| 1 | |
| 2 | |
| 3 | * kernel/cpuset.c
|
| 4 | *
|
| 5 | * Processor and Memory placement constraints for sets of tasks.
|
| 6 | *
|
| 7 | * Copyright (C) 2003 BULL SA.
|
| 8 | * Copyright (C) 2004-2007 Silicon Graphics, Inc.
|
| 9 | * Copyright (C) 2006 Google, Inc
|
| 10 | *
|
| 11 | * Portions derived from Patrick Mochel's sysfs code.
|
| 12 | * sysfs is Copyright (c) 2001-3 Patrick Mochel
|
| 13 | *
|
| 14 | * 2003-10-10 Written by Simon Derr.
|
| 15 | * 2003-10-22 Updates by Stephen Hemminger.
|
| 16 | * 2004 May-July Rework by Paul Jackson.
|
| 17 | * 2006 Rework by Paul Menage to use generic cgroups
|
| 18 | * 2008 Rework of the scheduler domains and CPU hotplug handling
|
| 19 | * by Max Krasnyansky
|
| 20 |
|
| 21 | #include "cpuset-internal.h"
|
| 22 |
|
| 23 | #include <linux/init.h>
|
| 24 | #include <linux/interrupt.h>
|
| 25 | #include <linux/kernel.h>
|
| 26 | #include <linux/mempolicy.h>
|
| 27 | #include <linux/mm.h>
|
| 28 | #include <linux/memory.h>
|
| 29 | #include <linux/rcupdate.h>
|
| 30 | #include <linux/sched.h>
|
| 31 | #include <linux/sched/deadline.h>
|
| 32 | #include <linux/sched/mm.h>
|
| 33 | #include <linux/sched/task.h>
|
| 34 | #include <linux/security.h>
|
| 35 | #include <linux/oom.h>
|
| 36 | #include <linux/sched/isolation.h>
|
| 37 | #include <linux/wait.h>
|
| 38 | #include <linux/workqueue.h>
|
| 39 | #include <linux/task_work.h>
|
| 40 | #include <linux/llist.h>
|
| 41 |
|
| 42 | DEFINE_STATIC_KEY_FALSE(cpusets_pre_enable_key);
|
| 43 | DEFINE_STATIC_KEY_FALSE(cpusets_enabled_key);
|
| 44 |
|
| 45 | |
| 46 | * There could be abnormal cpuset configurations for cpu or memory
|
| 47 | * node binding, add this key to provide a quick low-cost judgment
|
| 48 | * of the situation.
|
| 49 |
|
| 50 | DEFINE_STATIC_KEY_FALSE(cpusets_insane_config_key);
|
| 51 |
|
| 52 | static const char * const perr_strings[] = {
|
| 53 | [PERR_INVCPUS] = "Invalid cpu list in cpuset.cpus.exclusive",
|
| 54 | [PERR_INVPARENT] = "Parent is an invalid partition root",
|
| 55 | [PERR_NOTPART] = "Parent is not a partition root",
|
| 56 | [PERR_NOTEXCL] = "Cpu list in cpuset.cpus not exclusive",
|
| 57 | [PERR_NOCPUS] = "Parent unable to distribute cpu downstream",
|
| 58 | [PERR_HOTPLUG] = "No cpu available due to hotplug",
|
| 59 | [PERR_CPUSEMPTY] = "cpuset.cpus and cpuset.cpus.exclusive are empty",
|
| 60 | [PERR_HKEEPING] = "partition config conflicts with housekeeping setup",
|
| 61 | [PERR_ACCESS] = "Enable partition not permitted",
|
| 62 | [PERR_REMOTE] = "Have remote partition underneath",
|
| 63 | };
|
| 64 |
|
| 65 | |
| 66 | * CPUSET Locking Convention
|
| 67 | * -------------------------
|
| 68 | *
|
| 69 | * Below are the four global/local locks guarding cpuset structures in lock
|
| 70 | * acquisition order:
|
| 71 | * - cpuset_top_mutex
|
| 72 | * - cpu_hotplug_lock (cpus_read_lock/cpus_write_lock)
|
| 73 | * - cpuset_mutex
|
| 74 | * - callback_lock (raw spinlock)
|
| 75 | *
|
| 76 | * As cpuset will now indirectly flush a number of different workqueues in
|
| 77 | * housekeeping_update() to update housekeeping cpumasks when the set of
|
| 78 | * isolated CPUs is going to be changed, it may be vulnerable to deadlock
|
| 79 | * if we hold cpus_read_lock while calling into housekeeping_update().
|
| 80 | *
|
| 81 | * The first cpuset_top_mutex will be held except when calling into
|
| 82 | * cpuset_handle_hotplug() from the CPU hotplug code where cpus_write_lock
|
| 83 | * and cpuset_mutex will be held instead. The main purpose of this mutex
|
| 84 | * is to prevent regular cpuset control file write actions from interfering
|
| 85 | * with the call to housekeeping_update(), though CPU hotplug operation can
|
| 86 | * still happen in parallel. This mutex also provides protection for some
|
| 87 | * internal variables.
|
| 88 | *
|
| 89 | * A task must hold all the remaining three locks to modify externally visible
|
| 90 | * or used fields of cpusets, though some of the internally used cpuset fields
|
| 91 | * and internal variables can be modified without holding callback_lock. If only
|
| 92 | * reliable read access of the externally used fields are needed, a task can
|
| 93 | * hold either cpuset_mutex or callback_lock which are exposed to other
|
| 94 | * external subsystems.
|
| 95 | *
|
| 96 | * If a task holds cpu_hotplug_lock and cpuset_mutex, it blocks others,
|
| 97 | * ensuring that it is the only task able to also acquire callback_lock and
|
| 98 | * be able to modify cpusets. It can perform various checks on the cpuset
|
| 99 | * structure first, knowing nothing will change. It can also allocate memory
|
| 100 | * without holding callback_lock. While it is performing these checks, various
|
| 101 | * callback routines can briefly acquire callback_lock to query cpusets. Once
|
| 102 | * it is ready to make the changes, it takes callback_lock, blocking everyone
|
| 103 | * else.
|
| 104 | *
|
| 105 | * Calls to the kernel memory allocator cannot be made while holding
|
| 106 | * callback_lock which is a spinlock, as the memory allocator may sleep or
|
| 107 | * call back into cpuset code and acquire callback_lock.
|
| 108 | *
|
| 109 | * Now, the task_struct fields mems_allowed and mempolicy may be changed
|
| 110 | * by other task, we use alloc_lock in the task_struct fields to protect
|
| 111 | * them.
|
| 112 | *
|
| 113 | * The cpuset_common_seq_show() handlers only hold callback_lock across
|
| 114 | * small pieces of code, such as when reading out possibly multi-word
|
| 115 | * cpumasks and nodemasks.
|
| 116 |
|
| 117 |
|
| 118 | static DEFINE_MUTEX(cpuset_top_mutex);
|
| 119 | static DEFINE_MUTEX(cpuset_mutex);
|
| 120 |
|
| 121 | |
| 122 | * File level internal variables below follow one of the following exclusion
|
| 123 | * rules.
|
| 124 | *
|
| 125 | * RWCS: Read/write-able by holding either cpus_write_lock (and optionally
|
| 126 | * cpuset_mutex) or both cpus_read_lock and cpuset_mutex.
|
| 127 | *
|
| 128 | * CSCB: Readable by holding either cpuset_mutex or callback_lock. Writable
|
| 129 | * by holding both cpuset_mutex and callback_lock.
|
| 130 | *
|
| 131 | * T: Read/write-able by holding the cpuset_top_mutex.
|
| 132 |
|
| 133 |
|
| 134 | |
| 135 | * For local partitions, update to subpartitions_cpus & isolated_cpus is done
|
| 136 | * in update_parent_effective_cpumask(). For remote partitions, it is done in
|
| 137 | * the remote_partition_*() and remote_cpus_update() helpers.
|
| 138 |
|
| 139 | |
| 140 | * Exclusive CPUs distributed out to local or remote sub-partitions of
|
| 141 | * top_cpuset
|
| 142 |
|
| 143 | static cpumask_var_t subpartitions_cpus;
|
| 144 |
|
| 145 | |
| 146 | * Exclusive CPUs in isolated partitions (shown in cpuset.cpus.isolated)
|
| 147 |
|
| 148 | static cpumask_var_t isolated_cpus;
|
| 149 |
|
| 150 | |
| 151 | * Set if housekeeping cpumasks are to be updated.
|
| 152 |
|
| 153 | static bool update_housekeeping;
|
| 154 |
|
| 155 | |
| 156 | * Set if "cpuset_v2_mode" mount option is used
|
| 157 | * Cached at bind time and not lock protected; accessed via {READ,WRITE}_ONCE
|
| 158 |
|
| 159 | static bool cpuset_v2_mode;
|
| 160 |
|
| 161 | |
| 162 | * Copy of isolated_cpus to be passed to housekeeping_update()
|
| 163 |
|
| 164 | static cpumask_var_t isolated_hk_cpus;
|
| 165 |
|
| 166 | |
| 167 | * A flag to force sched domain rebuild at the end of an operation.
|
| 168 | * It can be set in
|
| 169 | * - update_partition_sd_lb()
|
| 170 | * - update_cpumasks_hier()
|
| 171 | * - cpuset_update_flag()
|
| 172 | * - cpuset_hotplug_update_tasks()
|
| 173 | * - cpuset_handle_hotplug()
|
| 174 | *
|
| 175 | * Protected by cpuset_mutex (with cpus_read_lock held) or cpus_write_lock.
|
| 176 | *
|
| 177 | * Note that update_relax_domain_level() in cpuset-v1.c can still call
|
| 178 | * rebuild_sched_domains_locked() directly without using this flag.
|
| 179 |
|
| 180 | static bool force_sd_rebuild;
|
| 181 |
|
| 182 | |
| 183 | * Partition root states:
|
| 184 | *
|
| 185 | * 0 - member (not a partition root)
|
| 186 | * 1 - partition root
|
| 187 | * 2 - partition root without load balancing (isolated)
|
| 188 | * -1 - invalid partition root
|
| 189 | * -2 - invalid isolated partition root
|
| 190 | *
|
| 191 | * There are 2 types of partitions - local or remote. Local partitions are
|
| 192 | * those whose parents are partition root themselves. Setting of
|
| 193 | * cpuset.cpus.exclusive are optional in setting up local partitions.
|
| 194 | * Remote partitions are those whose parents are not partition roots. Passing
|
| 195 | * down exclusive CPUs by setting cpuset.cpus.exclusive along its ancestor
|
| 196 | * nodes are mandatory in creating a remote partition.
|
| 197 | *
|
| 198 | * For simplicity, a local partition can be created under a local or remote
|
| 199 | * partition but a remote partition cannot have any partition root in its
|
| 200 | * ancestor chain except the cgroup root.
|
| 201 | *
|
| 202 | * A valid partition can be formed by setting exclusive_cpus or cpus_allowed
|
| 203 | * if exclusive_cpus is not set. In the case of partition with empty
|
| 204 | * exclusive_cpus, all the conflicting exclusive CPUs specified in the
|
| 205 | * following cpumasks of sibling cpusets will be removed from its
|
| 206 | * cpus_allowed in determining its effective_xcpus.
|
| 207 | * - effective_xcpus
|
| 208 | * - exclusive_cpus
|
| 209 | *
|
| 210 | * The "cpuset.cpus.exclusive" control file should be used for setting up
|
| 211 | * partition if the users want to get as many CPUs as possible.
|
| 212 |
|
| 213 | #define PRS_MEMBER 0
|
| 214 | #define PRS_ROOT 1
|
| 215 | #define PRS_ISOLATED 2
|
| 216 | #define PRS_INVALID_ROOT -1
|
| 217 | #define PRS_INVALID_ISOLATED -2
|
| 218 |
|
| 219 | |
| 220 | * Temporary cpumasks for working with partitions that are passed among
|
| 221 | * functions to avoid memory allocation in inner functions.
|
| 222 |
|
| 223 | struct tmpmasks {
|
| 224 | cpumask_var_t addmask, delmask;
|
| 225 | cpumask_var_t new_cpus;
|
| 226 | };
|
| 227 |
|
| 228 | void inc_dl_tasks_cs(struct task_struct *p)
|
| 229 | {
|
| 230 | struct cpuset *cs = task_cs(p);
|
| 231 |
|
| 232 | atomic_inc(&cs->nr_deadline_tasks);
|
| 233 | }
|
| 234 |
|
| 235 | void dec_dl_tasks_cs(struct task_struct *p)
|
| 236 | {
|
| 237 | struct cpuset *cs = task_cs(p);
|
| 238 |
|
| 239 | atomic_dec(&cs->nr_deadline_tasks);
|
| 240 | }
|
| 241 |
|
| 242 | static inline bool is_partition_valid(const struct cpuset *cs)
|
| 243 | {
|
| 244 | return cs->partition_root_state > 0;
|
| 245 | }
|
| 246 |
|
| 247 | static inline bool is_partition_invalid(const struct cpuset *cs)
|
| 248 | {
|
| 249 | return cs->partition_root_state < 0;
|
| 250 | }
|
| 251 |
|
| 252 | static inline bool cs_is_member(const struct cpuset *cs)
|
| 253 | {
|
| 254 | return cs->partition_root_state == PRS_MEMBER;
|
| 255 | }
|
| 256 |
|
| 257 | |
| 258 | * Callers should hold callback_lock to modify partition_root_state.
|
| 259 |
|
| 260 | static inline void make_partition_invalid(struct cpuset *cs)
|
| 261 | {
|
| 262 | if (cs->partition_root_state > 0)
|
| 263 | cs->partition_root_state = -cs->partition_root_state;
|
| 264 | }
|
| 265 |
|
| 266 | |
| 267 | * Send notification event of whenever partition_root_state changes.
|
| 268 |
|
| 269 | static inline void notify_partition_change(struct cpuset *cs, int old_prs)
|
| 270 | {
|
| 271 | if (old_prs == cs->partition_root_state)
|
| 272 | return;
|
| 273 | cgroup_file_notify(&cs->partition_file);
|
| 274 |
|
| 275 |
|
| 276 | if (is_partition_valid(cs))
|
| 277 | WRITE_ONCE(cs->prs_err, PERR_NONE);
|
| 278 | }
|
| 279 |
|
| 280 | |
| 281 | * The top_cpuset is always synchronized to cpu_active_mask and we should avoid
|
| 282 | * using cpu_online_mask as much as possible. An active CPU is always an online
|
| 283 | * CPU, but not vice versa. cpu_active_mask and cpu_online_mask can differ
|
| 284 | * during hotplug operations. A CPU is marked active at the last stage of CPU
|
| 285 | * bringup (CPUHP_AP_ACTIVE). It is also the stage where cpuset hotplug code
|
| 286 | * will be called to update the sched domains so that the scheduler can move
|
| 287 | * a normal task to a newly active CPU or remove tasks away from a newly
|
| 288 | * inactivated CPU. The online bit is set much earlier in the CPU bringup
|
| 289 | * process and cleared much later in CPU teardown.
|
| 290 | *
|
| 291 | * If cpu_online_mask is used while a hotunplug operation is happening in
|
| 292 | * parallel, we may leave an offline CPU in cpu_allowed or some other masks.
|
| 293 |
|
| 294 | struct cpuset top_cpuset = {
|
| 295 | .flags = BIT(CS_CPU_EXCLUSIVE) |
|
| 296 | BIT(CS_MEM_EXCLUSIVE) | BIT(CS_SCHED_LOAD_BALANCE),
|
| 297 | .partition_root_state = PRS_ROOT,
|
| 298 | .dl_bw_cpu = -1,
|
| 299 | };
|
| 300 |
|
| 301 | |
| 302 | * cpuset_lock - Acquire the global cpuset mutex
|
| 303 | *
|
| 304 | * This locks the global cpuset mutex to prevent modifications to cpuset
|
| 305 | * hierarchy and configurations. This helper is not enough to make modification.
|
| 306 |
|
| 307 | void cpuset_lock(void)
|
| 308 | {
|
| 309 | mutex_lock(&cpuset_mutex);
|
| 310 | }
|
| 311 |
|
| 312 | void cpuset_unlock(void)
|
| 313 | {
|
| 314 | mutex_unlock(&cpuset_mutex);
|
| 315 | }
|
| 316 |
|
| 317 | void lockdep_assert_cpuset_lock_held(void)
|
| 318 | {
|
| 319 | lockdep_assert_held(&cpuset_mutex);
|
| 320 | }
|
| 321 |
|
| 322 | |
| 323 | * cpuset_full_lock - Acquire full protection for cpuset modification
|
| 324 | *
|
| 325 | * Takes both CPU hotplug read lock (cpus_read_lock()) and cpuset mutex
|
| 326 | * to safely modify cpuset data.
|
| 327 |
|
| 328 | void cpuset_full_lock(void)
|
| 329 | {
|
| 330 | mutex_lock(&cpuset_top_mutex);
|
| 331 | cpus_read_lock();
|
| 332 | mutex_lock(&cpuset_mutex);
|
| 333 | }
|
| 334 |
|
| 335 | void cpuset_full_unlock(void)
|
| 336 | {
|
| 337 | mutex_unlock(&cpuset_mutex);
|
| 338 | cpus_read_unlock();
|
| 339 | mutex_unlock(&cpuset_top_mutex);
|
| 340 | }
|
| 341 |
|
| 342 | #ifdef CONFIG_LOCKDEP
|
| 343 | bool lockdep_is_cpuset_held(void)
|
| 344 | {
|
| 345 | return lockdep_is_held(&cpuset_mutex) ||
|
| 346 | lockdep_is_held(&cpuset_top_mutex);
|
| 347 | }
|
| 348 | #endif
|
| 349 |
|
| 350 | static DEFINE_SPINLOCK(callback_lock);
|
| 351 |
|
| 352 | void cpuset_callback_lock_irq(void)
|
| 353 | {
|
| 354 | spin_lock_irq(&callback_lock);
|
| 355 | }
|
| 356 |
|
| 357 | void cpuset_callback_unlock_irq(void)
|
| 358 | {
|
| 359 | spin_unlock_irq(&callback_lock);
|
| 360 | }
|
| 361 |
|
| 362 | static struct workqueue_struct *cpuset_migrate_mm_wq;
|
| 363 |
|
| 364 | static DECLARE_WAIT_QUEUE_HEAD(cpuset_attach_wq);
|
| 365 |
|
| 366 | |
| 367 | * Cpuset task attach context
|
| 368 | * Protected by cpuset_mutex
|
| 369 |
|
| 370 | static struct {
|
| 371 | int in_progress;
|
| 372 | bool cpus_updated;
|
| 373 | bool mems_updated;
|
| 374 | bool task_work_queued;
|
| 375 | bool many_dest_cs;
|
| 376 | struct cpuset *old_cs;
|
| 377 | nodemask_t nodemask_to;
|
| 378 | } attach_ctx;
|
| 379 | static LLIST_HEAD(src_cs_head);
|
| 380 | static LLIST_HEAD(dst_cs_head);
|
| 381 |
|
| 382 | |
| 383 | * Wait if task attach is in progress until it is done and then acquire
|
| 384 | * cpuset_mutex before returning.
|
| 385 |
|
| 386 | static void wait_attach_done_lock(void)
|
| 387 | __acquires(&cpuset_mutex)
|
| 388 | {
|
| 389 | for (;;) {
|
| 390 | mutex_lock(&cpuset_mutex);
|
| 391 | if (!attach_ctx.in_progress)
|
| 392 | return;
|
| 393 |
|
| 394 | mutex_unlock(&cpuset_mutex);
|
| 395 |
|
| 396 |
|
| 397 | wait_event(cpuset_attach_wq, attach_ctx.in_progress == 0);
|
| 398 | }
|
| 399 | }
|
| 400 |
|
| 401 | static inline void check_insane_mems_config(nodemask_t *nodes)
|
| 402 | {
|
| 403 | if (!cpusets_insane_config() &&
|
| 404 | movable_only_nodes(nodes)) {
|
| 405 | static_branch_enable_cpuslocked(&cpusets_insane_config_key);
|
| 406 | pr_info("Unsupported (movable nodes only) cpuset configuration detected (nmask=%*pbl)!\n"
|
| 407 | "Cpuset allocations might fail even with a lot of memory available.\n",
|
| 408 | nodemask_pr_args(nodes));
|
| 409 | }
|
| 410 | }
|
| 411 |
|
| 412 | |
| 413 | * decrease attach_ctx.in_progress.
|
| 414 | * wake_up cpuset_attach_wq if attach_ctx.in_progress==0.
|
| 415 |
|
| 416 | static inline void dec_attach_in_progress_locked(void)
|
| 417 | {
|
| 418 | lockdep_assert_cpuset_lock_held();
|
| 419 |
|
| 420 | attach_ctx.in_progress--;
|
| 421 | if (!attach_ctx.in_progress)
|
| 422 | wake_up(&cpuset_attach_wq);
|
| 423 | }
|
| 424 |
|
| 425 | static inline void dec_attach_in_progress(void)
|
| 426 | {
|
| 427 | mutex_lock(&cpuset_mutex);
|
| 428 | dec_attach_in_progress_locked();
|
| 429 | mutex_unlock(&cpuset_mutex);
|
| 430 | }
|
| 431 |
|
| 432 | static inline bool cpuset_v2(void)
|
| 433 | {
|
| 434 | return !IS_ENABLED(CONFIG_CPUSETS_V1) ||
|
| 435 | cgroup_subsys_on_dfl(cpuset_cgrp_subsys);
|
| 436 | }
|
| 437 |
|
| 438 | |
| 439 | * Cgroup v2 behavior is used on the "cpus" and "mems" control files when
|
| 440 | * on default hierarchy or when the cpuset_v2_mode flag is set by mounting
|
| 441 | * the v1 cpuset cgroup filesystem with the "cpuset_v2_mode" mount option.
|
| 442 | * With v2 behavior, "cpus" and "mems" are always what the users have
|
| 443 | * requested and won't be changed by hotplug events. Only the effective
|
| 444 | * cpus or mems will be affected.
|
| 445 |
|
| 446 | static inline bool is_in_v2_mode(void)
|
| 447 | {
|
| 448 | return cpuset_v2() || READ_ONCE(cpuset_v2_mode);
|
| 449 | }
|
| 450 |
|
| 451 | |
| 452 | * partition_is_populated - check if partition has tasks
|
| 453 | * @cs: partition root to be checked
|
| 454 | * @excluded_child: a child cpuset to be excluded in task checking
|
| 455 | * Return: true if there are tasks, false otherwise
|
| 456 | *
|
| 457 | * @cs should be a valid partition root or going to become a partition root.
|
| 458 | * @excluded_child should be non-NULL when this cpuset is going to become a
|
| 459 | * partition itself.
|
| 460 | *
|
| 461 | * Note that a remote partition is not allowed underneath a valid local
|
| 462 | * or remote partition. So if a non-partition root child is populated,
|
| 463 | * the whole partition is considered populated.
|
| 464 |
|
| 465 | static inline bool partition_is_populated(struct cpuset *cs,
|
| 466 | struct cpuset *excluded_child)
|
| 467 | {
|
| 468 | struct cpuset *cp;
|
| 469 | struct cgroup_subsys_state *pos_css;
|
| 470 |
|
| 471 | |
| 472 | * We cannot call cs_is_populated(cs) directly, as
|
| 473 | * nr_populated_domain_children may include populated
|
| 474 | * csets from descendants that are partitions.
|
| 475 |
|
| 476 | if (cgroup_has_tasks(cs->css.cgroup))
|
| 477 | return true;
|
| 478 |
|
| 479 | rcu_read_lock();
|
| 480 | cpuset_for_each_descendant_pre(cp, pos_css, cs) {
|
| 481 | if (cp == cs || cp == excluded_child)
|
| 482 | continue;
|
| 483 |
|
| 484 | if (is_partition_valid(cp)) {
|
| 485 | pos_css = css_rightmost_descendant(pos_css);
|
| 486 | continue;
|
| 487 | }
|
| 488 |
|
| 489 | if (cpuset_is_populated(cp)) {
|
| 490 | rcu_read_unlock();
|
| 491 | return true;
|
| 492 | }
|
| 493 | }
|
| 494 | rcu_read_unlock();
|
| 495 | return false;
|
| 496 | }
|
| 497 |
|
| 498 | |
| 499 | * Return in pmask the portion of a task's cpusets's cpus_allowed that
|
| 500 | * are online and are capable of running the task. If none are found,
|
| 501 | * walk up the cpuset hierarchy until we find one that does have some
|
| 502 | * appropriate cpus.
|
| 503 | *
|
| 504 | * One way or another, we guarantee to return some non-empty subset
|
| 505 | * of cpu_active_mask.
|
| 506 | *
|
| 507 | * Call with callback_lock or cpuset_mutex held.
|
| 508 |
|
| 509 | static void guarantee_active_cpus(struct task_struct *tsk,
|
| 510 | struct cpumask *pmask)
|
| 511 | {
|
| 512 | const struct cpumask *possible_mask = task_cpu_possible_mask(tsk);
|
| 513 | struct cpuset *cs;
|
| 514 |
|
| 515 | if (WARN_ON(!cpumask_and(pmask, possible_mask, cpu_active_mask)))
|
| 516 | cpumask_copy(pmask, cpu_active_mask);
|
| 517 |
|
| 518 | rcu_read_lock();
|
| 519 | cs = task_cs(tsk);
|
| 520 |
|
| 521 | while (!cpumask_intersects(cs->effective_cpus, pmask)) {
|
| 522 | cs = parent_cs(cs);
|
| 523 | if (unlikely(!cs)) {
|
| 524 | |
| 525 | * The top cpuset doesn't have any active cpu as a
|
| 526 | * consequence of a race between its caller and the cpu
|
| 527 | * hotplug operation where cpu_active_mask is updated
|
| 528 | * asynchronously before cpuset_handle_hotplug() is
|
| 529 | * being called to adjust the effective_cpus of the
|
| 530 | * affected cpusets. But we know the top cpuset's
|
| 531 | * effective_cpus is on its way to be identical to
|
| 532 | * cpu_active_mask minus the exclusive CPUs dedicated
|
| 533 | * to other valid cpuset partitions. Just pass back
|
| 534 | * the filtered cpu_active_mask in this case.
|
| 535 |
|
| 536 | goto out_unlock;
|
| 537 | }
|
| 538 | }
|
| 539 | cpumask_and(pmask, pmask, cs->effective_cpus);
|
| 540 | out_unlock:
|
| 541 | rcu_read_unlock();
|
| 542 | }
|
| 543 |
|
| 544 | |
| 545 | * Return in *pmask the portion of a cpusets's mems_allowed that
|
| 546 | * are online, with memory. If none are online with memory, walk
|
| 547 | * up the cpuset hierarchy until we find one that does have some
|
| 548 | * online mems. The top cpuset always has some mems online. With v2,
|
| 549 | * effective_mems should always contain online memory nodes except
|
| 550 | * during the transition period where a memory node hotunplug operation
|
| 551 | * is in progress.
|
| 552 | *
|
| 553 | * One way or another, we guarantee to return some non-empty subset
|
| 554 | * of node_states[N_MEMORY].
|
| 555 | *
|
| 556 | * Call with callback_lock or cpuset_mutex held.
|
| 557 |
|
| 558 | static void guarantee_online_mems(struct cpuset *cs, nodemask_t *pmask)
|
| 559 | {
|
| 560 | while (!nodes_and(*pmask, cs->effective_mems, node_states[N_MEMORY]))
|
| 561 | cs = parent_cs(cs);
|
| 562 | }
|
| 563 |
|
| 564 | |
| 565 | * alloc_cpumasks - Allocate an array of cpumask variables
|
| 566 | * @pmasks: Pointer to array of cpumask_var_t pointers
|
| 567 | * @size: Number of cpumasks to allocate
|
| 568 | * Return: 0 if successful, -ENOMEM otherwise.
|
| 569 | *
|
| 570 | * Allocates @size cpumasks and initializes them to empty. Returns 0 on
|
| 571 | * success, -ENOMEM on allocation failure. On failure, any previously
|
| 572 | * allocated cpumasks are freed.
|
| 573 |
|
| 574 | static inline int alloc_cpumasks(cpumask_var_t *pmasks[], u32 size)
|
| 575 | {
|
| 576 | int i;
|
| 577 |
|
| 578 | for (i = 0; i < size; i++) {
|
| 579 | if (!zalloc_cpumask_var(pmasks[i], GFP_KERNEL)) {
|
| 580 | while (--i >= 0)
|
| 581 | free_cpumask_var(*pmasks[i]);
|
| 582 | return -ENOMEM;
|
| 583 | }
|
| 584 | }
|
| 585 | return 0;
|
| 586 | }
|
| 587 |
|
| 588 | |
| 589 | * alloc_tmpmasks - Allocate temporary cpumasks for cpuset operations.
|
| 590 | * @tmp: Pointer to tmpmasks structure to populate
|
| 591 | * Return: 0 on success, -ENOMEM on allocation failure
|
| 592 |
|
| 593 | static inline int alloc_tmpmasks(struct tmpmasks *tmp)
|
| 594 | {
|
| 595 | |
| 596 | * Array of pointers to the three cpumask_var_t fields in tmpmasks.
|
| 597 | * Note: Array size must match actual number of masks (3)
|
| 598 |
|
| 599 | cpumask_var_t *pmask[3] = {
|
| 600 | &tmp->new_cpus,
|
| 601 | &tmp->addmask,
|
| 602 | &tmp->delmask
|
| 603 | };
|
| 604 |
|
| 605 | return alloc_cpumasks(pmask, ARRAY_SIZE(pmask));
|
| 606 | }
|
| 607 |
|
| 608 | |
| 609 | * free_tmpmasks - free cpumasks in a tmpmasks structure
|
| 610 | * @tmp: the tmpmasks structure pointer
|
| 611 |
|
| 612 | static inline void free_tmpmasks(struct tmpmasks *tmp)
|
| 613 | {
|
| 614 | if (!tmp)
|
| 615 | return;
|
| 616 |
|
| 617 | free_cpumask_var(tmp->new_cpus);
|
| 618 | free_cpumask_var(tmp->addmask);
|
| 619 | free_cpumask_var(tmp->delmask);
|
| 620 | }
|
| 621 |
|
| 622 | |
| 623 | * dup_or_alloc_cpuset - Duplicate or allocate a new cpuset
|
| 624 | * @cs: Source cpuset to duplicate (NULL for a fresh allocation)
|
| 625 | *
|
| 626 | * Creates a new cpuset by either:
|
| 627 | * 1. Duplicating an existing cpuset (if @cs is non-NULL), or
|
| 628 | * 2. Allocating a fresh cpuset with zero-initialized masks (if @cs is NULL)
|
| 629 | *
|
| 630 | * Return: Pointer to newly allocated cpuset on success, NULL on failure
|
| 631 |
|
| 632 | static struct cpuset *dup_or_alloc_cpuset(struct cpuset *cs)
|
| 633 | {
|
| 634 | struct cpuset *trial;
|
| 635 |
|
| 636 |
|
| 637 | trial = cs ? kmemdup(cs, sizeof(*cs), GFP_KERNEL) :
|
| 638 | kzalloc_obj(*cs);
|
| 639 | if (!trial)
|
| 640 | return NULL;
|
| 641 |
|
| 642 | trial->dl_bw_cpu = -1;
|
| 643 | init_llist_node(&trial->attach_node);
|
| 644 |
|
| 645 |
|
| 646 | cpumask_var_t *pmask[4] = {
|
| 647 | &trial->cpus_allowed,
|
| 648 | &trial->effective_cpus,
|
| 649 | &trial->effective_xcpus,
|
| 650 | &trial->exclusive_cpus
|
| 651 | };
|
| 652 |
|
| 653 | if (alloc_cpumasks(pmask, ARRAY_SIZE(pmask))) {
|
| 654 | kfree(trial);
|
| 655 | return NULL;
|
| 656 | }
|
| 657 |
|
| 658 |
|
| 659 | if (cs) {
|
| 660 | cpumask_copy(trial->cpus_allowed, cs->cpus_allowed);
|
| 661 | cpumask_copy(trial->effective_cpus, cs->effective_cpus);
|
| 662 | cpumask_copy(trial->effective_xcpus, cs->effective_xcpus);
|
| 663 | cpumask_copy(trial->exclusive_cpus, cs->exclusive_cpus);
|
| 664 | }
|
| 665 |
|
| 666 | return trial;
|
| 667 | }
|
| 668 |
|
| 669 | |
| 670 | * free_cpuset - free the cpuset
|
| 671 | * @cs: the cpuset to be freed
|
| 672 |
|
| 673 | static inline void free_cpuset(struct cpuset *cs)
|
| 674 | {
|
| 675 | free_cpumask_var(cs->cpus_allowed);
|
| 676 | free_cpumask_var(cs->effective_cpus);
|
| 677 | free_cpumask_var(cs->effective_xcpus);
|
| 678 | free_cpumask_var(cs->exclusive_cpus);
|
| 679 | kfree(cs);
|
| 680 | }
|
| 681 |
|
| 682 |
|
| 683 | static inline struct cpumask *user_xcpus(struct cpuset *cs)
|
| 684 | {
|
| 685 | return cpumask_empty(cs->exclusive_cpus) ? cs->cpus_allowed
|
| 686 | : cs->exclusive_cpus;
|
| 687 | }
|
| 688 |
|
| 689 | static inline bool xcpus_empty(struct cpuset *cs)
|
| 690 | {
|
| 691 | return cpumask_empty(cs->cpus_allowed) &&
|
| 692 | cpumask_empty(cs->exclusive_cpus);
|
| 693 | }
|
| 694 |
|
| 695 | |
| 696 | * cpusets_are_exclusive() - check if two cpusets are exclusive
|
| 697 | *
|
| 698 | * Return true if exclusive, false if not
|
| 699 |
|
| 700 | static inline bool cpusets_are_exclusive(struct cpuset *cs1, struct cpuset *cs2)
|
| 701 | {
|
| 702 | struct cpumask *xcpus1 = user_xcpus(cs1);
|
| 703 | struct cpumask *xcpus2 = user_xcpus(cs2);
|
| 704 |
|
| 705 | if (cpumask_intersects(xcpus1, xcpus2))
|
| 706 | return false;
|
| 707 | return true;
|
| 708 | }
|
| 709 |
|
| 710 | |
| 711 | * cpus_excl_conflict - Check if two cpusets have exclusive CPU conflicts
|
| 712 | * @trial: the trial cpuset to be checked
|
| 713 | * @sibling: a sibling cpuset to be checked against
|
| 714 | * @xcpus_changed: set if exclusive_cpus has been set
|
| 715 | *
|
| 716 | * Returns: true if CPU exclusivity conflict exists, false otherwise
|
| 717 | *
|
| 718 | * Conflict detection rules:
|
| 719 | * o cgroup v1
|
| 720 | * See cpuset1_cpus_excl_conflict()
|
| 721 | * o cgroup v2
|
| 722 | * - The exclusive_cpus values cannot overlap.
|
| 723 | * - New exclusive_cpus cannot be a superset of a sibling's cpus_allowed.
|
| 724 |
|
| 725 | static inline bool cpus_excl_conflict(struct cpuset *trial, struct cpuset *sibling,
|
| 726 | bool xcpus_changed)
|
| 727 | {
|
| 728 | if (!cpuset_v2())
|
| 729 | return cpuset1_cpus_excl_conflict(trial, sibling);
|
| 730 |
|
| 731 |
|
| 732 | if (xcpus_changed && !cpumask_empty(sibling->cpus_allowed) &&
|
| 733 | cpumask_subset(sibling->cpus_allowed, trial->exclusive_cpus))
|
| 734 | return true;
|
| 735 |
|
| 736 |
|
| 737 | return cpumask_intersects(trial->exclusive_cpus, sibling->exclusive_cpus);
|
| 738 | }
|
| 739 |
|
| 740 | static inline bool mems_excl_conflict(struct cpuset *cs1, struct cpuset *cs2)
|
| 741 | {
|
| 742 | if ((is_mem_exclusive(cs1) || is_mem_exclusive(cs2)))
|
| 743 | return nodes_intersects(cs1->mems_allowed, cs2->mems_allowed);
|
| 744 | return false;
|
| 745 | }
|
| 746 |
|
| 747 | |
| 748 | * validate_change() - Used to validate that any proposed cpuset change
|
| 749 | * follows the structural rules for cpusets.
|
| 750 | *
|
| 751 | * If we replaced the flag and mask values of the current cpuset
|
| 752 | * (cur) with those values in the trial cpuset (trial), would
|
| 753 | * our various subset and exclusive rules still be valid? Presumes
|
| 754 | * cpuset_mutex held.
|
| 755 | *
|
| 756 | * 'cur' is the address of an actual, in-use cpuset. Operations
|
| 757 | * such as list traversal that depend on the actual address of the
|
| 758 | * cpuset in the list must use cur below, not trial.
|
| 759 | *
|
| 760 | * 'trial' is the address of bulk structure copy of cur, with
|
| 761 | * perhaps one or more of the fields cpus_allowed, mems_allowed,
|
| 762 | * or flags changed to new, trial values.
|
| 763 | *
|
| 764 | * Return 0 if valid, -errno if not.
|
| 765 |
|
| 766 |
|
| 767 | static int validate_change(struct cpuset *cur, struct cpuset *trial)
|
| 768 | {
|
| 769 | struct cgroup_subsys_state *css;
|
| 770 | struct cpuset *c, *par;
|
| 771 | bool xcpus_changed;
|
| 772 | int ret = 0;
|
| 773 |
|
| 774 | rcu_read_lock();
|
| 775 |
|
| 776 | if (!is_in_v2_mode())
|
| 777 | ret = cpuset1_validate_change(cur, trial);
|
| 778 | if (ret)
|
| 779 | goto out;
|
| 780 |
|
| 781 |
|
| 782 | if (cur == &top_cpuset)
|
| 783 | goto out;
|
| 784 |
|
| 785 | par = parent_cs(cur);
|
| 786 |
|
| 787 | |
| 788 | * We can't shrink if we won't have enough room for SCHED_DEADLINE
|
| 789 | * tasks. This check is not done when scheduling is disabled as the
|
| 790 | * users should know what they are doing.
|
| 791 | *
|
| 792 | * For v1, effective_cpus == cpus_allowed & user_xcpus() returns
|
| 793 | * cpus_allowed.
|
| 794 | *
|
| 795 | * For v2, is_cpu_exclusive() & is_sched_load_balance() are true only
|
| 796 | * for non-isolated partition root. At this point, the target
|
| 797 | * effective_cpus isn't computed yet. user_xcpus() is the best
|
| 798 | * approximation.
|
| 799 | *
|
| 800 | * TBD: May need to precompute the real effective_cpus here in case
|
| 801 | * incorrect scheduling of SCHED_DEADLINE tasks in a partition
|
| 802 | * becomes an issue.
|
| 803 |
|
| 804 | ret = -EBUSY;
|
| 805 | if (is_cpu_exclusive(cur) && is_sched_load_balance(cur) &&
|
| 806 | !cpuset_cpumask_can_shrink(cur->effective_cpus, user_xcpus(trial)))
|
| 807 | goto out;
|
| 808 |
|
| 809 | |
| 810 | * If either I or some sibling (!= me) is exclusive, we can't
|
| 811 | * overlap. exclusive_cpus cannot overlap with each other if set.
|
| 812 |
|
| 813 | ret = -EINVAL;
|
| 814 | xcpus_changed = !cpumask_equal(cur->exclusive_cpus, trial->exclusive_cpus);
|
| 815 | cpuset_for_each_child(c, css, par) {
|
| 816 | if (c == cur)
|
| 817 | continue;
|
| 818 | if (cpus_excl_conflict(trial, c, xcpus_changed))
|
| 819 | goto out;
|
| 820 | if (mems_excl_conflict(trial, c))
|
| 821 | goto out;
|
| 822 | }
|
| 823 |
|
| 824 | ret = 0;
|
| 825 | out:
|
| 826 | rcu_read_unlock();
|
| 827 | return ret;
|
| 828 | }
|
| 829 |
|
| 830 | #ifdef CONFIG_SMP
|
| 831 |
|
| 832 | |
| 833 | * generate_sched_domains()
|
| 834 | *
|
| 835 | * This function builds a partial partition of the systems CPUs
|
| 836 | * A 'partial partition' is a set of non-overlapping subsets whose
|
| 837 | * union is a subset of that set.
|
| 838 | * The output of this function needs to be passed to kernel/sched/core.c
|
| 839 | * partition_sched_domains() routine, which will rebuild the scheduler's
|
| 840 | * load balancing domains (sched domains) as specified by that partial
|
| 841 | * partition.
|
| 842 | *
|
| 843 | * See "What is sched_load_balance" in Documentation/admin-guide/cgroup-v1/cpusets.rst
|
| 844 | * for a background explanation of this.
|
| 845 | *
|
| 846 | * Does not return errors, on the theory that the callers of this
|
| 847 | * routine would rather not worry about failures to rebuild sched
|
| 848 | * domains when operating in the severe memory shortage situations
|
| 849 | * that could cause allocation failures below.
|
| 850 | *
|
| 851 | * Must be called with cpuset_mutex held.
|
| 852 | *
|
| 853 | * The three key local variables below are:
|
| 854 | * cp - cpuset pointer, used (together with pos_css) to perform a
|
| 855 | * top-down scan of all cpusets. For our purposes, rebuilding
|
| 856 | * the schedulers sched domains, we can ignore !is_sched_load_
|
| 857 | * balance cpusets.
|
| 858 | * csa - (for CpuSet Array) Array of pointers to all the cpusets
|
| 859 | * that need to be load balanced, for convenient iterative
|
| 860 | * access by the subsequent code that finds the best partition,
|
| 861 | * i.e the set of domains (subsets) of CPUs such that the
|
| 862 | * cpus_allowed of every cpuset marked is_sched_load_balance
|
| 863 | * is a subset of one of these domains, while there are as
|
| 864 | * many such domains as possible, each as small as possible.
|
| 865 | * doms - Conversion of 'csa' to an array of cpumasks, for passing to
|
| 866 | * the kernel/sched/core.c routine partition_sched_domains() in a
|
| 867 | * convenient format, that can be easily compared to the prior
|
| 868 | * value to determine what partition elements (sched domains)
|
| 869 | * were changed (added or removed.)
|
| 870 |
|
| 871 | static int generate_sched_domains(cpumask_var_t **domains,
|
| 872 | struct sched_domain_attr **attributes)
|
| 873 | {
|
| 874 | struct cpuset *cp;
|
| 875 | struct cpuset **csa;
|
| 876 | int i, j;
|
| 877 | cpumask_var_t *doms;
|
| 878 | struct sched_domain_attr *dattr;
|
| 879 | int ndoms = 0;
|
| 880 | struct cgroup_subsys_state *pos_css;
|
| 881 |
|
| 882 | if (!cpuset_v2())
|
| 883 | return cpuset1_generate_sched_domains(domains, attributes);
|
| 884 |
|
| 885 | doms = NULL;
|
| 886 | dattr = NULL;
|
| 887 | csa = NULL;
|
| 888 |
|
| 889 |
|
| 890 | if (cpumask_empty(subpartitions_cpus)) {
|
| 891 | ndoms = 1;
|
| 892 |
|
| 893 | goto generate_doms;
|
| 894 | }
|
| 895 |
|
| 896 | csa = kmalloc_objs(cp, nr_cpusets());
|
| 897 | if (!csa)
|
| 898 | goto done;
|
| 899 |
|
| 900 |
|
| 901 | rcu_read_lock();
|
| 902 | cpuset_for_each_descendant_pre(cp, pos_css, &top_cpuset) {
|
| 903 | |
| 904 | * Only valid partition roots that are not isolated and with
|
| 905 | * non-empty effective_cpus will be saved into csa[].
|
| 906 |
|
| 907 | if ((cp->partition_root_state == PRS_ROOT) &&
|
| 908 | !cpumask_empty(cp->effective_cpus))
|
| 909 | csa[ndoms++] = cp;
|
| 910 |
|
| 911 | |
| 912 | * Skip @cp's subtree if not a partition root and has no
|
| 913 | * exclusive CPUs to be granted to child cpusets.
|
| 914 |
|
| 915 | if (!is_partition_valid(cp) && cpumask_empty(cp->exclusive_cpus))
|
| 916 | pos_css = css_rightmost_descendant(pos_css);
|
| 917 | }
|
| 918 | rcu_read_unlock();
|
| 919 |
|
| 920 | for (i = 0; i < ndoms; i++) {
|
| 921 | for (j = i + 1; j < ndoms; j++) {
|
| 922 | if (cpusets_overlap(csa[i], csa[j]))
|
| 923 | |
| 924 | * Cgroup v2 shouldn't pass down overlapping
|
| 925 | * partition root cpusets.
|
| 926 |
|
| 927 | WARN_ON_ONCE(1);
|
| 928 | }
|
| 929 | }
|
| 930 |
|
| 931 | generate_doms:
|
| 932 | doms = alloc_sched_domains(ndoms);
|
| 933 | if (!doms)
|
| 934 | goto done;
|
| 935 |
|
| 936 | |
| 937 | * The rest of the code, including the scheduler, can deal with
|
| 938 | * dattr==NULL case. No need to abort if alloc fails.
|
| 939 |
|
| 940 | dattr = kmalloc_objs(struct sched_domain_attr, ndoms);
|
| 941 |
|
| 942 | |
| 943 | * Cgroup v2 doesn't support domain attributes, just set all of them
|
| 944 | * to SD_ATTR_INIT. Also non-isolating partition root CPUs are a
|
| 945 | * subset of HK_TYPE_DOMAIN_BOOT housekeeping CPUs.
|
| 946 |
|
| 947 | for (i = 0; i < ndoms; i++) {
|
| 948 | |
| 949 | * The top cpuset may contain some boot time isolated
|
| 950 | * CPUs that need to be excluded from the sched domain.
|
| 951 |
|
| 952 | if (!csa || csa[i] == &top_cpuset)
|
| 953 | cpumask_and(doms[i], top_cpuset.effective_cpus,
|
| 954 | housekeeping_cpumask(HK_TYPE_DOMAIN_BOOT));
|
| 955 | else
|
| 956 | cpumask_copy(doms[i], csa[i]->effective_cpus);
|
| 957 | if (dattr)
|
| 958 | dattr[i] = SD_ATTR_INIT;
|
| 959 | }
|
| 960 |
|
| 961 | done:
|
| 962 | kfree(csa);
|
| 963 |
|
| 964 | |
| 965 | * Fallback to the default domain if kmalloc() failed.
|
| 966 | * See comments in partition_sched_domains().
|
| 967 |
|
| 968 | if (doms == NULL)
|
| 969 | ndoms = 1;
|
| 970 |
|
| 971 | *domains = doms;
|
| 972 | *attributes = dattr;
|
| 973 | return ndoms;
|
| 974 | }
|
| 975 |
|
| 976 | static void dl_update_tasks_root_domain(struct cpuset *cs)
|
| 977 | {
|
| 978 | struct css_task_iter it;
|
| 979 | struct task_struct *task;
|
| 980 |
|
| 981 | if (atomic_read(&cs->nr_deadline_tasks) == 0)
|
| 982 | return;
|
| 983 |
|
| 984 | css_task_iter_start(&cs->css, 0, &it);
|
| 985 |
|
| 986 | while ((task = css_task_iter_next(&it)))
|
| 987 | dl_add_task_root_domain(task);
|
| 988 |
|
| 989 | css_task_iter_end(&it);
|
| 990 | }
|
| 991 |
|
| 992 | void dl_rebuild_rd_accounting(void)
|
| 993 | {
|
| 994 | struct cpuset *cs = NULL;
|
| 995 | struct cgroup_subsys_state *pos_css;
|
| 996 | int cpu;
|
| 997 | u64 cookie = ++dl_cookie;
|
| 998 |
|
| 999 | lockdep_assert_cpuset_lock_held();
|
| 1000 | lockdep_assert_cpus_held();
|
| 1001 | lockdep_assert_held(&sched_domains_mutex);
|
| 1002 |
|
| 1003 | rcu_read_lock();
|
| 1004 |
|
| 1005 | for_each_possible_cpu(cpu) {
|
| 1006 | if (dl_bw_visited(cpu, cookie))
|
| 1007 | continue;
|
| 1008 |
|
| 1009 | dl_clear_root_domain_cpu(cpu);
|
| 1010 | }
|
| 1011 |
|
| 1012 | cpuset_for_each_descendant_pre(cs, pos_css, &top_cpuset) {
|
| 1013 |
|
| 1014 | if (cpumask_empty(cs->effective_cpus)) {
|
| 1015 | pos_css = css_rightmost_descendant(pos_css);
|
| 1016 | continue;
|
| 1017 | }
|
| 1018 |
|
| 1019 | css_get(&cs->css);
|
| 1020 |
|
| 1021 | rcu_read_unlock();
|
| 1022 |
|
| 1023 | dl_update_tasks_root_domain(cs);
|
| 1024 |
|
| 1025 | rcu_read_lock();
|
| 1026 | css_put(&cs->css);
|
| 1027 | }
|
| 1028 | rcu_read_unlock();
|
| 1029 | }
|
| 1030 |
|
| 1031 | |
| 1032 | * Rebuild scheduler domains.
|
| 1033 | *
|
| 1034 | * If the flag 'sched_load_balance' of any cpuset with non-empty
|
| 1035 | * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
|
| 1036 | * which has that flag enabled, or if any cpuset with a non-empty
|
| 1037 | * 'cpus' is removed, then call this routine to rebuild the
|
| 1038 | * scheduler's dynamic sched domains.
|
| 1039 | *
|
| 1040 | * Call with cpuset_mutex held. Takes cpus_read_lock().
|
| 1041 |
|
| 1042 | void rebuild_sched_domains_locked(void)
|
| 1043 | {
|
| 1044 | struct sched_domain_attr *attr;
|
| 1045 | cpumask_var_t *doms;
|
| 1046 | int ndoms;
|
| 1047 | int i;
|
| 1048 |
|
| 1049 | lockdep_assert_cpus_held();
|
| 1050 | lockdep_assert_cpuset_lock_held();
|
| 1051 | force_sd_rebuild = false;
|
| 1052 |
|
| 1053 |
|
| 1054 | ndoms = generate_sched_domains(&doms, &attr);
|
| 1055 |
|
| 1056 | |
| 1057 | * cpuset_hotplug_workfn is invoked synchronously now, thus this
|
| 1058 | * function should not race with CPU hotplug. And the effective CPUs
|
| 1059 | * must not include any offline CPUs. Passing an offline CPU in the
|
| 1060 | * doms to partition_sched_domains() will trigger a kernel panic.
|
| 1061 | *
|
| 1062 | * We perform a final check here: if the doms contains any
|
| 1063 | * offline CPUs, a warning is emitted and we return directly to
|
| 1064 | * prevent the panic.
|
| 1065 |
|
| 1066 | for (i = 0; doms && i < ndoms; i++) {
|
| 1067 | if (WARN_ON_ONCE(!cpumask_subset(doms[i], cpu_active_mask))) {
|
| 1068 | free_sched_domains(doms, ndoms);
|
| 1069 | kfree(attr);
|
| 1070 | return;
|
| 1071 | }
|
| 1072 | }
|
| 1073 |
|
| 1074 |
|
| 1075 | partition_sched_domains(ndoms, doms, attr);
|
| 1076 | }
|
| 1077 | #else
|
| 1078 | void rebuild_sched_domains_locked(void)
|
| 1079 | {
|
| 1080 | }
|
| 1081 | #endif
|
| 1082 |
|
| 1083 | static void rebuild_sched_domains_cpuslocked(void)
|
| 1084 | {
|
| 1085 | mutex_lock(&cpuset_mutex);
|
| 1086 | rebuild_sched_domains_locked();
|
| 1087 | mutex_unlock(&cpuset_mutex);
|
| 1088 | }
|
| 1089 |
|
| 1090 | void rebuild_sched_domains(void)
|
| 1091 | {
|
| 1092 | cpus_read_lock();
|
| 1093 | rebuild_sched_domains_cpuslocked();
|
| 1094 | cpus_read_unlock();
|
| 1095 | }
|
| 1096 |
|
| 1097 | void cpuset_reset_sched_domains(void)
|
| 1098 | {
|
| 1099 | mutex_lock(&cpuset_mutex);
|
| 1100 | partition_sched_domains(1, NULL, NULL);
|
| 1101 | mutex_unlock(&cpuset_mutex);
|
| 1102 | }
|
| 1103 |
|
| 1104 | |
| 1105 | * cpuset_update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
|
| 1106 | * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
|
| 1107 | * @new_cpus: the temp variable for the new effective_cpus mask
|
| 1108 | *
|
| 1109 | * Iterate through each task of @cs updating its cpus_allowed to the
|
| 1110 | * effective cpuset's. As this function is called with cpuset_mutex held,
|
| 1111 | * cpuset membership stays stable.
|
| 1112 | *
|
| 1113 | * For top_cpuset, task_cpu_possible_mask() is used instead of effective_cpus
|
| 1114 | * to make sure all offline CPUs are also included as hotplug code won't
|
| 1115 | * update cpumasks for tasks in top_cpuset.
|
| 1116 | *
|
| 1117 | * As task_cpu_possible_mask() can be task dependent in arm64, we have to
|
| 1118 | * do cpu masking per task instead of doing it once for all.
|
| 1119 |
|
| 1120 | void cpuset_update_tasks_cpumask(struct cpuset *cs, struct cpumask *new_cpus)
|
| 1121 | {
|
| 1122 | struct css_task_iter it;
|
| 1123 | struct task_struct *task;
|
| 1124 | bool top_cs = cs == &top_cpuset;
|
| 1125 |
|
| 1126 | css_task_iter_start(&cs->css, 0, &it);
|
| 1127 | while ((task = css_task_iter_next(&it))) {
|
| 1128 | const struct cpumask *possible_mask = task_cpu_possible_mask(task);
|
| 1129 |
|
| 1130 | if (top_cs) {
|
| 1131 | |
| 1132 | * PF_KTHREAD tasks are handled by housekeeping.
|
| 1133 | * PF_NO_SETAFFINITY tasks are ignored.
|
| 1134 |
|
| 1135 | if (task->flags & (PF_KTHREAD | PF_NO_SETAFFINITY))
|
| 1136 | continue;
|
| 1137 | cpumask_andnot(new_cpus, possible_mask, subpartitions_cpus);
|
| 1138 | } else {
|
| 1139 | cpumask_and(new_cpus, possible_mask, cs->effective_cpus);
|
| 1140 | }
|
| 1141 | set_cpus_allowed_ptr(task, new_cpus);
|
| 1142 | }
|
| 1143 | css_task_iter_end(&it);
|
| 1144 | }
|
| 1145 |
|
| 1146 | |
| 1147 | * compute_effective_cpumask - Compute the effective cpumask of the cpuset
|
| 1148 | * @new_cpus: the temp variable for the new effective_cpus mask
|
| 1149 | * @cs: the cpuset the need to recompute the new effective_cpus mask
|
| 1150 | * @parent: the parent cpuset
|
| 1151 | *
|
| 1152 | * For v2, the parent's effective_cpus is inherited if cpumask is empty.
|
| 1153 | * The result is valid only if the given cpuset isn't a partition root.
|
| 1154 |
|
| 1155 | static void compute_effective_cpumask(struct cpumask *new_cpus,
|
| 1156 | struct cpuset *cs, struct cpuset *parent)
|
| 1157 | {
|
| 1158 | bool has_cpus;
|
| 1159 |
|
| 1160 | has_cpus = cpumask_and(new_cpus, cs->cpus_allowed, parent->effective_cpus);
|
| 1161 | if (!has_cpus && is_in_v2_mode())
|
| 1162 | cpumask_copy(new_cpus, parent->effective_cpus);
|
| 1163 | }
|
| 1164 |
|
| 1165 | |
| 1166 | * compute_effective_nodemask - Compute the effective nodemask of the cpuset
|
| 1167 | * @new_mems: the temp variable for the new effective_mems mask
|
| 1168 | * @cs: the cpuset the need to recompute the new effective_mems mask
|
| 1169 | * @parent: the parent cpuset
|
| 1170 | *
|
| 1171 | * For v2, the parent's effective_mems is inherited if nodemask is empty.
|
| 1172 |
|
| 1173 | static void compute_effective_nodemask(nodemask_t *new_mems,
|
| 1174 | struct cpuset *cs, struct cpuset *parent)
|
| 1175 | {
|
| 1176 | bool has_mems;
|
| 1177 |
|
| 1178 | has_mems = nodes_and(*new_mems, cs->mems_allowed, parent->effective_mems);
|
| 1179 | if (!has_mems && is_in_v2_mode())
|
| 1180 | nodes_copy(*new_mems, parent->effective_mems);
|
| 1181 | }
|
| 1182 |
|
| 1183 | |
| 1184 | * Commands for update_parent_effective_cpumask
|
| 1185 |
|
| 1186 | enum partition_cmd {
|
| 1187 | partcmd_enable,
|
| 1188 | partcmd_enablei,
|
| 1189 | partcmd_disable,
|
| 1190 | partcmd_update,
|
| 1191 | partcmd_invalidate,
|
| 1192 | };
|
| 1193 |
|
| 1194 | static void update_sibling_cpumasks(struct cpuset *parent, struct cpuset *cs,
|
| 1195 | struct tmpmasks *tmp);
|
| 1196 |
|
| 1197 | |
| 1198 | * Update partition exclusive flag
|
| 1199 | *
|
| 1200 | * Return: 0 if successful, an error code otherwise
|
| 1201 |
|
| 1202 | static int update_partition_exclusive_flag(struct cpuset *cs, int new_prs)
|
| 1203 | {
|
| 1204 | bool exclusive = (new_prs > PRS_MEMBER);
|
| 1205 |
|
| 1206 | if (exclusive && !is_cpu_exclusive(cs)) {
|
| 1207 | if (cpuset_update_flag(CS_CPU_EXCLUSIVE, cs, 1))
|
| 1208 | return PERR_NOTEXCL;
|
| 1209 | } else if (!exclusive && is_cpu_exclusive(cs)) {
|
| 1210 |
|
| 1211 | cpuset_update_flag(CS_CPU_EXCLUSIVE, cs, 0);
|
| 1212 | }
|
| 1213 | return 0;
|
| 1214 | }
|
| 1215 |
|
| 1216 | |
| 1217 | * Update partition load balance flag and/or rebuild sched domain
|
| 1218 | *
|
| 1219 | * Changing load balance flag will automatically call
|
| 1220 | * rebuild_sched_domains_locked().
|
| 1221 | * This function is for cgroup v2 only.
|
| 1222 |
|
| 1223 | static void update_partition_sd_lb(struct cpuset *cs, int old_prs)
|
| 1224 | {
|
| 1225 | int new_prs = cs->partition_root_state;
|
| 1226 | bool rebuild_domains = (new_prs > 0) || (old_prs > 0);
|
| 1227 | bool new_lb;
|
| 1228 |
|
| 1229 | |
| 1230 | * If cs is not a valid partition root, the load balance state
|
| 1231 | * will follow its parent.
|
| 1232 |
|
| 1233 | if (new_prs > 0) {
|
| 1234 | new_lb = (new_prs != PRS_ISOLATED);
|
| 1235 | } else {
|
| 1236 | new_lb = is_sched_load_balance(parent_cs(cs));
|
| 1237 | }
|
| 1238 | if (new_lb != !!is_sched_load_balance(cs)) {
|
| 1239 | rebuild_domains = true;
|
| 1240 | if (new_lb)
|
| 1241 | set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
|
| 1242 | else
|
| 1243 | clear_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
|
| 1244 | }
|
| 1245 |
|
| 1246 | if (rebuild_domains)
|
| 1247 | cpuset_force_rebuild();
|
| 1248 | }
|
| 1249 |
|
| 1250 | |
| 1251 | * tasks_nocpu_error - Return true if tasks will have no effective_cpus
|
| 1252 |
|
| 1253 | static bool tasks_nocpu_error(struct cpuset *parent, struct cpuset *cs,
|
| 1254 | struct cpumask *xcpus)
|
| 1255 | {
|
| 1256 | |
| 1257 | * A populated partition (cs or parent) can't have empty effective_cpus
|
| 1258 |
|
| 1259 | return (cpumask_subset(parent->effective_cpus, xcpus) &&
|
| 1260 | partition_is_populated(parent, cs)) ||
|
| 1261 | (!cpumask_intersects(xcpus, cpu_active_mask) &&
|
| 1262 | partition_is_populated(cs, NULL));
|
| 1263 | }
|
| 1264 |
|
| 1265 | static void reset_partition_data(struct cpuset *cs)
|
| 1266 | {
|
| 1267 | struct cpuset *parent = parent_cs(cs);
|
| 1268 |
|
| 1269 | if (!cpuset_v2())
|
| 1270 | return;
|
| 1271 |
|
| 1272 | lockdep_assert_held(&callback_lock);
|
| 1273 |
|
| 1274 | if (cpumask_empty(cs->exclusive_cpus)) {
|
| 1275 | cpumask_clear(cs->effective_xcpus);
|
| 1276 | if (is_cpu_exclusive(cs))
|
| 1277 | clear_bit(CS_CPU_EXCLUSIVE, &cs->flags);
|
| 1278 | }
|
| 1279 | if (!cpumask_and(cs->effective_cpus, parent->effective_cpus, cs->cpus_allowed))
|
| 1280 | cpumask_copy(cs->effective_cpus, parent->effective_cpus);
|
| 1281 | }
|
| 1282 |
|
| 1283 |
|
| 1284 | static bool isolated_cpu_update(int new_prs, int cpu)
|
| 1285 | {
|
| 1286 | lockdep_assert_held(&callback_lock);
|
| 1287 | lockdep_assert_held(&cpuset_mutex);
|
| 1288 |
|
| 1289 | if (new_prs == PRS_ISOLATED) {
|
| 1290 | if (cpumask_test_cpu(cpu, isolated_cpus))
|
| 1291 | return false;
|
| 1292 | cpumask_set_cpu(cpu, isolated_cpus);
|
| 1293 | return true;
|
| 1294 | }
|
| 1295 |
|
| 1296 |
|
| 1297 | if (!cpumask_test_cpu(cpu,
|
| 1298 | housekeeping_cpumask(HK_TYPE_DOMAIN_BOOT)) ||
|
| 1299 | !cpumask_test_cpu(cpu, isolated_cpus))
|
| 1300 | return false;
|
| 1301 | cpumask_clear_cpu(cpu, isolated_cpus);
|
| 1302 | return true;
|
| 1303 | }
|
| 1304 |
|
| 1305 | |
| 1306 | * isolated_cpus_update - Update the isolated_cpus mask
|
| 1307 | * @old_prs: old partition_root_state
|
| 1308 | * @new_prs: new partition_root_state
|
| 1309 | * @xcpus: exclusive CPUs with state change
|
| 1310 |
|
| 1311 | static void isolated_cpus_update(int old_prs, int new_prs, struct cpumask *xcpus)
|
| 1312 | {
|
| 1313 | bool updated = false;
|
| 1314 | int cpu;
|
| 1315 |
|
| 1316 | WARN_ON_ONCE(old_prs == new_prs);
|
| 1317 | lockdep_assert_held(&callback_lock);
|
| 1318 | lockdep_assert_held(&cpuset_mutex);
|
| 1319 | for_each_cpu(cpu, xcpus)
|
| 1320 | updated |= isolated_cpu_update(new_prs, cpu);
|
| 1321 | if (updated)
|
| 1322 | update_housekeeping = true;
|
| 1323 | }
|
| 1324 |
|
| 1325 | |
| 1326 | * partition_xcpus_add - Add new exclusive CPUs to partition
|
| 1327 | * @new_prs: new partition_root_state
|
| 1328 | * @parent: parent cpuset
|
| 1329 | * @xcpus: exclusive CPUs to be added
|
| 1330 | *
|
| 1331 | * Remote partition if parent == NULL
|
| 1332 |
|
| 1333 | static void partition_xcpus_add(int new_prs, struct cpuset *parent,
|
| 1334 | struct cpumask *xcpus)
|
| 1335 | {
|
| 1336 | WARN_ON_ONCE(new_prs < 0);
|
| 1337 | lockdep_assert_held(&callback_lock);
|
| 1338 | if (!parent)
|
| 1339 | parent = &top_cpuset;
|
| 1340 |
|
| 1341 |
|
| 1342 | if (parent == &top_cpuset)
|
| 1343 | cpumask_or(subpartitions_cpus, subpartitions_cpus, xcpus);
|
| 1344 |
|
| 1345 | if (new_prs != parent->partition_root_state)
|
| 1346 | isolated_cpus_update(parent->partition_root_state, new_prs,
|
| 1347 | xcpus);
|
| 1348 |
|
| 1349 | cpumask_andnot(parent->effective_cpus, parent->effective_cpus, xcpus);
|
| 1350 | }
|
| 1351 |
|
| 1352 | |
| 1353 | * partition_xcpus_del - Remove exclusive CPUs from partition
|
| 1354 | * @old_prs: old partition_root_state
|
| 1355 | * @parent: parent cpuset
|
| 1356 | * @xcpus: exclusive CPUs to be removed
|
| 1357 | *
|
| 1358 | * Remote partition if parent == NULL
|
| 1359 |
|
| 1360 | static void partition_xcpus_del(int old_prs, struct cpuset *parent,
|
| 1361 | struct cpumask *xcpus)
|
| 1362 | {
|
| 1363 | WARN_ON_ONCE(old_prs < 0);
|
| 1364 | lockdep_assert_held(&callback_lock);
|
| 1365 | if (!parent)
|
| 1366 | parent = &top_cpuset;
|
| 1367 |
|
| 1368 | if (parent == &top_cpuset)
|
| 1369 | cpumask_andnot(subpartitions_cpus, subpartitions_cpus, xcpus);
|
| 1370 |
|
| 1371 | if (old_prs != parent->partition_root_state)
|
| 1372 | isolated_cpus_update(old_prs, parent->partition_root_state,
|
| 1373 | xcpus);
|
| 1374 |
|
| 1375 | cpumask_or(parent->effective_cpus, parent->effective_cpus, xcpus);
|
| 1376 | cpumask_and(parent->effective_cpus, parent->effective_cpus, cpu_active_mask);
|
| 1377 | }
|
| 1378 |
|
| 1379 | |
| 1380 | * isolated_cpus_can_update - check for isolated & nohz_full conflicts
|
| 1381 | * @add_cpus: cpu mask for cpus that are going to be isolated
|
| 1382 | * @del_cpus: cpu mask for cpus that are no longer isolated, can be NULL
|
| 1383 | * Return: false if there is conflict, true otherwise
|
| 1384 | *
|
| 1385 | * If nohz_full is enabled and we have isolated CPUs, their combination must
|
| 1386 | * still leave housekeeping CPUs.
|
| 1387 | *
|
| 1388 | * TBD: Should consider merging this function into
|
| 1389 | * prstate_housekeeping_conflict().
|
| 1390 |
|
| 1391 | static bool isolated_cpus_can_update(struct cpumask *add_cpus,
|
| 1392 | struct cpumask *del_cpus)
|
| 1393 | {
|
| 1394 | cpumask_var_t full_hk_cpus;
|
| 1395 | int res = true;
|
| 1396 |
|
| 1397 | if (!housekeeping_enabled(HK_TYPE_KERNEL_NOISE))
|
| 1398 | return true;
|
| 1399 |
|
| 1400 | if (del_cpus && cpumask_weight_and(del_cpus,
|
| 1401 | housekeeping_cpumask(HK_TYPE_KERNEL_NOISE)))
|
| 1402 | return true;
|
| 1403 |
|
| 1404 | if (!alloc_cpumask_var(&full_hk_cpus, GFP_KERNEL))
|
| 1405 | return false;
|
| 1406 |
|
| 1407 | cpumask_and(full_hk_cpus, housekeeping_cpumask(HK_TYPE_KERNEL_NOISE),
|
| 1408 | housekeeping_cpumask(HK_TYPE_DOMAIN));
|
| 1409 | cpumask_andnot(full_hk_cpus, full_hk_cpus, isolated_cpus);
|
| 1410 | cpumask_and(full_hk_cpus, full_hk_cpus, cpu_active_mask);
|
| 1411 | if (!cpumask_weight_andnot(full_hk_cpus, add_cpus))
|
| 1412 | res = false;
|
| 1413 |
|
| 1414 | free_cpumask_var(full_hk_cpus);
|
| 1415 | return res;
|
| 1416 | }
|
| 1417 |
|
| 1418 | |
| 1419 | * prstate_housekeeping_conflict - check for partition & housekeeping conflicts
|
| 1420 | * @prstate: partition root state to be checked
|
| 1421 | * @new_cpus: cpu mask
|
| 1422 | * Return: true if there is conflict, false otherwise
|
| 1423 | *
|
| 1424 | * CPUs outside of HK_TYPE_DOMAIN_BOOT, if defined, can only be used in an
|
| 1425 | * isolated partition.
|
| 1426 |
|
| 1427 | static bool prstate_housekeeping_conflict(int prstate, struct cpumask *new_cpus)
|
| 1428 | {
|
| 1429 | if (!housekeeping_enabled(HK_TYPE_DOMAIN_BOOT))
|
| 1430 | return false;
|
| 1431 |
|
| 1432 | if ((prstate != PRS_ISOLATED) &&
|
| 1433 | !cpumask_subset(new_cpus, housekeeping_cpumask(HK_TYPE_DOMAIN_BOOT)))
|
| 1434 | return true;
|
| 1435 |
|
| 1436 | return false;
|
| 1437 | }
|
| 1438 |
|
| 1439 | |
| 1440 | * cpuset_update_sd_hk_unlock - Rebuild sched domains, update HK & unlock
|
| 1441 | *
|
| 1442 | * Update housekeeping cpumasks and rebuild sched domains if necessary and
|
| 1443 | * then do a cpuset_full_unlock().
|
| 1444 | * This should be called at the end of cpuset operation.
|
| 1445 |
|
| 1446 | static void cpuset_update_sd_hk_unlock(void)
|
| 1447 | __releases(&cpuset_mutex)
|
| 1448 | __releases(&cpuset_top_mutex)
|
| 1449 | {
|
| 1450 |
|
| 1451 | if (force_sd_rebuild)
|
| 1452 | rebuild_sched_domains_locked();
|
| 1453 |
|
| 1454 | if (update_housekeeping) {
|
| 1455 | update_housekeeping = false;
|
| 1456 | cpumask_copy(isolated_hk_cpus, isolated_cpus);
|
| 1457 |
|
| 1458 | |
| 1459 | * housekeeping_update() is now called without holding
|
| 1460 | * cpus_read_lock and cpuset_mutex. Only cpuset_top_mutex
|
| 1461 | * is still being held for mutual exclusion.
|
| 1462 |
|
| 1463 | mutex_unlock(&cpuset_mutex);
|
| 1464 | cpus_read_unlock();
|
| 1465 | WARN_ON_ONCE(housekeeping_update(isolated_hk_cpus));
|
| 1466 | mutex_unlock(&cpuset_top_mutex);
|
| 1467 | } else {
|
| 1468 | cpuset_full_unlock();
|
| 1469 | }
|
| 1470 | }
|
| 1471 |
|
| 1472 | |
| 1473 | * Work function to invoke cpuset_update_sd_hk_unlock()
|
| 1474 |
|
| 1475 | static void hk_sd_workfn(struct work_struct *work)
|
| 1476 | {
|
| 1477 | cpuset_full_lock();
|
| 1478 | cpuset_update_sd_hk_unlock();
|
| 1479 | }
|
| 1480 |
|
| 1481 | |
| 1482 | * rm_siblings_excl_cpus - Remove exclusive CPUs that are used by sibling cpusets
|
| 1483 | * @parent: Parent cpuset containing all siblings
|
| 1484 | * @cs: Current cpuset (will be skipped)
|
| 1485 | * @excpus: exclusive effective CPU mask to modify
|
| 1486 | *
|
| 1487 | * This function ensures the given @excpus mask doesn't include any CPUs that
|
| 1488 | * are exclusively allocated to sibling cpusets. It walks through all siblings
|
| 1489 | * of @cs under @parent and removes their exclusive CPUs from @excpus.
|
| 1490 |
|
| 1491 | static int rm_siblings_excl_cpus(struct cpuset *parent, struct cpuset *cs,
|
| 1492 | struct cpumask *excpus)
|
| 1493 | {
|
| 1494 | struct cgroup_subsys_state *css;
|
| 1495 | struct cpuset *sibling;
|
| 1496 | int retval = 0;
|
| 1497 |
|
| 1498 | if (cpumask_empty(excpus))
|
| 1499 | return 0;
|
| 1500 |
|
| 1501 | |
| 1502 | * Remove exclusive CPUs from siblings
|
| 1503 |
|
| 1504 | rcu_read_lock();
|
| 1505 | cpuset_for_each_child(sibling, css, parent) {
|
| 1506 | struct cpumask *sibling_xcpus;
|
| 1507 |
|
| 1508 | if (sibling == cs)
|
| 1509 | continue;
|
| 1510 |
|
| 1511 | |
| 1512 | * If exclusive_cpus is defined, effective_xcpus will always
|
| 1513 | * be a subset. Otherwise, effective_xcpus will only be set
|
| 1514 | * in a valid partition root.
|
| 1515 |
|
| 1516 | sibling_xcpus = cpumask_empty(sibling->exclusive_cpus)
|
| 1517 | ? sibling->effective_xcpus
|
| 1518 | : sibling->exclusive_cpus;
|
| 1519 |
|
| 1520 | if (cpumask_intersects(excpus, sibling_xcpus)) {
|
| 1521 | cpumask_andnot(excpus, excpus, sibling_xcpus);
|
| 1522 | retval++;
|
| 1523 | }
|
| 1524 | }
|
| 1525 | rcu_read_unlock();
|
| 1526 |
|
| 1527 | return retval;
|
| 1528 | }
|
| 1529 |
|
| 1530 | |
| 1531 | * compute_excpus - compute effective exclusive CPUs
|
| 1532 | * @cs: cpuset
|
| 1533 | * @xcpus: effective exclusive CPUs value to be set
|
| 1534 | * Return: 0 if there is no sibling conflict, > 0 otherwise
|
| 1535 | *
|
| 1536 | * If exclusive_cpus isn't explicitly set , we have to scan the sibling cpusets
|
| 1537 | * and exclude their exclusive_cpus or effective_xcpus as well.
|
| 1538 |
|
| 1539 | static int compute_excpus(struct cpuset *cs, struct cpumask *excpus)
|
| 1540 | {
|
| 1541 | struct cpuset *parent = parent_cs(cs);
|
| 1542 |
|
| 1543 | cpumask_and(excpus, user_xcpus(cs), parent->effective_xcpus);
|
| 1544 |
|
| 1545 | if (!cpumask_empty(cs->exclusive_cpus))
|
| 1546 | return 0;
|
| 1547 |
|
| 1548 | return rm_siblings_excl_cpus(parent, cs, excpus);
|
| 1549 | }
|
| 1550 |
|
| 1551 | |
| 1552 | * compute_trialcs_excpus - Compute effective exclusive CPUs for a trial cpuset
|
| 1553 | * @trialcs: The trial cpuset containing the proposed new configuration
|
| 1554 | * @cs: The original cpuset that the trial configuration is based on
|
| 1555 | * Return: 0 if successful with no sibling conflict, >0 if a conflict is found
|
| 1556 | *
|
| 1557 | * Computes the effective_xcpus for a trial configuration. @cs is provided to represent
|
| 1558 | * the real cs.
|
| 1559 |
|
| 1560 | static int compute_trialcs_excpus(struct cpuset *trialcs, struct cpuset *cs)
|
| 1561 | {
|
| 1562 | struct cpuset *parent = parent_cs(trialcs);
|
| 1563 | struct cpumask *excpus = trialcs->effective_xcpus;
|
| 1564 |
|
| 1565 |
|
| 1566 | if (cs_is_member(cs))
|
| 1567 | cpumask_and(excpus, trialcs->exclusive_cpus,
|
| 1568 | parent->effective_xcpus);
|
| 1569 | else
|
| 1570 | cpumask_and(excpus, user_xcpus(trialcs), parent->effective_xcpus);
|
| 1571 |
|
| 1572 | return rm_siblings_excl_cpus(parent, cs, excpus);
|
| 1573 | }
|
| 1574 |
|
| 1575 | static inline bool is_remote_partition(struct cpuset *cs)
|
| 1576 | {
|
| 1577 | return cs->remote_partition;
|
| 1578 | }
|
| 1579 |
|
| 1580 | static inline bool is_local_partition(struct cpuset *cs)
|
| 1581 | {
|
| 1582 | return is_partition_valid(cs) && !is_remote_partition(cs);
|
| 1583 | }
|
| 1584 |
|
| 1585 | |
| 1586 | * remote_partition_enable - Enable current cpuset as a remote partition root
|
| 1587 | * @cs: the cpuset to update
|
| 1588 | * @new_prs: new partition_root_state
|
| 1589 | * @tmp: temporary masks
|
| 1590 | * Return: 0 if successful, errcode if error
|
| 1591 | *
|
| 1592 | * Enable the current cpuset to become a remote partition root taking CPUs
|
| 1593 | * directly from the top cpuset. cpuset_mutex must be held by the caller.
|
| 1594 |
|
| 1595 | static int remote_partition_enable(struct cpuset *cs, int new_prs,
|
| 1596 | struct tmpmasks *tmp)
|
| 1597 | {
|
| 1598 | |
| 1599 | * The user must have sysadmin privilege.
|
| 1600 |
|
| 1601 | if (!capable(CAP_SYS_ADMIN))
|
| 1602 | return PERR_ACCESS;
|
| 1603 |
|
| 1604 | |
| 1605 | * The requested exclusive_cpus must not be allocated to other
|
| 1606 | * partitions and it can't use up all the root's effective_cpus.
|
| 1607 | *
|
| 1608 | * The effective_xcpus mask can contain offline CPUs, but there must
|
| 1609 | * be at least one or more online CPUs present before it can be enabled.
|
| 1610 | *
|
| 1611 | * Note that creating a remote partition with any local partition root
|
| 1612 | * above it or remote partition root underneath it is not allowed.
|
| 1613 |
|
| 1614 | compute_excpus(cs, tmp->new_cpus);
|
| 1615 | if (!cpumask_intersects(tmp->new_cpus, cpu_active_mask) ||
|
| 1616 | cpumask_subset(top_cpuset.effective_cpus, tmp->new_cpus))
|
| 1617 | return PERR_INVCPUS;
|
| 1618 | if (cpumask_intersects(tmp->new_cpus, subpartitions_cpus))
|
| 1619 | return PERR_NOCPUS;
|
| 1620 | if (((new_prs == PRS_ISOLATED) &&
|
| 1621 | !isolated_cpus_can_update(tmp->new_cpus, NULL)) ||
|
| 1622 | prstate_housekeeping_conflict(new_prs, tmp->new_cpus))
|
| 1623 | return PERR_HKEEPING;
|
| 1624 |
|
| 1625 | spin_lock_irq(&callback_lock);
|
| 1626 | partition_xcpus_add(new_prs, NULL, tmp->new_cpus);
|
| 1627 | cs->remote_partition = true;
|
| 1628 | cpumask_copy(cs->effective_xcpus, tmp->new_cpus);
|
| 1629 | spin_unlock_irq(&callback_lock);
|
| 1630 | cpuset_force_rebuild();
|
| 1631 | WRITE_ONCE(cs->prs_err, 0);
|
| 1632 |
|
| 1633 | |
| 1634 | * Propagate changes in top_cpuset's effective_cpus down the hierarchy.
|
| 1635 |
|
| 1636 | cpuset_update_tasks_cpumask(&top_cpuset, tmp->new_cpus);
|
| 1637 | update_sibling_cpumasks(&top_cpuset, NULL, tmp);
|
| 1638 | return 0;
|
| 1639 | }
|
| 1640 |
|
| 1641 | |
| 1642 | * remote_partition_disable - Remove current cpuset from remote partition list
|
| 1643 | * @cs: the cpuset to update
|
| 1644 | * @tmp: temporary masks
|
| 1645 | *
|
| 1646 | * The effective_cpus is also updated.
|
| 1647 | *
|
| 1648 | * cpuset_mutex must be held by the caller.
|
| 1649 |
|
| 1650 | static void remote_partition_disable(struct cpuset *cs, struct tmpmasks *tmp)
|
| 1651 | {
|
| 1652 | WARN_ON_ONCE(!is_remote_partition(cs));
|
| 1653 | |
| 1654 | * When a CPU is offlined, top_cpuset may end up with no available CPUs,
|
| 1655 | * which should clear subpartitions_cpus. We should not emit a warning for this
|
| 1656 | * scenario: the hierarchy is updated from top to bottom, so subpartitions_cpus
|
| 1657 | * may already be cleared when disabling the partition.
|
| 1658 |
|
| 1659 | WARN_ON_ONCE(!cpumask_subset(cs->effective_xcpus, subpartitions_cpus) &&
|
| 1660 | !cpumask_empty(subpartitions_cpus));
|
| 1661 |
|
| 1662 | spin_lock_irq(&callback_lock);
|
| 1663 | cs->remote_partition = false;
|
| 1664 | partition_xcpus_del(cs->partition_root_state, NULL, cs->effective_xcpus);
|
| 1665 | if (cs->prs_err)
|
| 1666 | cs->partition_root_state = -cs->partition_root_state;
|
| 1667 | else
|
| 1668 | cs->partition_root_state = PRS_MEMBER;
|
| 1669 |
|
| 1670 |
|
| 1671 | compute_excpus(cs, cs->effective_xcpus);
|
| 1672 | reset_partition_data(cs);
|
| 1673 | spin_unlock_irq(&callback_lock);
|
| 1674 | cpuset_force_rebuild();
|
| 1675 |
|
| 1676 | |
| 1677 | * Propagate changes in top_cpuset's effective_cpus down the hierarchy.
|
| 1678 |
|
| 1679 | cpuset_update_tasks_cpumask(&top_cpuset, tmp->new_cpus);
|
| 1680 | update_sibling_cpumasks(&top_cpuset, NULL, tmp);
|
| 1681 | }
|
| 1682 |
|
| 1683 | |
| 1684 | * remote_cpus_update - cpus_exclusive change of remote partition
|
| 1685 | * @cs: the cpuset to be updated
|
| 1686 | * @xcpus: the new exclusive_cpus mask, if non-NULL
|
| 1687 | * @excpus: the new effective_xcpus mask
|
| 1688 | * @tmp: temporary masks
|
| 1689 | *
|
| 1690 | * top_cpuset and subpartitions_cpus will be updated or partition can be
|
| 1691 | * invalidated.
|
| 1692 |
|
| 1693 | static void remote_cpus_update(struct cpuset *cs, struct cpumask *xcpus,
|
| 1694 | struct cpumask *excpus, struct tmpmasks *tmp)
|
| 1695 | {
|
| 1696 | bool adding, deleting;
|
| 1697 | int prs = cs->partition_root_state;
|
| 1698 |
|
| 1699 | if (WARN_ON_ONCE(!is_remote_partition(cs)))
|
| 1700 | return;
|
| 1701 |
|
| 1702 | WARN_ON_ONCE(!cpumask_subset(cs->effective_xcpus, subpartitions_cpus));
|
| 1703 |
|
| 1704 | if (cpumask_empty(excpus)) {
|
| 1705 | WRITE_ONCE(cs->prs_err, PERR_CPUSEMPTY);
|
| 1706 | goto invalidate;
|
| 1707 | }
|
| 1708 |
|
| 1709 | adding = cpumask_andnot(tmp->addmask, excpus, cs->effective_xcpus);
|
| 1710 | deleting = cpumask_andnot(tmp->delmask, cs->effective_xcpus, excpus);
|
| 1711 |
|
| 1712 | |
| 1713 | * Additions of remote CPUs is only allowed if those CPUs are
|
| 1714 | * not allocated to other partitions and there are effective_cpus
|
| 1715 | * left in the top cpuset.
|
| 1716 |
|
| 1717 | if (adding) {
|
| 1718 | WARN_ON_ONCE(cpumask_intersects(tmp->addmask, subpartitions_cpus));
|
| 1719 | if (!capable(CAP_SYS_ADMIN))
|
| 1720 | WRITE_ONCE(cs->prs_err, PERR_ACCESS);
|
| 1721 | else if (cpumask_intersects(tmp->addmask, subpartitions_cpus) ||
|
| 1722 | cpumask_subset(top_cpuset.effective_cpus, tmp->addmask))
|
| 1723 | WRITE_ONCE(cs->prs_err, PERR_NOCPUS);
|
| 1724 | else if ((prs == PRS_ISOLATED) &&
|
| 1725 | !isolated_cpus_can_update(tmp->addmask, tmp->delmask))
|
| 1726 | WRITE_ONCE(cs->prs_err, PERR_HKEEPING);
|
| 1727 | if (cs->prs_err)
|
| 1728 | goto invalidate;
|
| 1729 | }
|
| 1730 |
|
| 1731 | spin_lock_irq(&callback_lock);
|
| 1732 | if (adding)
|
| 1733 | partition_xcpus_add(prs, NULL, tmp->addmask);
|
| 1734 | if (deleting)
|
| 1735 | partition_xcpus_del(prs, NULL, tmp->delmask);
|
| 1736 | |
| 1737 | * Need to update effective_xcpus and exclusive_cpus now as
|
| 1738 | * update_sibling_cpumasks() below may iterate back to the same cs.
|
| 1739 |
|
| 1740 | cpumask_copy(cs->effective_xcpus, excpus);
|
| 1741 | if (xcpus)
|
| 1742 | cpumask_copy(cs->exclusive_cpus, xcpus);
|
| 1743 | spin_unlock_irq(&callback_lock);
|
| 1744 | if (adding || deleting)
|
| 1745 | cpuset_force_rebuild();
|
| 1746 |
|
| 1747 | |
| 1748 | * Propagate changes in top_cpuset's effective_cpus down the hierarchy.
|
| 1749 |
|
| 1750 | cpuset_update_tasks_cpumask(&top_cpuset, tmp->new_cpus);
|
| 1751 | update_sibling_cpumasks(&top_cpuset, NULL, tmp);
|
| 1752 | return;
|
| 1753 |
|
| 1754 | invalidate:
|
| 1755 | remote_partition_disable(cs, tmp);
|
| 1756 | }
|
| 1757 |
|
| 1758 | |
| 1759 | * update_parent_effective_cpumask - update effective_cpus mask of parent cpuset
|
| 1760 | * @cs: The cpuset that requests change in partition root state
|
| 1761 | * @cmd: Partition root state change command
|
| 1762 | * @newmask: Optional new cpumask for partcmd_update
|
| 1763 | * @tmp: Temporary addmask and delmask
|
| 1764 | * Return: 0 or a partition root state error code
|
| 1765 | *
|
| 1766 | * For partcmd_enable*, the cpuset is being transformed from a non-partition
|
| 1767 | * root to a partition root. The effective_xcpus (cpus_allowed if
|
| 1768 | * effective_xcpus not set) mask of the given cpuset will be taken away from
|
| 1769 | * parent's effective_cpus. The function will return 0 if all the CPUs listed
|
| 1770 | * in effective_xcpus can be granted or an error code will be returned.
|
| 1771 | *
|
| 1772 | * For partcmd_disable, the cpuset is being transformed from a partition
|
| 1773 | * root back to a non-partition root. Any CPUs in effective_xcpus will be
|
| 1774 | * given back to parent's effective_cpus. 0 will always be returned.
|
| 1775 | *
|
| 1776 | * For partcmd_update, if the optional newmask is specified, the cpu list is
|
| 1777 | * to be changed from effective_xcpus to newmask. Otherwise, effective_xcpus is
|
| 1778 | * assumed to remain the same. The cpuset should either be a valid or invalid
|
| 1779 | * partition root. The partition root state may change from valid to invalid
|
| 1780 | * or vice versa. An error code will be returned if transitioning from
|
| 1781 | * invalid to valid violates the exclusivity rule.
|
| 1782 | *
|
| 1783 | * For partcmd_invalidate, the current partition will be made invalid.
|
| 1784 | *
|
| 1785 | * The partcmd_enable* and partcmd_disable commands are used by
|
| 1786 | * update_prstate(). An error code may be returned and the caller will check
|
| 1787 | * for error.
|
| 1788 | *
|
| 1789 | * The partcmd_update command is used by update_cpumasks_hier() with newmask
|
| 1790 | * NULL and update_cpumask() with newmask set. The partcmd_invalidate is used
|
| 1791 | * by update_cpumask() with NULL newmask. In both cases, the callers won't
|
| 1792 | * check for error and so partition_root_state and prs_err will be updated
|
| 1793 | * directly.
|
| 1794 |
|
| 1795 | static int update_parent_effective_cpumask(struct cpuset *cs, int cmd,
|
| 1796 | struct cpumask *newmask,
|
| 1797 | struct tmpmasks *tmp)
|
| 1798 | {
|
| 1799 | struct cpuset *parent = parent_cs(cs);
|
| 1800 | int adding;
|
| 1801 | int deleting;
|
| 1802 | int old_prs, new_prs;
|
| 1803 | int part_error = PERR_NONE;
|
| 1804 | struct cpumask *xcpus = user_xcpus(cs);
|
| 1805 | int parent_prs = parent->partition_root_state;
|
| 1806 | bool nocpu;
|
| 1807 |
|
| 1808 | lockdep_assert_cpuset_lock_held();
|
| 1809 | WARN_ON_ONCE(is_remote_partition(cs));
|
| 1810 |
|
| 1811 | |
| 1812 | * new_prs will only be changed for the partcmd_update and
|
| 1813 | * partcmd_invalidate commands.
|
| 1814 |
|
| 1815 | adding = deleting = false;
|
| 1816 | old_prs = new_prs = cs->partition_root_state;
|
| 1817 |
|
| 1818 | if (cmd == partcmd_invalidate) {
|
| 1819 | if (is_partition_invalid(cs))
|
| 1820 | return 0;
|
| 1821 |
|
| 1822 | |
| 1823 | * Make the current partition invalid.
|
| 1824 |
|
| 1825 | if (is_partition_valid(parent))
|
| 1826 | adding = cpumask_and(tmp->addmask,
|
| 1827 | cs->effective_xcpus,
|
| 1828 | parent->effective_xcpus);
|
| 1829 | if (old_prs > 0)
|
| 1830 | new_prs = -old_prs;
|
| 1831 |
|
| 1832 | goto write_error;
|
| 1833 | }
|
| 1834 |
|
| 1835 | |
| 1836 | * The parent must be a partition root.
|
| 1837 | * The new cpumask, if present, or the current cpus_allowed must
|
| 1838 | * not be empty.
|
| 1839 |
|
| 1840 | if (!is_partition_valid(parent)) {
|
| 1841 | return is_partition_invalid(parent)
|
| 1842 | ? PERR_INVPARENT : PERR_NOTPART;
|
| 1843 | }
|
| 1844 | if (!newmask && xcpus_empty(cs))
|
| 1845 | return PERR_CPUSEMPTY;
|
| 1846 |
|
| 1847 | nocpu = tasks_nocpu_error(parent, cs, xcpus);
|
| 1848 |
|
| 1849 | if ((cmd == partcmd_enable) || (cmd == partcmd_enablei)) {
|
| 1850 | |
| 1851 | * Need to call compute_excpus() in case
|
| 1852 | * exclusive_cpus not set. Sibling conflict should only happen
|
| 1853 | * if exclusive_cpus isn't set.
|
| 1854 |
|
| 1855 | xcpus = tmp->delmask;
|
| 1856 | if (compute_excpus(cs, xcpus))
|
| 1857 | WARN_ON_ONCE(!cpumask_empty(cs->exclusive_cpus));
|
| 1858 | new_prs = (cmd == partcmd_enable) ? PRS_ROOT : PRS_ISOLATED;
|
| 1859 |
|
| 1860 | |
| 1861 | * Enabling partition root is not allowed if its
|
| 1862 | * effective_xcpus is empty.
|
| 1863 |
|
| 1864 | if (cpumask_empty(xcpus))
|
| 1865 | return PERR_INVCPUS;
|
| 1866 |
|
| 1867 | if (prstate_housekeeping_conflict(new_prs, xcpus))
|
| 1868 | return PERR_HKEEPING;
|
| 1869 |
|
| 1870 | if ((new_prs == PRS_ISOLATED) && (new_prs != parent_prs) &&
|
| 1871 | !isolated_cpus_can_update(xcpus, NULL))
|
| 1872 | return PERR_HKEEPING;
|
| 1873 |
|
| 1874 | if (tasks_nocpu_error(parent, cs, xcpus))
|
| 1875 | return PERR_NOCPUS;
|
| 1876 |
|
| 1877 | |
| 1878 | * This function will only be called when all the preliminary
|
| 1879 | * checks have passed. At this point, the following condition
|
| 1880 | * should hold.
|
| 1881 | *
|
| 1882 | * (cs->effective_xcpus & cpu_active_mask) ⊆ parent->effective_cpus
|
| 1883 | *
|
| 1884 | * Warn if it is not the case.
|
| 1885 |
|
| 1886 | cpumask_and(tmp->new_cpus, xcpus, cpu_active_mask);
|
| 1887 | WARN_ON_ONCE(!cpumask_subset(tmp->new_cpus, parent->effective_cpus));
|
| 1888 |
|
| 1889 | deleting = true;
|
| 1890 | } else if (cmd == partcmd_disable) {
|
| 1891 | |
| 1892 | * May need to add cpus back to parent's effective_cpus
|
| 1893 | * (and maybe removed from subpartitions_cpus/isolated_cpus)
|
| 1894 | * for valid partition root. xcpus may contain CPUs that
|
| 1895 | * shouldn't be removed from the two global cpumasks.
|
| 1896 |
|
| 1897 | if (is_partition_valid(cs)) {
|
| 1898 | cpumask_copy(tmp->addmask, cs->effective_xcpus);
|
| 1899 | adding = true;
|
| 1900 | }
|
| 1901 | new_prs = PRS_MEMBER;
|
| 1902 | } else if (newmask) {
|
| 1903 | |
| 1904 | * Empty cpumask is not allowed
|
| 1905 |
|
| 1906 | if (cpumask_empty(newmask)) {
|
| 1907 | part_error = PERR_CPUSEMPTY;
|
| 1908 | goto write_error;
|
| 1909 | }
|
| 1910 |
|
| 1911 |
|
| 1912 | nocpu |= tasks_nocpu_error(parent, cs, newmask);
|
| 1913 |
|
| 1914 | |
| 1915 | * partcmd_update with newmask:
|
| 1916 | *
|
| 1917 | * Compute add/delete mask to/from effective_cpus
|
| 1918 | *
|
| 1919 | * For valid partition:
|
| 1920 | * addmask = effective_xcpus & ~newmask
|
| 1921 | * & parent->effective_xcpus
|
| 1922 | * delmask = newmask & ~effective_xcpus
|
| 1923 | * & parent->effective_xcpus
|
| 1924 | *
|
| 1925 | * For invalid partition:
|
| 1926 | * delmask = newmask & parent->effective_xcpus
|
| 1927 | * The partition may become valid soon.
|
| 1928 |
|
| 1929 | if (is_partition_invalid(cs)) {
|
| 1930 | adding = false;
|
| 1931 | deleting = cpumask_and(tmp->delmask,
|
| 1932 | newmask, parent->effective_xcpus);
|
| 1933 | } else {
|
| 1934 | cpumask_andnot(tmp->addmask, cs->effective_xcpus, newmask);
|
| 1935 | adding = cpumask_and(tmp->addmask, tmp->addmask,
|
| 1936 | parent->effective_xcpus);
|
| 1937 |
|
| 1938 | cpumask_andnot(tmp->delmask, newmask, cs->effective_xcpus);
|
| 1939 | deleting = cpumask_and(tmp->delmask, tmp->delmask,
|
| 1940 | parent->effective_xcpus);
|
| 1941 | }
|
| 1942 |
|
| 1943 | |
| 1944 | * TBD: Invalidate a currently valid child root partition may
|
| 1945 | * still break isolated_cpus_can_update() rule if parent is an
|
| 1946 | * isolated partition.
|
| 1947 |
|
| 1948 | if (is_partition_valid(cs) && (old_prs != parent_prs)) {
|
| 1949 | if ((parent_prs == PRS_ROOT) &&
|
| 1950 |
|
| 1951 | !isolated_cpus_can_update(tmp->delmask, tmp->addmask))
|
| 1952 | part_error = PERR_HKEEPING;
|
| 1953 | if ((parent_prs == PRS_ISOLATED) &&
|
| 1954 |
|
| 1955 | !isolated_cpus_can_update(tmp->addmask, tmp->delmask))
|
| 1956 | part_error = PERR_HKEEPING;
|
| 1957 | }
|
| 1958 |
|
| 1959 | |
| 1960 | * The new CPUs to be removed from parent's effective CPUs
|
| 1961 | * must be present.
|
| 1962 |
|
| 1963 | if (deleting) {
|
| 1964 | cpumask_and(tmp->new_cpus, tmp->delmask, cpu_active_mask);
|
| 1965 | WARN_ON_ONCE(!cpumask_subset(tmp->new_cpus, parent->effective_cpus));
|
| 1966 | }
|
| 1967 |
|
| 1968 | |
| 1969 | * Make partition invalid if parent's effective_cpus could
|
| 1970 | * become empty and there are tasks in the parent.
|
| 1971 |
|
| 1972 | if (nocpu && (!adding ||
|
| 1973 | !cpumask_intersects(tmp->addmask, cpu_active_mask))) {
|
| 1974 | part_error = PERR_NOCPUS;
|
| 1975 | deleting = false;
|
| 1976 | adding = cpumask_and(tmp->addmask,
|
| 1977 | cs->effective_xcpus, parent->effective_xcpus);
|
| 1978 | }
|
| 1979 | } else {
|
| 1980 | |
| 1981 | * partcmd_update w/o newmask
|
| 1982 | *
|
| 1983 | * delmask = effective_xcpus & parent->effective_cpus
|
| 1984 | *
|
| 1985 | * This can be called from:
|
| 1986 | * 1) update_cpumasks_hier()
|
| 1987 | * 2) cpuset_hotplug_update_tasks()
|
| 1988 | *
|
| 1989 | * Check to see if it can be transitioned from valid to
|
| 1990 | * invalid partition or vice versa.
|
| 1991 | *
|
| 1992 | * A partition error happens when parent has tasks and all
|
| 1993 | * its effective CPUs will have to be distributed out.
|
| 1994 |
|
| 1995 | if (nocpu) {
|
| 1996 | part_error = PERR_NOCPUS;
|
| 1997 | if (is_partition_valid(cs))
|
| 1998 | adding = cpumask_and(tmp->addmask,
|
| 1999 | cs->effective_xcpus,
|
| 2000 | parent->effective_xcpus);
|
| 2001 | } else if (is_partition_invalid(cs) && !cpumask_empty(xcpus) &&
|
| 2002 | cpumask_subset(xcpus, parent->effective_xcpus)) {
|
| 2003 | struct cgroup_subsys_state *css;
|
| 2004 | struct cpuset *child;
|
| 2005 | bool exclusive = true;
|
| 2006 |
|
| 2007 | |
| 2008 | * Convert invalid partition to valid has to
|
| 2009 | * pass the cpu exclusivity test.
|
| 2010 |
|
| 2011 | rcu_read_lock();
|
| 2012 | cpuset_for_each_child(child, css, parent) {
|
| 2013 | if (child == cs)
|
| 2014 | continue;
|
| 2015 | if (!cpusets_are_exclusive(cs, child)) {
|
| 2016 | exclusive = false;
|
| 2017 | break;
|
| 2018 | }
|
| 2019 | }
|
| 2020 | rcu_read_unlock();
|
| 2021 | if (exclusive)
|
| 2022 | deleting = cpumask_and(tmp->delmask,
|
| 2023 | xcpus, parent->effective_cpus);
|
| 2024 | else
|
| 2025 | part_error = PERR_NOTEXCL;
|
| 2026 | }
|
| 2027 | }
|
| 2028 |
|
| 2029 | write_error:
|
| 2030 | if (part_error)
|
| 2031 | WRITE_ONCE(cs->prs_err, part_error);
|
| 2032 |
|
| 2033 | if (cmd == partcmd_update) {
|
| 2034 | |
| 2035 | * Check for possible transition between valid and invalid
|
| 2036 | * partition root.
|
| 2037 |
|
| 2038 | switch (cs->partition_root_state) {
|
| 2039 | case PRS_ROOT:
|
| 2040 | case PRS_ISOLATED:
|
| 2041 | if (part_error)
|
| 2042 | new_prs = -old_prs;
|
| 2043 | break;
|
| 2044 | case PRS_INVALID_ROOT:
|
| 2045 | case PRS_INVALID_ISOLATED:
|
| 2046 | if (!part_error)
|
| 2047 | new_prs = -old_prs;
|
| 2048 | break;
|
| 2049 | }
|
| 2050 | }
|
| 2051 |
|
| 2052 | if (!adding && !deleting && (new_prs == old_prs))
|
| 2053 | return 0;
|
| 2054 |
|
| 2055 | |
| 2056 | * Transitioning between invalid to valid or vice versa may require
|
| 2057 | * changing CS_CPU_EXCLUSIVE. In the case of partcmd_update,
|
| 2058 | * validate_change() has already been successfully called and
|
| 2059 | * CPU lists in cs haven't been updated yet. So defer it to later.
|
| 2060 |
|
| 2061 | if ((old_prs != new_prs) && (cmd != partcmd_update)) {
|
| 2062 | int err = update_partition_exclusive_flag(cs, new_prs);
|
| 2063 |
|
| 2064 | if (err)
|
| 2065 | return err;
|
| 2066 | }
|
| 2067 |
|
| 2068 | |
| 2069 | * Change the parent's effective_cpus & effective_xcpus (top cpuset
|
| 2070 | * only).
|
| 2071 | *
|
| 2072 | * Newly added CPUs will be removed from effective_cpus and
|
| 2073 | * newly deleted ones will be added back to effective_cpus.
|
| 2074 |
|
| 2075 | spin_lock_irq(&callback_lock);
|
| 2076 | if (old_prs != new_prs)
|
| 2077 | cs->partition_root_state = new_prs;
|
| 2078 |
|
| 2079 | |
| 2080 | * Adding to parent's effective_cpus means deletion CPUs from cs
|
| 2081 | * and vice versa.
|
| 2082 |
|
| 2083 | if (adding)
|
| 2084 | partition_xcpus_del(old_prs, parent, tmp->addmask);
|
| 2085 | if (deleting)
|
| 2086 | partition_xcpus_add(new_prs, parent, tmp->delmask);
|
| 2087 |
|
| 2088 | spin_unlock_irq(&callback_lock);
|
| 2089 |
|
| 2090 | if ((old_prs != new_prs) && (cmd == partcmd_update))
|
| 2091 | update_partition_exclusive_flag(cs, new_prs);
|
| 2092 |
|
| 2093 | if (adding || deleting) {
|
| 2094 | cpuset_update_tasks_cpumask(parent, tmp->addmask);
|
| 2095 | update_sibling_cpumasks(parent, cs, tmp);
|
| 2096 | }
|
| 2097 |
|
| 2098 | |
| 2099 | * For partcmd_update without newmask, it is being called from
|
| 2100 | * cpuset_handle_hotplug(). Update the load balance flag and
|
| 2101 | * scheduling domain accordingly.
|
| 2102 |
|
| 2103 | if ((cmd == partcmd_update) && !newmask)
|
| 2104 | update_partition_sd_lb(cs, old_prs);
|
| 2105 |
|
| 2106 | notify_partition_change(cs, old_prs);
|
| 2107 | return 0;
|
| 2108 | }
|
| 2109 |
|
| 2110 | |
| 2111 | * compute_partition_effective_cpumask - compute effective_cpus for partition
|
| 2112 | * @cs: partition root cpuset
|
| 2113 | * @new_ecpus: previously computed effective_cpus to be updated
|
| 2114 | *
|
| 2115 | * Compute the effective_cpus of a partition root by scanning effective_xcpus
|
| 2116 | * of child partition roots and excluding their effective_xcpus.
|
| 2117 | *
|
| 2118 | * This has the side effect of invalidating valid child partition roots,
|
| 2119 | * if necessary. Since it is called from either cpuset_hotplug_update_tasks()
|
| 2120 | * or update_cpumasks_hier() where parent and children are modified
|
| 2121 | * successively, we don't need to call update_parent_effective_cpumask()
|
| 2122 | * and the child's effective_cpus will be updated in later iterations.
|
| 2123 | *
|
| 2124 | * Note that rcu_read_lock() is assumed to be held.
|
| 2125 |
|
| 2126 | static void compute_partition_effective_cpumask(struct cpuset *cs,
|
| 2127 | struct cpumask *new_ecpus)
|
| 2128 | {
|
| 2129 | struct cgroup_subsys_state *css;
|
| 2130 | struct cpuset *child;
|
| 2131 | bool populated = partition_is_populated(cs, NULL);
|
| 2132 |
|
| 2133 | |
| 2134 | * Check child partition roots to see if they should be
|
| 2135 | * invalidated when
|
| 2136 | * 1) child effective_xcpus not a subset of new
|
| 2137 | * excluisve_cpus
|
| 2138 | * 2) All the effective_cpus will be used up and cp
|
| 2139 | * has tasks
|
| 2140 |
|
| 2141 | compute_excpus(cs, new_ecpus);
|
| 2142 | cpumask_and(new_ecpus, new_ecpus, cpu_active_mask);
|
| 2143 |
|
| 2144 | rcu_read_lock();
|
| 2145 | cpuset_for_each_child(child, css, cs) {
|
| 2146 | if (!is_partition_valid(child))
|
| 2147 | continue;
|
| 2148 |
|
| 2149 | |
| 2150 | * There shouldn't be a remote partition underneath another
|
| 2151 | * partition root.
|
| 2152 |
|
| 2153 | WARN_ON_ONCE(is_remote_partition(child));
|
| 2154 | WRITE_ONCE(child->prs_err, 0);
|
| 2155 | if (!cpumask_subset(child->effective_xcpus,
|
| 2156 | cs->effective_xcpus))
|
| 2157 | WRITE_ONCE(child->prs_err, PERR_INVCPUS);
|
| 2158 | else if (populated &&
|
| 2159 | cpumask_subset(new_ecpus, child->effective_xcpus))
|
| 2160 | WRITE_ONCE(child->prs_err, PERR_NOCPUS);
|
| 2161 |
|
| 2162 | if (child->prs_err) {
|
| 2163 | int old_prs = child->partition_root_state;
|
| 2164 |
|
| 2165 | |
| 2166 | * Invalidate child partition
|
| 2167 |
|
| 2168 | spin_lock_irq(&callback_lock);
|
| 2169 | make_partition_invalid(child);
|
| 2170 | spin_unlock_irq(&callback_lock);
|
| 2171 | notify_partition_change(child, old_prs);
|
| 2172 | continue;
|
| 2173 | }
|
| 2174 | cpumask_andnot(new_ecpus, new_ecpus,
|
| 2175 | child->effective_xcpus);
|
| 2176 | }
|
| 2177 | rcu_read_unlock();
|
| 2178 | }
|
| 2179 |
|
| 2180 | |
| 2181 | * update_cpumasks_hier - Update effective cpumasks and tasks in the subtree
|
| 2182 | * @cs: the cpuset to consider
|
| 2183 | * @tmp: temp variables for calculating effective_cpus & partition setup
|
| 2184 | * @force: don't skip any descendant cpusets if set
|
| 2185 | *
|
| 2186 | * When configured cpumask is changed, the effective cpumasks of this cpuset
|
| 2187 | * and all its descendants need to be updated.
|
| 2188 | *
|
| 2189 | * On legacy hierarchy, effective_cpus will be the same with cpu_allowed.
|
| 2190 | *
|
| 2191 | * Called with cpuset_mutex held
|
| 2192 |
|
| 2193 | static void update_cpumasks_hier(struct cpuset *cs, struct tmpmasks *tmp,
|
| 2194 | bool force)
|
| 2195 | {
|
| 2196 | struct cpuset *cp;
|
| 2197 | struct cgroup_subsys_state *pos_css;
|
| 2198 | int old_prs, new_prs;
|
| 2199 |
|
| 2200 | rcu_read_lock();
|
| 2201 | cpuset_for_each_descendant_pre(cp, pos_css, cs) {
|
| 2202 | struct cpuset *parent = parent_cs(cp);
|
| 2203 | bool remote = is_remote_partition(cp);
|
| 2204 | bool update_parent = false;
|
| 2205 |
|
| 2206 | old_prs = new_prs = cp->partition_root_state;
|
| 2207 |
|
| 2208 | |
| 2209 | * For child remote partition root (!= cs), we need to call
|
| 2210 | * remote_cpus_update() if effective_xcpus will be changed.
|
| 2211 | * Otherwise, we can skip the whole subtree.
|
| 2212 | *
|
| 2213 | * remote_cpus_update() will reuse tmp->new_cpus only after
|
| 2214 | * its value is being processed.
|
| 2215 |
|
| 2216 | if (remote && (cp != cs)) {
|
| 2217 | compute_excpus(cp, tmp->new_cpus);
|
| 2218 | if (cpumask_equal(cp->effective_xcpus, tmp->new_cpus)) {
|
| 2219 | pos_css = css_rightmost_descendant(pos_css);
|
| 2220 | continue;
|
| 2221 | }
|
| 2222 | rcu_read_unlock();
|
| 2223 | remote_cpus_update(cp, NULL, tmp->new_cpus, tmp);
|
| 2224 | rcu_read_lock();
|
| 2225 |
|
| 2226 |
|
| 2227 | new_prs = cp->partition_root_state;
|
| 2228 | remote = (new_prs == old_prs);
|
| 2229 | }
|
| 2230 |
|
| 2231 | if (remote || (is_partition_valid(parent) && is_partition_valid(cp)))
|
| 2232 | compute_partition_effective_cpumask(cp, tmp->new_cpus);
|
| 2233 | else
|
| 2234 | compute_effective_cpumask(tmp->new_cpus, cp, parent);
|
| 2235 |
|
| 2236 | if (remote)
|
| 2237 | goto get_css;
|
| 2238 |
|
| 2239 | |
| 2240 | * A partition with no effective_cpus is allowed as long as
|
| 2241 | * there is no task associated with it. Call
|
| 2242 | * update_parent_effective_cpumask() to check it.
|
| 2243 |
|
| 2244 | if (is_partition_valid(cp) && cpumask_empty(tmp->new_cpus)) {
|
| 2245 | update_parent = true;
|
| 2246 | goto update_parent_effective;
|
| 2247 | }
|
| 2248 |
|
| 2249 | |
| 2250 | * Skip the whole subtree if
|
| 2251 | * 1) the cpumask remains the same,
|
| 2252 | * 2) has no partition root state,
|
| 2253 | * 3) force flag not set, and
|
| 2254 | * 4) for v2 load balance state same as its parent.
|
| 2255 |
|
| 2256 | if (!cp->partition_root_state && !force &&
|
| 2257 | cpumask_equal(tmp->new_cpus, cp->effective_cpus) &&
|
| 2258 | (!cpuset_v2() ||
|
| 2259 | (is_sched_load_balance(parent) == is_sched_load_balance(cp)))) {
|
| 2260 | pos_css = css_rightmost_descendant(pos_css);
|
| 2261 | continue;
|
| 2262 | }
|
| 2263 |
|
| 2264 | update_parent_effective:
|
| 2265 | |
| 2266 | * update_parent_effective_cpumask() should have been called
|
| 2267 | * for cs already in update_cpumask(). We should also call
|
| 2268 | * cpuset_update_tasks_cpumask() again for tasks in the parent
|
| 2269 | * cpuset if the parent's effective_cpus changes.
|
| 2270 |
|
| 2271 | if ((cp != cs) && old_prs) {
|
| 2272 | switch (parent->partition_root_state) {
|
| 2273 | case PRS_ROOT:
|
| 2274 | case PRS_ISOLATED:
|
| 2275 | update_parent = true;
|
| 2276 | break;
|
| 2277 |
|
| 2278 | default:
|
| 2279 | |
| 2280 | * When parent is not a partition root or is
|
| 2281 | * invalid, child partition roots become
|
| 2282 | * invalid too.
|
| 2283 |
|
| 2284 | if (is_partition_valid(cp))
|
| 2285 | new_prs = -cp->partition_root_state;
|
| 2286 | WRITE_ONCE(cp->prs_err,
|
| 2287 | is_partition_invalid(parent)
|
| 2288 | ? PERR_INVPARENT : PERR_NOTPART);
|
| 2289 | break;
|
| 2290 | }
|
| 2291 | }
|
| 2292 | get_css:
|
| 2293 | if (!css_tryget_online(&cp->css))
|
| 2294 | continue;
|
| 2295 | rcu_read_unlock();
|
| 2296 |
|
| 2297 | if (update_parent) {
|
| 2298 | update_parent_effective_cpumask(cp, partcmd_update, NULL, tmp);
|
| 2299 | |
| 2300 | * The cpuset partition_root_state may become
|
| 2301 | * invalid. Capture it.
|
| 2302 |
|
| 2303 | new_prs = cp->partition_root_state;
|
| 2304 | }
|
| 2305 |
|
| 2306 | spin_lock_irq(&callback_lock);
|
| 2307 | cpumask_copy(cp->effective_cpus, tmp->new_cpus);
|
| 2308 | cp->partition_root_state = new_prs;
|
| 2309 | |
| 2310 | * Need to compute effective_xcpus if either exclusive_cpus
|
| 2311 | * is non-empty or it is a valid partition root.
|
| 2312 |
|
| 2313 | if ((new_prs > 0) || !cpumask_empty(cp->exclusive_cpus))
|
| 2314 | compute_excpus(cp, cp->effective_xcpus);
|
| 2315 | if (new_prs <= 0)
|
| 2316 | reset_partition_data(cp);
|
| 2317 | spin_unlock_irq(&callback_lock);
|
| 2318 |
|
| 2319 | notify_partition_change(cp, old_prs);
|
| 2320 |
|
| 2321 | WARN_ON(!is_in_v2_mode() &&
|
| 2322 | !cpumask_equal(cp->cpus_allowed, cp->effective_cpus));
|
| 2323 |
|
| 2324 | cpuset_update_tasks_cpumask(cp, tmp->new_cpus);
|
| 2325 |
|
| 2326 | |
| 2327 | * On default hierarchy, inherit the CS_SCHED_LOAD_BALANCE
|
| 2328 | * from parent if current cpuset isn't a valid partition root
|
| 2329 | * and their load balance states differ.
|
| 2330 |
|
| 2331 | if (cpuset_v2() && !is_partition_valid(cp) &&
|
| 2332 | (is_sched_load_balance(parent) != is_sched_load_balance(cp))) {
|
| 2333 | if (is_sched_load_balance(parent))
|
| 2334 | set_bit(CS_SCHED_LOAD_BALANCE, &cp->flags);
|
| 2335 | else
|
| 2336 | clear_bit(CS_SCHED_LOAD_BALANCE, &cp->flags);
|
| 2337 | }
|
| 2338 |
|
| 2339 | |
| 2340 | * On legacy hierarchy, if the effective cpumask of any non-
|
| 2341 | * empty cpuset is changed, we need to rebuild sched domains.
|
| 2342 | * On default hierarchy, the cpuset needs to be a partition
|
| 2343 | * root as well.
|
| 2344 |
|
| 2345 | if (!cpumask_empty(cp->cpus_allowed) &&
|
| 2346 | is_sched_load_balance(cp) &&
|
| 2347 | (!cpuset_v2() || is_partition_valid(cp)))
|
| 2348 | cpuset_force_rebuild();
|
| 2349 |
|
| 2350 | rcu_read_lock();
|
| 2351 | css_put(&cp->css);
|
| 2352 | }
|
| 2353 | rcu_read_unlock();
|
| 2354 | }
|
| 2355 |
|
| 2356 | |
| 2357 | * update_sibling_cpumasks - Update siblings cpumasks
|
| 2358 | * @parent: Parent cpuset
|
| 2359 | * @cs: Current cpuset
|
| 2360 | * @tmp: Temp variables
|
| 2361 |
|
| 2362 | static void update_sibling_cpumasks(struct cpuset *parent, struct cpuset *cs,
|
| 2363 | struct tmpmasks *tmp)
|
| 2364 | {
|
| 2365 | struct cpuset *sibling;
|
| 2366 | struct cgroup_subsys_state *pos_css;
|
| 2367 |
|
| 2368 | lockdep_assert_cpuset_lock_held();
|
| 2369 |
|
| 2370 | |
| 2371 | * Check all its siblings and call update_cpumasks_hier()
|
| 2372 | * if their effective_cpus will need to be changed.
|
| 2373 | *
|
| 2374 | * It is possible a change in parent's effective_cpus
|
| 2375 | * due to a change in a child partition's effective_xcpus will impact
|
| 2376 | * its siblings even if they do not inherit parent's effective_cpus
|
| 2377 | * directly. It should not impact valid partition.
|
| 2378 | *
|
| 2379 | * The update_cpumasks_hier() function may sleep. So we have to
|
| 2380 | * release the RCU read lock before calling it.
|
| 2381 |
|
| 2382 | rcu_read_lock();
|
| 2383 | cpuset_for_each_child(sibling, pos_css, parent) {
|
| 2384 | if (sibling == cs || is_partition_valid(sibling))
|
| 2385 | continue;
|
| 2386 |
|
| 2387 | compute_effective_cpumask(tmp->new_cpus, sibling,
|
| 2388 | parent);
|
| 2389 | if (cpumask_equal(tmp->new_cpus, sibling->effective_cpus))
|
| 2390 | continue;
|
| 2391 |
|
| 2392 | if (!css_tryget_online(&sibling->css))
|
| 2393 | continue;
|
| 2394 |
|
| 2395 | rcu_read_unlock();
|
| 2396 | update_cpumasks_hier(sibling, tmp, false);
|
| 2397 | rcu_read_lock();
|
| 2398 | css_put(&sibling->css);
|
| 2399 | }
|
| 2400 | rcu_read_unlock();
|
| 2401 | }
|
| 2402 |
|
| 2403 | static int parse_cpuset_cpulist(const char *buf, struct cpumask *out_mask)
|
| 2404 | {
|
| 2405 | int retval;
|
| 2406 |
|
| 2407 | retval = cpulist_parse(buf, out_mask);
|
| 2408 | if (retval < 0)
|
| 2409 | return retval;
|
| 2410 | if (!cpumask_subset(out_mask, top_cpuset.cpus_allowed))
|
| 2411 | return -EINVAL;
|
| 2412 |
|
| 2413 | return 0;
|
| 2414 | }
|
| 2415 |
|
| 2416 | |
| 2417 | * validate_partition - Validate a cpuset partition configuration
|
| 2418 | * @cs: The cpuset to validate
|
| 2419 | * @trialcs: The trial cpuset containing proposed configuration changes
|
| 2420 | *
|
| 2421 | * If any validation check fails, the appropriate error code is set in the
|
| 2422 | * cpuset's prs_err field.
|
| 2423 | *
|
| 2424 | * Return: PRS error code (0 if valid, non-zero error code if invalid)
|
| 2425 |
|
| 2426 | static enum prs_errcode validate_partition(struct cpuset *cs, struct cpuset *trialcs)
|
| 2427 | {
|
| 2428 | struct cpuset *parent = parent_cs(cs);
|
| 2429 |
|
| 2430 | if (cs_is_member(trialcs))
|
| 2431 | return PERR_NONE;
|
| 2432 |
|
| 2433 | if (cpumask_empty(trialcs->effective_xcpus))
|
| 2434 | return PERR_INVCPUS;
|
| 2435 |
|
| 2436 | if (prstate_housekeeping_conflict(trialcs->partition_root_state,
|
| 2437 | trialcs->effective_xcpus))
|
| 2438 | return PERR_HKEEPING;
|
| 2439 |
|
| 2440 | if (tasks_nocpu_error(parent, cs, trialcs->effective_xcpus))
|
| 2441 | return PERR_NOCPUS;
|
| 2442 |
|
| 2443 | return PERR_NONE;
|
| 2444 | }
|
| 2445 |
|
| 2446 | |
| 2447 | * partition_cpus_change - Handle partition state changes due to CPU mask updates
|
| 2448 | * @cs: The target cpuset being modified
|
| 2449 | * @trialcs: The trial cpuset containing proposed configuration changes
|
| 2450 | * @tmp: Temporary masks for intermediate calculations
|
| 2451 | *
|
| 2452 | * This function handles partition state transitions triggered by CPU mask changes.
|
| 2453 | * CPU modifications may cause a partition to be disabled or require state updates.
|
| 2454 |
|
| 2455 | static void partition_cpus_change(struct cpuset *cs, struct cpuset *trialcs,
|
| 2456 | struct tmpmasks *tmp)
|
| 2457 | {
|
| 2458 | enum prs_errcode prs_err;
|
| 2459 |
|
| 2460 | if (cs_is_member(cs))
|
| 2461 | return;
|
| 2462 |
|
| 2463 | prs_err = validate_partition(cs, trialcs);
|
| 2464 | if (prs_err) {
|
| 2465 | WRITE_ONCE(cs->prs_err, prs_err);
|
| 2466 | trialcs->prs_err = prs_err;
|
| 2467 | }
|
| 2468 |
|
| 2469 | if (is_remote_partition(cs)) {
|
| 2470 | if (trialcs->prs_err)
|
| 2471 | remote_partition_disable(cs, tmp);
|
| 2472 | else
|
| 2473 | remote_cpus_update(cs, trialcs->exclusive_cpus,
|
| 2474 | trialcs->effective_xcpus, tmp);
|
| 2475 | } else {
|
| 2476 | if (trialcs->prs_err)
|
| 2477 | update_parent_effective_cpumask(cs, partcmd_invalidate,
|
| 2478 | NULL, tmp);
|
| 2479 | else
|
| 2480 | update_parent_effective_cpumask(cs, partcmd_update,
|
| 2481 | trialcs->effective_xcpus, tmp);
|
| 2482 | }
|
| 2483 | }
|
| 2484 |
|
| 2485 | |
| 2486 | * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
|
| 2487 | * @cs: the cpuset to consider
|
| 2488 | * @trialcs: trial cpuset
|
| 2489 | * @buf: buffer of cpu numbers written to this cpuset
|
| 2490 |
|
| 2491 | static int update_cpumask(struct cpuset *cs, struct cpuset *trialcs,
|
| 2492 | const char *buf)
|
| 2493 | {
|
| 2494 | int retval;
|
| 2495 | struct tmpmasks tmp;
|
| 2496 | bool force = false;
|
| 2497 | int old_prs = cs->partition_root_state;
|
| 2498 |
|
| 2499 | retval = parse_cpuset_cpulist(buf, trialcs->cpus_allowed);
|
| 2500 | if (retval < 0)
|
| 2501 | return retval;
|
| 2502 |
|
| 2503 |
|
| 2504 | if (cpumask_equal(cs->cpus_allowed, trialcs->cpus_allowed))
|
| 2505 | return 0;
|
| 2506 |
|
| 2507 | compute_trialcs_excpus(trialcs, cs);
|
| 2508 | trialcs->prs_err = PERR_NONE;
|
| 2509 |
|
| 2510 | retval = validate_change(cs, trialcs);
|
| 2511 | if (retval < 0)
|
| 2512 | return retval;
|
| 2513 |
|
| 2514 | if (alloc_tmpmasks(&tmp))
|
| 2515 | return -ENOMEM;
|
| 2516 |
|
| 2517 | |
| 2518 | * Check all the descendants in update_cpumasks_hier() if
|
| 2519 | * effective_xcpus is to be changed.
|
| 2520 |
|
| 2521 | force = !cpumask_equal(cs->effective_xcpus, trialcs->effective_xcpus);
|
| 2522 |
|
| 2523 | partition_cpus_change(cs, trialcs, &tmp);
|
| 2524 |
|
| 2525 | spin_lock_irq(&callback_lock);
|
| 2526 | cpumask_copy(cs->cpus_allowed, trialcs->cpus_allowed);
|
| 2527 | cpumask_copy(cs->effective_xcpus, trialcs->effective_xcpus);
|
| 2528 | if ((old_prs > 0) && !is_partition_valid(cs))
|
| 2529 | reset_partition_data(cs);
|
| 2530 | spin_unlock_irq(&callback_lock);
|
| 2531 |
|
| 2532 |
|
| 2533 | update_cpumasks_hier(cs, &tmp, force);
|
| 2534 |
|
| 2535 |
|
| 2536 | if (cs->partition_root_state)
|
| 2537 | update_partition_sd_lb(cs, old_prs);
|
| 2538 |
|
| 2539 | free_tmpmasks(&tmp);
|
| 2540 | return retval;
|
| 2541 | }
|
| 2542 |
|
| 2543 | |
| 2544 | * update_exclusive_cpumask - update the exclusive_cpus mask of a cpuset
|
| 2545 | * @cs: the cpuset to consider
|
| 2546 | * @trialcs: trial cpuset
|
| 2547 | * @buf: buffer of cpu numbers written to this cpuset
|
| 2548 | *
|
| 2549 | * The tasks' cpumask will be updated if cs is a valid partition root.
|
| 2550 |
|
| 2551 | static int update_exclusive_cpumask(struct cpuset *cs, struct cpuset *trialcs,
|
| 2552 | const char *buf)
|
| 2553 | {
|
| 2554 | int retval;
|
| 2555 | struct tmpmasks tmp;
|
| 2556 | bool force = false;
|
| 2557 | int old_prs = cs->partition_root_state;
|
| 2558 |
|
| 2559 | retval = parse_cpuset_cpulist(buf, trialcs->exclusive_cpus);
|
| 2560 | if (retval < 0)
|
| 2561 | return retval;
|
| 2562 |
|
| 2563 |
|
| 2564 | if (cpumask_equal(cs->exclusive_cpus, trialcs->exclusive_cpus))
|
| 2565 | return 0;
|
| 2566 |
|
| 2567 | |
| 2568 | * Reject the change if there is exclusive CPUs conflict with
|
| 2569 | * the siblings.
|
| 2570 |
|
| 2571 | if (compute_trialcs_excpus(trialcs, cs))
|
| 2572 | return -EINVAL;
|
| 2573 |
|
| 2574 | |
| 2575 | * Check all the descendants in update_cpumasks_hier() if
|
| 2576 | * effective_xcpus is to be changed.
|
| 2577 |
|
| 2578 | force = !cpumask_equal(cs->effective_xcpus, trialcs->effective_xcpus);
|
| 2579 |
|
| 2580 | retval = validate_change(cs, trialcs);
|
| 2581 | if (retval)
|
| 2582 | return retval;
|
| 2583 |
|
| 2584 | if (alloc_tmpmasks(&tmp))
|
| 2585 | return -ENOMEM;
|
| 2586 |
|
| 2587 | trialcs->prs_err = PERR_NONE;
|
| 2588 | partition_cpus_change(cs, trialcs, &tmp);
|
| 2589 |
|
| 2590 | spin_lock_irq(&callback_lock);
|
| 2591 | cpumask_copy(cs->exclusive_cpus, trialcs->exclusive_cpus);
|
| 2592 | cpumask_copy(cs->effective_xcpus, trialcs->effective_xcpus);
|
| 2593 | if ((old_prs > 0) && !is_partition_valid(cs))
|
| 2594 | reset_partition_data(cs);
|
| 2595 | spin_unlock_irq(&callback_lock);
|
| 2596 |
|
| 2597 | |
| 2598 | * Call update_cpumasks_hier() to update effective_cpus/effective_xcpus
|
| 2599 | * of the subtree when it is a valid partition root or effective_xcpus
|
| 2600 | * is updated.
|
| 2601 |
|
| 2602 | if (is_partition_valid(cs) || force)
|
| 2603 | update_cpumasks_hier(cs, &tmp, force);
|
| 2604 |
|
| 2605 |
|
| 2606 | if (cs->partition_root_state)
|
| 2607 | update_partition_sd_lb(cs, old_prs);
|
| 2608 |
|
| 2609 | free_tmpmasks(&tmp);
|
| 2610 | return 0;
|
| 2611 | }
|
| 2612 |
|
| 2613 | |
| 2614 | * Migrate memory region from one set of nodes to another. This is
|
| 2615 | * performed asynchronously as it can be called from process migration path
|
| 2616 | * holding locks involved in process management. All mm migrations are
|
| 2617 | * performed in the queued order and can be waited for by flushing
|
| 2618 | * cpuset_migrate_mm_wq.
|
| 2619 |
|
| 2620 |
|
| 2621 | struct cpuset_migrate_mm_work {
|
| 2622 | struct work_struct work;
|
| 2623 | struct mm_struct *mm;
|
| 2624 | nodemask_t from;
|
| 2625 | nodemask_t to;
|
| 2626 | };
|
| 2627 |
|
| 2628 | static void cpuset_migrate_mm_workfn(struct work_struct *work)
|
| 2629 | {
|
| 2630 | struct cpuset_migrate_mm_work *mwork =
|
| 2631 | container_of(work, struct cpuset_migrate_mm_work, work);
|
| 2632 |
|
| 2633 |
|
| 2634 | do_migrate_pages(mwork->mm, &mwork->from, &mwork->to, MPOL_MF_MOVE_ALL);
|
| 2635 | mmput(mwork->mm);
|
| 2636 | kfree(mwork);
|
| 2637 | }
|
| 2638 |
|
| 2639 | static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
|
| 2640 | const nodemask_t *to)
|
| 2641 | {
|
| 2642 | struct cpuset_migrate_mm_work *mwork;
|
| 2643 |
|
| 2644 | if (nodes_equal(*from, *to)) {
|
| 2645 | mmput(mm);
|
| 2646 | return;
|
| 2647 | }
|
| 2648 |
|
| 2649 | mwork = kzalloc_obj(*mwork);
|
| 2650 | if (mwork) {
|
| 2651 | mwork->mm = mm;
|
| 2652 | mwork->from = *from;
|
| 2653 | mwork->to = *to;
|
| 2654 | INIT_WORK(&mwork->work, cpuset_migrate_mm_workfn);
|
| 2655 | queue_work(cpuset_migrate_mm_wq, &mwork->work);
|
| 2656 | } else {
|
| 2657 | mmput(mm);
|
| 2658 | }
|
| 2659 | }
|
| 2660 |
|
| 2661 | static void flush_migrate_mm_task_workfn(struct callback_head *head)
|
| 2662 | {
|
| 2663 | flush_workqueue(cpuset_migrate_mm_wq);
|
| 2664 | kfree(head);
|
| 2665 | }
|
| 2666 |
|
| 2667 | static void schedule_flush_migrate_mm(void)
|
| 2668 | {
|
| 2669 | struct callback_head *flush_cb;
|
| 2670 |
|
| 2671 | flush_cb = kzalloc_obj(struct callback_head);
|
| 2672 | if (!flush_cb)
|
| 2673 | return;
|
| 2674 |
|
| 2675 | init_task_work(flush_cb, flush_migrate_mm_task_workfn);
|
| 2676 |
|
| 2677 | if (task_work_add(current, flush_cb, TWA_RESUME))
|
| 2678 | kfree(flush_cb);
|
| 2679 | }
|
| 2680 |
|
| 2681 | |
| 2682 | * cpuset_change_task_nodemask - change task's mems_allowed and mempolicy
|
| 2683 | * @tsk: the task to change
|
| 2684 | * @newmems: new nodes that the task will be set
|
| 2685 | *
|
| 2686 | * We use the mems_allowed_seq seqlock to safely update both tsk->mems_allowed
|
| 2687 | * and rebind an eventual tasks' mempolicy. If the task is allocating in
|
| 2688 | * parallel, it might temporarily see an empty intersection, which results in
|
| 2689 | * a seqlock check and retry before OOM or allocation failure.
|
| 2690 |
|
| 2691 | static void cpuset_change_task_nodemask(struct task_struct *tsk,
|
| 2692 | nodemask_t *newmems)
|
| 2693 | {
|
| 2694 | task_lock(tsk);
|
| 2695 |
|
| 2696 | local_irq_disable();
|
| 2697 | write_seqcount_begin(&tsk->mems_allowed_seq);
|
| 2698 |
|
| 2699 | nodes_or(tsk->mems_allowed, tsk->mems_allowed, *newmems);
|
| 2700 | mpol_rebind_task(tsk, newmems);
|
| 2701 | tsk->mems_allowed = *newmems;
|
| 2702 |
|
| 2703 | write_seqcount_end(&tsk->mems_allowed_seq);
|
| 2704 | local_irq_enable();
|
| 2705 |
|
| 2706 | task_unlock(tsk);
|
| 2707 | }
|
| 2708 |
|
| 2709 | static void *cpuset_being_rebound;
|
| 2710 |
|
| 2711 | |
| 2712 | * cpuset_update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
|
| 2713 | * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
|
| 2714 | *
|
| 2715 | * Iterate through each task of @cs updating its mems_allowed to the
|
| 2716 | * effective cpuset's. As this function is called with cpuset_mutex held,
|
| 2717 | * cpuset membership stays stable.
|
| 2718 | *
|
| 2719 | * - cpuset_change_task_nodemask(): guarantee_online_mems()
|
| 2720 | * - mpol_rebind_mm(): effective_mems
|
| 2721 | * - cpuset_migrate_mm(): guarantee_online_mems()
|
| 2722 | * - old_mems_allowed: guarantee_online_mems()
|
| 2723 | *
|
| 2724 | * For v2, guarantee_online_mems() should return a node mask that is the same
|
| 2725 | * as the effective_mems of current cpuset.
|
| 2726 |
|
| 2727 | void cpuset_update_tasks_nodemask(struct cpuset *cs)
|
| 2728 | {
|
| 2729 | static nodemask_t newmems;
|
| 2730 | struct css_task_iter it;
|
| 2731 | struct task_struct *task;
|
| 2732 |
|
| 2733 | cpuset_being_rebound = cs;
|
| 2734 | guarantee_online_mems(cs, &newmems);
|
| 2735 |
|
| 2736 | |
| 2737 | * The mpol_rebind_mm() call takes mmap_lock, which we couldn't
|
| 2738 | * take while holding tasklist_lock. Forks can happen - the
|
| 2739 | * mpol_dup() cpuset_being_rebound check will catch such forks,
|
| 2740 | * and rebind their vma mempolicies too. Because we still hold
|
| 2741 | * the global cpuset_mutex, we know that no other rebind effort
|
| 2742 | * will be contending for the global variable cpuset_being_rebound.
|
| 2743 | * It's ok if we rebind the same mm twice; mpol_rebind_mm()
|
| 2744 | * is idempotent. Also migrate pages in each mm to new nodes.
|
| 2745 |
|
| 2746 | css_task_iter_start(&cs->css, 0, &it);
|
| 2747 | while ((task = css_task_iter_next(&it))) {
|
| 2748 | struct mm_struct *mm;
|
| 2749 | bool migrate;
|
| 2750 |
|
| 2751 | cpuset_change_task_nodemask(task, &newmems);
|
| 2752 |
|
| 2753 |
|
| 2754 | if (!thread_group_leader(task))
|
| 2755 | continue;
|
| 2756 |
|
| 2757 | mm = get_task_mm(task);
|
| 2758 | if (!mm)
|
| 2759 | continue;
|
| 2760 |
|
| 2761 | migrate = is_memory_migrate(cs);
|
| 2762 |
|
| 2763 | |
| 2764 | * For v1 we can have empty effective_mems, but we cannot
|
| 2765 | * attach any tasks (see cpuset_can_attach_check()). For v2,
|
| 2766 | * effective_mems is guaranteed to not be empty.
|
| 2767 |
|
| 2768 | mpol_rebind_mm(mm, &cs->effective_mems);
|
| 2769 | if (migrate)
|
| 2770 | cpuset_migrate_mm(mm, &cs->old_mems_allowed, &newmems);
|
| 2771 | else
|
| 2772 | mmput(mm);
|
| 2773 | }
|
| 2774 | css_task_iter_end(&it);
|
| 2775 |
|
| 2776 | |
| 2777 | * All the tasks' nodemasks have been updated, update
|
| 2778 | * cs->old_mems_allowed.
|
| 2779 |
|
| 2780 | cs->old_mems_allowed = newmems;
|
| 2781 |
|
| 2782 |
|
| 2783 | cpuset_being_rebound = NULL;
|
| 2784 | }
|
| 2785 |
|
| 2786 | |
| 2787 | * update_nodemasks_hier - Update effective nodemasks and tasks in the subtree
|
| 2788 | * @cs: the cpuset to consider
|
| 2789 | * @new_mems: a temp variable for calculating new effective_mems
|
| 2790 | *
|
| 2791 | * When configured nodemask is changed, the effective nodemasks of this cpuset
|
| 2792 | * and all its descendants need to be updated.
|
| 2793 | *
|
| 2794 | * On legacy hierarchy, effective_mems will be the same with mems_allowed.
|
| 2795 | *
|
| 2796 | * Called with cpuset_mutex held
|
| 2797 |
|
| 2798 | static void update_nodemasks_hier(struct cpuset *cs, nodemask_t *new_mems)
|
| 2799 | {
|
| 2800 | struct cpuset *cp;
|
| 2801 | struct cgroup_subsys_state *pos_css;
|
| 2802 |
|
| 2803 | rcu_read_lock();
|
| 2804 | cpuset_for_each_descendant_pre(cp, pos_css, cs) {
|
| 2805 | struct cpuset *parent = parent_cs(cp);
|
| 2806 |
|
| 2807 | compute_effective_nodemask(new_mems, cp, parent);
|
| 2808 |
|
| 2809 |
|
| 2810 | if (nodes_equal(*new_mems, cp->effective_mems)) {
|
| 2811 | pos_css = css_rightmost_descendant(pos_css);
|
| 2812 | continue;
|
| 2813 | }
|
| 2814 |
|
| 2815 | if (!css_tryget_online(&cp->css))
|
| 2816 | continue;
|
| 2817 | rcu_read_unlock();
|
| 2818 |
|
| 2819 | spin_lock_irq(&callback_lock);
|
| 2820 | cp->effective_mems = *new_mems;
|
| 2821 | spin_unlock_irq(&callback_lock);
|
| 2822 |
|
| 2823 | WARN_ON(!is_in_v2_mode() &&
|
| 2824 | !nodes_equal(cp->mems_allowed, cp->effective_mems));
|
| 2825 |
|
| 2826 | cpuset_update_tasks_nodemask(cp);
|
| 2827 |
|
| 2828 | rcu_read_lock();
|
| 2829 | css_put(&cp->css);
|
| 2830 | }
|
| 2831 | rcu_read_unlock();
|
| 2832 | }
|
| 2833 |
|
| 2834 | |
| 2835 | * Handle user request to change the 'mems' memory placement
|
| 2836 | * of a cpuset. Needs to validate the request, update the
|
| 2837 | * cpusets mems_allowed, and for each task in the cpuset,
|
| 2838 | * update mems_allowed and rebind task's mempolicy and any vma
|
| 2839 | * mempolicies and if the cpuset is marked 'memory_migrate',
|
| 2840 | * migrate the tasks pages to the new memory.
|
| 2841 | *
|
| 2842 | * Call with cpuset_mutex held. May take callback_lock during call.
|
| 2843 | * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
|
| 2844 | * lock each such tasks mm->mmap_lock, scan its vma's and rebind
|
| 2845 | * their mempolicies to the cpusets new mems_allowed.
|
| 2846 |
|
| 2847 | static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
|
| 2848 | const char *buf)
|
| 2849 | {
|
| 2850 | int retval;
|
| 2851 |
|
| 2852 | |
| 2853 | * An empty mems_allowed is ok iff there are no tasks in the cpuset.
|
| 2854 | * The validate_change() call ensures that cpusets with tasks have memory.
|
| 2855 |
|
| 2856 | retval = nodelist_parse(buf, trialcs->mems_allowed);
|
| 2857 | if (retval < 0)
|
| 2858 | return retval;
|
| 2859 |
|
| 2860 | if (!nodes_subset(trialcs->mems_allowed,
|
| 2861 | top_cpuset.mems_allowed))
|
| 2862 | return -EINVAL;
|
| 2863 |
|
| 2864 |
|
| 2865 | if (nodes_equal(cs->mems_allowed, trialcs->mems_allowed))
|
| 2866 | return 0;
|
| 2867 |
|
| 2868 | retval = validate_change(cs, trialcs);
|
| 2869 | if (retval < 0)
|
| 2870 | return retval;
|
| 2871 |
|
| 2872 | check_insane_mems_config(&trialcs->mems_allowed);
|
| 2873 |
|
| 2874 | spin_lock_irq(&callback_lock);
|
| 2875 | cs->mems_allowed = trialcs->mems_allowed;
|
| 2876 | spin_unlock_irq(&callback_lock);
|
| 2877 |
|
| 2878 |
|
| 2879 | update_nodemasks_hier(cs, &trialcs->mems_allowed);
|
| 2880 | return 0;
|
| 2881 | }
|
| 2882 |
|
| 2883 | bool current_cpuset_is_being_rebound(void)
|
| 2884 | {
|
| 2885 | bool ret;
|
| 2886 |
|
| 2887 | rcu_read_lock();
|
| 2888 | ret = task_cs(current) == cpuset_being_rebound;
|
| 2889 | rcu_read_unlock();
|
| 2890 |
|
| 2891 | return ret;
|
| 2892 | }
|
| 2893 |
|
| 2894 | |
| 2895 | * cpuset_update_flag - read a 0 or a 1 in a file and update associated flag
|
| 2896 | * bit: the bit to update (see cpuset_flagbits_t)
|
| 2897 | * cs: the cpuset to update
|
| 2898 | * turning_on: whether the flag is being set or cleared
|
| 2899 | *
|
| 2900 | * Call with cpuset_mutex held.
|
| 2901 |
|
| 2902 |
|
| 2903 | int cpuset_update_flag(cpuset_flagbits_t bit, struct cpuset *cs,
|
| 2904 | int turning_on)
|
| 2905 | {
|
| 2906 | struct cpuset *trialcs;
|
| 2907 | int balance_flag_changed;
|
| 2908 | int spread_page_changed;
|
| 2909 | int err;
|
| 2910 |
|
| 2911 | trialcs = dup_or_alloc_cpuset(cs);
|
| 2912 | if (!trialcs)
|
| 2913 | return -ENOMEM;
|
| 2914 |
|
| 2915 | if (turning_on)
|
| 2916 | set_bit(bit, &trialcs->flags);
|
| 2917 | else
|
| 2918 | clear_bit(bit, &trialcs->flags);
|
| 2919 |
|
| 2920 | err = validate_change(cs, trialcs);
|
| 2921 | if (err < 0)
|
| 2922 | goto out;
|
| 2923 |
|
| 2924 | balance_flag_changed = (is_sched_load_balance(cs) !=
|
| 2925 | is_sched_load_balance(trialcs));
|
| 2926 |
|
| 2927 | spread_page_changed = is_spread_page(cs) != is_spread_page(trialcs);
|
| 2928 |
|
| 2929 | spin_lock_irq(&callback_lock);
|
| 2930 | cs->flags = trialcs->flags;
|
| 2931 | spin_unlock_irq(&callback_lock);
|
| 2932 |
|
| 2933 | if (!cpumask_empty(trialcs->cpus_allowed) && balance_flag_changed) {
|
| 2934 | if (cpuset_v2())
|
| 2935 | cpuset_force_rebuild();
|
| 2936 | else
|
| 2937 | rebuild_sched_domains_locked();
|
| 2938 | }
|
| 2939 |
|
| 2940 | if (spread_page_changed)
|
| 2941 | cpuset1_update_tasks_flags(cs);
|
| 2942 | out:
|
| 2943 | free_cpuset(trialcs);
|
| 2944 | return err;
|
| 2945 | }
|
| 2946 |
|
| 2947 | |
| 2948 | * update_prstate - update partition_root_state
|
| 2949 | * @cs: the cpuset to update
|
| 2950 | * @new_prs: new partition root state
|
| 2951 | * Return: 0 if successful, != 0 if error
|
| 2952 | *
|
| 2953 | * Call with cpuset_mutex held.
|
| 2954 |
|
| 2955 | static int update_prstate(struct cpuset *cs, int new_prs)
|
| 2956 | {
|
| 2957 | int err = PERR_NONE, old_prs = cs->partition_root_state;
|
| 2958 | struct cpuset *parent = parent_cs(cs);
|
| 2959 | struct tmpmasks tmpmask;
|
| 2960 | bool isolcpus_updated = false;
|
| 2961 |
|
| 2962 | if (old_prs == new_prs)
|
| 2963 | return 0;
|
| 2964 |
|
| 2965 | |
| 2966 | * Treat a previously invalid partition root as if it is a "member".
|
| 2967 |
|
| 2968 | if (new_prs && is_partition_invalid(cs))
|
| 2969 | old_prs = PRS_MEMBER;
|
| 2970 |
|
| 2971 | if (alloc_tmpmasks(&tmpmask))
|
| 2972 | return -ENOMEM;
|
| 2973 |
|
| 2974 | err = update_partition_exclusive_flag(cs, new_prs);
|
| 2975 | if (err)
|
| 2976 | goto out;
|
| 2977 |
|
| 2978 | if (!old_prs) {
|
| 2979 | |
| 2980 | * cpus_allowed and exclusive_cpus cannot be both empty.
|
| 2981 |
|
| 2982 | if (xcpus_empty(cs)) {
|
| 2983 | err = PERR_CPUSEMPTY;
|
| 2984 | goto out;
|
| 2985 | }
|
| 2986 |
|
| 2987 | |
| 2988 | * We don't support the creation of a new local partition with
|
| 2989 | * a remote partition underneath it. This unsupported
|
| 2990 | * setting can happen only if parent is the top_cpuset because
|
| 2991 | * a remote partition cannot be created underneath an existing
|
| 2992 | * local or remote partition.
|
| 2993 |
|
| 2994 | if ((parent == &top_cpuset) &&
|
| 2995 | cpumask_intersects(cs->exclusive_cpus, subpartitions_cpus)) {
|
| 2996 | err = PERR_REMOTE;
|
| 2997 | goto out;
|
| 2998 | }
|
| 2999 |
|
| 3000 | |
| 3001 | * If parent is valid partition, enable local partiion.
|
| 3002 | * Otherwise, enable a remote partition.
|
| 3003 |
|
| 3004 | if (is_partition_valid(parent)) {
|
| 3005 | enum partition_cmd cmd = (new_prs == PRS_ROOT)
|
| 3006 | ? partcmd_enable : partcmd_enablei;
|
| 3007 |
|
| 3008 | err = update_parent_effective_cpumask(cs, cmd, NULL, &tmpmask);
|
| 3009 | } else {
|
| 3010 | err = remote_partition_enable(cs, new_prs, &tmpmask);
|
| 3011 | }
|
| 3012 | } else if (old_prs && new_prs) {
|
| 3013 | |
| 3014 | * A change in load balance state only, no change in cpumasks.
|
| 3015 | * Need to update isolated_cpus.
|
| 3016 |
|
| 3017 | if (((new_prs == PRS_ISOLATED) &&
|
| 3018 | !isolated_cpus_can_update(cs->effective_xcpus, NULL)) ||
|
| 3019 | prstate_housekeeping_conflict(new_prs, cs->effective_xcpus))
|
| 3020 | err = PERR_HKEEPING;
|
| 3021 | else
|
| 3022 | isolcpus_updated = true;
|
| 3023 | } else {
|
| 3024 | |
| 3025 | * Switching back to member is always allowed even if it
|
| 3026 | * disables child partitions.
|
| 3027 |
|
| 3028 | if (is_remote_partition(cs))
|
| 3029 | remote_partition_disable(cs, &tmpmask);
|
| 3030 | else
|
| 3031 | update_parent_effective_cpumask(cs, partcmd_disable,
|
| 3032 | NULL, &tmpmask);
|
| 3033 |
|
| 3034 | |
| 3035 | * Invalidation of child partitions will be done in
|
| 3036 | * update_cpumasks_hier().
|
| 3037 |
|
| 3038 | }
|
| 3039 | out:
|
| 3040 | |
| 3041 | * Make partition invalid & disable CS_CPU_EXCLUSIVE if an error
|
| 3042 | * happens.
|
| 3043 |
|
| 3044 | if (err) {
|
| 3045 | new_prs = -new_prs;
|
| 3046 | update_partition_exclusive_flag(cs, new_prs);
|
| 3047 | }
|
| 3048 |
|
| 3049 | spin_lock_irq(&callback_lock);
|
| 3050 | cs->partition_root_state = new_prs;
|
| 3051 | WRITE_ONCE(cs->prs_err, err);
|
| 3052 | if (!is_partition_valid(cs))
|
| 3053 | reset_partition_data(cs);
|
| 3054 | else if (isolcpus_updated)
|
| 3055 | isolated_cpus_update(old_prs, new_prs, cs->effective_xcpus);
|
| 3056 | spin_unlock_irq(&callback_lock);
|
| 3057 |
|
| 3058 |
|
| 3059 | update_cpumasks_hier(cs, &tmpmask, !new_prs);
|
| 3060 |
|
| 3061 |
|
| 3062 | WARN_ON_ONCE(!old_prs && (new_prs > 0)
|
| 3063 | && cpumask_empty(cs->effective_xcpus));
|
| 3064 |
|
| 3065 |
|
| 3066 | update_partition_sd_lb(cs, old_prs);
|
| 3067 |
|
| 3068 | notify_partition_change(cs, old_prs);
|
| 3069 | if (force_sd_rebuild)
|
| 3070 | rebuild_sched_domains_locked();
|
| 3071 | free_tmpmasks(&tmpmask);
|
| 3072 | return 0;
|
| 3073 | }
|
| 3074 |
|
| 3075 | |
| 3076 | * Check to see if a cpuset can accept a new task
|
| 3077 | * For v1, cpus_allowed and mems_allowed can't be empty.
|
| 3078 | * For v2, effective_cpus can't be empty.
|
| 3079 | * Note that in v1, effective_cpus = cpus_allowed.
|
| 3080 | *
|
| 3081 | * Also set the boolean flag passed in by @psetsched depending on if
|
| 3082 | * security_task_setscheduler() call is needed and @oldcs is not NULL.
|
| 3083 |
|
| 3084 | static int cpuset_can_attach_check(struct cpuset *cs, struct cpuset *oldcs,
|
| 3085 | bool *psetsched)
|
| 3086 | {
|
| 3087 | bool cpus_updated, mems_updated;
|
| 3088 |
|
| 3089 | if (cpumask_empty(cs->effective_cpus) ||
|
| 3090 | (!is_in_v2_mode() && nodes_empty(cs->mems_allowed)))
|
| 3091 | return -ENOSPC;
|
| 3092 |
|
| 3093 | if (!oldcs)
|
| 3094 | return 0;
|
| 3095 |
|
| 3096 | if (!llist_on_list(&oldcs->attach_node))
|
| 3097 | llist_add(&oldcs->attach_node, &src_cs_head);
|
| 3098 |
|
| 3099 | if (!llist_on_list(&cs->attach_node))
|
| 3100 | llist_add(&cs->attach_node, &dst_cs_head);
|
| 3101 |
|
| 3102 | cpus_updated = !cpumask_equal(cs->effective_cpus, oldcs->effective_cpus);
|
| 3103 | mems_updated = !nodes_equal(cs->effective_mems, oldcs->effective_mems);
|
| 3104 |
|
| 3105 | if (cpus_updated)
|
| 3106 | attach_ctx.cpus_updated = true;
|
| 3107 | if (mems_updated)
|
| 3108 | attach_ctx.mems_updated = true;
|
| 3109 |
|
| 3110 | |
| 3111 | * Skip rights over task setsched check in v2 when nothing changes for
|
| 3112 | * the current oldcs/cs pair, migration permission derives from
|
| 3113 | * hierarchy ownership in cgroup_procs_write_permission()).
|
| 3114 |
|
| 3115 | *psetsched = !cpuset_v2() || cpus_updated || mems_updated;
|
| 3116 |
|
| 3117 | |
| 3118 | * A v1 cpuset with tasks will have no CPU left only when CPU hotplug
|
| 3119 | * brings the last online CPU offline as users are not allowed to empty
|
| 3120 | * cpuset.cpus when there are active tasks inside. When that happens,
|
| 3121 | * we should allow tasks to migrate out without security check to make
|
| 3122 | * sure they will be able to run after migration.
|
| 3123 |
|
| 3124 | if (!is_in_v2_mode() && cpumask_empty(oldcs->effective_cpus))
|
| 3125 | *psetsched = false;
|
| 3126 |
|
| 3127 | return 0;
|
| 3128 | }
|
| 3129 |
|
| 3130 | static int cpuset_reserve_dl_bw(void)
|
| 3131 | {
|
| 3132 | struct cpuset *cs;
|
| 3133 | int cpu, ret;
|
| 3134 |
|
| 3135 | llist_for_each_entry(cs, dst_cs_head.first, attach_node) {
|
| 3136 | if (!cs->sum_migrate_dl_bw)
|
| 3137 | continue;
|
| 3138 |
|
| 3139 | cpu = cpumask_any_and(cpu_active_mask, cs->effective_cpus);
|
| 3140 | if (unlikely(cpu >= nr_cpu_ids))
|
| 3141 | return -EINVAL;
|
| 3142 |
|
| 3143 | ret = dl_bw_alloc(cpu, cs->sum_migrate_dl_bw);
|
| 3144 | if (ret)
|
| 3145 | return ret;
|
| 3146 |
|
| 3147 | cs->dl_bw_cpu = cpu;
|
| 3148 | }
|
| 3149 | return 0;
|
| 3150 | }
|
| 3151 |
|
| 3152 | |
| 3153 | * Clear and optionally apply (@cancel is false) the attach related data in the
|
| 3154 | * source or destination cpuset.
|
| 3155 |
|
| 3156 | static void clear_attach_data(struct llist_head *head, bool cancel)
|
| 3157 | {
|
| 3158 | struct cpuset *cs, *next;
|
| 3159 | struct llist_node *lnode = __llist_del_all(head);
|
| 3160 |
|
| 3161 | llist_for_each_entry_safe(cs, next, lnode, attach_node) {
|
| 3162 | init_llist_node(&cs->attach_node);
|
| 3163 | if (cs->nr_migrate_dl_tasks) {
|
| 3164 | if (!cancel)
|
| 3165 | atomic_add(cs->nr_migrate_dl_tasks, &cs->nr_deadline_tasks);
|
| 3166 | else if (cs->dl_bw_cpu >= 0)
|
| 3167 | dl_bw_free(cs->dl_bw_cpu, cs->sum_migrate_dl_bw);
|
| 3168 | cs->nr_migrate_dl_tasks = 0;
|
| 3169 | cs->sum_migrate_dl_bw = 0;
|
| 3170 | cs->dl_bw_cpu = -1;
|
| 3171 | }
|
| 3172 | }
|
| 3173 | }
|
| 3174 |
|
| 3175 |
|
| 3176 | static int cpuset_can_attach(struct cgroup_taskset *tset)
|
| 3177 | {
|
| 3178 | struct cgroup_subsys_state *css;
|
| 3179 | struct cpuset *cs, *oldcs;
|
| 3180 | struct task_struct *task;
|
| 3181 | bool setsched_check;
|
| 3182 | int ret;
|
| 3183 |
|
| 3184 | cs = oldcs = NULL;
|
| 3185 | mutex_lock(&cpuset_mutex);
|
| 3186 | attach_ctx.old_cs = NULL;
|
| 3187 | attach_ctx.cpus_updated = false;
|
| 3188 | attach_ctx.mems_updated = false;
|
| 3189 | attach_ctx.many_dest_cs = false;
|
| 3190 |
|
| 3191 | |
| 3192 | * The attach_ctx.old_cs is used mainly by cpuset_migrate_mm() to get
|
| 3193 | * the old_mems_allowed value. There are two ways that many-to-one
|
| 3194 | * cpuset migration can happen:
|
| 3195 | * 1) A multithread application with threads in different cpusets is
|
| 3196 | * wholely migrated to a new cpuset.
|
| 3197 | * 2) Disabling v2 cpuset controller will move all the tasks in child
|
| 3198 | * cpusets to the parent cpuset.
|
| 3199 | *
|
| 3200 | * In the former case, it is the mm setting of the group leader that
|
| 3201 | * really matters. So attach_ctx.old_cs should track the oldcs of the
|
| 3202 | * group leader. It falls back to the oldcs of the first task if there
|
| 3203 | * is no group leader in the taskset. In the latter case, effective_mems
|
| 3204 | * of child cpusets must always be a subset of the parent. So no real
|
| 3205 | * page migration will be necessary no matter which child cpuset is
|
| 3206 | * selected as attach_ctx.old_cs.
|
| 3207 | *
|
| 3208 | * For a v2 threaded subtree where cpuset isn't enabled in some of the
|
| 3209 | * cgroups, it is possible that oldcs == cs for some of the tasks.
|
| 3210 | * In this case, we can skip checking on those tasks as there is no
|
| 3211 | * actual migration wrt cpuset.
|
| 3212 |
|
| 3213 | cgroup_taskset_for_each(task, css, tset) {
|
| 3214 | struct cpuset *new_cs = css_cs(css);
|
| 3215 | struct cpuset *new_oldcs = task_cs(task);
|
| 3216 |
|
| 3217 | if ((new_oldcs != oldcs) || (new_cs != cs)) {
|
| 3218 | if (cs && (new_cs != cs))
|
| 3219 | attach_ctx.many_dest_cs = true;
|
| 3220 | cs = new_cs;
|
| 3221 | oldcs = new_oldcs;
|
| 3222 | if (oldcs == cs)
|
| 3223 | continue;
|
| 3224 | if (!attach_ctx.old_cs)
|
| 3225 | attach_ctx.old_cs = oldcs;
|
| 3226 | ret = cpuset_can_attach_check(cs, oldcs, &setsched_check);
|
| 3227 | if (ret)
|
| 3228 | goto out_unlock;
|
| 3229 | }
|
| 3230 |
|
| 3231 | if (oldcs == cs)
|
| 3232 | continue;
|
| 3233 |
|
| 3234 | ret = task_can_attach(task);
|
| 3235 | if (ret)
|
| 3236 | goto out_unlock;
|
| 3237 |
|
| 3238 |
|
| 3239 | if (task == task->group_leader)
|
| 3240 | attach_ctx.old_cs = task_cs(task);
|
| 3241 |
|
| 3242 | if (setsched_check) {
|
| 3243 | ret = security_task_setscheduler(task);
|
| 3244 | if (ret)
|
| 3245 | goto out_unlock;
|
| 3246 | }
|
| 3247 |
|
| 3248 | if (dl_task(task)) {
|
| 3249 | |
| 3250 | * Count all migrating DL tasks for cpuset task accounting.
|
| 3251 | * Only tasks that need a root-domain bandwidth move
|
| 3252 | * contribute to sum_migrate_dl_bw.
|
| 3253 |
|
| 3254 | cs->nr_migrate_dl_tasks++;
|
| 3255 | oldcs->nr_migrate_dl_tasks--;
|
| 3256 | if (dl_task_needs_bw_move(task, cs->effective_cpus))
|
| 3257 | cs->sum_migrate_dl_bw += task->dl.dl_bw;
|
| 3258 | }
|
| 3259 | }
|
| 3260 |
|
| 3261 | |
| 3262 | * The only case where there are multiple destination cpusets for
|
| 3263 | * task migration is when enabling a v2 cpuset controllers where
|
| 3264 | * tasks will be migrated to multiple child cpusets from a parent
|
| 3265 | * cpuset with the same effective CPUs and memory nodes. IOW,
|
| 3266 | * both attach_cpus_updated and attach_mems_updated should be false.
|
| 3267 | * If not, it is a condition that the current code cannot handle.
|
| 3268 | * Print a warning and abort the attach operation as further code
|
| 3269 | * change may be needed.
|
| 3270 |
|
| 3271 | if (WARN_ON_ONCE(attach_ctx.many_dest_cs && (!cpuset_v2() ||
|
| 3272 | attach_ctx.cpus_updated || attach_ctx.mems_updated))) {
|
| 3273 | ret = -EINVAL;
|
| 3274 | goto out_unlock;
|
| 3275 | }
|
| 3276 |
|
| 3277 | ret = cpuset_reserve_dl_bw();
|
| 3278 |
|
| 3279 | out_unlock:
|
| 3280 | if (ret) {
|
| 3281 | clear_attach_data(&src_cs_head, true);
|
| 3282 | clear_attach_data(&dst_cs_head, true);
|
| 3283 | } else {
|
| 3284 | attach_ctx.in_progress++;
|
| 3285 | }
|
| 3286 |
|
| 3287 | mutex_unlock(&cpuset_mutex);
|
| 3288 | return ret;
|
| 3289 | }
|
| 3290 |
|
| 3291 | static void cpuset_cancel_attach(struct cgroup_taskset *tset)
|
| 3292 | {
|
| 3293 | mutex_lock(&cpuset_mutex);
|
| 3294 | dec_attach_in_progress_locked();
|
| 3295 | clear_attach_data(&src_cs_head, true);
|
| 3296 | clear_attach_data(&dst_cs_head, true);
|
| 3297 | mutex_unlock(&cpuset_mutex);
|
| 3298 | }
|
| 3299 |
|
| 3300 | |
| 3301 | * Protected by cpuset_mutex. cpus_attach is used only by cpuset_attach_task()
|
| 3302 | * but we can't allocate it dynamically there. Define it global and
|
| 3303 | * allocate from cpuset_init().
|
| 3304 |
|
| 3305 | static cpumask_var_t cpus_attach;
|
| 3306 |
|
| 3307 | static void cpuset_attach_task(struct cpuset *cs, struct task_struct *task)
|
| 3308 | {
|
| 3309 | struct mm_struct *mm;
|
| 3310 |
|
| 3311 | lockdep_assert_cpuset_lock_held();
|
| 3312 |
|
| 3313 | if (cs != &top_cpuset)
|
| 3314 | guarantee_active_cpus(task, cpus_attach);
|
| 3315 | else
|
| 3316 | cpumask_andnot(cpus_attach, task_cpu_possible_mask(task),
|
| 3317 | subpartitions_cpus);
|
| 3318 | |
| 3319 | * can_attach beforehand should guarantee that this doesn't
|
| 3320 | * fail. TODO: have a better way to handle failure here
|
| 3321 |
|
| 3322 | WARN_ON_ONCE(set_cpus_allowed_ptr(task, cpus_attach));
|
| 3323 |
|
| 3324 | if (cpuset_v2() && !attach_ctx.mems_updated)
|
| 3325 | return;
|
| 3326 |
|
| 3327 | cpuset_change_task_nodemask(task, &attach_ctx.nodemask_to);
|
| 3328 | cpuset1_update_task_spread_flags(cs, task);
|
| 3329 |
|
| 3330 | if ((task != task->group_leader) || !attach_ctx.mems_updated)
|
| 3331 | return;
|
| 3332 |
|
| 3333 | |
| 3334 | * Change mm for threadgroup leader. This is expensive and may
|
| 3335 | * sleep and should be moved outside migration path proper.
|
| 3336 |
|
| 3337 | mm = get_task_mm(task);
|
| 3338 | if (mm) {
|
| 3339 | struct cpuset *oldcs = attach_ctx.old_cs;
|
| 3340 |
|
| 3341 | mpol_rebind_mm(mm, &cs->effective_mems);
|
| 3342 |
|
| 3343 | |
| 3344 | * old_mems_allowed is the same with mems_allowed
|
| 3345 | * here, except if this task is being moved
|
| 3346 | * automatically due to hotplug. In that case
|
| 3347 | * @mems_allowed has been updated and is empty, so
|
| 3348 | * @old_mems_allowed is the right nodesets that we
|
| 3349 | * migrate mm from.
|
| 3350 |
|
| 3351 | if (is_memory_migrate(cs)) {
|
| 3352 | cpuset_migrate_mm(mm, &oldcs->old_mems_allowed,
|
| 3353 | &attach_ctx.nodemask_to);
|
| 3354 | attach_ctx.task_work_queued = true;
|
| 3355 | } else {
|
| 3356 | mmput(mm);
|
| 3357 | }
|
| 3358 | }
|
| 3359 | }
|
| 3360 |
|
| 3361 | static void cpuset_attach(struct cgroup_taskset *tset)
|
| 3362 | {
|
| 3363 | struct task_struct *task;
|
| 3364 | struct cgroup_subsys_state *css;
|
| 3365 | struct cpuset *cs;
|
| 3366 |
|
| 3367 | cgroup_taskset_first(tset, &css);
|
| 3368 | cs = css_cs(css);
|
| 3369 |
|
| 3370 | lockdep_assert_cpus_held();
|
| 3371 | mutex_lock(&cpuset_mutex);
|
| 3372 | attach_ctx.task_work_queued = false;
|
| 3373 | guarantee_online_mems(cs, &attach_ctx.nodemask_to);
|
| 3374 |
|
| 3375 | |
| 3376 | * attach_ctx.old_cs can only be NULL if no task is actually migrating.
|
| 3377 | * This is highly unlikely. If it happens at all, we can skip task
|
| 3378 | * iteration and setting old_mems_allowed.
|
| 3379 |
|
| 3380 | if (unlikely(!attach_ctx.old_cs))
|
| 3381 | goto out;
|
| 3382 |
|
| 3383 | |
| 3384 | * In the default hierarchy, enabling cpuset in the child cgroups
|
| 3385 | * will trigger a cpuset_attach() call with no change in effective cpus
|
| 3386 | * and mems. In that case, we can optimize out by skipping the task
|
| 3387 | * iteration and the destination cpuset list is iterated to set
|
| 3388 | * old_mems_allowed.
|
| 3389 |
|
| 3390 | if (cpuset_v2() && !attach_ctx.cpus_updated && !attach_ctx.mems_updated) {
|
| 3391 | llist_for_each_entry(cs, dst_cs_head.first, attach_node)
|
| 3392 | cs->old_mems_allowed = attach_ctx.nodemask_to;
|
| 3393 | goto out;
|
| 3394 | }
|
| 3395 |
|
| 3396 | cgroup_taskset_for_each(task, css, tset)
|
| 3397 | cpuset_attach_task(cs, task);
|
| 3398 |
|
| 3399 | if (attach_ctx.task_work_queued)
|
| 3400 | schedule_flush_migrate_mm();
|
| 3401 | cs->old_mems_allowed = attach_ctx.nodemask_to;
|
| 3402 | out:
|
| 3403 | clear_attach_data(&src_cs_head, false);
|
| 3404 | clear_attach_data(&dst_cs_head, false);
|
| 3405 | dec_attach_in_progress_locked();
|
| 3406 |
|
| 3407 | mutex_unlock(&cpuset_mutex);
|
| 3408 | }
|
| 3409 |
|
| 3410 | |
| 3411 | * Common handling for a write to a "cpus" or "mems" file.
|
| 3412 |
|
| 3413 | ssize_t cpuset_write_resmask(struct kernfs_open_file *of,
|
| 3414 | char *buf, size_t nbytes, loff_t off)
|
| 3415 | {
|
| 3416 | struct cpuset *cs = css_cs(of_css(of));
|
| 3417 | struct cpuset *trialcs;
|
| 3418 | int retval = -ENODEV;
|
| 3419 |
|
| 3420 |
|
| 3421 | if (cs == &top_cpuset)
|
| 3422 | return -EACCES;
|
| 3423 |
|
| 3424 | buf = strstrip(buf);
|
| 3425 |
|
| 3426 |
|
| 3427 | mutex_lock(&cpuset_top_mutex);
|
| 3428 | cpus_read_lock();
|
| 3429 | wait_attach_done_lock();
|
| 3430 |
|
| 3431 | if (!is_cpuset_online(cs))
|
| 3432 | goto out_unlock;
|
| 3433 |
|
| 3434 | trialcs = dup_or_alloc_cpuset(cs);
|
| 3435 | if (!trialcs) {
|
| 3436 | retval = -ENOMEM;
|
| 3437 | goto out_unlock;
|
| 3438 | }
|
| 3439 |
|
| 3440 | switch (of_cft(of)->private) {
|
| 3441 | case FILE_CPULIST:
|
| 3442 | retval = update_cpumask(cs, trialcs, buf);
|
| 3443 | break;
|
| 3444 | case FILE_EXCLUSIVE_CPULIST:
|
| 3445 | retval = update_exclusive_cpumask(cs, trialcs, buf);
|
| 3446 | break;
|
| 3447 | case FILE_MEMLIST:
|
| 3448 | retval = update_nodemask(cs, trialcs, buf);
|
| 3449 | break;
|
| 3450 | default:
|
| 3451 | retval = -EINVAL;
|
| 3452 | break;
|
| 3453 | }
|
| 3454 |
|
| 3455 | free_cpuset(trialcs);
|
| 3456 | out_unlock:
|
| 3457 | cpuset_update_sd_hk_unlock();
|
| 3458 | if (of_cft(of)->private == FILE_MEMLIST)
|
| 3459 | schedule_flush_migrate_mm();
|
| 3460 | return retval ?: nbytes;
|
| 3461 | }
|
| 3462 |
|
| 3463 | |
| 3464 | * These ascii lists should be read in a single call, by using a user
|
| 3465 | * buffer large enough to hold the entire map. If read in smaller
|
| 3466 | * chunks, there is no guarantee of atomicity. Since the display format
|
| 3467 | * used, list of ranges of sequential numbers, is variable length,
|
| 3468 | * and since these maps can change value dynamically, one could read
|
| 3469 | * gibberish by doing partial reads while a list was changing.
|
| 3470 |
|
| 3471 | int cpuset_common_seq_show(struct seq_file *sf, void *v)
|
| 3472 | {
|
| 3473 | struct cpuset *cs = css_cs(seq_css(sf));
|
| 3474 | cpuset_filetype_t type = seq_cft(sf)->private;
|
| 3475 | int ret = 0;
|
| 3476 |
|
| 3477 | spin_lock_irq(&callback_lock);
|
| 3478 |
|
| 3479 | switch (type) {
|
| 3480 | case FILE_CPULIST:
|
| 3481 | seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->cpus_allowed));
|
| 3482 | break;
|
| 3483 | case FILE_MEMLIST:
|
| 3484 | seq_printf(sf, "%*pbl\n", nodemask_pr_args(&cs->mems_allowed));
|
| 3485 | break;
|
| 3486 | case FILE_EFFECTIVE_CPULIST:
|
| 3487 | seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->effective_cpus));
|
| 3488 | break;
|
| 3489 | case FILE_EFFECTIVE_MEMLIST:
|
| 3490 | seq_printf(sf, "%*pbl\n", nodemask_pr_args(&cs->effective_mems));
|
| 3491 | break;
|
| 3492 | case FILE_EXCLUSIVE_CPULIST:
|
| 3493 | seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->exclusive_cpus));
|
| 3494 | break;
|
| 3495 | case FILE_EFFECTIVE_XCPULIST:
|
| 3496 | seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->effective_xcpus));
|
| 3497 | break;
|
| 3498 | case FILE_SUBPARTS_CPULIST:
|
| 3499 | seq_printf(sf, "%*pbl\n", cpumask_pr_args(subpartitions_cpus));
|
| 3500 | break;
|
| 3501 | case FILE_ISOLATED_CPULIST:
|
| 3502 | seq_printf(sf, "%*pbl\n", cpumask_pr_args(isolated_cpus));
|
| 3503 | break;
|
| 3504 | default:
|
| 3505 | ret = -EINVAL;
|
| 3506 | }
|
| 3507 |
|
| 3508 | spin_unlock_irq(&callback_lock);
|
| 3509 | return ret;
|
| 3510 | }
|
| 3511 |
|
| 3512 | static int cpuset_partition_show(struct seq_file *seq, void *v)
|
| 3513 | {
|
| 3514 | struct cpuset *cs = css_cs(seq_css(seq));
|
| 3515 | const char *err, *type = NULL;
|
| 3516 |
|
| 3517 | switch (cs->partition_root_state) {
|
| 3518 | case PRS_ROOT:
|
| 3519 | seq_puts(seq, "root\n");
|
| 3520 | break;
|
| 3521 | case PRS_ISOLATED:
|
| 3522 | seq_puts(seq, "isolated\n");
|
| 3523 | break;
|
| 3524 | case PRS_MEMBER:
|
| 3525 | seq_puts(seq, "member\n");
|
| 3526 | break;
|
| 3527 | case PRS_INVALID_ROOT:
|
| 3528 | type = "root";
|
| 3529 | fallthrough;
|
| 3530 | case PRS_INVALID_ISOLATED:
|
| 3531 | if (!type)
|
| 3532 | type = "isolated";
|
| 3533 | err = perr_strings[READ_ONCE(cs->prs_err)];
|
| 3534 | if (err)
|
| 3535 | seq_printf(seq, "%s invalid (%s)\n", type, err);
|
| 3536 | else
|
| 3537 | seq_printf(seq, "%s invalid\n", type);
|
| 3538 | break;
|
| 3539 | }
|
| 3540 | return 0;
|
| 3541 | }
|
| 3542 |
|
| 3543 | static ssize_t cpuset_partition_write(struct kernfs_open_file *of, char *buf,
|
| 3544 | size_t nbytes, loff_t off)
|
| 3545 | {
|
| 3546 | struct cpuset *cs = css_cs(of_css(of));
|
| 3547 | int val;
|
| 3548 | int retval = -ENODEV;
|
| 3549 |
|
| 3550 | buf = strstrip(buf);
|
| 3551 |
|
| 3552 | if (!strcmp(buf, "root"))
|
| 3553 | val = PRS_ROOT;
|
| 3554 | else if (!strcmp(buf, "member"))
|
| 3555 | val = PRS_MEMBER;
|
| 3556 | else if (!strcmp(buf, "isolated"))
|
| 3557 | val = PRS_ISOLATED;
|
| 3558 | else
|
| 3559 | return -EINVAL;
|
| 3560 |
|
| 3561 | mutex_lock(&cpuset_top_mutex);
|
| 3562 | cpus_read_lock();
|
| 3563 | wait_attach_done_lock();
|
| 3564 |
|
| 3565 | if (is_cpuset_online(cs))
|
| 3566 | retval = update_prstate(cs, val);
|
| 3567 | cpuset_update_sd_hk_unlock();
|
| 3568 | return retval ?: nbytes;
|
| 3569 | }
|
| 3570 |
|
| 3571 | |
| 3572 | * This is currently a minimal set for the default hierarchy. It can be
|
| 3573 | * expanded later on by migrating more features and control files from v1.
|
| 3574 |
|
| 3575 | static struct cftype dfl_files[] = {
|
| 3576 | {
|
| 3577 | .name = "cpus",
|
| 3578 | .seq_show = cpuset_common_seq_show,
|
| 3579 | .write = cpuset_write_resmask,
|
| 3580 | .max_write_len = (100U + 6 * NR_CPUS),
|
| 3581 | .private = FILE_CPULIST,
|
| 3582 | .flags = CFTYPE_NOT_ON_ROOT,
|
| 3583 | },
|
| 3584 |
|
| 3585 | {
|
| 3586 | .name = "mems",
|
| 3587 | .seq_show = cpuset_common_seq_show,
|
| 3588 | .write = cpuset_write_resmask,
|
| 3589 | .max_write_len = (100U + 6 * MAX_NUMNODES),
|
| 3590 | .private = FILE_MEMLIST,
|
| 3591 | .flags = CFTYPE_NOT_ON_ROOT,
|
| 3592 | },
|
| 3593 |
|
| 3594 | {
|
| 3595 | .name = "cpus.effective",
|
| 3596 | .seq_show = cpuset_common_seq_show,
|
| 3597 | .private = FILE_EFFECTIVE_CPULIST,
|
| 3598 | },
|
| 3599 |
|
| 3600 | {
|
| 3601 | .name = "mems.effective",
|
| 3602 | .seq_show = cpuset_common_seq_show,
|
| 3603 | .private = FILE_EFFECTIVE_MEMLIST,
|
| 3604 | },
|
| 3605 |
|
| 3606 | {
|
| 3607 | .name = "cpus.partition",
|
| 3608 | .seq_show = cpuset_partition_show,
|
| 3609 | .write = cpuset_partition_write,
|
| 3610 | .private = FILE_PARTITION_ROOT,
|
| 3611 | .flags = CFTYPE_NOT_ON_ROOT,
|
| 3612 | .file_offset = offsetof(struct cpuset, partition_file),
|
| 3613 | },
|
| 3614 |
|
| 3615 | {
|
| 3616 | .name = "cpus.exclusive",
|
| 3617 | .seq_show = cpuset_common_seq_show,
|
| 3618 | .write = cpuset_write_resmask,
|
| 3619 | .max_write_len = (100U + 6 * NR_CPUS),
|
| 3620 | .private = FILE_EXCLUSIVE_CPULIST,
|
| 3621 | .flags = CFTYPE_NOT_ON_ROOT,
|
| 3622 | },
|
| 3623 |
|
| 3624 | {
|
| 3625 | .name = "cpus.exclusive.effective",
|
| 3626 | .seq_show = cpuset_common_seq_show,
|
| 3627 | .private = FILE_EFFECTIVE_XCPULIST,
|
| 3628 | .flags = CFTYPE_NOT_ON_ROOT,
|
| 3629 | },
|
| 3630 |
|
| 3631 | {
|
| 3632 | .name = "cpus.subpartitions",
|
| 3633 | .seq_show = cpuset_common_seq_show,
|
| 3634 | .private = FILE_SUBPARTS_CPULIST,
|
| 3635 | .flags = CFTYPE_ONLY_ON_ROOT | CFTYPE_DEBUG,
|
| 3636 | },
|
| 3637 |
|
| 3638 | {
|
| 3639 | .name = "cpus.isolated",
|
| 3640 | .seq_show = cpuset_common_seq_show,
|
| 3641 | .private = FILE_ISOLATED_CPULIST,
|
| 3642 | .flags = CFTYPE_ONLY_ON_ROOT,
|
| 3643 | },
|
| 3644 |
|
| 3645 | { }
|
| 3646 | };
|
| 3647 |
|
| 3648 |
|
| 3649 | |
| 3650 | * cpuset_css_alloc - Allocate a cpuset css
|
| 3651 | * @parent_css: Parent css of the control group that the new cpuset will be
|
| 3652 | * part of
|
| 3653 | * Return: cpuset css on success, -ENOMEM on failure.
|
| 3654 | *
|
| 3655 | * Allocate and initialize a new cpuset css, for non-NULL @parent_css, return
|
| 3656 | * top cpuset css otherwise.
|
| 3657 |
|
| 3658 | static struct cgroup_subsys_state *
|
| 3659 | cpuset_css_alloc(struct cgroup_subsys_state *parent_css)
|
| 3660 | {
|
| 3661 | struct cpuset *cs;
|
| 3662 |
|
| 3663 | if (!parent_css)
|
| 3664 | return &top_cpuset.css;
|
| 3665 |
|
| 3666 | cs = dup_or_alloc_cpuset(NULL);
|
| 3667 | if (!cs)
|
| 3668 | return ERR_PTR(-ENOMEM);
|
| 3669 |
|
| 3670 | __set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
|
| 3671 | cpuset1_init(cs);
|
| 3672 |
|
| 3673 |
|
| 3674 | if (cpuset_v2())
|
| 3675 | __set_bit(CS_MEMORY_MIGRATE, &cs->flags);
|
| 3676 |
|
| 3677 | return &cs->css;
|
| 3678 | }
|
| 3679 |
|
| 3680 | static int cpuset_css_online(struct cgroup_subsys_state *css)
|
| 3681 | {
|
| 3682 | struct cpuset *cs = css_cs(css);
|
| 3683 | struct cpuset *parent = parent_cs(cs);
|
| 3684 |
|
| 3685 | if (!parent)
|
| 3686 | return 0;
|
| 3687 |
|
| 3688 | cpuset_full_lock();
|
| 3689 | |
| 3690 | * For v2, clear CS_SCHED_LOAD_BALANCE if parent is isolated
|
| 3691 |
|
| 3692 | if (cpuset_v2() && !is_sched_load_balance(parent))
|
| 3693 | clear_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
|
| 3694 |
|
| 3695 | cpuset_inc();
|
| 3696 |
|
| 3697 | spin_lock_irq(&callback_lock);
|
| 3698 | if (is_in_v2_mode()) {
|
| 3699 | cpumask_copy(cs->effective_cpus, parent->effective_cpus);
|
| 3700 | cs->effective_mems = parent->effective_mems;
|
| 3701 | }
|
| 3702 | spin_unlock_irq(&callback_lock);
|
| 3703 | cpuset1_online_css(css);
|
| 3704 |
|
| 3705 | cpuset_full_unlock();
|
| 3706 | return 0;
|
| 3707 | }
|
| 3708 |
|
| 3709 | |
| 3710 | * If the cpuset being removed has its flag 'sched_load_balance'
|
| 3711 | * enabled, then simulate turning sched_load_balance off, which
|
| 3712 | * will call rebuild_sched_domains_locked(). That is not needed
|
| 3713 | * in the default hierarchy where only changes in partition
|
| 3714 | * will cause repartitioning.
|
| 3715 |
|
| 3716 | static void cpuset_css_offline(struct cgroup_subsys_state *css)
|
| 3717 | {
|
| 3718 | struct cpuset *cs = css_cs(css);
|
| 3719 |
|
| 3720 | cpuset_full_lock();
|
| 3721 | if (!cpuset_v2() && is_sched_load_balance(cs))
|
| 3722 | cpuset_update_flag(CS_SCHED_LOAD_BALANCE, cs, 0);
|
| 3723 |
|
| 3724 | cpuset_dec();
|
| 3725 | cpuset_full_unlock();
|
| 3726 | }
|
| 3727 |
|
| 3728 | |
| 3729 | * If a dying cpuset has the 'cpus.partition' enabled, turn it off by
|
| 3730 | * changing it back to member to free its exclusive CPUs back to the pool to
|
| 3731 | * be used by other online cpusets.
|
| 3732 |
|
| 3733 | static void cpuset_css_killed(struct cgroup_subsys_state *css)
|
| 3734 | {
|
| 3735 | struct cpuset *cs = css_cs(css);
|
| 3736 |
|
| 3737 | cpuset_full_lock();
|
| 3738 |
|
| 3739 | if (is_partition_valid(cs))
|
| 3740 | update_prstate(cs, PRS_MEMBER);
|
| 3741 | cpuset_update_sd_hk_unlock();
|
| 3742 | }
|
| 3743 |
|
| 3744 | static void cpuset_css_free(struct cgroup_subsys_state *css)
|
| 3745 | {
|
| 3746 | struct cpuset *cs = css_cs(css);
|
| 3747 |
|
| 3748 | free_cpuset(cs);
|
| 3749 | }
|
| 3750 |
|
| 3751 | static void cpuset_bind(struct cgroup_subsys_state *root_css)
|
| 3752 | {
|
| 3753 | mutex_lock(&cpuset_mutex);
|
| 3754 | spin_lock_irq(&callback_lock);
|
| 3755 |
|
| 3756 | WRITE_ONCE(cpuset_v2_mode,
|
| 3757 | !!(cpuset_cgrp_subsys.root->flags & CGRP_ROOT_CPUSET_V2_MODE));
|
| 3758 | if (is_in_v2_mode()) {
|
| 3759 | cpumask_copy(top_cpuset.cpus_allowed, cpu_possible_mask);
|
| 3760 | cpumask_copy(top_cpuset.effective_xcpus, cpu_possible_mask);
|
| 3761 | top_cpuset.mems_allowed = node_possible_map;
|
| 3762 | } else {
|
| 3763 | cpumask_copy(top_cpuset.cpus_allowed,
|
| 3764 | top_cpuset.effective_cpus);
|
| 3765 | top_cpuset.mems_allowed = top_cpuset.effective_mems;
|
| 3766 | }
|
| 3767 |
|
| 3768 | spin_unlock_irq(&callback_lock);
|
| 3769 | mutex_unlock(&cpuset_mutex);
|
| 3770 | }
|
| 3771 |
|
| 3772 | |
| 3773 | * In case the child is cloned into a cpuset different from its parent,
|
| 3774 | * additional checks are done to see if the move is allowed.
|
| 3775 |
|
| 3776 | static int cpuset_can_fork(struct task_struct *task, struct css_set *cset)
|
| 3777 | {
|
| 3778 | struct cpuset *cs = css_cs(cset->subsys[cpuset_cgrp_id]);
|
| 3779 | bool same_cs;
|
| 3780 | int ret;
|
| 3781 |
|
| 3782 | rcu_read_lock();
|
| 3783 | same_cs = (cs == task_cs(current));
|
| 3784 | rcu_read_unlock();
|
| 3785 |
|
| 3786 | if (same_cs)
|
| 3787 | return 0;
|
| 3788 |
|
| 3789 | lockdep_assert_held(&cgroup_mutex);
|
| 3790 | mutex_lock(&cpuset_mutex);
|
| 3791 |
|
| 3792 |
|
| 3793 | ret = cpuset_can_attach_check(cs, NULL, NULL);
|
| 3794 | if (ret)
|
| 3795 | goto out_unlock;
|
| 3796 |
|
| 3797 | ret = task_can_attach(task);
|
| 3798 | if (ret)
|
| 3799 | goto out_unlock;
|
| 3800 |
|
| 3801 | ret = security_task_setscheduler(task);
|
| 3802 | if (ret)
|
| 3803 | goto out_unlock;
|
| 3804 |
|
| 3805 | attach_ctx.in_progress++;
|
| 3806 | out_unlock:
|
| 3807 | mutex_unlock(&cpuset_mutex);
|
| 3808 | return ret;
|
| 3809 | }
|
| 3810 |
|
| 3811 | static void cpuset_cancel_fork(struct task_struct *task, struct css_set *cset)
|
| 3812 | {
|
| 3813 | struct cpuset *cs = css_cs(cset->subsys[cpuset_cgrp_id]);
|
| 3814 | bool same_cs;
|
| 3815 |
|
| 3816 | rcu_read_lock();
|
| 3817 | same_cs = (cs == task_cs(current));
|
| 3818 | rcu_read_unlock();
|
| 3819 |
|
| 3820 | if (same_cs)
|
| 3821 | return;
|
| 3822 |
|
| 3823 | dec_attach_in_progress();
|
| 3824 | }
|
| 3825 |
|
| 3826 | |
| 3827 | * Make sure the new task conform to the current state of its parent,
|
| 3828 | * which could have been changed by cpuset just after it inherits the
|
| 3829 | * state from the parent and before it sits on the cgroup's task list.
|
| 3830 |
|
| 3831 | static void cpuset_fork(struct task_struct *task)
|
| 3832 | {
|
| 3833 | struct cpuset *cs, *oldcs;
|
| 3834 |
|
| 3835 | rcu_read_lock();
|
| 3836 | cs = task_cs(task);
|
| 3837 | oldcs = task_cs(current);
|
| 3838 | rcu_read_unlock();
|
| 3839 |
|
| 3840 | if (cs == oldcs) {
|
| 3841 | if (cs == &top_cpuset)
|
| 3842 | return;
|
| 3843 |
|
| 3844 | set_cpus_allowed_ptr(task, current->cpus_ptr);
|
| 3845 | task->mems_allowed = current->mems_allowed;
|
| 3846 | return;
|
| 3847 | }
|
| 3848 |
|
| 3849 |
|
| 3850 | mutex_lock(&cpuset_mutex);
|
| 3851 | guarantee_online_mems(cs, &attach_ctx.nodemask_to);
|
| 3852 | cs->old_mems_allowed = attach_ctx.nodemask_to;
|
| 3853 |
|
| 3854 | |
| 3855 | * Assume CPUs and memory nodes are updated
|
| 3856 | * A CLONE_INTO_CGROUP operation should have taken the cgroup mutex
|
| 3857 | * and so there shouldn't be a competing cpuset_attach() operation.
|
| 3858 |
|
| 3859 | attach_ctx.cpus_updated = attach_ctx.mems_updated = true;
|
| 3860 | attach_ctx.task_work_queued = false;
|
| 3861 | attach_ctx.old_cs = oldcs;
|
| 3862 | cpuset_attach_task(cs, task);
|
| 3863 | if (attach_ctx.task_work_queued)
|
| 3864 | schedule_flush_migrate_mm();
|
| 3865 |
|
| 3866 | dec_attach_in_progress_locked();
|
| 3867 | mutex_unlock(&cpuset_mutex);
|
| 3868 | }
|
| 3869 |
|
| 3870 | struct cgroup_subsys cpuset_cgrp_subsys = {
|
| 3871 | .css_alloc = cpuset_css_alloc,
|
| 3872 | .css_online = cpuset_css_online,
|
| 3873 | .css_offline = cpuset_css_offline,
|
| 3874 | .css_killed = cpuset_css_killed,
|
| 3875 | .css_free = cpuset_css_free,
|
| 3876 | .can_attach = cpuset_can_attach,
|
| 3877 | .cancel_attach = cpuset_cancel_attach,
|
| 3878 | .attach = cpuset_attach,
|
| 3879 | .bind = cpuset_bind,
|
| 3880 | .can_fork = cpuset_can_fork,
|
| 3881 | .cancel_fork = cpuset_cancel_fork,
|
| 3882 | .fork = cpuset_fork,
|
| 3883 | #ifdef CONFIG_CPUSETS_V1
|
| 3884 | .legacy_cftypes = cpuset1_files,
|
| 3885 | #endif
|
| 3886 | .dfl_cftypes = dfl_files,
|
| 3887 | .early_init = true,
|
| 3888 | .threaded = true,
|
| 3889 | };
|
| 3890 |
|
| 3891 | |
| 3892 | * cpuset_init - initialize cpusets at system boot
|
| 3893 | *
|
| 3894 | * Description: Initialize top_cpuset
|
| 3895 | *
|
| 3896 |
|
| 3897 | int __init cpuset_init(void)
|
| 3898 | {
|
| 3899 | BUG_ON(!alloc_cpumask_var(&top_cpuset.cpus_allowed, GFP_KERNEL));
|
| 3900 | BUG_ON(!alloc_cpumask_var(&top_cpuset.effective_cpus, GFP_KERNEL));
|
| 3901 | BUG_ON(!alloc_cpumask_var(&top_cpuset.effective_xcpus, GFP_KERNEL));
|
| 3902 | BUG_ON(!alloc_cpumask_var(&top_cpuset.exclusive_cpus, GFP_KERNEL));
|
| 3903 | BUG_ON(!zalloc_cpumask_var(&subpartitions_cpus, GFP_KERNEL));
|
| 3904 | BUG_ON(!zalloc_cpumask_var(&isolated_cpus, GFP_KERNEL));
|
| 3905 | BUG_ON(!zalloc_cpumask_var(&isolated_hk_cpus, GFP_KERNEL));
|
| 3906 |
|
| 3907 | cpumask_setall(top_cpuset.cpus_allowed);
|
| 3908 | nodes_setall(top_cpuset.mems_allowed);
|
| 3909 | cpumask_setall(top_cpuset.effective_cpus);
|
| 3910 | cpumask_setall(top_cpuset.effective_xcpus);
|
| 3911 | cpumask_setall(top_cpuset.exclusive_cpus);
|
| 3912 | nodes_setall(top_cpuset.effective_mems);
|
| 3913 | init_llist_node(&top_cpuset.attach_node);
|
| 3914 |
|
| 3915 | cpuset1_init(&top_cpuset);
|
| 3916 |
|
| 3917 | BUG_ON(!alloc_cpumask_var(&cpus_attach, GFP_KERNEL));
|
| 3918 |
|
| 3919 | if (housekeeping_enabled(HK_TYPE_DOMAIN_BOOT))
|
| 3920 | cpumask_andnot(isolated_cpus, cpu_possible_mask,
|
| 3921 | housekeeping_cpumask(HK_TYPE_DOMAIN_BOOT));
|
| 3922 |
|
| 3923 | return 0;
|
| 3924 | }
|
| 3925 |
|
| 3926 | static void
|
| 3927 | hotplug_update_tasks(struct cpuset *cs,
|
| 3928 | struct cpumask *new_cpus, nodemask_t *new_mems,
|
| 3929 | bool cpus_updated, bool mems_updated)
|
| 3930 | {
|
| 3931 |
|
| 3932 | if (cpumask_empty(new_cpus) && !is_partition_valid(cs))
|
| 3933 | cpumask_copy(new_cpus, parent_cs(cs)->effective_cpus);
|
| 3934 | if (nodes_empty(*new_mems))
|
| 3935 | *new_mems = parent_cs(cs)->effective_mems;
|
| 3936 |
|
| 3937 | spin_lock_irq(&callback_lock);
|
| 3938 | cpumask_copy(cs->effective_cpus, new_cpus);
|
| 3939 | cs->effective_mems = *new_mems;
|
| 3940 | spin_unlock_irq(&callback_lock);
|
| 3941 |
|
| 3942 | if (cpus_updated)
|
| 3943 | cpuset_update_tasks_cpumask(cs, new_cpus);
|
| 3944 | if (mems_updated)
|
| 3945 | cpuset_update_tasks_nodemask(cs);
|
| 3946 | }
|
| 3947 |
|
| 3948 | void cpuset_force_rebuild(void)
|
| 3949 | {
|
| 3950 | force_sd_rebuild = true;
|
| 3951 | }
|
| 3952 |
|
| 3953 | |
| 3954 | * cpuset_hotplug_update_tasks - update tasks in a cpuset for hotunplug
|
| 3955 | * @cs: cpuset in interest
|
| 3956 | * @tmp: the tmpmasks structure pointer
|
| 3957 | *
|
| 3958 | * Compare @cs's cpu and mem masks against top_cpuset and if some have gone
|
| 3959 | * offline, update @cs accordingly. If @cs ends up with no CPU or memory,
|
| 3960 | * all its tasks are moved to the nearest ancestor with both resources.
|
| 3961 |
|
| 3962 | static void cpuset_hotplug_update_tasks(struct cpuset *cs, struct tmpmasks *tmp)
|
| 3963 | {
|
| 3964 | static cpumask_t new_cpus;
|
| 3965 | static nodemask_t new_mems;
|
| 3966 | bool cpus_updated;
|
| 3967 | bool mems_updated;
|
| 3968 | bool remote;
|
| 3969 | int partcmd = -1;
|
| 3970 | struct cpuset *parent;
|
| 3971 |
|
| 3972 | wait_attach_done_lock();
|
| 3973 | parent = parent_cs(cs);
|
| 3974 | compute_effective_cpumask(&new_cpus, cs, parent);
|
| 3975 | compute_effective_nodemask(&new_mems, cs, parent);
|
| 3976 |
|
| 3977 | if (!tmp || !cs->partition_root_state)
|
| 3978 | goto update_tasks;
|
| 3979 |
|
| 3980 | |
| 3981 | * Compute effective_cpus for valid partition root, may invalidate
|
| 3982 | * child partition roots if necessary.
|
| 3983 |
|
| 3984 | remote = is_remote_partition(cs);
|
| 3985 | if (remote || (is_partition_valid(cs) && is_partition_valid(parent)))
|
| 3986 | compute_partition_effective_cpumask(cs, &new_cpus);
|
| 3987 |
|
| 3988 | if (remote && (cpumask_empty(subpartitions_cpus) ||
|
| 3989 | (cpumask_empty(&new_cpus) &&
|
| 3990 | partition_is_populated(cs, NULL)))) {
|
| 3991 | WRITE_ONCE(cs->prs_err, PERR_HOTPLUG);
|
| 3992 | remote_partition_disable(cs, tmp);
|
| 3993 | compute_effective_cpumask(&new_cpus, cs, parent);
|
| 3994 | remote = false;
|
| 3995 | }
|
| 3996 |
|
| 3997 | |
| 3998 | * Force the partition to become invalid if either one of
|
| 3999 | * the following conditions hold:
|
| 4000 | * 1) empty effective cpus but not valid empty partition.
|
| 4001 | * 2) parent is invalid or doesn't grant any cpus to child
|
| 4002 | * partitions.
|
| 4003 | * 3) subpartitions_cpus is empty.
|
| 4004 |
|
| 4005 | if (is_local_partition(cs) &&
|
| 4006 | (!is_partition_valid(parent) ||
|
| 4007 | tasks_nocpu_error(parent, cs, &new_cpus) ||
|
| 4008 | cpumask_empty(subpartitions_cpus)))
|
| 4009 | partcmd = partcmd_invalidate;
|
| 4010 | |
| 4011 | * On the other hand, an invalid partition root may be transitioned
|
| 4012 | * back to a regular one with a non-empty effective xcpus.
|
| 4013 |
|
| 4014 | else if (is_partition_valid(parent) && is_partition_invalid(cs) &&
|
| 4015 | !cpumask_empty(cs->effective_xcpus))
|
| 4016 | partcmd = partcmd_update;
|
| 4017 |
|
| 4018 | if (partcmd >= 0) {
|
| 4019 | update_parent_effective_cpumask(cs, partcmd, NULL, tmp);
|
| 4020 | if ((partcmd == partcmd_invalidate) || is_partition_valid(cs)) {
|
| 4021 | compute_partition_effective_cpumask(cs, &new_cpus);
|
| 4022 | cpuset_force_rebuild();
|
| 4023 | }
|
| 4024 | }
|
| 4025 |
|
| 4026 | update_tasks:
|
| 4027 | cpus_updated = !cpumask_equal(&new_cpus, cs->effective_cpus);
|
| 4028 | mems_updated = !nodes_equal(new_mems, cs->effective_mems);
|
| 4029 | if (!cpus_updated && !mems_updated)
|
| 4030 | goto unlock;
|
| 4031 |
|
| 4032 | if (mems_updated)
|
| 4033 | check_insane_mems_config(&new_mems);
|
| 4034 |
|
| 4035 | if (is_in_v2_mode())
|
| 4036 | hotplug_update_tasks(cs, &new_cpus, &new_mems,
|
| 4037 | cpus_updated, mems_updated);
|
| 4038 | else
|
| 4039 | cpuset1_hotplug_update_tasks(cs, &new_cpus, &new_mems,
|
| 4040 | cpus_updated, mems_updated);
|
| 4041 |
|
| 4042 | unlock:
|
| 4043 | mutex_unlock(&cpuset_mutex);
|
| 4044 | }
|
| 4045 |
|
| 4046 | |
| 4047 | * cpuset_handle_hotplug - handle CPU/memory hot{,un}plug for a cpuset
|
| 4048 | *
|
| 4049 | * This function is called after either CPU or memory configuration has
|
| 4050 | * changed and updates cpuset accordingly. The top_cpuset is always
|
| 4051 | * synchronized to cpu_active_mask and N_MEMORY, which is necessary in
|
| 4052 | * order to make cpusets transparent (of no affect) on systems that are
|
| 4053 | * actively using CPU hotplug but making no active use of cpusets.
|
| 4054 | *
|
| 4055 | * Non-root cpusets are only affected by offlining. If any CPUs or memory
|
| 4056 | * nodes have been taken down, cpuset_hotplug_update_tasks() is invoked on
|
| 4057 | * all descendants.
|
| 4058 | *
|
| 4059 | * Note that CPU offlining during suspend is ignored. We don't modify
|
| 4060 | * cpusets across suspend/resume cycles at all.
|
| 4061 | *
|
| 4062 | * CPU / memory hotplug is handled synchronously.
|
| 4063 |
|
| 4064 | static void cpuset_handle_hotplug(void)
|
| 4065 | {
|
| 4066 | static DECLARE_WORK(hk_sd_work, hk_sd_workfn);
|
| 4067 | static cpumask_t new_cpus;
|
| 4068 | static nodemask_t new_mems;
|
| 4069 | bool cpus_updated, mems_updated;
|
| 4070 | bool on_dfl;
|
| 4071 | struct tmpmasks tmp, *ptmp = NULL;
|
| 4072 |
|
| 4073 | lockdep_assert_cpus_held();
|
| 4074 | mutex_lock(&cpuset_mutex);
|
| 4075 |
|
| 4076 | on_dfl = is_in_v2_mode();
|
| 4077 | if (on_dfl && !alloc_tmpmasks(&tmp))
|
| 4078 | ptmp = &tmp;
|
| 4079 |
|
| 4080 |
|
| 4081 | cpumask_copy(&new_cpus, cpu_active_mask);
|
| 4082 | new_mems = node_states[N_MEMORY];
|
| 4083 |
|
| 4084 | |
| 4085 | * If subpartitions_cpus is populated, it is likely that the check
|
| 4086 | * below will produce a false positive on cpus_updated when the cpu
|
| 4087 | * list isn't changed. It is extra work, but it is better to be safe.
|
| 4088 |
|
| 4089 | cpus_updated = !cpumask_equal(top_cpuset.effective_cpus, &new_cpus) ||
|
| 4090 | !cpumask_empty(subpartitions_cpus);
|
| 4091 | mems_updated = !nodes_equal(top_cpuset.effective_mems, new_mems);
|
| 4092 |
|
| 4093 |
|
| 4094 | if (cpus_updated) {
|
| 4095 | cpuset_force_rebuild();
|
| 4096 | spin_lock_irq(&callback_lock);
|
| 4097 | if (!on_dfl)
|
| 4098 | cpumask_copy(top_cpuset.cpus_allowed, &new_cpus);
|
| 4099 | |
| 4100 | * Make sure that CPUs allocated to child partitions
|
| 4101 | * do not show up in effective_cpus. If no CPU is left,
|
| 4102 | * we clear the subpartitions_cpus & let the child partitions
|
| 4103 | * fight for the CPUs again.
|
| 4104 |
|
| 4105 | if (!cpumask_empty(subpartitions_cpus)) {
|
| 4106 | if (cpumask_subset(&new_cpus, subpartitions_cpus)) {
|
| 4107 | cpumask_clear(subpartitions_cpus);
|
| 4108 | } else {
|
| 4109 | cpumask_andnot(&new_cpus, &new_cpus,
|
| 4110 | subpartitions_cpus);
|
| 4111 | }
|
| 4112 | }
|
| 4113 | cpumask_copy(top_cpuset.effective_cpus, &new_cpus);
|
| 4114 | spin_unlock_irq(&callback_lock);
|
| 4115 |
|
| 4116 | }
|
| 4117 |
|
| 4118 |
|
| 4119 | if (mems_updated) {
|
| 4120 | spin_lock_irq(&callback_lock);
|
| 4121 | if (!on_dfl)
|
| 4122 | top_cpuset.mems_allowed = new_mems;
|
| 4123 | top_cpuset.effective_mems = new_mems;
|
| 4124 | spin_unlock_irq(&callback_lock);
|
| 4125 | cpuset_update_tasks_nodemask(&top_cpuset);
|
| 4126 | }
|
| 4127 |
|
| 4128 | mutex_unlock(&cpuset_mutex);
|
| 4129 |
|
| 4130 |
|
| 4131 | if (cpus_updated || mems_updated) {
|
| 4132 | struct cpuset *cs;
|
| 4133 | struct cgroup_subsys_state *pos_css;
|
| 4134 |
|
| 4135 | rcu_read_lock();
|
| 4136 | cpuset_for_each_descendant_pre(cs, pos_css, &top_cpuset) {
|
| 4137 | if (cs == &top_cpuset || !css_tryget_online(&cs->css))
|
| 4138 | continue;
|
| 4139 | rcu_read_unlock();
|
| 4140 |
|
| 4141 | cpuset_hotplug_update_tasks(cs, ptmp);
|
| 4142 |
|
| 4143 | rcu_read_lock();
|
| 4144 | css_put(&cs->css);
|
| 4145 | }
|
| 4146 | rcu_read_unlock();
|
| 4147 | }
|
| 4148 |
|
| 4149 | |
| 4150 | * rebuild_sched_domains() will always be called directly if needed
|
| 4151 | * to make sure that newly added or removed CPU will be reflected in
|
| 4152 | * the sched domains. However, if isolated partition invalidation
|
| 4153 | * or recreation is being done (update_housekeeping set), a work item
|
| 4154 | * will be queued to call housekeeping_update() to update the
|
| 4155 | * corresponding housekeeping cpumasks after some slight delay.
|
| 4156 | *
|
| 4157 | * We rely on WORK_STRUCT_PENDING_BIT to not requeue a work item that
|
| 4158 | * is still pending. Before the pending bit is cleared, the work data
|
| 4159 | * is copied out and work item dequeued. So it is possible to queue
|
| 4160 | * the work again before the hk_sd_workfn() is invoked to process the
|
| 4161 | * previously queued work. Since hk_sd_workfn() doesn't use the work
|
| 4162 | * item at all, this is not a problem.
|
| 4163 |
|
| 4164 | if (force_sd_rebuild)
|
| 4165 | rebuild_sched_domains_cpuslocked();
|
| 4166 | if (update_housekeeping)
|
| 4167 | queue_work(system_dfl_wq, &hk_sd_work);
|
| 4168 |
|
| 4169 | free_tmpmasks(ptmp);
|
| 4170 | }
|
| 4171 |
|
| 4172 | void cpuset_update_active_cpus(void)
|
| 4173 | {
|
| 4174 | |
| 4175 | * We're inside cpu hotplug critical region which usually nests
|
| 4176 | * inside cgroup synchronization. Bounce actual hotplug processing
|
| 4177 | * to a work item to avoid reverse locking order.
|
| 4178 |
|
| 4179 | cpuset_handle_hotplug();
|
| 4180 | }
|
| 4181 |
|
| 4182 | |
| 4183 | * Keep top_cpuset.mems_allowed tracking node_states[N_MEMORY].
|
| 4184 | * Call this routine anytime after node_states[N_MEMORY] changes.
|
| 4185 | * See cpuset_update_active_cpus() for CPU hotplug handling.
|
| 4186 |
|
| 4187 | static int cpuset_track_online_nodes(struct notifier_block *self,
|
| 4188 | unsigned long action, void *arg)
|
| 4189 | {
|
| 4190 | cpuset_handle_hotplug();
|
| 4191 | return NOTIFY_OK;
|
| 4192 | }
|
| 4193 |
|
| 4194 | |
| 4195 | * cpuset_init_smp - initialize cpus_allowed
|
| 4196 | *
|
| 4197 | * Description: Finish top cpuset after cpu, node maps are initialized
|
| 4198 |
|
| 4199 | void __init cpuset_init_smp(void)
|
| 4200 | {
|
| 4201 | |
| 4202 | * cpus_allowd/mems_allowed set to v2 values in the initial
|
| 4203 | * cpuset_bind() call will be reset to v1 values in another
|
| 4204 | * cpuset_bind() call when v1 cpuset is mounted.
|
| 4205 |
|
| 4206 | top_cpuset.old_mems_allowed = top_cpuset.mems_allowed;
|
| 4207 |
|
| 4208 | cpumask_copy(top_cpuset.effective_cpus, cpu_active_mask);
|
| 4209 | top_cpuset.effective_mems = node_states[N_MEMORY];
|
| 4210 |
|
| 4211 | hotplug_node_notifier(cpuset_track_online_nodes, CPUSET_CALLBACK_PRI);
|
| 4212 |
|
| 4213 | cpuset_migrate_mm_wq = alloc_ordered_workqueue("cpuset_migrate_mm", 0);
|
| 4214 | BUG_ON(!cpuset_migrate_mm_wq);
|
| 4215 | }
|
| 4216 |
|
| 4217 | |
| 4218 | * Return cpus_allowed mask from a task's cpuset.
|
| 4219 |
|
| 4220 | static void __cpuset_cpus_allowed_locked(struct task_struct *tsk, struct cpumask *pmask)
|
| 4221 | {
|
| 4222 | struct cpuset *cs;
|
| 4223 |
|
| 4224 | cs = task_cs(tsk);
|
| 4225 | if (cs != &top_cpuset)
|
| 4226 | guarantee_active_cpus(tsk, pmask);
|
| 4227 | |
| 4228 | * Tasks in the top cpuset won't get update to their cpumasks
|
| 4229 | * when a hotplug online/offline event happens. So we include all
|
| 4230 | * offline cpus in the allowed cpu list.
|
| 4231 |
|
| 4232 | if ((cs == &top_cpuset) || cpumask_empty(pmask)) {
|
| 4233 | const struct cpumask *possible_mask = task_cpu_possible_mask(tsk);
|
| 4234 |
|
| 4235 | |
| 4236 | * We first exclude cpus allocated to partitions. If there is no
|
| 4237 | * allowable online cpu left, we fall back to all possible cpus.
|
| 4238 |
|
| 4239 | cpumask_andnot(pmask, possible_mask, subpartitions_cpus);
|
| 4240 | if (!cpumask_intersects(pmask, cpu_active_mask))
|
| 4241 | cpumask_copy(pmask, possible_mask);
|
| 4242 | }
|
| 4243 | }
|
| 4244 |
|
| 4245 | |
| 4246 | * cpuset_cpus_allowed_locked - return cpus_allowed mask from a task's cpuset.
|
| 4247 | * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
|
| 4248 | * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
|
| 4249 | *
|
| 4250 | * Similir to cpuset_cpus_allowed() except that the caller must have acquired
|
| 4251 | * cpuset_mutex.
|
| 4252 |
|
| 4253 | void cpuset_cpus_allowed_locked(struct task_struct *tsk, struct cpumask *pmask)
|
| 4254 | {
|
| 4255 | lockdep_assert_cpuset_lock_held();
|
| 4256 | __cpuset_cpus_allowed_locked(tsk, pmask);
|
| 4257 | }
|
| 4258 |
|
| 4259 | |
| 4260 | * cpuset_cpus_allowed - return cpus_allowed mask from a task's cpuset.
|
| 4261 | * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
|
| 4262 | * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
|
| 4263 | *
|
| 4264 | * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
|
| 4265 | * attached to the specified @tsk. Guaranteed to return some non-empty
|
| 4266 | * subset of cpu_active_mask, even if this means going outside the
|
| 4267 | * tasks cpuset, except when the task is in the top cpuset.
|
| 4268 | *
|
| 4269 |
|
| 4270 | void cpuset_cpus_allowed(struct task_struct *tsk, struct cpumask *pmask)
|
| 4271 | {
|
| 4272 | unsigned long flags;
|
| 4273 |
|
| 4274 | spin_lock_irqsave(&callback_lock, flags);
|
| 4275 | __cpuset_cpus_allowed_locked(tsk, pmask);
|
| 4276 | spin_unlock_irqrestore(&callback_lock, flags);
|
| 4277 | }
|
| 4278 |
|
| 4279 | |
| 4280 | * cpuset_cpus_allowed_fallback - final fallback before complete catastrophe.
|
| 4281 | * @tsk: pointer to task_struct with which the scheduler is struggling
|
| 4282 | *
|
| 4283 | * Description: In the case that the scheduler cannot find an allowed cpu in
|
| 4284 | * tsk->cpus_allowed, we fall back to task_cs(tsk)->cpus_allowed. In legacy
|
| 4285 | * mode however, this value is the same as task_cs(tsk)->effective_cpus,
|
| 4286 | * which will not contain a sane cpumask during cases such as cpu hotplugging.
|
| 4287 | * This is the absolute last resort for the scheduler and it is only used if
|
| 4288 | * _every_ other avenue has been traveled.
|
| 4289 | *
|
| 4290 | * Returns true if the affinity of @tsk was changed, false otherwise.
|
| 4291 | *
|
| 4292 |
|
| 4293 | bool cpuset_cpus_allowed_fallback(struct task_struct *tsk)
|
| 4294 | {
|
| 4295 | const struct cpumask *possible_mask = task_cpu_possible_mask(tsk);
|
| 4296 | const struct cpumask *cs_mask;
|
| 4297 | bool changed = false;
|
| 4298 |
|
| 4299 | rcu_read_lock();
|
| 4300 | cs_mask = task_cs(tsk)->cpus_allowed;
|
| 4301 | if (is_in_v2_mode() && cpumask_subset(cs_mask, possible_mask)) {
|
| 4302 | set_cpus_allowed_force(tsk, cs_mask);
|
| 4303 | changed = true;
|
| 4304 | }
|
| 4305 | rcu_read_unlock();
|
| 4306 |
|
| 4307 | |
| 4308 | * We own tsk->cpus_allowed, nobody can change it under us.
|
| 4309 | *
|
| 4310 | * But we used cs && cs->cpus_allowed lockless and thus can
|
| 4311 | * race with cgroup_attach_task() or update_cpumask() and get
|
| 4312 | * the wrong tsk->cpus_allowed. However, both cases imply the
|
| 4313 | * subsequent cpuset_change_cpumask()->set_cpus_allowed_ptr()
|
| 4314 | * which takes task_rq_lock().
|
| 4315 | *
|
| 4316 | * If we are called after it dropped the lock we must see all
|
| 4317 | * changes in tsk_cs()->cpus_allowed. Otherwise we can temporary
|
| 4318 | * set any mask even if it is not right from task_cs() pov,
|
| 4319 | * the pending set_cpus_allowed_ptr() will fix things.
|
| 4320 | *
|
| 4321 | * select_fallback_rq() will fix things ups and set cpu_possible_mask
|
| 4322 | * if required.
|
| 4323 |
|
| 4324 | return changed;
|
| 4325 | }
|
| 4326 |
|
| 4327 | |
| 4328 | * Returns the number of CPUs available for this cgroup.
|
| 4329 | *
|
| 4330 | * This only really works for cgroup-v2 where all the controllers are mounted
|
| 4331 | * in the same hierarchy. If not cgroup-v2 or no cpuset controller is
|
| 4332 | * configured it reverts to num_online_cpus().
|
| 4333 |
|
| 4334 | int cpuset_num_cpus(struct cgroup *cgrp)
|
| 4335 | {
|
| 4336 | int nr = num_online_cpus();
|
| 4337 | struct cpuset *cs;
|
| 4338 |
|
| 4339 | if (is_in_v2_mode()) {
|
| 4340 | guard(rcu)();
|
| 4341 | cs = css_cs(cgroup_e_css(cgrp, &cpuset_cgrp_subsys));
|
| 4342 | if (cs)
|
| 4343 | nr = cpumask_weight(cs->effective_cpus);
|
| 4344 | }
|
| 4345 |
|
| 4346 | return nr;
|
| 4347 | }
|
| 4348 |
|
| 4349 | void __init cpuset_init_current_mems_allowed(void)
|
| 4350 | {
|
| 4351 | nodes_setall(current->mems_allowed);
|
| 4352 | }
|
| 4353 |
|
| 4354 | |
| 4355 | * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
|
| 4356 | * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
|
| 4357 | *
|
| 4358 | * Description: Returns the nodemask_t mems_allowed of the cpuset
|
| 4359 | * attached to the specified @tsk. Guaranteed to return some non-empty
|
| 4360 | * subset of node_states[N_MEMORY], even if this means going outside the
|
| 4361 | * tasks cpuset.
|
| 4362 | *
|
| 4363 |
|
| 4364 | nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
|
| 4365 | {
|
| 4366 | nodemask_t mask;
|
| 4367 | unsigned long flags;
|
| 4368 |
|
| 4369 | spin_lock_irqsave(&callback_lock, flags);
|
| 4370 | guarantee_online_mems(task_cs(tsk), &mask);
|
| 4371 | spin_unlock_irqrestore(&callback_lock, flags);
|
| 4372 |
|
| 4373 | return mask;
|
| 4374 | }
|
| 4375 |
|
| 4376 | |
| 4377 | * cpuset_nodemask_valid_mems_allowed - check nodemask vs. current mems_allowed
|
| 4378 | * @nodemask: the nodemask to be checked
|
| 4379 | *
|
| 4380 | * Are any of the nodes in the nodemask allowed in current->mems_allowed?
|
| 4381 |
|
| 4382 | int cpuset_nodemask_valid_mems_allowed(const nodemask_t *nodemask)
|
| 4383 | {
|
| 4384 | return nodes_intersects(*nodemask, current->mems_allowed);
|
| 4385 | }
|
| 4386 |
|
| 4387 | |
| 4388 | * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
|
| 4389 | * mem_hardwall ancestor to the specified cpuset. Call holding
|
| 4390 | * callback_lock. If no ancestor is mem_exclusive or mem_hardwall
|
| 4391 | * (an unusual configuration), then returns the root cpuset.
|
| 4392 |
|
| 4393 | static struct cpuset *nearest_hardwall_ancestor(struct cpuset *cs)
|
| 4394 | {
|
| 4395 | while (!(is_mem_exclusive(cs) || is_mem_hardwall(cs)) && parent_cs(cs))
|
| 4396 | cs = parent_cs(cs);
|
| 4397 | return cs;
|
| 4398 | }
|
| 4399 |
|
| 4400 | |
| 4401 | * cpuset_current_node_allowed - Can current task allocate on a memory node?
|
| 4402 | * @node: is this an allowed node?
|
| 4403 | * @gfp_mask: memory allocation flags
|
| 4404 | *
|
| 4405 | * If we're in interrupt, yes, we can always allocate. If @node is set in
|
| 4406 | * current's mems_allowed, yes. If it's not a __GFP_HARDWALL request and this
|
| 4407 | * node is set in the nearest hardwalled cpuset ancestor to current's cpuset,
|
| 4408 | * yes. If current has access to memory reserves as an oom victim, yes.
|
| 4409 | * If the current task is PF_EXITING, yes. Otherwise, no.
|
| 4410 | *
|
| 4411 | * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
|
| 4412 | * and do not allow allocations outside the current tasks cpuset
|
| 4413 | * unless the task has been OOM killed or is exiting.
|
| 4414 | * GFP_KERNEL allocations are not so marked, so can escape to the
|
| 4415 | * nearest enclosing hardwalled ancestor cpuset.
|
| 4416 | *
|
| 4417 | * Scanning up parent cpusets requires callback_lock. The
|
| 4418 | * page allocator only calls here with __GFP_HARDWALL bit
|
| 4419 | * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
|
| 4420 | * current tasks mems_allowed came up empty on the first pass over
|
| 4421 | * the zonelist. So only GFP_KERNEL allocations, if all nodes in the
|
| 4422 | * cpuset are short of memory, might require taking the callback_lock.
|
| 4423 | *
|
| 4424 | * The first call here from mm/page_alloc:get_page_from_freelist()
|
| 4425 | * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
|
| 4426 | * so no allocation on a node outside the cpuset is allowed (unless
|
| 4427 | * in interrupt, of course). The PF_EXITING check must therefore
|
| 4428 | * come before the __GFP_HARDWALL check, otherwise a dying task
|
| 4429 | * would be blocked on the fast path.
|
| 4430 | *
|
| 4431 | * The second pass through get_page_from_freelist() doesn't even call here for
|
| 4432 | * GFP_ATOMIC calls. That, and the checks below have the combined affect that:
|
| 4433 | * in_interrupt - any node ok (current task context irrelevant)
|
| 4434 | * GFP_ATOMIC - any node ok
|
| 4435 | * tsk_is_oom_victim - any node ok
|
| 4436 | * PF_EXITING - any node ok (let dying task exit quickly)
|
| 4437 | * GFP_KERNEL - any node in enclosing hardwalled cpuset ok
|
| 4438 | * GFP_USER - only nodes in current tasks mems allowed ok.
|
| 4439 |
|
| 4440 | bool cpuset_current_node_allowed(int node, gfp_t gfp_mask)
|
| 4441 | {
|
| 4442 | struct cpuset *cs;
|
| 4443 | bool allowed;
|
| 4444 | unsigned long flags;
|
| 4445 |
|
| 4446 | if (in_interrupt())
|
| 4447 | return true;
|
| 4448 | if (node_isset(node, current->mems_allowed))
|
| 4449 | return true;
|
| 4450 | |
| 4451 | * Allow tasks that have access to memory reserves because they have
|
| 4452 | * been OOM killed to get memory anywhere.
|
| 4453 |
|
| 4454 | if (unlikely(tsk_is_oom_victim(current)))
|
| 4455 | return true;
|
| 4456 | if (current->flags & PF_EXITING)
|
| 4457 | return true;
|
| 4458 | if (gfp_mask & __GFP_HARDWALL)
|
| 4459 | return false;
|
| 4460 |
|
| 4461 | if (cpuset_v2())
|
| 4462 | return true;
|
| 4463 |
|
| 4464 |
|
| 4465 | spin_lock_irqsave(&callback_lock, flags);
|
| 4466 |
|
| 4467 | cs = nearest_hardwall_ancestor(task_cs(current));
|
| 4468 | allowed = node_isset(node, cs->mems_allowed);
|
| 4469 |
|
| 4470 | spin_unlock_irqrestore(&callback_lock, flags);
|
| 4471 | return allowed;
|
| 4472 | }
|
| 4473 |
|
| 4474 | |
| 4475 | * cpuset_nodes_allowed - return effective_mems mask from a cgroup cpuset.
|
| 4476 | * @cgroup: pointer to struct cgroup.
|
| 4477 | * @mask: pointer to struct nodemask_t to be returned.
|
| 4478 | *
|
| 4479 | * Returns effective_mems mask from a cgroup cpuset if it is cgroup v2 and
|
| 4480 | * has cpuset subsys. Otherwise, returns node_states[N_MEMORY].
|
| 4481 | *
|
| 4482 | * This function intentionally avoids taking the cpuset_mutex or callback_lock
|
| 4483 | * when accessing effective_mems. This is because the obtained effective_mems
|
| 4484 | * is stale immediately after the query anyway (e.g., effective_mems is updated
|
| 4485 | * immediately after releasing the lock but before returning).
|
| 4486 | *
|
| 4487 | * As a result, returned @mask may be empty because cs->effective_mems can be
|
| 4488 | * rebound during this call. Besides, nodes in @mask are not guaranteed to be
|
| 4489 | * online due to hot plugins. Callers should check the mask for validity on
|
| 4490 | * return based on its subsequent use.
|
| 4491 | *
|
| 4492 | void cpuset_nodes_allowed(struct cgroup *cgroup, nodemask_t *mask)
|
| 4493 | {
|
| 4494 | struct cgroup_subsys_state *css;
|
| 4495 | struct cpuset *cs;
|
| 4496 |
|
| 4497 | |
| 4498 | * In v1, mem_cgroup and cpuset are unlikely in the same hierarchy
|
| 4499 | * and mems_allowed is likely to be empty even if we could get to it,
|
| 4500 | * so return directly to avoid taking a global lock on the empty check.
|
| 4501 |
|
| 4502 | if (!cgroup || !cpuset_v2()) {
|
| 4503 | nodes_copy(*mask, node_states[N_MEMORY]);
|
| 4504 | return;
|
| 4505 | }
|
| 4506 |
|
| 4507 | css = cgroup_get_e_css(cgroup, &cpuset_cgrp_subsys);
|
| 4508 | if (!css) {
|
| 4509 | nodes_copy(*mask, node_states[N_MEMORY]);
|
| 4510 | return;
|
| 4511 | }
|
| 4512 |
|
| 4513 | |
| 4514 | * The reference taken via cgroup_get_e_css is sufficient to
|
| 4515 | * protect css, but it does not imply safe accesses to effective_mems.
|
| 4516 | *
|
| 4517 | * Normally, accessing effective_mems would require the cpuset_mutex
|
| 4518 | * or callback_lock - but the correctness of this information is stale
|
| 4519 | * immediately after the query anyway. We do not acquire the lock
|
| 4520 | * during this process to save lock contention in exchange for racing
|
| 4521 | * against mems_allowed rebinds.
|
| 4522 |
|
| 4523 | cs = container_of(css, struct cpuset, css);
|
| 4524 | nodes_copy(*mask, cs->effective_mems);
|
| 4525 | css_put(css);
|
| 4526 | }
|
| 4527 |
|
| 4528 | |
| 4529 | * cpuset_spread_node() - On which node to begin search for a page
|
| 4530 | * @rotor: round robin rotor
|
| 4531 | *
|
| 4532 | * If a task is marked PFA_SPREAD_PAGE and a page cache allocation uses
|
| 4533 | * cpuset_mem_spread_node() to determine where to start looking, spread the
|
| 4534 | * starting node around the task's mems_allowed nodes instead of starting on
|
| 4535 | * the local node.
|
| 4536 | *
|
| 4537 | * We don't have to worry about the returned node being offline
|
| 4538 | * because "it can't happen", and even if it did, it would be ok.
|
| 4539 | *
|
| 4540 | * The routines calling guarantee_online_mems() are careful to
|
| 4541 | * only set nodes in task->mems_allowed that are online. So it
|
| 4542 | * should not be possible for the following code to return an
|
| 4543 | * offline node. But if it did, that would be ok, as this routine
|
| 4544 | * is not returning the node where the allocation must be, only
|
| 4545 | * the node where the search should start. The zonelist used by
|
| 4546 | * the allocator will include all nodes. If the slab allocator
|
| 4547 | * is passed an offline node, it will fall back to the local node.
|
| 4548 | * See kmem_cache_alloc_node().
|
| 4549 |
|
| 4550 | static int cpuset_spread_node(int *rotor)
|
| 4551 | {
|
| 4552 | return *rotor = next_node_in(*rotor, current->mems_allowed);
|
| 4553 | }
|
| 4554 |
|
| 4555 | |
| 4556 | * cpuset_mem_spread_node() - On which node to begin search for a file page
|
| 4557 |
|
| 4558 | int cpuset_mem_spread_node(void)
|
| 4559 | {
|
| 4560 | if (current->cpuset_mem_spread_rotor == NUMA_NO_NODE)
|
| 4561 | current->cpuset_mem_spread_rotor =
|
| 4562 | node_random(¤t->mems_allowed);
|
| 4563 |
|
| 4564 | return cpuset_spread_node(¤t->cpuset_mem_spread_rotor);
|
| 4565 | }
|
| 4566 |
|
| 4567 | |
| 4568 | * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
|
| 4569 | * @tsk1: pointer to task_struct of some task.
|
| 4570 | * @tsk2: pointer to task_struct of some other task.
|
| 4571 | *
|
| 4572 | * Description: Return true if @tsk1's mems_allowed intersects the
|
| 4573 | * mems_allowed of @tsk2. Used by the OOM killer to determine if
|
| 4574 | * one of the task's memory usage might impact the memory available
|
| 4575 | * to the other.
|
| 4576 | *
|
| 4577 |
|
| 4578 | int cpuset_mems_allowed_intersects(const struct task_struct *tsk1,
|
| 4579 | const struct task_struct *tsk2)
|
| 4580 | {
|
| 4581 | return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed);
|
| 4582 | }
|
| 4583 |
|
| 4584 | |
| 4585 | * cpuset_print_current_mems_allowed - prints current's cpuset and mems_allowed
|
| 4586 | *
|
| 4587 | * Description: Prints current's name, cpuset name, and cached copy of its
|
| 4588 | * mems_allowed to the kernel log.
|
| 4589 |
|
| 4590 | void cpuset_print_current_mems_allowed(void)
|
| 4591 | {
|
| 4592 | struct cgroup *cgrp;
|
| 4593 |
|
| 4594 | rcu_read_lock();
|
| 4595 |
|
| 4596 | cgrp = task_cs(current)->css.cgroup;
|
| 4597 | pr_cont(",cpuset=");
|
| 4598 | pr_cont_cgroup_name(cgrp);
|
| 4599 | pr_cont(",mems_allowed=%*pbl",
|
| 4600 | nodemask_pr_args(¤t->mems_allowed));
|
| 4601 |
|
| 4602 | rcu_read_unlock();
|
| 4603 | }
|
| 4604 |
|
| 4605 |
|
| 4606 | void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task)
|
| 4607 | {
|
| 4608 | seq_printf(m, "Mems_allowed:\t%*pb\n",
|
| 4609 | nodemask_pr_args(&task->mems_allowed));
|
| 4610 | seq_printf(m, "Mems_allowed_list:\t%*pbl\n",
|
| 4611 | nodemask_pr_args(&task->mems_allowed));
|
| 4612 | }
|