgroup *new_cgroup = scan->data;
1934 
1935  cgroup_attach_task(new_cgroup, tsk);
1936 }
1937 
1949 static void move_member_tasks_to_cpuset(struct cpuset *from, struct cpuset *to)
1950 {
1951  struct cgroup_scanner scan;
1952 
1953  scan.cg = from->css.cgroup;
1954  scan.test_task = NULL; /* select all tasks in cgroup */
1955  scan.process_task = cpuset_do_move_task;
1956  scan.heap = NULL;
1957  scan.data = to->css.cgroup;
1958 
1959  if (cgroup_scan_tasks(&scan))
1960  printk(KERN_ERR "move_member_tasks_to_cpuset: "
1961  "cgroup_scan_tasks failed\n");
1962 }
1963 
1964 /*
1965  * If CPU and/or memory hotplug handlers, below, unplug any CPUs
1966  * or memory nodes, we need to walk over the cpuset hierarchy,
1967  * removing that CPU or node from all cpusets. If this removes the
1968  * last CPU or node from a cpuset, then move the tasks in the empty
1969  * cpuset to its next-highest non-empty parent.
1970  *
1971  * Called with cgroup_mutex held
1972  * callback_mutex must not be held, as cpuset_attach() will take it.
1973  */
1974 static void remove_tasks_in_empty_cpuset(struct cpuset *cs)
1975 {
1976  struct cpuset *parent;
1977 
1978  /*
1979  * The cgroup's css_sets list is in use if there are tasks
1980  * in the cpuset; the list is empty if there are none;
1981  * the cs->css.refcnt seems always 0.
1982  */
1983  if (list_empty(&cs->css.cgroup->css_sets))
1984  return;
1985 
1986  /*
1987  * Find its next-highest non-empty parent, (top cpuset
1988  * has online cpus, so can't be empty).
1989  */
1990  parent = cs->parent;
1991  while (cpumask_empty(parent->cpus_allowed) ||
1992  nodes_empty(parent->mems_allowed))
1993  parent = parent->parent;
1994 
1995  move_member_tasks_to_cpuset(cs, parent);
1996 }
1997 
1998 /*
1999  * Helper function to traverse cpusets.
2000  * It can be used to walk the cpuset tree from top to bottom, completing
2001  * one layer before dropping down to the next (thus always processing a
2002  * node before any of its children).
2003  */
2004 static struct cpuset *cpuset_next(struct list_head *queue)
2005 {
2006  struct cpuset *cp;
2007  struct cpuset *child; /* scans child cpusets of cp */
2008  struct cgroup *cont;
2009 
2010  if (list_empty(queue))
2011  return NULL;
2012 
2013  cp = list_first_entry(queue, struct cpuset, stack_list);
2014  list_del(queue->next);
2015  list_for_each_entry(cont, &cp->css.cgroup->children, sibling) {
2016  child = cgroup_cs(cont);
2017  list_add_tail(&child->stack_list, queue);
2018  }
2019 
2020  return cp;
2021 }
2022 
2023 
2024 /*
2025  * Walk the specified cpuset subtree upon a hotplug operation (CPU/Memory
2026  * online/offline) and update the cpusets accordingly.
2027  * For regular CPU/Mem hotplug, look for empty cpusets; the tasks of such
2028  * cpuset must be moved to a parent cpuset.
2029  *
2030  * Called with cgroup_mutex held. We take callback_mutex to modify
2031  * cpus_allowed and mems_allowed.
2032  *
2033  * This walk processes the tree from top to bottom, completing one layer
2034  * before dropping down to the next. It always processes a node before
2035  * any of its children.
2036  *
2037  * In the case of memory hot-unplug, it will remove nodes from N_HIGH_MEMORY
2038  * if all present pages from a node are offlined.
2039  */
2040 static void
2041 scan_cpusets_upon_hotplug(struct cpuset *root, enum hotplug_event event)
2042 {
2043  LIST_HEAD(queue);
2044  struct cpuset *cp; /* scans cpusets being updated */
2045  static nodemask_t oldmems; /* protected by cgroup_mutex */
2046 
2047  list_add_tail((struct list_head *)&root->stack_list, &queue);
2048 
2049  switch (event) {
2050  case CPUSET_CPU_OFFLINE:
2051  while ((cp = cpuset_next(&queue)) != NULL) {
2052 
2053  /* Continue past cpusets with all cpus online */
2054  if (cpumask_subset(cp->cpus_allowed, cpu_active_mask))
2055  continue;
2056 
2057  /* Remove offline cpus from this cpuset. */
2058  mutex_lock(&callback_mutex);
2059  cpumask_and(cp->cpus_allowed, cp->cpus_allowed,
2060  cpu_active_mask);
2061  mutex_unlock(&callback_mutex);
2062 
2063  /* Move tasks from the empty cpuset to a parent */
2064  if (cpumask_empty(cp->cpus_allowed))
2065  remove_tasks_in_empty_cpuset(cp);
2066  else
2067  update_tasks_cpumask(cp, NULL);
2068  }
2069  break;
2070 
2071  case CPUSET_MEM_OFFLINE:
2072  while ((cp = cpuset_next(&queue)) != NULL) {
2073 
2074  /* Continue past cpusets with all mems online */
2075  if (nodes_subset(cp->mems_allowed,
2076  node_states[N_HIGH_MEMORY]))
2077  continue;
2078 
2079  oldmems = cp->mems_allowed;
2080 
2081  /* Remove offline mems from this cpuset. */
2082  mutex_lock(&callback_mutex);
2084  node_states[N_HIGH_MEMORY]);
2085  mutex_unlock(&callback_mutex);
2086 
2087  /* Move tasks from the empty cpuset to a parent */
2088  if (nodes_empty(cp->mems_allowed))
2089  remove_tasks_in_empty_cpuset(cp);
2090  else
2091  update_tasks_nodemask(cp, &oldmems, NULL);
2092  }
2093  }
2094 }
2095 
2096 /*
2097  * The top_cpuset tracks what CPUs and Memory Nodes are online,
2098  * period. This is necessary in order to make cpusets transparent
2099  * (of no affect) on systems that are actively using CPU hotplug
2100  * but making no active use of cpusets.
2101  *
2102  * The only exception to this is suspend/resume, where we don't
2103  * modify cpusets at all.
2104  *
2105  * This routine ensures that top_cpuset.cpus_allowed tracks
2106  * cpu_active_mask on each CPU hotplug (cpuhp) event.
2107  *
2108  * Called within get_online_cpus(). Needs to call cgroup_lock()
2109  * before calling generate_sched_domains().
2110  *
2111  * @cpu_online: Indicates whether this is a CPU online event (true) or
2112  * a CPU offline event (false).
2113  */
2115 {
2116  struct sched_domain_attr *attr;
2117  cpumask_var_t *doms;
2118  int ndoms;
2119 
2120  cgroup_lock();
2121  mutex_lock(&callback_mutex);
2122  cpumask_copy(top_cpuset.cpus_allowed, cpu_active_mask);
2123  mutex_unlock(&callback_mutex);
2124 
2125  if (!cpu_online)
2126  scan_cpusets_upon_hotplug(&top_cpuset, CPUSET_CPU_OFFLINE);
2127 
2128  ndoms = generate_sched_domains(&doms, &attr);
2129  cgroup_unlock();
2130 
2131  /* Have scheduler rebuild the domains */
2132  partition_sched_domains(ndoms, doms, attr);
2133 }
2134 
2135 #ifdef CONFIG_MEMORY_HOTPLUG
2136 /*
2137  * Keep top_cpuset.mems_allowed tracking node_states[N_HIGH_MEMORY].
2138  * Call this routine anytime after node_states[N_HIGH_MEMORY] changes.
2139  * See cpuset_update_active_cpus() for CPU hotplug handling.
2140  */
2141 static int cpuset_track_online_nodes(struct notifier_block *self,
2142  unsigned long action, void *arg)
2143 {
2144  static nodemask_t oldmems; /* protected by cgroup_mutex */
2145 
2146  cgroup_lock();
2147  switch (action) {
2148  case MEM_ONLINE:
2149  oldmems = top_cpuset.mems_allowed;
2150  mutex_lock(&callback_mutex);
2151  top_cpuset.mems_allowed = node_states[N_HIGH_MEMORY];
2152  mutex_unlock(&callback_mutex);
2153  update_tasks_nodemask(&top_cpuset, &oldmems, NULL);
2154  break;
2155  case MEM_OFFLINE:
2156  /*
2157  * needn't update top_cpuset.mems_allowed explicitly because
2158  * scan_cpusets_upon_hotplug() will update it.
2159  */
2160  scan_cpusets_upon_hotplug(&top_cpuset, CPUSET_MEM_OFFLINE);
2161  break;
2162  default:
2163  break;
2164  }
2165  cgroup_unlock();
2166 
2167  return NOTIFY_OK;
2168 }
2169 #endif
2170 
2178 {
2179  cpumask_copy(top_cpuset.cpus_allowed, cpu_active_mask);
2180  top_cpuset.mems_allowed = node_states[N_HIGH_MEMORY];
2181 
2182  hotplug_memory_notifier(cpuset_track_online_nodes, 10);
2183 
2184  cpuset_wq = create_singlethread_workqueue("cpuset");
2185  BUG_ON(!cpuset_wq);
2186 }
2187 
2199 void cpuset_cpus_allowed(struct task_struct *tsk, struct cpumask *pmask)
2200 {
2201  mutex_lock(&callback_mutex);
2202  task_lock(tsk);
2203  guarantee_online_cpus(task_cs(tsk), pmask);
2204  task_unlock(tsk);
2205  mutex_unlock(&callback_mutex);
2206 }
2207 
2209 {
2210  const struct cpuset *cs;
2211 
2212  rcu_read_lock();
2213  cs = task_cs(tsk);
2214  if (cs)
2215  do_set_cpus_allowed(tsk, cs->cpus_allowed);
2216  rcu_read_unlock();
2217 
2218  /*
2219  * We own tsk->cpus_allowed, nobody can change it under us.
2220  *
2221  * But we used cs && cs->cpus_allowed lockless and thus can
2222  * race with cgroup_attach_task() or update_cpumask() and get
2223  * the wrong tsk->cpus_allowed. However, both cases imply the
2224  * subsequent cpuset_change_cpumask()->set_cpus_allowed_ptr()
2225  * which takes task_rq_lock().
2226  *
2227  * If we are called after it dropped the lock we must see all
2228  * changes in tsk_cs()->cpus_allowed. Otherwise we can temporary
2229  * set any mask even if it is not right from task_cs() pov,
2230  * the pending set_cpus_allowed_ptr() will fix things.
2231  *
2232  * select_fallback_rq() will fix things ups and set cpu_possible_mask
2233  * if required.
2234  */
2235 }
2236 
2238 {
2239  nodes_setall(current->mems_allowed);
2240 }
2241 
2253 {
2254  nodemask_t mask;
2255 
2256  mutex_lock(&callback_mutex);
2257  task_lock(tsk);
2258  guarantee_online_mems(task_cs(tsk), &mask);
2259  task_unlock(tsk);
2260  mutex_unlock(&callback_mutex);
2261 
2262  return mask;
2263 }
2264 
2272 {
2273  return nodes_intersects(*nodemask, current->mems_allowed);
2274 }
2275 
2276 /*
2277  * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
2278  * mem_hardwall ancestor to the specified cpuset. Call holding
2279  * callback_mutex. If no ancestor is mem_exclusive or mem_hardwall
2280  * (an unusual configuration), then returns the root cpuset.
2281  */
2282 static const struct cpuset *nearest_hardwall_ancestor(const struct cpuset *cs)
2283 {
2284  while (!(is_mem_exclusive(cs) || is_mem_hardwall(cs)) && cs->parent)
2285  cs = cs->parent;
2286  return cs;
2287 }
2288 
2351 {
2352  const struct cpuset *cs; /* current cpuset ancestors */
2353  int allowed; /* is allocation in zone z allowed? */
2354 
2355  if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
2356  return 1;
2357  might_sleep_if(!(gfp_mask & __GFP_HARDWALL));
2358  if (node_isset(node, current->mems_allowed))
2359  return 1;
2360  /*
2361  * Allow tasks that have access to memory reserves because they have
2362  * been OOM killed to get memory anywhere.
2363  */
2364  if (unlikely(test_thread_flag(TIF_MEMDIE)))
2365  return 1;
2366  if (gfp_mask & __GFP_HARDWALL) /* If hardwall request, stop here */
2367  return 0;
2368 
2369  if (current->flags & PF_EXITING) /* Let dying task have memory */
2370  return 1;
2371 
2372  /* Not hardwall and node outside mems_allowed: scan up cpusets */
2373  mutex_lock(&callback_mutex);
2374 
2375  task_lock(current);
2376  cs = nearest_hardwall_ancestor(task_cs(current));
2377  task_unlock(current);
2378 
2379  allowed = node_isset(node, cs->mems_allowed);
2380  mutex_unlock(&callback_mutex);
2381  return allowed;
2382 }
2383 
2384 /*
2385  * cpuset_node_allowed_hardwall - Can we allocate on a memory node?
2386  * @node: is this an allowed node?
2387  * @gfp_mask: memory allocation flags
2388  *
2389  * If we're in interrupt, yes, we can always allocate. If __GFP_THISNODE is
2390  * set, yes, we can always allocate. If node is in our task's mems_allowed,
2391  * yes. If the task has been OOM killed and has access to memory reserves as
2392  * specified by the TIF_MEMDIE flag, yes.
2393  * Otherwise, no.
2394  *
2395  * The __GFP_THISNODE placement logic is really handled elsewhere,
2396  * by forcibly using a zonelist starting at a specified node, and by
2397  * (in get_page_from_freelist()) refusing to consider the zones for
2398  * any node on the zonelist except the first. By the time any such
2399  * calls get to this routine, we should just shut up and say 'yes'.
2400  *
2401  * Unlike the cpuset_node_allowed_softwall() variant, above,
2402  * this variant requires that the node be in the current task's
2403  * mems_allowed or that we're in interrupt. It does not scan up the
2404  * cpuset hierarchy for the nearest enclosing mem_exclusive cpuset.
2405  * It never sleeps.
2406  */
2408 {
2409  if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
2410  return 1;
2411  if (node_isset(node, current->mems_allowed))
2412  return 1;
2413  /*
2414  * Allow tasks that have access to memory reserves because they have
2415  * been OOM killed to get memory anywhere.
2416  */
2417  if (unlikely(test_thread_flag(TIF_MEMDIE)))
2418  return 1;
2419  return 0;
2420 }
2421 
2428 void cpuset_unlock(void)
2429 {
2430  mutex_unlock(&callback_mutex);
2431 }
2432 
2460 static int cpuset_spread_node(int *rotor)
2461 {
2462  int node;
2463 
2464  node = next_node(*rotor, current->mems_allowed);
2465  if (node == MAX_NUMNODES)
2466  node = first_node(current->mems_allowed);
2467  *rotor = node;
2468  return node;
2469 }
2470 
2472 {
2473  if (current->cpuset_mem_spread_rotor == NUMA_NO_NODE)
2474  current->cpuset_mem_spread_rotor =
2475  node_random(&current->mems_allowed);
2476 
2477  return cpuset_spread_node(&current->cpuset_mem_spread_rotor);
2478 }
2479 
2481 {
2482  if (current->cpuset_slab_spread_rotor == NUMA_NO_NODE)
2483  current->cpuset_slab_spread_rotor =
2484  node_random(&current->mems_allowed);
2485 
2486  return cpuset_spread_node(&current->cpuset_slab_spread_rotor);
2487 }
2488 
2490 
2503  const struct task_struct *tsk2)
2504 {
2505  return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed);
2506 }
2507 
2517 {
2518  struct dentry *dentry;
2519 
2520  dentry = task_cs(tsk)->css.cgroup->dentry;
2521  spin_lock(&cpuset_buffer_lock);
2522  snprintf(cpuset_name, CPUSET_NAME_LEN,
2523  dentry ? (const char *)dentry->d_name.name : "/");
2524  nodelist_scnprintf(cpuset_nodelist, CPUSET_NODELIST_LEN,
2525  tsk->mems_allowed);
2526  printk(KERN_INFO "%s cpuset=%s mems_allowed=%s\n",
2527  tsk->comm, cpuset_name, cpuset_nodelist);
2528  spin_unlock(&cpuset_buffer_lock);
2529 }
2530 
2531 /*
2532  * Collection of memory_pressure is suppressed unless
2533  * this flag is enabled by writing "1" to the special
2534  * cpuset file 'memory_pressure_enabled' in the root cpuset.
2535  */
2536 
2537 int cpuset_memory_pressure_enabled __read_mostly;
2538 
2558 {
2559  task_lock(current);
2560  fmeter_markevent(&task_cs(current)->fmeter);
2561  task_unlock(current);
2562 }
2563 
2564 #ifdef CONFIG_PROC_PID_CPUSET
2565 /*
2566  * proc_cpuset_show()
2567  * - Print tasks cpuset path into seq_file.
2568  * - Used for /proc/<pid>/cpuset.
2569  * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
2570  * doesn't really matter if tsk->cpuset changes after we read it,
2571  * and we take cgroup_mutex, keeping cpuset_attach() from changing it
2572  * anyway.
2573  */
2574 static int proc_cpuset_show(struct seq_file *m, void *unused_v)
2575 {
2576  struct pid *pid;
2577  struct task_struct *tsk;
2578  char *buf;
2579  struct cgroup_subsys_state *css;
2580  int retval;
2581 
2582  retval = -ENOMEM;
2583  buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
2584  if (!buf)
2585  goto out;
2586 
2587  retval = -ESRCH;
2588  pid = m->private;
2589  tsk = get_pid_task(pid, PIDTYPE_PID);
2590  if (!tsk)
2591  goto out_free;
2592 
2593  retval = -EINVAL;
2594  cgroup_lock();
2595  css = task_subsys_state(tsk, cpuset_subsys_id);
2596  retval = cgroup_path(css->cgroup, buf, PAGE_SIZE);
2597  if (retval < 0)
2598  goto out_unlock;
2599  seq_puts(m, buf);
2600  seq_putc(m, '\n');
2601 out_unlock:
2602  cgroup_unlock();
2603  put_task_struct(tsk);
2604 out_free:
2605  kfree(buf);
2606 out:
2607  return retval;
2608 }
2609 
2610 static int cpuset_open(struct inode *inode, struct file *file)
2611 {
2612  struct pid *pid = PROC_I(inode)->pid;
2613  return single_open(file, proc_cpuset_show, pid);
2614 }
2615 
2616 const struct file_operations proc_cpuset_operations = {
2617  .open = cpuset_open,
2618  .read = seq_read,
2619  .llseek = seq_lseek,
2620  .release = single_release,
2621 };
2622 #endif /* CONFIG_PROC_PID_CPUSET */
2623 
2624 /* Display task mems_allowed in /proc/<pid>/status file. */
2625 void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task)
2626 {
2627  seq_printf(m, "Mems_allowed:\t");
2628  seq_nodemask(m, &task->mems_allowed);
2629  seq_printf(m, "\n");
2630  seq_printf(m, "Mems_allowed_list:\t");
2631  seq_nodemask_list(m, &task->mems_allowed);
2632  seq_printf(m, "\n");
2633 }