cpuset.c 75 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767768769770771772773774775776777778779780781782783784785786787788789790791792793794795796797798799800801802803804805806807808809810811812813814815816817818819820821822823824825826827828829830831832833834835836837838839840841842843844845846847848849850851852853854855856857858859860861862863864865866867868869870871872873874875876877878879880881882883884885886887888889890891892893894895896897898899900901902903904905906907908909910911912913914915916917918919920921922923924925926927928929930931932933934935936937938939940941942943944945946947948949950951952953954955956957958959960961962963964965966967968969970971972973974975976977978979980981982983984985986987988989990991992993994995996997998999100010011002100310041005100610071008100910101011101210131014101510161017101810191020102110221023102410251026102710281029103010311032103310341035103610371038103910401041104210431044104510461047104810491050105110521053105410551056105710581059106010611062106310641065106610671068106910701071107210731074107510761077107810791080108110821083108410851086108710881089109010911092109310941095109610971098109911001101110211031104110511061107110811091110111111121113111411151116111711181119112011211122112311241125112611271128112911301131113211331134113511361137113811391140114111421143114411451146114711481149115011511152115311541155115611571158115911601161116211631164116511661167116811691170117111721173117411751176117711781179118011811182118311841185118611871188118911901191119211931194119511961197119811991200120112021203120412051206120712081209121012111212121312141215121612171218121912201221122212231224122512261227122812291230123112321233123412351236123712381239124012411242124312441245124612471248124912501251125212531254125512561257125812591260126112621263126412651266126712681269127012711272127312741275127612771278127912801281128212831284128512861287128812891290129112921293129412951296129712981299130013011302130313041305130613071308130913101311131213131314131513161317131813191320132113221323132413251326132713281329133013311332133313341335133613371338133913401341134213431344134513461347134813491350135113521353135413551356135713581359136013611362136313641365136613671368136913701371137213731374137513761377137813791380138113821383138413851386138713881389139013911392139313941395139613971398139914001401140214031404140514061407140814091410141114121413141414151416141714181419142014211422142314241425142614271428142914301431143214331434143514361437143814391440144114421443144414451446144714481449145014511452145314541455145614571458145914601461146214631464146514661467146814691470147114721473147414751476147714781479148014811482148314841485148614871488148914901491149214931494149514961497149814991500150115021503150415051506150715081509151015111512151315141515151615171518151915201521152215231524152515261527152815291530153115321533153415351536153715381539154015411542154315441545154615471548154915501551155215531554155515561557155815591560156115621563156415651566156715681569157015711572157315741575157615771578157915801581158215831584158515861587158815891590159115921593159415951596159715981599160016011602160316041605160616071608160916101611161216131614161516161617161816191620162116221623162416251626162716281629163016311632163316341635163616371638163916401641164216431644164516461647164816491650165116521653165416551656165716581659166016611662166316641665166616671668166916701671167216731674167516761677167816791680168116821683168416851686168716881689169016911692169316941695169616971698169917001701170217031704170517061707170817091710171117121713171417151716171717181719172017211722172317241725172617271728172917301731173217331734173517361737173817391740174117421743174417451746174717481749175017511752175317541755175617571758175917601761176217631764176517661767176817691770177117721773177417751776177717781779178017811782178317841785178617871788178917901791179217931794179517961797179817991800180118021803180418051806180718081809181018111812181318141815181618171818181918201821182218231824182518261827182818291830183118321833183418351836183718381839184018411842184318441845184618471848184918501851185218531854185518561857185818591860186118621863186418651866186718681869187018711872187318741875187618771878187918801881188218831884188518861887188818891890189118921893189418951896189718981899190019011902190319041905190619071908190919101911191219131914191519161917191819191920192119221923192419251926192719281929193019311932193319341935193619371938193919401941194219431944194519461947194819491950195119521953195419551956195719581959196019611962196319641965196619671968196919701971197219731974197519761977197819791980198119821983198419851986198719881989199019911992199319941995199619971998199920002001200220032004200520062007200820092010201120122013201420152016201720182019202020212022202320242025202620272028202920302031203220332034203520362037203820392040204120422043204420452046204720482049205020512052205320542055205620572058205920602061206220632064206520662067206820692070207120722073207420752076207720782079208020812082208320842085208620872088208920902091209220932094209520962097209820992100210121022103210421052106210721082109211021112112211321142115211621172118211921202121212221232124212521262127212821292130213121322133213421352136213721382139214021412142214321442145214621472148214921502151215221532154215521562157215821592160216121622163216421652166216721682169217021712172217321742175217621772178217921802181218221832184218521862187218821892190219121922193219421952196219721982199220022012202220322042205220622072208220922102211221222132214221522162217221822192220222122222223222422252226222722282229223022312232223322342235223622372238223922402241224222432244224522462247224822492250225122522253225422552256225722582259226022612262226322642265226622672268226922702271227222732274227522762277227822792280228122822283228422852286228722882289229022912292229322942295229622972298229923002301230223032304230523062307230823092310231123122313231423152316231723182319232023212322232323242325232623272328232923302331233223332334233523362337233823392340234123422343234423452346234723482349235023512352235323542355235623572358235923602361236223632364236523662367236823692370237123722373237423752376237723782379238023812382238323842385238623872388238923902391239223932394239523962397239823992400240124022403240424052406240724082409241024112412241324142415241624172418241924202421242224232424242524262427242824292430243124322433243424352436243724382439244024412442244324442445244624472448244924502451245224532454245524562457245824592460246124622463246424652466246724682469247024712472247324742475247624772478247924802481248224832484248524862487248824892490249124922493249424952496249724982499250025012502250325042505250625072508250925102511251225132514251525162517251825192520252125222523252425252526252725282529253025312532253325342535253625372538253925402541254225432544254525462547254825492550255125522553255425552556255725582559256025612562256325642565256625672568256925702571257225732574257525762577257825792580258125822583258425852586258725882589259025912592259325942595259625972598259926002601260226032604260526062607260826092610261126122613261426152616261726182619262026212622262326242625262626272628262926302631263226332634263526362637263826392640264126422643264426452646264726482649265026512652265326542655265626572658265926602661266226632664266526662667
  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 fmeter fmeter; /* memory_pressure filter */
  82. /*
  83. * Tasks are being attached to this cpuset. Used to prevent
  84. * zeroing cpus/mems_allowed between ->can_attach() and ->attach().
  85. */
  86. int attach_in_progress;
  87. /* partition number for rebuild_sched_domains() */
  88. int pn;
  89. /* for custom sched domain */
  90. int relax_domain_level;
  91. struct work_struct hotplug_work;
  92. };
  93. /* Retrieve the cpuset for a cgroup */
  94. static inline struct cpuset *cgroup_cs(struct cgroup *cont)
  95. {
  96. return container_of(cgroup_subsys_state(cont, cpuset_subsys_id),
  97. struct cpuset, css);
  98. }
  99. /* Retrieve the cpuset for a task */
  100. static inline struct cpuset *task_cs(struct task_struct *task)
  101. {
  102. return container_of(task_subsys_state(task, cpuset_subsys_id),
  103. struct cpuset, css);
  104. }
  105. static inline struct cpuset *parent_cs(const struct cpuset *cs)
  106. {
  107. struct cgroup *pcgrp = cs->css.cgroup->parent;
  108. if (pcgrp)
  109. return cgroup_cs(pcgrp);
  110. return NULL;
  111. }
  112. #ifdef CONFIG_NUMA
  113. static inline bool task_has_mempolicy(struct task_struct *task)
  114. {
  115. return task->mempolicy;
  116. }
  117. #else
  118. static inline bool task_has_mempolicy(struct task_struct *task)
  119. {
  120. return false;
  121. }
  122. #endif
  123. /* bits in struct cpuset flags field */
  124. typedef enum {
  125. CS_ONLINE,
  126. CS_CPU_EXCLUSIVE,
  127. CS_MEM_EXCLUSIVE,
  128. CS_MEM_HARDWALL,
  129. CS_MEMORY_MIGRATE,
  130. CS_SCHED_LOAD_BALANCE,
  131. CS_SPREAD_PAGE,
  132. CS_SPREAD_SLAB,
  133. } cpuset_flagbits_t;
  134. /* convenient tests for these bits */
  135. static inline bool is_cpuset_online(const struct cpuset *cs)
  136. {
  137. return test_bit(CS_ONLINE, &cs->flags);
  138. }
  139. static inline int is_cpu_exclusive(const struct cpuset *cs)
  140. {
  141. return test_bit(CS_CPU_EXCLUSIVE, &cs->flags);
  142. }
  143. static inline int is_mem_exclusive(const struct cpuset *cs)
  144. {
  145. return test_bit(CS_MEM_EXCLUSIVE, &cs->flags);
  146. }
  147. static inline int is_mem_hardwall(const struct cpuset *cs)
  148. {
  149. return test_bit(CS_MEM_HARDWALL, &cs->flags);
  150. }
  151. static inline int is_sched_load_balance(const struct cpuset *cs)
  152. {
  153. return test_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
  154. }
  155. static inline int is_memory_migrate(const struct cpuset *cs)
  156. {
  157. return test_bit(CS_MEMORY_MIGRATE, &cs->flags);
  158. }
  159. static inline int is_spread_page(const struct cpuset *cs)
  160. {
  161. return test_bit(CS_SPREAD_PAGE, &cs->flags);
  162. }
  163. static inline int is_spread_slab(const struct cpuset *cs)
  164. {
  165. return test_bit(CS_SPREAD_SLAB, &cs->flags);
  166. }
  167. static struct cpuset top_cpuset = {
  168. .flags = ((1 << CS_ONLINE) | (1 << CS_CPU_EXCLUSIVE) |
  169. (1 << CS_MEM_EXCLUSIVE)),
  170. };
  171. /**
  172. * cpuset_for_each_child - traverse online children of a cpuset
  173. * @child_cs: loop cursor pointing to the current child
  174. * @pos_cgrp: used for iteration
  175. * @parent_cs: target cpuset to walk children of
  176. *
  177. * Walk @child_cs through the online children of @parent_cs. Must be used
  178. * with RCU read locked.
  179. */
  180. #define cpuset_for_each_child(child_cs, pos_cgrp, parent_cs) \
  181. cgroup_for_each_child((pos_cgrp), (parent_cs)->css.cgroup) \
  182. if (is_cpuset_online(((child_cs) = cgroup_cs((pos_cgrp)))))
  183. /**
  184. * cpuset_for_each_descendant_pre - pre-order walk of a cpuset's descendants
  185. * @des_cs: loop cursor pointing to the current descendant
  186. * @pos_cgrp: used for iteration
  187. * @root_cs: target cpuset to walk ancestor of
  188. *
  189. * Walk @des_cs through the online descendants of @root_cs. Must be used
  190. * with RCU read locked. The caller may modify @pos_cgrp by calling
  191. * cgroup_rightmost_descendant() to skip subtree.
  192. */
  193. #define cpuset_for_each_descendant_pre(des_cs, pos_cgrp, root_cs) \
  194. cgroup_for_each_descendant_pre((pos_cgrp), (root_cs)->css.cgroup) \
  195. if (is_cpuset_online(((des_cs) = cgroup_cs((pos_cgrp)))))
  196. /*
  197. * There are two global mutexes guarding cpuset structures - cpuset_mutex
  198. * and callback_mutex. The latter may nest inside the former. We also
  199. * require taking task_lock() when dereferencing a task's cpuset pointer.
  200. * See "The task_lock() exception", at the end of this comment.
  201. *
  202. * A task must hold both mutexes to modify cpusets. If a task holds
  203. * cpuset_mutex, then it blocks others wanting that mutex, ensuring that it
  204. * is the only task able to also acquire callback_mutex and be able to
  205. * modify cpusets. It can perform various checks on the cpuset structure
  206. * first, knowing nothing will change. It can also allocate memory while
  207. * just holding cpuset_mutex. While it is performing these checks, various
  208. * callback routines can briefly acquire callback_mutex to query cpusets.
  209. * Once it is ready to make the changes, it takes callback_mutex, blocking
  210. * everyone else.
  211. *
  212. * Calls to the kernel memory allocator can not be made while holding
  213. * callback_mutex, as that would risk double tripping on callback_mutex
  214. * from one of the callbacks into the cpuset code from within
  215. * __alloc_pages().
  216. *
  217. * If a task is only holding callback_mutex, then it has read-only
  218. * access to cpusets.
  219. *
  220. * Now, the task_struct fields mems_allowed and mempolicy may be changed
  221. * by other task, we use alloc_lock in the task_struct fields to protect
  222. * them.
  223. *
  224. * The cpuset_common_file_read() handlers only hold callback_mutex across
  225. * small pieces of code, such as when reading out possibly multi-word
  226. * cpumasks and nodemasks.
  227. *
  228. * Accessing a task's cpuset should be done in accordance with the
  229. * guidelines for accessing subsystem state in kernel/cgroup.c
  230. */
  231. static DEFINE_MUTEX(cpuset_mutex);
  232. static DEFINE_MUTEX(callback_mutex);
  233. /*
  234. * CPU / memory hotplug is handled asynchronously.
  235. */
  236. static struct workqueue_struct *cpuset_propagate_hotplug_wq;
  237. static void cpuset_hotplug_workfn(struct work_struct *work);
  238. static void cpuset_propagate_hotplug_workfn(struct work_struct *work);
  239. static void schedule_cpuset_propagate_hotplug(struct cpuset *cs);
  240. static DECLARE_WORK(cpuset_hotplug_work, cpuset_hotplug_workfn);
  241. /*
  242. * This is ugly, but preserves the userspace API for existing cpuset
  243. * users. If someone tries to mount the "cpuset" filesystem, we
  244. * silently switch it to mount "cgroup" instead
  245. */
  246. static struct dentry *cpuset_mount(struct file_system_type *fs_type,
  247. int flags, const char *unused_dev_name, void *data)
  248. {
  249. struct file_system_type *cgroup_fs = get_fs_type("cgroup");
  250. struct dentry *ret = ERR_PTR(-ENODEV);
  251. if (cgroup_fs) {
  252. char mountopts[] =
  253. "cpuset,noprefix,"
  254. "release_agent=/sbin/cpuset_release_agent";
  255. ret = cgroup_fs->mount(cgroup_fs, flags,
  256. unused_dev_name, mountopts);
  257. put_filesystem(cgroup_fs);
  258. }
  259. return ret;
  260. }
  261. static struct file_system_type cpuset_fs_type = {
  262. .name = "cpuset",
  263. .mount = cpuset_mount,
  264. };
  265. /*
  266. * Return in pmask the portion of a cpusets's cpus_allowed that
  267. * are online. If none are online, walk up the cpuset hierarchy
  268. * until we find one that does have some online cpus. If we get
  269. * all the way to the top and still haven't found any online cpus,
  270. * return cpu_online_mask. Or if passed a NULL cs from an exit'ing
  271. * task, return cpu_online_mask.
  272. *
  273. * One way or another, we guarantee to return some non-empty subset
  274. * of cpu_online_mask.
  275. *
  276. * Call with callback_mutex held.
  277. */
  278. static void guarantee_online_cpus(const struct cpuset *cs,
  279. struct cpumask *pmask)
  280. {
  281. while (cs && !cpumask_intersects(cs->cpus_allowed, cpu_online_mask))
  282. cs = parent_cs(cs);
  283. if (cs)
  284. cpumask_and(pmask, cs->cpus_allowed, cpu_online_mask);
  285. else
  286. cpumask_copy(pmask, cpu_online_mask);
  287. BUG_ON(!cpumask_intersects(pmask, cpu_online_mask));
  288. }
  289. /*
  290. * Return in *pmask the portion of a cpusets's mems_allowed that
  291. * are online, with memory. If none are online with memory, walk
  292. * up the cpuset hierarchy until we find one that does have some
  293. * online mems. If we get all the way to the top and still haven't
  294. * found any online mems, return node_states[N_MEMORY].
  295. *
  296. * One way or another, we guarantee to return some non-empty subset
  297. * of node_states[N_MEMORY].
  298. *
  299. * Call with callback_mutex held.
  300. */
  301. static void guarantee_online_mems(const struct cpuset *cs, nodemask_t *pmask)
  302. {
  303. while (cs && !nodes_intersects(cs->mems_allowed,
  304. node_states[N_MEMORY]))
  305. cs = parent_cs(cs);
  306. if (cs)
  307. nodes_and(*pmask, cs->mems_allowed,
  308. node_states[N_MEMORY]);
  309. else
  310. *pmask = node_states[N_MEMORY];
  311. BUG_ON(!nodes_intersects(*pmask, node_states[N_MEMORY]));
  312. }
  313. /*
  314. * update task's spread flag if cpuset's page/slab spread flag is set
  315. *
  316. * Called with callback_mutex/cpuset_mutex held
  317. */
  318. static void cpuset_update_task_spread_flag(struct cpuset *cs,
  319. struct task_struct *tsk)
  320. {
  321. if (is_spread_page(cs))
  322. tsk->flags |= PF_SPREAD_PAGE;
  323. else
  324. tsk->flags &= ~PF_SPREAD_PAGE;
  325. if (is_spread_slab(cs))
  326. tsk->flags |= PF_SPREAD_SLAB;
  327. else
  328. tsk->flags &= ~PF_SPREAD_SLAB;
  329. }
  330. /*
  331. * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
  332. *
  333. * One cpuset is a subset of another if all its allowed CPUs and
  334. * Memory Nodes are a subset of the other, and its exclusive flags
  335. * are only set if the other's are set. Call holding cpuset_mutex.
  336. */
  337. static int is_cpuset_subset(const struct cpuset *p, const struct cpuset *q)
  338. {
  339. return cpumask_subset(p->cpus_allowed, q->cpus_allowed) &&
  340. nodes_subset(p->mems_allowed, q->mems_allowed) &&
  341. is_cpu_exclusive(p) <= is_cpu_exclusive(q) &&
  342. is_mem_exclusive(p) <= is_mem_exclusive(q);
  343. }
  344. /**
  345. * alloc_trial_cpuset - allocate a trial cpuset
  346. * @cs: the cpuset that the trial cpuset duplicates
  347. */
  348. static struct cpuset *alloc_trial_cpuset(const struct cpuset *cs)
  349. {
  350. struct cpuset *trial;
  351. trial = kmemdup(cs, sizeof(*cs), GFP_KERNEL);
  352. if (!trial)
  353. return NULL;
  354. if (!alloc_cpumask_var(&trial->cpus_allowed, GFP_KERNEL)) {
  355. kfree(trial);
  356. return NULL;
  357. }
  358. cpumask_copy(trial->cpus_allowed, cs->cpus_allowed);
  359. return trial;
  360. }
  361. /**
  362. * free_trial_cpuset - free the trial cpuset
  363. * @trial: the trial cpuset to be freed
  364. */
  365. static void free_trial_cpuset(struct cpuset *trial)
  366. {
  367. free_cpumask_var(trial->cpus_allowed);
  368. kfree(trial);
  369. }
  370. /*
  371. * validate_change() - Used to validate that any proposed cpuset change
  372. * follows the structural rules for cpusets.
  373. *
  374. * If we replaced the flag and mask values of the current cpuset
  375. * (cur) with those values in the trial cpuset (trial), would
  376. * our various subset and exclusive rules still be valid? Presumes
  377. * cpuset_mutex held.
  378. *
  379. * 'cur' is the address of an actual, in-use cpuset. Operations
  380. * such as list traversal that depend on the actual address of the
  381. * cpuset in the list must use cur below, not trial.
  382. *
  383. * 'trial' is the address of bulk structure copy of cur, with
  384. * perhaps one or more of the fields cpus_allowed, mems_allowed,
  385. * or flags changed to new, trial values.
  386. *
  387. * Return 0 if valid, -errno if not.
  388. */
  389. static int validate_change(const struct cpuset *cur, const struct cpuset *trial)
  390. {
  391. struct cgroup *cont;
  392. struct cpuset *c, *par;
  393. int ret;
  394. rcu_read_lock();
  395. /* Each of our child cpusets must be a subset of us */
  396. ret = -EBUSY;
  397. cpuset_for_each_child(c, cont, cur)
  398. if (!is_cpuset_subset(c, trial))
  399. goto out;
  400. /* Remaining checks don't apply to root cpuset */
  401. ret = 0;
  402. if (cur == &top_cpuset)
  403. goto out;
  404. par = parent_cs(cur);
  405. /* We must be a subset of our parent cpuset */
  406. ret = -EACCES;
  407. if (!is_cpuset_subset(trial, par))
  408. goto out;
  409. /*
  410. * If either I or some sibling (!= me) is exclusive, we can't
  411. * overlap
  412. */
  413. ret = -EINVAL;
  414. cpuset_for_each_child(c, cont, par) {
  415. if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
  416. c != cur &&
  417. cpumask_intersects(trial->cpus_allowed, c->cpus_allowed))
  418. goto out;
  419. if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) &&
  420. c != cur &&
  421. nodes_intersects(trial->mems_allowed, c->mems_allowed))
  422. goto out;
  423. }
  424. /*
  425. * Cpusets with tasks - existing or newly being attached - can't
  426. * have empty cpus_allowed or mems_allowed.
  427. */
  428. ret = -ENOSPC;
  429. if ((cgroup_task_count(cur->css.cgroup) || cur->attach_in_progress) &&
  430. (cpumask_empty(trial->cpus_allowed) ||
  431. nodes_empty(trial->mems_allowed)))
  432. goto out;
  433. ret = 0;
  434. out:
  435. rcu_read_unlock();
  436. return ret;
  437. }
  438. #ifdef CONFIG_SMP
  439. /*
  440. * Helper routine for generate_sched_domains().
  441. * Do cpusets a, b have overlapping cpus_allowed masks?
  442. */
  443. static int cpusets_overlap(struct cpuset *a, struct cpuset *b)
  444. {
  445. return cpumask_intersects(a->cpus_allowed, b->cpus_allowed);
  446. }
  447. static void
  448. update_domain_attr(struct sched_domain_attr *dattr, struct cpuset *c)
  449. {
  450. if (dattr->relax_domain_level < c->relax_domain_level)
  451. dattr->relax_domain_level = c->relax_domain_level;
  452. return;
  453. }
  454. static void update_domain_attr_tree(struct sched_domain_attr *dattr,
  455. struct cpuset *root_cs)
  456. {
  457. struct cpuset *cp;
  458. struct cgroup *pos_cgrp;
  459. rcu_read_lock();
  460. cpuset_for_each_descendant_pre(cp, pos_cgrp, root_cs) {
  461. /* skip the whole subtree if @cp doesn't have any CPU */
  462. if (cpumask_empty(cp->cpus_allowed)) {
  463. pos_cgrp = cgroup_rightmost_descendant(pos_cgrp);
  464. continue;
  465. }
  466. if (is_sched_load_balance(cp))
  467. update_domain_attr(dattr, cp);
  468. }
  469. rcu_read_unlock();
  470. }
  471. /*
  472. * generate_sched_domains()
  473. *
  474. * This function builds a partial partition of the systems CPUs
  475. * A 'partial partition' is a set of non-overlapping subsets whose
  476. * union is a subset of that set.
  477. * The output of this function needs to be passed to kernel/sched.c
  478. * partition_sched_domains() routine, which will rebuild the scheduler's
  479. * load balancing domains (sched domains) as specified by that partial
  480. * partition.
  481. *
  482. * See "What is sched_load_balance" in Documentation/cgroups/cpusets.txt
  483. * for a background explanation of this.
  484. *
  485. * Does not return errors, on the theory that the callers of this
  486. * routine would rather not worry about failures to rebuild sched
  487. * domains when operating in the severe memory shortage situations
  488. * that could cause allocation failures below.
  489. *
  490. * Must be called with cpuset_mutex held.
  491. *
  492. * The three key local variables below are:
  493. * q - a linked-list queue of cpuset pointers, used to implement a
  494. * top-down scan of all cpusets. This scan loads a pointer
  495. * to each cpuset marked is_sched_load_balance into the
  496. * array 'csa'. For our purposes, rebuilding the schedulers
  497. * sched domains, we can ignore !is_sched_load_balance cpusets.
  498. * csa - (for CpuSet Array) Array of pointers to all the cpusets
  499. * that need to be load balanced, for convenient iterative
  500. * access by the subsequent code that finds the best partition,
  501. * i.e the set of domains (subsets) of CPUs such that the
  502. * cpus_allowed of every cpuset marked is_sched_load_balance
  503. * is a subset of one of these domains, while there are as
  504. * many such domains as possible, each as small as possible.
  505. * doms - Conversion of 'csa' to an array of cpumasks, for passing to
  506. * the kernel/sched.c routine partition_sched_domains() in a
  507. * convenient format, that can be easily compared to the prior
  508. * value to determine what partition elements (sched domains)
  509. * were changed (added or removed.)
  510. *
  511. * Finding the best partition (set of domains):
  512. * The triple nested loops below over i, j, k scan over the
  513. * load balanced cpusets (using the array of cpuset pointers in
  514. * csa[]) looking for pairs of cpusets that have overlapping
  515. * cpus_allowed, but which don't have the same 'pn' partition
  516. * number and gives them in the same partition number. It keeps
  517. * looping on the 'restart' label until it can no longer find
  518. * any such pairs.
  519. *
  520. * The union of the cpus_allowed masks from the set of
  521. * all cpusets having the same 'pn' value then form the one
  522. * element of the partition (one sched domain) to be passed to
  523. * partition_sched_domains().
  524. */
  525. static int generate_sched_domains(cpumask_var_t **domains,
  526. struct sched_domain_attr **attributes)
  527. {
  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. struct cgroup *pos_cgrp;
  537. doms = NULL;
  538. dattr = NULL;
  539. csa = NULL;
  540. /* Special case for the 99% of systems with one, full, sched domain */
  541. if (is_sched_load_balance(&top_cpuset)) {
  542. ndoms = 1;
  543. doms = alloc_sched_domains(ndoms);
  544. if (!doms)
  545. goto done;
  546. dattr = kmalloc(sizeof(struct sched_domain_attr), GFP_KERNEL);
  547. if (dattr) {
  548. *dattr = SD_ATTR_INIT;
  549. update_domain_attr_tree(dattr, &top_cpuset);
  550. }
  551. cpumask_copy(doms[0], top_cpuset.cpus_allowed);
  552. goto done;
  553. }
  554. csa = kmalloc(number_of_cpusets * sizeof(cp), GFP_KERNEL);
  555. if (!csa)
  556. goto done;
  557. csn = 0;
  558. rcu_read_lock();
  559. cpuset_for_each_descendant_pre(cp, pos_cgrp, &top_cpuset) {
  560. /*
  561. * Continue traversing beyond @cp iff @cp has some CPUs and
  562. * isn't load balancing. The former is obvious. The
  563. * latter: All child cpusets contain a subset of the
  564. * parent's cpus, so just skip them, and then we call
  565. * update_domain_attr_tree() to calc relax_domain_level of
  566. * the corresponding sched domain.
  567. */
  568. if (!cpumask_empty(cp->cpus_allowed) &&
  569. !is_sched_load_balance(cp))
  570. continue;
  571. if (is_sched_load_balance(cp))
  572. csa[csn++] = cp;
  573. /* skip @cp's subtree */
  574. pos_cgrp = cgroup_rightmost_descendant(pos_cgrp);
  575. }
  576. rcu_read_unlock();
  577. for (i = 0; i < csn; i++)
  578. csa[i]->pn = i;
  579. ndoms = csn;
  580. restart:
  581. /* Find the best partition (set of sched domains) */
  582. for (i = 0; i < csn; i++) {
  583. struct cpuset *a = csa[i];
  584. int apn = a->pn;
  585. for (j = 0; j < csn; j++) {
  586. struct cpuset *b = csa[j];
  587. int bpn = b->pn;
  588. if (apn != bpn && cpusets_overlap(a, b)) {
  589. for (k = 0; k < csn; k++) {
  590. struct cpuset *c = csa[k];
  591. if (c->pn == bpn)
  592. c->pn = apn;
  593. }
  594. ndoms--; /* one less element */
  595. goto restart;
  596. }
  597. }
  598. }
  599. /*
  600. * Now we know how many domains to create.
  601. * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
  602. */
  603. doms = alloc_sched_domains(ndoms);
  604. if (!doms)
  605. goto done;
  606. /*
  607. * The rest of the code, including the scheduler, can deal with
  608. * dattr==NULL case. No need to abort if alloc fails.
  609. */
  610. dattr = kmalloc(ndoms * sizeof(struct sched_domain_attr), GFP_KERNEL);
  611. for (nslot = 0, i = 0; i < csn; i++) {
  612. struct cpuset *a = csa[i];
  613. struct cpumask *dp;
  614. int apn = a->pn;
  615. if (apn < 0) {
  616. /* Skip completed partitions */
  617. continue;
  618. }
  619. dp = doms[nslot];
  620. if (nslot == ndoms) {
  621. static int warnings = 10;
  622. if (warnings) {
  623. printk(KERN_WARNING
  624. "rebuild_sched_domains confused:"
  625. " nslot %d, ndoms %d, csn %d, i %d,"
  626. " apn %d\n",
  627. nslot, ndoms, csn, i, apn);
  628. warnings--;
  629. }
  630. continue;
  631. }
  632. cpumask_clear(dp);
  633. if (dattr)
  634. *(dattr + nslot) = SD_ATTR_INIT;
  635. for (j = i; j < csn; j++) {
  636. struct cpuset *b = csa[j];
  637. if (apn == b->pn) {
  638. cpumask_or(dp, dp, b->cpus_allowed);
  639. if (dattr)
  640. update_domain_attr_tree(dattr + nslot, b);
  641. /* Done with this partition */
  642. b->pn = -1;
  643. }
  644. }
  645. nslot++;
  646. }
  647. BUG_ON(nslot != ndoms);
  648. done:
  649. kfree(csa);
  650. /*
  651. * Fallback to the default domain if kmalloc() failed.
  652. * See comments in partition_sched_domains().
  653. */
  654. if (doms == NULL)
  655. ndoms = 1;
  656. *domains = doms;
  657. *attributes = dattr;
  658. return ndoms;
  659. }
  660. /*
  661. * Rebuild scheduler domains.
  662. *
  663. * If the flag 'sched_load_balance' of any cpuset with non-empty
  664. * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
  665. * which has that flag enabled, or if any cpuset with a non-empty
  666. * 'cpus' is removed, then call this routine to rebuild the
  667. * scheduler's dynamic sched domains.
  668. *
  669. * Call with cpuset_mutex held. Takes get_online_cpus().
  670. */
  671. static void rebuild_sched_domains_locked(void)
  672. {
  673. struct sched_domain_attr *attr;
  674. cpumask_var_t *doms;
  675. int ndoms;
  676. lockdep_assert_held(&cpuset_mutex);
  677. get_online_cpus();
  678. /* Generate domain masks and attrs */
  679. ndoms = generate_sched_domains(&doms, &attr);
  680. /* Have scheduler rebuild the domains */
  681. partition_sched_domains(ndoms, doms, attr);
  682. put_online_cpus();
  683. }
  684. #else /* !CONFIG_SMP */
  685. static void rebuild_sched_domains_locked(void)
  686. {
  687. }
  688. static int generate_sched_domains(cpumask_var_t **domains,
  689. struct sched_domain_attr **attributes)
  690. {
  691. *domains = NULL;
  692. return 1;
  693. }
  694. #endif /* CONFIG_SMP */
  695. void rebuild_sched_domains(void)
  696. {
  697. mutex_lock(&cpuset_mutex);
  698. rebuild_sched_domains_locked();
  699. mutex_unlock(&cpuset_mutex);
  700. }
  701. /**
  702. * cpuset_test_cpumask - test a task's cpus_allowed versus its cpuset's
  703. * @tsk: task to test
  704. * @scan: struct cgroup_scanner contained in its struct cpuset_hotplug_scanner
  705. *
  706. * Call with cpuset_mutex held. May take callback_mutex during call.
  707. * Called for each task in a cgroup by cgroup_scan_tasks().
  708. * Return nonzero if this tasks's cpus_allowed mask should be changed (in other
  709. * words, if its mask is not equal to its cpuset's mask).
  710. */
  711. static int cpuset_test_cpumask(struct task_struct *tsk,
  712. struct cgroup_scanner *scan)
  713. {
  714. return !cpumask_equal(&tsk->cpus_allowed,
  715. (cgroup_cs(scan->cg))->cpus_allowed);
  716. }
  717. /**
  718. * cpuset_change_cpumask - make a task's cpus_allowed the same as its cpuset's
  719. * @tsk: task to test
  720. * @scan: struct cgroup_scanner containing the cgroup of the task
  721. *
  722. * Called by cgroup_scan_tasks() for each task in a cgroup whose
  723. * cpus_allowed mask needs to be changed.
  724. *
  725. * We don't need to re-check for the cgroup/cpuset membership, since we're
  726. * holding cpuset_mutex at this point.
  727. */
  728. static void cpuset_change_cpumask(struct task_struct *tsk,
  729. struct cgroup_scanner *scan)
  730. {
  731. set_cpus_allowed_ptr(tsk, ((cgroup_cs(scan->cg))->cpus_allowed));
  732. }
  733. /**
  734. * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
  735. * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
  736. * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
  737. *
  738. * Called with cpuset_mutex held
  739. *
  740. * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
  741. * calling callback functions for each.
  742. *
  743. * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
  744. * if @heap != NULL.
  745. */
  746. static void update_tasks_cpumask(struct cpuset *cs, struct ptr_heap *heap)
  747. {
  748. struct cgroup_scanner scan;
  749. scan.cg = cs->css.cgroup;
  750. scan.test_task = cpuset_test_cpumask;
  751. scan.process_task = cpuset_change_cpumask;
  752. scan.heap = heap;
  753. cgroup_scan_tasks(&scan);
  754. }
  755. /**
  756. * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
  757. * @cs: the cpuset to consider
  758. * @buf: buffer of cpu numbers written to this cpuset
  759. */
  760. static int update_cpumask(struct cpuset *cs, struct cpuset *trialcs,
  761. const char *buf)
  762. {
  763. struct ptr_heap heap;
  764. int retval;
  765. int is_load_balanced;
  766. /* top_cpuset.cpus_allowed tracks cpu_online_mask; it's read-only */
  767. if (cs == &top_cpuset)
  768. return -EACCES;
  769. /*
  770. * An empty cpus_allowed is ok only if the cpuset has no tasks.
  771. * Since cpulist_parse() fails on an empty mask, we special case
  772. * that parsing. The validate_change() call ensures that cpusets
  773. * with tasks have cpus.
  774. */
  775. if (!*buf) {
  776. cpumask_clear(trialcs->cpus_allowed);
  777. } else {
  778. retval = cpulist_parse(buf, trialcs->cpus_allowed);
  779. if (retval < 0)
  780. return retval;
  781. if (!cpumask_subset(trialcs->cpus_allowed, cpu_active_mask))
  782. return -EINVAL;
  783. }
  784. retval = validate_change(cs, trialcs);
  785. if (retval < 0)
  786. return retval;
  787. /* Nothing to do if the cpus didn't change */
  788. if (cpumask_equal(cs->cpus_allowed, trialcs->cpus_allowed))
  789. return 0;
  790. retval = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
  791. if (retval)
  792. return retval;
  793. is_load_balanced = is_sched_load_balance(trialcs);
  794. mutex_lock(&callback_mutex);
  795. cpumask_copy(cs->cpus_allowed, trialcs->cpus_allowed);
  796. mutex_unlock(&callback_mutex);
  797. /*
  798. * Scan tasks in the cpuset, and update the cpumasks of any
  799. * that need an update.
  800. */
  801. update_tasks_cpumask(cs, &heap);
  802. heap_free(&heap);
  803. if (is_load_balanced)
  804. rebuild_sched_domains_locked();
  805. return 0;
  806. }
  807. /*
  808. * cpuset_migrate_mm
  809. *
  810. * Migrate memory region from one set of nodes to another.
  811. *
  812. * Temporarilly set tasks mems_allowed to target nodes of migration,
  813. * so that the migration code can allocate pages on these nodes.
  814. *
  815. * Call holding cpuset_mutex, so current's cpuset won't change
  816. * during this call, as manage_mutex holds off any cpuset_attach()
  817. * calls. Therefore we don't need to take task_lock around the
  818. * call to guarantee_online_mems(), as we know no one is changing
  819. * our task's cpuset.
  820. *
  821. * While the mm_struct we are migrating is typically from some
  822. * other task, the task_struct mems_allowed that we are hacking
  823. * is for our current task, which must allocate new pages for that
  824. * migrating memory region.
  825. */
  826. static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
  827. const nodemask_t *to)
  828. {
  829. struct task_struct *tsk = current;
  830. tsk->mems_allowed = *to;
  831. do_migrate_pages(mm, from, to, MPOL_MF_MOVE_ALL);
  832. guarantee_online_mems(task_cs(tsk),&tsk->mems_allowed);
  833. }
  834. /*
  835. * cpuset_change_task_nodemask - change task's mems_allowed and mempolicy
  836. * @tsk: the task to change
  837. * @newmems: new nodes that the task will be set
  838. *
  839. * In order to avoid seeing no nodes if the old and new nodes are disjoint,
  840. * we structure updates as setting all new allowed nodes, then clearing newly
  841. * disallowed ones.
  842. */
  843. static void cpuset_change_task_nodemask(struct task_struct *tsk,
  844. nodemask_t *newmems)
  845. {
  846. bool need_loop;
  847. /*
  848. * Allow tasks that have access to memory reserves because they have
  849. * been OOM killed to get memory anywhere.
  850. */
  851. if (unlikely(test_thread_flag(TIF_MEMDIE)))
  852. return;
  853. if (current->flags & PF_EXITING) /* Let dying task have memory */
  854. return;
  855. task_lock(tsk);
  856. /*
  857. * Determine if a loop is necessary if another thread is doing
  858. * get_mems_allowed(). If at least one node remains unchanged and
  859. * tsk does not have a mempolicy, then an empty nodemask will not be
  860. * possible when mems_allowed is larger than a word.
  861. */
  862. need_loop = task_has_mempolicy(tsk) ||
  863. !nodes_intersects(*newmems, tsk->mems_allowed);
  864. if (need_loop)
  865. write_seqcount_begin(&tsk->mems_allowed_seq);
  866. nodes_or(tsk->mems_allowed, tsk->mems_allowed, *newmems);
  867. mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP1);
  868. mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP2);
  869. tsk->mems_allowed = *newmems;
  870. if (need_loop)
  871. write_seqcount_end(&tsk->mems_allowed_seq);
  872. task_unlock(tsk);
  873. }
  874. /*
  875. * Update task's mems_allowed and rebind its mempolicy and vmas' mempolicy
  876. * of it to cpuset's new mems_allowed, and migrate pages to new nodes if
  877. * memory_migrate flag is set. Called with cpuset_mutex held.
  878. */
  879. static void cpuset_change_nodemask(struct task_struct *p,
  880. struct cgroup_scanner *scan)
  881. {
  882. struct mm_struct *mm;
  883. struct cpuset *cs;
  884. int migrate;
  885. const nodemask_t *oldmem = scan->data;
  886. static nodemask_t newmems; /* protected by cpuset_mutex */
  887. cs = cgroup_cs(scan->cg);
  888. guarantee_online_mems(cs, &newmems);
  889. cpuset_change_task_nodemask(p, &newmems);
  890. mm = get_task_mm(p);
  891. if (!mm)
  892. return;
  893. migrate = is_memory_migrate(cs);
  894. mpol_rebind_mm(mm, &cs->mems_allowed);
  895. if (migrate)
  896. cpuset_migrate_mm(mm, oldmem, &cs->mems_allowed);
  897. mmput(mm);
  898. }
  899. static void *cpuset_being_rebound;
  900. /**
  901. * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
  902. * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
  903. * @oldmem: old mems_allowed of cpuset cs
  904. * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
  905. *
  906. * Called with cpuset_mutex held
  907. * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
  908. * if @heap != NULL.
  909. */
  910. static void update_tasks_nodemask(struct cpuset *cs, const nodemask_t *oldmem,
  911. struct ptr_heap *heap)
  912. {
  913. struct cgroup_scanner scan;
  914. cpuset_being_rebound = cs; /* causes mpol_dup() rebind */
  915. scan.cg = cs->css.cgroup;
  916. scan.test_task = NULL;
  917. scan.process_task = cpuset_change_nodemask;
  918. scan.heap = heap;
  919. scan.data = (nodemask_t *)oldmem;
  920. /*
  921. * The mpol_rebind_mm() call takes mmap_sem, which we couldn't
  922. * take while holding tasklist_lock. Forks can happen - the
  923. * mpol_dup() cpuset_being_rebound check will catch such forks,
  924. * and rebind their vma mempolicies too. Because we still hold
  925. * the global cpuset_mutex, we know that no other rebind effort
  926. * will be contending for the global variable cpuset_being_rebound.
  927. * It's ok if we rebind the same mm twice; mpol_rebind_mm()
  928. * is idempotent. Also migrate pages in each mm to new nodes.
  929. */
  930. cgroup_scan_tasks(&scan);
  931. /* We're done rebinding vmas to this cpuset's new mems_allowed. */
  932. cpuset_being_rebound = NULL;
  933. }
  934. /*
  935. * Handle user request to change the 'mems' memory placement
  936. * of a cpuset. Needs to validate the request, update the
  937. * cpusets mems_allowed, and for each task in the cpuset,
  938. * update mems_allowed and rebind task's mempolicy and any vma
  939. * mempolicies and if the cpuset is marked 'memory_migrate',
  940. * migrate the tasks pages to the new memory.
  941. *
  942. * Call with cpuset_mutex held. May take callback_mutex during call.
  943. * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
  944. * lock each such tasks mm->mmap_sem, scan its vma's and rebind
  945. * their mempolicies to the cpusets new mems_allowed.
  946. */
  947. static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
  948. const char *buf)
  949. {
  950. NODEMASK_ALLOC(nodemask_t, oldmem, GFP_KERNEL);
  951. int retval;
  952. struct ptr_heap heap;
  953. if (!oldmem)
  954. return -ENOMEM;
  955. /*
  956. * top_cpuset.mems_allowed tracks node_stats[N_MEMORY];
  957. * it's read-only
  958. */
  959. if (cs == &top_cpuset) {
  960. retval = -EACCES;
  961. goto done;
  962. }
  963. /*
  964. * An empty mems_allowed is ok iff there are no tasks in the cpuset.
  965. * Since nodelist_parse() fails on an empty mask, we special case
  966. * that parsing. The validate_change() call ensures that cpusets
  967. * with tasks have memory.
  968. */
  969. if (!*buf) {
  970. nodes_clear(trialcs->mems_allowed);
  971. } else {
  972. retval = nodelist_parse(buf, trialcs->mems_allowed);
  973. if (retval < 0)
  974. goto done;
  975. if (!nodes_subset(trialcs->mems_allowed,
  976. node_states[N_MEMORY])) {
  977. retval = -EINVAL;
  978. goto done;
  979. }
  980. }
  981. *oldmem = cs->mems_allowed;
  982. if (nodes_equal(*oldmem, trialcs->mems_allowed)) {
  983. retval = 0; /* Too easy - nothing to do */
  984. goto done;
  985. }
  986. retval = validate_change(cs, trialcs);
  987. if (retval < 0)
  988. goto done;
  989. retval = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
  990. if (retval < 0)
  991. goto done;
  992. mutex_lock(&callback_mutex);
  993. cs->mems_allowed = trialcs->mems_allowed;
  994. mutex_unlock(&callback_mutex);
  995. update_tasks_nodemask(cs, oldmem, &heap);
  996. heap_free(&heap);
  997. done:
  998. NODEMASK_FREE(oldmem);
  999. return retval;
  1000. }
  1001. int current_cpuset_is_being_rebound(void)
  1002. {
  1003. return task_cs(current) == cpuset_being_rebound;
  1004. }
  1005. static int update_relax_domain_level(struct cpuset *cs, s64 val)
  1006. {
  1007. #ifdef CONFIG_SMP
  1008. if (val < -1 || val >= sched_domain_level_max)
  1009. return -EINVAL;
  1010. #endif
  1011. if (val != cs->relax_domain_level) {
  1012. cs->relax_domain_level = val;
  1013. if (!cpumask_empty(cs->cpus_allowed) &&
  1014. is_sched_load_balance(cs))
  1015. rebuild_sched_domains_locked();
  1016. }
  1017. return 0;
  1018. }
  1019. /*
  1020. * cpuset_change_flag - make a task's spread flags the same as its cpuset's
  1021. * @tsk: task to be updated
  1022. * @scan: struct cgroup_scanner containing the cgroup of the task
  1023. *
  1024. * Called by cgroup_scan_tasks() for each task in a cgroup.
  1025. *
  1026. * We don't need to re-check for the cgroup/cpuset membership, since we're
  1027. * holding cpuset_mutex at this point.
  1028. */
  1029. static void cpuset_change_flag(struct task_struct *tsk,
  1030. struct cgroup_scanner *scan)
  1031. {
  1032. cpuset_update_task_spread_flag(cgroup_cs(scan->cg), tsk);
  1033. }
  1034. /*
  1035. * update_tasks_flags - update the spread flags of tasks in the cpuset.
  1036. * @cs: the cpuset in which each task's spread flags needs to be changed
  1037. * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
  1038. *
  1039. * Called with cpuset_mutex held
  1040. *
  1041. * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
  1042. * calling callback functions for each.
  1043. *
  1044. * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
  1045. * if @heap != NULL.
  1046. */
  1047. static void update_tasks_flags(struct cpuset *cs, struct ptr_heap *heap)
  1048. {
  1049. struct cgroup_scanner scan;
  1050. scan.cg = cs->css.cgroup;
  1051. scan.test_task = NULL;
  1052. scan.process_task = cpuset_change_flag;
  1053. scan.heap = heap;
  1054. cgroup_scan_tasks(&scan);
  1055. }
  1056. /*
  1057. * update_flag - read a 0 or a 1 in a file and update associated flag
  1058. * bit: the bit to update (see cpuset_flagbits_t)
  1059. * cs: the cpuset to update
  1060. * turning_on: whether the flag is being set or cleared
  1061. *
  1062. * Call with cpuset_mutex held.
  1063. */
  1064. static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs,
  1065. int turning_on)
  1066. {
  1067. struct cpuset *trialcs;
  1068. int balance_flag_changed;
  1069. int spread_flag_changed;
  1070. struct ptr_heap heap;
  1071. int err;
  1072. trialcs = alloc_trial_cpuset(cs);
  1073. if (!trialcs)
  1074. return -ENOMEM;
  1075. if (turning_on)
  1076. set_bit(bit, &trialcs->flags);
  1077. else
  1078. clear_bit(bit, &trialcs->flags);
  1079. err = validate_change(cs, trialcs);
  1080. if (err < 0)
  1081. goto out;
  1082. err = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
  1083. if (err < 0)
  1084. goto out;
  1085. balance_flag_changed = (is_sched_load_balance(cs) !=
  1086. is_sched_load_balance(trialcs));
  1087. spread_flag_changed = ((is_spread_slab(cs) != is_spread_slab(trialcs))
  1088. || (is_spread_page(cs) != is_spread_page(trialcs)));
  1089. mutex_lock(&callback_mutex);
  1090. cs->flags = trialcs->flags;
  1091. mutex_unlock(&callback_mutex);
  1092. if (!cpumask_empty(trialcs->cpus_allowed) && balance_flag_changed)
  1093. rebuild_sched_domains_locked();
  1094. if (spread_flag_changed)
  1095. update_tasks_flags(cs, &heap);
  1096. heap_free(&heap);
  1097. out:
  1098. free_trial_cpuset(trialcs);
  1099. return err;
  1100. }
  1101. /*
  1102. * Frequency meter - How fast is some event occurring?
  1103. *
  1104. * These routines manage a digitally filtered, constant time based,
  1105. * event frequency meter. There are four routines:
  1106. * fmeter_init() - initialize a frequency meter.
  1107. * fmeter_markevent() - called each time the event happens.
  1108. * fmeter_getrate() - returns the recent rate of such events.
  1109. * fmeter_update() - internal routine used to update fmeter.
  1110. *
  1111. * A common data structure is passed to each of these routines,
  1112. * which is used to keep track of the state required to manage the
  1113. * frequency meter and its digital filter.
  1114. *
  1115. * The filter works on the number of events marked per unit time.
  1116. * The filter is single-pole low-pass recursive (IIR). The time unit
  1117. * is 1 second. Arithmetic is done using 32-bit integers scaled to
  1118. * simulate 3 decimal digits of precision (multiplied by 1000).
  1119. *
  1120. * With an FM_COEF of 933, and a time base of 1 second, the filter
  1121. * has a half-life of 10 seconds, meaning that if the events quit
  1122. * happening, then the rate returned from the fmeter_getrate()
  1123. * will be cut in half each 10 seconds, until it converges to zero.
  1124. *
  1125. * It is not worth doing a real infinitely recursive filter. If more
  1126. * than FM_MAXTICKS ticks have elapsed since the last filter event,
  1127. * just compute FM_MAXTICKS ticks worth, by which point the level
  1128. * will be stable.
  1129. *
  1130. * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
  1131. * arithmetic overflow in the fmeter_update() routine.
  1132. *
  1133. * Given the simple 32 bit integer arithmetic used, this meter works
  1134. * best for reporting rates between one per millisecond (msec) and
  1135. * one per 32 (approx) seconds. At constant rates faster than one
  1136. * per msec it maxes out at values just under 1,000,000. At constant
  1137. * rates between one per msec, and one per second it will stabilize
  1138. * to a value N*1000, where N is the rate of events per second.
  1139. * At constant rates between one per second and one per 32 seconds,
  1140. * it will be choppy, moving up on the seconds that have an event,
  1141. * and then decaying until the next event. At rates slower than
  1142. * about one in 32 seconds, it decays all the way back to zero between
  1143. * each event.
  1144. */
  1145. #define FM_COEF 933 /* coefficient for half-life of 10 secs */
  1146. #define FM_MAXTICKS ((time_t)99) /* useless computing more ticks than this */
  1147. #define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
  1148. #define FM_SCALE 1000 /* faux fixed point scale */
  1149. /* Initialize a frequency meter */
  1150. static void fmeter_init(struct fmeter *fmp)
  1151. {
  1152. fmp->cnt = 0;
  1153. fmp->val = 0;
  1154. fmp->time = 0;
  1155. spin_lock_init(&fmp->lock);
  1156. }
  1157. /* Internal meter update - process cnt events and update value */
  1158. static void fmeter_update(struct fmeter *fmp)
  1159. {
  1160. time_t now = get_seconds();
  1161. time_t ticks = now - fmp->time;
  1162. if (ticks == 0)
  1163. return;
  1164. ticks = min(FM_MAXTICKS, ticks);
  1165. while (ticks-- > 0)
  1166. fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
  1167. fmp->time = now;
  1168. fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
  1169. fmp->cnt = 0;
  1170. }
  1171. /* Process any previous ticks, then bump cnt by one (times scale). */
  1172. static void fmeter_markevent(struct fmeter *fmp)
  1173. {
  1174. spin_lock(&fmp->lock);
  1175. fmeter_update(fmp);
  1176. fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
  1177. spin_unlock(&fmp->lock);
  1178. }
  1179. /* Process any previous ticks, then return current value. */
  1180. static int fmeter_getrate(struct fmeter *fmp)
  1181. {
  1182. int val;
  1183. spin_lock(&fmp->lock);
  1184. fmeter_update(fmp);
  1185. val = fmp->val;
  1186. spin_unlock(&fmp->lock);
  1187. return val;
  1188. }
  1189. /* Called by cgroups to determine if a cpuset is usable; cpuset_mutex held */
  1190. static int cpuset_can_attach(struct cgroup *cgrp, struct cgroup_taskset *tset)
  1191. {
  1192. struct cpuset *cs = cgroup_cs(cgrp);
  1193. struct task_struct *task;
  1194. int ret;
  1195. mutex_lock(&cpuset_mutex);
  1196. ret = -ENOSPC;
  1197. if (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed))
  1198. goto out_unlock;
  1199. cgroup_taskset_for_each(task, cgrp, tset) {
  1200. /*
  1201. * Kthreads bound to specific cpus cannot be moved to a new
  1202. * cpuset; we cannot change their cpu affinity and
  1203. * isolating such threads by their set of allowed nodes is
  1204. * unnecessary. Thus, cpusets are not applicable for such
  1205. * threads. This prevents checking for success of
  1206. * set_cpus_allowed_ptr() on all attached tasks before
  1207. * cpus_allowed may be changed.
  1208. */
  1209. ret = -EINVAL;
  1210. if (task->flags & PF_THREAD_BOUND)
  1211. goto out_unlock;
  1212. ret = security_task_setscheduler(task);
  1213. if (ret)
  1214. goto out_unlock;
  1215. }
  1216. /*
  1217. * Mark attach is in progress. This makes validate_change() fail
  1218. * changes which zero cpus/mems_allowed.
  1219. */
  1220. cs->attach_in_progress++;
  1221. ret = 0;
  1222. out_unlock:
  1223. mutex_unlock(&cpuset_mutex);
  1224. return ret;
  1225. }
  1226. static void cpuset_cancel_attach(struct cgroup *cgrp,
  1227. struct cgroup_taskset *tset)
  1228. {
  1229. mutex_lock(&cpuset_mutex);
  1230. cgroup_cs(cgrp)->attach_in_progress--;
  1231. mutex_unlock(&cpuset_mutex);
  1232. }
  1233. /*
  1234. * Protected by cpuset_mutex. cpus_attach is used only by cpuset_attach()
  1235. * but we can't allocate it dynamically there. Define it global and
  1236. * allocate from cpuset_init().
  1237. */
  1238. static cpumask_var_t cpus_attach;
  1239. static void cpuset_attach(struct cgroup *cgrp, struct cgroup_taskset *tset)
  1240. {
  1241. /* static bufs protected by cpuset_mutex */
  1242. static nodemask_t cpuset_attach_nodemask_from;
  1243. static nodemask_t cpuset_attach_nodemask_to;
  1244. struct mm_struct *mm;
  1245. struct task_struct *task;
  1246. struct task_struct *leader = cgroup_taskset_first(tset);
  1247. struct cgroup *oldcgrp = cgroup_taskset_cur_cgroup(tset);
  1248. struct cpuset *cs = cgroup_cs(cgrp);
  1249. struct cpuset *oldcs = cgroup_cs(oldcgrp);
  1250. mutex_lock(&cpuset_mutex);
  1251. /* prepare for attach */
  1252. if (cs == &top_cpuset)
  1253. cpumask_copy(cpus_attach, cpu_possible_mask);
  1254. else
  1255. guarantee_online_cpus(cs, cpus_attach);
  1256. guarantee_online_mems(cs, &cpuset_attach_nodemask_to);
  1257. cgroup_taskset_for_each(task, cgrp, tset) {
  1258. /*
  1259. * can_attach beforehand should guarantee that this doesn't
  1260. * fail. TODO: have a better way to handle failure here
  1261. */
  1262. WARN_ON_ONCE(set_cpus_allowed_ptr(task, cpus_attach));
  1263. cpuset_change_task_nodemask(task, &cpuset_attach_nodemask_to);
  1264. cpuset_update_task_spread_flag(cs, task);
  1265. }
  1266. /*
  1267. * Change mm, possibly for multiple threads in a threadgroup. This is
  1268. * expensive and may sleep.
  1269. */
  1270. cpuset_attach_nodemask_from = oldcs->mems_allowed;
  1271. cpuset_attach_nodemask_to = cs->mems_allowed;
  1272. mm = get_task_mm(leader);
  1273. if (mm) {
  1274. mpol_rebind_mm(mm, &cpuset_attach_nodemask_to);
  1275. if (is_memory_migrate(cs))
  1276. cpuset_migrate_mm(mm, &cpuset_attach_nodemask_from,
  1277. &cpuset_attach_nodemask_to);
  1278. mmput(mm);
  1279. }
  1280. cs->attach_in_progress--;
  1281. /*
  1282. * We may have raced with CPU/memory hotunplug. Trigger hotplug
  1283. * propagation if @cs doesn't have any CPU or memory. It will move
  1284. * the newly added tasks to the nearest parent which can execute.
  1285. */
  1286. if (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed))
  1287. schedule_cpuset_propagate_hotplug(cs);
  1288. mutex_unlock(&cpuset_mutex);
  1289. }
  1290. /* The various types of files and directories in a cpuset file system */
  1291. typedef enum {
  1292. FILE_MEMORY_MIGRATE,
  1293. FILE_CPULIST,
  1294. FILE_MEMLIST,
  1295. FILE_CPU_EXCLUSIVE,
  1296. FILE_MEM_EXCLUSIVE,
  1297. FILE_MEM_HARDWALL,
  1298. FILE_SCHED_LOAD_BALANCE,
  1299. FILE_SCHED_RELAX_DOMAIN_LEVEL,
  1300. FILE_MEMORY_PRESSURE_ENABLED,
  1301. FILE_MEMORY_PRESSURE,
  1302. FILE_SPREAD_PAGE,
  1303. FILE_SPREAD_SLAB,
  1304. } cpuset_filetype_t;
  1305. static int cpuset_write_u64(struct cgroup *cgrp, struct cftype *cft, u64 val)
  1306. {
  1307. struct cpuset *cs = cgroup_cs(cgrp);
  1308. cpuset_filetype_t type = cft->private;
  1309. int retval = -ENODEV;
  1310. mutex_lock(&cpuset_mutex);
  1311. if (!is_cpuset_online(cs))
  1312. goto out_unlock;
  1313. switch (type) {
  1314. case FILE_CPU_EXCLUSIVE:
  1315. retval = update_flag(CS_CPU_EXCLUSIVE, cs, val);
  1316. break;
  1317. case FILE_MEM_EXCLUSIVE:
  1318. retval = update_flag(CS_MEM_EXCLUSIVE, cs, val);
  1319. break;
  1320. case FILE_MEM_HARDWALL:
  1321. retval = update_flag(CS_MEM_HARDWALL, cs, val);
  1322. break;
  1323. case FILE_SCHED_LOAD_BALANCE:
  1324. retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, val);
  1325. break;
  1326. case FILE_MEMORY_MIGRATE:
  1327. retval = update_flag(CS_MEMORY_MIGRATE, cs, val);
  1328. break;
  1329. case FILE_MEMORY_PRESSURE_ENABLED:
  1330. cpuset_memory_pressure_enabled = !!val;
  1331. break;
  1332. case FILE_MEMORY_PRESSURE:
  1333. retval = -EACCES;
  1334. break;
  1335. case FILE_SPREAD_PAGE:
  1336. retval = update_flag(CS_SPREAD_PAGE, cs, val);
  1337. break;
  1338. case FILE_SPREAD_SLAB:
  1339. retval = update_flag(CS_SPREAD_SLAB, cs, val);
  1340. break;
  1341. default:
  1342. retval = -EINVAL;
  1343. break;
  1344. }
  1345. out_unlock:
  1346. mutex_unlock(&cpuset_mutex);
  1347. return retval;
  1348. }
  1349. static int cpuset_write_s64(struct cgroup *cgrp, struct cftype *cft, s64 val)
  1350. {
  1351. struct cpuset *cs = cgroup_cs(cgrp);
  1352. cpuset_filetype_t type = cft->private;
  1353. int retval = -ENODEV;
  1354. mutex_lock(&cpuset_mutex);
  1355. if (!is_cpuset_online(cs))
  1356. goto out_unlock;
  1357. switch (type) {
  1358. case FILE_SCHED_RELAX_DOMAIN_LEVEL:
  1359. retval = update_relax_domain_level(cs, val);
  1360. break;
  1361. default:
  1362. retval = -EINVAL;
  1363. break;
  1364. }
  1365. out_unlock:
  1366. mutex_unlock(&cpuset_mutex);
  1367. return retval;
  1368. }
  1369. /*
  1370. * Common handling for a write to a "cpus" or "mems" file.
  1371. */
  1372. static int cpuset_write_resmask(struct cgroup *cgrp, struct cftype *cft,
  1373. const char *buf)
  1374. {
  1375. struct cpuset *cs = cgroup_cs(cgrp);
  1376. struct cpuset *trialcs;
  1377. int retval = -ENODEV;
  1378. /*
  1379. * CPU or memory hotunplug may leave @cs w/o any execution
  1380. * resources, in which case the hotplug code asynchronously updates
  1381. * configuration and transfers all tasks to the nearest ancestor
  1382. * which can execute.
  1383. *
  1384. * As writes to "cpus" or "mems" may restore @cs's execution
  1385. * resources, wait for the previously scheduled operations before
  1386. * proceeding, so that we don't end up keep removing tasks added
  1387. * after execution capability is restored.
  1388. *
  1389. * Flushing cpuset_hotplug_work is enough to synchronize against
  1390. * hotplug hanlding; however, cpuset_attach() may schedule
  1391. * propagation work directly. Flush the workqueue too.
  1392. */
  1393. flush_work(&cpuset_hotplug_work);
  1394. flush_workqueue(cpuset_propagate_hotplug_wq);
  1395. mutex_lock(&cpuset_mutex);
  1396. if (!is_cpuset_online(cs))
  1397. goto out_unlock;
  1398. trialcs = alloc_trial_cpuset(cs);
  1399. if (!trialcs) {
  1400. retval = -ENOMEM;
  1401. goto out_unlock;
  1402. }
  1403. switch (cft->private) {
  1404. case FILE_CPULIST:
  1405. retval = update_cpumask(cs, trialcs, buf);
  1406. break;
  1407. case FILE_MEMLIST:
  1408. retval = update_nodemask(cs, trialcs, buf);
  1409. break;
  1410. default:
  1411. retval = -EINVAL;
  1412. break;
  1413. }
  1414. free_trial_cpuset(trialcs);
  1415. out_unlock:
  1416. mutex_unlock(&cpuset_mutex);
  1417. return retval;
  1418. }
  1419. /*
  1420. * These ascii lists should be read in a single call, by using a user
  1421. * buffer large enough to hold the entire map. If read in smaller
  1422. * chunks, there is no guarantee of atomicity. Since the display format
  1423. * used, list of ranges of sequential numbers, is variable length,
  1424. * and since these maps can change value dynamically, one could read
  1425. * gibberish by doing partial reads while a list was changing.
  1426. * A single large read to a buffer that crosses a page boundary is
  1427. * ok, because the result being copied to user land is not recomputed
  1428. * across a page fault.
  1429. */
  1430. static size_t cpuset_sprintf_cpulist(char *page, struct cpuset *cs)
  1431. {
  1432. size_t count;
  1433. mutex_lock(&callback_mutex);
  1434. count = cpulist_scnprintf(page, PAGE_SIZE, cs->cpus_allowed);
  1435. mutex_unlock(&callback_mutex);
  1436. return count;
  1437. }
  1438. static size_t cpuset_sprintf_memlist(char *page, struct cpuset *cs)
  1439. {
  1440. size_t count;
  1441. mutex_lock(&callback_mutex);
  1442. count = nodelist_scnprintf(page, PAGE_SIZE, cs->mems_allowed);
  1443. mutex_unlock(&callback_mutex);
  1444. return count;
  1445. }
  1446. static ssize_t cpuset_common_file_read(struct cgroup *cont,
  1447. struct cftype *cft,
  1448. struct file *file,
  1449. char __user *buf,
  1450. size_t nbytes, loff_t *ppos)
  1451. {
  1452. struct cpuset *cs = cgroup_cs(cont);
  1453. cpuset_filetype_t type = cft->private;
  1454. char *page;
  1455. ssize_t retval = 0;
  1456. char *s;
  1457. if (!(page = (char *)__get_free_page(GFP_TEMPORARY)))
  1458. return -ENOMEM;
  1459. s = page;
  1460. switch (type) {
  1461. case FILE_CPULIST:
  1462. s += cpuset_sprintf_cpulist(s, cs);
  1463. break;
  1464. case FILE_MEMLIST:
  1465. s += cpuset_sprintf_memlist(s, cs);
  1466. break;
  1467. default:
  1468. retval = -EINVAL;
  1469. goto out;
  1470. }
  1471. *s++ = '\n';
  1472. retval = simple_read_from_buffer(buf, nbytes, ppos, page, s - page);
  1473. out:
  1474. free_page((unsigned long)page);
  1475. return retval;
  1476. }
  1477. static u64 cpuset_read_u64(struct cgroup *cont, struct cftype *cft)
  1478. {
  1479. struct cpuset *cs = cgroup_cs(cont);
  1480. cpuset_filetype_t type = cft->private;
  1481. switch (type) {
  1482. case FILE_CPU_EXCLUSIVE:
  1483. return is_cpu_exclusive(cs);
  1484. case FILE_MEM_EXCLUSIVE:
  1485. return is_mem_exclusive(cs);
  1486. case FILE_MEM_HARDWALL:
  1487. return is_mem_hardwall(cs);
  1488. case FILE_SCHED_LOAD_BALANCE:
  1489. return is_sched_load_balance(cs);
  1490. case FILE_MEMORY_MIGRATE:
  1491. return is_memory_migrate(cs);
  1492. case FILE_MEMORY_PRESSURE_ENABLED:
  1493. return cpuset_memory_pressure_enabled;
  1494. case FILE_MEMORY_PRESSURE:
  1495. return fmeter_getrate(&cs->fmeter);
  1496. case FILE_SPREAD_PAGE:
  1497. return is_spread_page(cs);
  1498. case FILE_SPREAD_SLAB:
  1499. return is_spread_slab(cs);
  1500. default:
  1501. BUG();
  1502. }
  1503. /* Unreachable but makes gcc happy */
  1504. return 0;
  1505. }
  1506. static s64 cpuset_read_s64(struct cgroup *cont, struct cftype *cft)
  1507. {
  1508. struct cpuset *cs = cgroup_cs(cont);
  1509. cpuset_filetype_t type = cft->private;
  1510. switch (type) {
  1511. case FILE_SCHED_RELAX_DOMAIN_LEVEL:
  1512. return cs->relax_domain_level;
  1513. default:
  1514. BUG();
  1515. }
  1516. /* Unrechable but makes gcc happy */
  1517. return 0;
  1518. }
  1519. /*
  1520. * for the common functions, 'private' gives the type of file
  1521. */
  1522. static struct cftype files[] = {
  1523. {
  1524. .name = "cpus",
  1525. .read = cpuset_common_file_read,
  1526. .write_string = cpuset_write_resmask,
  1527. .max_write_len = (100U + 6 * NR_CPUS),
  1528. .private = FILE_CPULIST,
  1529. },
  1530. {
  1531. .name = "mems",
  1532. .read = cpuset_common_file_read,
  1533. .write_string = cpuset_write_resmask,
  1534. .max_write_len = (100U + 6 * MAX_NUMNODES),
  1535. .private = FILE_MEMLIST,
  1536. },
  1537. {
  1538. .name = "cpu_exclusive",
  1539. .read_u64 = cpuset_read_u64,
  1540. .write_u64 = cpuset_write_u64,
  1541. .private = FILE_CPU_EXCLUSIVE,
  1542. },
  1543. {
  1544. .name = "mem_exclusive",
  1545. .read_u64 = cpuset_read_u64,
  1546. .write_u64 = cpuset_write_u64,
  1547. .private = FILE_MEM_EXCLUSIVE,
  1548. },
  1549. {
  1550. .name = "mem_hardwall",
  1551. .read_u64 = cpuset_read_u64,
  1552. .write_u64 = cpuset_write_u64,
  1553. .private = FILE_MEM_HARDWALL,
  1554. },
  1555. {
  1556. .name = "sched_load_balance",
  1557. .read_u64 = cpuset_read_u64,
  1558. .write_u64 = cpuset_write_u64,
  1559. .private = FILE_SCHED_LOAD_BALANCE,
  1560. },
  1561. {
  1562. .name = "sched_relax_domain_level",
  1563. .read_s64 = cpuset_read_s64,
  1564. .write_s64 = cpuset_write_s64,
  1565. .private = FILE_SCHED_RELAX_DOMAIN_LEVEL,
  1566. },
  1567. {
  1568. .name = "memory_migrate",
  1569. .read_u64 = cpuset_read_u64,
  1570. .write_u64 = cpuset_write_u64,
  1571. .private = FILE_MEMORY_MIGRATE,
  1572. },
  1573. {
  1574. .name = "memory_pressure",
  1575. .read_u64 = cpuset_read_u64,
  1576. .write_u64 = cpuset_write_u64,
  1577. .private = FILE_MEMORY_PRESSURE,
  1578. .mode = S_IRUGO,
  1579. },
  1580. {
  1581. .name = "memory_spread_page",
  1582. .read_u64 = cpuset_read_u64,
  1583. .write_u64 = cpuset_write_u64,
  1584. .private = FILE_SPREAD_PAGE,
  1585. },
  1586. {
  1587. .name = "memory_spread_slab",
  1588. .read_u64 = cpuset_read_u64,
  1589. .write_u64 = cpuset_write_u64,
  1590. .private = FILE_SPREAD_SLAB,
  1591. },
  1592. {
  1593. .name = "memory_pressure_enabled",
  1594. .flags = CFTYPE_ONLY_ON_ROOT,
  1595. .read_u64 = cpuset_read_u64,
  1596. .write_u64 = cpuset_write_u64,
  1597. .private = FILE_MEMORY_PRESSURE_ENABLED,
  1598. },
  1599. { } /* terminate */
  1600. };
  1601. /*
  1602. * cpuset_css_alloc - allocate a cpuset css
  1603. * cont: control group that the new cpuset will be part of
  1604. */
  1605. static struct cgroup_subsys_state *cpuset_css_alloc(struct cgroup *cont)
  1606. {
  1607. struct cpuset *cs;
  1608. if (!cont->parent)
  1609. return &top_cpuset.css;
  1610. cs = kzalloc(sizeof(*cs), GFP_KERNEL);
  1611. if (!cs)
  1612. return ERR_PTR(-ENOMEM);
  1613. if (!alloc_cpumask_var(&cs->cpus_allowed, GFP_KERNEL)) {
  1614. kfree(cs);
  1615. return ERR_PTR(-ENOMEM);
  1616. }
  1617. set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
  1618. cpumask_clear(cs->cpus_allowed);
  1619. nodes_clear(cs->mems_allowed);
  1620. fmeter_init(&cs->fmeter);
  1621. INIT_WORK(&cs->hotplug_work, cpuset_propagate_hotplug_workfn);
  1622. cs->relax_domain_level = -1;
  1623. return &cs->css;
  1624. }
  1625. static int cpuset_css_online(struct cgroup *cgrp)
  1626. {
  1627. struct cpuset *cs = cgroup_cs(cgrp);
  1628. struct cpuset *parent = parent_cs(cs);
  1629. struct cpuset *tmp_cs;
  1630. struct cgroup *pos_cg;
  1631. if (!parent)
  1632. return 0;
  1633. mutex_lock(&cpuset_mutex);
  1634. set_bit(CS_ONLINE, &cs->flags);
  1635. if (is_spread_page(parent))
  1636. set_bit(CS_SPREAD_PAGE, &cs->flags);
  1637. if (is_spread_slab(parent))
  1638. set_bit(CS_SPREAD_SLAB, &cs->flags);
  1639. number_of_cpusets++;
  1640. if (!test_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags))
  1641. goto out_unlock;
  1642. /*
  1643. * Clone @parent's configuration if CGRP_CPUSET_CLONE_CHILDREN is
  1644. * set. This flag handling is implemented in cgroup core for
  1645. * histrical reasons - the flag may be specified during mount.
  1646. *
  1647. * Currently, if any sibling cpusets have exclusive cpus or mem, we
  1648. * refuse to clone the configuration - thereby refusing the task to
  1649. * be entered, and as a result refusing the sys_unshare() or
  1650. * clone() which initiated it. If this becomes a problem for some
  1651. * users who wish to allow that scenario, then this could be
  1652. * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
  1653. * (and likewise for mems) to the new cgroup.
  1654. */
  1655. rcu_read_lock();
  1656. cpuset_for_each_child(tmp_cs, pos_cg, parent) {
  1657. if (is_mem_exclusive(tmp_cs) || is_cpu_exclusive(tmp_cs)) {
  1658. rcu_read_unlock();
  1659. goto out_unlock;
  1660. }
  1661. }
  1662. rcu_read_unlock();
  1663. mutex_lock(&callback_mutex);
  1664. cs->mems_allowed = parent->mems_allowed;
  1665. cpumask_copy(cs->cpus_allowed, parent->cpus_allowed);
  1666. mutex_unlock(&callback_mutex);
  1667. out_unlock:
  1668. mutex_unlock(&cpuset_mutex);
  1669. return 0;
  1670. }
  1671. static void cpuset_css_offline(struct cgroup *cgrp)
  1672. {
  1673. struct cpuset *cs = cgroup_cs(cgrp);
  1674. mutex_lock(&cpuset_mutex);
  1675. if (is_sched_load_balance(cs))
  1676. update_flag(CS_SCHED_LOAD_BALANCE, cs, 0);
  1677. number_of_cpusets--;
  1678. clear_bit(CS_ONLINE, &cs->flags);
  1679. mutex_unlock(&cpuset_mutex);
  1680. }
  1681. /*
  1682. * If the cpuset being removed has its flag 'sched_load_balance'
  1683. * enabled, then simulate turning sched_load_balance off, which
  1684. * will call rebuild_sched_domains_locked().
  1685. */
  1686. static void cpuset_css_free(struct cgroup *cont)
  1687. {
  1688. struct cpuset *cs = cgroup_cs(cont);
  1689. free_cpumask_var(cs->cpus_allowed);
  1690. kfree(cs);
  1691. }
  1692. struct cgroup_subsys cpuset_subsys = {
  1693. .name = "cpuset",
  1694. .css_alloc = cpuset_css_alloc,
  1695. .css_online = cpuset_css_online,
  1696. .css_offline = cpuset_css_offline,
  1697. .css_free = cpuset_css_free,
  1698. .can_attach = cpuset_can_attach,
  1699. .cancel_attach = cpuset_cancel_attach,
  1700. .attach = cpuset_attach,
  1701. .subsys_id = cpuset_subsys_id,
  1702. .base_cftypes = files,
  1703. .early_init = 1,
  1704. };
  1705. /**
  1706. * cpuset_init - initialize cpusets at system boot
  1707. *
  1708. * Description: Initialize top_cpuset and the cpuset internal file system,
  1709. **/
  1710. int __init cpuset_init(void)
  1711. {
  1712. int err = 0;
  1713. if (!alloc_cpumask_var(&top_cpuset.cpus_allowed, GFP_KERNEL))
  1714. BUG();
  1715. cpumask_setall(top_cpuset.cpus_allowed);
  1716. nodes_setall(top_cpuset.mems_allowed);
  1717. fmeter_init(&top_cpuset.fmeter);
  1718. set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags);
  1719. top_cpuset.relax_domain_level = -1;
  1720. err = register_filesystem(&cpuset_fs_type);
  1721. if (err < 0)
  1722. return err;
  1723. if (!alloc_cpumask_var(&cpus_attach, GFP_KERNEL))
  1724. BUG();
  1725. number_of_cpusets = 1;
  1726. return 0;
  1727. }
  1728. /*
  1729. * If CPU and/or memory hotplug handlers, below, unplug any CPUs
  1730. * or memory nodes, we need to walk over the cpuset hierarchy,
  1731. * removing that CPU or node from all cpusets. If this removes the
  1732. * last CPU or node from a cpuset, then move the tasks in the empty
  1733. * cpuset to its next-highest non-empty parent.
  1734. */
  1735. static void remove_tasks_in_empty_cpuset(struct cpuset *cs)
  1736. {
  1737. struct cpuset *parent;
  1738. /*
  1739. * Find its next-highest non-empty parent, (top cpuset
  1740. * has online cpus, so can't be empty).
  1741. */
  1742. parent = parent_cs(cs);
  1743. while (cpumask_empty(parent->cpus_allowed) ||
  1744. nodes_empty(parent->mems_allowed))
  1745. parent = parent_cs(parent);
  1746. if (cgroup_transfer_tasks(parent->css.cgroup, cs->css.cgroup)) {
  1747. rcu_read_lock();
  1748. printk(KERN_ERR "cpuset: failed to transfer tasks out of empty cpuset %s\n",
  1749. cgroup_name(cs->css.cgroup));
  1750. rcu_read_unlock();
  1751. }
  1752. }
  1753. /**
  1754. * cpuset_propagate_hotplug_workfn - propagate CPU/memory hotplug to a cpuset
  1755. * @cs: cpuset in interest
  1756. *
  1757. * Compare @cs's cpu and mem masks against top_cpuset and if some have gone
  1758. * offline, update @cs accordingly. If @cs ends up with no CPU or memory,
  1759. * all its tasks are moved to the nearest ancestor with both resources.
  1760. */
  1761. static void cpuset_propagate_hotplug_workfn(struct work_struct *work)
  1762. {
  1763. static cpumask_t off_cpus;
  1764. static nodemask_t off_mems, tmp_mems;
  1765. struct cpuset *cs = container_of(work, struct cpuset, hotplug_work);
  1766. bool is_empty;
  1767. mutex_lock(&cpuset_mutex);
  1768. cpumask_andnot(&off_cpus, cs->cpus_allowed, top_cpuset.cpus_allowed);
  1769. nodes_andnot(off_mems, cs->mems_allowed, top_cpuset.mems_allowed);
  1770. /* remove offline cpus from @cs */
  1771. if (!cpumask_empty(&off_cpus)) {
  1772. mutex_lock(&callback_mutex);
  1773. cpumask_andnot(cs->cpus_allowed, cs->cpus_allowed, &off_cpus);
  1774. mutex_unlock(&callback_mutex);
  1775. update_tasks_cpumask(cs, NULL);
  1776. }
  1777. /* remove offline mems from @cs */
  1778. if (!nodes_empty(off_mems)) {
  1779. tmp_mems = cs->mems_allowed;
  1780. mutex_lock(&callback_mutex);
  1781. nodes_andnot(cs->mems_allowed, cs->mems_allowed, off_mems);
  1782. mutex_unlock(&callback_mutex);
  1783. update_tasks_nodemask(cs, &tmp_mems, NULL);
  1784. }
  1785. is_empty = cpumask_empty(cs->cpus_allowed) ||
  1786. nodes_empty(cs->mems_allowed);
  1787. mutex_unlock(&cpuset_mutex);
  1788. /*
  1789. * If @cs became empty, move tasks to the nearest ancestor with
  1790. * execution resources. This is full cgroup operation which will
  1791. * also call back into cpuset. Should be done outside any lock.
  1792. */
  1793. if (is_empty)
  1794. remove_tasks_in_empty_cpuset(cs);
  1795. /* the following may free @cs, should be the last operation */
  1796. css_put(&cs->css);
  1797. }
  1798. /**
  1799. * schedule_cpuset_propagate_hotplug - schedule hotplug propagation to a cpuset
  1800. * @cs: cpuset of interest
  1801. *
  1802. * Schedule cpuset_propagate_hotplug_workfn() which will update CPU and
  1803. * memory masks according to top_cpuset.
  1804. */
  1805. static void schedule_cpuset_propagate_hotplug(struct cpuset *cs)
  1806. {
  1807. /*
  1808. * Pin @cs. The refcnt will be released when the work item
  1809. * finishes executing.
  1810. */
  1811. if (!css_tryget(&cs->css))
  1812. return;
  1813. /*
  1814. * Queue @cs->hotplug_work. If already pending, lose the css ref.
  1815. * cpuset_propagate_hotplug_wq is ordered and propagation will
  1816. * happen in the order this function is called.
  1817. */
  1818. if (!queue_work(cpuset_propagate_hotplug_wq, &cs->hotplug_work))
  1819. css_put(&cs->css);
  1820. }
  1821. /**
  1822. * cpuset_hotplug_workfn - handle CPU/memory hotunplug for a cpuset
  1823. *
  1824. * This function is called after either CPU or memory configuration has
  1825. * changed and updates cpuset accordingly. The top_cpuset is always
  1826. * synchronized to cpu_active_mask and N_MEMORY, which is necessary in
  1827. * order to make cpusets transparent (of no affect) on systems that are
  1828. * actively using CPU hotplug but making no active use of cpusets.
  1829. *
  1830. * Non-root cpusets are only affected by offlining. If any CPUs or memory
  1831. * nodes have been taken down, cpuset_propagate_hotplug() is invoked on all
  1832. * descendants.
  1833. *
  1834. * Note that CPU offlining during suspend is ignored. We don't modify
  1835. * cpusets across suspend/resume cycles at all.
  1836. */
  1837. static void cpuset_hotplug_workfn(struct work_struct *work)
  1838. {
  1839. static cpumask_t new_cpus, tmp_cpus;
  1840. static nodemask_t new_mems, tmp_mems;
  1841. bool cpus_updated, mems_updated;
  1842. bool cpus_offlined, mems_offlined;
  1843. mutex_lock(&cpuset_mutex);
  1844. /* fetch the available cpus/mems and find out which changed how */
  1845. cpumask_copy(&new_cpus, cpu_active_mask);
  1846. new_mems = node_states[N_MEMORY];
  1847. cpus_updated = !cpumask_equal(top_cpuset.cpus_allowed, &new_cpus);
  1848. cpus_offlined = cpumask_andnot(&tmp_cpus, top_cpuset.cpus_allowed,
  1849. &new_cpus);
  1850. mems_updated = !nodes_equal(top_cpuset.mems_allowed, new_mems);
  1851. nodes_andnot(tmp_mems, top_cpuset.mems_allowed, new_mems);
  1852. mems_offlined = !nodes_empty(tmp_mems);
  1853. /* synchronize cpus_allowed to cpu_active_mask */
  1854. if (cpus_updated) {
  1855. mutex_lock(&callback_mutex);
  1856. cpumask_copy(top_cpuset.cpus_allowed, &new_cpus);
  1857. mutex_unlock(&callback_mutex);
  1858. /* we don't mess with cpumasks of tasks in top_cpuset */
  1859. }
  1860. /* synchronize mems_allowed to N_MEMORY */
  1861. if (mems_updated) {
  1862. tmp_mems = top_cpuset.mems_allowed;
  1863. mutex_lock(&callback_mutex);
  1864. top_cpuset.mems_allowed = new_mems;
  1865. mutex_unlock(&callback_mutex);
  1866. update_tasks_nodemask(&top_cpuset, &tmp_mems, NULL);
  1867. }
  1868. /* if cpus or mems went down, we need to propagate to descendants */
  1869. if (cpus_offlined || mems_offlined) {
  1870. struct cpuset *cs;
  1871. struct cgroup *pos_cgrp;
  1872. rcu_read_lock();
  1873. cpuset_for_each_descendant_pre(cs, pos_cgrp, &top_cpuset)
  1874. schedule_cpuset_propagate_hotplug(cs);
  1875. rcu_read_unlock();
  1876. }
  1877. mutex_unlock(&cpuset_mutex);
  1878. /* wait for propagations to finish */
  1879. flush_workqueue(cpuset_propagate_hotplug_wq);
  1880. /* rebuild sched domains if cpus_allowed has changed */
  1881. if (cpus_updated)
  1882. rebuild_sched_domains();
  1883. }
  1884. void cpuset_update_active_cpus(bool cpu_online)
  1885. {
  1886. /*
  1887. * We're inside cpu hotplug critical region which usually nests
  1888. * inside cgroup synchronization. Bounce actual hotplug processing
  1889. * to a work item to avoid reverse locking order.
  1890. *
  1891. * We still need to do partition_sched_domains() synchronously;
  1892. * otherwise, the scheduler will get confused and put tasks to the
  1893. * dead CPU. Fall back to the default single domain.
  1894. * cpuset_hotplug_workfn() will rebuild it as necessary.
  1895. */
  1896. partition_sched_domains(1, NULL, NULL);
  1897. schedule_work(&cpuset_hotplug_work);
  1898. }
  1899. #ifdef CONFIG_MEMORY_HOTPLUG
  1900. /*
  1901. * Keep top_cpuset.mems_allowed tracking node_states[N_MEMORY].
  1902. * Call this routine anytime after node_states[N_MEMORY] changes.
  1903. * See cpuset_update_active_cpus() for CPU hotplug handling.
  1904. */
  1905. static int cpuset_track_online_nodes(struct notifier_block *self,
  1906. unsigned long action, void *arg)
  1907. {
  1908. schedule_work(&cpuset_hotplug_work);
  1909. return NOTIFY_OK;
  1910. }
  1911. #endif
  1912. /**
  1913. * cpuset_init_smp - initialize cpus_allowed
  1914. *
  1915. * Description: Finish top cpuset after cpu, node maps are initialized
  1916. **/
  1917. void __init cpuset_init_smp(void)
  1918. {
  1919. cpumask_copy(top_cpuset.cpus_allowed, cpu_active_mask);
  1920. top_cpuset.mems_allowed = node_states[N_MEMORY];
  1921. hotplug_memory_notifier(cpuset_track_online_nodes, 10);
  1922. cpuset_propagate_hotplug_wq =
  1923. alloc_ordered_workqueue("cpuset_hotplug", 0);
  1924. BUG_ON(!cpuset_propagate_hotplug_wq);
  1925. }
  1926. /**
  1927. * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
  1928. * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
  1929. * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
  1930. *
  1931. * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
  1932. * attached to the specified @tsk. Guaranteed to return some non-empty
  1933. * subset of cpu_online_mask, even if this means going outside the
  1934. * tasks cpuset.
  1935. **/
  1936. void cpuset_cpus_allowed(struct task_struct *tsk, struct cpumask *pmask)
  1937. {
  1938. mutex_lock(&callback_mutex);
  1939. task_lock(tsk);
  1940. guarantee_online_cpus(task_cs(tsk), pmask);
  1941. task_unlock(tsk);
  1942. mutex_unlock(&callback_mutex);
  1943. }
  1944. void cpuset_cpus_allowed_fallback(struct task_struct *tsk)
  1945. {
  1946. const struct cpuset *cs;
  1947. rcu_read_lock();
  1948. cs = task_cs(tsk);
  1949. if (cs)
  1950. do_set_cpus_allowed(tsk, cs->cpus_allowed);
  1951. rcu_read_unlock();
  1952. /*
  1953. * We own tsk->cpus_allowed, nobody can change it under us.
  1954. *
  1955. * But we used cs && cs->cpus_allowed lockless and thus can
  1956. * race with cgroup_attach_task() or update_cpumask() and get
  1957. * the wrong tsk->cpus_allowed. However, both cases imply the
  1958. * subsequent cpuset_change_cpumask()->set_cpus_allowed_ptr()
  1959. * which takes task_rq_lock().
  1960. *
  1961. * If we are called after it dropped the lock we must see all
  1962. * changes in tsk_cs()->cpus_allowed. Otherwise we can temporary
  1963. * set any mask even if it is not right from task_cs() pov,
  1964. * the pending set_cpus_allowed_ptr() will fix things.
  1965. *
  1966. * select_fallback_rq() will fix things ups and set cpu_possible_mask
  1967. * if required.
  1968. */
  1969. }
  1970. void cpuset_init_current_mems_allowed(void)
  1971. {
  1972. nodes_setall(current->mems_allowed);
  1973. }
  1974. /**
  1975. * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
  1976. * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
  1977. *
  1978. * Description: Returns the nodemask_t mems_allowed of the cpuset
  1979. * attached to the specified @tsk. Guaranteed to return some non-empty
  1980. * subset of node_states[N_MEMORY], even if this means going outside the
  1981. * tasks cpuset.
  1982. **/
  1983. nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
  1984. {
  1985. nodemask_t mask;
  1986. mutex_lock(&callback_mutex);
  1987. task_lock(tsk);
  1988. guarantee_online_mems(task_cs(tsk), &mask);
  1989. task_unlock(tsk);
  1990. mutex_unlock(&callback_mutex);
  1991. return mask;
  1992. }
  1993. /**
  1994. * cpuset_nodemask_valid_mems_allowed - check nodemask vs. curremt mems_allowed
  1995. * @nodemask: the nodemask to be checked
  1996. *
  1997. * Are any of the nodes in the nodemask allowed in current->mems_allowed?
  1998. */
  1999. int cpuset_nodemask_valid_mems_allowed(nodemask_t *nodemask)
  2000. {
  2001. return nodes_intersects(*nodemask, current->mems_allowed);
  2002. }
  2003. /*
  2004. * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
  2005. * mem_hardwall ancestor to the specified cpuset. Call holding
  2006. * callback_mutex. If no ancestor is mem_exclusive or mem_hardwall
  2007. * (an unusual configuration), then returns the root cpuset.
  2008. */
  2009. static const struct cpuset *nearest_hardwall_ancestor(const struct cpuset *cs)
  2010. {
  2011. while (!(is_mem_exclusive(cs) || is_mem_hardwall(cs)) && parent_cs(cs))
  2012. cs = parent_cs(cs);
  2013. return cs;
  2014. }
  2015. /**
  2016. * cpuset_node_allowed_softwall - Can we allocate on a memory node?
  2017. * @node: is this an allowed node?
  2018. * @gfp_mask: memory allocation flags
  2019. *
  2020. * If we're in interrupt, yes, we can always allocate. If __GFP_THISNODE is
  2021. * set, yes, we can always allocate. If node is in our task's mems_allowed,
  2022. * yes. If it's not a __GFP_HARDWALL request and this node is in the nearest
  2023. * hardwalled cpuset ancestor to this task's cpuset, yes. If the task has been
  2024. * OOM killed and has access to memory reserves as specified by the TIF_MEMDIE
  2025. * flag, yes.
  2026. * Otherwise, no.
  2027. *
  2028. * If __GFP_HARDWALL is set, cpuset_node_allowed_softwall() reduces to
  2029. * cpuset_node_allowed_hardwall(). Otherwise, cpuset_node_allowed_softwall()
  2030. * might sleep, and might allow a node from an enclosing cpuset.
  2031. *
  2032. * cpuset_node_allowed_hardwall() only handles the simpler case of hardwall
  2033. * cpusets, and never sleeps.
  2034. *
  2035. * The __GFP_THISNODE placement logic is really handled elsewhere,
  2036. * by forcibly using a zonelist starting at a specified node, and by
  2037. * (in get_page_from_freelist()) refusing to consider the zones for
  2038. * any node on the zonelist except the first. By the time any such
  2039. * calls get to this routine, we should just shut up and say 'yes'.
  2040. *
  2041. * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
  2042. * and do not allow allocations outside the current tasks cpuset
  2043. * unless the task has been OOM killed as is marked TIF_MEMDIE.
  2044. * GFP_KERNEL allocations are not so marked, so can escape to the
  2045. * nearest enclosing hardwalled ancestor cpuset.
  2046. *
  2047. * Scanning up parent cpusets requires callback_mutex. The
  2048. * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
  2049. * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
  2050. * current tasks mems_allowed came up empty on the first pass over
  2051. * the zonelist. So only GFP_KERNEL allocations, if all nodes in the
  2052. * cpuset are short of memory, might require taking the callback_mutex
  2053. * mutex.
  2054. *
  2055. * The first call here from mm/page_alloc:get_page_from_freelist()
  2056. * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
  2057. * so no allocation on a node outside the cpuset is allowed (unless
  2058. * in interrupt, of course).
  2059. *
  2060. * The second pass through get_page_from_freelist() doesn't even call
  2061. * here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
  2062. * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
  2063. * in alloc_flags. That logic and the checks below have the combined
  2064. * affect that:
  2065. * in_interrupt - any node ok (current task context irrelevant)
  2066. * GFP_ATOMIC - any node ok
  2067. * TIF_MEMDIE - any node ok
  2068. * GFP_KERNEL - any node in enclosing hardwalled cpuset ok
  2069. * GFP_USER - only nodes in current tasks mems allowed ok.
  2070. *
  2071. * Rule:
  2072. * Don't call cpuset_node_allowed_softwall if you can't sleep, unless you
  2073. * pass in the __GFP_HARDWALL flag set in gfp_flag, which disables
  2074. * the code that might scan up ancestor cpusets and sleep.
  2075. */
  2076. int __cpuset_node_allowed_softwall(int node, gfp_t gfp_mask)
  2077. {
  2078. const struct cpuset *cs; /* current cpuset ancestors */
  2079. int allowed; /* is allocation in zone z allowed? */
  2080. if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
  2081. return 1;
  2082. might_sleep_if(!(gfp_mask & __GFP_HARDWALL));
  2083. if (node_isset(node, current->mems_allowed))
  2084. return 1;
  2085. /*
  2086. * Allow tasks that have access to memory reserves because they have
  2087. * been OOM killed to get memory anywhere.
  2088. */
  2089. if (unlikely(test_thread_flag(TIF_MEMDIE)))
  2090. return 1;
  2091. if (gfp_mask & __GFP_HARDWALL) /* If hardwall request, stop here */
  2092. return 0;
  2093. if (current->flags & PF_EXITING) /* Let dying task have memory */
  2094. return 1;
  2095. /* Not hardwall and node outside mems_allowed: scan up cpusets */
  2096. mutex_lock(&callback_mutex);
  2097. task_lock(current);
  2098. cs = nearest_hardwall_ancestor(task_cs(current));
  2099. task_unlock(current);
  2100. allowed = node_isset(node, cs->mems_allowed);
  2101. mutex_unlock(&callback_mutex);
  2102. return allowed;
  2103. }
  2104. /*
  2105. * cpuset_node_allowed_hardwall - Can we allocate on a memory node?
  2106. * @node: is this an allowed node?
  2107. * @gfp_mask: memory allocation flags
  2108. *
  2109. * If we're in interrupt, yes, we can always allocate. If __GFP_THISNODE is
  2110. * set, yes, we can always allocate. If node is in our task's mems_allowed,
  2111. * yes. If the task has been OOM killed and has access to memory reserves as
  2112. * specified by the TIF_MEMDIE flag, yes.
  2113. * Otherwise, no.
  2114. *
  2115. * The __GFP_THISNODE placement logic is really handled elsewhere,
  2116. * by forcibly using a zonelist starting at a specified node, and by
  2117. * (in get_page_from_freelist()) refusing to consider the zones for
  2118. * any node on the zonelist except the first. By the time any such
  2119. * calls get to this routine, we should just shut up and say 'yes'.
  2120. *
  2121. * Unlike the cpuset_node_allowed_softwall() variant, above,
  2122. * this variant requires that the node be in the current task's
  2123. * mems_allowed or that we're in interrupt. It does not scan up the
  2124. * cpuset hierarchy for the nearest enclosing mem_exclusive cpuset.
  2125. * It never sleeps.
  2126. */
  2127. int __cpuset_node_allowed_hardwall(int node, gfp_t gfp_mask)
  2128. {
  2129. if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
  2130. return 1;
  2131. if (node_isset(node, current->mems_allowed))
  2132. return 1;
  2133. /*
  2134. * Allow tasks that have access to memory reserves because they have
  2135. * been OOM killed to get memory anywhere.
  2136. */
  2137. if (unlikely(test_thread_flag(TIF_MEMDIE)))
  2138. return 1;
  2139. return 0;
  2140. }
  2141. /**
  2142. * cpuset_mem_spread_node() - On which node to begin search for a file page
  2143. * cpuset_slab_spread_node() - On which node to begin search for a slab page
  2144. *
  2145. * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
  2146. * tasks in a cpuset with is_spread_page or is_spread_slab set),
  2147. * and if the memory allocation used cpuset_mem_spread_node()
  2148. * to determine on which node to start looking, as it will for
  2149. * certain page cache or slab cache pages such as used for file
  2150. * system buffers and inode caches, then instead of starting on the
  2151. * local node to look for a free page, rather spread the starting
  2152. * node around the tasks mems_allowed nodes.
  2153. *
  2154. * We don't have to worry about the returned node being offline
  2155. * because "it can't happen", and even if it did, it would be ok.
  2156. *
  2157. * The routines calling guarantee_online_mems() are careful to
  2158. * only set nodes in task->mems_allowed that are online. So it
  2159. * should not be possible for the following code to return an
  2160. * offline node. But if it did, that would be ok, as this routine
  2161. * is not returning the node where the allocation must be, only
  2162. * the node where the search should start. The zonelist passed to
  2163. * __alloc_pages() will include all nodes. If the slab allocator
  2164. * is passed an offline node, it will fall back to the local node.
  2165. * See kmem_cache_alloc_node().
  2166. */
  2167. static int cpuset_spread_node(int *rotor)
  2168. {
  2169. int node;
  2170. node = next_node(*rotor, current->mems_allowed);
  2171. if (node == MAX_NUMNODES)
  2172. node = first_node(current->mems_allowed);
  2173. *rotor = node;
  2174. return node;
  2175. }
  2176. int cpuset_mem_spread_node(void)
  2177. {
  2178. if (current->cpuset_mem_spread_rotor == NUMA_NO_NODE)
  2179. current->cpuset_mem_spread_rotor =
  2180. node_random(&current->mems_allowed);
  2181. return cpuset_spread_node(&current->cpuset_mem_spread_rotor);
  2182. }
  2183. int cpuset_slab_spread_node(void)
  2184. {
  2185. if (current->cpuset_slab_spread_rotor == NUMA_NO_NODE)
  2186. current->cpuset_slab_spread_rotor =
  2187. node_random(&current->mems_allowed);
  2188. return cpuset_spread_node(&current->cpuset_slab_spread_rotor);
  2189. }
  2190. EXPORT_SYMBOL_GPL(cpuset_mem_spread_node);
  2191. /**
  2192. * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
  2193. * @tsk1: pointer to task_struct of some task.
  2194. * @tsk2: pointer to task_struct of some other task.
  2195. *
  2196. * Description: Return true if @tsk1's mems_allowed intersects the
  2197. * mems_allowed of @tsk2. Used by the OOM killer to determine if
  2198. * one of the task's memory usage might impact the memory available
  2199. * to the other.
  2200. **/
  2201. int cpuset_mems_allowed_intersects(const struct task_struct *tsk1,
  2202. const struct task_struct *tsk2)
  2203. {
  2204. return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed);
  2205. }
  2206. #define CPUSET_NODELIST_LEN (256)
  2207. /**
  2208. * cpuset_print_task_mems_allowed - prints task's cpuset and mems_allowed
  2209. * @task: pointer to task_struct of some task.
  2210. *
  2211. * Description: Prints @task's name, cpuset name, and cached copy of its
  2212. * mems_allowed to the kernel log. Must hold task_lock(task) to allow
  2213. * dereferencing task_cs(task).
  2214. */
  2215. void cpuset_print_task_mems_allowed(struct task_struct *tsk)
  2216. {
  2217. /* Statically allocated to prevent using excess stack. */
  2218. static char cpuset_nodelist[CPUSET_NODELIST_LEN];
  2219. static DEFINE_SPINLOCK(cpuset_buffer_lock);
  2220. struct cgroup *cgrp = task_cs(tsk)->css.cgroup;
  2221. rcu_read_lock();
  2222. spin_lock(&cpuset_buffer_lock);
  2223. nodelist_scnprintf(cpuset_nodelist, CPUSET_NODELIST_LEN,
  2224. tsk->mems_allowed);
  2225. printk(KERN_INFO "%s cpuset=%s mems_allowed=%s\n",
  2226. tsk->comm, cgroup_name(cgrp), cpuset_nodelist);
  2227. spin_unlock(&cpuset_buffer_lock);
  2228. rcu_read_unlock();
  2229. }
  2230. /*
  2231. * Collection of memory_pressure is suppressed unless
  2232. * this flag is enabled by writing "1" to the special
  2233. * cpuset file 'memory_pressure_enabled' in the root cpuset.
  2234. */
  2235. int cpuset_memory_pressure_enabled __read_mostly;
  2236. /**
  2237. * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
  2238. *
  2239. * Keep a running average of the rate of synchronous (direct)
  2240. * page reclaim efforts initiated by tasks in each cpuset.
  2241. *
  2242. * This represents the rate at which some task in the cpuset
  2243. * ran low on memory on all nodes it was allowed to use, and
  2244. * had to enter the kernels page reclaim code in an effort to
  2245. * create more free memory by tossing clean pages or swapping
  2246. * or writing dirty pages.
  2247. *
  2248. * Display to user space in the per-cpuset read-only file
  2249. * "memory_pressure". Value displayed is an integer
  2250. * representing the recent rate of entry into the synchronous
  2251. * (direct) page reclaim by any task attached to the cpuset.
  2252. **/
  2253. void __cpuset_memory_pressure_bump(void)
  2254. {
  2255. task_lock(current);
  2256. fmeter_markevent(&task_cs(current)->fmeter);
  2257. task_unlock(current);
  2258. }
  2259. #ifdef CONFIG_PROC_PID_CPUSET
  2260. /*
  2261. * proc_cpuset_show()
  2262. * - Print tasks cpuset path into seq_file.
  2263. * - Used for /proc/<pid>/cpuset.
  2264. * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
  2265. * doesn't really matter if tsk->cpuset changes after we read it,
  2266. * and we take cpuset_mutex, keeping cpuset_attach() from changing it
  2267. * anyway.
  2268. */
  2269. static int proc_cpuset_show(struct seq_file *m, void *unused_v)
  2270. {
  2271. struct pid *pid;
  2272. struct task_struct *tsk;
  2273. char *buf;
  2274. struct cgroup_subsys_state *css;
  2275. int retval;
  2276. retval = -ENOMEM;
  2277. buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
  2278. if (!buf)
  2279. goto out;
  2280. retval = -ESRCH;
  2281. pid = m->private;
  2282. tsk = get_pid_task(pid, PIDTYPE_PID);
  2283. if (!tsk)
  2284. goto out_free;
  2285. rcu_read_lock();
  2286. css = task_subsys_state(tsk, cpuset_subsys_id);
  2287. retval = cgroup_path(css->cgroup, buf, PAGE_SIZE);
  2288. rcu_read_unlock();
  2289. if (retval < 0)
  2290. goto out_put_task;
  2291. seq_puts(m, buf);
  2292. seq_putc(m, '\n');
  2293. out_put_task:
  2294. put_task_struct(tsk);
  2295. out_free:
  2296. kfree(buf);
  2297. out:
  2298. return retval;
  2299. }
  2300. static int cpuset_open(struct inode *inode, struct file *file)
  2301. {
  2302. struct pid *pid = PROC_I(inode)->pid;
  2303. return single_open(file, proc_cpuset_show, pid);
  2304. }
  2305. const struct file_operations proc_cpuset_operations = {
  2306. .open = cpuset_open,
  2307. .read = seq_read,
  2308. .llseek = seq_lseek,
  2309. .release = single_release,
  2310. };
  2311. #endif /* CONFIG_PROC_PID_CPUSET */
  2312. /* Display task mems_allowed in /proc/<pid>/status file. */
  2313. void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task)
  2314. {
  2315. seq_printf(m, "Mems_allowed:\t");
  2316. seq_nodemask(m, &task->mems_allowed);
  2317. seq_printf(m, "\n");
  2318. seq_printf(m, "Mems_allowed_list:\t");
  2319. seq_nodemask_list(m, &task->mems_allowed);
  2320. seq_printf(m, "\n");
  2321. }