| 1 | /* |
| 2 | * kernel/cpuset.c |
| 3 | * |
| 4 | * Processor and Memory placement constraints for sets of tasks. |
| 5 | * |
| 6 | * Copyright (C) 2003 BULL SA. |
| 7 | * Copyright (C) 2004-2007 Silicon Graphics, Inc. |
| 8 | * Copyright (C) 2006 Google, Inc |
| 9 | * |
| 10 | * Portions derived from Patrick Mochel's sysfs code. |
| 11 | * sysfs is Copyright (c) 2001-3 Patrick Mochel |
| 12 | * |
| 13 | * 2003-10-10 Written by Simon Derr. |
| 14 | * 2003-10-22 Updates by Stephen Hemminger. |
| 15 | * 2004 May-July Rework by Paul Jackson. |
| 16 | * 2006 Rework by Paul Menage to use generic cgroups |
| 17 | * 2008 Rework of the scheduler domains and CPU hotplug handling |
| 18 | * by Max Krasnyansky |
| 19 | * |
| 20 | * This file is subject to the terms and conditions of the GNU General Public |
| 21 | * License. See the file COPYING in the main directory of the Linux |
| 22 | * distribution for more details. |
| 23 | */ |
| 24 | |
| 25 | #include <linux/cpu.h> |
| 26 | #include <linux/cpumask.h> |
| 27 | #include <linux/cpuset.h> |
| 28 | #include <linux/err.h> |
| 29 | #include <linux/errno.h> |
| 30 | #include <linux/file.h> |
| 31 | #include <linux/fs.h> |
| 32 | #include <linux/init.h> |
| 33 | #include <linux/interrupt.h> |
| 34 | #include <linux/kernel.h> |
| 35 | #include <linux/kmod.h> |
| 36 | #include <linux/list.h> |
| 37 | #include <linux/mempolicy.h> |
| 38 | #include <linux/mm.h> |
| 39 | #include <linux/memory.h> |
| 40 | #include <linux/export.h> |
| 41 | #include <linux/mount.h> |
| 42 | #include <linux/namei.h> |
| 43 | #include <linux/pagemap.h> |
| 44 | #include <linux/proc_fs.h> |
| 45 | #include <linux/rcupdate.h> |
| 46 | #include <linux/sched.h> |
| 47 | #include <linux/seq_file.h> |
| 48 | #include <linux/security.h> |
| 49 | #include <linux/slab.h> |
| 50 | #include <linux/spinlock.h> |
| 51 | #include <linux/stat.h> |
| 52 | #include <linux/string.h> |
| 53 | #include <linux/time.h> |
| 54 | #include <linux/backing-dev.h> |
| 55 | #include <linux/sort.h> |
| 56 | |
| 57 | #include <asm/uaccess.h> |
| 58 | #include <linux/atomic.h> |
| 59 | #include <linux/mutex.h> |
| 60 | #include <linux/workqueue.h> |
| 61 | #include <linux/cgroup.h> |
| 62 | #include <linux/wait.h> |
| 63 | |
| 64 | /* |
| 65 | * Tracks how many cpusets are currently defined in system. |
| 66 | * When there is only one cpuset (the root cpuset) we can |
| 67 | * short circuit some hooks. |
| 68 | */ |
| 69 | int number_of_cpusets __read_mostly; |
| 70 | |
| 71 | /* Forward declare cgroup structures */ |
| 72 | struct cgroup_subsys cpuset_subsys; |
| 73 | |
| 74 | /* See "Frequency meter" comments, below. */ |
| 75 | |
| 76 | struct fmeter { |
| 77 | int cnt; /* unprocessed events count */ |
| 78 | int val; /* most recent output value */ |
| 79 | time_t time; /* clock (secs) when val computed */ |
| 80 | spinlock_t lock; /* guards read or write of above */ |
| 81 | }; |
| 82 | |
| 83 | struct cpuset { |
| 84 | struct cgroup_subsys_state css; |
| 85 | |
| 86 | unsigned long flags; /* "unsigned long" so bitops work */ |
| 87 | cpumask_var_t cpus_allowed; /* CPUs allowed to tasks in cpuset */ |
| 88 | nodemask_t mems_allowed; /* Memory Nodes allowed to tasks */ |
| 89 | |
| 90 | /* |
| 91 | * This is old Memory Nodes tasks took on. |
| 92 | * |
| 93 | * - top_cpuset.old_mems_allowed is initialized to mems_allowed. |
| 94 | * - A new cpuset's old_mems_allowed is initialized when some |
| 95 | * task is moved into it. |
| 96 | * - old_mems_allowed is used in cpuset_migrate_mm() when we change |
| 97 | * cpuset.mems_allowed and have tasks' nodemask updated, and |
| 98 | * then old_mems_allowed is updated to mems_allowed. |
| 99 | */ |
| 100 | nodemask_t old_mems_allowed; |
| 101 | |
| 102 | struct fmeter fmeter; /* memory_pressure filter */ |
| 103 | |
| 104 | /* |
| 105 | * Tasks are being attached to this cpuset. Used to prevent |
| 106 | * zeroing cpus/mems_allowed between ->can_attach() and ->attach(). |
| 107 | */ |
| 108 | int attach_in_progress; |
| 109 | |
| 110 | /* partition number for rebuild_sched_domains() */ |
| 111 | int pn; |
| 112 | |
| 113 | /* for custom sched domain */ |
| 114 | int relax_domain_level; |
| 115 | }; |
| 116 | |
| 117 | static inline struct cpuset *css_cs(struct cgroup_subsys_state *css) |
| 118 | { |
| 119 | return css ? container_of(css, struct cpuset, css) : NULL; |
| 120 | } |
| 121 | |
| 122 | /* Retrieve the cpuset for a cgroup */ |
| 123 | static inline struct cpuset *cgroup_cs(struct cgroup *cgrp) |
| 124 | { |
| 125 | return css_cs(cgroup_css(cgrp, cpuset_subsys_id)); |
| 126 | } |
| 127 | |
| 128 | /* Retrieve the cpuset for a task */ |
| 129 | static inline struct cpuset *task_cs(struct task_struct *task) |
| 130 | { |
| 131 | return css_cs(task_css(task, cpuset_subsys_id)); |
| 132 | } |
| 133 | |
| 134 | static inline struct cpuset *parent_cs(struct cpuset *cs) |
| 135 | { |
| 136 | return css_cs(css_parent(&cs->css)); |
| 137 | } |
| 138 | |
| 139 | #ifdef CONFIG_NUMA |
| 140 | static inline bool task_has_mempolicy(struct task_struct *task) |
| 141 | { |
| 142 | return task->mempolicy; |
| 143 | } |
| 144 | #else |
| 145 | static inline bool task_has_mempolicy(struct task_struct *task) |
| 146 | { |
| 147 | return false; |
| 148 | } |
| 149 | #endif |
| 150 | |
| 151 | |
| 152 | /* bits in struct cpuset flags field */ |
| 153 | typedef enum { |
| 154 | CS_ONLINE, |
| 155 | CS_CPU_EXCLUSIVE, |
| 156 | CS_MEM_EXCLUSIVE, |
| 157 | CS_MEM_HARDWALL, |
| 158 | CS_MEMORY_MIGRATE, |
| 159 | CS_SCHED_LOAD_BALANCE, |
| 160 | CS_SPREAD_PAGE, |
| 161 | CS_SPREAD_SLAB, |
| 162 | } cpuset_flagbits_t; |
| 163 | |
| 164 | /* convenient tests for these bits */ |
| 165 | static inline bool is_cpuset_online(const struct cpuset *cs) |
| 166 | { |
| 167 | return test_bit(CS_ONLINE, &cs->flags); |
| 168 | } |
| 169 | |
| 170 | static inline int is_cpu_exclusive(const struct cpuset *cs) |
| 171 | { |
| 172 | return test_bit(CS_CPU_EXCLUSIVE, &cs->flags); |
| 173 | } |
| 174 | |
| 175 | static inline int is_mem_exclusive(const struct cpuset *cs) |
| 176 | { |
| 177 | return test_bit(CS_MEM_EXCLUSIVE, &cs->flags); |
| 178 | } |
| 179 | |
| 180 | static inline int is_mem_hardwall(const struct cpuset *cs) |
| 181 | { |
| 182 | return test_bit(CS_MEM_HARDWALL, &cs->flags); |
| 183 | } |
| 184 | |
| 185 | static inline int is_sched_load_balance(const struct cpuset *cs) |
| 186 | { |
| 187 | return test_bit(CS_SCHED_LOAD_BALANCE, &cs->flags); |
| 188 | } |
| 189 | |
| 190 | static inline int is_memory_migrate(const struct cpuset *cs) |
| 191 | { |
| 192 | return test_bit(CS_MEMORY_MIGRATE, &cs->flags); |
| 193 | } |
| 194 | |
| 195 | static inline int is_spread_page(const struct cpuset *cs) |
| 196 | { |
| 197 | return test_bit(CS_SPREAD_PAGE, &cs->flags); |
| 198 | } |
| 199 | |
| 200 | static inline int is_spread_slab(const struct cpuset *cs) |
| 201 | { |
| 202 | return test_bit(CS_SPREAD_SLAB, &cs->flags); |
| 203 | } |
| 204 | |
| 205 | static struct cpuset top_cpuset = { |
| 206 | .flags = ((1 << CS_ONLINE) | (1 << CS_CPU_EXCLUSIVE) | |
| 207 | (1 << CS_MEM_EXCLUSIVE)), |
| 208 | }; |
| 209 | |
| 210 | /** |
| 211 | * cpuset_for_each_child - traverse online children of a cpuset |
| 212 | * @child_cs: loop cursor pointing to the current child |
| 213 | * @pos_cgrp: used for iteration |
| 214 | * @parent_cs: target cpuset to walk children of |
| 215 | * |
| 216 | * Walk @child_cs through the online children of @parent_cs. Must be used |
| 217 | * with RCU read locked. |
| 218 | */ |
| 219 | #define cpuset_for_each_child(child_cs, pos_cgrp, parent_cs) \ |
| 220 | cgroup_for_each_child((pos_cgrp), (parent_cs)->css.cgroup) \ |
| 221 | if (is_cpuset_online(((child_cs) = cgroup_cs((pos_cgrp))))) |
| 222 | |
| 223 | /** |
| 224 | * cpuset_for_each_descendant_pre - pre-order walk of a cpuset's descendants |
| 225 | * @des_cs: loop cursor pointing to the current descendant |
| 226 | * @pos_cgrp: used for iteration |
| 227 | * @root_cs: target cpuset to walk ancestor of |
| 228 | * |
| 229 | * Walk @des_cs through the online descendants of @root_cs. Must be used |
| 230 | * with RCU read locked. The caller may modify @pos_cgrp by calling |
| 231 | * cgroup_rightmost_descendant() to skip subtree. |
| 232 | */ |
| 233 | #define cpuset_for_each_descendant_pre(des_cs, pos_cgrp, root_cs) \ |
| 234 | cgroup_for_each_descendant_pre((pos_cgrp), (root_cs)->css.cgroup) \ |
| 235 | if (is_cpuset_online(((des_cs) = cgroup_cs((pos_cgrp))))) |
| 236 | |
| 237 | /* |
| 238 | * There are two global mutexes guarding cpuset structures - cpuset_mutex |
| 239 | * and callback_mutex. The latter may nest inside the former. We also |
| 240 | * require taking task_lock() when dereferencing a task's cpuset pointer. |
| 241 | * See "The task_lock() exception", at the end of this comment. |
| 242 | * |
| 243 | * A task must hold both mutexes to modify cpusets. If a task holds |
| 244 | * cpuset_mutex, then it blocks others wanting that mutex, ensuring that it |
| 245 | * is the only task able to also acquire callback_mutex and be able to |
| 246 | * modify cpusets. It can perform various checks on the cpuset structure |
| 247 | * first, knowing nothing will change. It can also allocate memory while |
| 248 | * just holding cpuset_mutex. While it is performing these checks, various |
| 249 | * callback routines can briefly acquire callback_mutex to query cpusets. |
| 250 | * Once it is ready to make the changes, it takes callback_mutex, blocking |
| 251 | * everyone else. |
| 252 | * |
| 253 | * Calls to the kernel memory allocator can not be made while holding |
| 254 | * callback_mutex, as that would risk double tripping on callback_mutex |
| 255 | * from one of the callbacks into the cpuset code from within |
| 256 | * __alloc_pages(). |
| 257 | * |
| 258 | * If a task is only holding callback_mutex, then it has read-only |
| 259 | * access to cpusets. |
| 260 | * |
| 261 | * Now, the task_struct fields mems_allowed and mempolicy may be changed |
| 262 | * by other task, we use alloc_lock in the task_struct fields to protect |
| 263 | * them. |
| 264 | * |
| 265 | * The cpuset_common_file_read() handlers only hold callback_mutex across |
| 266 | * small pieces of code, such as when reading out possibly multi-word |
| 267 | * cpumasks and nodemasks. |
| 268 | * |
| 269 | * Accessing a task's cpuset should be done in accordance with the |
| 270 | * guidelines for accessing subsystem state in kernel/cgroup.c |
| 271 | */ |
| 272 | |
| 273 | static DEFINE_MUTEX(cpuset_mutex); |
| 274 | static DEFINE_MUTEX(callback_mutex); |
| 275 | |
| 276 | /* |
| 277 | * CPU / memory hotplug is handled asynchronously. |
| 278 | */ |
| 279 | static void cpuset_hotplug_workfn(struct work_struct *work); |
| 280 | static DECLARE_WORK(cpuset_hotplug_work, cpuset_hotplug_workfn); |
| 281 | |
| 282 | static DECLARE_WAIT_QUEUE_HEAD(cpuset_attach_wq); |
| 283 | |
| 284 | /* |
| 285 | * This is ugly, but preserves the userspace API for existing cpuset |
| 286 | * users. If someone tries to mount the "cpuset" filesystem, we |
| 287 | * silently switch it to mount "cgroup" instead |
| 288 | */ |
| 289 | static struct dentry *cpuset_mount(struct file_system_type *fs_type, |
| 290 | int flags, const char *unused_dev_name, void *data) |
| 291 | { |
| 292 | struct file_system_type *cgroup_fs = get_fs_type("cgroup"); |
| 293 | struct dentry *ret = ERR_PTR(-ENODEV); |
| 294 | if (cgroup_fs) { |
| 295 | char mountopts[] = |
| 296 | "cpuset,noprefix," |
| 297 | "release_agent=/sbin/cpuset_release_agent"; |
| 298 | ret = cgroup_fs->mount(cgroup_fs, flags, |
| 299 | unused_dev_name, mountopts); |
| 300 | put_filesystem(cgroup_fs); |
| 301 | } |
| 302 | return ret; |
| 303 | } |
| 304 | |
| 305 | static struct file_system_type cpuset_fs_type = { |
| 306 | .name = "cpuset", |
| 307 | .mount = cpuset_mount, |
| 308 | }; |
| 309 | |
| 310 | /* |
| 311 | * Return in pmask the portion of a cpusets's cpus_allowed that |
| 312 | * are online. If none are online, walk up the cpuset hierarchy |
| 313 | * until we find one that does have some online cpus. The top |
| 314 | * cpuset always has some cpus online. |
| 315 | * |
| 316 | * One way or another, we guarantee to return some non-empty subset |
| 317 | * of cpu_online_mask. |
| 318 | * |
| 319 | * Call with callback_mutex held. |
| 320 | */ |
| 321 | static void guarantee_online_cpus(struct cpuset *cs, struct cpumask *pmask) |
| 322 | { |
| 323 | while (!cpumask_intersects(cs->cpus_allowed, cpu_online_mask)) |
| 324 | cs = parent_cs(cs); |
| 325 | cpumask_and(pmask, cs->cpus_allowed, cpu_online_mask); |
| 326 | } |
| 327 | |
| 328 | /* |
| 329 | * Return in *pmask the portion of a cpusets's mems_allowed that |
| 330 | * are online, with memory. If none are online with memory, walk |
| 331 | * up the cpuset hierarchy until we find one that does have some |
| 332 | * online mems. The top cpuset always has some mems online. |
| 333 | * |
| 334 | * One way or another, we guarantee to return some non-empty subset |
| 335 | * of node_states[N_MEMORY]. |
| 336 | * |
| 337 | * Call with callback_mutex held. |
| 338 | */ |
| 339 | static void guarantee_online_mems(struct cpuset *cs, nodemask_t *pmask) |
| 340 | { |
| 341 | while (!nodes_intersects(cs->mems_allowed, node_states[N_MEMORY])) |
| 342 | cs = parent_cs(cs); |
| 343 | nodes_and(*pmask, cs->mems_allowed, node_states[N_MEMORY]); |
| 344 | } |
| 345 | |
| 346 | /* |
| 347 | * update task's spread flag if cpuset's page/slab spread flag is set |
| 348 | * |
| 349 | * Called with callback_mutex/cpuset_mutex held |
| 350 | */ |
| 351 | static void cpuset_update_task_spread_flag(struct cpuset *cs, |
| 352 | struct task_struct *tsk) |
| 353 | { |
| 354 | if (is_spread_page(cs)) |
| 355 | tsk->flags |= PF_SPREAD_PAGE; |
| 356 | else |
| 357 | tsk->flags &= ~PF_SPREAD_PAGE; |
| 358 | if (is_spread_slab(cs)) |
| 359 | tsk->flags |= PF_SPREAD_SLAB; |
| 360 | else |
| 361 | tsk->flags &= ~PF_SPREAD_SLAB; |
| 362 | } |
| 363 | |
| 364 | /* |
| 365 | * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q? |
| 366 | * |
| 367 | * One cpuset is a subset of another if all its allowed CPUs and |
| 368 | * Memory Nodes are a subset of the other, and its exclusive flags |
| 369 | * are only set if the other's are set. Call holding cpuset_mutex. |
| 370 | */ |
| 371 | |
| 372 | static int is_cpuset_subset(const struct cpuset *p, const struct cpuset *q) |
| 373 | { |
| 374 | return cpumask_subset(p->cpus_allowed, q->cpus_allowed) && |
| 375 | nodes_subset(p->mems_allowed, q->mems_allowed) && |
| 376 | is_cpu_exclusive(p) <= is_cpu_exclusive(q) && |
| 377 | is_mem_exclusive(p) <= is_mem_exclusive(q); |
| 378 | } |
| 379 | |
| 380 | /** |
| 381 | * alloc_trial_cpuset - allocate a trial cpuset |
| 382 | * @cs: the cpuset that the trial cpuset duplicates |
| 383 | */ |
| 384 | static struct cpuset *alloc_trial_cpuset(struct cpuset *cs) |
| 385 | { |
| 386 | struct cpuset *trial; |
| 387 | |
| 388 | trial = kmemdup(cs, sizeof(*cs), GFP_KERNEL); |
| 389 | if (!trial) |
| 390 | return NULL; |
| 391 | |
| 392 | if (!alloc_cpumask_var(&trial->cpus_allowed, GFP_KERNEL)) { |
| 393 | kfree(trial); |
| 394 | return NULL; |
| 395 | } |
| 396 | cpumask_copy(trial->cpus_allowed, cs->cpus_allowed); |
| 397 | |
| 398 | return trial; |
| 399 | } |
| 400 | |
| 401 | /** |
| 402 | * free_trial_cpuset - free the trial cpuset |
| 403 | * @trial: the trial cpuset to be freed |
| 404 | */ |
| 405 | static void free_trial_cpuset(struct cpuset *trial) |
| 406 | { |
| 407 | free_cpumask_var(trial->cpus_allowed); |
| 408 | kfree(trial); |
| 409 | } |
| 410 | |
| 411 | /* |
| 412 | * validate_change() - Used to validate that any proposed cpuset change |
| 413 | * follows the structural rules for cpusets. |
| 414 | * |
| 415 | * If we replaced the flag and mask values of the current cpuset |
| 416 | * (cur) with those values in the trial cpuset (trial), would |
| 417 | * our various subset and exclusive rules still be valid? Presumes |
| 418 | * cpuset_mutex held. |
| 419 | * |
| 420 | * 'cur' is the address of an actual, in-use cpuset. Operations |
| 421 | * such as list traversal that depend on the actual address of the |
| 422 | * cpuset in the list must use cur below, not trial. |
| 423 | * |
| 424 | * 'trial' is the address of bulk structure copy of cur, with |
| 425 | * perhaps one or more of the fields cpus_allowed, mems_allowed, |
| 426 | * or flags changed to new, trial values. |
| 427 | * |
| 428 | * Return 0 if valid, -errno if not. |
| 429 | */ |
| 430 | |
| 431 | static int validate_change(struct cpuset *cur, struct cpuset *trial) |
| 432 | { |
| 433 | struct cgroup *cgrp; |
| 434 | struct cpuset *c, *par; |
| 435 | int ret; |
| 436 | |
| 437 | rcu_read_lock(); |
| 438 | |
| 439 | /* Each of our child cpusets must be a subset of us */ |
| 440 | ret = -EBUSY; |
| 441 | cpuset_for_each_child(c, cgrp, cur) |
| 442 | if (!is_cpuset_subset(c, trial)) |
| 443 | goto out; |
| 444 | |
| 445 | /* Remaining checks don't apply to root cpuset */ |
| 446 | ret = 0; |
| 447 | if (cur == &top_cpuset) |
| 448 | goto out; |
| 449 | |
| 450 | par = parent_cs(cur); |
| 451 | |
| 452 | /* We must be a subset of our parent cpuset */ |
| 453 | ret = -EACCES; |
| 454 | if (!is_cpuset_subset(trial, par)) |
| 455 | goto out; |
| 456 | |
| 457 | /* |
| 458 | * If either I or some sibling (!= me) is exclusive, we can't |
| 459 | * overlap |
| 460 | */ |
| 461 | ret = -EINVAL; |
| 462 | cpuset_for_each_child(c, cgrp, par) { |
| 463 | if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) && |
| 464 | c != cur && |
| 465 | cpumask_intersects(trial->cpus_allowed, c->cpus_allowed)) |
| 466 | goto out; |
| 467 | if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) && |
| 468 | c != cur && |
| 469 | nodes_intersects(trial->mems_allowed, c->mems_allowed)) |
| 470 | goto out; |
| 471 | } |
| 472 | |
| 473 | /* |
| 474 | * Cpusets with tasks - existing or newly being attached - can't |
| 475 | * have empty cpus_allowed or mems_allowed. |
| 476 | */ |
| 477 | ret = -ENOSPC; |
| 478 | if ((cgroup_task_count(cur->css.cgroup) || cur->attach_in_progress) && |
| 479 | (cpumask_empty(trial->cpus_allowed) && |
| 480 | nodes_empty(trial->mems_allowed))) |
| 481 | goto out; |
| 482 | |
| 483 | ret = 0; |
| 484 | out: |
| 485 | rcu_read_unlock(); |
| 486 | return ret; |
| 487 | } |
| 488 | |
| 489 | #ifdef CONFIG_SMP |
| 490 | /* |
| 491 | * Helper routine for generate_sched_domains(). |
| 492 | * Do cpusets a, b have overlapping cpus_allowed masks? |
| 493 | */ |
| 494 | static int cpusets_overlap(struct cpuset *a, struct cpuset *b) |
| 495 | { |
| 496 | return cpumask_intersects(a->cpus_allowed, b->cpus_allowed); |
| 497 | } |
| 498 | |
| 499 | static void |
| 500 | update_domain_attr(struct sched_domain_attr *dattr, struct cpuset *c) |
| 501 | { |
| 502 | if (dattr->relax_domain_level < c->relax_domain_level) |
| 503 | dattr->relax_domain_level = c->relax_domain_level; |
| 504 | return; |
| 505 | } |
| 506 | |
| 507 | static void update_domain_attr_tree(struct sched_domain_attr *dattr, |
| 508 | struct cpuset *root_cs) |
| 509 | { |
| 510 | struct cpuset *cp; |
| 511 | struct cgroup *pos_cgrp; |
| 512 | |
| 513 | rcu_read_lock(); |
| 514 | cpuset_for_each_descendant_pre(cp, pos_cgrp, root_cs) { |
| 515 | /* skip the whole subtree if @cp doesn't have any CPU */ |
| 516 | if (cpumask_empty(cp->cpus_allowed)) { |
| 517 | pos_cgrp = cgroup_rightmost_descendant(pos_cgrp); |
| 518 | continue; |
| 519 | } |
| 520 | |
| 521 | if (is_sched_load_balance(cp)) |
| 522 | update_domain_attr(dattr, cp); |
| 523 | } |
| 524 | rcu_read_unlock(); |
| 525 | } |
| 526 | |
| 527 | /* |
| 528 | * generate_sched_domains() |
| 529 | * |
| 530 | * This function builds a partial partition of the systems CPUs |
| 531 | * A 'partial partition' is a set of non-overlapping subsets whose |
| 532 | * union is a subset of that set. |
| 533 | * The output of this function needs to be passed to kernel/sched/core.c |
| 534 | * partition_sched_domains() routine, which will rebuild the scheduler's |
| 535 | * load balancing domains (sched domains) as specified by that partial |
| 536 | * partition. |
| 537 | * |
| 538 | * See "What is sched_load_balance" in Documentation/cgroups/cpusets.txt |
| 539 | * for a background explanation of this. |
| 540 | * |
| 541 | * Does not return errors, on the theory that the callers of this |
| 542 | * routine would rather not worry about failures to rebuild sched |
| 543 | * domains when operating in the severe memory shortage situations |
| 544 | * that could cause allocation failures below. |
| 545 | * |
| 546 | * Must be called with cpuset_mutex held. |
| 547 | * |
| 548 | * The three key local variables below are: |
| 549 | * q - a linked-list queue of cpuset pointers, used to implement a |
| 550 | * top-down scan of all cpusets. This scan loads a pointer |
| 551 | * to each cpuset marked is_sched_load_balance into the |
| 552 | * array 'csa'. For our purposes, rebuilding the schedulers |
| 553 | * sched domains, we can ignore !is_sched_load_balance cpusets. |
| 554 | * csa - (for CpuSet Array) Array of pointers to all the cpusets |
| 555 | * that need to be load balanced, for convenient iterative |
| 556 | * access by the subsequent code that finds the best partition, |
| 557 | * i.e the set of domains (subsets) of CPUs such that the |
| 558 | * cpus_allowed of every cpuset marked is_sched_load_balance |
| 559 | * is a subset of one of these domains, while there are as |
| 560 | * many such domains as possible, each as small as possible. |
| 561 | * doms - Conversion of 'csa' to an array of cpumasks, for passing to |
| 562 | * the kernel/sched/core.c routine partition_sched_domains() in a |
| 563 | * convenient format, that can be easily compared to the prior |
| 564 | * value to determine what partition elements (sched domains) |
| 565 | * were changed (added or removed.) |
| 566 | * |
| 567 | * Finding the best partition (set of domains): |
| 568 | * The triple nested loops below over i, j, k scan over the |
| 569 | * load balanced cpusets (using the array of cpuset pointers in |
| 570 | * csa[]) looking for pairs of cpusets that have overlapping |
| 571 | * cpus_allowed, but which don't have the same 'pn' partition |
| 572 | * number and gives them in the same partition number. It keeps |
| 573 | * looping on the 'restart' label until it can no longer find |
| 574 | * any such pairs. |
| 575 | * |
| 576 | * The union of the cpus_allowed masks from the set of |
| 577 | * all cpusets having the same 'pn' value then form the one |
| 578 | * element of the partition (one sched domain) to be passed to |
| 579 | * partition_sched_domains(). |
| 580 | */ |
| 581 | static int generate_sched_domains(cpumask_var_t **domains, |
| 582 | struct sched_domain_attr **attributes) |
| 583 | { |
| 584 | struct cpuset *cp; /* scans q */ |
| 585 | struct cpuset **csa; /* array of all cpuset ptrs */ |
| 586 | int csn; /* how many cpuset ptrs in csa so far */ |
| 587 | int i, j, k; /* indices for partition finding loops */ |
| 588 | cpumask_var_t *doms; /* resulting partition; i.e. sched domains */ |
| 589 | struct sched_domain_attr *dattr; /* attributes for custom domains */ |
| 590 | int ndoms = 0; /* number of sched domains in result */ |
| 591 | int nslot; /* next empty doms[] struct cpumask slot */ |
| 592 | struct cgroup *pos_cgrp; |
| 593 | |
| 594 | doms = NULL; |
| 595 | dattr = NULL; |
| 596 | csa = NULL; |
| 597 | |
| 598 | /* Special case for the 99% of systems with one, full, sched domain */ |
| 599 | if (is_sched_load_balance(&top_cpuset)) { |
| 600 | ndoms = 1; |
| 601 | doms = alloc_sched_domains(ndoms); |
| 602 | if (!doms) |
| 603 | goto done; |
| 604 | |
| 605 | dattr = kmalloc(sizeof(struct sched_domain_attr), GFP_KERNEL); |
| 606 | if (dattr) { |
| 607 | *dattr = SD_ATTR_INIT; |
| 608 | update_domain_attr_tree(dattr, &top_cpuset); |
| 609 | } |
| 610 | cpumask_copy(doms[0], top_cpuset.cpus_allowed); |
| 611 | |
| 612 | goto done; |
| 613 | } |
| 614 | |
| 615 | csa = kmalloc(number_of_cpusets * sizeof(cp), GFP_KERNEL); |
| 616 | if (!csa) |
| 617 | goto done; |
| 618 | csn = 0; |
| 619 | |
| 620 | rcu_read_lock(); |
| 621 | cpuset_for_each_descendant_pre(cp, pos_cgrp, &top_cpuset) { |
| 622 | /* |
| 623 | * Continue traversing beyond @cp iff @cp has some CPUs and |
| 624 | * isn't load balancing. The former is obvious. The |
| 625 | * latter: All child cpusets contain a subset of the |
| 626 | * parent's cpus, so just skip them, and then we call |
| 627 | * update_domain_attr_tree() to calc relax_domain_level of |
| 628 | * the corresponding sched domain. |
| 629 | */ |
| 630 | if (!cpumask_empty(cp->cpus_allowed) && |
| 631 | !is_sched_load_balance(cp)) |
| 632 | continue; |
| 633 | |
| 634 | if (is_sched_load_balance(cp)) |
| 635 | csa[csn++] = cp; |
| 636 | |
| 637 | /* skip @cp's subtree */ |
| 638 | pos_cgrp = cgroup_rightmost_descendant(pos_cgrp); |
| 639 | } |
| 640 | rcu_read_unlock(); |
| 641 | |
| 642 | for (i = 0; i < csn; i++) |
| 643 | csa[i]->pn = i; |
| 644 | ndoms = csn; |
| 645 | |
| 646 | restart: |
| 647 | /* Find the best partition (set of sched domains) */ |
| 648 | for (i = 0; i < csn; i++) { |
| 649 | struct cpuset *a = csa[i]; |
| 650 | int apn = a->pn; |
| 651 | |
| 652 | for (j = 0; j < csn; j++) { |
| 653 | struct cpuset *b = csa[j]; |
| 654 | int bpn = b->pn; |
| 655 | |
| 656 | if (apn != bpn && cpusets_overlap(a, b)) { |
| 657 | for (k = 0; k < csn; k++) { |
| 658 | struct cpuset *c = csa[k]; |
| 659 | |
| 660 | if (c->pn == bpn) |
| 661 | c->pn = apn; |
| 662 | } |
| 663 | ndoms--; /* one less element */ |
| 664 | goto restart; |
| 665 | } |
| 666 | } |
| 667 | } |
| 668 | |
| 669 | /* |
| 670 | * Now we know how many domains to create. |
| 671 | * Convert <csn, csa> to <ndoms, doms> and populate cpu masks. |
| 672 | */ |
| 673 | doms = alloc_sched_domains(ndoms); |
| 674 | if (!doms) |
| 675 | goto done; |
| 676 | |
| 677 | /* |
| 678 | * The rest of the code, including the scheduler, can deal with |
| 679 | * dattr==NULL case. No need to abort if alloc fails. |
| 680 | */ |
| 681 | dattr = kmalloc(ndoms * sizeof(struct sched_domain_attr), GFP_KERNEL); |
| 682 | |
| 683 | for (nslot = 0, i = 0; i < csn; i++) { |
| 684 | struct cpuset *a = csa[i]; |
| 685 | struct cpumask *dp; |
| 686 | int apn = a->pn; |
| 687 | |
| 688 | if (apn < 0) { |
| 689 | /* Skip completed partitions */ |
| 690 | continue; |
| 691 | } |
| 692 | |
| 693 | dp = doms[nslot]; |
| 694 | |
| 695 | if (nslot == ndoms) { |
| 696 | static int warnings = 10; |
| 697 | if (warnings) { |
| 698 | printk(KERN_WARNING |
| 699 | "rebuild_sched_domains confused:" |
| 700 | " nslot %d, ndoms %d, csn %d, i %d," |
| 701 | " apn %d\n", |
| 702 | nslot, ndoms, csn, i, apn); |
| 703 | warnings--; |
| 704 | } |
| 705 | continue; |
| 706 | } |
| 707 | |
| 708 | cpumask_clear(dp); |
| 709 | if (dattr) |
| 710 | *(dattr + nslot) = SD_ATTR_INIT; |
| 711 | for (j = i; j < csn; j++) { |
| 712 | struct cpuset *b = csa[j]; |
| 713 | |
| 714 | if (apn == b->pn) { |
| 715 | cpumask_or(dp, dp, b->cpus_allowed); |
| 716 | if (dattr) |
| 717 | update_domain_attr_tree(dattr + nslot, b); |
| 718 | |
| 719 | /* Done with this partition */ |
| 720 | b->pn = -1; |
| 721 | } |
| 722 | } |
| 723 | nslot++; |
| 724 | } |
| 725 | BUG_ON(nslot != ndoms); |
| 726 | |
| 727 | done: |
| 728 | kfree(csa); |
| 729 | |
| 730 | /* |
| 731 | * Fallback to the default domain if kmalloc() failed. |
| 732 | * See comments in partition_sched_domains(). |
| 733 | */ |
| 734 | if (doms == NULL) |
| 735 | ndoms = 1; |
| 736 | |
| 737 | *domains = doms; |
| 738 | *attributes = dattr; |
| 739 | return ndoms; |
| 740 | } |
| 741 | |
| 742 | /* |
| 743 | * Rebuild scheduler domains. |
| 744 | * |
| 745 | * If the flag 'sched_load_balance' of any cpuset with non-empty |
| 746 | * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset |
| 747 | * which has that flag enabled, or if any cpuset with a non-empty |
| 748 | * 'cpus' is removed, then call this routine to rebuild the |
| 749 | * scheduler's dynamic sched domains. |
| 750 | * |
| 751 | * Call with cpuset_mutex held. Takes get_online_cpus(). |
| 752 | */ |
| 753 | static void rebuild_sched_domains_locked(void) |
| 754 | { |
| 755 | struct sched_domain_attr *attr; |
| 756 | cpumask_var_t *doms; |
| 757 | int ndoms; |
| 758 | |
| 759 | lockdep_assert_held(&cpuset_mutex); |
| 760 | get_online_cpus(); |
| 761 | |
| 762 | /* |
| 763 | * We have raced with CPU hotplug. Don't do anything to avoid |
| 764 | * passing doms with offlined cpu to partition_sched_domains(). |
| 765 | * Anyways, hotplug work item will rebuild sched domains. |
| 766 | */ |
| 767 | if (!cpumask_equal(top_cpuset.cpus_allowed, cpu_active_mask)) |
| 768 | goto out; |
| 769 | |
| 770 | /* Generate domain masks and attrs */ |
| 771 | ndoms = generate_sched_domains(&doms, &attr); |
| 772 | |
| 773 | /* Have scheduler rebuild the domains */ |
| 774 | partition_sched_domains(ndoms, doms, attr); |
| 775 | out: |
| 776 | put_online_cpus(); |
| 777 | } |
| 778 | #else /* !CONFIG_SMP */ |
| 779 | static void rebuild_sched_domains_locked(void) |
| 780 | { |
| 781 | } |
| 782 | #endif /* CONFIG_SMP */ |
| 783 | |
| 784 | void rebuild_sched_domains(void) |
| 785 | { |
| 786 | mutex_lock(&cpuset_mutex); |
| 787 | rebuild_sched_domains_locked(); |
| 788 | mutex_unlock(&cpuset_mutex); |
| 789 | } |
| 790 | |
| 791 | /* |
| 792 | * effective_cpumask_cpuset - return nearest ancestor with non-empty cpus |
| 793 | * @cs: the cpuset in interest |
| 794 | * |
| 795 | * A cpuset's effective cpumask is the cpumask of the nearest ancestor |
| 796 | * with non-empty cpus. We use effective cpumask whenever: |
| 797 | * - we update tasks' cpus_allowed. (they take on the ancestor's cpumask |
| 798 | * if the cpuset they reside in has no cpus) |
| 799 | * - we want to retrieve task_cs(tsk)'s cpus_allowed. |
| 800 | * |
| 801 | * Called with cpuset_mutex held. cpuset_cpus_allowed_fallback() is an |
| 802 | * exception. See comments there. |
| 803 | */ |
| 804 | static struct cpuset *effective_cpumask_cpuset(struct cpuset *cs) |
| 805 | { |
| 806 | while (cpumask_empty(cs->cpus_allowed)) |
| 807 | cs = parent_cs(cs); |
| 808 | return cs; |
| 809 | } |
| 810 | |
| 811 | /* |
| 812 | * effective_nodemask_cpuset - return nearest ancestor with non-empty mems |
| 813 | * @cs: the cpuset in interest |
| 814 | * |
| 815 | * A cpuset's effective nodemask is the nodemask of the nearest ancestor |
| 816 | * with non-empty memss. We use effective nodemask whenever: |
| 817 | * - we update tasks' mems_allowed. (they take on the ancestor's nodemask |
| 818 | * if the cpuset they reside in has no mems) |
| 819 | * - we want to retrieve task_cs(tsk)'s mems_allowed. |
| 820 | * |
| 821 | * Called with cpuset_mutex held. |
| 822 | */ |
| 823 | static struct cpuset *effective_nodemask_cpuset(struct cpuset *cs) |
| 824 | { |
| 825 | while (nodes_empty(cs->mems_allowed)) |
| 826 | cs = parent_cs(cs); |
| 827 | return cs; |
| 828 | } |
| 829 | |
| 830 | /** |
| 831 | * cpuset_change_cpumask - make a task's cpus_allowed the same as its cpuset's |
| 832 | * @tsk: task to test |
| 833 | * @scan: struct cgroup_scanner containing the cgroup of the task |
| 834 | * |
| 835 | * Called by cgroup_scan_tasks() for each task in a cgroup whose |
| 836 | * cpus_allowed mask needs to be changed. |
| 837 | * |
| 838 | * We don't need to re-check for the cgroup/cpuset membership, since we're |
| 839 | * holding cpuset_mutex at this point. |
| 840 | */ |
| 841 | static void cpuset_change_cpumask(struct task_struct *tsk, |
| 842 | struct cgroup_scanner *scan) |
| 843 | { |
| 844 | struct cpuset *cpus_cs; |
| 845 | |
| 846 | cpus_cs = effective_cpumask_cpuset(cgroup_cs(scan->cgrp)); |
| 847 | set_cpus_allowed_ptr(tsk, cpus_cs->cpus_allowed); |
| 848 | } |
| 849 | |
| 850 | /** |
| 851 | * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset. |
| 852 | * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed |
| 853 | * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks() |
| 854 | * |
| 855 | * Called with cpuset_mutex held |
| 856 | * |
| 857 | * The cgroup_scan_tasks() function will scan all the tasks in a cgroup, |
| 858 | * calling callback functions for each. |
| 859 | * |
| 860 | * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0 |
| 861 | * if @heap != NULL. |
| 862 | */ |
| 863 | static void update_tasks_cpumask(struct cpuset *cs, struct ptr_heap *heap) |
| 864 | { |
| 865 | struct cgroup_scanner scan; |
| 866 | |
| 867 | scan.cgrp = cs->css.cgroup; |
| 868 | scan.test_task = NULL; |
| 869 | scan.process_task = cpuset_change_cpumask; |
| 870 | scan.heap = heap; |
| 871 | cgroup_scan_tasks(&scan); |
| 872 | } |
| 873 | |
| 874 | /* |
| 875 | * update_tasks_cpumask_hier - Update the cpumasks of tasks in the hierarchy. |
| 876 | * @root_cs: the root cpuset of the hierarchy |
| 877 | * @update_root: update root cpuset or not? |
| 878 | * @heap: the heap used by cgroup_scan_tasks() |
| 879 | * |
| 880 | * This will update cpumasks of tasks in @root_cs and all other empty cpusets |
| 881 | * which take on cpumask of @root_cs. |
| 882 | * |
| 883 | * Called with cpuset_mutex held |
| 884 | */ |
| 885 | static void update_tasks_cpumask_hier(struct cpuset *root_cs, |
| 886 | bool update_root, struct ptr_heap *heap) |
| 887 | { |
| 888 | struct cpuset *cp; |
| 889 | struct cgroup *pos_cgrp; |
| 890 | |
| 891 | if (update_root) |
| 892 | update_tasks_cpumask(root_cs, heap); |
| 893 | |
| 894 | rcu_read_lock(); |
| 895 | cpuset_for_each_descendant_pre(cp, pos_cgrp, root_cs) { |
| 896 | /* skip the whole subtree if @cp have some CPU */ |
| 897 | if (!cpumask_empty(cp->cpus_allowed)) { |
| 898 | pos_cgrp = cgroup_rightmost_descendant(pos_cgrp); |
| 899 | continue; |
| 900 | } |
| 901 | if (!css_tryget(&cp->css)) |
| 902 | continue; |
| 903 | rcu_read_unlock(); |
| 904 | |
| 905 | update_tasks_cpumask(cp, heap); |
| 906 | |
| 907 | rcu_read_lock(); |
| 908 | css_put(&cp->css); |
| 909 | } |
| 910 | rcu_read_unlock(); |
| 911 | } |
| 912 | |
| 913 | /** |
| 914 | * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it |
| 915 | * @cs: the cpuset to consider |
| 916 | * @buf: buffer of cpu numbers written to this cpuset |
| 917 | */ |
| 918 | static int update_cpumask(struct cpuset *cs, struct cpuset *trialcs, |
| 919 | const char *buf) |
| 920 | { |
| 921 | struct ptr_heap heap; |
| 922 | int retval; |
| 923 | int is_load_balanced; |
| 924 | |
| 925 | /* top_cpuset.cpus_allowed tracks cpu_online_mask; it's read-only */ |
| 926 | if (cs == &top_cpuset) |
| 927 | return -EACCES; |
| 928 | |
| 929 | /* |
| 930 | * An empty cpus_allowed is ok only if the cpuset has no tasks. |
| 931 | * Since cpulist_parse() fails on an empty mask, we special case |
| 932 | * that parsing. The validate_change() call ensures that cpusets |
| 933 | * with tasks have cpus. |
| 934 | */ |
| 935 | if (!*buf) { |
| 936 | cpumask_clear(trialcs->cpus_allowed); |
| 937 | } else { |
| 938 | retval = cpulist_parse(buf, trialcs->cpus_allowed); |
| 939 | if (retval < 0) |
| 940 | return retval; |
| 941 | |
| 942 | if (!cpumask_subset(trialcs->cpus_allowed, cpu_active_mask)) |
| 943 | return -EINVAL; |
| 944 | } |
| 945 | |
| 946 | /* Nothing to do if the cpus didn't change */ |
| 947 | if (cpumask_equal(cs->cpus_allowed, trialcs->cpus_allowed)) |
| 948 | return 0; |
| 949 | |
| 950 | retval = validate_change(cs, trialcs); |
| 951 | if (retval < 0) |
| 952 | return retval; |
| 953 | |
| 954 | retval = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL); |
| 955 | if (retval) |
| 956 | return retval; |
| 957 | |
| 958 | is_load_balanced = is_sched_load_balance(trialcs); |
| 959 | |
| 960 | mutex_lock(&callback_mutex); |
| 961 | cpumask_copy(cs->cpus_allowed, trialcs->cpus_allowed); |
| 962 | mutex_unlock(&callback_mutex); |
| 963 | |
| 964 | update_tasks_cpumask_hier(cs, true, &heap); |
| 965 | |
| 966 | heap_free(&heap); |
| 967 | |
| 968 | if (is_load_balanced) |
| 969 | rebuild_sched_domains_locked(); |
| 970 | return 0; |
| 971 | } |
| 972 | |
| 973 | /* |
| 974 | * cpuset_migrate_mm |
| 975 | * |
| 976 | * Migrate memory region from one set of nodes to another. |
| 977 | * |
| 978 | * Temporarilly set tasks mems_allowed to target nodes of migration, |
| 979 | * so that the migration code can allocate pages on these nodes. |
| 980 | * |
| 981 | * Call holding cpuset_mutex, so current's cpuset won't change |
| 982 | * during this call, as manage_mutex holds off any cpuset_attach() |
| 983 | * calls. Therefore we don't need to take task_lock around the |
| 984 | * call to guarantee_online_mems(), as we know no one is changing |
| 985 | * our task's cpuset. |
| 986 | * |
| 987 | * While the mm_struct we are migrating is typically from some |
| 988 | * other task, the task_struct mems_allowed that we are hacking |
| 989 | * is for our current task, which must allocate new pages for that |
| 990 | * migrating memory region. |
| 991 | */ |
| 992 | |
| 993 | static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from, |
| 994 | const nodemask_t *to) |
| 995 | { |
| 996 | struct task_struct *tsk = current; |
| 997 | struct cpuset *mems_cs; |
| 998 | |
| 999 | tsk->mems_allowed = *to; |
| 1000 | |
| 1001 | do_migrate_pages(mm, from, to, MPOL_MF_MOVE_ALL); |
| 1002 | |
| 1003 | mems_cs = effective_nodemask_cpuset(task_cs(tsk)); |
| 1004 | guarantee_online_mems(mems_cs, &tsk->mems_allowed); |
| 1005 | } |
| 1006 | |
| 1007 | /* |
| 1008 | * cpuset_change_task_nodemask - change task's mems_allowed and mempolicy |
| 1009 | * @tsk: the task to change |
| 1010 | * @newmems: new nodes that the task will be set |
| 1011 | * |
| 1012 | * In order to avoid seeing no nodes if the old and new nodes are disjoint, |
| 1013 | * we structure updates as setting all new allowed nodes, then clearing newly |
| 1014 | * disallowed ones. |
| 1015 | */ |
| 1016 | static void cpuset_change_task_nodemask(struct task_struct *tsk, |
| 1017 | nodemask_t *newmems) |
| 1018 | { |
| 1019 | bool need_loop; |
| 1020 | |
| 1021 | /* |
| 1022 | * Allow tasks that have access to memory reserves because they have |
| 1023 | * been OOM killed to get memory anywhere. |
| 1024 | */ |
| 1025 | if (unlikely(test_thread_flag(TIF_MEMDIE))) |
| 1026 | return; |
| 1027 | if (current->flags & PF_EXITING) /* Let dying task have memory */ |
| 1028 | return; |
| 1029 | |
| 1030 | task_lock(tsk); |
| 1031 | /* |
| 1032 | * Determine if a loop is necessary if another thread is doing |
| 1033 | * get_mems_allowed(). If at least one node remains unchanged and |
| 1034 | * tsk does not have a mempolicy, then an empty nodemask will not be |
| 1035 | * possible when mems_allowed is larger than a word. |
| 1036 | */ |
| 1037 | need_loop = task_has_mempolicy(tsk) || |
| 1038 | !nodes_intersects(*newmems, tsk->mems_allowed); |
| 1039 | |
| 1040 | if (need_loop) |
| 1041 | write_seqcount_begin(&tsk->mems_allowed_seq); |
| 1042 | |
| 1043 | nodes_or(tsk->mems_allowed, tsk->mems_allowed, *newmems); |
| 1044 | mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP1); |
| 1045 | |
| 1046 | mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP2); |
| 1047 | tsk->mems_allowed = *newmems; |
| 1048 | |
| 1049 | if (need_loop) |
| 1050 | write_seqcount_end(&tsk->mems_allowed_seq); |
| 1051 | |
| 1052 | task_unlock(tsk); |
| 1053 | } |
| 1054 | |
| 1055 | /* |
| 1056 | * Update task's mems_allowed and rebind its mempolicy and vmas' mempolicy |
| 1057 | * of it to cpuset's new mems_allowed, and migrate pages to new nodes if |
| 1058 | * memory_migrate flag is set. Called with cpuset_mutex held. |
| 1059 | */ |
| 1060 | static void cpuset_change_nodemask(struct task_struct *p, |
| 1061 | struct cgroup_scanner *scan) |
| 1062 | { |
| 1063 | struct cpuset *cs = cgroup_cs(scan->cgrp); |
| 1064 | struct mm_struct *mm; |
| 1065 | int migrate; |
| 1066 | nodemask_t *newmems = scan->data; |
| 1067 | |
| 1068 | cpuset_change_task_nodemask(p, newmems); |
| 1069 | |
| 1070 | mm = get_task_mm(p); |
| 1071 | if (!mm) |
| 1072 | return; |
| 1073 | |
| 1074 | migrate = is_memory_migrate(cs); |
| 1075 | |
| 1076 | mpol_rebind_mm(mm, &cs->mems_allowed); |
| 1077 | if (migrate) |
| 1078 | cpuset_migrate_mm(mm, &cs->old_mems_allowed, newmems); |
| 1079 | mmput(mm); |
| 1080 | } |
| 1081 | |
| 1082 | static void *cpuset_being_rebound; |
| 1083 | |
| 1084 | /** |
| 1085 | * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset. |
| 1086 | * @cs: the cpuset in which each task's mems_allowed mask needs to be changed |
| 1087 | * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks() |
| 1088 | * |
| 1089 | * Called with cpuset_mutex held |
| 1090 | * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0 |
| 1091 | * if @heap != NULL. |
| 1092 | */ |
| 1093 | static void update_tasks_nodemask(struct cpuset *cs, struct ptr_heap *heap) |
| 1094 | { |
| 1095 | static nodemask_t newmems; /* protected by cpuset_mutex */ |
| 1096 | struct cgroup_scanner scan; |
| 1097 | struct cpuset *mems_cs = effective_nodemask_cpuset(cs); |
| 1098 | |
| 1099 | cpuset_being_rebound = cs; /* causes mpol_dup() rebind */ |
| 1100 | |
| 1101 | guarantee_online_mems(mems_cs, &newmems); |
| 1102 | |
| 1103 | scan.cgrp = cs->css.cgroup; |
| 1104 | scan.test_task = NULL; |
| 1105 | scan.process_task = cpuset_change_nodemask; |
| 1106 | scan.heap = heap; |
| 1107 | scan.data = &newmems; |
| 1108 | |
| 1109 | /* |
| 1110 | * The mpol_rebind_mm() call takes mmap_sem, which we couldn't |
| 1111 | * take while holding tasklist_lock. Forks can happen - the |
| 1112 | * mpol_dup() cpuset_being_rebound check will catch such forks, |
| 1113 | * and rebind their vma mempolicies too. Because we still hold |
| 1114 | * the global cpuset_mutex, we know that no other rebind effort |
| 1115 | * will be contending for the global variable cpuset_being_rebound. |
| 1116 | * It's ok if we rebind the same mm twice; mpol_rebind_mm() |
| 1117 | * is idempotent. Also migrate pages in each mm to new nodes. |
| 1118 | */ |
| 1119 | cgroup_scan_tasks(&scan); |
| 1120 | |
| 1121 | /* |
| 1122 | * All the tasks' nodemasks have been updated, update |
| 1123 | * cs->old_mems_allowed. |
| 1124 | */ |
| 1125 | cs->old_mems_allowed = newmems; |
| 1126 | |
| 1127 | /* We're done rebinding vmas to this cpuset's new mems_allowed. */ |
| 1128 | cpuset_being_rebound = NULL; |
| 1129 | } |
| 1130 | |
| 1131 | /* |
| 1132 | * update_tasks_nodemask_hier - Update the nodemasks of tasks in the hierarchy. |
| 1133 | * @cs: the root cpuset of the hierarchy |
| 1134 | * @update_root: update the root cpuset or not? |
| 1135 | * @heap: the heap used by cgroup_scan_tasks() |
| 1136 | * |
| 1137 | * This will update nodemasks of tasks in @root_cs and all other empty cpusets |
| 1138 | * which take on nodemask of @root_cs. |
| 1139 | * |
| 1140 | * Called with cpuset_mutex held |
| 1141 | */ |
| 1142 | static void update_tasks_nodemask_hier(struct cpuset *root_cs, |
| 1143 | bool update_root, struct ptr_heap *heap) |
| 1144 | { |
| 1145 | struct cpuset *cp; |
| 1146 | struct cgroup *pos_cgrp; |
| 1147 | |
| 1148 | if (update_root) |
| 1149 | update_tasks_nodemask(root_cs, heap); |
| 1150 | |
| 1151 | rcu_read_lock(); |
| 1152 | cpuset_for_each_descendant_pre(cp, pos_cgrp, root_cs) { |
| 1153 | /* skip the whole subtree if @cp have some CPU */ |
| 1154 | if (!nodes_empty(cp->mems_allowed)) { |
| 1155 | pos_cgrp = cgroup_rightmost_descendant(pos_cgrp); |
| 1156 | continue; |
| 1157 | } |
| 1158 | if (!css_tryget(&cp->css)) |
| 1159 | continue; |
| 1160 | rcu_read_unlock(); |
| 1161 | |
| 1162 | update_tasks_nodemask(cp, heap); |
| 1163 | |
| 1164 | rcu_read_lock(); |
| 1165 | css_put(&cp->css); |
| 1166 | } |
| 1167 | rcu_read_unlock(); |
| 1168 | } |
| 1169 | |
| 1170 | /* |
| 1171 | * Handle user request to change the 'mems' memory placement |
| 1172 | * of a cpuset. Needs to validate the request, update the |
| 1173 | * cpusets mems_allowed, and for each task in the cpuset, |
| 1174 | * update mems_allowed and rebind task's mempolicy and any vma |
| 1175 | * mempolicies and if the cpuset is marked 'memory_migrate', |
| 1176 | * migrate the tasks pages to the new memory. |
| 1177 | * |
| 1178 | * Call with cpuset_mutex held. May take callback_mutex during call. |
| 1179 | * Will take tasklist_lock, scan tasklist for tasks in cpuset cs, |
| 1180 | * lock each such tasks mm->mmap_sem, scan its vma's and rebind |
| 1181 | * their mempolicies to the cpusets new mems_allowed. |
| 1182 | */ |
| 1183 | static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs, |
| 1184 | const char *buf) |
| 1185 | { |
| 1186 | int retval; |
| 1187 | struct ptr_heap heap; |
| 1188 | |
| 1189 | /* |
| 1190 | * top_cpuset.mems_allowed tracks node_stats[N_MEMORY]; |
| 1191 | * it's read-only |
| 1192 | */ |
| 1193 | if (cs == &top_cpuset) { |
| 1194 | retval = -EACCES; |
| 1195 | goto done; |
| 1196 | } |
| 1197 | |
| 1198 | /* |
| 1199 | * An empty mems_allowed is ok iff there are no tasks in the cpuset. |
| 1200 | * Since nodelist_parse() fails on an empty mask, we special case |
| 1201 | * that parsing. The validate_change() call ensures that cpusets |
| 1202 | * with tasks have memory. |
| 1203 | */ |
| 1204 | if (!*buf) { |
| 1205 | nodes_clear(trialcs->mems_allowed); |
| 1206 | } else { |
| 1207 | retval = nodelist_parse(buf, trialcs->mems_allowed); |
| 1208 | if (retval < 0) |
| 1209 | goto done; |
| 1210 | |
| 1211 | if (!nodes_subset(trialcs->mems_allowed, |
| 1212 | node_states[N_MEMORY])) { |
| 1213 | retval = -EINVAL; |
| 1214 | goto done; |
| 1215 | } |
| 1216 | } |
| 1217 | |
| 1218 | if (nodes_equal(cs->mems_allowed, trialcs->mems_allowed)) { |
| 1219 | retval = 0; /* Too easy - nothing to do */ |
| 1220 | goto done; |
| 1221 | } |
| 1222 | retval = validate_change(cs, trialcs); |
| 1223 | if (retval < 0) |
| 1224 | goto done; |
| 1225 | |
| 1226 | retval = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL); |
| 1227 | if (retval < 0) |
| 1228 | goto done; |
| 1229 | |
| 1230 | mutex_lock(&callback_mutex); |
| 1231 | cs->mems_allowed = trialcs->mems_allowed; |
| 1232 | mutex_unlock(&callback_mutex); |
| 1233 | |
| 1234 | update_tasks_nodemask_hier(cs, true, &heap); |
| 1235 | |
| 1236 | heap_free(&heap); |
| 1237 | done: |
| 1238 | return retval; |
| 1239 | } |
| 1240 | |
| 1241 | int current_cpuset_is_being_rebound(void) |
| 1242 | { |
| 1243 | return task_cs(current) == cpuset_being_rebound; |
| 1244 | } |
| 1245 | |
| 1246 | static int update_relax_domain_level(struct cpuset *cs, s64 val) |
| 1247 | { |
| 1248 | #ifdef CONFIG_SMP |
| 1249 | if (val < -1 || val >= sched_domain_level_max) |
| 1250 | return -EINVAL; |
| 1251 | #endif |
| 1252 | |
| 1253 | if (val != cs->relax_domain_level) { |
| 1254 | cs->relax_domain_level = val; |
| 1255 | if (!cpumask_empty(cs->cpus_allowed) && |
| 1256 | is_sched_load_balance(cs)) |
| 1257 | rebuild_sched_domains_locked(); |
| 1258 | } |
| 1259 | |
| 1260 | return 0; |
| 1261 | } |
| 1262 | |
| 1263 | /* |
| 1264 | * cpuset_change_flag - make a task's spread flags the same as its cpuset's |
| 1265 | * @tsk: task to be updated |
| 1266 | * @scan: struct cgroup_scanner containing the cgroup of the task |
| 1267 | * |
| 1268 | * Called by cgroup_scan_tasks() for each task in a cgroup. |
| 1269 | * |
| 1270 | * We don't need to re-check for the cgroup/cpuset membership, since we're |
| 1271 | * holding cpuset_mutex at this point. |
| 1272 | */ |
| 1273 | static void cpuset_change_flag(struct task_struct *tsk, |
| 1274 | struct cgroup_scanner *scan) |
| 1275 | { |
| 1276 | cpuset_update_task_spread_flag(cgroup_cs(scan->cgrp), tsk); |
| 1277 | } |
| 1278 | |
| 1279 | /* |
| 1280 | * update_tasks_flags - update the spread flags of tasks in the cpuset. |
| 1281 | * @cs: the cpuset in which each task's spread flags needs to be changed |
| 1282 | * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks() |
| 1283 | * |
| 1284 | * Called with cpuset_mutex held |
| 1285 | * |
| 1286 | * The cgroup_scan_tasks() function will scan all the tasks in a cgroup, |
| 1287 | * calling callback functions for each. |
| 1288 | * |
| 1289 | * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0 |
| 1290 | * if @heap != NULL. |
| 1291 | */ |
| 1292 | static void update_tasks_flags(struct cpuset *cs, struct ptr_heap *heap) |
| 1293 | { |
| 1294 | struct cgroup_scanner scan; |
| 1295 | |
| 1296 | scan.cgrp = cs->css.cgroup; |
| 1297 | scan.test_task = NULL; |
| 1298 | scan.process_task = cpuset_change_flag; |
| 1299 | scan.heap = heap; |
| 1300 | cgroup_scan_tasks(&scan); |
| 1301 | } |
| 1302 | |
| 1303 | /* |
| 1304 | * update_flag - read a 0 or a 1 in a file and update associated flag |
| 1305 | * bit: the bit to update (see cpuset_flagbits_t) |
| 1306 | * cs: the cpuset to update |
| 1307 | * turning_on: whether the flag is being set or cleared |
| 1308 | * |
| 1309 | * Call with cpuset_mutex held. |
| 1310 | */ |
| 1311 | |
| 1312 | static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs, |
| 1313 | int turning_on) |
| 1314 | { |
| 1315 | struct cpuset *trialcs; |
| 1316 | int balance_flag_changed; |
| 1317 | int spread_flag_changed; |
| 1318 | struct ptr_heap heap; |
| 1319 | int err; |
| 1320 | |
| 1321 | trialcs = alloc_trial_cpuset(cs); |
| 1322 | if (!trialcs) |
| 1323 | return -ENOMEM; |
| 1324 | |
| 1325 | if (turning_on) |
| 1326 | set_bit(bit, &trialcs->flags); |
| 1327 | else |
| 1328 | clear_bit(bit, &trialcs->flags); |
| 1329 | |
| 1330 | err = validate_change(cs, trialcs); |
| 1331 | if (err < 0) |
| 1332 | goto out; |
| 1333 | |
| 1334 | err = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL); |
| 1335 | if (err < 0) |
| 1336 | goto out; |
| 1337 | |
| 1338 | balance_flag_changed = (is_sched_load_balance(cs) != |
| 1339 | is_sched_load_balance(trialcs)); |
| 1340 | |
| 1341 | spread_flag_changed = ((is_spread_slab(cs) != is_spread_slab(trialcs)) |
| 1342 | || (is_spread_page(cs) != is_spread_page(trialcs))); |
| 1343 | |
| 1344 | mutex_lock(&callback_mutex); |
| 1345 | cs->flags = trialcs->flags; |
| 1346 | mutex_unlock(&callback_mutex); |
| 1347 | |
| 1348 | if (!cpumask_empty(trialcs->cpus_allowed) && balance_flag_changed) |
| 1349 | rebuild_sched_domains_locked(); |
| 1350 | |
| 1351 | if (spread_flag_changed) |
| 1352 | update_tasks_flags(cs, &heap); |
| 1353 | heap_free(&heap); |
| 1354 | out: |
| 1355 | free_trial_cpuset(trialcs); |
| 1356 | return err; |
| 1357 | } |
| 1358 | |
| 1359 | /* |
| 1360 | * Frequency meter - How fast is some event occurring? |
| 1361 | * |
| 1362 | * These routines manage a digitally filtered, constant time based, |
| 1363 | * event frequency meter. There are four routines: |
| 1364 | * fmeter_init() - initialize a frequency meter. |
| 1365 | * fmeter_markevent() - called each time the event happens. |
| 1366 | * fmeter_getrate() - returns the recent rate of such events. |
| 1367 | * fmeter_update() - internal routine used to update fmeter. |
| 1368 | * |
| 1369 | * A common data structure is passed to each of these routines, |
| 1370 | * which is used to keep track of the state required to manage the |
| 1371 | * frequency meter and its digital filter. |
| 1372 | * |
| 1373 | * The filter works on the number of events marked per unit time. |
| 1374 | * The filter is single-pole low-pass recursive (IIR). The time unit |
| 1375 | * is 1 second. Arithmetic is done using 32-bit integers scaled to |
| 1376 | * simulate 3 decimal digits of precision (multiplied by 1000). |
| 1377 | * |
| 1378 | * With an FM_COEF of 933, and a time base of 1 second, the filter |
| 1379 | * has a half-life of 10 seconds, meaning that if the events quit |
| 1380 | * happening, then the rate returned from the fmeter_getrate() |
| 1381 | * will be cut in half each 10 seconds, until it converges to zero. |
| 1382 | * |
| 1383 | * It is not worth doing a real infinitely recursive filter. If more |
| 1384 | * than FM_MAXTICKS ticks have elapsed since the last filter event, |
| 1385 | * just compute FM_MAXTICKS ticks worth, by which point the level |
| 1386 | * will be stable. |
| 1387 | * |
| 1388 | * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid |
| 1389 | * arithmetic overflow in the fmeter_update() routine. |
| 1390 | * |
| 1391 | * Given the simple 32 bit integer arithmetic used, this meter works |
| 1392 | * best for reporting rates between one per millisecond (msec) and |
| 1393 | * one per 32 (approx) seconds. At constant rates faster than one |
| 1394 | * per msec it maxes out at values just under 1,000,000. At constant |
| 1395 | * rates between one per msec, and one per second it will stabilize |
| 1396 | * to a value N*1000, where N is the rate of events per second. |
| 1397 | * At constant rates between one per second and one per 32 seconds, |
| 1398 | * it will be choppy, moving up on the seconds that have an event, |
| 1399 | * and then decaying until the next event. At rates slower than |
| 1400 | * about one in 32 seconds, it decays all the way back to zero between |
| 1401 | * each event. |
| 1402 | */ |
| 1403 | |
| 1404 | #define FM_COEF 933 /* coefficient for half-life of 10 secs */ |
| 1405 | #define FM_MAXTICKS ((time_t)99) /* useless computing more ticks than this */ |
| 1406 | #define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */ |
| 1407 | #define FM_SCALE 1000 /* faux fixed point scale */ |
| 1408 | |
| 1409 | /* Initialize a frequency meter */ |
| 1410 | static void fmeter_init(struct fmeter *fmp) |
| 1411 | { |
| 1412 | fmp->cnt = 0; |
| 1413 | fmp->val = 0; |
| 1414 | fmp->time = 0; |
| 1415 | spin_lock_init(&fmp->lock); |
| 1416 | } |
| 1417 | |
| 1418 | /* Internal meter update - process cnt events and update value */ |
| 1419 | static void fmeter_update(struct fmeter *fmp) |
| 1420 | { |
| 1421 | time_t now = get_seconds(); |
| 1422 | time_t ticks = now - fmp->time; |
| 1423 | |
| 1424 | if (ticks == 0) |
| 1425 | return; |
| 1426 | |
| 1427 | ticks = min(FM_MAXTICKS, ticks); |
| 1428 | while (ticks-- > 0) |
| 1429 | fmp->val = (FM_COEF * fmp->val) / FM_SCALE; |
| 1430 | fmp->time = now; |
| 1431 | |
| 1432 | fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE; |
| 1433 | fmp->cnt = 0; |
| 1434 | } |
| 1435 | |
| 1436 | /* Process any previous ticks, then bump cnt by one (times scale). */ |
| 1437 | static void fmeter_markevent(struct fmeter *fmp) |
| 1438 | { |
| 1439 | spin_lock(&fmp->lock); |
| 1440 | fmeter_update(fmp); |
| 1441 | fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE); |
| 1442 | spin_unlock(&fmp->lock); |
| 1443 | } |
| 1444 | |
| 1445 | /* Process any previous ticks, then return current value. */ |
| 1446 | static int fmeter_getrate(struct fmeter *fmp) |
| 1447 | { |
| 1448 | int val; |
| 1449 | |
| 1450 | spin_lock(&fmp->lock); |
| 1451 | fmeter_update(fmp); |
| 1452 | val = fmp->val; |
| 1453 | spin_unlock(&fmp->lock); |
| 1454 | return val; |
| 1455 | } |
| 1456 | |
| 1457 | /* Called by cgroups to determine if a cpuset is usable; cpuset_mutex held */ |
| 1458 | static int cpuset_can_attach(struct cgroup *cgrp, struct cgroup_taskset *tset) |
| 1459 | { |
| 1460 | struct cpuset *cs = cgroup_cs(cgrp); |
| 1461 | struct task_struct *task; |
| 1462 | int ret; |
| 1463 | |
| 1464 | mutex_lock(&cpuset_mutex); |
| 1465 | |
| 1466 | /* |
| 1467 | * We allow to move tasks into an empty cpuset if sane_behavior |
| 1468 | * flag is set. |
| 1469 | */ |
| 1470 | ret = -ENOSPC; |
| 1471 | if (!cgroup_sane_behavior(cgrp) && |
| 1472 | (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed))) |
| 1473 | goto out_unlock; |
| 1474 | |
| 1475 | cgroup_taskset_for_each(task, cgrp, tset) { |
| 1476 | /* |
| 1477 | * Kthreads which disallow setaffinity shouldn't be moved |
| 1478 | * to a new cpuset; we don't want to change their cpu |
| 1479 | * affinity and isolating such threads by their set of |
| 1480 | * allowed nodes is unnecessary. Thus, cpusets are not |
| 1481 | * applicable for such threads. This prevents checking for |
| 1482 | * success of set_cpus_allowed_ptr() on all attached tasks |
| 1483 | * before cpus_allowed may be changed. |
| 1484 | */ |
| 1485 | ret = -EINVAL; |
| 1486 | if (task->flags & PF_NO_SETAFFINITY) |
| 1487 | goto out_unlock; |
| 1488 | ret = security_task_setscheduler(task); |
| 1489 | if (ret) |
| 1490 | goto out_unlock; |
| 1491 | } |
| 1492 | |
| 1493 | /* |
| 1494 | * Mark attach is in progress. This makes validate_change() fail |
| 1495 | * changes which zero cpus/mems_allowed. |
| 1496 | */ |
| 1497 | cs->attach_in_progress++; |
| 1498 | ret = 0; |
| 1499 | out_unlock: |
| 1500 | mutex_unlock(&cpuset_mutex); |
| 1501 | return ret; |
| 1502 | } |
| 1503 | |
| 1504 | static void cpuset_cancel_attach(struct cgroup *cgrp, |
| 1505 | struct cgroup_taskset *tset) |
| 1506 | { |
| 1507 | mutex_lock(&cpuset_mutex); |
| 1508 | cgroup_cs(cgrp)->attach_in_progress--; |
| 1509 | mutex_unlock(&cpuset_mutex); |
| 1510 | } |
| 1511 | |
| 1512 | /* |
| 1513 | * Protected by cpuset_mutex. cpus_attach is used only by cpuset_attach() |
| 1514 | * but we can't allocate it dynamically there. Define it global and |
| 1515 | * allocate from cpuset_init(). |
| 1516 | */ |
| 1517 | static cpumask_var_t cpus_attach; |
| 1518 | |
| 1519 | static void cpuset_attach(struct cgroup *cgrp, struct cgroup_taskset *tset) |
| 1520 | { |
| 1521 | /* static buf protected by cpuset_mutex */ |
| 1522 | static nodemask_t cpuset_attach_nodemask_to; |
| 1523 | struct mm_struct *mm; |
| 1524 | struct task_struct *task; |
| 1525 | struct task_struct *leader = cgroup_taskset_first(tset); |
| 1526 | struct cgroup *oldcgrp = cgroup_taskset_cur_cgroup(tset); |
| 1527 | struct cpuset *cs = cgroup_cs(cgrp); |
| 1528 | struct cpuset *oldcs = cgroup_cs(oldcgrp); |
| 1529 | struct cpuset *cpus_cs = effective_cpumask_cpuset(cs); |
| 1530 | struct cpuset *mems_cs = effective_nodemask_cpuset(cs); |
| 1531 | |
| 1532 | mutex_lock(&cpuset_mutex); |
| 1533 | |
| 1534 | /* prepare for attach */ |
| 1535 | if (cs == &top_cpuset) |
| 1536 | cpumask_copy(cpus_attach, cpu_possible_mask); |
| 1537 | else |
| 1538 | guarantee_online_cpus(cpus_cs, cpus_attach); |
| 1539 | |
| 1540 | guarantee_online_mems(mems_cs, &cpuset_attach_nodemask_to); |
| 1541 | |
| 1542 | cgroup_taskset_for_each(task, cgrp, tset) { |
| 1543 | /* |
| 1544 | * can_attach beforehand should guarantee that this doesn't |
| 1545 | * fail. TODO: have a better way to handle failure here |
| 1546 | */ |
| 1547 | WARN_ON_ONCE(set_cpus_allowed_ptr(task, cpus_attach)); |
| 1548 | |
| 1549 | cpuset_change_task_nodemask(task, &cpuset_attach_nodemask_to); |
| 1550 | cpuset_update_task_spread_flag(cs, task); |
| 1551 | } |
| 1552 | |
| 1553 | /* |
| 1554 | * Change mm, possibly for multiple threads in a threadgroup. This is |
| 1555 | * expensive and may sleep. |
| 1556 | */ |
| 1557 | cpuset_attach_nodemask_to = cs->mems_allowed; |
| 1558 | mm = get_task_mm(leader); |
| 1559 | if (mm) { |
| 1560 | struct cpuset *mems_oldcs = effective_nodemask_cpuset(oldcs); |
| 1561 | |
| 1562 | mpol_rebind_mm(mm, &cpuset_attach_nodemask_to); |
| 1563 | |
| 1564 | /* |
| 1565 | * old_mems_allowed is the same with mems_allowed here, except |
| 1566 | * if this task is being moved automatically due to hotplug. |
| 1567 | * In that case @mems_allowed has been updated and is empty, |
| 1568 | * so @old_mems_allowed is the right nodesets that we migrate |
| 1569 | * mm from. |
| 1570 | */ |
| 1571 | if (is_memory_migrate(cs)) { |
| 1572 | cpuset_migrate_mm(mm, &mems_oldcs->old_mems_allowed, |
| 1573 | &cpuset_attach_nodemask_to); |
| 1574 | } |
| 1575 | mmput(mm); |
| 1576 | } |
| 1577 | |
| 1578 | cs->old_mems_allowed = cpuset_attach_nodemask_to; |
| 1579 | |
| 1580 | cs->attach_in_progress--; |
| 1581 | if (!cs->attach_in_progress) |
| 1582 | wake_up(&cpuset_attach_wq); |
| 1583 | |
| 1584 | mutex_unlock(&cpuset_mutex); |
| 1585 | } |
| 1586 | |
| 1587 | /* The various types of files and directories in a cpuset file system */ |
| 1588 | |
| 1589 | typedef enum { |
| 1590 | FILE_MEMORY_MIGRATE, |
| 1591 | FILE_CPULIST, |
| 1592 | FILE_MEMLIST, |
| 1593 | FILE_CPU_EXCLUSIVE, |
| 1594 | FILE_MEM_EXCLUSIVE, |
| 1595 | FILE_MEM_HARDWALL, |
| 1596 | FILE_SCHED_LOAD_BALANCE, |
| 1597 | FILE_SCHED_RELAX_DOMAIN_LEVEL, |
| 1598 | FILE_MEMORY_PRESSURE_ENABLED, |
| 1599 | FILE_MEMORY_PRESSURE, |
| 1600 | FILE_SPREAD_PAGE, |
| 1601 | FILE_SPREAD_SLAB, |
| 1602 | } cpuset_filetype_t; |
| 1603 | |
| 1604 | static int cpuset_write_u64(struct cgroup *cgrp, struct cftype *cft, u64 val) |
| 1605 | { |
| 1606 | struct cpuset *cs = cgroup_cs(cgrp); |
| 1607 | cpuset_filetype_t type = cft->private; |
| 1608 | int retval = -ENODEV; |
| 1609 | |
| 1610 | mutex_lock(&cpuset_mutex); |
| 1611 | if (!is_cpuset_online(cs)) |
| 1612 | goto out_unlock; |
| 1613 | |
| 1614 | switch (type) { |
| 1615 | case FILE_CPU_EXCLUSIVE: |
| 1616 | retval = update_flag(CS_CPU_EXCLUSIVE, cs, val); |
| 1617 | break; |
| 1618 | case FILE_MEM_EXCLUSIVE: |
| 1619 | retval = update_flag(CS_MEM_EXCLUSIVE, cs, val); |
| 1620 | break; |
| 1621 | case FILE_MEM_HARDWALL: |
| 1622 | retval = update_flag(CS_MEM_HARDWALL, cs, val); |
| 1623 | break; |
| 1624 | case FILE_SCHED_LOAD_BALANCE: |
| 1625 | retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, val); |
| 1626 | break; |
| 1627 | case FILE_MEMORY_MIGRATE: |
| 1628 | retval = update_flag(CS_MEMORY_MIGRATE, cs, val); |
| 1629 | break; |
| 1630 | case FILE_MEMORY_PRESSURE_ENABLED: |
| 1631 | cpuset_memory_pressure_enabled = !!val; |
| 1632 | break; |
| 1633 | case FILE_MEMORY_PRESSURE: |
| 1634 | retval = -EACCES; |
| 1635 | break; |
| 1636 | case FILE_SPREAD_PAGE: |
| 1637 | retval = update_flag(CS_SPREAD_PAGE, cs, val); |
| 1638 | break; |
| 1639 | case FILE_SPREAD_SLAB: |
| 1640 | retval = update_flag(CS_SPREAD_SLAB, cs, val); |
| 1641 | break; |
| 1642 | default: |
| 1643 | retval = -EINVAL; |
| 1644 | break; |
| 1645 | } |
| 1646 | out_unlock: |
| 1647 | mutex_unlock(&cpuset_mutex); |
| 1648 | return retval; |
| 1649 | } |
| 1650 | |
| 1651 | static int cpuset_write_s64(struct cgroup *cgrp, struct cftype *cft, s64 val) |
| 1652 | { |
| 1653 | struct cpuset *cs = cgroup_cs(cgrp); |
| 1654 | cpuset_filetype_t type = cft->private; |
| 1655 | int retval = -ENODEV; |
| 1656 | |
| 1657 | mutex_lock(&cpuset_mutex); |
| 1658 | if (!is_cpuset_online(cs)) |
| 1659 | goto out_unlock; |
| 1660 | |
| 1661 | switch (type) { |
| 1662 | case FILE_SCHED_RELAX_DOMAIN_LEVEL: |
| 1663 | retval = update_relax_domain_level(cs, val); |
| 1664 | break; |
| 1665 | default: |
| 1666 | retval = -EINVAL; |
| 1667 | break; |
| 1668 | } |
| 1669 | out_unlock: |
| 1670 | mutex_unlock(&cpuset_mutex); |
| 1671 | return retval; |
| 1672 | } |
| 1673 | |
| 1674 | /* |
| 1675 | * Common handling for a write to a "cpus" or "mems" file. |
| 1676 | */ |
| 1677 | static int cpuset_write_resmask(struct cgroup *cgrp, struct cftype *cft, |
| 1678 | const char *buf) |
| 1679 | { |
| 1680 | struct cpuset *cs = cgroup_cs(cgrp); |
| 1681 | struct cpuset *trialcs; |
| 1682 | int retval = -ENODEV; |
| 1683 | |
| 1684 | /* |
| 1685 | * CPU or memory hotunplug may leave @cs w/o any execution |
| 1686 | * resources, in which case the hotplug code asynchronously updates |
| 1687 | * configuration and transfers all tasks to the nearest ancestor |
| 1688 | * which can execute. |
| 1689 | * |
| 1690 | * As writes to "cpus" or "mems" may restore @cs's execution |
| 1691 | * resources, wait for the previously scheduled operations before |
| 1692 | * proceeding, so that we don't end up keep removing tasks added |
| 1693 | * after execution capability is restored. |
| 1694 | */ |
| 1695 | flush_work(&cpuset_hotplug_work); |
| 1696 | |
| 1697 | mutex_lock(&cpuset_mutex); |
| 1698 | if (!is_cpuset_online(cs)) |
| 1699 | goto out_unlock; |
| 1700 | |
| 1701 | trialcs = alloc_trial_cpuset(cs); |
| 1702 | if (!trialcs) { |
| 1703 | retval = -ENOMEM; |
| 1704 | goto out_unlock; |
| 1705 | } |
| 1706 | |
| 1707 | switch (cft->private) { |
| 1708 | case FILE_CPULIST: |
| 1709 | retval = update_cpumask(cs, trialcs, buf); |
| 1710 | break; |
| 1711 | case FILE_MEMLIST: |
| 1712 | retval = update_nodemask(cs, trialcs, buf); |
| 1713 | break; |
| 1714 | default: |
| 1715 | retval = -EINVAL; |
| 1716 | break; |
| 1717 | } |
| 1718 | |
| 1719 | free_trial_cpuset(trialcs); |
| 1720 | out_unlock: |
| 1721 | mutex_unlock(&cpuset_mutex); |
| 1722 | return retval; |
| 1723 | } |
| 1724 | |
| 1725 | /* |
| 1726 | * These ascii lists should be read in a single call, by using a user |
| 1727 | * buffer large enough to hold the entire map. If read in smaller |
| 1728 | * chunks, there is no guarantee of atomicity. Since the display format |
| 1729 | * used, list of ranges of sequential numbers, is variable length, |
| 1730 | * and since these maps can change value dynamically, one could read |
| 1731 | * gibberish by doing partial reads while a list was changing. |
| 1732 | * A single large read to a buffer that crosses a page boundary is |
| 1733 | * ok, because the result being copied to user land is not recomputed |
| 1734 | * across a page fault. |
| 1735 | */ |
| 1736 | |
| 1737 | static size_t cpuset_sprintf_cpulist(char *page, struct cpuset *cs) |
| 1738 | { |
| 1739 | size_t count; |
| 1740 | |
| 1741 | mutex_lock(&callback_mutex); |
| 1742 | count = cpulist_scnprintf(page, PAGE_SIZE, cs->cpus_allowed); |
| 1743 | mutex_unlock(&callback_mutex); |
| 1744 | |
| 1745 | return count; |
| 1746 | } |
| 1747 | |
| 1748 | static size_t cpuset_sprintf_memlist(char *page, struct cpuset *cs) |
| 1749 | { |
| 1750 | size_t count; |
| 1751 | |
| 1752 | mutex_lock(&callback_mutex); |
| 1753 | count = nodelist_scnprintf(page, PAGE_SIZE, cs->mems_allowed); |
| 1754 | mutex_unlock(&callback_mutex); |
| 1755 | |
| 1756 | return count; |
| 1757 | } |
| 1758 | |
| 1759 | static ssize_t cpuset_common_file_read(struct cgroup *cgrp, |
| 1760 | struct cftype *cft, |
| 1761 | struct file *file, |
| 1762 | char __user *buf, |
| 1763 | size_t nbytes, loff_t *ppos) |
| 1764 | { |
| 1765 | struct cpuset *cs = cgroup_cs(cgrp); |
| 1766 | cpuset_filetype_t type = cft->private; |
| 1767 | char *page; |
| 1768 | ssize_t retval = 0; |
| 1769 | char *s; |
| 1770 | |
| 1771 | if (!(page = (char *)__get_free_page(GFP_TEMPORARY))) |
| 1772 | return -ENOMEM; |
| 1773 | |
| 1774 | s = page; |
| 1775 | |
| 1776 | switch (type) { |
| 1777 | case FILE_CPULIST: |
| 1778 | s += cpuset_sprintf_cpulist(s, cs); |
| 1779 | break; |
| 1780 | case FILE_MEMLIST: |
| 1781 | s += cpuset_sprintf_memlist(s, cs); |
| 1782 | break; |
| 1783 | default: |
| 1784 | retval = -EINVAL; |
| 1785 | goto out; |
| 1786 | } |
| 1787 | *s++ = '\n'; |
| 1788 | |
| 1789 | retval = simple_read_from_buffer(buf, nbytes, ppos, page, s - page); |
| 1790 | out: |
| 1791 | free_page((unsigned long)page); |
| 1792 | return retval; |
| 1793 | } |
| 1794 | |
| 1795 | static u64 cpuset_read_u64(struct cgroup *cgrp, struct cftype *cft) |
| 1796 | { |
| 1797 | struct cpuset *cs = cgroup_cs(cgrp); |
| 1798 | cpuset_filetype_t type = cft->private; |
| 1799 | switch (type) { |
| 1800 | case FILE_CPU_EXCLUSIVE: |
| 1801 | return is_cpu_exclusive(cs); |
| 1802 | case FILE_MEM_EXCLUSIVE: |
| 1803 | return is_mem_exclusive(cs); |
| 1804 | case FILE_MEM_HARDWALL: |
| 1805 | return is_mem_hardwall(cs); |
| 1806 | case FILE_SCHED_LOAD_BALANCE: |
| 1807 | return is_sched_load_balance(cs); |
| 1808 | case FILE_MEMORY_MIGRATE: |
| 1809 | return is_memory_migrate(cs); |
| 1810 | case FILE_MEMORY_PRESSURE_ENABLED: |
| 1811 | return cpuset_memory_pressure_enabled; |
| 1812 | case FILE_MEMORY_PRESSURE: |
| 1813 | return fmeter_getrate(&cs->fmeter); |
| 1814 | case FILE_SPREAD_PAGE: |
| 1815 | return is_spread_page(cs); |
| 1816 | case FILE_SPREAD_SLAB: |
| 1817 | return is_spread_slab(cs); |
| 1818 | default: |
| 1819 | BUG(); |
| 1820 | } |
| 1821 | |
| 1822 | /* Unreachable but makes gcc happy */ |
| 1823 | return 0; |
| 1824 | } |
| 1825 | |
| 1826 | static s64 cpuset_read_s64(struct cgroup *cgrp, struct cftype *cft) |
| 1827 | { |
| 1828 | struct cpuset *cs = cgroup_cs(cgrp); |
| 1829 | cpuset_filetype_t type = cft->private; |
| 1830 | switch (type) { |
| 1831 | case FILE_SCHED_RELAX_DOMAIN_LEVEL: |
| 1832 | return cs->relax_domain_level; |
| 1833 | default: |
| 1834 | BUG(); |
| 1835 | } |
| 1836 | |
| 1837 | /* Unrechable but makes gcc happy */ |
| 1838 | return 0; |
| 1839 | } |
| 1840 | |
| 1841 | |
| 1842 | /* |
| 1843 | * for the common functions, 'private' gives the type of file |
| 1844 | */ |
| 1845 | |
| 1846 | static struct cftype files[] = { |
| 1847 | { |
| 1848 | .name = "cpus", |
| 1849 | .read = cpuset_common_file_read, |
| 1850 | .write_string = cpuset_write_resmask, |
| 1851 | .max_write_len = (100U + 6 * NR_CPUS), |
| 1852 | .private = FILE_CPULIST, |
| 1853 | }, |
| 1854 | |
| 1855 | { |
| 1856 | .name = "mems", |
| 1857 | .read = cpuset_common_file_read, |
| 1858 | .write_string = cpuset_write_resmask, |
| 1859 | .max_write_len = (100U + 6 * MAX_NUMNODES), |
| 1860 | .private = FILE_MEMLIST, |
| 1861 | }, |
| 1862 | |
| 1863 | { |
| 1864 | .name = "cpu_exclusive", |
| 1865 | .read_u64 = cpuset_read_u64, |
| 1866 | .write_u64 = cpuset_write_u64, |
| 1867 | .private = FILE_CPU_EXCLUSIVE, |
| 1868 | }, |
| 1869 | |
| 1870 | { |
| 1871 | .name = "mem_exclusive", |
| 1872 | .read_u64 = cpuset_read_u64, |
| 1873 | .write_u64 = cpuset_write_u64, |
| 1874 | .private = FILE_MEM_EXCLUSIVE, |
| 1875 | }, |
| 1876 | |
| 1877 | { |
| 1878 | .name = "mem_hardwall", |
| 1879 | .read_u64 = cpuset_read_u64, |
| 1880 | .write_u64 = cpuset_write_u64, |
| 1881 | .private = FILE_MEM_HARDWALL, |
| 1882 | }, |
| 1883 | |
| 1884 | { |
| 1885 | .name = "sched_load_balance", |
| 1886 | .read_u64 = cpuset_read_u64, |
| 1887 | .write_u64 = cpuset_write_u64, |
| 1888 | .private = FILE_SCHED_LOAD_BALANCE, |
| 1889 | }, |
| 1890 | |
| 1891 | { |
| 1892 | .name = "sched_relax_domain_level", |
| 1893 | .read_s64 = cpuset_read_s64, |
| 1894 | .write_s64 = cpuset_write_s64, |
| 1895 | .private = FILE_SCHED_RELAX_DOMAIN_LEVEL, |
| 1896 | }, |
| 1897 | |
| 1898 | { |
| 1899 | .name = "memory_migrate", |
| 1900 | .read_u64 = cpuset_read_u64, |
| 1901 | .write_u64 = cpuset_write_u64, |
| 1902 | .private = FILE_MEMORY_MIGRATE, |
| 1903 | }, |
| 1904 | |
| 1905 | { |
| 1906 | .name = "memory_pressure", |
| 1907 | .read_u64 = cpuset_read_u64, |
| 1908 | .write_u64 = cpuset_write_u64, |
| 1909 | .private = FILE_MEMORY_PRESSURE, |
| 1910 | .mode = S_IRUGO, |
| 1911 | }, |
| 1912 | |
| 1913 | { |
| 1914 | .name = "memory_spread_page", |
| 1915 | .read_u64 = cpuset_read_u64, |
| 1916 | .write_u64 = cpuset_write_u64, |
| 1917 | .private = FILE_SPREAD_PAGE, |
| 1918 | }, |
| 1919 | |
| 1920 | { |
| 1921 | .name = "memory_spread_slab", |
| 1922 | .read_u64 = cpuset_read_u64, |
| 1923 | .write_u64 = cpuset_write_u64, |
| 1924 | .private = FILE_SPREAD_SLAB, |
| 1925 | }, |
| 1926 | |
| 1927 | { |
| 1928 | .name = "memory_pressure_enabled", |
| 1929 | .flags = CFTYPE_ONLY_ON_ROOT, |
| 1930 | .read_u64 = cpuset_read_u64, |
| 1931 | .write_u64 = cpuset_write_u64, |
| 1932 | .private = FILE_MEMORY_PRESSURE_ENABLED, |
| 1933 | }, |
| 1934 | |
| 1935 | { } /* terminate */ |
| 1936 | }; |
| 1937 | |
| 1938 | /* |
| 1939 | * cpuset_css_alloc - allocate a cpuset css |
| 1940 | * cgrp: control group that the new cpuset will be part of |
| 1941 | */ |
| 1942 | |
| 1943 | static struct cgroup_subsys_state *cpuset_css_alloc(struct cgroup *cgrp) |
| 1944 | { |
| 1945 | struct cpuset *cs; |
| 1946 | |
| 1947 | if (!cgrp->parent) |
| 1948 | return &top_cpuset.css; |
| 1949 | |
| 1950 | cs = kzalloc(sizeof(*cs), GFP_KERNEL); |
| 1951 | if (!cs) |
| 1952 | return ERR_PTR(-ENOMEM); |
| 1953 | if (!alloc_cpumask_var(&cs->cpus_allowed, GFP_KERNEL)) { |
| 1954 | kfree(cs); |
| 1955 | return ERR_PTR(-ENOMEM); |
| 1956 | } |
| 1957 | |
| 1958 | set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags); |
| 1959 | cpumask_clear(cs->cpus_allowed); |
| 1960 | nodes_clear(cs->mems_allowed); |
| 1961 | fmeter_init(&cs->fmeter); |
| 1962 | cs->relax_domain_level = -1; |
| 1963 | |
| 1964 | return &cs->css; |
| 1965 | } |
| 1966 | |
| 1967 | static int cpuset_css_online(struct cgroup *cgrp) |
| 1968 | { |
| 1969 | struct cpuset *cs = cgroup_cs(cgrp); |
| 1970 | struct cpuset *parent = parent_cs(cs); |
| 1971 | struct cpuset *tmp_cs; |
| 1972 | struct cgroup *pos_cgrp; |
| 1973 | |
| 1974 | if (!parent) |
| 1975 | return 0; |
| 1976 | |
| 1977 | mutex_lock(&cpuset_mutex); |
| 1978 | |
| 1979 | set_bit(CS_ONLINE, &cs->flags); |
| 1980 | if (is_spread_page(parent)) |
| 1981 | set_bit(CS_SPREAD_PAGE, &cs->flags); |
| 1982 | if (is_spread_slab(parent)) |
| 1983 | set_bit(CS_SPREAD_SLAB, &cs->flags); |
| 1984 | |
| 1985 | number_of_cpusets++; |
| 1986 | |
| 1987 | if (!test_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags)) |
| 1988 | goto out_unlock; |
| 1989 | |
| 1990 | /* |
| 1991 | * Clone @parent's configuration if CGRP_CPUSET_CLONE_CHILDREN is |
| 1992 | * set. This flag handling is implemented in cgroup core for |
| 1993 | * histrical reasons - the flag may be specified during mount. |
| 1994 | * |
| 1995 | * Currently, if any sibling cpusets have exclusive cpus or mem, we |
| 1996 | * refuse to clone the configuration - thereby refusing the task to |
| 1997 | * be entered, and as a result refusing the sys_unshare() or |
| 1998 | * clone() which initiated it. If this becomes a problem for some |
| 1999 | * users who wish to allow that scenario, then this could be |
| 2000 | * changed to grant parent->cpus_allowed-sibling_cpus_exclusive |
| 2001 | * (and likewise for mems) to the new cgroup. |
| 2002 | */ |
| 2003 | rcu_read_lock(); |
| 2004 | cpuset_for_each_child(tmp_cs, pos_cgrp, parent) { |
| 2005 | if (is_mem_exclusive(tmp_cs) || is_cpu_exclusive(tmp_cs)) { |
| 2006 | rcu_read_unlock(); |
| 2007 | goto out_unlock; |
| 2008 | } |
| 2009 | } |
| 2010 | rcu_read_unlock(); |
| 2011 | |
| 2012 | mutex_lock(&callback_mutex); |
| 2013 | cs->mems_allowed = parent->mems_allowed; |
| 2014 | cpumask_copy(cs->cpus_allowed, parent->cpus_allowed); |
| 2015 | mutex_unlock(&callback_mutex); |
| 2016 | out_unlock: |
| 2017 | mutex_unlock(&cpuset_mutex); |
| 2018 | return 0; |
| 2019 | } |
| 2020 | |
| 2021 | /* |
| 2022 | * If the cpuset being removed has its flag 'sched_load_balance' |
| 2023 | * enabled, then simulate turning sched_load_balance off, which |
| 2024 | * will call rebuild_sched_domains_locked(). |
| 2025 | */ |
| 2026 | |
| 2027 | static void cpuset_css_offline(struct cgroup *cgrp) |
| 2028 | { |
| 2029 | struct cpuset *cs = cgroup_cs(cgrp); |
| 2030 | |
| 2031 | mutex_lock(&cpuset_mutex); |
| 2032 | |
| 2033 | if (is_sched_load_balance(cs)) |
| 2034 | update_flag(CS_SCHED_LOAD_BALANCE, cs, 0); |
| 2035 | |
| 2036 | number_of_cpusets--; |
| 2037 | clear_bit(CS_ONLINE, &cs->flags); |
| 2038 | |
| 2039 | mutex_unlock(&cpuset_mutex); |
| 2040 | } |
| 2041 | |
| 2042 | static void cpuset_css_free(struct cgroup *cgrp) |
| 2043 | { |
| 2044 | struct cpuset *cs = cgroup_cs(cgrp); |
| 2045 | |
| 2046 | free_cpumask_var(cs->cpus_allowed); |
| 2047 | kfree(cs); |
| 2048 | } |
| 2049 | |
| 2050 | struct cgroup_subsys cpuset_subsys = { |
| 2051 | .name = "cpuset", |
| 2052 | .css_alloc = cpuset_css_alloc, |
| 2053 | .css_online = cpuset_css_online, |
| 2054 | .css_offline = cpuset_css_offline, |
| 2055 | .css_free = cpuset_css_free, |
| 2056 | .can_attach = cpuset_can_attach, |
| 2057 | .cancel_attach = cpuset_cancel_attach, |
| 2058 | .attach = cpuset_attach, |
| 2059 | .subsys_id = cpuset_subsys_id, |
| 2060 | .base_cftypes = files, |
| 2061 | .early_init = 1, |
| 2062 | }; |
| 2063 | |
| 2064 | /** |
| 2065 | * cpuset_init - initialize cpusets at system boot |
| 2066 | * |
| 2067 | * Description: Initialize top_cpuset and the cpuset internal file system, |
| 2068 | **/ |
| 2069 | |
| 2070 | int __init cpuset_init(void) |
| 2071 | { |
| 2072 | int err = 0; |
| 2073 | |
| 2074 | if (!alloc_cpumask_var(&top_cpuset.cpus_allowed, GFP_KERNEL)) |
| 2075 | BUG(); |
| 2076 | |
| 2077 | cpumask_setall(top_cpuset.cpus_allowed); |
| 2078 | nodes_setall(top_cpuset.mems_allowed); |
| 2079 | |
| 2080 | fmeter_init(&top_cpuset.fmeter); |
| 2081 | set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags); |
| 2082 | top_cpuset.relax_domain_level = -1; |
| 2083 | |
| 2084 | err = register_filesystem(&cpuset_fs_type); |
| 2085 | if (err < 0) |
| 2086 | return err; |
| 2087 | |
| 2088 | if (!alloc_cpumask_var(&cpus_attach, GFP_KERNEL)) |
| 2089 | BUG(); |
| 2090 | |
| 2091 | number_of_cpusets = 1; |
| 2092 | return 0; |
| 2093 | } |
| 2094 | |
| 2095 | /* |
| 2096 | * If CPU and/or memory hotplug handlers, below, unplug any CPUs |
| 2097 | * or memory nodes, we need to walk over the cpuset hierarchy, |
| 2098 | * removing that CPU or node from all cpusets. If this removes the |
| 2099 | * last CPU or node from a cpuset, then move the tasks in the empty |
| 2100 | * cpuset to its next-highest non-empty parent. |
| 2101 | */ |
| 2102 | static void remove_tasks_in_empty_cpuset(struct cpuset *cs) |
| 2103 | { |
| 2104 | struct cpuset *parent; |
| 2105 | |
| 2106 | /* |
| 2107 | * Find its next-highest non-empty parent, (top cpuset |
| 2108 | * has online cpus, so can't be empty). |
| 2109 | */ |
| 2110 | parent = parent_cs(cs); |
| 2111 | while (cpumask_empty(parent->cpus_allowed) || |
| 2112 | nodes_empty(parent->mems_allowed)) |
| 2113 | parent = parent_cs(parent); |
| 2114 | |
| 2115 | if (cgroup_transfer_tasks(parent->css.cgroup, cs->css.cgroup)) { |
| 2116 | rcu_read_lock(); |
| 2117 | printk(KERN_ERR "cpuset: failed to transfer tasks out of empty cpuset %s\n", |
| 2118 | cgroup_name(cs->css.cgroup)); |
| 2119 | rcu_read_unlock(); |
| 2120 | } |
| 2121 | } |
| 2122 | |
| 2123 | /** |
| 2124 | * cpuset_hotplug_update_tasks - update tasks in a cpuset for hotunplug |
| 2125 | * @cs: cpuset in interest |
| 2126 | * |
| 2127 | * Compare @cs's cpu and mem masks against top_cpuset and if some have gone |
| 2128 | * offline, update @cs accordingly. If @cs ends up with no CPU or memory, |
| 2129 | * all its tasks are moved to the nearest ancestor with both resources. |
| 2130 | */ |
| 2131 | static void cpuset_hotplug_update_tasks(struct cpuset *cs) |
| 2132 | { |
| 2133 | static cpumask_t off_cpus; |
| 2134 | static nodemask_t off_mems; |
| 2135 | bool is_empty; |
| 2136 | bool sane = cgroup_sane_behavior(cs->css.cgroup); |
| 2137 | |
| 2138 | retry: |
| 2139 | wait_event(cpuset_attach_wq, cs->attach_in_progress == 0); |
| 2140 | |
| 2141 | mutex_lock(&cpuset_mutex); |
| 2142 | |
| 2143 | /* |
| 2144 | * We have raced with task attaching. We wait until attaching |
| 2145 | * is finished, so we won't attach a task to an empty cpuset. |
| 2146 | */ |
| 2147 | if (cs->attach_in_progress) { |
| 2148 | mutex_unlock(&cpuset_mutex); |
| 2149 | goto retry; |
| 2150 | } |
| 2151 | |
| 2152 | cpumask_andnot(&off_cpus, cs->cpus_allowed, top_cpuset.cpus_allowed); |
| 2153 | nodes_andnot(off_mems, cs->mems_allowed, top_cpuset.mems_allowed); |
| 2154 | |
| 2155 | mutex_lock(&callback_mutex); |
| 2156 | cpumask_andnot(cs->cpus_allowed, cs->cpus_allowed, &off_cpus); |
| 2157 | mutex_unlock(&callback_mutex); |
| 2158 | |
| 2159 | /* |
| 2160 | * If sane_behavior flag is set, we need to update tasks' cpumask |
| 2161 | * for empty cpuset to take on ancestor's cpumask. Otherwise, don't |
| 2162 | * call update_tasks_cpumask() if the cpuset becomes empty, as |
| 2163 | * the tasks in it will be migrated to an ancestor. |
| 2164 | */ |
| 2165 | if ((sane && cpumask_empty(cs->cpus_allowed)) || |
| 2166 | (!cpumask_empty(&off_cpus) && !cpumask_empty(cs->cpus_allowed))) |
| 2167 | update_tasks_cpumask(cs, NULL); |
| 2168 | |
| 2169 | mutex_lock(&callback_mutex); |
| 2170 | nodes_andnot(cs->mems_allowed, cs->mems_allowed, off_mems); |
| 2171 | mutex_unlock(&callback_mutex); |
| 2172 | |
| 2173 | /* |
| 2174 | * If sane_behavior flag is set, we need to update tasks' nodemask |
| 2175 | * for empty cpuset to take on ancestor's nodemask. Otherwise, don't |
| 2176 | * call update_tasks_nodemask() if the cpuset becomes empty, as |
| 2177 | * the tasks in it will be migratd to an ancestor. |
| 2178 | */ |
| 2179 | if ((sane && nodes_empty(cs->mems_allowed)) || |
| 2180 | (!nodes_empty(off_mems) && !nodes_empty(cs->mems_allowed))) |
| 2181 | update_tasks_nodemask(cs, NULL); |
| 2182 | |
| 2183 | is_empty = cpumask_empty(cs->cpus_allowed) || |
| 2184 | nodes_empty(cs->mems_allowed); |
| 2185 | |
| 2186 | mutex_unlock(&cpuset_mutex); |
| 2187 | |
| 2188 | /* |
| 2189 | * If sane_behavior flag is set, we'll keep tasks in empty cpusets. |
| 2190 | * |
| 2191 | * Otherwise move tasks to the nearest ancestor with execution |
| 2192 | * resources. This is full cgroup operation which will |
| 2193 | * also call back into cpuset. Should be done outside any lock. |
| 2194 | */ |
| 2195 | if (!sane && is_empty) |
| 2196 | remove_tasks_in_empty_cpuset(cs); |
| 2197 | } |
| 2198 | |
| 2199 | /** |
| 2200 | * cpuset_hotplug_workfn - handle CPU/memory hotunplug for a cpuset |
| 2201 | * |
| 2202 | * This function is called after either CPU or memory configuration has |
| 2203 | * changed and updates cpuset accordingly. The top_cpuset is always |
| 2204 | * synchronized to cpu_active_mask and N_MEMORY, which is necessary in |
| 2205 | * order to make cpusets transparent (of no affect) on systems that are |
| 2206 | * actively using CPU hotplug but making no active use of cpusets. |
| 2207 | * |
| 2208 | * Non-root cpusets are only affected by offlining. If any CPUs or memory |
| 2209 | * nodes have been taken down, cpuset_hotplug_update_tasks() is invoked on |
| 2210 | * all descendants. |
| 2211 | * |
| 2212 | * Note that CPU offlining during suspend is ignored. We don't modify |
| 2213 | * cpusets across suspend/resume cycles at all. |
| 2214 | */ |
| 2215 | static void cpuset_hotplug_workfn(struct work_struct *work) |
| 2216 | { |
| 2217 | static cpumask_t new_cpus; |
| 2218 | static nodemask_t new_mems; |
| 2219 | bool cpus_updated, mems_updated; |
| 2220 | |
| 2221 | mutex_lock(&cpuset_mutex); |
| 2222 | |
| 2223 | /* fetch the available cpus/mems and find out which changed how */ |
| 2224 | cpumask_copy(&new_cpus, cpu_active_mask); |
| 2225 | new_mems = node_states[N_MEMORY]; |
| 2226 | |
| 2227 | cpus_updated = !cpumask_equal(top_cpuset.cpus_allowed, &new_cpus); |
| 2228 | mems_updated = !nodes_equal(top_cpuset.mems_allowed, new_mems); |
| 2229 | |
| 2230 | /* synchronize cpus_allowed to cpu_active_mask */ |
| 2231 | if (cpus_updated) { |
| 2232 | mutex_lock(&callback_mutex); |
| 2233 | cpumask_copy(top_cpuset.cpus_allowed, &new_cpus); |
| 2234 | mutex_unlock(&callback_mutex); |
| 2235 | /* we don't mess with cpumasks of tasks in top_cpuset */ |
| 2236 | } |
| 2237 | |
| 2238 | /* synchronize mems_allowed to N_MEMORY */ |
| 2239 | if (mems_updated) { |
| 2240 | mutex_lock(&callback_mutex); |
| 2241 | top_cpuset.mems_allowed = new_mems; |
| 2242 | mutex_unlock(&callback_mutex); |
| 2243 | update_tasks_nodemask(&top_cpuset, NULL); |
| 2244 | } |
| 2245 | |
| 2246 | mutex_unlock(&cpuset_mutex); |
| 2247 | |
| 2248 | /* if cpus or mems changed, we need to propagate to descendants */ |
| 2249 | if (cpus_updated || mems_updated) { |
| 2250 | struct cpuset *cs; |
| 2251 | struct cgroup *pos_cgrp; |
| 2252 | |
| 2253 | rcu_read_lock(); |
| 2254 | cpuset_for_each_descendant_pre(cs, pos_cgrp, &top_cpuset) { |
| 2255 | if (!css_tryget(&cs->css)) |
| 2256 | continue; |
| 2257 | rcu_read_unlock(); |
| 2258 | |
| 2259 | cpuset_hotplug_update_tasks(cs); |
| 2260 | |
| 2261 | rcu_read_lock(); |
| 2262 | css_put(&cs->css); |
| 2263 | } |
| 2264 | rcu_read_unlock(); |
| 2265 | } |
| 2266 | |
| 2267 | /* rebuild sched domains if cpus_allowed has changed */ |
| 2268 | if (cpus_updated) |
| 2269 | rebuild_sched_domains(); |
| 2270 | } |
| 2271 | |
| 2272 | void cpuset_update_active_cpus(bool cpu_online) |
| 2273 | { |
| 2274 | /* |
| 2275 | * We're inside cpu hotplug critical region which usually nests |
| 2276 | * inside cgroup synchronization. Bounce actual hotplug processing |
| 2277 | * to a work item to avoid reverse locking order. |
| 2278 | * |
| 2279 | * We still need to do partition_sched_domains() synchronously; |
| 2280 | * otherwise, the scheduler will get confused and put tasks to the |
| 2281 | * dead CPU. Fall back to the default single domain. |
| 2282 | * cpuset_hotplug_workfn() will rebuild it as necessary. |
| 2283 | */ |
| 2284 | partition_sched_domains(1, NULL, NULL); |
| 2285 | schedule_work(&cpuset_hotplug_work); |
| 2286 | } |
| 2287 | |
| 2288 | /* |
| 2289 | * Keep top_cpuset.mems_allowed tracking node_states[N_MEMORY]. |
| 2290 | * Call this routine anytime after node_states[N_MEMORY] changes. |
| 2291 | * See cpuset_update_active_cpus() for CPU hotplug handling. |
| 2292 | */ |
| 2293 | static int cpuset_track_online_nodes(struct notifier_block *self, |
| 2294 | unsigned long action, void *arg) |
| 2295 | { |
| 2296 | schedule_work(&cpuset_hotplug_work); |
| 2297 | return NOTIFY_OK; |
| 2298 | } |
| 2299 | |
| 2300 | static struct notifier_block cpuset_track_online_nodes_nb = { |
| 2301 | .notifier_call = cpuset_track_online_nodes, |
| 2302 | .priority = 10, /* ??! */ |
| 2303 | }; |
| 2304 | |
| 2305 | /** |
| 2306 | * cpuset_init_smp - initialize cpus_allowed |
| 2307 | * |
| 2308 | * Description: Finish top cpuset after cpu, node maps are initialized |
| 2309 | */ |
| 2310 | void __init cpuset_init_smp(void) |
| 2311 | { |
| 2312 | cpumask_copy(top_cpuset.cpus_allowed, cpu_active_mask); |
| 2313 | top_cpuset.mems_allowed = node_states[N_MEMORY]; |
| 2314 | top_cpuset.old_mems_allowed = top_cpuset.mems_allowed; |
| 2315 | |
| 2316 | register_hotmemory_notifier(&cpuset_track_online_nodes_nb); |
| 2317 | } |
| 2318 | |
| 2319 | /** |
| 2320 | * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset. |
| 2321 | * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed. |
| 2322 | * @pmask: pointer to struct cpumask variable to receive cpus_allowed set. |
| 2323 | * |
| 2324 | * Description: Returns the cpumask_var_t cpus_allowed of the cpuset |
| 2325 | * attached to the specified @tsk. Guaranteed to return some non-empty |
| 2326 | * subset of cpu_online_mask, even if this means going outside the |
| 2327 | * tasks cpuset. |
| 2328 | **/ |
| 2329 | |
| 2330 | void cpuset_cpus_allowed(struct task_struct *tsk, struct cpumask *pmask) |
| 2331 | { |
| 2332 | struct cpuset *cpus_cs; |
| 2333 | |
| 2334 | mutex_lock(&callback_mutex); |
| 2335 | task_lock(tsk); |
| 2336 | cpus_cs = effective_cpumask_cpuset(task_cs(tsk)); |
| 2337 | guarantee_online_cpus(cpus_cs, pmask); |
| 2338 | task_unlock(tsk); |
| 2339 | mutex_unlock(&callback_mutex); |
| 2340 | } |
| 2341 | |
| 2342 | void cpuset_cpus_allowed_fallback(struct task_struct *tsk) |
| 2343 | { |
| 2344 | struct cpuset *cpus_cs; |
| 2345 | |
| 2346 | rcu_read_lock(); |
| 2347 | cpus_cs = effective_cpumask_cpuset(task_cs(tsk)); |
| 2348 | do_set_cpus_allowed(tsk, cpus_cs->cpus_allowed); |
| 2349 | rcu_read_unlock(); |
| 2350 | |
| 2351 | /* |
| 2352 | * We own tsk->cpus_allowed, nobody can change it under us. |
| 2353 | * |
| 2354 | * But we used cs && cs->cpus_allowed lockless and thus can |
| 2355 | * race with cgroup_attach_task() or update_cpumask() and get |
| 2356 | * the wrong tsk->cpus_allowed. However, both cases imply the |
| 2357 | * subsequent cpuset_change_cpumask()->set_cpus_allowed_ptr() |
| 2358 | * which takes task_rq_lock(). |
| 2359 | * |
| 2360 | * If we are called after it dropped the lock we must see all |
| 2361 | * changes in tsk_cs()->cpus_allowed. Otherwise we can temporary |
| 2362 | * set any mask even if it is not right from task_cs() pov, |
| 2363 | * the pending set_cpus_allowed_ptr() will fix things. |
| 2364 | * |
| 2365 | * select_fallback_rq() will fix things ups and set cpu_possible_mask |
| 2366 | * if required. |
| 2367 | */ |
| 2368 | } |
| 2369 | |
| 2370 | void cpuset_init_current_mems_allowed(void) |
| 2371 | { |
| 2372 | nodes_setall(current->mems_allowed); |
| 2373 | } |
| 2374 | |
| 2375 | /** |
| 2376 | * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset. |
| 2377 | * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed. |
| 2378 | * |
| 2379 | * Description: Returns the nodemask_t mems_allowed of the cpuset |
| 2380 | * attached to the specified @tsk. Guaranteed to return some non-empty |
| 2381 | * subset of node_states[N_MEMORY], even if this means going outside the |
| 2382 | * tasks cpuset. |
| 2383 | **/ |
| 2384 | |
| 2385 | nodemask_t cpuset_mems_allowed(struct task_struct *tsk) |
| 2386 | { |
| 2387 | struct cpuset *mems_cs; |
| 2388 | nodemask_t mask; |
| 2389 | |
| 2390 | mutex_lock(&callback_mutex); |
| 2391 | task_lock(tsk); |
| 2392 | mems_cs = effective_nodemask_cpuset(task_cs(tsk)); |
| 2393 | guarantee_online_mems(mems_cs, &mask); |
| 2394 | task_unlock(tsk); |
| 2395 | mutex_unlock(&callback_mutex); |
| 2396 | |
| 2397 | return mask; |
| 2398 | } |
| 2399 | |
| 2400 | /** |
| 2401 | * cpuset_nodemask_valid_mems_allowed - check nodemask vs. curremt mems_allowed |
| 2402 | * @nodemask: the nodemask to be checked |
| 2403 | * |
| 2404 | * Are any of the nodes in the nodemask allowed in current->mems_allowed? |
| 2405 | */ |
| 2406 | int cpuset_nodemask_valid_mems_allowed(nodemask_t *nodemask) |
| 2407 | { |
| 2408 | return nodes_intersects(*nodemask, current->mems_allowed); |
| 2409 | } |
| 2410 | |
| 2411 | /* |
| 2412 | * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or |
| 2413 | * mem_hardwall ancestor to the specified cpuset. Call holding |
| 2414 | * callback_mutex. If no ancestor is mem_exclusive or mem_hardwall |
| 2415 | * (an unusual configuration), then returns the root cpuset. |
| 2416 | */ |
| 2417 | static struct cpuset *nearest_hardwall_ancestor(struct cpuset *cs) |
| 2418 | { |
| 2419 | while (!(is_mem_exclusive(cs) || is_mem_hardwall(cs)) && parent_cs(cs)) |
| 2420 | cs = parent_cs(cs); |
| 2421 | return cs; |
| 2422 | } |
| 2423 | |
| 2424 | /** |
| 2425 | * cpuset_node_allowed_softwall - Can we allocate on a memory node? |
| 2426 | * @node: is this an allowed node? |
| 2427 | * @gfp_mask: memory allocation flags |
| 2428 | * |
| 2429 | * If we're in interrupt, yes, we can always allocate. If __GFP_THISNODE is |
| 2430 | * set, yes, we can always allocate. If node is in our task's mems_allowed, |
| 2431 | * yes. If it's not a __GFP_HARDWALL request and this node is in the nearest |
| 2432 | * hardwalled cpuset ancestor to this task's cpuset, yes. If the task has been |
| 2433 | * OOM killed and has access to memory reserves as specified by the TIF_MEMDIE |
| 2434 | * flag, yes. |
| 2435 | * Otherwise, no. |
| 2436 | * |
| 2437 | * If __GFP_HARDWALL is set, cpuset_node_allowed_softwall() reduces to |
| 2438 | * cpuset_node_allowed_hardwall(). Otherwise, cpuset_node_allowed_softwall() |
| 2439 | * might sleep, and might allow a node from an enclosing cpuset. |
| 2440 | * |
| 2441 | * cpuset_node_allowed_hardwall() only handles the simpler case of hardwall |
| 2442 | * cpusets, and never sleeps. |
| 2443 | * |
| 2444 | * The __GFP_THISNODE placement logic is really handled elsewhere, |
| 2445 | * by forcibly using a zonelist starting at a specified node, and by |
| 2446 | * (in get_page_from_freelist()) refusing to consider the zones for |
| 2447 | * any node on the zonelist except the first. By the time any such |
| 2448 | * calls get to this routine, we should just shut up and say 'yes'. |
| 2449 | * |
| 2450 | * GFP_USER allocations are marked with the __GFP_HARDWALL bit, |
| 2451 | * and do not allow allocations outside the current tasks cpuset |
| 2452 | * unless the task has been OOM killed as is marked TIF_MEMDIE. |
| 2453 | * GFP_KERNEL allocations are not so marked, so can escape to the |
| 2454 | * nearest enclosing hardwalled ancestor cpuset. |
| 2455 | * |
| 2456 | * Scanning up parent cpusets requires callback_mutex. The |
| 2457 | * __alloc_pages() routine only calls here with __GFP_HARDWALL bit |
| 2458 | * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the |
| 2459 | * current tasks mems_allowed came up empty on the first pass over |
| 2460 | * the zonelist. So only GFP_KERNEL allocations, if all nodes in the |
| 2461 | * cpuset are short of memory, might require taking the callback_mutex |
| 2462 | * mutex. |
| 2463 | * |
| 2464 | * The first call here from mm/page_alloc:get_page_from_freelist() |
| 2465 | * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets, |
| 2466 | * so no allocation on a node outside the cpuset is allowed (unless |
| 2467 | * in interrupt, of course). |
| 2468 | * |
| 2469 | * The second pass through get_page_from_freelist() doesn't even call |
| 2470 | * here for GFP_ATOMIC calls. For those calls, the __alloc_pages() |
| 2471 | * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set |
| 2472 | * in alloc_flags. That logic and the checks below have the combined |
| 2473 | * affect that: |
| 2474 | * in_interrupt - any node ok (current task context irrelevant) |
| 2475 | * GFP_ATOMIC - any node ok |
| 2476 | * TIF_MEMDIE - any node ok |
| 2477 | * GFP_KERNEL - any node in enclosing hardwalled cpuset ok |
| 2478 | * GFP_USER - only nodes in current tasks mems allowed ok. |
| 2479 | * |
| 2480 | * Rule: |
| 2481 | * Don't call cpuset_node_allowed_softwall if you can't sleep, unless you |
| 2482 | * pass in the __GFP_HARDWALL flag set in gfp_flag, which disables |
| 2483 | * the code that might scan up ancestor cpusets and sleep. |
| 2484 | */ |
| 2485 | int __cpuset_node_allowed_softwall(int node, gfp_t gfp_mask) |
| 2486 | { |
| 2487 | struct cpuset *cs; /* current cpuset ancestors */ |
| 2488 | int allowed; /* is allocation in zone z allowed? */ |
| 2489 | |
| 2490 | if (in_interrupt() || (gfp_mask & __GFP_THISNODE)) |
| 2491 | return 1; |
| 2492 | might_sleep_if(!(gfp_mask & __GFP_HARDWALL)); |
| 2493 | if (node_isset(node, current->mems_allowed)) |
| 2494 | return 1; |
| 2495 | /* |
| 2496 | * Allow tasks that have access to memory reserves because they have |
| 2497 | * been OOM killed to get memory anywhere. |
| 2498 | */ |
| 2499 | if (unlikely(test_thread_flag(TIF_MEMDIE))) |
| 2500 | return 1; |
| 2501 | if (gfp_mask & __GFP_HARDWALL) /* If hardwall request, stop here */ |
| 2502 | return 0; |
| 2503 | |
| 2504 | if (current->flags & PF_EXITING) /* Let dying task have memory */ |
| 2505 | return 1; |
| 2506 | |
| 2507 | /* Not hardwall and node outside mems_allowed: scan up cpusets */ |
| 2508 | mutex_lock(&callback_mutex); |
| 2509 | |
| 2510 | task_lock(current); |
| 2511 | cs = nearest_hardwall_ancestor(task_cs(current)); |
| 2512 | task_unlock(current); |
| 2513 | |
| 2514 | allowed = node_isset(node, cs->mems_allowed); |
| 2515 | mutex_unlock(&callback_mutex); |
| 2516 | return allowed; |
| 2517 | } |
| 2518 | |
| 2519 | /* |
| 2520 | * cpuset_node_allowed_hardwall - Can we allocate on a memory node? |
| 2521 | * @node: is this an allowed node? |
| 2522 | * @gfp_mask: memory allocation flags |
| 2523 | * |
| 2524 | * If we're in interrupt, yes, we can always allocate. If __GFP_THISNODE is |
| 2525 | * set, yes, we can always allocate. If node is in our task's mems_allowed, |
| 2526 | * yes. If the task has been OOM killed and has access to memory reserves as |
| 2527 | * specified by the TIF_MEMDIE flag, yes. |
| 2528 | * Otherwise, no. |
| 2529 | * |
| 2530 | * The __GFP_THISNODE placement logic is really handled elsewhere, |
| 2531 | * by forcibly using a zonelist starting at a specified node, and by |
| 2532 | * (in get_page_from_freelist()) refusing to consider the zones for |
| 2533 | * any node on the zonelist except the first. By the time any such |
| 2534 | * calls get to this routine, we should just shut up and say 'yes'. |
| 2535 | * |
| 2536 | * Unlike the cpuset_node_allowed_softwall() variant, above, |
| 2537 | * this variant requires that the node be in the current task's |
| 2538 | * mems_allowed or that we're in interrupt. It does not scan up the |
| 2539 | * cpuset hierarchy for the nearest enclosing mem_exclusive cpuset. |
| 2540 | * It never sleeps. |
| 2541 | */ |
| 2542 | int __cpuset_node_allowed_hardwall(int node, gfp_t gfp_mask) |
| 2543 | { |
| 2544 | if (in_interrupt() || (gfp_mask & __GFP_THISNODE)) |
| 2545 | return 1; |
| 2546 | if (node_isset(node, current->mems_allowed)) |
| 2547 | return 1; |
| 2548 | /* |
| 2549 | * Allow tasks that have access to memory reserves because they have |
| 2550 | * been OOM killed to get memory anywhere. |
| 2551 | */ |
| 2552 | if (unlikely(test_thread_flag(TIF_MEMDIE))) |
| 2553 | return 1; |
| 2554 | return 0; |
| 2555 | } |
| 2556 | |
| 2557 | /** |
| 2558 | * cpuset_mem_spread_node() - On which node to begin search for a file page |
| 2559 | * cpuset_slab_spread_node() - On which node to begin search for a slab page |
| 2560 | * |
| 2561 | * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for |
| 2562 | * tasks in a cpuset with is_spread_page or is_spread_slab set), |
| 2563 | * and if the memory allocation used cpuset_mem_spread_node() |
| 2564 | * to determine on which node to start looking, as it will for |
| 2565 | * certain page cache or slab cache pages such as used for file |
| 2566 | * system buffers and inode caches, then instead of starting on the |
| 2567 | * local node to look for a free page, rather spread the starting |
| 2568 | * node around the tasks mems_allowed nodes. |
| 2569 | * |
| 2570 | * We don't have to worry about the returned node being offline |
| 2571 | * because "it can't happen", and even if it did, it would be ok. |
| 2572 | * |
| 2573 | * The routines calling guarantee_online_mems() are careful to |
| 2574 | * only set nodes in task->mems_allowed that are online. So it |
| 2575 | * should not be possible for the following code to return an |
| 2576 | * offline node. But if it did, that would be ok, as this routine |
| 2577 | * is not returning the node where the allocation must be, only |
| 2578 | * the node where the search should start. The zonelist passed to |
| 2579 | * __alloc_pages() will include all nodes. If the slab allocator |
| 2580 | * is passed an offline node, it will fall back to the local node. |
| 2581 | * See kmem_cache_alloc_node(). |
| 2582 | */ |
| 2583 | |
| 2584 | static int cpuset_spread_node(int *rotor) |
| 2585 | { |
| 2586 | int node; |
| 2587 | |
| 2588 | node = next_node(*rotor, current->mems_allowed); |
| 2589 | if (node == MAX_NUMNODES) |
| 2590 | node = first_node(current->mems_allowed); |
| 2591 | *rotor = node; |
| 2592 | return node; |
| 2593 | } |
| 2594 | |
| 2595 | int cpuset_mem_spread_node(void) |
| 2596 | { |
| 2597 | if (current->cpuset_mem_spread_rotor == NUMA_NO_NODE) |
| 2598 | current->cpuset_mem_spread_rotor = |
| 2599 | node_random(¤t->mems_allowed); |
| 2600 | |
| 2601 | return cpuset_spread_node(¤t->cpuset_mem_spread_rotor); |
| 2602 | } |
| 2603 | |
| 2604 | int cpuset_slab_spread_node(void) |
| 2605 | { |
| 2606 | if (current->cpuset_slab_spread_rotor == NUMA_NO_NODE) |
| 2607 | current->cpuset_slab_spread_rotor = |
| 2608 | node_random(¤t->mems_allowed); |
| 2609 | |
| 2610 | return cpuset_spread_node(¤t->cpuset_slab_spread_rotor); |
| 2611 | } |
| 2612 | |
| 2613 | EXPORT_SYMBOL_GPL(cpuset_mem_spread_node); |
| 2614 | |
| 2615 | /** |
| 2616 | * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's? |
| 2617 | * @tsk1: pointer to task_struct of some task. |
| 2618 | * @tsk2: pointer to task_struct of some other task. |
| 2619 | * |
| 2620 | * Description: Return true if @tsk1's mems_allowed intersects the |
| 2621 | * mems_allowed of @tsk2. Used by the OOM killer to determine if |
| 2622 | * one of the task's memory usage might impact the memory available |
| 2623 | * to the other. |
| 2624 | **/ |
| 2625 | |
| 2626 | int cpuset_mems_allowed_intersects(const struct task_struct *tsk1, |
| 2627 | const struct task_struct *tsk2) |
| 2628 | { |
| 2629 | return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed); |
| 2630 | } |
| 2631 | |
| 2632 | #define CPUSET_NODELIST_LEN (256) |
| 2633 | |
| 2634 | /** |
| 2635 | * cpuset_print_task_mems_allowed - prints task's cpuset and mems_allowed |
| 2636 | * @task: pointer to task_struct of some task. |
| 2637 | * |
| 2638 | * Description: Prints @task's name, cpuset name, and cached copy of its |
| 2639 | * mems_allowed to the kernel log. Must hold task_lock(task) to allow |
| 2640 | * dereferencing task_cs(task). |
| 2641 | */ |
| 2642 | void cpuset_print_task_mems_allowed(struct task_struct *tsk) |
| 2643 | { |
| 2644 | /* Statically allocated to prevent using excess stack. */ |
| 2645 | static char cpuset_nodelist[CPUSET_NODELIST_LEN]; |
| 2646 | static DEFINE_SPINLOCK(cpuset_buffer_lock); |
| 2647 | |
| 2648 | struct cgroup *cgrp = task_cs(tsk)->css.cgroup; |
| 2649 | |
| 2650 | rcu_read_lock(); |
| 2651 | spin_lock(&cpuset_buffer_lock); |
| 2652 | |
| 2653 | nodelist_scnprintf(cpuset_nodelist, CPUSET_NODELIST_LEN, |
| 2654 | tsk->mems_allowed); |
| 2655 | printk(KERN_INFO "%s cpuset=%s mems_allowed=%s\n", |
| 2656 | tsk->comm, cgroup_name(cgrp), cpuset_nodelist); |
| 2657 | |
| 2658 | spin_unlock(&cpuset_buffer_lock); |
| 2659 | rcu_read_unlock(); |
| 2660 | } |
| 2661 | |
| 2662 | /* |
| 2663 | * Collection of memory_pressure is suppressed unless |
| 2664 | * this flag is enabled by writing "1" to the special |
| 2665 | * cpuset file 'memory_pressure_enabled' in the root cpuset. |
| 2666 | */ |
| 2667 | |
| 2668 | int cpuset_memory_pressure_enabled __read_mostly; |
| 2669 | |
| 2670 | /** |
| 2671 | * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims. |
| 2672 | * |
| 2673 | * Keep a running average of the rate of synchronous (direct) |
| 2674 | * page reclaim efforts initiated by tasks in each cpuset. |
| 2675 | * |
| 2676 | * This represents the rate at which some task in the cpuset |
| 2677 | * ran low on memory on all nodes it was allowed to use, and |
| 2678 | * had to enter the kernels page reclaim code in an effort to |
| 2679 | * create more free memory by tossing clean pages or swapping |
| 2680 | * or writing dirty pages. |
| 2681 | * |
| 2682 | * Display to user space in the per-cpuset read-only file |
| 2683 | * "memory_pressure". Value displayed is an integer |
| 2684 | * representing the recent rate of entry into the synchronous |
| 2685 | * (direct) page reclaim by any task attached to the cpuset. |
| 2686 | **/ |
| 2687 | |
| 2688 | void __cpuset_memory_pressure_bump(void) |
| 2689 | { |
| 2690 | task_lock(current); |
| 2691 | fmeter_markevent(&task_cs(current)->fmeter); |
| 2692 | task_unlock(current); |
| 2693 | } |
| 2694 | |
| 2695 | #ifdef CONFIG_PROC_PID_CPUSET |
| 2696 | /* |
| 2697 | * proc_cpuset_show() |
| 2698 | * - Print tasks cpuset path into seq_file. |
| 2699 | * - Used for /proc/<pid>/cpuset. |
| 2700 | * - No need to task_lock(tsk) on this tsk->cpuset reference, as it |
| 2701 | * doesn't really matter if tsk->cpuset changes after we read it, |
| 2702 | * and we take cpuset_mutex, keeping cpuset_attach() from changing it |
| 2703 | * anyway. |
| 2704 | */ |
| 2705 | int proc_cpuset_show(struct seq_file *m, void *unused_v) |
| 2706 | { |
| 2707 | struct pid *pid; |
| 2708 | struct task_struct *tsk; |
| 2709 | char *buf; |
| 2710 | struct cgroup_subsys_state *css; |
| 2711 | int retval; |
| 2712 | |
| 2713 | retval = -ENOMEM; |
| 2714 | buf = kmalloc(PAGE_SIZE, GFP_KERNEL); |
| 2715 | if (!buf) |
| 2716 | goto out; |
| 2717 | |
| 2718 | retval = -ESRCH; |
| 2719 | pid = m->private; |
| 2720 | tsk = get_pid_task(pid, PIDTYPE_PID); |
| 2721 | if (!tsk) |
| 2722 | goto out_free; |
| 2723 | |
| 2724 | rcu_read_lock(); |
| 2725 | css = task_css(tsk, cpuset_subsys_id); |
| 2726 | retval = cgroup_path(css->cgroup, buf, PAGE_SIZE); |
| 2727 | rcu_read_unlock(); |
| 2728 | if (retval < 0) |
| 2729 | goto out_put_task; |
| 2730 | seq_puts(m, buf); |
| 2731 | seq_putc(m, '\n'); |
| 2732 | out_put_task: |
| 2733 | put_task_struct(tsk); |
| 2734 | out_free: |
| 2735 | kfree(buf); |
| 2736 | out: |
| 2737 | return retval; |
| 2738 | } |
| 2739 | #endif /* CONFIG_PROC_PID_CPUSET */ |
| 2740 | |
| 2741 | /* Display task mems_allowed in /proc/<pid>/status file. */ |
| 2742 | void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task) |
| 2743 | { |
| 2744 | seq_printf(m, "Mems_allowed:\t"); |
| 2745 | seq_nodemask(m, &task->mems_allowed); |
| 2746 | seq_printf(m, "\n"); |
| 2747 | seq_printf(m, "Mems_allowed_list:\t"); |
| 2748 | seq_nodemask_list(m, &task->mems_allowed); |
| 2749 | seq_printf(m, "\n"); |
| 2750 | } |