| 1 | |
| 2 | * kernel/cpuset.c
|
| 3 | *
|
| 4 | * Processor and Memory placement constraints for sets of tasks.
|
| 5 | *
|
| 6 | * Copyright (C) 2003 BULL SA.
|
| 7 | * Copyright (C) 2004-2007 Silicon Graphics, Inc.
|
| 8 | * Copyright (C) 2006 Google, Inc
|
| 9 | *
|
| 10 | * Portions derived from Patrick Mochel's sysfs code.
|
| 11 | * sysfs is Copyright (c) 2001-3 Patrick Mochel
|
| 12 | *
|
| 13 | * 2003-10-10 Written by Simon Derr.
|
| 14 | * 2003-10-22 Updates by Stephen Hemminger.
|
| 15 | * 2004 May-July Rework by Paul Jackson.
|
| 16 | * 2006 Rework by Paul Menage to use generic cgroups
|
| 17 | * 2008 Rework of the scheduler domains and CPU hotplug handling
|
| 18 | * by Max Krasnyansky
|
| 19 | *
|
| 20 | * This file is subject to the terms and conditions of the GNU General Public
|
| 21 | * License. See the file COPYING in the main directory of the Linux
|
| 22 | * distribution for more details.
|
| 23 |
|
| 24 |
|
| 25 | #include <linux/cpu.h>
|
| 26 | #include <linux/cpumask.h>
|
| 27 | #include <linux/cpuset.h>
|
| 28 | #include <linux/err.h>
|
| 29 | #include <linux/errno.h>
|
| 30 | #include <linux/file.h>
|
| 31 | #include <linux/fs.h>
|
| 32 | #include <linux/init.h>
|
| 33 | #include <linux/interrupt.h>
|
| 34 | #include <linux/kernel.h>
|
| 35 | #include <linux/kmod.h>
|
| 36 | #include <linux/list.h>
|
| 37 | #include <linux/mempolicy.h>
|
| 38 | #include <linux/mm.h>
|
| 39 | #include <linux/memory.h>
|
| 40 | #include <linux/export.h>
|
| 41 | #include <linux/mount.h>
|
| 42 | #include <linux/namei.h>
|
| 43 | #include <linux/pagemap.h>
|
| 44 | #include <linux/proc_fs.h>
|
| 45 | #include <linux/rcupdate.h>
|
| 46 | #include <linux/sched.h>
|
| 47 | #include <linux/seq_file.h>
|
| 48 | #include <linux/security.h>
|
| 49 | #include <linux/slab.h>
|
| 50 | #include <linux/spinlock.h>
|
| 51 | #include <linux/stat.h>
|
| 52 | #include <linux/string.h>
|
| 53 | #include <linux/time.h>
|
| 54 | #include <linux/backing-dev.h>
|
| 55 | #include <linux/sort.h>
|
| 56 |
|
| 57 | #include <asm/uaccess.h>
|
| 58 | #include <linux/atomic.h>
|
| 59 | #include <linux/mutex.h>
|
| 60 | #include <linux/workqueue.h>
|
| 61 | #include <linux/cgroup.h>
|
| 62 | #include <linux/wait.h>
|
| 63 |
|
| 64 | |
| 65 | * Tracks how many cpusets are currently defined in system.
|
| 66 | * When there is only one cpuset (the root cpuset) we can
|
| 67 | * short circuit some hooks.
|
| 68 |
|
| 69 | int number_of_cpusets __read_mostly;
|
| 70 |
|
| 71 |
|
| 72 |
|
| 73 | struct fmeter {
|
| 74 | int cnt;
|
| 75 | int val;
|
| 76 | time_t time;
|
| 77 | spinlock_t lock;
|
| 78 | };
|
| 79 |
|
| 80 | struct cpuset {
|
| 81 | struct cgroup_subsys_state css;
|
| 82 |
|
| 83 | unsigned long flags;
|
| 84 | cpumask_var_t cpus_allowed;
|
| 85 | nodemask_t mems_allowed;
|
| 86 |
|
| 87 | |
| 88 | * This is old Memory Nodes tasks took on.
|
| 89 | *
|
| 90 | * - top_cpuset.old_mems_allowed is initialized to mems_allowed.
|
| 91 | * - A new cpuset's old_mems_allowed is initialized when some
|
| 92 | * task is moved into it.
|
| 93 | * - old_mems_allowed is used in cpuset_migrate_mm() when we change
|
| 94 | * cpuset.mems_allowed and have tasks' nodemask updated, and
|
| 95 | * then old_mems_allowed is updated to mems_allowed.
|
| 96 |
|
| 97 | nodemask_t old_mems_allowed;
|
| 98 |
|
| 99 | struct fmeter fmeter;
|
| 100 |
|
| 101 | |
| 102 | * Tasks are being attached to this cpuset. Used to prevent
|
| 103 | * zeroing cpus/mems_allowed between ->can_attach() and ->attach().
|
| 104 |
|
| 105 | int attach_in_progress;
|
| 106 |
|
| 107 |
|
| 108 | int pn;
|
| 109 |
|
| 110 |
|
| 111 | int relax_domain_level;
|
| 112 | };
|
| 113 |
|
| 114 | static inline struct cpuset *css_cs(struct cgroup_subsys_state *css)
|
| 115 | {
|
| 116 | return css ? container_of(css, struct cpuset, css) : NULL;
|
| 117 | }
|
| 118 |
|
| 119 |
|
| 120 | static inline struct cpuset *task_cs(struct task_struct *task)
|
| 121 | {
|
| 122 | return css_cs(task_css(task, cpuset_subsys_id));
|
| 123 | }
|
| 124 |
|
| 125 | static inline struct cpuset *parent_cs(struct cpuset *cs)
|
| 126 | {
|
| 127 | return css_cs(css_parent(&cs->css));
|
| 128 | }
|
| 129 |
|
| 130 | #ifdef CONFIG_NUMA
|
| 131 | static inline bool task_has_mempolicy(struct task_struct *task)
|
| 132 | {
|
| 133 | return task->mempolicy;
|
| 134 | }
|
| 135 | #else
|
| 136 | static inline bool task_has_mempolicy(struct task_struct *task)
|
| 137 | {
|
| 138 | return false;
|
| 139 | }
|
| 140 | #endif
|
| 141 |
|
| 142 |
|
| 143 |
|
| 144 | typedef enum {
|
| 145 | CS_ONLINE,
|
| 146 | CS_CPU_EXCLUSIVE,
|
| 147 | CS_MEM_EXCLUSIVE,
|
| 148 | CS_MEM_HARDWALL,
|
| 149 | CS_MEMORY_MIGRATE,
|
| 150 | CS_SCHED_LOAD_BALANCE,
|
| 151 | CS_SPREAD_PAGE,
|
| 152 | CS_SPREAD_SLAB,
|
| 153 | } cpuset_flagbits_t;
|
| 154 |
|
| 155 |
|
| 156 | static inline bool is_cpuset_online(const struct cpuset *cs)
|
| 157 | {
|
| 158 | return test_bit(CS_ONLINE, &cs->flags);
|
| 159 | }
|
| 160 |
|
| 161 | static inline int is_cpu_exclusive(const struct cpuset *cs)
|
| 162 | {
|
| 163 | return test_bit(CS_CPU_EXCLUSIVE, &cs->flags);
|
| 164 | }
|
| 165 |
|
| 166 | static inline int is_mem_exclusive(const struct cpuset *cs)
|
| 167 | {
|
| 168 | return test_bit(CS_MEM_EXCLUSIVE, &cs->flags);
|
| 169 | }
|
| 170 |
|
| 171 | static inline int is_mem_hardwall(const struct cpuset *cs)
|
| 172 | {
|
| 173 | return test_bit(CS_MEM_HARDWALL, &cs->flags);
|
| 174 | }
|
| 175 |
|
| 176 | static inline int is_sched_load_balance(const struct cpuset *cs)
|
| 177 | {
|
| 178 | return test_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
|
| 179 | }
|
| 180 |
|
| 181 | static inline int is_memory_migrate(const struct cpuset *cs)
|
| 182 | {
|
| 183 | return test_bit(CS_MEMORY_MIGRATE, &cs->flags);
|
| 184 | }
|
| 185 |
|
| 186 | static inline int is_spread_page(const struct cpuset *cs)
|
| 187 | {
|
| 188 | return test_bit(CS_SPREAD_PAGE, &cs->flags);
|
| 189 | }
|
| 190 |
|
| 191 | static inline int is_spread_slab(const struct cpuset *cs)
|
| 192 | {
|
| 193 | return test_bit(CS_SPREAD_SLAB, &cs->flags);
|
| 194 | }
|
| 195 |
|
| 196 | static struct cpuset top_cpuset = {
|
| 197 | .flags = ((1 << CS_ONLINE) | (1 << CS_CPU_EXCLUSIVE) |
|
| 198 | (1 << CS_MEM_EXCLUSIVE)),
|
| 199 | };
|
| 200 |
|
| 201 | |
| 202 | * cpuset_for_each_child - traverse online children of a cpuset
|
| 203 | * @child_cs: loop cursor pointing to the current child
|
| 204 | * @pos_css: used for iteration
|
| 205 | * @parent_cs: target cpuset to walk children of
|
| 206 | *
|
| 207 | * Walk @child_cs through the online children of @parent_cs. Must be used
|
| 208 | * with RCU read locked.
|
| 209 |
|
| 210 | #define cpuset_for_each_child(child_cs, pos_css, parent_cs) \
|
| 211 | css_for_each_child((pos_css), &(parent_cs)->css) \
|
| 212 | if (is_cpuset_online(((child_cs) = css_cs((pos_css)))))
|
| 213 |
|
| 214 | |
| 215 | * cpuset_for_each_descendant_pre - pre-order walk of a cpuset's descendants
|
| 216 | * @des_cs: loop cursor pointing to the current descendant
|
| 217 | * @pos_css: used for iteration
|
| 218 | * @root_cs: target cpuset to walk ancestor of
|
| 219 | *
|
| 220 | * Walk @des_cs through the online descendants of @root_cs. Must be used
|
| 221 | * with RCU read locked. The caller may modify @pos_css by calling
|
| 222 | * css_rightmost_descendant() to skip subtree. @root_cs is included in the
|
| 223 | * iteration and the first node to be visited.
|
| 224 |
|
| 225 | #define cpuset_for_each_descendant_pre(des_cs, pos_css, root_cs) \
|
| 226 | css_for_each_descendant_pre((pos_css), &(root_cs)->css) \
|
| 227 | if (is_cpuset_online(((des_cs) = css_cs((pos_css)))))
|
| 228 |
|
| 229 | |
| 230 | * There are two global mutexes guarding cpuset structures - cpuset_mutex
|
| 231 | * and callback_mutex. The latter may nest inside the former. We also
|
| 232 | * require taking task_lock() when dereferencing a task's cpuset pointer.
|
| 233 | * See "The task_lock() exception", at the end of this comment.
|
| 234 | *
|
| 235 | * A task must hold both mutexes to modify cpusets. If a task holds
|
| 236 | * cpuset_mutex, then it blocks others wanting that mutex, ensuring that it
|
| 237 | * is the only task able to also acquire callback_mutex and be able to
|
| 238 | * modify cpusets. It can perform various checks on the cpuset structure
|
| 239 | * first, knowing nothing will change. It can also allocate memory while
|
| 240 | * just holding cpuset_mutex. While it is performing these checks, various
|
| 241 | * callback routines can briefly acquire callback_mutex to query cpusets.
|
| 242 | * Once it is ready to make the changes, it takes callback_mutex, blocking
|
| 243 | * everyone else.
|
| 244 | *
|
| 245 | * Calls to the kernel memory allocator can not be made while holding
|
| 246 | * callback_mutex, as that would risk double tripping on callback_mutex
|
| 247 | * from one of the callbacks into the cpuset code from within
|
| 248 | * __alloc_pages().
|
| 249 | *
|
| 250 | * If a task is only holding callback_mutex, then it has read-only
|
| 251 | * access to cpusets.
|
| 252 | *
|
| 253 | * Now, the task_struct fields mems_allowed and mempolicy may be changed
|
| 254 | * by other task, we use alloc_lock in the task_struct fields to protect
|
| 255 | * them.
|
| 256 | *
|
| 257 | * The cpuset_common_file_read() handlers only hold callback_mutex across
|
| 258 | * small pieces of code, such as when reading out possibly multi-word
|
| 259 | * cpumasks and nodemasks.
|
| 260 | *
|
| 261 | * Accessing a task's cpuset should be done in accordance with the
|
| 262 | * guidelines for accessing subsystem state in kernel/cgroup.c
|
| 263 |
|
| 264 |
|
| 265 | static DEFINE_MUTEX(cpuset_mutex);
|
| 266 | static DEFINE_MUTEX(callback_mutex);
|
| 267 |
|
| 268 | |
| 269 | * CPU / memory hotplug is handled asynchronously.
|
| 270 |
|
| 271 | static void cpuset_hotplug_workfn(struct work_struct *work);
|
| 272 | static DECLARE_WORK(cpuset_hotplug_work, cpuset_hotplug_workfn);
|
| 273 |
|
| 274 | static DECLARE_WAIT_QUEUE_HEAD(cpuset_attach_wq);
|
| 275 |
|
| 276 | |
| 277 | * This is ugly, but preserves the userspace API for existing cpuset
|
| 278 | * users. If someone tries to mount the "cpuset" filesystem, we
|
| 279 | * silently switch it to mount "cgroup" instead
|
| 280 |
|
| 281 | static struct dentry *cpuset_mount(struct file_system_type *fs_type,
|
| 282 | int flags, const char *unused_dev_name, void *data)
|
| 283 | {
|
| 284 | struct file_system_type *cgroup_fs = get_fs_type("cgroup");
|
| 285 | struct dentry *ret = ERR_PTR(-ENODEV);
|
| 286 | if (cgroup_fs) {
|
| 287 | char mountopts[] =
|
| 288 | "cpuset,noprefix,"
|
| 289 | "release_agent=/sbin/cpuset_release_agent";
|
| 290 | ret = cgroup_fs->mount(cgroup_fs, flags,
|
| 291 | unused_dev_name, mountopts);
|
| 292 | put_filesystem(cgroup_fs);
|
| 293 | }
|
| 294 | return ret;
|
| 295 | }
|
| 296 |
|
| 297 | static struct file_system_type cpuset_fs_type = {
|
| 298 | .name = "cpuset",
|
| 299 | .mount = cpuset_mount,
|
| 300 | };
|
| 301 |
|
| 302 | |
| 303 | * Return in pmask the portion of a cpusets's cpus_allowed that
|
| 304 | * are online. If none are online, walk up the cpuset hierarchy
|
| 305 | * until we find one that does have some online cpus. The top
|
| 306 | * cpuset always has some cpus online.
|
| 307 | *
|
| 308 | * One way or another, we guarantee to return some non-empty subset
|
| 309 | * of cpu_online_mask.
|
| 310 | *
|
| 311 | * Call with callback_mutex held.
|
| 312 |
|
| 313 | static void guarantee_online_cpus(struct cpuset *cs, struct cpumask *pmask)
|
| 314 | {
|
| 315 | while (!cpumask_intersects(cs->cpus_allowed, cpu_online_mask))
|
| 316 | cs = parent_cs(cs);
|
| 317 | cpumask_and(pmask, cs->cpus_allowed, cpu_online_mask);
|
| 318 | }
|
| 319 |
|
| 320 | |
| 321 | * Return in *pmask the portion of a cpusets's mems_allowed that
|
| 322 | * are online, with memory. If none are online with memory, walk
|
| 323 | * up the cpuset hierarchy until we find one that does have some
|
| 324 | * online mems. The top cpuset always has some mems online.
|
| 325 | *
|
| 326 | * One way or another, we guarantee to return some non-empty subset
|
| 327 | * of node_states[N_MEMORY].
|
| 328 | *
|
| 329 | * Call with callback_mutex held.
|
| 330 |
|
| 331 | static void guarantee_online_mems(struct cpuset *cs, nodemask_t *pmask)
|
| 332 | {
|
| 333 | while (!nodes_intersects(cs->mems_allowed, node_states[N_MEMORY]))
|
| 334 | cs = parent_cs(cs);
|
| 335 | nodes_and(*pmask, cs->mems_allowed, node_states[N_MEMORY]);
|
| 336 | }
|
| 337 |
|
| 338 | |
| 339 | * update task's spread flag if cpuset's page/slab spread flag is set
|
| 340 | *
|
| 341 | * Called with callback_mutex/cpuset_mutex held
|
| 342 |
|
| 343 | static void cpuset_update_task_spread_flag(struct cpuset *cs,
|
| 344 | struct task_struct *tsk)
|
| 345 | {
|
| 346 | if (is_spread_page(cs))
|
| 347 | tsk->flags |= PF_SPREAD_PAGE;
|
| 348 | else
|
| 349 | tsk->flags &= ~PF_SPREAD_PAGE;
|
| 350 | if (is_spread_slab(cs))
|
| 351 | tsk->flags |= PF_SPREAD_SLAB;
|
| 352 | else
|
| 353 | tsk->flags &= ~PF_SPREAD_SLAB;
|
| 354 | }
|
| 355 |
|
| 356 | |
| 357 | * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
|
| 358 | *
|
| 359 | * One cpuset is a subset of another if all its allowed CPUs and
|
| 360 | * Memory Nodes are a subset of the other, and its exclusive flags
|
| 361 | * are only set if the other's are set. Call holding cpuset_mutex.
|
| 362 |
|
| 363 |
|
| 364 | static int is_cpuset_subset(const struct cpuset *p, const struct cpuset *q)
|
| 365 | {
|
| 366 | return cpumask_subset(p->cpus_allowed, q->cpus_allowed) &&
|
| 367 | nodes_subset(p->mems_allowed, q->mems_allowed) &&
|
| 368 | is_cpu_exclusive(p) <= is_cpu_exclusive(q) &&
|
| 369 | is_mem_exclusive(p) <= is_mem_exclusive(q);
|
| 370 | }
|
| 371 |
|
| 372 | |
| 373 | * alloc_trial_cpuset - allocate a trial cpuset
|
| 374 | * @cs: the cpuset that the trial cpuset duplicates
|
| 375 |
|
| 376 | static struct cpuset *alloc_trial_cpuset(struct cpuset *cs)
|
| 377 | {
|
| 378 | struct cpuset *trial;
|
| 379 |
|
| 380 | trial = kmemdup(cs, sizeof(*cs), GFP_KERNEL);
|
| 381 | if (!trial)
|
| 382 | return NULL;
|
| 383 |
|
| 384 | if (!alloc_cpumask_var(&trial->cpus_allowed, GFP_KERNEL)) {
|
| 385 | kfree(trial);
|
| 386 | return NULL;
|
| 387 | }
|
| 388 | cpumask_copy(trial->cpus_allowed, cs->cpus_allowed);
|
| 389 |
|
| 390 | return trial;
|
| 391 | }
|
| 392 |
|
| 393 | |
| 394 | * free_trial_cpuset - free the trial cpuset
|
| 395 | * @trial: the trial cpuset to be freed
|
| 396 |
|
| 397 | static void free_trial_cpuset(struct cpuset *trial)
|
| 398 | {
|
| 399 | free_cpumask_var(trial->cpus_allowed);
|
| 400 | kfree(trial);
|
| 401 | }
|
| 402 |
|
| 403 | |
| 404 | * validate_change() - Used to validate that any proposed cpuset change
|
| 405 | * follows the structural rules for cpusets.
|
| 406 | *
|
| 407 | * If we replaced the flag and mask values of the current cpuset
|
| 408 | * (cur) with those values in the trial cpuset (trial), would
|
| 409 | * our various subset and exclusive rules still be valid? Presumes
|
| 410 | * cpuset_mutex held.
|
| 411 | *
|
| 412 | * 'cur' is the address of an actual, in-use cpuset. Operations
|
| 413 | * such as list traversal that depend on the actual address of the
|
| 414 | * cpuset in the list must use cur below, not trial.
|
| 415 | *
|
| 416 | * 'trial' is the address of bulk structure copy of cur, with
|
| 417 | * perhaps one or more of the fields cpus_allowed, mems_allowed,
|
| 418 | * or flags changed to new, trial values.
|
| 419 | *
|
| 420 | * Return 0 if valid, -errno if not.
|
| 421 |
|
| 422 |
|
| 423 | static int validate_change(struct cpuset *cur, struct cpuset *trial)
|
| 424 | {
|
| 425 | struct cgroup_subsys_state *css;
|
| 426 | struct cpuset *c, *par;
|
| 427 | int ret;
|
| 428 |
|
| 429 | rcu_read_lock();
|
| 430 |
|
| 431 |
|
| 432 | ret = -EBUSY;
|
| 433 | cpuset_for_each_child(c, css, cur)
|
| 434 | if (!is_cpuset_subset(c, trial))
|
| 435 | goto out;
|
| 436 |
|
| 437 |
|
| 438 | ret = 0;
|
| 439 | if (cur == &top_cpuset)
|
| 440 | goto out;
|
| 441 |
|
| 442 | par = parent_cs(cur);
|
| 443 |
|
| 444 |
|
| 445 | ret = -EACCES;
|
| 446 | if (!is_cpuset_subset(trial, par))
|
| 447 | goto out;
|
| 448 |
|
| 449 | |
| 450 | * If either I or some sibling (!= me) is exclusive, we can't
|
| 451 | * overlap
|
| 452 |
|
| 453 | ret = -EINVAL;
|
| 454 | cpuset_for_each_child(c, css, par) {
|
| 455 | if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
|
| 456 | c != cur &&
|
| 457 | cpumask_intersects(trial->cpus_allowed, c->cpus_allowed))
|
| 458 | goto out;
|
| 459 | if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) &&
|
| 460 | c != cur &&
|
| 461 | nodes_intersects(trial->mems_allowed, c->mems_allowed))
|
| 462 | goto out;
|
| 463 | }
|
| 464 |
|
| 465 | |
| 466 | * Cpusets with tasks - existing or newly being attached - can't
|
| 467 | * be changed to have empty cpus_allowed or mems_allowed.
|
| 468 |
|
| 469 | ret = -ENOSPC;
|
| 470 | if ((cgroup_task_count(cur->css.cgroup) || cur->attach_in_progress)) {
|
| 471 | if (!cpumask_empty(cur->cpus_allowed) &&
|
| 472 | cpumask_empty(trial->cpus_allowed))
|
| 473 | goto out;
|
| 474 | if (!nodes_empty(cur->mems_allowed) &&
|
| 475 | nodes_empty(trial->mems_allowed))
|
| 476 | goto out;
|
| 477 | }
|
| 478 |
|
| 479 | ret = 0;
|
| 480 | out:
|
| 481 | rcu_read_unlock();
|
| 482 | return ret;
|
| 483 | }
|
| 484 |
|
| 485 | #ifdef CONFIG_SMP
|
| 486 | |
| 487 | * Helper routine for generate_sched_domains().
|
| 488 | * Do cpusets a, b have overlapping cpus_allowed masks?
|
| 489 |
|
| 490 | static int cpusets_overlap(struct cpuset *a, struct cpuset *b)
|
| 491 | {
|
| 492 | return cpumask_intersects(a->cpus_allowed, b->cpus_allowed);
|
| 493 | }
|
| 494 |
|
| 495 | static void
|
| 496 | update_domain_attr(struct sched_domain_attr *dattr, struct cpuset *c)
|
| 497 | {
|
| 498 | if (dattr->relax_domain_level < c->relax_domain_level)
|
| 499 | dattr->relax_domain_level = c->relax_domain_level;
|
| 500 | return;
|
| 501 | }
|
| 502 |
|
| 503 | static void update_domain_attr_tree(struct sched_domain_attr *dattr,
|
| 504 | struct cpuset *root_cs)
|
| 505 | {
|
| 506 | struct cpuset *cp;
|
| 507 | struct cgroup_subsys_state *pos_css;
|
| 508 |
|
| 509 | rcu_read_lock();
|
| 510 | cpuset_for_each_descendant_pre(cp, pos_css, root_cs) {
|
| 511 | if (cp == root_cs)
|
| 512 | continue;
|
| 513 |
|
| 514 |
|
| 515 | if (cpumask_empty(cp->cpus_allowed)) {
|
| 516 | pos_css = css_rightmost_descendant(pos_css);
|
| 517 | continue;
|
| 518 | }
|
| 519 |
|
| 520 | if (is_sched_load_balance(cp))
|
| 521 | update_domain_attr(dattr, cp);
|
| 522 | }
|
| 523 | rcu_read_unlock();
|
| 524 | }
|
| 525 |
|
| 526 | |
| 527 | * generate_sched_domains()
|
| 528 | *
|
| 529 | * This function builds a partial partition of the systems CPUs
|
| 530 | * A 'partial partition' is a set of non-overlapping subsets whose
|
| 531 | * union is a subset of that set.
|
| 532 | * The output of this function needs to be passed to kernel/sched/core.c
|
| 533 | * partition_sched_domains() routine, which will rebuild the scheduler's
|
| 534 | * load balancing domains (sched domains) as specified by that partial
|
| 535 | * partition.
|
| 536 | *
|
| 537 | * See "What is sched_load_balance" in Documentation/cgroups/cpusets.txt
|
| 538 | * for a background explanation of this.
|
| 539 | *
|
| 540 | * Does not return errors, on the theory that the callers of this
|
| 541 | * routine would rather not worry about failures to rebuild sched
|
| 542 | * domains when operating in the severe memory shortage situations
|
| 543 | * that could cause allocation failures below.
|
| 544 | *
|
| 545 | * Must be called with cpuset_mutex held.
|
| 546 | *
|
| 547 | * The three key local variables below are:
|
| 548 | * q - a linked-list queue of cpuset pointers, used to implement a
|
| 549 | * top-down scan of all cpusets. This scan loads a pointer
|
| 550 | * to each cpuset marked is_sched_load_balance into the
|
| 551 | * array 'csa'. For our purposes, rebuilding the schedulers
|
| 552 | * sched domains, we can ignore !is_sched_load_balance cpusets.
|
| 553 | * csa - (for CpuSet Array) Array of pointers to all the cpusets
|
| 554 | * that need to be load balanced, for convenient iterative
|
| 555 | * access by the subsequent code that finds the best partition,
|
| 556 | * i.e the set of domains (subsets) of CPUs such that the
|
| 557 | * cpus_allowed of every cpuset marked is_sched_load_balance
|
| 558 | * is a subset of one of these domains, while there are as
|
| 559 | * many such domains as possible, each as small as possible.
|
| 560 | * doms - Conversion of 'csa' to an array of cpumasks, for passing to
|
| 561 | * the kernel/sched/core.c routine partition_sched_domains() in a
|
| 562 | * convenient format, that can be easily compared to the prior
|
| 563 | * value to determine what partition elements (sched domains)
|
| 564 | * were changed (added or removed.)
|
| 565 | *
|
| 566 | * Finding the best partition (set of domains):
|
| 567 | * The triple nested loops below over i, j, k scan over the
|
| 568 | * load balanced cpusets (using the array of cpuset pointers in
|
| 569 | * csa[]) looking for pairs of cpusets that have overlapping
|
| 570 | * cpus_allowed, but which don't have the same 'pn' partition
|
| 571 | * number and gives them in the same partition number. It keeps
|
| 572 | * looping on the 'restart' label until it can no longer find
|
| 573 | * any such pairs.
|
| 574 | *
|
| 575 | * The union of the cpus_allowed masks from the set of
|
| 576 | * all cpusets having the same 'pn' value then form the one
|
| 577 | * element of the partition (one sched domain) to be passed to
|
| 578 | * partition_sched_domains().
|
| 579 |
|
| 580 | static int generate_sched_domains(cpumask_var_t **domains,
|
| 581 | struct sched_domain_attr **attributes)
|
| 582 | {
|
| 583 | struct cpuset *cp;
|
| 584 | struct cpuset **csa;
|
| 585 | int csn;
|
| 586 | int i, j, k;
|
| 587 | cpumask_var_t *doms;
|
| 588 | struct sched_domain_attr *dattr;
|
| 589 | int ndoms = 0;
|
| 590 | int nslot;
|
| 591 | struct cgroup_subsys_state *pos_css;
|
| 592 |
|
| 593 | doms = NULL;
|
| 594 | dattr = NULL;
|
| 595 | csa = NULL;
|
| 596 |
|
| 597 |
|
| 598 | if (is_sched_load_balance(&top_cpuset)) {
|
| 599 | ndoms = 1;
|
| 600 | doms = alloc_sched_domains(ndoms);
|
| 601 | if (!doms)
|
| 602 | goto done;
|
| 603 |
|
| 604 | dattr = kmalloc(sizeof(struct sched_domain_attr), GFP_KERNEL);
|
| 605 | if (dattr) {
|
| 606 | *dattr = SD_ATTR_INIT;
|
| 607 | update_domain_attr_tree(dattr, &top_cpuset);
|
| 608 | }
|
| 609 | cpumask_copy(doms[0], top_cpuset.cpus_allowed);
|
| 610 |
|
| 611 | goto done;
|
| 612 | }
|
| 613 |
|
| 614 | csa = kmalloc(number_of_cpusets * sizeof(cp), GFP_KERNEL);
|
| 615 | if (!csa)
|
| 616 | goto done;
|
| 617 | csn = 0;
|
| 618 |
|
| 619 | rcu_read_lock();
|
| 620 | cpuset_for_each_descendant_pre(cp, pos_css, &top_cpuset) {
|
| 621 | if (cp == &top_cpuset)
|
| 622 | continue;
|
| 623 | |
| 624 | * Continue traversing beyond @cp iff @cp has some CPUs and
|
| 625 | * isn't load balancing. The former is obvious. The
|
| 626 | * latter: All child cpusets contain a subset of the
|
| 627 | * parent's cpus, so just skip them, and then we call
|
| 628 | * update_domain_attr_tree() to calc relax_domain_level of
|
| 629 | * the corresponding sched domain.
|
| 630 |
|
| 631 | if (!cpumask_empty(cp->cpus_allowed) &&
|
| 632 | !is_sched_load_balance(cp))
|
| 633 | continue;
|
| 634 |
|
| 635 | if (is_sched_load_balance(cp))
|
| 636 | csa[csn++] = cp;
|
| 637 |
|
| 638 |
|
| 639 | pos_css = css_rightmost_descendant(pos_css);
|
| 640 | }
|
| 641 | rcu_read_unlock();
|
| 642 |
|
| 643 | for (i = 0; i < csn; i++)
|
| 644 | csa[i]->pn = i;
|
| 645 | ndoms = csn;
|
| 646 |
|
| 647 | restart:
|
| 648 |
|
| 649 | for (i = 0; i < csn; i++) {
|
| 650 | struct cpuset *a = csa[i];
|
| 651 | int apn = a->pn;
|
| 652 |
|
| 653 | for (j = 0; j < csn; j++) {
|
| 654 | struct cpuset *b = csa[j];
|
| 655 | int bpn = b->pn;
|
| 656 |
|
| 657 | if (apn != bpn && cpusets_overlap(a, b)) {
|
| 658 | for (k = 0; k < csn; k++) {
|
| 659 | struct cpuset *c = csa[k];
|
| 660 |
|
| 661 | if (c->pn == bpn)
|
| 662 | c->pn = apn;
|
| 663 | }
|
| 664 | ndoms--;
|
| 665 | goto restart;
|
| 666 | }
|
| 667 | }
|
| 668 | }
|
| 669 |
|
| 670 | |
| 671 | * Now we know how many domains to create.
|
| 672 | * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
|
| 673 |
|
| 674 | doms = alloc_sched_domains(ndoms);
|
| 675 | if (!doms)
|
| 676 | goto done;
|
| 677 |
|
| 678 | |
| 679 | * The rest of the code, including the scheduler, can deal with
|
| 680 | * dattr==NULL case. No need to abort if alloc fails.
|
| 681 |
|
| 682 | dattr = kmalloc(ndoms * sizeof(struct sched_domain_attr), GFP_KERNEL);
|
| 683 |
|
| 684 | for (nslot = 0, i = 0; i < csn; i++) {
|
| 685 | struct cpuset *a = csa[i];
|
| 686 | struct cpumask *dp;
|
| 687 | int apn = a->pn;
|
| 688 |
|
| 689 | if (apn < 0) {
|
| 690 |
|
| 691 | continue;
|
| 692 | }
|
| 693 |
|
| 694 | dp = doms[nslot];
|
| 695 |
|
| 696 | if (nslot == ndoms) {
|
| 697 | static int warnings = 10;
|
| 698 | if (warnings) {
|
| 699 | printk(KERN_WARNING
|
| 700 | "rebuild_sched_domains confused:"
|
| 701 | " nslot %d, ndoms %d, csn %d, i %d,"
|
| 702 | " apn %d\n",
|
| 703 | nslot, ndoms, csn, i, apn);
|
| 704 | warnings--;
|
| 705 | }
|
| 706 | continue;
|
| 707 | }
|
| 708 |
|
| 709 | cpumask_clear(dp);
|
| 710 | if (dattr)
|
| 711 | *(dattr + nslot) = SD_ATTR_INIT;
|
| 712 | for (j = i; j < csn; j++) {
|
| 713 | struct cpuset *b = csa[j];
|
| 714 |
|
| 715 | if (apn == b->pn) {
|
| 716 | cpumask_or(dp, dp, b->cpus_allowed);
|
| 717 | if (dattr)
|
| 718 | update_domain_attr_tree(dattr + nslot, b);
|
| 719 |
|
| 720 |
|
| 721 | b->pn = -1;
|
| 722 | }
|
| 723 | }
|
| 724 | nslot++;
|
| 725 | }
|
| 726 | BUG_ON(nslot != ndoms);
|
| 727 |
|
| 728 | done:
|
| 729 | kfree(csa);
|
| 730 |
|
| 731 | |
| 732 | * Fallback to the default domain if kmalloc() failed.
|
| 733 | * See comments in partition_sched_domains().
|
| 734 |
|
| 735 | if (doms == NULL)
|
| 736 | ndoms = 1;
|
| 737 |
|
| 738 | *domains = doms;
|
| 739 | *attributes = dattr;
|
| 740 | return ndoms;
|
| 741 | }
|
| 742 |
|
| 743 | |
| 744 | * Rebuild scheduler domains.
|
| 745 | *
|
| 746 | * If the flag 'sched_load_balance' of any cpuset with non-empty
|
| 747 | * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
|
| 748 | * which has that flag enabled, or if any cpuset with a non-empty
|
| 749 | * 'cpus' is removed, then call this routine to rebuild the
|
| 750 | * scheduler's dynamic sched domains.
|
| 751 | *
|
| 752 | * Call with cpuset_mutex held. Takes get_online_cpus().
|
| 753 |
|
| 754 | static void rebuild_sched_domains_locked(void)
|
| 755 | {
|
| 756 | struct sched_domain_attr *attr;
|
| 757 | cpumask_var_t *doms;
|
| 758 | int ndoms;
|
| 759 |
|
| 760 | lockdep_assert_held(&cpuset_mutex);
|
| 761 | get_online_cpus();
|
| 762 |
|
| 763 | |
| 764 | * We have raced with CPU hotplug. Don't do anything to avoid
|
| 765 | * passing doms with offlined cpu to partition_sched_domains().
|
| 766 | * Anyways, hotplug work item will rebuild sched domains.
|
| 767 |
|
| 768 | if (!cpumask_equal(top_cpuset.cpus_allowed, cpu_active_mask))
|
| 769 | goto out;
|
| 770 |
|
| 771 |
|
| 772 | ndoms = generate_sched_domains(&doms, &attr);
|
| 773 |
|
| 774 |
|
| 775 | partition_sched_domains(ndoms, doms, attr);
|
| 776 | out:
|
| 777 | put_online_cpus();
|
| 778 | }
|
| 779 | #else
|
| 780 | static void rebuild_sched_domains_locked(void)
|
| 781 | {
|
| 782 | }
|
| 783 | #endif
|
| 784 |
|
| 785 | void rebuild_sched_domains(void)
|
| 786 | {
|
| 787 | mutex_lock(&cpuset_mutex);
|
| 788 | rebuild_sched_domains_locked();
|
| 789 | mutex_unlock(&cpuset_mutex);
|
| 790 | }
|
| 791 |
|
| 792 | |
| 793 | * effective_cpumask_cpuset - return nearest ancestor with non-empty cpus
|
| 794 | * @cs: the cpuset in interest
|
| 795 | *
|
| 796 | * A cpuset's effective cpumask is the cpumask of the nearest ancestor
|
| 797 | * with non-empty cpus. We use effective cpumask whenever:
|
| 798 | * - we update tasks' cpus_allowed. (they take on the ancestor's cpumask
|
| 799 | * if the cpuset they reside in has no cpus)
|
| 800 | * - we want to retrieve task_cs(tsk)'s cpus_allowed.
|
| 801 | *
|
| 802 | * Called with cpuset_mutex held. cpuset_cpus_allowed_fallback() is an
|
| 803 | * exception. See comments there.
|
| 804 |
|
| 805 | static struct cpuset *effective_cpumask_cpuset(struct cpuset *cs)
|
| 806 | {
|
| 807 | while (cpumask_empty(cs->cpus_allowed))
|
| 808 | cs = parent_cs(cs);
|
| 809 | return cs;
|
| 810 | }
|
| 811 |
|
| 812 | |
| 813 | * effective_nodemask_cpuset - return nearest ancestor with non-empty mems
|
| 814 | * @cs: the cpuset in interest
|
| 815 | *
|
| 816 | * A cpuset's effective nodemask is the nodemask of the nearest ancestor
|
| 817 | * with non-empty memss. We use effective nodemask whenever:
|
| 818 | * - we update tasks' mems_allowed. (they take on the ancestor's nodemask
|
| 819 | * if the cpuset they reside in has no mems)
|
| 820 | * - we want to retrieve task_cs(tsk)'s mems_allowed.
|
| 821 | *
|
| 822 | * Called with cpuset_mutex held.
|
| 823 |
|
| 824 | static struct cpuset *effective_nodemask_cpuset(struct cpuset *cs)
|
| 825 | {
|
| 826 | while (nodes_empty(cs->mems_allowed))
|
| 827 | cs = parent_cs(cs);
|
| 828 | return cs;
|
| 829 | }
|
| 830 |
|
| 831 | |
| 832 | * cpuset_change_cpumask - make a task's cpus_allowed the same as its cpuset's
|
| 833 | * @tsk: task to test
|
| 834 | * @data: cpuset to @tsk belongs to
|
| 835 | *
|
| 836 | * Called by css_scan_tasks() for each task in a cgroup whose cpus_allowed
|
| 837 | * mask needs to be changed.
|
| 838 | *
|
| 839 | * We don't need to re-check for the cgroup/cpuset membership, since we're
|
| 840 | * holding cpuset_mutex at this point.
|
| 841 |
|
| 842 | static void cpuset_change_cpumask(struct task_struct *tsk, void *data)
|
| 843 | {
|
| 844 | struct cpuset *cs = data;
|
| 845 | struct cpuset *cpus_cs = effective_cpumask_cpuset(cs);
|
| 846 |
|
| 847 | set_cpus_allowed_ptr(tsk, cpus_cs->cpus_allowed);
|
| 848 | }
|
| 849 |
|
| 850 | |
| 851 | * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
|
| 852 | * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
|
| 853 | * @heap: if NULL, defer allocating heap memory to css_scan_tasks()
|
| 854 | *
|
| 855 | * Called with cpuset_mutex held
|
| 856 | *
|
| 857 | * The css_scan_tasks() function will scan all the tasks in a cgroup,
|
| 858 | * calling callback functions for each.
|
| 859 | *
|
| 860 | * No return value. It's guaranteed that css_scan_tasks() always returns 0
|
| 861 | * if @heap != NULL.
|
| 862 |
|
| 863 | static void update_tasks_cpumask(struct cpuset *cs, struct ptr_heap *heap)
|
| 864 | {
|
| 865 | css_scan_tasks(&cs->css, NULL, cpuset_change_cpumask, cs, heap);
|
| 866 | }
|
| 867 |
|
| 868 | |
| 869 | * update_tasks_cpumask_hier - Update the cpumasks of tasks in the hierarchy.
|
| 870 | * @root_cs: the root cpuset of the hierarchy
|
| 871 | * @update_root: update root cpuset or not?
|
| 872 | * @heap: the heap used by css_scan_tasks()
|
| 873 | *
|
| 874 | * This will update cpumasks of tasks in @root_cs and all other empty cpusets
|
| 875 | * which take on cpumask of @root_cs.
|
| 876 | *
|
| 877 | * Called with cpuset_mutex held
|
| 878 |
|
| 879 | static void update_tasks_cpumask_hier(struct cpuset *root_cs,
|
| 880 | bool update_root, struct ptr_heap *heap)
|
| 881 | {
|
| 882 | struct cpuset *cp;
|
| 883 | struct cgroup_subsys_state *pos_css;
|
| 884 |
|
| 885 | rcu_read_lock();
|
| 886 | cpuset_for_each_descendant_pre(cp, pos_css, root_cs) {
|
| 887 | if (cp == root_cs) {
|
| 888 | if (!update_root)
|
| 889 | continue;
|
| 890 | } else {
|
| 891 |
|
| 892 | if (!cpumask_empty(cp->cpus_allowed)) {
|
| 893 | pos_css = css_rightmost_descendant(pos_css);
|
| 894 | continue;
|
| 895 | }
|
| 896 | }
|
| 897 | if (!css_tryget(&cp->css))
|
| 898 | continue;
|
| 899 | rcu_read_unlock();
|
| 900 |
|
| 901 | update_tasks_cpumask(cp, heap);
|
| 902 |
|
| 903 | rcu_read_lock();
|
| 904 | css_put(&cp->css);
|
| 905 | }
|
| 906 | rcu_read_unlock();
|
| 907 | }
|
| 908 |
|
| 909 | |
| 910 | * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
|
| 911 | * @cs: the cpuset to consider
|
| 912 | * @buf: buffer of cpu numbers written to this cpuset
|
| 913 |
|
| 914 | static int update_cpumask(struct cpuset *cs, struct cpuset *trialcs,
|
| 915 | const char *buf)
|
| 916 | {
|
| 917 | struct ptr_heap heap;
|
| 918 | int retval;
|
| 919 | int is_load_balanced;
|
| 920 |
|
| 921 |
|
| 922 | if (cs == &top_cpuset)
|
| 923 | return -EACCES;
|
| 924 |
|
| 925 | |
| 926 | * An empty cpus_allowed is ok only if the cpuset has no tasks.
|
| 927 | * Since cpulist_parse() fails on an empty mask, we special case
|
| 928 | * that parsing. The validate_change() call ensures that cpusets
|
| 929 | * with tasks have cpus.
|
| 930 |
|
| 931 | if (!*buf) {
|
| 932 | cpumask_clear(trialcs->cpus_allowed);
|
| 933 | } else {
|
| 934 | retval = cpulist_parse(buf, trialcs->cpus_allowed);
|
| 935 | if (retval < 0)
|
| 936 | return retval;
|
| 937 |
|
| 938 | if (!cpumask_subset(trialcs->cpus_allowed, cpu_active_mask))
|
| 939 | return -EINVAL;
|
| 940 | }
|
| 941 |
|
| 942 |
|
| 943 | if (cpumask_equal(cs->cpus_allowed, trialcs->cpus_allowed))
|
| 944 | return 0;
|
| 945 |
|
| 946 | retval = validate_change(cs, trialcs);
|
| 947 | if (retval < 0)
|
| 948 | return retval;
|
| 949 |
|
| 950 | retval = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
|
| 951 | if (retval)
|
| 952 | return retval;
|
| 953 |
|
| 954 | is_load_balanced = is_sched_load_balance(trialcs);
|
| 955 |
|
| 956 | mutex_lock(&callback_mutex);
|
| 957 | cpumask_copy(cs->cpus_allowed, trialcs->cpus_allowed);
|
| 958 | mutex_unlock(&callback_mutex);
|
| 959 |
|
| 960 | update_tasks_cpumask_hier(cs, true, &heap);
|
| 961 |
|
| 962 | heap_free(&heap);
|
| 963 |
|
| 964 | if (is_load_balanced)
|
| 965 | rebuild_sched_domains_locked();
|
| 966 | return 0;
|
| 967 | }
|
| 968 |
|
| 969 | |
| 970 | * cpuset_migrate_mm
|
| 971 | *
|
| 972 | * Migrate memory region from one set of nodes to another.
|
| 973 | *
|
| 974 | * Temporarilly set tasks mems_allowed to target nodes of migration,
|
| 975 | * so that the migration code can allocate pages on these nodes.
|
| 976 | *
|
| 977 | * Call holding cpuset_mutex, so current's cpuset won't change
|
| 978 | * during this call, as manage_mutex holds off any cpuset_attach()
|
| 979 | * calls. Therefore we don't need to take task_lock around the
|
| 980 | * call to guarantee_online_mems(), as we know no one is changing
|
| 981 | * our task's cpuset.
|
| 982 | *
|
| 983 | * While the mm_struct we are migrating is typically from some
|
| 984 | * other task, the task_struct mems_allowed that we are hacking
|
| 985 | * is for our current task, which must allocate new pages for that
|
| 986 | * migrating memory region.
|
| 987 |
|
| 988 |
|
| 989 | static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
|
| 990 | const nodemask_t *to)
|
| 991 | {
|
| 992 | struct task_struct *tsk = current;
|
| 993 | struct cpuset *mems_cs;
|
| 994 |
|
| 995 | tsk->mems_allowed = *to;
|
| 996 |
|
| 997 | do_migrate_pages(mm, from, to, MPOL_MF_MOVE_ALL);
|
| 998 |
|
| 999 | mems_cs = effective_nodemask_cpuset(task_cs(tsk));
|
| 1000 | guarantee_online_mems(mems_cs, &tsk->mems_allowed);
|
| 1001 | }
|
| 1002 |
|
| 1003 | |
| 1004 | * cpuset_change_task_nodemask - change task's mems_allowed and mempolicy
|
| 1005 | * @tsk: the task to change
|
| 1006 | * @newmems: new nodes that the task will be set
|
| 1007 | *
|
| 1008 | * In order to avoid seeing no nodes if the old and new nodes are disjoint,
|
| 1009 | * we structure updates as setting all new allowed nodes, then clearing newly
|
| 1010 | * disallowed ones.
|
| 1011 |
|
| 1012 | static void cpuset_change_task_nodemask(struct task_struct *tsk,
|
| 1013 | nodemask_t *newmems)
|
| 1014 | {
|
| 1015 | bool need_loop;
|
| 1016 |
|
| 1017 | |
| 1018 | * Allow tasks that have access to memory reserves because they have
|
| 1019 | * been OOM killed to get memory anywhere.
|
| 1020 |
|
| 1021 | if (unlikely(test_thread_flag(TIF_MEMDIE)))
|
| 1022 | return;
|
| 1023 | if (current->flags & PF_EXITING)
|
| 1024 | return;
|
| 1025 |
|
| 1026 | task_lock(tsk);
|
| 1027 | |
| 1028 | * Determine if a loop is necessary if another thread is doing
|
| 1029 | * get_mems_allowed(). If at least one node remains unchanged and
|
| 1030 | * tsk does not have a mempolicy, then an empty nodemask will not be
|
| 1031 | * possible when mems_allowed is larger than a word.
|
| 1032 |
|
| 1033 | need_loop = task_has_mempolicy(tsk) ||
|
| 1034 | !nodes_intersects(*newmems, tsk->mems_allowed);
|
| 1035 |
|
| 1036 | if (need_loop) {
|
| 1037 | local_irq_disable();
|
| 1038 | write_seqcount_begin(&tsk->mems_allowed_seq);
|
| 1039 | }
|
| 1040 |
|
| 1041 | nodes_or(tsk->mems_allowed, tsk->mems_allowed, *newmems);
|
| 1042 | mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP1);
|
| 1043 |
|
| 1044 | mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP2);
|
| 1045 | tsk->mems_allowed = *newmems;
|
| 1046 |
|
| 1047 | if (need_loop) {
|
| 1048 | write_seqcount_end(&tsk->mems_allowed_seq);
|
| 1049 | local_irq_enable();
|
| 1050 | }
|
| 1051 |
|
| 1052 | task_unlock(tsk);
|
| 1053 | }
|
| 1054 |
|
| 1055 | struct cpuset_change_nodemask_arg {
|
| 1056 | struct cpuset *cs;
|
| 1057 | nodemask_t *newmems;
|
| 1058 | };
|
| 1059 |
|
| 1060 | |
| 1061 | * Update task's mems_allowed and rebind its mempolicy and vmas' mempolicy
|
| 1062 | * of it to cpuset's new mems_allowed, and migrate pages to new nodes if
|
| 1063 | * memory_migrate flag is set. Called with cpuset_mutex held.
|
| 1064 |
|
| 1065 | static void cpuset_change_nodemask(struct task_struct *p, void *data)
|
| 1066 | {
|
| 1067 | struct cpuset_change_nodemask_arg *arg = data;
|
| 1068 | struct cpuset *cs = arg->cs;
|
| 1069 | struct mm_struct *mm;
|
| 1070 | int migrate;
|
| 1071 |
|
| 1072 | cpuset_change_task_nodemask(p, arg->newmems);
|
| 1073 |
|
| 1074 | mm = get_task_mm(p);
|
| 1075 | if (!mm)
|
| 1076 | return;
|
| 1077 |
|
| 1078 | migrate = is_memory_migrate(cs);
|
| 1079 |
|
| 1080 | mpol_rebind_mm(mm, &cs->mems_allowed);
|
| 1081 | if (migrate)
|
| 1082 | cpuset_migrate_mm(mm, &cs->old_mems_allowed, arg->newmems);
|
| 1083 | mmput(mm);
|
| 1084 | }
|
| 1085 |
|
| 1086 | static void *cpuset_being_rebound;
|
| 1087 |
|
| 1088 | |
| 1089 | * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
|
| 1090 | * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
|
| 1091 | * @heap: if NULL, defer allocating heap memory to css_scan_tasks()
|
| 1092 | *
|
| 1093 | * Called with cpuset_mutex held. No return value. It's guaranteed that
|
| 1094 | * css_scan_tasks() always returns 0 if @heap != NULL.
|
| 1095 |
|
| 1096 | static void update_tasks_nodemask(struct cpuset *cs, struct ptr_heap *heap)
|
| 1097 | {
|
| 1098 | static nodemask_t newmems;
|
| 1099 | struct cpuset *mems_cs = effective_nodemask_cpuset(cs);
|
| 1100 | struct cpuset_change_nodemask_arg arg = { .cs = cs,
|
| 1101 | .newmems = &newmems };
|
| 1102 |
|
| 1103 | cpuset_being_rebound = cs;
|
| 1104 |
|
| 1105 | guarantee_online_mems(mems_cs, &newmems);
|
| 1106 |
|
| 1107 | |
| 1108 | * The mpol_rebind_mm() call takes mmap_sem, which we couldn't
|
| 1109 | * take while holding tasklist_lock. Forks can happen - the
|
| 1110 | * mpol_dup() cpuset_being_rebound check will catch such forks,
|
| 1111 | * and rebind their vma mempolicies too. Because we still hold
|
| 1112 | * the global cpuset_mutex, we know that no other rebind effort
|
| 1113 | * will be contending for the global variable cpuset_being_rebound.
|
| 1114 | * It's ok if we rebind the same mm twice; mpol_rebind_mm()
|
| 1115 | * is idempotent. Also migrate pages in each mm to new nodes.
|
| 1116 |
|
| 1117 | css_scan_tasks(&cs->css, NULL, cpuset_change_nodemask, &arg, heap);
|
| 1118 |
|
| 1119 | |
| 1120 | * All the tasks' nodemasks have been updated, update
|
| 1121 | * cs->old_mems_allowed.
|
| 1122 |
|
| 1123 | cs->old_mems_allowed = newmems;
|
| 1124 |
|
| 1125 |
|
| 1126 | cpuset_being_rebound = NULL;
|
| 1127 | }
|
| 1128 |
|
| 1129 | |
| 1130 | * update_tasks_nodemask_hier - Update the nodemasks of tasks in the hierarchy.
|
| 1131 | * @cs: the root cpuset of the hierarchy
|
| 1132 | * @update_root: update the root cpuset or not?
|
| 1133 | * @heap: the heap used by css_scan_tasks()
|
| 1134 | *
|
| 1135 | * This will update nodemasks of tasks in @root_cs and all other empty cpusets
|
| 1136 | * which take on nodemask of @root_cs.
|
| 1137 | *
|
| 1138 | * Called with cpuset_mutex held
|
| 1139 |
|
| 1140 | static void update_tasks_nodemask_hier(struct cpuset *root_cs,
|
| 1141 | bool update_root, struct ptr_heap *heap)
|
| 1142 | {
|
| 1143 | struct cpuset *cp;
|
| 1144 | struct cgroup_subsys_state *pos_css;
|
| 1145 |
|
| 1146 | rcu_read_lock();
|
| 1147 | cpuset_for_each_descendant_pre(cp, pos_css, root_cs) {
|
| 1148 | if (cp == root_cs) {
|
| 1149 | if (!update_root)
|
| 1150 | continue;
|
| 1151 | } else {
|
| 1152 |
|
| 1153 | if (!nodes_empty(cp->mems_allowed)) {
|
| 1154 | pos_css = css_rightmost_descendant(pos_css);
|
| 1155 | continue;
|
| 1156 | }
|
| 1157 | }
|
| 1158 | if (!css_tryget(&cp->css))
|
| 1159 | continue;
|
| 1160 | rcu_read_unlock();
|
| 1161 |
|
| 1162 | update_tasks_nodemask(cp, heap);
|
| 1163 |
|
| 1164 | rcu_read_lock();
|
| 1165 | css_put(&cp->css);
|
| 1166 | }
|
| 1167 | rcu_read_unlock();
|
| 1168 | }
|
| 1169 |
|
| 1170 | |
| 1171 | * Handle user request to change the 'mems' memory placement
|
| 1172 | * of a cpuset. Needs to validate the request, update the
|
| 1173 | * cpusets mems_allowed, and for each task in the cpuset,
|
| 1174 | * update mems_allowed and rebind task's mempolicy and any vma
|
| 1175 | * mempolicies and if the cpuset is marked 'memory_migrate',
|
| 1176 | * migrate the tasks pages to the new memory.
|
| 1177 | *
|
| 1178 | * Call with cpuset_mutex held. May take callback_mutex during call.
|
| 1179 | * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
|
| 1180 | * lock each such tasks mm->mmap_sem, scan its vma's and rebind
|
| 1181 | * their mempolicies to the cpusets new mems_allowed.
|
| 1182 |
|
| 1183 | static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
|
| 1184 | const char *buf)
|
| 1185 | {
|
| 1186 | int retval;
|
| 1187 | struct ptr_heap heap;
|
| 1188 |
|
| 1189 | |
| 1190 | * top_cpuset.mems_allowed tracks node_stats[N_MEMORY];
|
| 1191 | * it's read-only
|
| 1192 |
|
| 1193 | if (cs == &top_cpuset) {
|
| 1194 | retval = -EACCES;
|
| 1195 | goto done;
|
| 1196 | }
|
| 1197 |
|
| 1198 | |
| 1199 | * An empty mems_allowed is ok iff there are no tasks in the cpuset.
|
| 1200 | * Since nodelist_parse() fails on an empty mask, we special case
|
| 1201 | * that parsing. The validate_change() call ensures that cpusets
|
| 1202 | * with tasks have memory.
|
| 1203 |
|
| 1204 | if (!*buf) {
|
| 1205 | nodes_clear(trialcs->mems_allowed);
|
| 1206 | } else {
|
| 1207 | retval = nodelist_parse(buf, trialcs->mems_allowed);
|
| 1208 | if (retval < 0)
|
| 1209 | goto done;
|
| 1210 |
|
| 1211 | if (!nodes_subset(trialcs->mems_allowed,
|
| 1212 | node_states[N_MEMORY])) {
|
| 1213 | retval = -EINVAL;
|
| 1214 | goto done;
|
| 1215 | }
|
| 1216 | }
|
| 1217 |
|
| 1218 | if (nodes_equal(cs->mems_allowed, trialcs->mems_allowed)) {
|
| 1219 | retval = 0;
|
| 1220 | goto done;
|
| 1221 | }
|
| 1222 | retval = validate_change(cs, trialcs);
|
| 1223 | if (retval < 0)
|
| 1224 | goto done;
|
| 1225 |
|
| 1226 | retval = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
|
| 1227 | if (retval < 0)
|
| 1228 | goto done;
|
| 1229 |
|
| 1230 | mutex_lock(&callback_mutex);
|
| 1231 | cs->mems_allowed = trialcs->mems_allowed;
|
| 1232 | mutex_unlock(&callback_mutex);
|
| 1233 |
|
| 1234 | update_tasks_nodemask_hier(cs, true, &heap);
|
| 1235 |
|
| 1236 | heap_free(&heap);
|
| 1237 | done:
|
| 1238 | return retval;
|
| 1239 | }
|
| 1240 |
|
| 1241 | int current_cpuset_is_being_rebound(void)
|
| 1242 | {
|
| 1243 | return task_cs(current) == cpuset_being_rebound;
|
| 1244 | }
|
| 1245 |
|
| 1246 | static int update_relax_domain_level(struct cpuset *cs, s64 val)
|
| 1247 | {
|
| 1248 | #ifdef CONFIG_SMP
|
| 1249 | if (val < -1 || val >= sched_domain_level_max)
|
| 1250 | return -EINVAL;
|
| 1251 | #endif
|
| 1252 |
|
| 1253 | if (val != cs->relax_domain_level) {
|
| 1254 | cs->relax_domain_level = val;
|
| 1255 | if (!cpumask_empty(cs->cpus_allowed) &&
|
| 1256 | is_sched_load_balance(cs))
|
| 1257 | rebuild_sched_domains_locked();
|
| 1258 | }
|
| 1259 |
|
| 1260 | return 0;
|
| 1261 | }
|
| 1262 |
|
| 1263 | |
| 1264 | * cpuset_change_flag - make a task's spread flags the same as its cpuset's
|
| 1265 | * @tsk: task to be updated
|
| 1266 | * @data: cpuset to @tsk belongs to
|
| 1267 | *
|
| 1268 | * Called by css_scan_tasks() for each task in a cgroup.
|
| 1269 | *
|
| 1270 | * We don't need to re-check for the cgroup/cpuset membership, since we're
|
| 1271 | * holding cpuset_mutex at this point.
|
| 1272 |
|
| 1273 | static void cpuset_change_flag(struct task_struct *tsk, void *data)
|
| 1274 | {
|
| 1275 | struct cpuset *cs = data;
|
| 1276 |
|
| 1277 | cpuset_update_task_spread_flag(cs, tsk);
|
| 1278 | }
|
| 1279 |
|
| 1280 | |
| 1281 | * update_tasks_flags - update the spread flags of tasks in the cpuset.
|
| 1282 | * @cs: the cpuset in which each task's spread flags needs to be changed
|
| 1283 | * @heap: if NULL, defer allocating heap memory to css_scan_tasks()
|
| 1284 | *
|
| 1285 | * Called with cpuset_mutex held
|
| 1286 | *
|
| 1287 | * The css_scan_tasks() function will scan all the tasks in a cgroup,
|
| 1288 | * calling callback functions for each.
|
| 1289 | *
|
| 1290 | * No return value. It's guaranteed that css_scan_tasks() always returns 0
|
| 1291 | * if @heap != NULL.
|
| 1292 |
|
| 1293 | static void update_tasks_flags(struct cpuset *cs, struct ptr_heap *heap)
|
| 1294 | {
|
| 1295 | css_scan_tasks(&cs->css, NULL, cpuset_change_flag, cs, heap);
|
| 1296 | }
|
| 1297 |
|
| 1298 | |
| 1299 | * update_flag - read a 0 or a 1 in a file and update associated flag
|
| 1300 | * bit: the bit to update (see cpuset_flagbits_t)
|
| 1301 | * cs: the cpuset to update
|
| 1302 | * turning_on: whether the flag is being set or cleared
|
| 1303 | *
|
| 1304 | * Call with cpuset_mutex held.
|
| 1305 |
|
| 1306 |
|
| 1307 | static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs,
|
| 1308 | int turning_on)
|
| 1309 | {
|
| 1310 | struct cpuset *trialcs;
|
| 1311 | int balance_flag_changed;
|
| 1312 | int spread_flag_changed;
|
| 1313 | struct ptr_heap heap;
|
| 1314 | int err;
|
| 1315 |
|
| 1316 | trialcs = alloc_trial_cpuset(cs);
|
| 1317 | if (!trialcs)
|
| 1318 | return -ENOMEM;
|
| 1319 |
|
| 1320 | if (turning_on)
|
| 1321 | set_bit(bit, &trialcs->flags);
|
| 1322 | else
|
| 1323 | clear_bit(bit, &trialcs->flags);
|
| 1324 |
|
| 1325 | err = validate_change(cs, trialcs);
|
| 1326 | if (err < 0)
|
| 1327 | goto out;
|
| 1328 |
|
| 1329 | err = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
|
| 1330 | if (err < 0)
|
| 1331 | goto out;
|
| 1332 |
|
| 1333 | balance_flag_changed = (is_sched_load_balance(cs) !=
|
| 1334 | is_sched_load_balance(trialcs));
|
| 1335 |
|
| 1336 | spread_flag_changed = ((is_spread_slab(cs) != is_spread_slab(trialcs))
|
| 1337 | || (is_spread_page(cs) != is_spread_page(trialcs)));
|
| 1338 |
|
| 1339 | mutex_lock(&callback_mutex);
|
| 1340 | cs->flags = trialcs->flags;
|
| 1341 | mutex_unlock(&callback_mutex);
|
| 1342 |
|
| 1343 | if (!cpumask_empty(trialcs->cpus_allowed) && balance_flag_changed)
|
| 1344 | rebuild_sched_domains_locked();
|
| 1345 |
|
| 1346 | if (spread_flag_changed)
|
| 1347 | update_tasks_flags(cs, &heap);
|
| 1348 | heap_free(&heap);
|
| 1349 | out:
|
| 1350 | free_trial_cpuset(trialcs);
|
| 1351 | return err;
|
| 1352 | }
|
| 1353 |
|
| 1354 | |
| 1355 | * Frequency meter - How fast is some event occurring?
|
| 1356 | *
|
| 1357 | * These routines manage a digitally filtered, constant time based,
|
| 1358 | * event frequency meter. There are four routines:
|
| 1359 | * fmeter_init() - initialize a frequency meter.
|
| 1360 | * fmeter_markevent() - called each time the event happens.
|
| 1361 | * fmeter_getrate() - returns the recent rate of such events.
|
| 1362 | * fmeter_update() - internal routine used to update fmeter.
|
| 1363 | *
|
| 1364 | * A common data structure is passed to each of these routines,
|
| 1365 | * which is used to keep track of the state required to manage the
|
| 1366 | * frequency meter and its digital filter.
|
| 1367 | *
|
| 1368 | * The filter works on the number of events marked per unit time.
|
| 1369 | * The filter is single-pole low-pass recursive (IIR). The time unit
|
| 1370 | * is 1 second. Arithmetic is done using 32-bit integers scaled to
|
| 1371 | * simulate 3 decimal digits of precision (multiplied by 1000).
|
| 1372 | *
|
| 1373 | * With an FM_COEF of 933, and a time base of 1 second, the filter
|
| 1374 | * has a half-life of 10 seconds, meaning that if the events quit
|
| 1375 | * happening, then the rate returned from the fmeter_getrate()
|
| 1376 | * will be cut in half each 10 seconds, until it converges to zero.
|
| 1377 | *
|
| 1378 | * It is not worth doing a real infinitely recursive filter. If more
|
| 1379 | * than FM_MAXTICKS ticks have elapsed since the last filter event,
|
| 1380 | * just compute FM_MAXTICKS ticks worth, by which point the level
|
| 1381 | * will be stable.
|
| 1382 | *
|
| 1383 | * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
|
| 1384 | * arithmetic overflow in the fmeter_update() routine.
|
| 1385 | *
|
| 1386 | * Given the simple 32 bit integer arithmetic used, this meter works
|
| 1387 | * best for reporting rates between one per millisecond (msec) and
|
| 1388 | * one per 32 (approx) seconds. At constant rates faster than one
|
| 1389 | * per msec it maxes out at values just under 1,000,000. At constant
|
| 1390 | * rates between one per msec, and one per second it will stabilize
|
| 1391 | * to a value N*1000, where N is the rate of events per second.
|
| 1392 | * At constant rates between one per second and one per 32 seconds,
|
| 1393 | * it will be choppy, moving up on the seconds that have an event,
|
| 1394 | * and then decaying until the next event. At rates slower than
|
| 1395 | * about one in 32 seconds, it decays all the way back to zero between
|
| 1396 | * each event.
|
| 1397 |
|
| 1398 |
|
| 1399 | #define FM_COEF 933
|
| 1400 | #define FM_MAXTICKS ((time_t)99)
|
| 1401 | #define FM_MAXCNT 1000000
|
| 1402 | #define FM_SCALE 1000
|
| 1403 |
|
| 1404 |
|
| 1405 | static void fmeter_init(struct fmeter *fmp)
|
| 1406 | {
|
| 1407 | fmp->cnt = 0;
|
| 1408 | fmp->val = 0;
|
| 1409 | fmp->time = 0;
|
| 1410 | spin_lock_init(&fmp->lock);
|
| 1411 | }
|
| 1412 |
|
| 1413 |
|
| 1414 | static void fmeter_update(struct fmeter *fmp)
|
| 1415 | {
|
| 1416 | time_t now = get_seconds();
|
| 1417 | time_t ticks = now - fmp->time;
|
| 1418 |
|
| 1419 | if (ticks == 0)
|
| 1420 | return;
|
| 1421 |
|
| 1422 | ticks = min(FM_MAXTICKS, ticks);
|
| 1423 | while (ticks-- > 0)
|
| 1424 | fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
|
| 1425 | fmp->time = now;
|
| 1426 |
|
| 1427 | fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
|
| 1428 | fmp->cnt = 0;
|
| 1429 | }
|
| 1430 |
|
| 1431 |
|
| 1432 | static void fmeter_markevent(struct fmeter *fmp)
|
| 1433 | {
|
| 1434 | spin_lock(&fmp->lock);
|
| 1435 | fmeter_update(fmp);
|
| 1436 | fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
|
| 1437 | spin_unlock(&fmp->lock);
|
| 1438 | }
|
| 1439 |
|
| 1440 |
|
| 1441 | static int fmeter_getrate(struct fmeter *fmp)
|
| 1442 | {
|
| 1443 | int val;
|
| 1444 |
|
| 1445 | spin_lock(&fmp->lock);
|
| 1446 | fmeter_update(fmp);
|
| 1447 | val = fmp->val;
|
| 1448 | spin_unlock(&fmp->lock);
|
| 1449 | return val;
|
| 1450 | }
|
| 1451 |
|
| 1452 |
|
| 1453 | static int cpuset_can_attach(struct cgroup_subsys_state *css,
|
| 1454 | struct cgroup_taskset *tset)
|
| 1455 | {
|
| 1456 | struct cpuset *cs = css_cs(css);
|
| 1457 | struct task_struct *task;
|
| 1458 | int ret;
|
| 1459 |
|
| 1460 | mutex_lock(&cpuset_mutex);
|
| 1461 |
|
| 1462 | |
| 1463 | * We allow to move tasks into an empty cpuset if sane_behavior
|
| 1464 | * flag is set.
|
| 1465 |
|
| 1466 | ret = -ENOSPC;
|
| 1467 | if (!cgroup_sane_behavior(css->cgroup) &&
|
| 1468 | (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed)))
|
| 1469 | goto out_unlock;
|
| 1470 |
|
| 1471 | cgroup_taskset_for_each(task, css, tset) {
|
| 1472 | |
| 1473 | * Kthreads which disallow setaffinity shouldn't be moved
|
| 1474 | * to a new cpuset; we don't want to change their cpu
|
| 1475 | * affinity and isolating such threads by their set of
|
| 1476 | * allowed nodes is unnecessary. Thus, cpusets are not
|
| 1477 | * applicable for such threads. This prevents checking for
|
| 1478 | * success of set_cpus_allowed_ptr() on all attached tasks
|
| 1479 | * before cpus_allowed may be changed.
|
| 1480 |
|
| 1481 | ret = -EINVAL;
|
| 1482 | if (task->flags & PF_NO_SETAFFINITY)
|
| 1483 | goto out_unlock;
|
| 1484 | ret = security_task_setscheduler(task);
|
| 1485 | if (ret)
|
| 1486 | goto out_unlock;
|
| 1487 | }
|
| 1488 |
|
| 1489 | |
| 1490 | * Mark attach is in progress. This makes validate_change() fail
|
| 1491 | * changes which zero cpus/mems_allowed.
|
| 1492 |
|
| 1493 | cs->attach_in_progress++;
|
| 1494 | ret = 0;
|
| 1495 | out_unlock:
|
| 1496 | mutex_unlock(&cpuset_mutex);
|
| 1497 | return ret;
|
| 1498 | }
|
| 1499 |
|
| 1500 | static void cpuset_cancel_attach(struct cgroup_subsys_state *css,
|
| 1501 | struct cgroup_taskset *tset)
|
| 1502 | {
|
| 1503 | mutex_lock(&cpuset_mutex);
|
| 1504 | css_cs(css)->attach_in_progress--;
|
| 1505 | mutex_unlock(&cpuset_mutex);
|
| 1506 | }
|
| 1507 |
|
| 1508 | |
| 1509 | * Protected by cpuset_mutex. cpus_attach is used only by cpuset_attach()
|
| 1510 | * but we can't allocate it dynamically there. Define it global and
|
| 1511 | * allocate from cpuset_init().
|
| 1512 |
|
| 1513 | static cpumask_var_t cpus_attach;
|
| 1514 |
|
| 1515 | static void cpuset_attach(struct cgroup_subsys_state *css,
|
| 1516 | struct cgroup_taskset *tset)
|
| 1517 | {
|
| 1518 |
|
| 1519 | static nodemask_t cpuset_attach_nodemask_to;
|
| 1520 | struct mm_struct *mm;
|
| 1521 | struct task_struct *task;
|
| 1522 | struct task_struct *leader = cgroup_taskset_first(tset);
|
| 1523 | struct cgroup_subsys_state *oldcss = cgroup_taskset_cur_css(tset,
|
| 1524 | cpuset_subsys_id);
|
| 1525 | struct cpuset *cs = css_cs(css);
|
| 1526 | struct cpuset *oldcs = css_cs(oldcss);
|
| 1527 | struct cpuset *cpus_cs = effective_cpumask_cpuset(cs);
|
| 1528 | struct cpuset *mems_cs = effective_nodemask_cpuset(cs);
|
| 1529 |
|
| 1530 | mutex_lock(&cpuset_mutex);
|
| 1531 |
|
| 1532 |
|
| 1533 | if (cs == &top_cpuset)
|
| 1534 | cpumask_copy(cpus_attach, cpu_possible_mask);
|
| 1535 | else
|
| 1536 | guarantee_online_cpus(cpus_cs, cpus_attach);
|
| 1537 |
|
| 1538 | guarantee_online_mems(mems_cs, &cpuset_attach_nodemask_to);
|
| 1539 |
|
| 1540 | cgroup_taskset_for_each(task, css, tset) {
|
| 1541 | |
| 1542 | * can_attach beforehand should guarantee that this doesn't
|
| 1543 | * fail. TODO: have a better way to handle failure here
|
| 1544 |
|
| 1545 | WARN_ON_ONCE(set_cpus_allowed_ptr(task, cpus_attach));
|
| 1546 |
|
| 1547 | cpuset_change_task_nodemask(task, &cpuset_attach_nodemask_to);
|
| 1548 | cpuset_update_task_spread_flag(cs, task);
|
| 1549 | }
|
| 1550 |
|
| 1551 | |
| 1552 | * Change mm, possibly for multiple threads in a threadgroup. This is
|
| 1553 | * expensive and may sleep.
|
| 1554 |
|
| 1555 | cpuset_attach_nodemask_to = cs->mems_allowed;
|
| 1556 | mm = get_task_mm(leader);
|
| 1557 | if (mm) {
|
| 1558 | struct cpuset *mems_oldcs = effective_nodemask_cpuset(oldcs);
|
| 1559 |
|
| 1560 | mpol_rebind_mm(mm, &cpuset_attach_nodemask_to);
|
| 1561 |
|
| 1562 | |
| 1563 | * old_mems_allowed is the same with mems_allowed here, except
|
| 1564 | * if this task is being moved automatically due to hotplug.
|
| 1565 | * In that case @mems_allowed has been updated and is empty,
|
| 1566 | * so @old_mems_allowed is the right nodesets that we migrate
|
| 1567 | * mm from.
|
| 1568 |
|
| 1569 | if (is_memory_migrate(cs)) {
|
| 1570 | cpuset_migrate_mm(mm, &mems_oldcs->old_mems_allowed,
|
| 1571 | &cpuset_attach_nodemask_to);
|
| 1572 | }
|
| 1573 | mmput(mm);
|
| 1574 | }
|
| 1575 |
|
| 1576 | cs->old_mems_allowed = cpuset_attach_nodemask_to;
|
| 1577 |
|
| 1578 | cs->attach_in_progress--;
|
| 1579 | if (!cs->attach_in_progress)
|
| 1580 | wake_up(&cpuset_attach_wq);
|
| 1581 |
|
| 1582 | mutex_unlock(&cpuset_mutex);
|
| 1583 | }
|
| 1584 |
|
| 1585 |
|
| 1586 |
|
| 1587 | typedef enum {
|
| 1588 | FILE_MEMORY_MIGRATE,
|
| 1589 | FILE_CPULIST,
|
| 1590 | FILE_MEMLIST,
|
| 1591 | FILE_CPU_EXCLUSIVE,
|
| 1592 | FILE_MEM_EXCLUSIVE,
|
| 1593 | FILE_MEM_HARDWALL,
|
| 1594 | FILE_SCHED_LOAD_BALANCE,
|
| 1595 | FILE_SCHED_RELAX_DOMAIN_LEVEL,
|
| 1596 | FILE_MEMORY_PRESSURE_ENABLED,
|
| 1597 | FILE_MEMORY_PRESSURE,
|
| 1598 | FILE_SPREAD_PAGE,
|
| 1599 | FILE_SPREAD_SLAB,
|
| 1600 | } cpuset_filetype_t;
|
| 1601 |
|
| 1602 | static int cpuset_write_u64(struct cgroup_subsys_state *css, struct cftype *cft,
|
| 1603 | u64 val)
|
| 1604 | {
|
| 1605 | struct cpuset *cs = css_cs(css);
|
| 1606 | cpuset_filetype_t type = cft->private;
|
| 1607 | int retval = 0;
|
| 1608 |
|
| 1609 | mutex_lock(&cpuset_mutex);
|
| 1610 | if (!is_cpuset_online(cs)) {
|
| 1611 | retval = -ENODEV;
|
| 1612 | goto out_unlock;
|
| 1613 | }
|
| 1614 |
|
| 1615 | switch (type) {
|
| 1616 | case FILE_CPU_EXCLUSIVE:
|
| 1617 | retval = update_flag(CS_CPU_EXCLUSIVE, cs, val);
|
| 1618 | break;
|
| 1619 | case FILE_MEM_EXCLUSIVE:
|
| 1620 | retval = update_flag(CS_MEM_EXCLUSIVE, cs, val);
|
| 1621 | break;
|
| 1622 | case FILE_MEM_HARDWALL:
|
| 1623 | retval = update_flag(CS_MEM_HARDWALL, cs, val);
|
| 1624 | break;
|
| 1625 | case FILE_SCHED_LOAD_BALANCE:
|
| 1626 | retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, val);
|
| 1627 | break;
|
| 1628 | case FILE_MEMORY_MIGRATE:
|
| 1629 | retval = update_flag(CS_MEMORY_MIGRATE, cs, val);
|
| 1630 | break;
|
| 1631 | case FILE_MEMORY_PRESSURE_ENABLED:
|
| 1632 | cpuset_memory_pressure_enabled = !!val;
|
| 1633 | break;
|
| 1634 | case FILE_MEMORY_PRESSURE:
|
| 1635 | retval = -EACCES;
|
| 1636 | break;
|
| 1637 | case FILE_SPREAD_PAGE:
|
| 1638 | retval = update_flag(CS_SPREAD_PAGE, cs, val);
|
| 1639 | break;
|
| 1640 | case FILE_SPREAD_SLAB:
|
| 1641 | retval = update_flag(CS_SPREAD_SLAB, cs, val);
|
| 1642 | break;
|
| 1643 | default:
|
| 1644 | retval = -EINVAL;
|
| 1645 | break;
|
| 1646 | }
|
| 1647 | out_unlock:
|
| 1648 | mutex_unlock(&cpuset_mutex);
|
| 1649 | return retval;
|
| 1650 | }
|
| 1651 |
|
| 1652 | static int cpuset_write_s64(struct cgroup_subsys_state *css, struct cftype *cft,
|
| 1653 | s64 val)
|
| 1654 | {
|
| 1655 | struct cpuset *cs = css_cs(css);
|
| 1656 | cpuset_filetype_t type = cft->private;
|
| 1657 | int retval = -ENODEV;
|
| 1658 |
|
| 1659 | mutex_lock(&cpuset_mutex);
|
| 1660 | if (!is_cpuset_online(cs))
|
| 1661 | goto out_unlock;
|
| 1662 |
|
| 1663 | switch (type) {
|
| 1664 | case FILE_SCHED_RELAX_DOMAIN_LEVEL:
|
| 1665 | retval = update_relax_domain_level(cs, val);
|
| 1666 | break;
|
| 1667 | default:
|
| 1668 | retval = -EINVAL;
|
| 1669 | break;
|
| 1670 | }
|
| 1671 | out_unlock:
|
| 1672 | mutex_unlock(&cpuset_mutex);
|
| 1673 | return retval;
|
| 1674 | }
|
| 1675 |
|
| 1676 | |
| 1677 | * Common handling for a write to a "cpus" or "mems" file.
|
| 1678 |
|
| 1679 | static int cpuset_write_resmask(struct cgroup_subsys_state *css,
|
| 1680 | struct cftype *cft, const char *buf)
|
| 1681 | {
|
| 1682 | struct cpuset *cs = css_cs(css);
|
| 1683 | struct cpuset *trialcs;
|
| 1684 | int retval = -ENODEV;
|
| 1685 |
|
| 1686 | |
| 1687 | * CPU or memory hotunplug may leave @cs w/o any execution
|
| 1688 | * resources, in which case the hotplug code asynchronously updates
|
| 1689 | * configuration and transfers all tasks to the nearest ancestor
|
| 1690 | * which can execute.
|
| 1691 | *
|
| 1692 | * As writes to "cpus" or "mems" may restore @cs's execution
|
| 1693 | * resources, wait for the previously scheduled operations before
|
| 1694 | * proceeding, so that we don't end up keep removing tasks added
|
| 1695 | * after execution capability is restored.
|
| 1696 |
|
| 1697 | flush_work(&cpuset_hotplug_work);
|
| 1698 |
|
| 1699 | mutex_lock(&cpuset_mutex);
|
| 1700 | if (!is_cpuset_online(cs))
|
| 1701 | goto out_unlock;
|
| 1702 |
|
| 1703 | trialcs = alloc_trial_cpuset(cs);
|
| 1704 | if (!trialcs) {
|
| 1705 | retval = -ENOMEM;
|
| 1706 | goto out_unlock;
|
| 1707 | }
|
| 1708 |
|
| 1709 | switch (cft->private) {
|
| 1710 | case FILE_CPULIST:
|
| 1711 | retval = update_cpumask(cs, trialcs, buf);
|
| 1712 | break;
|
| 1713 | case FILE_MEMLIST:
|
| 1714 | retval = update_nodemask(cs, trialcs, buf);
|
| 1715 | break;
|
| 1716 | default:
|
| 1717 | retval = -EINVAL;
|
| 1718 | break;
|
| 1719 | }
|
| 1720 |
|
| 1721 | free_trial_cpuset(trialcs);
|
| 1722 | out_unlock:
|
| 1723 | mutex_unlock(&cpuset_mutex);
|
| 1724 | return retval;
|
| 1725 | }
|
| 1726 |
|
| 1727 | |
| 1728 | * These ascii lists should be read in a single call, by using a user
|
| 1729 | * buffer large enough to hold the entire map. If read in smaller
|
| 1730 | * chunks, there is no guarantee of atomicity. Since the display format
|
| 1731 | * used, list of ranges of sequential numbers, is variable length,
|
| 1732 | * and since these maps can change value dynamically, one could read
|
| 1733 | * gibberish by doing partial reads while a list was changing.
|
| 1734 | * A single large read to a buffer that crosses a page boundary is
|
| 1735 | * ok, because the result being copied to user land is not recomputed
|
| 1736 | * across a page fault.
|
| 1737 |
|
| 1738 |
|
| 1739 | static size_t cpuset_sprintf_cpulist(char *page, struct cpuset *cs)
|
| 1740 | {
|
| 1741 | size_t count;
|
| 1742 |
|
| 1743 | mutex_lock(&callback_mutex);
|
| 1744 | count = cpulist_scnprintf(page, PAGE_SIZE, cs->cpus_allowed);
|
| 1745 | mutex_unlock(&callback_mutex);
|
| 1746 |
|
| 1747 | return count;
|
| 1748 | }
|
| 1749 |
|
| 1750 | static size_t cpuset_sprintf_memlist(char *page, struct cpuset *cs)
|
| 1751 | {
|
| 1752 | size_t count;
|
| 1753 |
|
| 1754 | mutex_lock(&callback_mutex);
|
| 1755 | count = nodelist_scnprintf(page, PAGE_SIZE, cs->mems_allowed);
|
| 1756 | mutex_unlock(&callback_mutex);
|
| 1757 |
|
| 1758 | return count;
|
| 1759 | }
|
| 1760 |
|
| 1761 | static ssize_t cpuset_common_file_read(struct cgroup_subsys_state *css,
|
| 1762 | struct cftype *cft, struct file *file,
|
| 1763 | char __user *buf, size_t nbytes,
|
| 1764 | loff_t *ppos)
|
| 1765 | {
|
| 1766 | struct cpuset *cs = css_cs(css);
|
| 1767 | cpuset_filetype_t type = cft->private;
|
| 1768 | char *page;
|
| 1769 | ssize_t retval = 0;
|
| 1770 | char *s;
|
| 1771 |
|
| 1772 | if (!(page = (char *)__get_free_page(GFP_TEMPORARY)))
|
| 1773 | return -ENOMEM;
|
| 1774 |
|
| 1775 | s = page;
|
| 1776 |
|
| 1777 | switch (type) {
|
| 1778 | case FILE_CPULIST:
|
| 1779 | s += cpuset_sprintf_cpulist(s, cs);
|
| 1780 | break;
|
| 1781 | case FILE_MEMLIST:
|
| 1782 | s += cpuset_sprintf_memlist(s, cs);
|
| 1783 | break;
|
| 1784 | default:
|
| 1785 | retval = -EINVAL;
|
| 1786 | goto out;
|
| 1787 | }
|
| 1788 | *s++ = '\n';
|
| 1789 |
|
| 1790 | retval = simple_read_from_buffer(buf, nbytes, ppos, page, s - page);
|
| 1791 | out:
|
| 1792 | free_page((unsigned long)page);
|
| 1793 | return retval;
|
| 1794 | }
|
| 1795 |
|
| 1796 | static u64 cpuset_read_u64(struct cgroup_subsys_state *css, struct cftype *cft)
|
| 1797 | {
|
| 1798 | struct cpuset *cs = css_cs(css);
|
| 1799 | cpuset_filetype_t type = cft->private;
|
| 1800 | switch (type) {
|
| 1801 | case FILE_CPU_EXCLUSIVE:
|
| 1802 | return is_cpu_exclusive(cs);
|
| 1803 | case FILE_MEM_EXCLUSIVE:
|
| 1804 | return is_mem_exclusive(cs);
|
| 1805 | case FILE_MEM_HARDWALL:
|
| 1806 | return is_mem_hardwall(cs);
|
| 1807 | case FILE_SCHED_LOAD_BALANCE:
|
| 1808 | return is_sched_load_balance(cs);
|
| 1809 | case FILE_MEMORY_MIGRATE:
|
| 1810 | return is_memory_migrate(cs);
|
| 1811 | case FILE_MEMORY_PRESSURE_ENABLED:
|
| 1812 | return cpuset_memory_pressure_enabled;
|
| 1813 | case FILE_MEMORY_PRESSURE:
|
| 1814 | return fmeter_getrate(&cs->fmeter);
|
| 1815 | case FILE_SPREAD_PAGE:
|
| 1816 | return is_spread_page(cs);
|
| 1817 | case FILE_SPREAD_SLAB:
|
| 1818 | return is_spread_slab(cs);
|
| 1819 | default:
|
| 1820 | BUG();
|
| 1821 | }
|
| 1822 |
|
| 1823 |
|
| 1824 | return 0;
|
| 1825 | }
|
| 1826 |
|
| 1827 | static s64 cpuset_read_s64(struct cgroup_subsys_state *css, struct cftype *cft)
|
| 1828 | {
|
| 1829 | struct cpuset *cs = css_cs(css);
|
| 1830 | cpuset_filetype_t type = cft->private;
|
| 1831 | switch (type) {
|
| 1832 | case FILE_SCHED_RELAX_DOMAIN_LEVEL:
|
| 1833 | return cs->relax_domain_level;
|
| 1834 | default:
|
| 1835 | BUG();
|
| 1836 | }
|
| 1837 |
|
| 1838 |
|
| 1839 | return 0;
|
| 1840 | }
|
| 1841 |
|
| 1842 |
|
| 1843 | |
| 1844 | * for the common functions, 'private' gives the type of file
|
| 1845 |
|
| 1846 |
|
| 1847 | static struct cftype files[] = {
|
| 1848 | {
|
| 1849 | .name = "cpus",
|
| 1850 | .read = cpuset_common_file_read,
|
| 1851 | .write_string = cpuset_write_resmask,
|
| 1852 | .max_write_len = (100U + 6 * NR_CPUS),
|
| 1853 | .private = FILE_CPULIST,
|
| 1854 | },
|
| 1855 |
|
| 1856 | {
|
| 1857 | .name = "mems",
|
| 1858 | .read = cpuset_common_file_read,
|
| 1859 | .write_string = cpuset_write_resmask,
|
| 1860 | .max_write_len = (100U + 6 * MAX_NUMNODES),
|
| 1861 | .private = FILE_MEMLIST,
|
| 1862 | },
|
| 1863 |
|
| 1864 | {
|
| 1865 | .name = "cpu_exclusive",
|
| 1866 | .read_u64 = cpuset_read_u64,
|
| 1867 | .write_u64 = cpuset_write_u64,
|
| 1868 | .private = FILE_CPU_EXCLUSIVE,
|
| 1869 | },
|
| 1870 |
|
| 1871 | {
|
| 1872 | .name = "mem_exclusive",
|
| 1873 | .read_u64 = cpuset_read_u64,
|
| 1874 | .write_u64 = cpuset_write_u64,
|
| 1875 | .private = FILE_MEM_EXCLUSIVE,
|
| 1876 | },
|
| 1877 |
|
| 1878 | {
|
| 1879 | .name = "mem_hardwall",
|
| 1880 | .read_u64 = cpuset_read_u64,
|
| 1881 | .write_u64 = cpuset_write_u64,
|
| 1882 | .private = FILE_MEM_HARDWALL,
|
| 1883 | },
|
| 1884 |
|
| 1885 | {
|
| 1886 | .name = "sched_load_balance",
|
| 1887 | .read_u64 = cpuset_read_u64,
|
| 1888 | .write_u64 = cpuset_write_u64,
|
| 1889 | .private = FILE_SCHED_LOAD_BALANCE,
|
| 1890 | },
|
| 1891 |
|
| 1892 | {
|
| 1893 | .name = "sched_relax_domain_level",
|
| 1894 | .read_s64 = cpuset_read_s64,
|
| 1895 | .write_s64 = cpuset_write_s64,
|
| 1896 | .private = FILE_SCHED_RELAX_DOMAIN_LEVEL,
|
| 1897 | },
|
| 1898 |
|
| 1899 | {
|
| 1900 | .name = "memory_migrate",
|
| 1901 | .read_u64 = cpuset_read_u64,
|
| 1902 | .write_u64 = cpuset_write_u64,
|
| 1903 | .private = FILE_MEMORY_MIGRATE,
|
| 1904 | },
|
| 1905 |
|
| 1906 | {
|
| 1907 | .name = "memory_pressure",
|
| 1908 | .read_u64 = cpuset_read_u64,
|
| 1909 | .write_u64 = cpuset_write_u64,
|
| 1910 | .private = FILE_MEMORY_PRESSURE,
|
| 1911 | .mode = S_IRUGO,
|
| 1912 | },
|
| 1913 |
|
| 1914 | {
|
| 1915 | .name = "memory_spread_page",
|
| 1916 | .read_u64 = cpuset_read_u64,
|
| 1917 | .write_u64 = cpuset_write_u64,
|
| 1918 | .private = FILE_SPREAD_PAGE,
|
| 1919 | },
|
| 1920 |
|
| 1921 | {
|
| 1922 | .name = "memory_spread_slab",
|
| 1923 | .read_u64 = cpuset_read_u64,
|
| 1924 | .write_u64 = cpuset_write_u64,
|
| 1925 | .private = FILE_SPREAD_SLAB,
|
| 1926 | },
|
| 1927 |
|
| 1928 | {
|
| 1929 | .name = "memory_pressure_enabled",
|
| 1930 | .flags = CFTYPE_ONLY_ON_ROOT,
|
| 1931 | .read_u64 = cpuset_read_u64,
|
| 1932 | .write_u64 = cpuset_write_u64,
|
| 1933 | .private = FILE_MEMORY_PRESSURE_ENABLED,
|
| 1934 | },
|
| 1935 |
|
| 1936 | { }
|
| 1937 | };
|
| 1938 |
|
| 1939 | |
| 1940 | * cpuset_css_alloc - allocate a cpuset css
|
| 1941 | * cgrp: control group that the new cpuset will be part of
|
| 1942 |
|
| 1943 |
|
| 1944 | static struct cgroup_subsys_state *
|
| 1945 | cpuset_css_alloc(struct cgroup_subsys_state *parent_css)
|
| 1946 | {
|
| 1947 | struct cpuset *cs;
|
| 1948 |
|
| 1949 | if (!parent_css)
|
| 1950 | return &top_cpuset.css;
|
| 1951 |
|
| 1952 | cs = kzalloc(sizeof(*cs), GFP_KERNEL);
|
| 1953 | if (!cs)
|
| 1954 | return ERR_PTR(-ENOMEM);
|
| 1955 | if (!alloc_cpumask_var(&cs->cpus_allowed, GFP_KERNEL)) {
|
| 1956 | kfree(cs);
|
| 1957 | return ERR_PTR(-ENOMEM);
|
| 1958 | }
|
| 1959 |
|
| 1960 | set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
|
| 1961 | cpumask_clear(cs->cpus_allowed);
|
| 1962 | nodes_clear(cs->mems_allowed);
|
| 1963 | fmeter_init(&cs->fmeter);
|
| 1964 | cs->relax_domain_level = -1;
|
| 1965 |
|
| 1966 | return &cs->css;
|
| 1967 | }
|
| 1968 |
|
| 1969 | static int cpuset_css_online(struct cgroup_subsys_state *css)
|
| 1970 | {
|
| 1971 | struct cpuset *cs = css_cs(css);
|
| 1972 | struct cpuset *parent = parent_cs(cs);
|
| 1973 | struct cpuset *tmp_cs;
|
| 1974 | struct cgroup_subsys_state *pos_css;
|
| 1975 |
|
| 1976 | if (!parent)
|
| 1977 | return 0;
|
| 1978 |
|
| 1979 | mutex_lock(&cpuset_mutex);
|
| 1980 |
|
| 1981 | set_bit(CS_ONLINE, &cs->flags);
|
| 1982 | if (is_spread_page(parent))
|
| 1983 | set_bit(CS_SPREAD_PAGE, &cs->flags);
|
| 1984 | if (is_spread_slab(parent))
|
| 1985 | set_bit(CS_SPREAD_SLAB, &cs->flags);
|
| 1986 |
|
| 1987 | number_of_cpusets++;
|
| 1988 |
|
| 1989 | if (!test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags))
|
| 1990 | goto out_unlock;
|
| 1991 |
|
| 1992 | |
| 1993 | * Clone @parent's configuration if CGRP_CPUSET_CLONE_CHILDREN is
|
| 1994 | * set. This flag handling is implemented in cgroup core for
|
| 1995 | * histrical reasons - the flag may be specified during mount.
|
| 1996 | *
|
| 1997 | * Currently, if any sibling cpusets have exclusive cpus or mem, we
|
| 1998 | * refuse to clone the configuration - thereby refusing the task to
|
| 1999 | * be entered, and as a result refusing the sys_unshare() or
|
| 2000 | * clone() which initiated it. If this becomes a problem for some
|
| 2001 | * users who wish to allow that scenario, then this could be
|
| 2002 | * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
|
| 2003 | * (and likewise for mems) to the new cgroup.
|
| 2004 |
|
| 2005 | rcu_read_lock();
|
| 2006 | cpuset_for_each_child(tmp_cs, pos_css, parent) {
|
| 2007 | if (is_mem_exclusive(tmp_cs) || is_cpu_exclusive(tmp_cs)) {
|
| 2008 | rcu_read_unlock();
|
| 2009 | goto out_unlock;
|
| 2010 | }
|
| 2011 | }
|
| 2012 | rcu_read_unlock();
|
| 2013 |
|
| 2014 | mutex_lock(&callback_mutex);
|
| 2015 | cs->mems_allowed = parent->mems_allowed;
|
| 2016 | cpumask_copy(cs->cpus_allowed, parent->cpus_allowed);
|
| 2017 | mutex_unlock(&callback_mutex);
|
| 2018 | out_unlock:
|
| 2019 | mutex_unlock(&cpuset_mutex);
|
| 2020 | return 0;
|
| 2021 | }
|
| 2022 |
|
| 2023 | |
| 2024 | * If the cpuset being removed has its flag 'sched_load_balance'
|
| 2025 | * enabled, then simulate turning sched_load_balance off, which
|
| 2026 | * will call rebuild_sched_domains_locked().
|
| 2027 |
|
| 2028 |
|
| 2029 | static void cpuset_css_offline(struct cgroup_subsys_state *css)
|
| 2030 | {
|
| 2031 | struct cpuset *cs = css_cs(css);
|
| 2032 |
|
| 2033 | mutex_lock(&cpuset_mutex);
|
| 2034 |
|
| 2035 | if (is_sched_load_balance(cs))
|
| 2036 | update_flag(CS_SCHED_LOAD_BALANCE, cs, 0);
|
| 2037 |
|
| 2038 | number_of_cpusets--;
|
| 2039 | clear_bit(CS_ONLINE, &cs->flags);
|
| 2040 |
|
| 2041 | mutex_unlock(&cpuset_mutex);
|
| 2042 | }
|
| 2043 |
|
| 2044 | static void cpuset_css_free(struct cgroup_subsys_state *css)
|
| 2045 | {
|
| 2046 | struct cpuset *cs = css_cs(css);
|
| 2047 |
|
| 2048 | free_cpumask_var(cs->cpus_allowed);
|
| 2049 | kfree(cs);
|
| 2050 | }
|
| 2051 |
|
| 2052 | struct cgroup_subsys cpuset_subsys = {
|
| 2053 | .name = "cpuset",
|
| 2054 | .css_alloc = cpuset_css_alloc,
|
| 2055 | .css_online = cpuset_css_online,
|
| 2056 | .css_offline = cpuset_css_offline,
|
| 2057 | .css_free = cpuset_css_free,
|
| 2058 | .can_attach = cpuset_can_attach,
|
| 2059 | .cancel_attach = cpuset_cancel_attach,
|
| 2060 | .attach = cpuset_attach,
|
| 2061 | .subsys_id = cpuset_subsys_id,
|
| 2062 | .base_cftypes = files,
|
| 2063 | .early_init = 1,
|
| 2064 | };
|
| 2065 |
|
| 2066 | |
| 2067 | * cpuset_init - initialize cpusets at system boot
|
| 2068 | *
|
| 2069 | * Description: Initialize top_cpuset and the cpuset internal file system,
|
| 2070 | *
|
| 2071 |
|
| 2072 | int __init cpuset_init(void)
|
| 2073 | {
|
| 2074 | int err = 0;
|
| 2075 |
|
| 2076 | if (!alloc_cpumask_var(&top_cpuset.cpus_allowed, GFP_KERNEL))
|
| 2077 | BUG();
|
| 2078 |
|
| 2079 | cpumask_setall(top_cpuset.cpus_allowed);
|
| 2080 | nodes_setall(top_cpuset.mems_allowed);
|
| 2081 |
|
| 2082 | fmeter_init(&top_cpuset.fmeter);
|
| 2083 | set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags);
|
| 2084 | top_cpuset.relax_domain_level = -1;
|
| 2085 |
|
| 2086 | err = register_filesystem(&cpuset_fs_type);
|
| 2087 | if (err < 0)
|
| 2088 | return err;
|
| 2089 |
|
| 2090 | if (!alloc_cpumask_var(&cpus_attach, GFP_KERNEL))
|
| 2091 | BUG();
|
| 2092 |
|
| 2093 | number_of_cpusets = 1;
|
| 2094 | return 0;
|
| 2095 | }
|
| 2096 |
|
| 2097 | |
| 2098 | * If CPU and/or memory hotplug handlers, below, unplug any CPUs
|
| 2099 | * or memory nodes, we need to walk over the cpuset hierarchy,
|
| 2100 | * removing that CPU or node from all cpusets. If this removes the
|
| 2101 | * last CPU or node from a cpuset, then move the tasks in the empty
|
| 2102 | * cpuset to its next-highest non-empty parent.
|
| 2103 |
|
| 2104 | static void remove_tasks_in_empty_cpuset(struct cpuset *cs)
|
| 2105 | {
|
| 2106 | struct cpuset *parent;
|
| 2107 |
|
| 2108 | |
| 2109 | * Find its next-highest non-empty parent, (top cpuset
|
| 2110 | * has online cpus, so can't be empty).
|
| 2111 |
|
| 2112 | parent = parent_cs(cs);
|
| 2113 | while (cpumask_empty(parent->cpus_allowed) ||
|
| 2114 | nodes_empty(parent->mems_allowed))
|
| 2115 | parent = parent_cs(parent);
|
| 2116 |
|
| 2117 | if (cgroup_transfer_tasks(parent->css.cgroup, cs->css.cgroup)) {
|
| 2118 | rcu_read_lock();
|
| 2119 | printk(KERN_ERR "cpuset: failed to transfer tasks out of empty cpuset %s\n",
|
| 2120 | cgroup_name(cs->css.cgroup));
|
| 2121 | rcu_read_unlock();
|
| 2122 | }
|
| 2123 | }
|
| 2124 |
|
| 2125 | |
| 2126 | * cpuset_hotplug_update_tasks - update tasks in a cpuset for hotunplug
|
| 2127 | * @cs: cpuset in interest
|
| 2128 | *
|
| 2129 | * Compare @cs's cpu and mem masks against top_cpuset and if some have gone
|
| 2130 | * offline, update @cs accordingly. If @cs ends up with no CPU or memory,
|
| 2131 | * all its tasks are moved to the nearest ancestor with both resources.
|
| 2132 |
|
| 2133 | static void cpuset_hotplug_update_tasks(struct cpuset *cs)
|
| 2134 | {
|
| 2135 | static cpumask_t off_cpus;
|
| 2136 | static nodemask_t off_mems;
|
| 2137 | bool is_empty;
|
| 2138 | bool sane = cgroup_sane_behavior(cs->css.cgroup);
|
| 2139 |
|
| 2140 | retry:
|
| 2141 | wait_event(cpuset_attach_wq, cs->attach_in_progress == 0);
|
| 2142 |
|
| 2143 | mutex_lock(&cpuset_mutex);
|
| 2144 |
|
| 2145 | |
| 2146 | * We have raced with task attaching. We wait until attaching
|
| 2147 | * is finished, so we won't attach a task to an empty cpuset.
|
| 2148 |
|
| 2149 | if (cs->attach_in_progress) {
|
| 2150 | mutex_unlock(&cpuset_mutex);
|
| 2151 | goto retry;
|
| 2152 | }
|
| 2153 |
|
| 2154 | cpumask_andnot(&off_cpus, cs->cpus_allowed, top_cpuset.cpus_allowed);
|
| 2155 | nodes_andnot(off_mems, cs->mems_allowed, top_cpuset.mems_allowed);
|
| 2156 |
|
| 2157 | mutex_lock(&callback_mutex);
|
| 2158 | cpumask_andnot(cs->cpus_allowed, cs->cpus_allowed, &off_cpus);
|
| 2159 | mutex_unlock(&callback_mutex);
|
| 2160 |
|
| 2161 | |
| 2162 | * If sane_behavior flag is set, we need to update tasks' cpumask
|
| 2163 | * for empty cpuset to take on ancestor's cpumask. Otherwise, don't
|
| 2164 | * call update_tasks_cpumask() if the cpuset becomes empty, as
|
| 2165 | * the tasks in it will be migrated to an ancestor.
|
| 2166 |
|
| 2167 | if ((sane && cpumask_empty(cs->cpus_allowed)) ||
|
| 2168 | (!cpumask_empty(&off_cpus) && !cpumask_empty(cs->cpus_allowed)))
|
| 2169 | update_tasks_cpumask(cs, NULL);
|
| 2170 |
|
| 2171 | mutex_lock(&callback_mutex);
|
| 2172 | nodes_andnot(cs->mems_allowed, cs->mems_allowed, off_mems);
|
| 2173 | mutex_unlock(&callback_mutex);
|
| 2174 |
|
| 2175 | |
| 2176 | * If sane_behavior flag is set, we need to update tasks' nodemask
|
| 2177 | * for empty cpuset to take on ancestor's nodemask. Otherwise, don't
|
| 2178 | * call update_tasks_nodemask() if the cpuset becomes empty, as
|
| 2179 | * the tasks in it will be migratd to an ancestor.
|
| 2180 |
|
| 2181 | if ((sane && nodes_empty(cs->mems_allowed)) ||
|
| 2182 | (!nodes_empty(off_mems) && !nodes_empty(cs->mems_allowed)))
|
| 2183 | update_tasks_nodemask(cs, NULL);
|
| 2184 |
|
| 2185 | is_empty = cpumask_empty(cs->cpus_allowed) ||
|
| 2186 | nodes_empty(cs->mems_allowed);
|
| 2187 |
|
| 2188 | mutex_unlock(&cpuset_mutex);
|
| 2189 |
|
| 2190 | |
| 2191 | * If sane_behavior flag is set, we'll keep tasks in empty cpusets.
|
| 2192 | *
|
| 2193 | * Otherwise move tasks to the nearest ancestor with execution
|
| 2194 | * resources. This is full cgroup operation which will
|
| 2195 | * also call back into cpuset. Should be done outside any lock.
|
| 2196 |
|
| 2197 | if (!sane && is_empty)
|
| 2198 | remove_tasks_in_empty_cpuset(cs);
|
| 2199 | }
|
| 2200 |
|
| 2201 | |
| 2202 | * cpuset_hotplug_workfn - handle CPU/memory hotunplug for a cpuset
|
| 2203 | *
|
| 2204 | * This function is called after either CPU or memory configuration has
|
| 2205 | * changed and updates cpuset accordingly. The top_cpuset is always
|
| 2206 | * synchronized to cpu_active_mask and N_MEMORY, which is necessary in
|
| 2207 | * order to make cpusets transparent (of no affect) on systems that are
|
| 2208 | * actively using CPU hotplug but making no active use of cpusets.
|
| 2209 | *
|
| 2210 | * Non-root cpusets are only affected by offlining. If any CPUs or memory
|
| 2211 | * nodes have been taken down, cpuset_hotplug_update_tasks() is invoked on
|
| 2212 | * all descendants.
|
| 2213 | *
|
| 2214 | * Note that CPU offlining during suspend is ignored. We don't modify
|
| 2215 | * cpusets across suspend/resume cycles at all.
|
| 2216 |
|
| 2217 | static void cpuset_hotplug_workfn(struct work_struct *work)
|
| 2218 | {
|
| 2219 | static cpumask_t new_cpus;
|
| 2220 | static nodemask_t new_mems;
|
| 2221 | bool cpus_updated, mems_updated;
|
| 2222 |
|
| 2223 | mutex_lock(&cpuset_mutex);
|
| 2224 |
|
| 2225 |
|
| 2226 | cpumask_copy(&new_cpus, cpu_active_mask);
|
| 2227 | new_mems = node_states[N_MEMORY];
|
| 2228 |
|
| 2229 | cpus_updated = !cpumask_equal(top_cpuset.cpus_allowed, &new_cpus);
|
| 2230 | mems_updated = !nodes_equal(top_cpuset.mems_allowed, new_mems);
|
| 2231 |
|
| 2232 |
|
| 2233 | if (cpus_updated) {
|
| 2234 | mutex_lock(&callback_mutex);
|
| 2235 | cpumask_copy(top_cpuset.cpus_allowed, &new_cpus);
|
| 2236 | mutex_unlock(&callback_mutex);
|
| 2237 |
|
| 2238 | }
|
| 2239 |
|
| 2240 |
|
| 2241 | if (mems_updated) {
|
| 2242 | mutex_lock(&callback_mutex);
|
| 2243 | top_cpuset.mems_allowed = new_mems;
|
| 2244 | mutex_unlock(&callback_mutex);
|
| 2245 | update_tasks_nodemask(&top_cpuset, NULL);
|
| 2246 | }
|
| 2247 |
|
| 2248 | mutex_unlock(&cpuset_mutex);
|
| 2249 |
|
| 2250 |
|
| 2251 | if (cpus_updated || mems_updated) {
|
| 2252 | struct cpuset *cs;
|
| 2253 | struct cgroup_subsys_state *pos_css;
|
| 2254 |
|
| 2255 | rcu_read_lock();
|
| 2256 | cpuset_for_each_descendant_pre(cs, pos_css, &top_cpuset) {
|
| 2257 | if (cs == &top_cpuset || !css_tryget(&cs->css))
|
| 2258 | continue;
|
| 2259 | rcu_read_unlock();
|
| 2260 |
|
| 2261 | cpuset_hotplug_update_tasks(cs);
|
| 2262 |
|
| 2263 | rcu_read_lock();
|
| 2264 | css_put(&cs->css);
|
| 2265 | }
|
| 2266 | rcu_read_unlock();
|
| 2267 | }
|
| 2268 |
|
| 2269 |
|
| 2270 | if (cpus_updated)
|
| 2271 | rebuild_sched_domains();
|
| 2272 | }
|
| 2273 |
|
| 2274 | void cpuset_update_active_cpus(bool cpu_online)
|
| 2275 | {
|
| 2276 | |
| 2277 | * We're inside cpu hotplug critical region which usually nests
|
| 2278 | * inside cgroup synchronization. Bounce actual hotplug processing
|
| 2279 | * to a work item to avoid reverse locking order.
|
| 2280 | *
|
| 2281 | * We still need to do partition_sched_domains() synchronously;
|
| 2282 | * otherwise, the scheduler will get confused and put tasks to the
|
| 2283 | * dead CPU. Fall back to the default single domain.
|
| 2284 | * cpuset_hotplug_workfn() will rebuild it as necessary.
|
| 2285 |
|
| 2286 | partition_sched_domains(1, NULL, NULL);
|
| 2287 | schedule_work(&cpuset_hotplug_work);
|
| 2288 | }
|
| 2289 |
|
| 2290 | |
| 2291 | * Keep top_cpuset.mems_allowed tracking node_states[N_MEMORY].
|
| 2292 | * Call this routine anytime after node_states[N_MEMORY] changes.
|
| 2293 | * See cpuset_update_active_cpus() for CPU hotplug handling.
|
| 2294 |
|
| 2295 | static int cpuset_track_online_nodes(struct notifier_block *self,
|
| 2296 | unsigned long action, void *arg)
|
| 2297 | {
|
| 2298 | schedule_work(&cpuset_hotplug_work);
|
| 2299 | return NOTIFY_OK;
|
| 2300 | }
|
| 2301 |
|
| 2302 | static struct notifier_block cpuset_track_online_nodes_nb = {
|
| 2303 | .notifier_call = cpuset_track_online_nodes,
|
| 2304 | .priority = 10,
|
| 2305 | };
|
| 2306 |
|
| 2307 | |
| 2308 | * cpuset_init_smp - initialize cpus_allowed
|
| 2309 | *
|
| 2310 | * Description: Finish top cpuset after cpu, node maps are initialized
|
| 2311 |
|
| 2312 | void __init cpuset_init_smp(void)
|
| 2313 | {
|
| 2314 | cpumask_copy(top_cpuset.cpus_allowed, cpu_active_mask);
|
| 2315 | top_cpuset.mems_allowed = node_states[N_MEMORY];
|
| 2316 | top_cpuset.old_mems_allowed = top_cpuset.mems_allowed;
|
| 2317 |
|
| 2318 | register_hotmemory_notifier(&cpuset_track_online_nodes_nb);
|
| 2319 | }
|
| 2320 |
|
| 2321 | |
| 2322 | * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
|
| 2323 | * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
|
| 2324 | * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
|
| 2325 | *
|
| 2326 | * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
|
| 2327 | * attached to the specified @tsk. Guaranteed to return some non-empty
|
| 2328 | * subset of cpu_online_mask, even if this means going outside the
|
| 2329 | * tasks cpuset.
|
| 2330 | *
|
| 2331 |
|
| 2332 | void cpuset_cpus_allowed(struct task_struct *tsk, struct cpumask *pmask)
|
| 2333 | {
|
| 2334 | struct cpuset *cpus_cs;
|
| 2335 |
|
| 2336 | mutex_lock(&callback_mutex);
|
| 2337 | task_lock(tsk);
|
| 2338 | cpus_cs = effective_cpumask_cpuset(task_cs(tsk));
|
| 2339 | guarantee_online_cpus(cpus_cs, pmask);
|
| 2340 | task_unlock(tsk);
|
| 2341 | mutex_unlock(&callback_mutex);
|
| 2342 | }
|
| 2343 |
|
| 2344 | void cpuset_cpus_allowed_fallback(struct task_struct *tsk)
|
| 2345 | {
|
| 2346 | struct cpuset *cpus_cs;
|
| 2347 |
|
| 2348 | rcu_read_lock();
|
| 2349 | cpus_cs = effective_cpumask_cpuset(task_cs(tsk));
|
| 2350 | do_set_cpus_allowed(tsk, cpus_cs->cpus_allowed);
|
| 2351 | rcu_read_unlock();
|
| 2352 |
|
| 2353 | |
| 2354 | * We own tsk->cpus_allowed, nobody can change it under us.
|
| 2355 | *
|
| 2356 | * But we used cs && cs->cpus_allowed lockless and thus can
|
| 2357 | * race with cgroup_attach_task() or update_cpumask() and get
|
| 2358 | * the wrong tsk->cpus_allowed. However, both cases imply the
|
| 2359 | * subsequent cpuset_change_cpumask()->set_cpus_allowed_ptr()
|
| 2360 | * which takes task_rq_lock().
|
| 2361 | *
|
| 2362 | * If we are called after it dropped the lock we must see all
|
| 2363 | * changes in tsk_cs()->cpus_allowed. Otherwise we can temporary
|
| 2364 | * set any mask even if it is not right from task_cs() pov,
|
| 2365 | * the pending set_cpus_allowed_ptr() will fix things.
|
| 2366 | *
|
| 2367 | * select_fallback_rq() will fix things ups and set cpu_possible_mask
|
| 2368 | * if required.
|
| 2369 |
|
| 2370 | }
|
| 2371 |
|
| 2372 | void cpuset_init_current_mems_allowed(void)
|
| 2373 | {
|
| 2374 | nodes_setall(current->mems_allowed);
|
| 2375 | }
|
| 2376 |
|
| 2377 | |
| 2378 | * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
|
| 2379 | * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
|
| 2380 | *
|
| 2381 | * Description: Returns the nodemask_t mems_allowed of the cpuset
|
| 2382 | * attached to the specified @tsk. Guaranteed to return some non-empty
|
| 2383 | * subset of node_states[N_MEMORY], even if this means going outside the
|
| 2384 | * tasks cpuset.
|
| 2385 | *
|
| 2386 |
|
| 2387 | nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
|
| 2388 | {
|
| 2389 | struct cpuset *mems_cs;
|
| 2390 | nodemask_t mask;
|
| 2391 |
|
| 2392 | mutex_lock(&callback_mutex);
|
| 2393 | task_lock(tsk);
|
| 2394 | mems_cs = effective_nodemask_cpuset(task_cs(tsk));
|
| 2395 | guarantee_online_mems(mems_cs, &mask);
|
| 2396 | task_unlock(tsk);
|
| 2397 | mutex_unlock(&callback_mutex);
|
| 2398 |
|
| 2399 | return mask;
|
| 2400 | }
|
| 2401 |
|
| 2402 | |
| 2403 | * cpuset_nodemask_valid_mems_allowed - check nodemask vs. curremt mems_allowed
|
| 2404 | * @nodemask: the nodemask to be checked
|
| 2405 | *
|
| 2406 | * Are any of the nodes in the nodemask allowed in current->mems_allowed?
|
| 2407 |
|
| 2408 | int cpuset_nodemask_valid_mems_allowed(nodemask_t *nodemask)
|
| 2409 | {
|
| 2410 | return nodes_intersects(*nodemask, current->mems_allowed);
|
| 2411 | }
|
| 2412 |
|
| 2413 | |
| 2414 | * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
|
| 2415 | * mem_hardwall ancestor to the specified cpuset. Call holding
|
| 2416 | * callback_mutex. If no ancestor is mem_exclusive or mem_hardwall
|
| 2417 | * (an unusual configuration), then returns the root cpuset.
|
| 2418 |
|
| 2419 | static struct cpuset *nearest_hardwall_ancestor(struct cpuset *cs)
|
| 2420 | {
|
| 2421 | while (!(is_mem_exclusive(cs) || is_mem_hardwall(cs)) && parent_cs(cs))
|
| 2422 | cs = parent_cs(cs);
|
| 2423 | return cs;
|
| 2424 | }
|
| 2425 |
|
| 2426 | |
| 2427 | * cpuset_node_allowed_softwall - Can we allocate on a memory node?
|
| 2428 | * @node: is this an allowed node?
|
| 2429 | * @gfp_mask: memory allocation flags
|
| 2430 | *
|
| 2431 | * If we're in interrupt, yes, we can always allocate. If __GFP_THISNODE is
|
| 2432 | * set, yes, we can always allocate. If node is in our task's mems_allowed,
|
| 2433 | * yes. If it's not a __GFP_HARDWALL request and this node is in the nearest
|
| 2434 | * hardwalled cpuset ancestor to this task's cpuset, yes. If the task has been
|
| 2435 | * OOM killed and has access to memory reserves as specified by the TIF_MEMDIE
|
| 2436 | * flag, yes.
|
| 2437 | * Otherwise, no.
|
| 2438 | *
|
| 2439 | * If __GFP_HARDWALL is set, cpuset_node_allowed_softwall() reduces to
|
| 2440 | * cpuset_node_allowed_hardwall(). Otherwise, cpuset_node_allowed_softwall()
|
| 2441 | * might sleep, and might allow a node from an enclosing cpuset.
|
| 2442 | *
|
| 2443 | * cpuset_node_allowed_hardwall() only handles the simpler case of hardwall
|
| 2444 | * cpusets, and never sleeps.
|
| 2445 | *
|
| 2446 | * The __GFP_THISNODE placement logic is really handled elsewhere,
|
| 2447 | * by forcibly using a zonelist starting at a specified node, and by
|
| 2448 | * (in get_page_from_freelist()) refusing to consider the zones for
|
| 2449 | * any node on the zonelist except the first. By the time any such
|
| 2450 | * calls get to this routine, we should just shut up and say 'yes'.
|
| 2451 | *
|
| 2452 | * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
|
| 2453 | * and do not allow allocations outside the current tasks cpuset
|
| 2454 | * unless the task has been OOM killed as is marked TIF_MEMDIE.
|
| 2455 | * GFP_KERNEL allocations are not so marked, so can escape to the
|
| 2456 | * nearest enclosing hardwalled ancestor cpuset.
|
| 2457 | *
|
| 2458 | * Scanning up parent cpusets requires callback_mutex. The
|
| 2459 | * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
|
| 2460 | * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
|
| 2461 | * current tasks mems_allowed came up empty on the first pass over
|
| 2462 | * the zonelist. So only GFP_KERNEL allocations, if all nodes in the
|
| 2463 | * cpuset are short of memory, might require taking the callback_mutex
|
| 2464 | * mutex.
|
| 2465 | *
|
| 2466 | * The first call here from mm/page_alloc:get_page_from_freelist()
|
| 2467 | * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
|
| 2468 | * so no allocation on a node outside the cpuset is allowed (unless
|
| 2469 | * in interrupt, of course).
|
| 2470 | *
|
| 2471 | * The second pass through get_page_from_freelist() doesn't even call
|
| 2472 | * here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
|
| 2473 | * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
|
| 2474 | * in alloc_flags. That logic and the checks below have the combined
|
| 2475 | * affect that:
|
| 2476 | * in_interrupt - any node ok (current task context irrelevant)
|
| 2477 | * GFP_ATOMIC - any node ok
|
| 2478 | * TIF_MEMDIE - any node ok
|
| 2479 | * GFP_KERNEL - any node in enclosing hardwalled cpuset ok
|
| 2480 | * GFP_USER - only nodes in current tasks mems allowed ok.
|
| 2481 | *
|
| 2482 | * Rule:
|
| 2483 | * Don't call cpuset_node_allowed_softwall if you can't sleep, unless you
|
| 2484 | * pass in the __GFP_HARDWALL flag set in gfp_flag, which disables
|
| 2485 | * the code that might scan up ancestor cpusets and sleep.
|
| 2486 |
|
| 2487 | int __cpuset_node_allowed_softwall(int node, gfp_t gfp_mask)
|
| 2488 | {
|
| 2489 | struct cpuset *cs;
|
| 2490 | int allowed;
|
| 2491 |
|
| 2492 | if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
|
| 2493 | return 1;
|
| 2494 | might_sleep_if(!(gfp_mask & __GFP_HARDWALL));
|
| 2495 | if (node_isset(node, current->mems_allowed))
|
| 2496 | return 1;
|
| 2497 | |
| 2498 | * Allow tasks that have access to memory reserves because they have
|
| 2499 | * been OOM killed to get memory anywhere.
|
| 2500 |
|
| 2501 | if (unlikely(test_thread_flag(TIF_MEMDIE)))
|
| 2502 | return 1;
|
| 2503 | if (gfp_mask & __GFP_HARDWALL)
|
| 2504 | return 0;
|
| 2505 |
|
| 2506 | if (current->flags & PF_EXITING)
|
| 2507 | return 1;
|
| 2508 |
|
| 2509 |
|
| 2510 | mutex_lock(&callback_mutex);
|
| 2511 |
|
| 2512 | task_lock(current);
|
| 2513 | cs = nearest_hardwall_ancestor(task_cs(current));
|
| 2514 | task_unlock(current);
|
| 2515 |
|
| 2516 | allowed = node_isset(node, cs->mems_allowed);
|
| 2517 | mutex_unlock(&callback_mutex);
|
| 2518 | return allowed;
|
| 2519 | }
|
| 2520 |
|
| 2521 | |
| 2522 | * cpuset_node_allowed_hardwall - Can we allocate on a memory node?
|
| 2523 | * @node: is this an allowed node?
|
| 2524 | * @gfp_mask: memory allocation flags
|
| 2525 | *
|
| 2526 | * If we're in interrupt, yes, we can always allocate. If __GFP_THISNODE is
|
| 2527 | * set, yes, we can always allocate. If node is in our task's mems_allowed,
|
| 2528 | * yes. If the task has been OOM killed and has access to memory reserves as
|
| 2529 | * specified by the TIF_MEMDIE flag, yes.
|
| 2530 | * Otherwise, no.
|
| 2531 | *
|
| 2532 | * The __GFP_THISNODE placement logic is really handled elsewhere,
|
| 2533 | * by forcibly using a zonelist starting at a specified node, and by
|
| 2534 | * (in get_page_from_freelist()) refusing to consider the zones for
|
| 2535 | * any node on the zonelist except the first. By the time any such
|
| 2536 | * calls get to this routine, we should just shut up and say 'yes'.
|
| 2537 | *
|
| 2538 | * Unlike the cpuset_node_allowed_softwall() variant, above,
|
| 2539 | * this variant requires that the node be in the current task's
|
| 2540 | * mems_allowed or that we're in interrupt. It does not scan up the
|
| 2541 | * cpuset hierarchy for the nearest enclosing mem_exclusive cpuset.
|
| 2542 | * It never sleeps.
|
| 2543 |
|
| 2544 | int __cpuset_node_allowed_hardwall(int node, gfp_t gfp_mask)
|
| 2545 | {
|
| 2546 | if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
|
| 2547 | return 1;
|
| 2548 | if (node_isset(node, current->mems_allowed))
|
| 2549 | return 1;
|
| 2550 | |
| 2551 | * Allow tasks that have access to memory reserves because they have
|
| 2552 | * been OOM killed to get memory anywhere.
|
| 2553 |
|
| 2554 | if (unlikely(test_thread_flag(TIF_MEMDIE)))
|
| 2555 | return 1;
|
| 2556 | return 0;
|
| 2557 | }
|
| 2558 |
|
| 2559 | |
| 2560 | * cpuset_mem_spread_node() - On which node to begin search for a file page
|
| 2561 | * cpuset_slab_spread_node() - On which node to begin search for a slab page
|
| 2562 | *
|
| 2563 | * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
|
| 2564 | * tasks in a cpuset with is_spread_page or is_spread_slab set),
|
| 2565 | * and if the memory allocation used cpuset_mem_spread_node()
|
| 2566 | * to determine on which node to start looking, as it will for
|
| 2567 | * certain page cache or slab cache pages such as used for file
|
| 2568 | * system buffers and inode caches, then instead of starting on the
|
| 2569 | * local node to look for a free page, rather spread the starting
|
| 2570 | * node around the tasks mems_allowed nodes.
|
| 2571 | *
|
| 2572 | * We don't have to worry about the returned node being offline
|
| 2573 | * because "it can't happen", and even if it did, it would be ok.
|
| 2574 | *
|
| 2575 | * The routines calling guarantee_online_mems() are careful to
|
| 2576 | * only set nodes in task->mems_allowed that are online. So it
|
| 2577 | * should not be possible for the following code to return an
|
| 2578 | * offline node. But if it did, that would be ok, as this routine
|
| 2579 | * is not returning the node where the allocation must be, only
|
| 2580 | * the node where the search should start. The zonelist passed to
|
| 2581 | * __alloc_pages() will include all nodes. If the slab allocator
|
| 2582 | * is passed an offline node, it will fall back to the local node.
|
| 2583 | * See kmem_cache_alloc_node().
|
| 2584 |
|
| 2585 |
|
| 2586 | static int cpuset_spread_node(int *rotor)
|
| 2587 | {
|
| 2588 | int node;
|
| 2589 |
|
| 2590 | node = next_node(*rotor, current->mems_allowed);
|
| 2591 | if (node == MAX_NUMNODES)
|
| 2592 | node = first_node(current->mems_allowed);
|
| 2593 | *rotor = node;
|
| 2594 | return node;
|
| 2595 | }
|
| 2596 |
|
| 2597 | int cpuset_mem_spread_node(void)
|
| 2598 | {
|
| 2599 | if (current->cpuset_mem_spread_rotor == NUMA_NO_NODE)
|
| 2600 | current->cpuset_mem_spread_rotor =
|
| 2601 | node_random(¤t->mems_allowed);
|
| 2602 |
|
| 2603 | return cpuset_spread_node(¤t->cpuset_mem_spread_rotor);
|
| 2604 | }
|
| 2605 |
|
| 2606 | int cpuset_slab_spread_node(void)
|
| 2607 | {
|
| 2608 | if (current->cpuset_slab_spread_rotor == NUMA_NO_NODE)
|
| 2609 | current->cpuset_slab_spread_rotor =
|
| 2610 | node_random(¤t->mems_allowed);
|
| 2611 |
|
| 2612 | return cpuset_spread_node(¤t->cpuset_slab_spread_rotor);
|
| 2613 | }
|
| 2614 |
|
| 2615 | EXPORT_SYMBOL_GPL(cpuset_mem_spread_node);
|
| 2616 |
|
| 2617 | |
| 2618 | * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
|
| 2619 | * @tsk1: pointer to task_struct of some task.
|
| 2620 | * @tsk2: pointer to task_struct of some other task.
|
| 2621 | *
|
| 2622 | * Description: Return true if @tsk1's mems_allowed intersects the
|
| 2623 | * mems_allowed of @tsk2. Used by the OOM killer to determine if
|
| 2624 | * one of the task's memory usage might impact the memory available
|
| 2625 | * to the other.
|
| 2626 | *
|
| 2627 |
|
| 2628 | int cpuset_mems_allowed_intersects(const struct task_struct *tsk1,
|
| 2629 | const struct task_struct *tsk2)
|
| 2630 | {
|
| 2631 | return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed);
|
| 2632 | }
|
| 2633 |
|
| 2634 | #define CPUSET_NODELIST_LEN (256)
|
| 2635 |
|
| 2636 | |
| 2637 | * cpuset_print_task_mems_allowed - prints task's cpuset and mems_allowed
|
| 2638 | * @task: pointer to task_struct of some task.
|
| 2639 | *
|
| 2640 | * Description: Prints @task's name, cpuset name, and cached copy of its
|
| 2641 | * mems_allowed to the kernel log. Must hold task_lock(task) to allow
|
| 2642 | * dereferencing task_cs(task).
|
| 2643 |
|
| 2644 | void cpuset_print_task_mems_allowed(struct task_struct *tsk)
|
| 2645 | {
|
| 2646 |
|
| 2647 | static char cpuset_nodelist[CPUSET_NODELIST_LEN];
|
| 2648 | static DEFINE_SPINLOCK(cpuset_buffer_lock);
|
| 2649 |
|
| 2650 | struct cgroup *cgrp = task_cs(tsk)->css.cgroup;
|
| 2651 |
|
| 2652 | rcu_read_lock();
|
| 2653 | spin_lock(&cpuset_buffer_lock);
|
| 2654 |
|
| 2655 | nodelist_scnprintf(cpuset_nodelist, CPUSET_NODELIST_LEN,
|
| 2656 | tsk->mems_allowed);
|
| 2657 | printk(KERN_INFO "%s cpuset=%s mems_allowed=%s\n",
|
| 2658 | tsk->comm, cgroup_name(cgrp), cpuset_nodelist);
|
| 2659 |
|
| 2660 | spin_unlock(&cpuset_buffer_lock);
|
| 2661 | rcu_read_unlock();
|
| 2662 | }
|
| 2663 |
|
| 2664 | |
| 2665 | * Collection of memory_pressure is suppressed unless
|
| 2666 | * this flag is enabled by writing "1" to the special
|
| 2667 | * cpuset file 'memory_pressure_enabled' in the root cpuset.
|
| 2668 |
|
| 2669 |
|
| 2670 | int cpuset_memory_pressure_enabled __read_mostly;
|
| 2671 |
|
| 2672 | |
| 2673 | * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
|
| 2674 | *
|
| 2675 | * Keep a running average of the rate of synchronous (direct)
|
| 2676 | * page reclaim efforts initiated by tasks in each cpuset.
|
| 2677 | *
|
| 2678 | * This represents the rate at which some task in the cpuset
|
| 2679 | * ran low on memory on all nodes it was allowed to use, and
|
| 2680 | * had to enter the kernels page reclaim code in an effort to
|
| 2681 | * create more free memory by tossing clean pages or swapping
|
| 2682 | * or writing dirty pages.
|
| 2683 | *
|
| 2684 | * Display to user space in the per-cpuset read-only file
|
| 2685 | * "memory_pressure". Value displayed is an integer
|
| 2686 | * representing the recent rate of entry into the synchronous
|
| 2687 | * (direct) page reclaim by any task attached to the cpuset.
|
| 2688 | *
|
| 2689 |
|
| 2690 | void __cpuset_memory_pressure_bump(void)
|
| 2691 | {
|
| 2692 | task_lock(current);
|
| 2693 | fmeter_markevent(&task_cs(current)->fmeter);
|
| 2694 | task_unlock(current);
|
| 2695 | }
|
| 2696 |
|
| 2697 | #ifdef CONFIG_PROC_PID_CPUSET
|
| 2698 | |
| 2699 | * proc_cpuset_show()
|
| 2700 | * - Print tasks cpuset path into seq_file.
|
| 2701 | * - Used for /proc/<pid>/cpuset.
|
| 2702 | * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
|
| 2703 | * doesn't really matter if tsk->cpuset changes after we read it,
|
| 2704 | * and we take cpuset_mutex, keeping cpuset_attach() from changing it
|
| 2705 | * anyway.
|
| 2706 |
|
| 2707 | int proc_cpuset_show(struct seq_file *m, void *unused_v)
|
| 2708 | {
|
| 2709 | struct pid *pid;
|
| 2710 | struct task_struct *tsk;
|
| 2711 | char *buf;
|
| 2712 | struct cgroup_subsys_state *css;
|
| 2713 | int retval;
|
| 2714 |
|
| 2715 | retval = -ENOMEM;
|
| 2716 | buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
|
| 2717 | if (!buf)
|
| 2718 | goto out;
|
| 2719 |
|
| 2720 | retval = -ESRCH;
|
| 2721 | pid = m->private;
|
| 2722 | tsk = get_pid_task(pid, PIDTYPE_PID);
|
| 2723 | if (!tsk)
|
| 2724 | goto out_free;
|
| 2725 |
|
| 2726 | rcu_read_lock();
|
| 2727 | css = task_css(tsk, cpuset_subsys_id);
|
| 2728 | retval = cgroup_path(css->cgroup, buf, PAGE_SIZE);
|
| 2729 | rcu_read_unlock();
|
| 2730 | if (retval < 0)
|
| 2731 | goto out_put_task;
|
| 2732 | seq_puts(m, buf);
|
| 2733 | seq_putc(m, '\n');
|
| 2734 | out_put_task:
|
| 2735 | put_task_struct(tsk);
|
| 2736 | out_free:
|
| 2737 | kfree(buf);
|
| 2738 | out:
|
| 2739 | return retval;
|
| 2740 | }
|
| 2741 | #endif
|
| 2742 |
|
| 2743 |
|
| 2744 | void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task)
|
| 2745 | {
|
| 2746 | seq_printf(m, "Mems_allowed:\t");
|
| 2747 | seq_nodemask(m, &task->mems_allowed);
|
| 2748 | seq_printf(m, "\n");
|
| 2749 | seq_printf(m, "Mems_allowed_list:\t");
|
| 2750 | seq_nodemask_list(m, &task->mems_allowed);
|
| 2751 | seq_printf(m, "\n");
|
| 2752 | }
|