cpuset.c 76 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767768769770771772773774775776777778779780781782783784785786787788789790791792793794795796797798799800801802803804805806807808809810811812813814815816817818819820821822823824825826827828829830831832833834835836837838839840841842843844845846847848849850851852853854855856857858859860861862863864865866867868869870871872873874875876877878879880881882883884885886887888889890891892893894895896897898899900901902903904905906907908909910911912913914915916917918919920921922923924925926927928929930931932933934935936937938939940941942943944945946947948949950951952953954955956957958959960961962963964965966967968969970971972973974975976977978979980981982983984985986987988989990991992993994995996997998999100010011002100310041005100610071008100910101011101210131014101510161017101810191020102110221023102410251026102710281029103010311032103310341035103610371038103910401041104210431044104510461047104810491050105110521053105410551056105710581059106010611062106310641065106610671068106910701071107210731074107510761077107810791080108110821083108410851086108710881089109010911092109310941095109610971098109911001101110211031104110511061107110811091110111111121113111411151116111711181119112011211122112311241125112611271128112911301131113211331134113511361137113811391140114111421143114411451146114711481149115011511152115311541155115611571158115911601161116211631164116511661167116811691170117111721173117411751176117711781179118011811182118311841185118611871188118911901191119211931194119511961197119811991200120112021203120412051206120712081209121012111212121312141215121612171218121912201221122212231224122512261227122812291230123112321233123412351236123712381239124012411242124312441245124612471248124912501251125212531254125512561257125812591260126112621263126412651266126712681269127012711272127312741275127612771278127912801281128212831284128512861287128812891290129112921293129412951296129712981299130013011302130313041305130613071308130913101311131213131314131513161317131813191320132113221323132413251326132713281329133013311332133313341335133613371338133913401341134213431344134513461347134813491350135113521353135413551356135713581359136013611362136313641365136613671368136913701371137213731374137513761377137813791380138113821383138413851386138713881389139013911392139313941395139613971398139914001401140214031404140514061407140814091410141114121413141414151416141714181419142014211422142314241425142614271428142914301431143214331434143514361437143814391440144114421443144414451446144714481449145014511452145314541455145614571458145914601461146214631464146514661467146814691470147114721473147414751476147714781479148014811482148314841485148614871488148914901491149214931494149514961497149814991500150115021503150415051506150715081509151015111512151315141515151615171518151915201521152215231524152515261527152815291530153115321533153415351536153715381539154015411542154315441545154615471548154915501551155215531554155515561557155815591560156115621563156415651566156715681569157015711572157315741575157615771578157915801581158215831584158515861587158815891590159115921593159415951596159715981599160016011602160316041605160616071608160916101611161216131614161516161617161816191620162116221623162416251626162716281629163016311632163316341635163616371638163916401641164216431644164516461647164816491650165116521653165416551656165716581659166016611662166316641665166616671668166916701671167216731674167516761677167816791680168116821683168416851686168716881689169016911692169316941695169616971698169917001701170217031704170517061707170817091710171117121713171417151716171717181719172017211722172317241725172617271728172917301731173217331734173517361737173817391740174117421743174417451746174717481749175017511752175317541755175617571758175917601761176217631764176517661767176817691770177117721773177417751776177717781779178017811782178317841785178617871788178917901791179217931794179517961797179817991800180118021803180418051806180718081809181018111812181318141815181618171818181918201821182218231824182518261827182818291830183118321833183418351836183718381839184018411842184318441845184618471848184918501851185218531854185518561857185818591860186118621863186418651866186718681869187018711872187318741875187618771878187918801881188218831884188518861887188818891890189118921893189418951896189718981899190019011902190319041905190619071908190919101911191219131914191519161917191819191920192119221923192419251926192719281929193019311932193319341935193619371938193919401941194219431944194519461947194819491950195119521953195419551956195719581959196019611962196319641965196619671968196919701971197219731974197519761977197819791980198119821983198419851986198719881989199019911992199319941995199619971998199920002001200220032004200520062007200820092010201120122013201420152016201720182019202020212022202320242025202620272028202920302031203220332034203520362037203820392040204120422043204420452046204720482049205020512052205320542055205620572058205920602061206220632064206520662067206820692070207120722073207420752076207720782079208020812082208320842085208620872088208920902091209220932094209520962097209820992100210121022103210421052106210721082109211021112112211321142115211621172118211921202121212221232124212521262127212821292130213121322133213421352136213721382139214021412142214321442145214621472148214921502151215221532154215521562157215821592160216121622163216421652166216721682169217021712172217321742175217621772178217921802181218221832184218521862187218821892190219121922193219421952196219721982199220022012202220322042205220622072208220922102211221222132214221522162217221822192220222122222223222422252226222722282229223022312232223322342235223622372238223922402241224222432244224522462247224822492250225122522253225422552256225722582259226022612262226322642265226622672268226922702271227222732274227522762277227822792280228122822283228422852286228722882289229022912292229322942295229622972298229923002301230223032304230523062307230823092310231123122313231423152316231723182319232023212322232323242325232623272328232923302331233223332334233523362337233823392340234123422343234423452346234723482349235023512352235323542355235623572358235923602361236223632364236523662367236823692370237123722373237423752376237723782379238023812382238323842385238623872388238923902391239223932394239523962397239823992400240124022403240424052406240724082409241024112412241324142415241624172418241924202421242224232424242524262427242824292430243124322433243424352436243724382439244024412442244324442445244624472448244924502451245224532454245524562457245824592460246124622463246424652466246724682469247024712472247324742475247624772478247924802481248224832484248524862487248824892490249124922493249424952496249724982499250025012502250325042505250625072508250925102511251225132514251525162517251825192520252125222523252425252526252725282529253025312532253325342535253625372538253925402541254225432544254525462547254825492550255125522553255425552556255725582559256025612562256325642565256625672568256925702571257225732574257525762577257825792580258125822583258425852586258725882589259025912592259325942595259625972598259926002601260226032604260526062607260826092610261126122613261426152616261726182619262026212622262326242625262626272628262926302631263226332634263526362637263826392640264126422643264426452646264726482649265026512652265326542655265626572658265926602661266226632664266526662667266826692670267126722673267426752676267726782679268026812682268326842685268626872688268926902691269226932694269526962697269826992700270127022703270427052706270727082709271027112712271327142715
  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. #include <linux/cpu.h>
  25. #include <linux/cpumask.h>
  26. #include <linux/cpuset.h>
  27. #include <linux/err.h>
  28. #include <linux/errno.h>
  29. #include <linux/file.h>
  30. #include <linux/fs.h>
  31. #include <linux/init.h>
  32. #include <linux/interrupt.h>
  33. #include <linux/kernel.h>
  34. #include <linux/kmod.h>
  35. #include <linux/list.h>
  36. #include <linux/mempolicy.h>
  37. #include <linux/mm.h>
  38. #include <linux/memory.h>
  39. #include <linux/export.h>
  40. #include <linux/mount.h>
  41. #include <linux/namei.h>
  42. #include <linux/pagemap.h>
  43. #include <linux/proc_fs.h>
  44. #include <linux/rcupdate.h>
  45. #include <linux/sched.h>
  46. #include <linux/seq_file.h>
  47. #include <linux/security.h>
  48. #include <linux/slab.h>
  49. #include <linux/spinlock.h>
  50. #include <linux/stat.h>
  51. #include <linux/string.h>
  52. #include <linux/time.h>
  53. #include <linux/backing-dev.h>
  54. #include <linux/sort.h>
  55. #include <asm/uaccess.h>
  56. #include <linux/atomic.h>
  57. #include <linux/mutex.h>
  58. #include <linux/workqueue.h>
  59. #include <linux/cgroup.h>
  60. /*
  61. * Tracks how many cpusets are currently defined in system.
  62. * When there is only one cpuset (the root cpuset) we can
  63. * short circuit some hooks.
  64. */
  65. int number_of_cpusets __read_mostly;
  66. /* Forward declare cgroup structures */
  67. struct cgroup_subsys cpuset_subsys;
  68. struct cpuset;
  69. /* See "Frequency meter" comments, below. */
  70. struct fmeter {
  71. int cnt; /* unprocessed events count */
  72. int val; /* most recent output value */
  73. time_t time; /* clock (secs) when val computed */
  74. spinlock_t lock; /* guards read or write of above */
  75. };
  76. struct cpuset {
  77. struct cgroup_subsys_state css;
  78. unsigned long flags; /* "unsigned long" so bitops work */
  79. cpumask_var_t cpus_allowed; /* CPUs allowed to tasks in cpuset */
  80. nodemask_t mems_allowed; /* Memory Nodes allowed to tasks */
  81. struct cpuset *parent; /* my parent */
  82. struct fmeter fmeter; /* memory_pressure filter */
  83. /*
  84. * Tasks are being attached to this cpuset. Used to prevent
  85. * zeroing cpus/mems_allowed between ->can_attach() and ->attach().
  86. */
  87. int attach_in_progress;
  88. /* partition number for rebuild_sched_domains() */
  89. int pn;
  90. /* for custom sched domain */
  91. int relax_domain_level;
  92. /* used for walking a cpuset hierarchy */
  93. struct list_head stack_list;
  94. struct work_struct hotplug_work;
  95. };
  96. /* Retrieve the cpuset for a cgroup */
  97. static inline struct cpuset *cgroup_cs(struct cgroup *cont)
  98. {
  99. return container_of(cgroup_subsys_state(cont, cpuset_subsys_id),
  100. struct cpuset, css);
  101. }
  102. /* Retrieve the cpuset for a task */
  103. static inline struct cpuset *task_cs(struct task_struct *task)
  104. {
  105. return container_of(task_subsys_state(task, cpuset_subsys_id),
  106. struct cpuset, css);
  107. }
  108. #ifdef CONFIG_NUMA
  109. static inline bool task_has_mempolicy(struct task_struct *task)
  110. {
  111. return task->mempolicy;
  112. }
  113. #else
  114. static inline bool task_has_mempolicy(struct task_struct *task)
  115. {
  116. return false;
  117. }
  118. #endif
  119. /* bits in struct cpuset flags field */
  120. typedef enum {
  121. CS_ONLINE,
  122. CS_CPU_EXCLUSIVE,
  123. CS_MEM_EXCLUSIVE,
  124. CS_MEM_HARDWALL,
  125. CS_MEMORY_MIGRATE,
  126. CS_SCHED_LOAD_BALANCE,
  127. CS_SPREAD_PAGE,
  128. CS_SPREAD_SLAB,
  129. } cpuset_flagbits_t;
  130. /* convenient tests for these bits */
  131. static inline bool is_cpuset_online(const struct cpuset *cs)
  132. {
  133. return test_bit(CS_ONLINE, &cs->flags);
  134. }
  135. static inline int is_cpu_exclusive(const struct cpuset *cs)
  136. {
  137. return test_bit(CS_CPU_EXCLUSIVE, &cs->flags);
  138. }
  139. static inline int is_mem_exclusive(const struct cpuset *cs)
  140. {
  141. return test_bit(CS_MEM_EXCLUSIVE, &cs->flags);
  142. }
  143. static inline int is_mem_hardwall(const struct cpuset *cs)
  144. {
  145. return test_bit(CS_MEM_HARDWALL, &cs->flags);
  146. }
  147. static inline int is_sched_load_balance(const struct cpuset *cs)
  148. {
  149. return test_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
  150. }
  151. static inline int is_memory_migrate(const struct cpuset *cs)
  152. {
  153. return test_bit(CS_MEMORY_MIGRATE, &cs->flags);
  154. }
  155. static inline int is_spread_page(const struct cpuset *cs)
  156. {
  157. return test_bit(CS_SPREAD_PAGE, &cs->flags);
  158. }
  159. static inline int is_spread_slab(const struct cpuset *cs)
  160. {
  161. return test_bit(CS_SPREAD_SLAB, &cs->flags);
  162. }
  163. static struct cpuset top_cpuset = {
  164. .flags = ((1 << CS_ONLINE) | (1 << CS_CPU_EXCLUSIVE) |
  165. (1 << CS_MEM_EXCLUSIVE)),
  166. };
  167. /**
  168. * cpuset_for_each_child - traverse online children of a cpuset
  169. * @child_cs: loop cursor pointing to the current child
  170. * @pos_cgrp: used for iteration
  171. * @parent_cs: target cpuset to walk children of
  172. *
  173. * Walk @child_cs through the online children of @parent_cs. Must be used
  174. * with RCU read locked.
  175. */
  176. #define cpuset_for_each_child(child_cs, pos_cgrp, parent_cs) \
  177. cgroup_for_each_child((pos_cgrp), (parent_cs)->css.cgroup) \
  178. if (is_cpuset_online(((child_cs) = cgroup_cs((pos_cgrp)))))
  179. /*
  180. * There are two global mutexes guarding cpuset structures. The first
  181. * is the main control groups cgroup_mutex, accessed via
  182. * cgroup_lock()/cgroup_unlock(). The second is the cpuset-specific
  183. * callback_mutex, below. They can nest. It is ok to first take
  184. * cgroup_mutex, then nest callback_mutex. We also require taking
  185. * task_lock() when dereferencing a task's cpuset pointer. See "The
  186. * task_lock() exception", at the end of this comment.
  187. *
  188. * A task must hold both mutexes to modify cpusets. If a task
  189. * holds cgroup_mutex, then it blocks others wanting that mutex,
  190. * ensuring that it is the only task able to also acquire callback_mutex
  191. * and be able to modify cpusets. It can perform various checks on
  192. * the cpuset structure first, knowing nothing will change. It can
  193. * also allocate memory while just holding cgroup_mutex. While it is
  194. * performing these checks, various callback routines can briefly
  195. * acquire callback_mutex to query cpusets. Once it is ready to make
  196. * the changes, it takes callback_mutex, blocking everyone else.
  197. *
  198. * Calls to the kernel memory allocator can not be made while holding
  199. * callback_mutex, as that would risk double tripping on callback_mutex
  200. * from one of the callbacks into the cpuset code from within
  201. * __alloc_pages().
  202. *
  203. * If a task is only holding callback_mutex, then it has read-only
  204. * access to cpusets.
  205. *
  206. * Now, the task_struct fields mems_allowed and mempolicy may be changed
  207. * by other task, we use alloc_lock in the task_struct fields to protect
  208. * them.
  209. *
  210. * The cpuset_common_file_read() handlers only hold callback_mutex across
  211. * small pieces of code, such as when reading out possibly multi-word
  212. * cpumasks and nodemasks.
  213. *
  214. * Accessing a task's cpuset should be done in accordance with the
  215. * guidelines for accessing subsystem state in kernel/cgroup.c
  216. */
  217. static DEFINE_MUTEX(callback_mutex);
  218. /*
  219. * cpuset_buffer_lock protects both the cpuset_name and cpuset_nodelist
  220. * buffers. They are statically allocated to prevent using excess stack
  221. * when calling cpuset_print_task_mems_allowed().
  222. */
  223. #define CPUSET_NAME_LEN (128)
  224. #define CPUSET_NODELIST_LEN (256)
  225. static char cpuset_name[CPUSET_NAME_LEN];
  226. static char cpuset_nodelist[CPUSET_NODELIST_LEN];
  227. static DEFINE_SPINLOCK(cpuset_buffer_lock);
  228. /*
  229. * CPU / memory hotplug is handled asynchronously.
  230. */
  231. static struct workqueue_struct *cpuset_propagate_hotplug_wq;
  232. static void cpuset_hotplug_workfn(struct work_struct *work);
  233. static void cpuset_propagate_hotplug_workfn(struct work_struct *work);
  234. static void schedule_cpuset_propagate_hotplug(struct cpuset *cs);
  235. static DECLARE_WORK(cpuset_hotplug_work, cpuset_hotplug_workfn);
  236. /*
  237. * This is ugly, but preserves the userspace API for existing cpuset
  238. * users. If someone tries to mount the "cpuset" filesystem, we
  239. * silently switch it to mount "cgroup" instead
  240. */
  241. static struct dentry *cpuset_mount(struct file_system_type *fs_type,
  242. int flags, const char *unused_dev_name, void *data)
  243. {
  244. struct file_system_type *cgroup_fs = get_fs_type("cgroup");
  245. struct dentry *ret = ERR_PTR(-ENODEV);
  246. if (cgroup_fs) {
  247. char mountopts[] =
  248. "cpuset,noprefix,"
  249. "release_agent=/sbin/cpuset_release_agent";
  250. ret = cgroup_fs->mount(cgroup_fs, flags,
  251. unused_dev_name, mountopts);
  252. put_filesystem(cgroup_fs);
  253. }
  254. return ret;
  255. }
  256. static struct file_system_type cpuset_fs_type = {
  257. .name = "cpuset",
  258. .mount = cpuset_mount,
  259. };
  260. /*
  261. * Return in pmask the portion of a cpusets's cpus_allowed that
  262. * are online. If none are online, walk up the cpuset hierarchy
  263. * until we find one that does have some online cpus. If we get
  264. * all the way to the top and still haven't found any online cpus,
  265. * return cpu_online_mask. Or if passed a NULL cs from an exit'ing
  266. * task, return cpu_online_mask.
  267. *
  268. * One way or another, we guarantee to return some non-empty subset
  269. * of cpu_online_mask.
  270. *
  271. * Call with callback_mutex held.
  272. */
  273. static void guarantee_online_cpus(const struct cpuset *cs,
  274. struct cpumask *pmask)
  275. {
  276. while (cs && !cpumask_intersects(cs->cpus_allowed, cpu_online_mask))
  277. cs = cs->parent;
  278. if (cs)
  279. cpumask_and(pmask, cs->cpus_allowed, cpu_online_mask);
  280. else
  281. cpumask_copy(pmask, cpu_online_mask);
  282. BUG_ON(!cpumask_intersects(pmask, cpu_online_mask));
  283. }
  284. /*
  285. * Return in *pmask the portion of a cpusets's mems_allowed that
  286. * are online, with memory. If none are online with memory, walk
  287. * up the cpuset hierarchy until we find one that does have some
  288. * online mems. If we get all the way to the top and still haven't
  289. * found any online mems, return node_states[N_MEMORY].
  290. *
  291. * One way or another, we guarantee to return some non-empty subset
  292. * of node_states[N_MEMORY].
  293. *
  294. * Call with callback_mutex held.
  295. */
  296. static void guarantee_online_mems(const struct cpuset *cs, nodemask_t *pmask)
  297. {
  298. while (cs && !nodes_intersects(cs->mems_allowed,
  299. node_states[N_MEMORY]))
  300. cs = cs->parent;
  301. if (cs)
  302. nodes_and(*pmask, cs->mems_allowed,
  303. node_states[N_MEMORY]);
  304. else
  305. *pmask = node_states[N_MEMORY];
  306. BUG_ON(!nodes_intersects(*pmask, node_states[N_MEMORY]));
  307. }
  308. /*
  309. * update task's spread flag if cpuset's page/slab spread flag is set
  310. *
  311. * Called with callback_mutex/cgroup_mutex held
  312. */
  313. static void cpuset_update_task_spread_flag(struct cpuset *cs,
  314. struct task_struct *tsk)
  315. {
  316. if (is_spread_page(cs))
  317. tsk->flags |= PF_SPREAD_PAGE;
  318. else
  319. tsk->flags &= ~PF_SPREAD_PAGE;
  320. if (is_spread_slab(cs))
  321. tsk->flags |= PF_SPREAD_SLAB;
  322. else
  323. tsk->flags &= ~PF_SPREAD_SLAB;
  324. }
  325. /*
  326. * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
  327. *
  328. * One cpuset is a subset of another if all its allowed CPUs and
  329. * Memory Nodes are a subset of the other, and its exclusive flags
  330. * are only set if the other's are set. Call holding cgroup_mutex.
  331. */
  332. static int is_cpuset_subset(const struct cpuset *p, const struct cpuset *q)
  333. {
  334. return cpumask_subset(p->cpus_allowed, q->cpus_allowed) &&
  335. nodes_subset(p->mems_allowed, q->mems_allowed) &&
  336. is_cpu_exclusive(p) <= is_cpu_exclusive(q) &&
  337. is_mem_exclusive(p) <= is_mem_exclusive(q);
  338. }
  339. /**
  340. * alloc_trial_cpuset - allocate a trial cpuset
  341. * @cs: the cpuset that the trial cpuset duplicates
  342. */
  343. static struct cpuset *alloc_trial_cpuset(const struct cpuset *cs)
  344. {
  345. struct cpuset *trial;
  346. trial = kmemdup(cs, sizeof(*cs), GFP_KERNEL);
  347. if (!trial)
  348. return NULL;
  349. if (!alloc_cpumask_var(&trial->cpus_allowed, GFP_KERNEL)) {
  350. kfree(trial);
  351. return NULL;
  352. }
  353. cpumask_copy(trial->cpus_allowed, cs->cpus_allowed);
  354. return trial;
  355. }
  356. /**
  357. * free_trial_cpuset - free the trial cpuset
  358. * @trial: the trial cpuset to be freed
  359. */
  360. static void free_trial_cpuset(struct cpuset *trial)
  361. {
  362. free_cpumask_var(trial->cpus_allowed);
  363. kfree(trial);
  364. }
  365. /*
  366. * validate_change() - Used to validate that any proposed cpuset change
  367. * follows the structural rules for cpusets.
  368. *
  369. * If we replaced the flag and mask values of the current cpuset
  370. * (cur) with those values in the trial cpuset (trial), would
  371. * our various subset and exclusive rules still be valid? Presumes
  372. * cgroup_mutex held.
  373. *
  374. * 'cur' is the address of an actual, in-use cpuset. Operations
  375. * such as list traversal that depend on the actual address of the
  376. * cpuset in the list must use cur below, not trial.
  377. *
  378. * 'trial' is the address of bulk structure copy of cur, with
  379. * perhaps one or more of the fields cpus_allowed, mems_allowed,
  380. * or flags changed to new, trial values.
  381. *
  382. * Return 0 if valid, -errno if not.
  383. */
  384. static int validate_change(const struct cpuset *cur, const struct cpuset *trial)
  385. {
  386. struct cgroup *cont;
  387. struct cpuset *c, *par;
  388. int ret;
  389. rcu_read_lock();
  390. /* Each of our child cpusets must be a subset of us */
  391. ret = -EBUSY;
  392. cpuset_for_each_child(c, cont, cur)
  393. if (!is_cpuset_subset(c, trial))
  394. goto out;
  395. /* Remaining checks don't apply to root cpuset */
  396. ret = 0;
  397. if (cur == &top_cpuset)
  398. goto out;
  399. par = cur->parent;
  400. /* We must be a subset of our parent cpuset */
  401. ret = -EACCES;
  402. if (!is_cpuset_subset(trial, par))
  403. goto out;
  404. /*
  405. * If either I or some sibling (!= me) is exclusive, we can't
  406. * overlap
  407. */
  408. ret = -EINVAL;
  409. cpuset_for_each_child(c, cont, par) {
  410. if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
  411. c != cur &&
  412. cpumask_intersects(trial->cpus_allowed, c->cpus_allowed))
  413. goto out;
  414. if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) &&
  415. c != cur &&
  416. nodes_intersects(trial->mems_allowed, c->mems_allowed))
  417. goto out;
  418. }
  419. /*
  420. * Cpusets with tasks - existing or newly being attached - can't
  421. * have empty cpus_allowed or mems_allowed.
  422. */
  423. ret = -ENOSPC;
  424. if ((cgroup_task_count(cur->css.cgroup) || cur->attach_in_progress) &&
  425. (cpumask_empty(trial->cpus_allowed) ||
  426. nodes_empty(trial->mems_allowed)))
  427. goto out;
  428. ret = 0;
  429. out:
  430. rcu_read_unlock();
  431. return ret;
  432. }
  433. #ifdef CONFIG_SMP
  434. /*
  435. * Helper routine for generate_sched_domains().
  436. * Do cpusets a, b have overlapping cpus_allowed masks?
  437. */
  438. static int cpusets_overlap(struct cpuset *a, struct cpuset *b)
  439. {
  440. return cpumask_intersects(a->cpus_allowed, b->cpus_allowed);
  441. }
  442. static void
  443. update_domain_attr(struct sched_domain_attr *dattr, struct cpuset *c)
  444. {
  445. if (dattr->relax_domain_level < c->relax_domain_level)
  446. dattr->relax_domain_level = c->relax_domain_level;
  447. return;
  448. }
  449. static void
  450. update_domain_attr_tree(struct sched_domain_attr *dattr, struct cpuset *c)
  451. {
  452. LIST_HEAD(q);
  453. list_add(&c->stack_list, &q);
  454. while (!list_empty(&q)) {
  455. struct cpuset *cp;
  456. struct cgroup *cont;
  457. struct cpuset *child;
  458. cp = list_first_entry(&q, struct cpuset, stack_list);
  459. list_del(q.next);
  460. if (cpumask_empty(cp->cpus_allowed))
  461. continue;
  462. if (is_sched_load_balance(cp))
  463. update_domain_attr(dattr, cp);
  464. rcu_read_lock();
  465. cpuset_for_each_child(child, cont, cp)
  466. list_add_tail(&child->stack_list, &q);
  467. rcu_read_unlock();
  468. }
  469. }
  470. /*
  471. * generate_sched_domains()
  472. *
  473. * This function builds a partial partition of the systems CPUs
  474. * A 'partial partition' is a set of non-overlapping subsets whose
  475. * union is a subset of that set.
  476. * The output of this function needs to be passed to kernel/sched.c
  477. * partition_sched_domains() routine, which will rebuild the scheduler's
  478. * load balancing domains (sched domains) as specified by that partial
  479. * partition.
  480. *
  481. * See "What is sched_load_balance" in Documentation/cgroups/cpusets.txt
  482. * for a background explanation of this.
  483. *
  484. * Does not return errors, on the theory that the callers of this
  485. * routine would rather not worry about failures to rebuild sched
  486. * domains when operating in the severe memory shortage situations
  487. * that could cause allocation failures below.
  488. *
  489. * Must be called with cgroup_lock held.
  490. *
  491. * The three key local variables below are:
  492. * q - a linked-list queue of cpuset pointers, used to implement a
  493. * top-down scan of all cpusets. This scan loads a pointer
  494. * to each cpuset marked is_sched_load_balance into the
  495. * array 'csa'. For our purposes, rebuilding the schedulers
  496. * sched domains, we can ignore !is_sched_load_balance cpusets.
  497. * csa - (for CpuSet Array) Array of pointers to all the cpusets
  498. * that need to be load balanced, for convenient iterative
  499. * access by the subsequent code that finds the best partition,
  500. * i.e the set of domains (subsets) of CPUs such that the
  501. * cpus_allowed of every cpuset marked is_sched_load_balance
  502. * is a subset of one of these domains, while there are as
  503. * many such domains as possible, each as small as possible.
  504. * doms - Conversion of 'csa' to an array of cpumasks, for passing to
  505. * the kernel/sched.c routine partition_sched_domains() in a
  506. * convenient format, that can be easily compared to the prior
  507. * value to determine what partition elements (sched domains)
  508. * were changed (added or removed.)
  509. *
  510. * Finding the best partition (set of domains):
  511. * The triple nested loops below over i, j, k scan over the
  512. * load balanced cpusets (using the array of cpuset pointers in
  513. * csa[]) looking for pairs of cpusets that have overlapping
  514. * cpus_allowed, but which don't have the same 'pn' partition
  515. * number and gives them in the same partition number. It keeps
  516. * looping on the 'restart' label until it can no longer find
  517. * any such pairs.
  518. *
  519. * The union of the cpus_allowed masks from the set of
  520. * all cpusets having the same 'pn' value then form the one
  521. * element of the partition (one sched domain) to be passed to
  522. * partition_sched_domains().
  523. */
  524. static int generate_sched_domains(cpumask_var_t **domains,
  525. struct sched_domain_attr **attributes)
  526. {
  527. LIST_HEAD(q); /* queue of cpusets to be scanned */
  528. struct cpuset *cp; /* scans q */
  529. struct cpuset **csa; /* array of all cpuset ptrs */
  530. int csn; /* how many cpuset ptrs in csa so far */
  531. int i, j, k; /* indices for partition finding loops */
  532. cpumask_var_t *doms; /* resulting partition; i.e. sched domains */
  533. struct sched_domain_attr *dattr; /* attributes for custom domains */
  534. int ndoms = 0; /* number of sched domains in result */
  535. int nslot; /* next empty doms[] struct cpumask slot */
  536. doms = NULL;
  537. dattr = NULL;
  538. csa = NULL;
  539. /* Special case for the 99% of systems with one, full, sched domain */
  540. if (is_sched_load_balance(&top_cpuset)) {
  541. ndoms = 1;
  542. doms = alloc_sched_domains(ndoms);
  543. if (!doms)
  544. goto done;
  545. dattr = kmalloc(sizeof(struct sched_domain_attr), GFP_KERNEL);
  546. if (dattr) {
  547. *dattr = SD_ATTR_INIT;
  548. update_domain_attr_tree(dattr, &top_cpuset);
  549. }
  550. cpumask_copy(doms[0], top_cpuset.cpus_allowed);
  551. goto done;
  552. }
  553. csa = kmalloc(number_of_cpusets * sizeof(cp), GFP_KERNEL);
  554. if (!csa)
  555. goto done;
  556. csn = 0;
  557. list_add(&top_cpuset.stack_list, &q);
  558. while (!list_empty(&q)) {
  559. struct cgroup *cont;
  560. struct cpuset *child; /* scans child cpusets of cp */
  561. cp = list_first_entry(&q, struct cpuset, stack_list);
  562. list_del(q.next);
  563. if (cpumask_empty(cp->cpus_allowed))
  564. continue;
  565. /*
  566. * All child cpusets contain a subset of the parent's cpus, so
  567. * just skip them, and then we call update_domain_attr_tree()
  568. * to calc relax_domain_level of the corresponding sched
  569. * domain.
  570. */
  571. if (is_sched_load_balance(cp)) {
  572. csa[csn++] = cp;
  573. continue;
  574. }
  575. rcu_read_lock();
  576. cpuset_for_each_child(child, cont, cp)
  577. list_add_tail(&child->stack_list, &q);
  578. rcu_read_unlock();
  579. }
  580. for (i = 0; i < csn; i++)
  581. csa[i]->pn = i;
  582. ndoms = csn;
  583. restart:
  584. /* Find the best partition (set of sched domains) */
  585. for (i = 0; i < csn; i++) {
  586. struct cpuset *a = csa[i];
  587. int apn = a->pn;
  588. for (j = 0; j < csn; j++) {
  589. struct cpuset *b = csa[j];
  590. int bpn = b->pn;
  591. if (apn != bpn && cpusets_overlap(a, b)) {
  592. for (k = 0; k < csn; k++) {
  593. struct cpuset *c = csa[k];
  594. if (c->pn == bpn)
  595. c->pn = apn;
  596. }
  597. ndoms--; /* one less element */
  598. goto restart;
  599. }
  600. }
  601. }
  602. /*
  603. * Now we know how many domains to create.
  604. * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
  605. */
  606. doms = alloc_sched_domains(ndoms);
  607. if (!doms)
  608. goto done;
  609. /*
  610. * The rest of the code, including the scheduler, can deal with
  611. * dattr==NULL case. No need to abort if alloc fails.
  612. */
  613. dattr = kmalloc(ndoms * sizeof(struct sched_domain_attr), GFP_KERNEL);
  614. for (nslot = 0, i = 0; i < csn; i++) {
  615. struct cpuset *a = csa[i];
  616. struct cpumask *dp;
  617. int apn = a->pn;
  618. if (apn < 0) {
  619. /* Skip completed partitions */
  620. continue;
  621. }
  622. dp = doms[nslot];
  623. if (nslot == ndoms) {
  624. static int warnings = 10;
  625. if (warnings) {
  626. printk(KERN_WARNING
  627. "rebuild_sched_domains confused:"
  628. " nslot %d, ndoms %d, csn %d, i %d,"
  629. " apn %d\n",
  630. nslot, ndoms, csn, i, apn);
  631. warnings--;
  632. }
  633. continue;
  634. }
  635. cpumask_clear(dp);
  636. if (dattr)
  637. *(dattr + nslot) = SD_ATTR_INIT;
  638. for (j = i; j < csn; j++) {
  639. struct cpuset *b = csa[j];
  640. if (apn == b->pn) {
  641. cpumask_or(dp, dp, b->cpus_allowed);
  642. if (dattr)
  643. update_domain_attr_tree(dattr + nslot, b);
  644. /* Done with this partition */
  645. b->pn = -1;
  646. }
  647. }
  648. nslot++;
  649. }
  650. BUG_ON(nslot != ndoms);
  651. done:
  652. kfree(csa);
  653. /*
  654. * Fallback to the default domain if kmalloc() failed.
  655. * See comments in partition_sched_domains().
  656. */
  657. if (doms == NULL)
  658. ndoms = 1;
  659. *domains = doms;
  660. *attributes = dattr;
  661. return ndoms;
  662. }
  663. /*
  664. * Rebuild scheduler domains.
  665. *
  666. * If the flag 'sched_load_balance' of any cpuset with non-empty
  667. * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
  668. * which has that flag enabled, or if any cpuset with a non-empty
  669. * 'cpus' is removed, then call this routine to rebuild the
  670. * scheduler's dynamic sched domains.
  671. *
  672. * Call with cgroup_mutex held. Takes get_online_cpus().
  673. */
  674. static void rebuild_sched_domains_locked(void)
  675. {
  676. struct sched_domain_attr *attr;
  677. cpumask_var_t *doms;
  678. int ndoms;
  679. WARN_ON_ONCE(!cgroup_lock_is_held());
  680. get_online_cpus();
  681. /* Generate domain masks and attrs */
  682. ndoms = generate_sched_domains(&doms, &attr);
  683. /* Have scheduler rebuild the domains */
  684. partition_sched_domains(ndoms, doms, attr);
  685. put_online_cpus();
  686. }
  687. #else /* !CONFIG_SMP */
  688. static void rebuild_sched_domains_locked(void)
  689. {
  690. }
  691. static int generate_sched_domains(cpumask_var_t **domains,
  692. struct sched_domain_attr **attributes)
  693. {
  694. *domains = NULL;
  695. return 1;
  696. }
  697. #endif /* CONFIG_SMP */
  698. void rebuild_sched_domains(void)
  699. {
  700. cgroup_lock();
  701. rebuild_sched_domains_locked();
  702. cgroup_unlock();
  703. }
  704. /**
  705. * cpuset_test_cpumask - test a task's cpus_allowed versus its cpuset's
  706. * @tsk: task to test
  707. * @scan: struct cgroup_scanner contained in its struct cpuset_hotplug_scanner
  708. *
  709. * Call with cgroup_mutex held. May take callback_mutex during call.
  710. * Called for each task in a cgroup by cgroup_scan_tasks().
  711. * Return nonzero if this tasks's cpus_allowed mask should be changed (in other
  712. * words, if its mask is not equal to its cpuset's mask).
  713. */
  714. static int cpuset_test_cpumask(struct task_struct *tsk,
  715. struct cgroup_scanner *scan)
  716. {
  717. return !cpumask_equal(&tsk->cpus_allowed,
  718. (cgroup_cs(scan->cg))->cpus_allowed);
  719. }
  720. /**
  721. * cpuset_change_cpumask - make a task's cpus_allowed the same as its cpuset's
  722. * @tsk: task to test
  723. * @scan: struct cgroup_scanner containing the cgroup of the task
  724. *
  725. * Called by cgroup_scan_tasks() for each task in a cgroup whose
  726. * cpus_allowed mask needs to be changed.
  727. *
  728. * We don't need to re-check for the cgroup/cpuset membership, since we're
  729. * holding cgroup_lock() at this point.
  730. */
  731. static void cpuset_change_cpumask(struct task_struct *tsk,
  732. struct cgroup_scanner *scan)
  733. {
  734. set_cpus_allowed_ptr(tsk, ((cgroup_cs(scan->cg))->cpus_allowed));
  735. }
  736. /**
  737. * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
  738. * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
  739. * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
  740. *
  741. * Called with cgroup_mutex held
  742. *
  743. * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
  744. * calling callback functions for each.
  745. *
  746. * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
  747. * if @heap != NULL.
  748. */
  749. static void update_tasks_cpumask(struct cpuset *cs, struct ptr_heap *heap)
  750. {
  751. struct cgroup_scanner scan;
  752. scan.cg = cs->css.cgroup;
  753. scan.test_task = cpuset_test_cpumask;
  754. scan.process_task = cpuset_change_cpumask;
  755. scan.heap = heap;
  756. cgroup_scan_tasks(&scan);
  757. }
  758. /**
  759. * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
  760. * @cs: the cpuset to consider
  761. * @buf: buffer of cpu numbers written to this cpuset
  762. */
  763. static int update_cpumask(struct cpuset *cs, struct cpuset *trialcs,
  764. const char *buf)
  765. {
  766. struct ptr_heap heap;
  767. int retval;
  768. int is_load_balanced;
  769. /* top_cpuset.cpus_allowed tracks cpu_online_mask; it's read-only */
  770. if (cs == &top_cpuset)
  771. return -EACCES;
  772. /*
  773. * An empty cpus_allowed is ok only if the cpuset has no tasks.
  774. * Since cpulist_parse() fails on an empty mask, we special case
  775. * that parsing. The validate_change() call ensures that cpusets
  776. * with tasks have cpus.
  777. */
  778. if (!*buf) {
  779. cpumask_clear(trialcs->cpus_allowed);
  780. } else {
  781. retval = cpulist_parse(buf, trialcs->cpus_allowed);
  782. if (retval < 0)
  783. return retval;
  784. if (!cpumask_subset(trialcs->cpus_allowed, cpu_active_mask))
  785. return -EINVAL;
  786. }
  787. retval = validate_change(cs, trialcs);
  788. if (retval < 0)
  789. return retval;
  790. /* Nothing to do if the cpus didn't change */
  791. if (cpumask_equal(cs->cpus_allowed, trialcs->cpus_allowed))
  792. return 0;
  793. retval = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
  794. if (retval)
  795. return retval;
  796. is_load_balanced = is_sched_load_balance(trialcs);
  797. mutex_lock(&callback_mutex);
  798. cpumask_copy(cs->cpus_allowed, trialcs->cpus_allowed);
  799. mutex_unlock(&callback_mutex);
  800. /*
  801. * Scan tasks in the cpuset, and update the cpumasks of any
  802. * that need an update.
  803. */
  804. update_tasks_cpumask(cs, &heap);
  805. heap_free(&heap);
  806. if (is_load_balanced)
  807. rebuild_sched_domains_locked();
  808. return 0;
  809. }
  810. /*
  811. * cpuset_migrate_mm
  812. *
  813. * Migrate memory region from one set of nodes to another.
  814. *
  815. * Temporarilly set tasks mems_allowed to target nodes of migration,
  816. * so that the migration code can allocate pages on these nodes.
  817. *
  818. * Call holding cgroup_mutex, so current's cpuset won't change
  819. * during this call, as manage_mutex holds off any cpuset_attach()
  820. * calls. Therefore we don't need to take task_lock around the
  821. * call to guarantee_online_mems(), as we know no one is changing
  822. * our task's cpuset.
  823. *
  824. * While the mm_struct we are migrating is typically from some
  825. * other task, the task_struct mems_allowed that we are hacking
  826. * is for our current task, which must allocate new pages for that
  827. * migrating memory region.
  828. */
  829. static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
  830. const nodemask_t *to)
  831. {
  832. struct task_struct *tsk = current;
  833. tsk->mems_allowed = *to;
  834. do_migrate_pages(mm, from, to, MPOL_MF_MOVE_ALL);
  835. guarantee_online_mems(task_cs(tsk),&tsk->mems_allowed);
  836. }
  837. /*
  838. * cpuset_change_task_nodemask - change task's mems_allowed and mempolicy
  839. * @tsk: the task to change
  840. * @newmems: new nodes that the task will be set
  841. *
  842. * In order to avoid seeing no nodes if the old and new nodes are disjoint,
  843. * we structure updates as setting all new allowed nodes, then clearing newly
  844. * disallowed ones.
  845. */
  846. static void cpuset_change_task_nodemask(struct task_struct *tsk,
  847. nodemask_t *newmems)
  848. {
  849. bool need_loop;
  850. /*
  851. * Allow tasks that have access to memory reserves because they have
  852. * been OOM killed to get memory anywhere.
  853. */
  854. if (unlikely(test_thread_flag(TIF_MEMDIE)))
  855. return;
  856. if (current->flags & PF_EXITING) /* Let dying task have memory */
  857. return;
  858. task_lock(tsk);
  859. /*
  860. * Determine if a loop is necessary if another thread is doing
  861. * get_mems_allowed(). If at least one node remains unchanged and
  862. * tsk does not have a mempolicy, then an empty nodemask will not be
  863. * possible when mems_allowed is larger than a word.
  864. */
  865. need_loop = task_has_mempolicy(tsk) ||
  866. !nodes_intersects(*newmems, tsk->mems_allowed);
  867. if (need_loop)
  868. write_seqcount_begin(&tsk->mems_allowed_seq);
  869. nodes_or(tsk->mems_allowed, tsk->mems_allowed, *newmems);
  870. mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP1);
  871. mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP2);
  872. tsk->mems_allowed = *newmems;
  873. if (need_loop)
  874. write_seqcount_end(&tsk->mems_allowed_seq);
  875. task_unlock(tsk);
  876. }
  877. /*
  878. * Update task's mems_allowed and rebind its mempolicy and vmas' mempolicy
  879. * of it to cpuset's new mems_allowed, and migrate pages to new nodes if
  880. * memory_migrate flag is set. Called with cgroup_mutex held.
  881. */
  882. static void cpuset_change_nodemask(struct task_struct *p,
  883. struct cgroup_scanner *scan)
  884. {
  885. struct mm_struct *mm;
  886. struct cpuset *cs;
  887. int migrate;
  888. const nodemask_t *oldmem = scan->data;
  889. static nodemask_t newmems; /* protected by cgroup_mutex */
  890. cs = cgroup_cs(scan->cg);
  891. guarantee_online_mems(cs, &newmems);
  892. cpuset_change_task_nodemask(p, &newmems);
  893. mm = get_task_mm(p);
  894. if (!mm)
  895. return;
  896. migrate = is_memory_migrate(cs);
  897. mpol_rebind_mm(mm, &cs->mems_allowed);
  898. if (migrate)
  899. cpuset_migrate_mm(mm, oldmem, &cs->mems_allowed);
  900. mmput(mm);
  901. }
  902. static void *cpuset_being_rebound;
  903. /**
  904. * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
  905. * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
  906. * @oldmem: old mems_allowed of cpuset cs
  907. * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
  908. *
  909. * Called with cgroup_mutex held
  910. * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
  911. * if @heap != NULL.
  912. */
  913. static void update_tasks_nodemask(struct cpuset *cs, const nodemask_t *oldmem,
  914. struct ptr_heap *heap)
  915. {
  916. struct cgroup_scanner scan;
  917. cpuset_being_rebound = cs; /* causes mpol_dup() rebind */
  918. scan.cg = cs->css.cgroup;
  919. scan.test_task = NULL;
  920. scan.process_task = cpuset_change_nodemask;
  921. scan.heap = heap;
  922. scan.data = (nodemask_t *)oldmem;
  923. /*
  924. * The mpol_rebind_mm() call takes mmap_sem, which we couldn't
  925. * take while holding tasklist_lock. Forks can happen - the
  926. * mpol_dup() cpuset_being_rebound check will catch such forks,
  927. * and rebind their vma mempolicies too. Because we still hold
  928. * the global cgroup_mutex, we know that no other rebind effort
  929. * will be contending for the global variable cpuset_being_rebound.
  930. * It's ok if we rebind the same mm twice; mpol_rebind_mm()
  931. * is idempotent. Also migrate pages in each mm to new nodes.
  932. */
  933. cgroup_scan_tasks(&scan);
  934. /* We're done rebinding vmas to this cpuset's new mems_allowed. */
  935. cpuset_being_rebound = NULL;
  936. }
  937. /*
  938. * Handle user request to change the 'mems' memory placement
  939. * of a cpuset. Needs to validate the request, update the
  940. * cpusets mems_allowed, and for each task in the cpuset,
  941. * update mems_allowed and rebind task's mempolicy and any vma
  942. * mempolicies and if the cpuset is marked 'memory_migrate',
  943. * migrate the tasks pages to the new memory.
  944. *
  945. * Call with cgroup_mutex held. May take callback_mutex during call.
  946. * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
  947. * lock each such tasks mm->mmap_sem, scan its vma's and rebind
  948. * their mempolicies to the cpusets new mems_allowed.
  949. */
  950. static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
  951. const char *buf)
  952. {
  953. NODEMASK_ALLOC(nodemask_t, oldmem, GFP_KERNEL);
  954. int retval;
  955. struct ptr_heap heap;
  956. if (!oldmem)
  957. return -ENOMEM;
  958. /*
  959. * top_cpuset.mems_allowed tracks node_stats[N_MEMORY];
  960. * it's read-only
  961. */
  962. if (cs == &top_cpuset) {
  963. retval = -EACCES;
  964. goto done;
  965. }
  966. /*
  967. * An empty mems_allowed is ok iff there are no tasks in the cpuset.
  968. * Since nodelist_parse() fails on an empty mask, we special case
  969. * that parsing. The validate_change() call ensures that cpusets
  970. * with tasks have memory.
  971. */
  972. if (!*buf) {
  973. nodes_clear(trialcs->mems_allowed);
  974. } else {
  975. retval = nodelist_parse(buf, trialcs->mems_allowed);
  976. if (retval < 0)
  977. goto done;
  978. if (!nodes_subset(trialcs->mems_allowed,
  979. node_states[N_MEMORY])) {
  980. retval = -EINVAL;
  981. goto done;
  982. }
  983. }
  984. *oldmem = cs->mems_allowed;
  985. if (nodes_equal(*oldmem, trialcs->mems_allowed)) {
  986. retval = 0; /* Too easy - nothing to do */
  987. goto done;
  988. }
  989. retval = validate_change(cs, trialcs);
  990. if (retval < 0)
  991. goto done;
  992. retval = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
  993. if (retval < 0)
  994. goto done;
  995. mutex_lock(&callback_mutex);
  996. cs->mems_allowed = trialcs->mems_allowed;
  997. mutex_unlock(&callback_mutex);
  998. update_tasks_nodemask(cs, oldmem, &heap);
  999. heap_free(&heap);
  1000. done:
  1001. NODEMASK_FREE(oldmem);
  1002. return retval;
  1003. }
  1004. int current_cpuset_is_being_rebound(void)
  1005. {
  1006. return task_cs(current) == cpuset_being_rebound;
  1007. }
  1008. static int update_relax_domain_level(struct cpuset *cs, s64 val)
  1009. {
  1010. #ifdef CONFIG_SMP
  1011. if (val < -1 || val >= sched_domain_level_max)
  1012. return -EINVAL;
  1013. #endif
  1014. if (val != cs->relax_domain_level) {
  1015. cs->relax_domain_level = val;
  1016. if (!cpumask_empty(cs->cpus_allowed) &&
  1017. is_sched_load_balance(cs))
  1018. rebuild_sched_domains_locked();
  1019. }
  1020. return 0;
  1021. }
  1022. /*
  1023. * cpuset_change_flag - make a task's spread flags the same as its cpuset's
  1024. * @tsk: task to be updated
  1025. * @scan: struct cgroup_scanner containing the cgroup of the task
  1026. *
  1027. * Called by cgroup_scan_tasks() for each task in a cgroup.
  1028. *
  1029. * We don't need to re-check for the cgroup/cpuset membership, since we're
  1030. * holding cgroup_lock() at this point.
  1031. */
  1032. static void cpuset_change_flag(struct task_struct *tsk,
  1033. struct cgroup_scanner *scan)
  1034. {
  1035. cpuset_update_task_spread_flag(cgroup_cs(scan->cg), tsk);
  1036. }
  1037. /*
  1038. * update_tasks_flags - update the spread flags of tasks in the cpuset.
  1039. * @cs: the cpuset in which each task's spread flags needs to be changed
  1040. * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
  1041. *
  1042. * Called with cgroup_mutex held
  1043. *
  1044. * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
  1045. * calling callback functions for each.
  1046. *
  1047. * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
  1048. * if @heap != NULL.
  1049. */
  1050. static void update_tasks_flags(struct cpuset *cs, struct ptr_heap *heap)
  1051. {
  1052. struct cgroup_scanner scan;
  1053. scan.cg = cs->css.cgroup;
  1054. scan.test_task = NULL;
  1055. scan.process_task = cpuset_change_flag;
  1056. scan.heap = heap;
  1057. cgroup_scan_tasks(&scan);
  1058. }
  1059. /*
  1060. * update_flag - read a 0 or a 1 in a file and update associated flag
  1061. * bit: the bit to update (see cpuset_flagbits_t)
  1062. * cs: the cpuset to update
  1063. * turning_on: whether the flag is being set or cleared
  1064. *
  1065. * Call with cgroup_mutex held.
  1066. */
  1067. static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs,
  1068. int turning_on)
  1069. {
  1070. struct cpuset *trialcs;
  1071. int balance_flag_changed;
  1072. int spread_flag_changed;
  1073. struct ptr_heap heap;
  1074. int err;
  1075. trialcs = alloc_trial_cpuset(cs);
  1076. if (!trialcs)
  1077. return -ENOMEM;
  1078. if (turning_on)
  1079. set_bit(bit, &trialcs->flags);
  1080. else
  1081. clear_bit(bit, &trialcs->flags);
  1082. err = validate_change(cs, trialcs);
  1083. if (err < 0)
  1084. goto out;
  1085. err = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
  1086. if (err < 0)
  1087. goto out;
  1088. balance_flag_changed = (is_sched_load_balance(cs) !=
  1089. is_sched_load_balance(trialcs));
  1090. spread_flag_changed = ((is_spread_slab(cs) != is_spread_slab(trialcs))
  1091. || (is_spread_page(cs) != is_spread_page(trialcs)));
  1092. mutex_lock(&callback_mutex);
  1093. cs->flags = trialcs->flags;
  1094. mutex_unlock(&callback_mutex);
  1095. if (!cpumask_empty(trialcs->cpus_allowed) && balance_flag_changed)
  1096. rebuild_sched_domains_locked();
  1097. if (spread_flag_changed)
  1098. update_tasks_flags(cs, &heap);
  1099. heap_free(&heap);
  1100. out:
  1101. free_trial_cpuset(trialcs);
  1102. return err;
  1103. }
  1104. /*
  1105. * Frequency meter - How fast is some event occurring?
  1106. *
  1107. * These routines manage a digitally filtered, constant time based,
  1108. * event frequency meter. There are four routines:
  1109. * fmeter_init() - initialize a frequency meter.
  1110. * fmeter_markevent() - called each time the event happens.
  1111. * fmeter_getrate() - returns the recent rate of such events.
  1112. * fmeter_update() - internal routine used to update fmeter.
  1113. *
  1114. * A common data structure is passed to each of these routines,
  1115. * which is used to keep track of the state required to manage the
  1116. * frequency meter and its digital filter.
  1117. *
  1118. * The filter works on the number of events marked per unit time.
  1119. * The filter is single-pole low-pass recursive (IIR). The time unit
  1120. * is 1 second. Arithmetic is done using 32-bit integers scaled to
  1121. * simulate 3 decimal digits of precision (multiplied by 1000).
  1122. *
  1123. * With an FM_COEF of 933, and a time base of 1 second, the filter
  1124. * has a half-life of 10 seconds, meaning that if the events quit
  1125. * happening, then the rate returned from the fmeter_getrate()
  1126. * will be cut in half each 10 seconds, until it converges to zero.
  1127. *
  1128. * It is not worth doing a real infinitely recursive filter. If more
  1129. * than FM_MAXTICKS ticks have elapsed since the last filter event,
  1130. * just compute FM_MAXTICKS ticks worth, by which point the level
  1131. * will be stable.
  1132. *
  1133. * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
  1134. * arithmetic overflow in the fmeter_update() routine.
  1135. *
  1136. * Given the simple 32 bit integer arithmetic used, this meter works
  1137. * best for reporting rates between one per millisecond (msec) and
  1138. * one per 32 (approx) seconds. At constant rates faster than one
  1139. * per msec it maxes out at values just under 1,000,000. At constant
  1140. * rates between one per msec, and one per second it will stabilize
  1141. * to a value N*1000, where N is the rate of events per second.
  1142. * At constant rates between one per second and one per 32 seconds,
  1143. * it will be choppy, moving up on the seconds that have an event,
  1144. * and then decaying until the next event. At rates slower than
  1145. * about one in 32 seconds, it decays all the way back to zero between
  1146. * each event.
  1147. */
  1148. #define FM_COEF 933 /* coefficient for half-life of 10 secs */
  1149. #define FM_MAXTICKS ((time_t)99) /* useless computing more ticks than this */
  1150. #define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
  1151. #define FM_SCALE 1000 /* faux fixed point scale */
  1152. /* Initialize a frequency meter */
  1153. static void fmeter_init(struct fmeter *fmp)
  1154. {
  1155. fmp->cnt = 0;
  1156. fmp->val = 0;
  1157. fmp->time = 0;
  1158. spin_lock_init(&fmp->lock);
  1159. }
  1160. /* Internal meter update - process cnt events and update value */
  1161. static void fmeter_update(struct fmeter *fmp)
  1162. {
  1163. time_t now = get_seconds();
  1164. time_t ticks = now - fmp->time;
  1165. if (ticks == 0)
  1166. return;
  1167. ticks = min(FM_MAXTICKS, ticks);
  1168. while (ticks-- > 0)
  1169. fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
  1170. fmp->time = now;
  1171. fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
  1172. fmp->cnt = 0;
  1173. }
  1174. /* Process any previous ticks, then bump cnt by one (times scale). */
  1175. static void fmeter_markevent(struct fmeter *fmp)
  1176. {
  1177. spin_lock(&fmp->lock);
  1178. fmeter_update(fmp);
  1179. fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
  1180. spin_unlock(&fmp->lock);
  1181. }
  1182. /* Process any previous ticks, then return current value. */
  1183. static int fmeter_getrate(struct fmeter *fmp)
  1184. {
  1185. int val;
  1186. spin_lock(&fmp->lock);
  1187. fmeter_update(fmp);
  1188. val = fmp->val;
  1189. spin_unlock(&fmp->lock);
  1190. return val;
  1191. }
  1192. /* Called by cgroups to determine if a cpuset is usable; cgroup_mutex held */
  1193. static int cpuset_can_attach(struct cgroup *cgrp, struct cgroup_taskset *tset)
  1194. {
  1195. struct cpuset *cs = cgroup_cs(cgrp);
  1196. struct task_struct *task;
  1197. int ret;
  1198. if (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed))
  1199. return -ENOSPC;
  1200. cgroup_taskset_for_each(task, cgrp, tset) {
  1201. /*
  1202. * Kthreads bound to specific cpus cannot be moved to a new
  1203. * cpuset; we cannot change their cpu affinity and
  1204. * isolating such threads by their set of allowed nodes is
  1205. * unnecessary. Thus, cpusets are not applicable for such
  1206. * threads. This prevents checking for success of
  1207. * set_cpus_allowed_ptr() on all attached tasks before
  1208. * cpus_allowed may be changed.
  1209. */
  1210. if (task->flags & PF_THREAD_BOUND)
  1211. return -EINVAL;
  1212. if ((ret = security_task_setscheduler(task)))
  1213. return ret;
  1214. }
  1215. /*
  1216. * Mark attach is in progress. This makes validate_change() fail
  1217. * changes which zero cpus/mems_allowed.
  1218. */
  1219. cs->attach_in_progress++;
  1220. return 0;
  1221. }
  1222. static void cpuset_cancel_attach(struct cgroup *cgrp,
  1223. struct cgroup_taskset *tset)
  1224. {
  1225. cgroup_cs(cgrp)->attach_in_progress--;
  1226. }
  1227. /*
  1228. * Protected by cgroup_mutex. cpus_attach is used only by cpuset_attach()
  1229. * but we can't allocate it dynamically there. Define it global and
  1230. * allocate from cpuset_init().
  1231. */
  1232. static cpumask_var_t cpus_attach;
  1233. static void cpuset_attach(struct cgroup *cgrp, struct cgroup_taskset *tset)
  1234. {
  1235. /* static bufs protected by cgroup_mutex */
  1236. static nodemask_t cpuset_attach_nodemask_from;
  1237. static nodemask_t cpuset_attach_nodemask_to;
  1238. struct mm_struct *mm;
  1239. struct task_struct *task;
  1240. struct task_struct *leader = cgroup_taskset_first(tset);
  1241. struct cgroup *oldcgrp = cgroup_taskset_cur_cgroup(tset);
  1242. struct cpuset *cs = cgroup_cs(cgrp);
  1243. struct cpuset *oldcs = cgroup_cs(oldcgrp);
  1244. /* prepare for attach */
  1245. if (cs == &top_cpuset)
  1246. cpumask_copy(cpus_attach, cpu_possible_mask);
  1247. else
  1248. guarantee_online_cpus(cs, cpus_attach);
  1249. guarantee_online_mems(cs, &cpuset_attach_nodemask_to);
  1250. cgroup_taskset_for_each(task, cgrp, tset) {
  1251. /*
  1252. * can_attach beforehand should guarantee that this doesn't
  1253. * fail. TODO: have a better way to handle failure here
  1254. */
  1255. WARN_ON_ONCE(set_cpus_allowed_ptr(task, cpus_attach));
  1256. cpuset_change_task_nodemask(task, &cpuset_attach_nodemask_to);
  1257. cpuset_update_task_spread_flag(cs, task);
  1258. }
  1259. /*
  1260. * Change mm, possibly for multiple threads in a threadgroup. This is
  1261. * expensive and may sleep.
  1262. */
  1263. cpuset_attach_nodemask_from = oldcs->mems_allowed;
  1264. cpuset_attach_nodemask_to = cs->mems_allowed;
  1265. mm = get_task_mm(leader);
  1266. if (mm) {
  1267. mpol_rebind_mm(mm, &cpuset_attach_nodemask_to);
  1268. if (is_memory_migrate(cs))
  1269. cpuset_migrate_mm(mm, &cpuset_attach_nodemask_from,
  1270. &cpuset_attach_nodemask_to);
  1271. mmput(mm);
  1272. }
  1273. cs->attach_in_progress--;
  1274. /*
  1275. * We may have raced with CPU/memory hotunplug. Trigger hotplug
  1276. * propagation if @cs doesn't have any CPU or memory. It will move
  1277. * the newly added tasks to the nearest parent which can execute.
  1278. */
  1279. if (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed))
  1280. schedule_cpuset_propagate_hotplug(cs);
  1281. }
  1282. /* The various types of files and directories in a cpuset file system */
  1283. typedef enum {
  1284. FILE_MEMORY_MIGRATE,
  1285. FILE_CPULIST,
  1286. FILE_MEMLIST,
  1287. FILE_CPU_EXCLUSIVE,
  1288. FILE_MEM_EXCLUSIVE,
  1289. FILE_MEM_HARDWALL,
  1290. FILE_SCHED_LOAD_BALANCE,
  1291. FILE_SCHED_RELAX_DOMAIN_LEVEL,
  1292. FILE_MEMORY_PRESSURE_ENABLED,
  1293. FILE_MEMORY_PRESSURE,
  1294. FILE_SPREAD_PAGE,
  1295. FILE_SPREAD_SLAB,
  1296. } cpuset_filetype_t;
  1297. static int cpuset_write_u64(struct cgroup *cgrp, struct cftype *cft, u64 val)
  1298. {
  1299. int retval = 0;
  1300. struct cpuset *cs = cgroup_cs(cgrp);
  1301. cpuset_filetype_t type = cft->private;
  1302. if (!cgroup_lock_live_group(cgrp))
  1303. return -ENODEV;
  1304. switch (type) {
  1305. case FILE_CPU_EXCLUSIVE:
  1306. retval = update_flag(CS_CPU_EXCLUSIVE, cs, val);
  1307. break;
  1308. case FILE_MEM_EXCLUSIVE:
  1309. retval = update_flag(CS_MEM_EXCLUSIVE, cs, val);
  1310. break;
  1311. case FILE_MEM_HARDWALL:
  1312. retval = update_flag(CS_MEM_HARDWALL, cs, val);
  1313. break;
  1314. case FILE_SCHED_LOAD_BALANCE:
  1315. retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, val);
  1316. break;
  1317. case FILE_MEMORY_MIGRATE:
  1318. retval = update_flag(CS_MEMORY_MIGRATE, cs, val);
  1319. break;
  1320. case FILE_MEMORY_PRESSURE_ENABLED:
  1321. cpuset_memory_pressure_enabled = !!val;
  1322. break;
  1323. case FILE_MEMORY_PRESSURE:
  1324. retval = -EACCES;
  1325. break;
  1326. case FILE_SPREAD_PAGE:
  1327. retval = update_flag(CS_SPREAD_PAGE, cs, val);
  1328. break;
  1329. case FILE_SPREAD_SLAB:
  1330. retval = update_flag(CS_SPREAD_SLAB, cs, val);
  1331. break;
  1332. default:
  1333. retval = -EINVAL;
  1334. break;
  1335. }
  1336. cgroup_unlock();
  1337. return retval;
  1338. }
  1339. static int cpuset_write_s64(struct cgroup *cgrp, struct cftype *cft, s64 val)
  1340. {
  1341. int retval = 0;
  1342. struct cpuset *cs = cgroup_cs(cgrp);
  1343. cpuset_filetype_t type = cft->private;
  1344. if (!cgroup_lock_live_group(cgrp))
  1345. return -ENODEV;
  1346. switch (type) {
  1347. case FILE_SCHED_RELAX_DOMAIN_LEVEL:
  1348. retval = update_relax_domain_level(cs, val);
  1349. break;
  1350. default:
  1351. retval = -EINVAL;
  1352. break;
  1353. }
  1354. cgroup_unlock();
  1355. return retval;
  1356. }
  1357. /*
  1358. * Common handling for a write to a "cpus" or "mems" file.
  1359. */
  1360. static int cpuset_write_resmask(struct cgroup *cgrp, struct cftype *cft,
  1361. const char *buf)
  1362. {
  1363. int retval = 0;
  1364. struct cpuset *cs = cgroup_cs(cgrp);
  1365. struct cpuset *trialcs;
  1366. /*
  1367. * CPU or memory hotunplug may leave @cs w/o any execution
  1368. * resources, in which case the hotplug code asynchronously updates
  1369. * configuration and transfers all tasks to the nearest ancestor
  1370. * which can execute.
  1371. *
  1372. * As writes to "cpus" or "mems" may restore @cs's execution
  1373. * resources, wait for the previously scheduled operations before
  1374. * proceeding, so that we don't end up keep removing tasks added
  1375. * after execution capability is restored.
  1376. *
  1377. * Flushing cpuset_hotplug_work is enough to synchronize against
  1378. * hotplug hanlding; however, cpuset_attach() may schedule
  1379. * propagation work directly. Flush the workqueue too.
  1380. */
  1381. flush_work(&cpuset_hotplug_work);
  1382. flush_workqueue(cpuset_propagate_hotplug_wq);
  1383. if (!cgroup_lock_live_group(cgrp))
  1384. return -ENODEV;
  1385. trialcs = alloc_trial_cpuset(cs);
  1386. if (!trialcs) {
  1387. retval = -ENOMEM;
  1388. goto out;
  1389. }
  1390. switch (cft->private) {
  1391. case FILE_CPULIST:
  1392. retval = update_cpumask(cs, trialcs, buf);
  1393. break;
  1394. case FILE_MEMLIST:
  1395. retval = update_nodemask(cs, trialcs, buf);
  1396. break;
  1397. default:
  1398. retval = -EINVAL;
  1399. break;
  1400. }
  1401. free_trial_cpuset(trialcs);
  1402. out:
  1403. cgroup_unlock();
  1404. return retval;
  1405. }
  1406. /*
  1407. * These ascii lists should be read in a single call, by using a user
  1408. * buffer large enough to hold the entire map. If read in smaller
  1409. * chunks, there is no guarantee of atomicity. Since the display format
  1410. * used, list of ranges of sequential numbers, is variable length,
  1411. * and since these maps can change value dynamically, one could read
  1412. * gibberish by doing partial reads while a list was changing.
  1413. * A single large read to a buffer that crosses a page boundary is
  1414. * ok, because the result being copied to user land is not recomputed
  1415. * across a page fault.
  1416. */
  1417. static size_t cpuset_sprintf_cpulist(char *page, struct cpuset *cs)
  1418. {
  1419. size_t count;
  1420. mutex_lock(&callback_mutex);
  1421. count = cpulist_scnprintf(page, PAGE_SIZE, cs->cpus_allowed);
  1422. mutex_unlock(&callback_mutex);
  1423. return count;
  1424. }
  1425. static size_t cpuset_sprintf_memlist(char *page, struct cpuset *cs)
  1426. {
  1427. size_t count;
  1428. mutex_lock(&callback_mutex);
  1429. count = nodelist_scnprintf(page, PAGE_SIZE, cs->mems_allowed);
  1430. mutex_unlock(&callback_mutex);
  1431. return count;
  1432. }
  1433. static ssize_t cpuset_common_file_read(struct cgroup *cont,
  1434. struct cftype *cft,
  1435. struct file *file,
  1436. char __user *buf,
  1437. size_t nbytes, loff_t *ppos)
  1438. {
  1439. struct cpuset *cs = cgroup_cs(cont);
  1440. cpuset_filetype_t type = cft->private;
  1441. char *page;
  1442. ssize_t retval = 0;
  1443. char *s;
  1444. if (!(page = (char *)__get_free_page(GFP_TEMPORARY)))
  1445. return -ENOMEM;
  1446. s = page;
  1447. switch (type) {
  1448. case FILE_CPULIST:
  1449. s += cpuset_sprintf_cpulist(s, cs);
  1450. break;
  1451. case FILE_MEMLIST:
  1452. s += cpuset_sprintf_memlist(s, cs);
  1453. break;
  1454. default:
  1455. retval = -EINVAL;
  1456. goto out;
  1457. }
  1458. *s++ = '\n';
  1459. retval = simple_read_from_buffer(buf, nbytes, ppos, page, s - page);
  1460. out:
  1461. free_page((unsigned long)page);
  1462. return retval;
  1463. }
  1464. static u64 cpuset_read_u64(struct cgroup *cont, struct cftype *cft)
  1465. {
  1466. struct cpuset *cs = cgroup_cs(cont);
  1467. cpuset_filetype_t type = cft->private;
  1468. switch (type) {
  1469. case FILE_CPU_EXCLUSIVE:
  1470. return is_cpu_exclusive(cs);
  1471. case FILE_MEM_EXCLUSIVE:
  1472. return is_mem_exclusive(cs);
  1473. case FILE_MEM_HARDWALL:
  1474. return is_mem_hardwall(cs);
  1475. case FILE_SCHED_LOAD_BALANCE:
  1476. return is_sched_load_balance(cs);
  1477. case FILE_MEMORY_MIGRATE:
  1478. return is_memory_migrate(cs);
  1479. case FILE_MEMORY_PRESSURE_ENABLED:
  1480. return cpuset_memory_pressure_enabled;
  1481. case FILE_MEMORY_PRESSURE:
  1482. return fmeter_getrate(&cs->fmeter);
  1483. case FILE_SPREAD_PAGE:
  1484. return is_spread_page(cs);
  1485. case FILE_SPREAD_SLAB:
  1486. return is_spread_slab(cs);
  1487. default:
  1488. BUG();
  1489. }
  1490. /* Unreachable but makes gcc happy */
  1491. return 0;
  1492. }
  1493. static s64 cpuset_read_s64(struct cgroup *cont, struct cftype *cft)
  1494. {
  1495. struct cpuset *cs = cgroup_cs(cont);
  1496. cpuset_filetype_t type = cft->private;
  1497. switch (type) {
  1498. case FILE_SCHED_RELAX_DOMAIN_LEVEL:
  1499. return cs->relax_domain_level;
  1500. default:
  1501. BUG();
  1502. }
  1503. /* Unrechable but makes gcc happy */
  1504. return 0;
  1505. }
  1506. /*
  1507. * for the common functions, 'private' gives the type of file
  1508. */
  1509. static struct cftype files[] = {
  1510. {
  1511. .name = "cpus",
  1512. .read = cpuset_common_file_read,
  1513. .write_string = cpuset_write_resmask,
  1514. .max_write_len = (100U + 6 * NR_CPUS),
  1515. .private = FILE_CPULIST,
  1516. },
  1517. {
  1518. .name = "mems",
  1519. .read = cpuset_common_file_read,
  1520. .write_string = cpuset_write_resmask,
  1521. .max_write_len = (100U + 6 * MAX_NUMNODES),
  1522. .private = FILE_MEMLIST,
  1523. },
  1524. {
  1525. .name = "cpu_exclusive",
  1526. .read_u64 = cpuset_read_u64,
  1527. .write_u64 = cpuset_write_u64,
  1528. .private = FILE_CPU_EXCLUSIVE,
  1529. },
  1530. {
  1531. .name = "mem_exclusive",
  1532. .read_u64 = cpuset_read_u64,
  1533. .write_u64 = cpuset_write_u64,
  1534. .private = FILE_MEM_EXCLUSIVE,
  1535. },
  1536. {
  1537. .name = "mem_hardwall",
  1538. .read_u64 = cpuset_read_u64,
  1539. .write_u64 = cpuset_write_u64,
  1540. .private = FILE_MEM_HARDWALL,
  1541. },
  1542. {
  1543. .name = "sched_load_balance",
  1544. .read_u64 = cpuset_read_u64,
  1545. .write_u64 = cpuset_write_u64,
  1546. .private = FILE_SCHED_LOAD_BALANCE,
  1547. },
  1548. {
  1549. .name = "sched_relax_domain_level",
  1550. .read_s64 = cpuset_read_s64,
  1551. .write_s64 = cpuset_write_s64,
  1552. .private = FILE_SCHED_RELAX_DOMAIN_LEVEL,
  1553. },
  1554. {
  1555. .name = "memory_migrate",
  1556. .read_u64 = cpuset_read_u64,
  1557. .write_u64 = cpuset_write_u64,
  1558. .private = FILE_MEMORY_MIGRATE,
  1559. },
  1560. {
  1561. .name = "memory_pressure",
  1562. .read_u64 = cpuset_read_u64,
  1563. .write_u64 = cpuset_write_u64,
  1564. .private = FILE_MEMORY_PRESSURE,
  1565. .mode = S_IRUGO,
  1566. },
  1567. {
  1568. .name = "memory_spread_page",
  1569. .read_u64 = cpuset_read_u64,
  1570. .write_u64 = cpuset_write_u64,
  1571. .private = FILE_SPREAD_PAGE,
  1572. },
  1573. {
  1574. .name = "memory_spread_slab",
  1575. .read_u64 = cpuset_read_u64,
  1576. .write_u64 = cpuset_write_u64,
  1577. .private = FILE_SPREAD_SLAB,
  1578. },
  1579. {
  1580. .name = "memory_pressure_enabled",
  1581. .flags = CFTYPE_ONLY_ON_ROOT,
  1582. .read_u64 = cpuset_read_u64,
  1583. .write_u64 = cpuset_write_u64,
  1584. .private = FILE_MEMORY_PRESSURE_ENABLED,
  1585. },
  1586. { } /* terminate */
  1587. };
  1588. /*
  1589. * cpuset_css_alloc - allocate a cpuset css
  1590. * cont: control group that the new cpuset will be part of
  1591. */
  1592. static struct cgroup_subsys_state *cpuset_css_alloc(struct cgroup *cont)
  1593. {
  1594. struct cpuset *cs;
  1595. if (!cont->parent)
  1596. return &top_cpuset.css;
  1597. cs = kzalloc(sizeof(*cs), GFP_KERNEL);
  1598. if (!cs)
  1599. return ERR_PTR(-ENOMEM);
  1600. if (!alloc_cpumask_var(&cs->cpus_allowed, GFP_KERNEL)) {
  1601. kfree(cs);
  1602. return ERR_PTR(-ENOMEM);
  1603. }
  1604. set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
  1605. cpumask_clear(cs->cpus_allowed);
  1606. nodes_clear(cs->mems_allowed);
  1607. fmeter_init(&cs->fmeter);
  1608. INIT_WORK(&cs->hotplug_work, cpuset_propagate_hotplug_workfn);
  1609. cs->relax_domain_level = -1;
  1610. cs->parent = cgroup_cs(cont->parent);
  1611. return &cs->css;
  1612. }
  1613. static int cpuset_css_online(struct cgroup *cgrp)
  1614. {
  1615. struct cpuset *cs = cgroup_cs(cgrp);
  1616. struct cpuset *parent = cs->parent;
  1617. struct cpuset *tmp_cs;
  1618. struct cgroup *pos_cg;
  1619. if (!parent)
  1620. return 0;
  1621. set_bit(CS_ONLINE, &cs->flags);
  1622. if (is_spread_page(parent))
  1623. set_bit(CS_SPREAD_PAGE, &cs->flags);
  1624. if (is_spread_slab(parent))
  1625. set_bit(CS_SPREAD_SLAB, &cs->flags);
  1626. number_of_cpusets++;
  1627. if (!test_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags))
  1628. return 0;
  1629. /*
  1630. * Clone @parent's configuration if CGRP_CPUSET_CLONE_CHILDREN is
  1631. * set. This flag handling is implemented in cgroup core for
  1632. * histrical reasons - the flag may be specified during mount.
  1633. *
  1634. * Currently, if any sibling cpusets have exclusive cpus or mem, we
  1635. * refuse to clone the configuration - thereby refusing the task to
  1636. * be entered, and as a result refusing the sys_unshare() or
  1637. * clone() which initiated it. If this becomes a problem for some
  1638. * users who wish to allow that scenario, then this could be
  1639. * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
  1640. * (and likewise for mems) to the new cgroup.
  1641. */
  1642. rcu_read_lock();
  1643. cpuset_for_each_child(tmp_cs, pos_cg, parent) {
  1644. if (is_mem_exclusive(tmp_cs) || is_cpu_exclusive(tmp_cs)) {
  1645. rcu_read_unlock();
  1646. return 0;
  1647. }
  1648. }
  1649. rcu_read_unlock();
  1650. mutex_lock(&callback_mutex);
  1651. cs->mems_allowed = parent->mems_allowed;
  1652. cpumask_copy(cs->cpus_allowed, parent->cpus_allowed);
  1653. mutex_unlock(&callback_mutex);
  1654. return 0;
  1655. }
  1656. static void cpuset_css_offline(struct cgroup *cgrp)
  1657. {
  1658. struct cpuset *cs = cgroup_cs(cgrp);
  1659. /* css_offline is called w/o cgroup_mutex, grab it */
  1660. cgroup_lock();
  1661. if (is_sched_load_balance(cs))
  1662. update_flag(CS_SCHED_LOAD_BALANCE, cs, 0);
  1663. number_of_cpusets--;
  1664. clear_bit(CS_ONLINE, &cs->flags);
  1665. cgroup_unlock();
  1666. }
  1667. /*
  1668. * If the cpuset being removed has its flag 'sched_load_balance'
  1669. * enabled, then simulate turning sched_load_balance off, which
  1670. * will call rebuild_sched_domains_locked().
  1671. */
  1672. static void cpuset_css_free(struct cgroup *cont)
  1673. {
  1674. struct cpuset *cs = cgroup_cs(cont);
  1675. free_cpumask_var(cs->cpus_allowed);
  1676. kfree(cs);
  1677. }
  1678. struct cgroup_subsys cpuset_subsys = {
  1679. .name = "cpuset",
  1680. .css_alloc = cpuset_css_alloc,
  1681. .css_online = cpuset_css_online,
  1682. .css_offline = cpuset_css_offline,
  1683. .css_free = cpuset_css_free,
  1684. .can_attach = cpuset_can_attach,
  1685. .cancel_attach = cpuset_cancel_attach,
  1686. .attach = cpuset_attach,
  1687. .subsys_id = cpuset_subsys_id,
  1688. .base_cftypes = files,
  1689. .early_init = 1,
  1690. };
  1691. /**
  1692. * cpuset_init - initialize cpusets at system boot
  1693. *
  1694. * Description: Initialize top_cpuset and the cpuset internal file system,
  1695. **/
  1696. int __init cpuset_init(void)
  1697. {
  1698. int err = 0;
  1699. if (!alloc_cpumask_var(&top_cpuset.cpus_allowed, GFP_KERNEL))
  1700. BUG();
  1701. cpumask_setall(top_cpuset.cpus_allowed);
  1702. nodes_setall(top_cpuset.mems_allowed);
  1703. fmeter_init(&top_cpuset.fmeter);
  1704. set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags);
  1705. top_cpuset.relax_domain_level = -1;
  1706. err = register_filesystem(&cpuset_fs_type);
  1707. if (err < 0)
  1708. return err;
  1709. if (!alloc_cpumask_var(&cpus_attach, GFP_KERNEL))
  1710. BUG();
  1711. number_of_cpusets = 1;
  1712. return 0;
  1713. }
  1714. /**
  1715. * cpuset_do_move_task - move a given task to another cpuset
  1716. * @tsk: pointer to task_struct the task to move
  1717. * @scan: struct cgroup_scanner contained in its struct cpuset_hotplug_scanner
  1718. *
  1719. * Called by cgroup_scan_tasks() for each task in a cgroup.
  1720. * Return nonzero to stop the walk through the tasks.
  1721. */
  1722. static void cpuset_do_move_task(struct task_struct *tsk,
  1723. struct cgroup_scanner *scan)
  1724. {
  1725. struct cgroup *new_cgroup = scan->data;
  1726. cgroup_attach_task(new_cgroup, tsk);
  1727. }
  1728. /**
  1729. * move_member_tasks_to_cpuset - move tasks from one cpuset to another
  1730. * @from: cpuset in which the tasks currently reside
  1731. * @to: cpuset to which the tasks will be moved
  1732. *
  1733. * Called with cgroup_mutex held
  1734. * callback_mutex must not be held, as cpuset_attach() will take it.
  1735. *
  1736. * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
  1737. * calling callback functions for each.
  1738. */
  1739. static void move_member_tasks_to_cpuset(struct cpuset *from, struct cpuset *to)
  1740. {
  1741. struct cgroup_scanner scan;
  1742. scan.cg = from->css.cgroup;
  1743. scan.test_task = NULL; /* select all tasks in cgroup */
  1744. scan.process_task = cpuset_do_move_task;
  1745. scan.heap = NULL;
  1746. scan.data = to->css.cgroup;
  1747. if (cgroup_scan_tasks(&scan))
  1748. printk(KERN_ERR "move_member_tasks_to_cpuset: "
  1749. "cgroup_scan_tasks failed\n");
  1750. }
  1751. /*
  1752. * If CPU and/or memory hotplug handlers, below, unplug any CPUs
  1753. * or memory nodes, we need to walk over the cpuset hierarchy,
  1754. * removing that CPU or node from all cpusets. If this removes the
  1755. * last CPU or node from a cpuset, then move the tasks in the empty
  1756. * cpuset to its next-highest non-empty parent.
  1757. *
  1758. * Called with cgroup_mutex held
  1759. * callback_mutex must not be held, as cpuset_attach() will take it.
  1760. */
  1761. static void remove_tasks_in_empty_cpuset(struct cpuset *cs)
  1762. {
  1763. struct cpuset *parent;
  1764. /*
  1765. * Find its next-highest non-empty parent, (top cpuset
  1766. * has online cpus, so can't be empty).
  1767. */
  1768. parent = cs->parent;
  1769. while (cpumask_empty(parent->cpus_allowed) ||
  1770. nodes_empty(parent->mems_allowed))
  1771. parent = parent->parent;
  1772. move_member_tasks_to_cpuset(cs, parent);
  1773. }
  1774. /*
  1775. * Helper function to traverse cpusets.
  1776. * It can be used to walk the cpuset tree from top to bottom, completing
  1777. * one layer before dropping down to the next (thus always processing a
  1778. * node before any of its children).
  1779. */
  1780. static struct cpuset *cpuset_next(struct list_head *queue)
  1781. {
  1782. struct cpuset *cp;
  1783. struct cpuset *child; /* scans child cpusets of cp */
  1784. struct cgroup *cont;
  1785. if (list_empty(queue))
  1786. return NULL;
  1787. cp = list_first_entry(queue, struct cpuset, stack_list);
  1788. list_del(queue->next);
  1789. rcu_read_lock();
  1790. cpuset_for_each_child(child, cont, cp)
  1791. list_add_tail(&child->stack_list, queue);
  1792. rcu_read_unlock();
  1793. return cp;
  1794. }
  1795. /**
  1796. * cpuset_propagate_hotplug_workfn - propagate CPU/memory hotplug to a cpuset
  1797. * @cs: cpuset in interest
  1798. *
  1799. * Compare @cs's cpu and mem masks against top_cpuset and if some have gone
  1800. * offline, update @cs accordingly. If @cs ends up with no CPU or memory,
  1801. * all its tasks are moved to the nearest ancestor with both resources.
  1802. */
  1803. static void cpuset_propagate_hotplug_workfn(struct work_struct *work)
  1804. {
  1805. static cpumask_t off_cpus;
  1806. static nodemask_t off_mems, tmp_mems;
  1807. struct cpuset *cs = container_of(work, struct cpuset, hotplug_work);
  1808. cgroup_lock();
  1809. cpumask_andnot(&off_cpus, cs->cpus_allowed, top_cpuset.cpus_allowed);
  1810. nodes_andnot(off_mems, cs->mems_allowed, top_cpuset.mems_allowed);
  1811. /* remove offline cpus from @cs */
  1812. if (!cpumask_empty(&off_cpus)) {
  1813. mutex_lock(&callback_mutex);
  1814. cpumask_andnot(cs->cpus_allowed, cs->cpus_allowed, &off_cpus);
  1815. mutex_unlock(&callback_mutex);
  1816. update_tasks_cpumask(cs, NULL);
  1817. }
  1818. /* remove offline mems from @cs */
  1819. if (!nodes_empty(off_mems)) {
  1820. tmp_mems = cs->mems_allowed;
  1821. mutex_lock(&callback_mutex);
  1822. nodes_andnot(cs->mems_allowed, cs->mems_allowed, off_mems);
  1823. mutex_unlock(&callback_mutex);
  1824. update_tasks_nodemask(cs, &tmp_mems, NULL);
  1825. }
  1826. if (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed))
  1827. remove_tasks_in_empty_cpuset(cs);
  1828. cgroup_unlock();
  1829. /* the following may free @cs, should be the last operation */
  1830. css_put(&cs->css);
  1831. }
  1832. /**
  1833. * schedule_cpuset_propagate_hotplug - schedule hotplug propagation to a cpuset
  1834. * @cs: cpuset of interest
  1835. *
  1836. * Schedule cpuset_propagate_hotplug_workfn() which will update CPU and
  1837. * memory masks according to top_cpuset.
  1838. */
  1839. static void schedule_cpuset_propagate_hotplug(struct cpuset *cs)
  1840. {
  1841. /*
  1842. * Pin @cs. The refcnt will be released when the work item
  1843. * finishes executing.
  1844. */
  1845. if (!css_tryget(&cs->css))
  1846. return;
  1847. /*
  1848. * Queue @cs->hotplug_work. If already pending, lose the css ref.
  1849. * cpuset_propagate_hotplug_wq is ordered and propagation will
  1850. * happen in the order this function is called.
  1851. */
  1852. if (!queue_work(cpuset_propagate_hotplug_wq, &cs->hotplug_work))
  1853. css_put(&cs->css);
  1854. }
  1855. /**
  1856. * cpuset_hotplug_workfn - handle CPU/memory hotunplug for a cpuset
  1857. *
  1858. * This function is called after either CPU or memory configuration has
  1859. * changed and updates cpuset accordingly. The top_cpuset is always
  1860. * synchronized to cpu_active_mask and N_MEMORY, which is necessary in
  1861. * order to make cpusets transparent (of no affect) on systems that are
  1862. * actively using CPU hotplug but making no active use of cpusets.
  1863. *
  1864. * Non-root cpusets are only affected by offlining. If any CPUs or memory
  1865. * nodes have been taken down, cpuset_propagate_hotplug() is invoked on all
  1866. * descendants.
  1867. *
  1868. * Note that CPU offlining during suspend is ignored. We don't modify
  1869. * cpusets across suspend/resume cycles at all.
  1870. */
  1871. static void cpuset_hotplug_workfn(struct work_struct *work)
  1872. {
  1873. static cpumask_t new_cpus, tmp_cpus;
  1874. static nodemask_t new_mems, tmp_mems;
  1875. bool cpus_updated, mems_updated;
  1876. bool cpus_offlined, mems_offlined;
  1877. cgroup_lock();
  1878. /* fetch the available cpus/mems and find out which changed how */
  1879. cpumask_copy(&new_cpus, cpu_active_mask);
  1880. new_mems = node_states[N_MEMORY];
  1881. cpus_updated = !cpumask_equal(top_cpuset.cpus_allowed, &new_cpus);
  1882. cpus_offlined = cpumask_andnot(&tmp_cpus, top_cpuset.cpus_allowed,
  1883. &new_cpus);
  1884. mems_updated = !nodes_equal(top_cpuset.mems_allowed, new_mems);
  1885. nodes_andnot(tmp_mems, top_cpuset.mems_allowed, new_mems);
  1886. mems_offlined = !nodes_empty(tmp_mems);
  1887. /* synchronize cpus_allowed to cpu_active_mask */
  1888. if (cpus_updated) {
  1889. mutex_lock(&callback_mutex);
  1890. cpumask_copy(top_cpuset.cpus_allowed, &new_cpus);
  1891. mutex_unlock(&callback_mutex);
  1892. /* we don't mess with cpumasks of tasks in top_cpuset */
  1893. }
  1894. /* synchronize mems_allowed to N_MEMORY */
  1895. if (mems_updated) {
  1896. tmp_mems = top_cpuset.mems_allowed;
  1897. mutex_lock(&callback_mutex);
  1898. top_cpuset.mems_allowed = new_mems;
  1899. mutex_unlock(&callback_mutex);
  1900. update_tasks_nodemask(&top_cpuset, &tmp_mems, NULL);
  1901. }
  1902. /* if cpus or mems went down, we need to propagate to descendants */
  1903. if (cpus_offlined || mems_offlined) {
  1904. struct cpuset *cs;
  1905. LIST_HEAD(queue);
  1906. list_add_tail(&top_cpuset.stack_list, &queue);
  1907. while ((cs = cpuset_next(&queue)))
  1908. if (cs != &top_cpuset)
  1909. schedule_cpuset_propagate_hotplug(cs);
  1910. }
  1911. cgroup_unlock();
  1912. /* wait for propagations to finish */
  1913. flush_workqueue(cpuset_propagate_hotplug_wq);
  1914. /* rebuild sched domains if cpus_allowed has changed */
  1915. if (cpus_updated) {
  1916. struct sched_domain_attr *attr;
  1917. cpumask_var_t *doms;
  1918. int ndoms;
  1919. cgroup_lock();
  1920. ndoms = generate_sched_domains(&doms, &attr);
  1921. cgroup_unlock();
  1922. partition_sched_domains(ndoms, doms, attr);
  1923. }
  1924. }
  1925. void cpuset_update_active_cpus(bool cpu_online)
  1926. {
  1927. /*
  1928. * We're inside cpu hotplug critical region which usually nests
  1929. * inside cgroup synchronization. Bounce actual hotplug processing
  1930. * to a work item to avoid reverse locking order.
  1931. *
  1932. * We still need to do partition_sched_domains() synchronously;
  1933. * otherwise, the scheduler will get confused and put tasks to the
  1934. * dead CPU. Fall back to the default single domain.
  1935. * cpuset_hotplug_workfn() will rebuild it as necessary.
  1936. */
  1937. partition_sched_domains(1, NULL, NULL);
  1938. schedule_work(&cpuset_hotplug_work);
  1939. }
  1940. #ifdef CONFIG_MEMORY_HOTPLUG
  1941. /*
  1942. * Keep top_cpuset.mems_allowed tracking node_states[N_MEMORY].
  1943. * Call this routine anytime after node_states[N_MEMORY] changes.
  1944. * See cpuset_update_active_cpus() for CPU hotplug handling.
  1945. */
  1946. static int cpuset_track_online_nodes(struct notifier_block *self,
  1947. unsigned long action, void *arg)
  1948. {
  1949. schedule_work(&cpuset_hotplug_work);
  1950. return NOTIFY_OK;
  1951. }
  1952. #endif
  1953. /**
  1954. * cpuset_init_smp - initialize cpus_allowed
  1955. *
  1956. * Description: Finish top cpuset after cpu, node maps are initialized
  1957. **/
  1958. void __init cpuset_init_smp(void)
  1959. {
  1960. cpumask_copy(top_cpuset.cpus_allowed, cpu_active_mask);
  1961. top_cpuset.mems_allowed = node_states[N_MEMORY];
  1962. hotplug_memory_notifier(cpuset_track_online_nodes, 10);
  1963. cpuset_propagate_hotplug_wq =
  1964. alloc_ordered_workqueue("cpuset_hotplug", 0);
  1965. BUG_ON(!cpuset_propagate_hotplug_wq);
  1966. }
  1967. /**
  1968. * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
  1969. * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
  1970. * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
  1971. *
  1972. * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
  1973. * attached to the specified @tsk. Guaranteed to return some non-empty
  1974. * subset of cpu_online_mask, even if this means going outside the
  1975. * tasks cpuset.
  1976. **/
  1977. void cpuset_cpus_allowed(struct task_struct *tsk, struct cpumask *pmask)
  1978. {
  1979. mutex_lock(&callback_mutex);
  1980. task_lock(tsk);
  1981. guarantee_online_cpus(task_cs(tsk), pmask);
  1982. task_unlock(tsk);
  1983. mutex_unlock(&callback_mutex);
  1984. }
  1985. void cpuset_cpus_allowed_fallback(struct task_struct *tsk)
  1986. {
  1987. const struct cpuset *cs;
  1988. rcu_read_lock();
  1989. cs = task_cs(tsk);
  1990. if (cs)
  1991. do_set_cpus_allowed(tsk, cs->cpus_allowed);
  1992. rcu_read_unlock();
  1993. /*
  1994. * We own tsk->cpus_allowed, nobody can change it under us.
  1995. *
  1996. * But we used cs && cs->cpus_allowed lockless and thus can
  1997. * race with cgroup_attach_task() or update_cpumask() and get
  1998. * the wrong tsk->cpus_allowed. However, both cases imply the
  1999. * subsequent cpuset_change_cpumask()->set_cpus_allowed_ptr()
  2000. * which takes task_rq_lock().
  2001. *
  2002. * If we are called after it dropped the lock we must see all
  2003. * changes in tsk_cs()->cpus_allowed. Otherwise we can temporary
  2004. * set any mask even if it is not right from task_cs() pov,
  2005. * the pending set_cpus_allowed_ptr() will fix things.
  2006. *
  2007. * select_fallback_rq() will fix things ups and set cpu_possible_mask
  2008. * if required.
  2009. */
  2010. }
  2011. void cpuset_init_current_mems_allowed(void)
  2012. {
  2013. nodes_setall(current->mems_allowed);
  2014. }
  2015. /**
  2016. * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
  2017. * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
  2018. *
  2019. * Description: Returns the nodemask_t mems_allowed of the cpuset
  2020. * attached to the specified @tsk. Guaranteed to return some non-empty
  2021. * subset of node_states[N_MEMORY], even if this means going outside the
  2022. * tasks cpuset.
  2023. **/
  2024. nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
  2025. {
  2026. nodemask_t mask;
  2027. mutex_lock(&callback_mutex);
  2028. task_lock(tsk);
  2029. guarantee_online_mems(task_cs(tsk), &mask);
  2030. task_unlock(tsk);
  2031. mutex_unlock(&callback_mutex);
  2032. return mask;
  2033. }
  2034. /**
  2035. * cpuset_nodemask_valid_mems_allowed - check nodemask vs. curremt mems_allowed
  2036. * @nodemask: the nodemask to be checked
  2037. *
  2038. * Are any of the nodes in the nodemask allowed in current->mems_allowed?
  2039. */
  2040. int cpuset_nodemask_valid_mems_allowed(nodemask_t *nodemask)
  2041. {
  2042. return nodes_intersects(*nodemask, current->mems_allowed);
  2043. }
  2044. /*
  2045. * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
  2046. * mem_hardwall ancestor to the specified cpuset. Call holding
  2047. * callback_mutex. If no ancestor is mem_exclusive or mem_hardwall
  2048. * (an unusual configuration), then returns the root cpuset.
  2049. */
  2050. static const struct cpuset *nearest_hardwall_ancestor(const struct cpuset *cs)
  2051. {
  2052. while (!(is_mem_exclusive(cs) || is_mem_hardwall(cs)) && cs->parent)
  2053. cs = cs->parent;
  2054. return cs;
  2055. }
  2056. /**
  2057. * cpuset_node_allowed_softwall - Can we allocate on a memory node?
  2058. * @node: is this an allowed node?
  2059. * @gfp_mask: memory allocation flags
  2060. *
  2061. * If we're in interrupt, yes, we can always allocate. If __GFP_THISNODE is
  2062. * set, yes, we can always allocate. If node is in our task's mems_allowed,
  2063. * yes. If it's not a __GFP_HARDWALL request and this node is in the nearest
  2064. * hardwalled cpuset ancestor to this task's cpuset, yes. If the task has been
  2065. * OOM killed and has access to memory reserves as specified by the TIF_MEMDIE
  2066. * flag, yes.
  2067. * Otherwise, no.
  2068. *
  2069. * If __GFP_HARDWALL is set, cpuset_node_allowed_softwall() reduces to
  2070. * cpuset_node_allowed_hardwall(). Otherwise, cpuset_node_allowed_softwall()
  2071. * might sleep, and might allow a node from an enclosing cpuset.
  2072. *
  2073. * cpuset_node_allowed_hardwall() only handles the simpler case of hardwall
  2074. * cpusets, and never sleeps.
  2075. *
  2076. * The __GFP_THISNODE placement logic is really handled elsewhere,
  2077. * by forcibly using a zonelist starting at a specified node, and by
  2078. * (in get_page_from_freelist()) refusing to consider the zones for
  2079. * any node on the zonelist except the first. By the time any such
  2080. * calls get to this routine, we should just shut up and say 'yes'.
  2081. *
  2082. * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
  2083. * and do not allow allocations outside the current tasks cpuset
  2084. * unless the task has been OOM killed as is marked TIF_MEMDIE.
  2085. * GFP_KERNEL allocations are not so marked, so can escape to the
  2086. * nearest enclosing hardwalled ancestor cpuset.
  2087. *
  2088. * Scanning up parent cpusets requires callback_mutex. The
  2089. * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
  2090. * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
  2091. * current tasks mems_allowed came up empty on the first pass over
  2092. * the zonelist. So only GFP_KERNEL allocations, if all nodes in the
  2093. * cpuset are short of memory, might require taking the callback_mutex
  2094. * mutex.
  2095. *
  2096. * The first call here from mm/page_alloc:get_page_from_freelist()
  2097. * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
  2098. * so no allocation on a node outside the cpuset is allowed (unless
  2099. * in interrupt, of course).
  2100. *
  2101. * The second pass through get_page_from_freelist() doesn't even call
  2102. * here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
  2103. * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
  2104. * in alloc_flags. That logic and the checks below have the combined
  2105. * affect that:
  2106. * in_interrupt - any node ok (current task context irrelevant)
  2107. * GFP_ATOMIC - any node ok
  2108. * TIF_MEMDIE - any node ok
  2109. * GFP_KERNEL - any node in enclosing hardwalled cpuset ok
  2110. * GFP_USER - only nodes in current tasks mems allowed ok.
  2111. *
  2112. * Rule:
  2113. * Don't call cpuset_node_allowed_softwall if you can't sleep, unless you
  2114. * pass in the __GFP_HARDWALL flag set in gfp_flag, which disables
  2115. * the code that might scan up ancestor cpusets and sleep.
  2116. */
  2117. int __cpuset_node_allowed_softwall(int node, gfp_t gfp_mask)
  2118. {
  2119. const struct cpuset *cs; /* current cpuset ancestors */
  2120. int allowed; /* is allocation in zone z allowed? */
  2121. if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
  2122. return 1;
  2123. might_sleep_if(!(gfp_mask & __GFP_HARDWALL));
  2124. if (node_isset(node, current->mems_allowed))
  2125. return 1;
  2126. /*
  2127. * Allow tasks that have access to memory reserves because they have
  2128. * been OOM killed to get memory anywhere.
  2129. */
  2130. if (unlikely(test_thread_flag(TIF_MEMDIE)))
  2131. return 1;
  2132. if (gfp_mask & __GFP_HARDWALL) /* If hardwall request, stop here */
  2133. return 0;
  2134. if (current->flags & PF_EXITING) /* Let dying task have memory */
  2135. return 1;
  2136. /* Not hardwall and node outside mems_allowed: scan up cpusets */
  2137. mutex_lock(&callback_mutex);
  2138. task_lock(current);
  2139. cs = nearest_hardwall_ancestor(task_cs(current));
  2140. task_unlock(current);
  2141. allowed = node_isset(node, cs->mems_allowed);
  2142. mutex_unlock(&callback_mutex);
  2143. return allowed;
  2144. }
  2145. /*
  2146. * cpuset_node_allowed_hardwall - Can we allocate on a memory node?
  2147. * @node: is this an allowed node?
  2148. * @gfp_mask: memory allocation flags
  2149. *
  2150. * If we're in interrupt, yes, we can always allocate. If __GFP_THISNODE is
  2151. * set, yes, we can always allocate. If node is in our task's mems_allowed,
  2152. * yes. If the task has been OOM killed and has access to memory reserves as
  2153. * specified by the TIF_MEMDIE flag, yes.
  2154. * Otherwise, no.
  2155. *
  2156. * The __GFP_THISNODE placement logic is really handled elsewhere,
  2157. * by forcibly using a zonelist starting at a specified node, and by
  2158. * (in get_page_from_freelist()) refusing to consider the zones for
  2159. * any node on the zonelist except the first. By the time any such
  2160. * calls get to this routine, we should just shut up and say 'yes'.
  2161. *
  2162. * Unlike the cpuset_node_allowed_softwall() variant, above,
  2163. * this variant requires that the node be in the current task's
  2164. * mems_allowed or that we're in interrupt. It does not scan up the
  2165. * cpuset hierarchy for the nearest enclosing mem_exclusive cpuset.
  2166. * It never sleeps.
  2167. */
  2168. int __cpuset_node_allowed_hardwall(int node, gfp_t gfp_mask)
  2169. {
  2170. if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
  2171. return 1;
  2172. if (node_isset(node, current->mems_allowed))
  2173. return 1;
  2174. /*
  2175. * Allow tasks that have access to memory reserves because they have
  2176. * been OOM killed to get memory anywhere.
  2177. */
  2178. if (unlikely(test_thread_flag(TIF_MEMDIE)))
  2179. return 1;
  2180. return 0;
  2181. }
  2182. /**
  2183. * cpuset_mem_spread_node() - On which node to begin search for a file page
  2184. * cpuset_slab_spread_node() - On which node to begin search for a slab page
  2185. *
  2186. * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
  2187. * tasks in a cpuset with is_spread_page or is_spread_slab set),
  2188. * and if the memory allocation used cpuset_mem_spread_node()
  2189. * to determine on which node to start looking, as it will for
  2190. * certain page cache or slab cache pages such as used for file
  2191. * system buffers and inode caches, then instead of starting on the
  2192. * local node to look for a free page, rather spread the starting
  2193. * node around the tasks mems_allowed nodes.
  2194. *
  2195. * We don't have to worry about the returned node being offline
  2196. * because "it can't happen", and even if it did, it would be ok.
  2197. *
  2198. * The routines calling guarantee_online_mems() are careful to
  2199. * only set nodes in task->mems_allowed that are online. So it
  2200. * should not be possible for the following code to return an
  2201. * offline node. But if it did, that would be ok, as this routine
  2202. * is not returning the node where the allocation must be, only
  2203. * the node where the search should start. The zonelist passed to
  2204. * __alloc_pages() will include all nodes. If the slab allocator
  2205. * is passed an offline node, it will fall back to the local node.
  2206. * See kmem_cache_alloc_node().
  2207. */
  2208. static int cpuset_spread_node(int *rotor)
  2209. {
  2210. int node;
  2211. node = next_node(*rotor, current->mems_allowed);
  2212. if (node == MAX_NUMNODES)
  2213. node = first_node(current->mems_allowed);
  2214. *rotor = node;
  2215. return node;
  2216. }
  2217. int cpuset_mem_spread_node(void)
  2218. {
  2219. if (current->cpuset_mem_spread_rotor == NUMA_NO_NODE)
  2220. current->cpuset_mem_spread_rotor =
  2221. node_random(&current->mems_allowed);
  2222. return cpuset_spread_node(&current->cpuset_mem_spread_rotor);
  2223. }
  2224. int cpuset_slab_spread_node(void)
  2225. {
  2226. if (current->cpuset_slab_spread_rotor == NUMA_NO_NODE)
  2227. current->cpuset_slab_spread_rotor =
  2228. node_random(&current->mems_allowed);
  2229. return cpuset_spread_node(&current->cpuset_slab_spread_rotor);
  2230. }
  2231. EXPORT_SYMBOL_GPL(cpuset_mem_spread_node);
  2232. /**
  2233. * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
  2234. * @tsk1: pointer to task_struct of some task.
  2235. * @tsk2: pointer to task_struct of some other task.
  2236. *
  2237. * Description: Return true if @tsk1's mems_allowed intersects the
  2238. * mems_allowed of @tsk2. Used by the OOM killer to determine if
  2239. * one of the task's memory usage might impact the memory available
  2240. * to the other.
  2241. **/
  2242. int cpuset_mems_allowed_intersects(const struct task_struct *tsk1,
  2243. const struct task_struct *tsk2)
  2244. {
  2245. return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed);
  2246. }
  2247. /**
  2248. * cpuset_print_task_mems_allowed - prints task's cpuset and mems_allowed
  2249. * @task: pointer to task_struct of some task.
  2250. *
  2251. * Description: Prints @task's name, cpuset name, and cached copy of its
  2252. * mems_allowed to the kernel log. Must hold task_lock(task) to allow
  2253. * dereferencing task_cs(task).
  2254. */
  2255. void cpuset_print_task_mems_allowed(struct task_struct *tsk)
  2256. {
  2257. struct dentry *dentry;
  2258. dentry = task_cs(tsk)->css.cgroup->dentry;
  2259. spin_lock(&cpuset_buffer_lock);
  2260. snprintf(cpuset_name, CPUSET_NAME_LEN,
  2261. dentry ? (const char *)dentry->d_name.name : "/");
  2262. nodelist_scnprintf(cpuset_nodelist, CPUSET_NODELIST_LEN,
  2263. tsk->mems_allowed);
  2264. printk(KERN_INFO "%s cpuset=%s mems_allowed=%s\n",
  2265. tsk->comm, cpuset_name, cpuset_nodelist);
  2266. spin_unlock(&cpuset_buffer_lock);
  2267. }
  2268. /*
  2269. * Collection of memory_pressure is suppressed unless
  2270. * this flag is enabled by writing "1" to the special
  2271. * cpuset file 'memory_pressure_enabled' in the root cpuset.
  2272. */
  2273. int cpuset_memory_pressure_enabled __read_mostly;
  2274. /**
  2275. * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
  2276. *
  2277. * Keep a running average of the rate of synchronous (direct)
  2278. * page reclaim efforts initiated by tasks in each cpuset.
  2279. *
  2280. * This represents the rate at which some task in the cpuset
  2281. * ran low on memory on all nodes it was allowed to use, and
  2282. * had to enter the kernels page reclaim code in an effort to
  2283. * create more free memory by tossing clean pages or swapping
  2284. * or writing dirty pages.
  2285. *
  2286. * Display to user space in the per-cpuset read-only file
  2287. * "memory_pressure". Value displayed is an integer
  2288. * representing the recent rate of entry into the synchronous
  2289. * (direct) page reclaim by any task attached to the cpuset.
  2290. **/
  2291. void __cpuset_memory_pressure_bump(void)
  2292. {
  2293. task_lock(current);
  2294. fmeter_markevent(&task_cs(current)->fmeter);
  2295. task_unlock(current);
  2296. }
  2297. #ifdef CONFIG_PROC_PID_CPUSET
  2298. /*
  2299. * proc_cpuset_show()
  2300. * - Print tasks cpuset path into seq_file.
  2301. * - Used for /proc/<pid>/cpuset.
  2302. * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
  2303. * doesn't really matter if tsk->cpuset changes after we read it,
  2304. * and we take cgroup_mutex, keeping cpuset_attach() from changing it
  2305. * anyway.
  2306. */
  2307. static int proc_cpuset_show(struct seq_file *m, void *unused_v)
  2308. {
  2309. struct pid *pid;
  2310. struct task_struct *tsk;
  2311. char *buf;
  2312. struct cgroup_subsys_state *css;
  2313. int retval;
  2314. retval = -ENOMEM;
  2315. buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
  2316. if (!buf)
  2317. goto out;
  2318. retval = -ESRCH;
  2319. pid = m->private;
  2320. tsk = get_pid_task(pid, PIDTYPE_PID);
  2321. if (!tsk)
  2322. goto out_free;
  2323. retval = -EINVAL;
  2324. cgroup_lock();
  2325. css = task_subsys_state(tsk, cpuset_subsys_id);
  2326. retval = cgroup_path(css->cgroup, buf, PAGE_SIZE);
  2327. if (retval < 0)
  2328. goto out_unlock;
  2329. seq_puts(m, buf);
  2330. seq_putc(m, '\n');
  2331. out_unlock:
  2332. cgroup_unlock();
  2333. put_task_struct(tsk);
  2334. out_free:
  2335. kfree(buf);
  2336. out:
  2337. return retval;
  2338. }
  2339. static int cpuset_open(struct inode *inode, struct file *file)
  2340. {
  2341. struct pid *pid = PROC_I(inode)->pid;
  2342. return single_open(file, proc_cpuset_show, pid);
  2343. }
  2344. const struct file_operations proc_cpuset_operations = {
  2345. .open = cpuset_open,
  2346. .read = seq_read,
  2347. .llseek = seq_lseek,
  2348. .release = single_release,
  2349. };
  2350. #endif /* CONFIG_PROC_PID_CPUSET */
  2351. /* Display task mems_allowed in /proc/<pid>/status file. */
  2352. void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task)
  2353. {
  2354. seq_printf(m, "Mems_allowed:\t");
  2355. seq_nodemask(m, &task->mems_allowed);
  2356. seq_printf(m, "\n");
  2357. seq_printf(m, "Mems_allowed_list:\t");
  2358. seq_nodemask_list(m, &task->mems_allowed);
  2359. seq_printf(m, "\n");
  2360. }