| 1 | // SPDX-License-Identifier: GPL-2.0-only |
| 2 | /* |
| 3 | * kernel/workqueue.c - generic async execution with shared worker pool |
| 4 | * |
| 5 | * Copyright (C) 2002 Ingo Molnar |
| 6 | * |
| 7 | * Derived from the taskqueue/keventd code by: |
| 8 | * David Woodhouse <dwmw2@infradead.org> |
| 9 | * Andrew Morton |
| 10 | * Kai Petzke <wpp@marie.physik.tu-berlin.de> |
| 11 | * Theodore Ts'o <tytso@mit.edu> |
| 12 | * |
| 13 | * Made to use alloc_percpu by Christoph Lameter. |
| 14 | * |
| 15 | * Copyright (C) 2010 SUSE Linux Products GmbH |
| 16 | * Copyright (C) 2010 Tejun Heo <tj@kernel.org> |
| 17 | * |
| 18 | * This is the generic async execution mechanism. Work items as are |
| 19 | * executed in process context. The worker pool is shared and |
| 20 | * automatically managed. There are two worker pools for each CPU (one for |
| 21 | * normal work items and the other for high priority ones) and some extra |
| 22 | * pools for workqueues which are not bound to any specific CPU - the |
| 23 | * number of these backing pools is dynamic. |
| 24 | * |
| 25 | * Please read Documentation/core-api/workqueue.rst for details. |
| 26 | */ |
| 27 | |
| 28 | #include <linux/export.h> |
| 29 | #include <linux/kernel.h> |
| 30 | #include <linux/sched.h> |
| 31 | #include <linux/init.h> |
| 32 | #include <linux/signal.h> |
| 33 | #include <linux/completion.h> |
| 34 | #include <linux/workqueue.h> |
| 35 | #include <linux/slab.h> |
| 36 | #include <linux/cpu.h> |
| 37 | #include <linux/notifier.h> |
| 38 | #include <linux/kthread.h> |
| 39 | #include <linux/hardirq.h> |
| 40 | #include <linux/mempolicy.h> |
| 41 | #include <linux/freezer.h> |
| 42 | #include <linux/debug_locks.h> |
| 43 | #include <linux/lockdep.h> |
| 44 | #include <linux/idr.h> |
| 45 | #include <linux/jhash.h> |
| 46 | #include <linux/hashtable.h> |
| 47 | #include <linux/rculist.h> |
| 48 | #include <linux/nodemask.h> |
| 49 | #include <linux/moduleparam.h> |
| 50 | #include <linux/uaccess.h> |
| 51 | #include <linux/sched/isolation.h> |
| 52 | #include <linux/nmi.h> |
| 53 | #include <linux/kvm_para.h> |
| 54 | |
| 55 | #include "workqueue_internal.h" |
| 56 | |
| 57 | enum { |
| 58 | /* |
| 59 | * worker_pool flags |
| 60 | * |
| 61 | * A bound pool is either associated or disassociated with its CPU. |
| 62 | * While associated (!DISASSOCIATED), all workers are bound to the |
| 63 | * CPU and none has %WORKER_UNBOUND set and concurrency management |
| 64 | * is in effect. |
| 65 | * |
| 66 | * While DISASSOCIATED, the cpu may be offline and all workers have |
| 67 | * %WORKER_UNBOUND set and concurrency management disabled, and may |
| 68 | * be executing on any CPU. The pool behaves as an unbound one. |
| 69 | * |
| 70 | * Note that DISASSOCIATED should be flipped only while holding |
| 71 | * wq_pool_attach_mutex to avoid changing binding state while |
| 72 | * worker_attach_to_pool() is in progress. |
| 73 | */ |
| 74 | POOL_MANAGER_ACTIVE = 1 << 0, /* being managed */ |
| 75 | POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */ |
| 76 | |
| 77 | /* worker flags */ |
| 78 | WORKER_DIE = 1 << 1, /* die die die */ |
| 79 | WORKER_IDLE = 1 << 2, /* is idle */ |
| 80 | WORKER_PREP = 1 << 3, /* preparing to run works */ |
| 81 | WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */ |
| 82 | WORKER_UNBOUND = 1 << 7, /* worker is unbound */ |
| 83 | WORKER_REBOUND = 1 << 8, /* worker was rebound */ |
| 84 | |
| 85 | WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE | |
| 86 | WORKER_UNBOUND | WORKER_REBOUND, |
| 87 | |
| 88 | NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */ |
| 89 | |
| 90 | UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */ |
| 91 | BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */ |
| 92 | |
| 93 | MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */ |
| 94 | IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */ |
| 95 | |
| 96 | MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2, |
| 97 | /* call for help after 10ms |
| 98 | (min two ticks) */ |
| 99 | MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */ |
| 100 | CREATE_COOLDOWN = HZ, /* time to breath after fail */ |
| 101 | |
| 102 | /* |
| 103 | * Rescue workers are used only on emergencies and shared by |
| 104 | * all cpus. Give MIN_NICE. |
| 105 | */ |
| 106 | RESCUER_NICE_LEVEL = MIN_NICE, |
| 107 | HIGHPRI_NICE_LEVEL = MIN_NICE, |
| 108 | |
| 109 | WQ_NAME_LEN = 24, |
| 110 | }; |
| 111 | |
| 112 | /* |
| 113 | * Structure fields follow one of the following exclusion rules. |
| 114 | * |
| 115 | * I: Modifiable by initialization/destruction paths and read-only for |
| 116 | * everyone else. |
| 117 | * |
| 118 | * P: Preemption protected. Disabling preemption is enough and should |
| 119 | * only be modified and accessed from the local cpu. |
| 120 | * |
| 121 | * L: pool->lock protected. Access with pool->lock held. |
| 122 | * |
| 123 | * X: During normal operation, modification requires pool->lock and should |
| 124 | * be done only from local cpu. Either disabling preemption on local |
| 125 | * cpu or grabbing pool->lock is enough for read access. If |
| 126 | * POOL_DISASSOCIATED is set, it's identical to L. |
| 127 | * |
| 128 | * A: wq_pool_attach_mutex protected. |
| 129 | * |
| 130 | * PL: wq_pool_mutex protected. |
| 131 | * |
| 132 | * PR: wq_pool_mutex protected for writes. RCU protected for reads. |
| 133 | * |
| 134 | * PW: wq_pool_mutex and wq->mutex protected for writes. Either for reads. |
| 135 | * |
| 136 | * PWR: wq_pool_mutex and wq->mutex protected for writes. Either or |
| 137 | * RCU for reads. |
| 138 | * |
| 139 | * WQ: wq->mutex protected. |
| 140 | * |
| 141 | * WR: wq->mutex protected for writes. RCU protected for reads. |
| 142 | * |
| 143 | * MD: wq_mayday_lock protected. |
| 144 | */ |
| 145 | |
| 146 | /* struct worker is defined in workqueue_internal.h */ |
| 147 | |
| 148 | struct worker_pool { |
| 149 | raw_spinlock_t lock; /* the pool lock */ |
| 150 | int cpu; /* I: the associated cpu */ |
| 151 | int node; /* I: the associated node ID */ |
| 152 | int id; /* I: pool ID */ |
| 153 | unsigned int flags; /* X: flags */ |
| 154 | |
| 155 | unsigned long watchdog_ts; /* L: watchdog timestamp */ |
| 156 | |
| 157 | /* |
| 158 | * The counter is incremented in a process context on the associated CPU |
| 159 | * w/ preemption disabled, and decremented or reset in the same context |
| 160 | * but w/ pool->lock held. The readers grab pool->lock and are |
| 161 | * guaranteed to see if the counter reached zero. |
| 162 | */ |
| 163 | int nr_running; |
| 164 | |
| 165 | struct list_head worklist; /* L: list of pending works */ |
| 166 | |
| 167 | int nr_workers; /* L: total number of workers */ |
| 168 | int nr_idle; /* L: currently idle workers */ |
| 169 | |
| 170 | struct list_head idle_list; /* L: list of idle workers */ |
| 171 | struct timer_list idle_timer; /* L: worker idle timeout */ |
| 172 | struct timer_list mayday_timer; /* L: SOS timer for workers */ |
| 173 | |
| 174 | /* a workers is either on busy_hash or idle_list, or the manager */ |
| 175 | DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER); |
| 176 | /* L: hash of busy workers */ |
| 177 | |
| 178 | struct worker *manager; /* L: purely informational */ |
| 179 | struct list_head workers; /* A: attached workers */ |
| 180 | struct completion *detach_completion; /* all workers detached */ |
| 181 | |
| 182 | struct ida worker_ida; /* worker IDs for task name */ |
| 183 | |
| 184 | struct workqueue_attrs *attrs; /* I: worker attributes */ |
| 185 | struct hlist_node hash_node; /* PL: unbound_pool_hash node */ |
| 186 | int refcnt; /* PL: refcnt for unbound pools */ |
| 187 | |
| 188 | /* |
| 189 | * Destruction of pool is RCU protected to allow dereferences |
| 190 | * from get_work_pool(). |
| 191 | */ |
| 192 | struct rcu_head rcu; |
| 193 | }; |
| 194 | |
| 195 | /* |
| 196 | * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS |
| 197 | * of work_struct->data are used for flags and the remaining high bits |
| 198 | * point to the pwq; thus, pwqs need to be aligned at two's power of the |
| 199 | * number of flag bits. |
| 200 | */ |
| 201 | struct pool_workqueue { |
| 202 | struct worker_pool *pool; /* I: the associated pool */ |
| 203 | struct workqueue_struct *wq; /* I: the owning workqueue */ |
| 204 | int work_color; /* L: current color */ |
| 205 | int flush_color; /* L: flushing color */ |
| 206 | int refcnt; /* L: reference count */ |
| 207 | int nr_in_flight[WORK_NR_COLORS]; |
| 208 | /* L: nr of in_flight works */ |
| 209 | |
| 210 | /* |
| 211 | * nr_active management and WORK_STRUCT_INACTIVE: |
| 212 | * |
| 213 | * When pwq->nr_active >= max_active, new work item is queued to |
| 214 | * pwq->inactive_works instead of pool->worklist and marked with |
| 215 | * WORK_STRUCT_INACTIVE. |
| 216 | * |
| 217 | * All work items marked with WORK_STRUCT_INACTIVE do not participate |
| 218 | * in pwq->nr_active and all work items in pwq->inactive_works are |
| 219 | * marked with WORK_STRUCT_INACTIVE. But not all WORK_STRUCT_INACTIVE |
| 220 | * work items are in pwq->inactive_works. Some of them are ready to |
| 221 | * run in pool->worklist or worker->scheduled. Those work itmes are |
| 222 | * only struct wq_barrier which is used for flush_work() and should |
| 223 | * not participate in pwq->nr_active. For non-barrier work item, it |
| 224 | * is marked with WORK_STRUCT_INACTIVE iff it is in pwq->inactive_works. |
| 225 | */ |
| 226 | int nr_active; /* L: nr of active works */ |
| 227 | int max_active; /* L: max active works */ |
| 228 | struct list_head inactive_works; /* L: inactive works */ |
| 229 | struct list_head pwqs_node; /* WR: node on wq->pwqs */ |
| 230 | struct list_head mayday_node; /* MD: node on wq->maydays */ |
| 231 | |
| 232 | /* |
| 233 | * Release of unbound pwq is punted to system_wq. See put_pwq() |
| 234 | * and pwq_unbound_release_workfn() for details. pool_workqueue |
| 235 | * itself is also RCU protected so that the first pwq can be |
| 236 | * determined without grabbing wq->mutex. |
| 237 | */ |
| 238 | struct work_struct unbound_release_work; |
| 239 | struct rcu_head rcu; |
| 240 | } __aligned(1 << WORK_STRUCT_FLAG_BITS); |
| 241 | |
| 242 | /* |
| 243 | * Structure used to wait for workqueue flush. |
| 244 | */ |
| 245 | struct wq_flusher { |
| 246 | struct list_head list; /* WQ: list of flushers */ |
| 247 | int flush_color; /* WQ: flush color waiting for */ |
| 248 | struct completion done; /* flush completion */ |
| 249 | }; |
| 250 | |
| 251 | struct wq_device; |
| 252 | |
| 253 | /* |
| 254 | * The externally visible workqueue. It relays the issued work items to |
| 255 | * the appropriate worker_pool through its pool_workqueues. |
| 256 | */ |
| 257 | struct workqueue_struct { |
| 258 | struct list_head pwqs; /* WR: all pwqs of this wq */ |
| 259 | struct list_head list; /* PR: list of all workqueues */ |
| 260 | |
| 261 | struct mutex mutex; /* protects this wq */ |
| 262 | int work_color; /* WQ: current work color */ |
| 263 | int flush_color; /* WQ: current flush color */ |
| 264 | atomic_t nr_pwqs_to_flush; /* flush in progress */ |
| 265 | struct wq_flusher *first_flusher; /* WQ: first flusher */ |
| 266 | struct list_head flusher_queue; /* WQ: flush waiters */ |
| 267 | struct list_head flusher_overflow; /* WQ: flush overflow list */ |
| 268 | |
| 269 | struct list_head maydays; /* MD: pwqs requesting rescue */ |
| 270 | struct worker *rescuer; /* MD: rescue worker */ |
| 271 | |
| 272 | int nr_drainers; /* WQ: drain in progress */ |
| 273 | int saved_max_active; /* WQ: saved pwq max_active */ |
| 274 | |
| 275 | struct workqueue_attrs *unbound_attrs; /* PW: only for unbound wqs */ |
| 276 | struct pool_workqueue *dfl_pwq; /* PW: only for unbound wqs */ |
| 277 | |
| 278 | #ifdef CONFIG_SYSFS |
| 279 | struct wq_device *wq_dev; /* I: for sysfs interface */ |
| 280 | #endif |
| 281 | #ifdef CONFIG_LOCKDEP |
| 282 | char *lock_name; |
| 283 | struct lock_class_key key; |
| 284 | struct lockdep_map lockdep_map; |
| 285 | #endif |
| 286 | char name[WQ_NAME_LEN]; /* I: workqueue name */ |
| 287 | |
| 288 | /* |
| 289 | * Destruction of workqueue_struct is RCU protected to allow walking |
| 290 | * the workqueues list without grabbing wq_pool_mutex. |
| 291 | * This is used to dump all workqueues from sysrq. |
| 292 | */ |
| 293 | struct rcu_head rcu; |
| 294 | |
| 295 | /* hot fields used during command issue, aligned to cacheline */ |
| 296 | unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */ |
| 297 | struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwqs */ |
| 298 | struct pool_workqueue __rcu *numa_pwq_tbl[]; /* PWR: unbound pwqs indexed by node */ |
| 299 | }; |
| 300 | |
| 301 | static struct kmem_cache *pwq_cache; |
| 302 | |
| 303 | static cpumask_var_t *wq_numa_possible_cpumask; |
| 304 | /* possible CPUs of each node */ |
| 305 | |
| 306 | static bool wq_disable_numa; |
| 307 | module_param_named(disable_numa, wq_disable_numa, bool, 0444); |
| 308 | |
| 309 | /* see the comment above the definition of WQ_POWER_EFFICIENT */ |
| 310 | static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT); |
| 311 | module_param_named(power_efficient, wq_power_efficient, bool, 0444); |
| 312 | |
| 313 | static bool wq_online; /* can kworkers be created yet? */ |
| 314 | |
| 315 | static bool wq_numa_enabled; /* unbound NUMA affinity enabled */ |
| 316 | |
| 317 | /* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */ |
| 318 | static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf; |
| 319 | |
| 320 | static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */ |
| 321 | static DEFINE_MUTEX(wq_pool_attach_mutex); /* protects worker attach/detach */ |
| 322 | static DEFINE_RAW_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */ |
| 323 | /* wait for manager to go away */ |
| 324 | static struct rcuwait manager_wait = __RCUWAIT_INITIALIZER(manager_wait); |
| 325 | |
| 326 | static LIST_HEAD(workqueues); /* PR: list of all workqueues */ |
| 327 | static bool workqueue_freezing; /* PL: have wqs started freezing? */ |
| 328 | |
| 329 | /* PL: allowable cpus for unbound wqs and work items */ |
| 330 | static cpumask_var_t wq_unbound_cpumask; |
| 331 | |
| 332 | /* CPU where unbound work was last round robin scheduled from this CPU */ |
| 333 | static DEFINE_PER_CPU(int, wq_rr_cpu_last); |
| 334 | |
| 335 | /* |
| 336 | * Local execution of unbound work items is no longer guaranteed. The |
| 337 | * following always forces round-robin CPU selection on unbound work items |
| 338 | * to uncover usages which depend on it. |
| 339 | */ |
| 340 | #ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU |
| 341 | static bool wq_debug_force_rr_cpu = true; |
| 342 | #else |
| 343 | static bool wq_debug_force_rr_cpu = false; |
| 344 | #endif |
| 345 | module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644); |
| 346 | |
| 347 | /* the per-cpu worker pools */ |
| 348 | static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], cpu_worker_pools); |
| 349 | |
| 350 | static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */ |
| 351 | |
| 352 | /* PL: hash of all unbound pools keyed by pool->attrs */ |
| 353 | static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER); |
| 354 | |
| 355 | /* I: attributes used when instantiating standard unbound pools on demand */ |
| 356 | static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS]; |
| 357 | |
| 358 | /* I: attributes used when instantiating ordered pools on demand */ |
| 359 | static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS]; |
| 360 | |
| 361 | struct workqueue_struct *system_wq __read_mostly; |
| 362 | EXPORT_SYMBOL(system_wq); |
| 363 | struct workqueue_struct *system_highpri_wq __read_mostly; |
| 364 | EXPORT_SYMBOL_GPL(system_highpri_wq); |
| 365 | struct workqueue_struct *system_long_wq __read_mostly; |
| 366 | EXPORT_SYMBOL_GPL(system_long_wq); |
| 367 | struct workqueue_struct *system_unbound_wq __read_mostly; |
| 368 | EXPORT_SYMBOL_GPL(system_unbound_wq); |
| 369 | struct workqueue_struct *system_freezable_wq __read_mostly; |
| 370 | EXPORT_SYMBOL_GPL(system_freezable_wq); |
| 371 | struct workqueue_struct *system_power_efficient_wq __read_mostly; |
| 372 | EXPORT_SYMBOL_GPL(system_power_efficient_wq); |
| 373 | struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly; |
| 374 | EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq); |
| 375 | |
| 376 | static int worker_thread(void *__worker); |
| 377 | static void workqueue_sysfs_unregister(struct workqueue_struct *wq); |
| 378 | static void show_pwq(struct pool_workqueue *pwq); |
| 379 | static void show_one_worker_pool(struct worker_pool *pool); |
| 380 | |
| 381 | #define CREATE_TRACE_POINTS |
| 382 | #include <trace/events/workqueue.h> |
| 383 | |
| 384 | #define assert_rcu_or_pool_mutex() \ |
| 385 | RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \ |
| 386 | !lockdep_is_held(&wq_pool_mutex), \ |
| 387 | "RCU or wq_pool_mutex should be held") |
| 388 | |
| 389 | #define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \ |
| 390 | RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \ |
| 391 | !lockdep_is_held(&wq->mutex) && \ |
| 392 | !lockdep_is_held(&wq_pool_mutex), \ |
| 393 | "RCU, wq->mutex or wq_pool_mutex should be held") |
| 394 | |
| 395 | #define for_each_cpu_worker_pool(pool, cpu) \ |
| 396 | for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \ |
| 397 | (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \ |
| 398 | (pool)++) |
| 399 | |
| 400 | /** |
| 401 | * for_each_pool - iterate through all worker_pools in the system |
| 402 | * @pool: iteration cursor |
| 403 | * @pi: integer used for iteration |
| 404 | * |
| 405 | * This must be called either with wq_pool_mutex held or RCU read |
| 406 | * locked. If the pool needs to be used beyond the locking in effect, the |
| 407 | * caller is responsible for guaranteeing that the pool stays online. |
| 408 | * |
| 409 | * The if/else clause exists only for the lockdep assertion and can be |
| 410 | * ignored. |
| 411 | */ |
| 412 | #define for_each_pool(pool, pi) \ |
| 413 | idr_for_each_entry(&worker_pool_idr, pool, pi) \ |
| 414 | if (({ assert_rcu_or_pool_mutex(); false; })) { } \ |
| 415 | else |
| 416 | |
| 417 | /** |
| 418 | * for_each_pool_worker - iterate through all workers of a worker_pool |
| 419 | * @worker: iteration cursor |
| 420 | * @pool: worker_pool to iterate workers of |
| 421 | * |
| 422 | * This must be called with wq_pool_attach_mutex. |
| 423 | * |
| 424 | * The if/else clause exists only for the lockdep assertion and can be |
| 425 | * ignored. |
| 426 | */ |
| 427 | #define for_each_pool_worker(worker, pool) \ |
| 428 | list_for_each_entry((worker), &(pool)->workers, node) \ |
| 429 | if (({ lockdep_assert_held(&wq_pool_attach_mutex); false; })) { } \ |
| 430 | else |
| 431 | |
| 432 | /** |
| 433 | * for_each_pwq - iterate through all pool_workqueues of the specified workqueue |
| 434 | * @pwq: iteration cursor |
| 435 | * @wq: the target workqueue |
| 436 | * |
| 437 | * This must be called either with wq->mutex held or RCU read locked. |
| 438 | * If the pwq needs to be used beyond the locking in effect, the caller is |
| 439 | * responsible for guaranteeing that the pwq stays online. |
| 440 | * |
| 441 | * The if/else clause exists only for the lockdep assertion and can be |
| 442 | * ignored. |
| 443 | */ |
| 444 | #define for_each_pwq(pwq, wq) \ |
| 445 | list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node, \ |
| 446 | lockdep_is_held(&(wq->mutex))) |
| 447 | |
| 448 | #ifdef CONFIG_DEBUG_OBJECTS_WORK |
| 449 | |
| 450 | static const struct debug_obj_descr work_debug_descr; |
| 451 | |
| 452 | static void *work_debug_hint(void *addr) |
| 453 | { |
| 454 | return ((struct work_struct *) addr)->func; |
| 455 | } |
| 456 | |
| 457 | static bool work_is_static_object(void *addr) |
| 458 | { |
| 459 | struct work_struct *work = addr; |
| 460 | |
| 461 | return test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work)); |
| 462 | } |
| 463 | |
| 464 | /* |
| 465 | * fixup_init is called when: |
| 466 | * - an active object is initialized |
| 467 | */ |
| 468 | static bool work_fixup_init(void *addr, enum debug_obj_state state) |
| 469 | { |
| 470 | struct work_struct *work = addr; |
| 471 | |
| 472 | switch (state) { |
| 473 | case ODEBUG_STATE_ACTIVE: |
| 474 | cancel_work_sync(work); |
| 475 | debug_object_init(work, &work_debug_descr); |
| 476 | return true; |
| 477 | default: |
| 478 | return false; |
| 479 | } |
| 480 | } |
| 481 | |
| 482 | /* |
| 483 | * fixup_free is called when: |
| 484 | * - an active object is freed |
| 485 | */ |
| 486 | static bool work_fixup_free(void *addr, enum debug_obj_state state) |
| 487 | { |
| 488 | struct work_struct *work = addr; |
| 489 | |
| 490 | switch (state) { |
| 491 | case ODEBUG_STATE_ACTIVE: |
| 492 | cancel_work_sync(work); |
| 493 | debug_object_free(work, &work_debug_descr); |
| 494 | return true; |
| 495 | default: |
| 496 | return false; |
| 497 | } |
| 498 | } |
| 499 | |
| 500 | static const struct debug_obj_descr work_debug_descr = { |
| 501 | .name = "work_struct", |
| 502 | .debug_hint = work_debug_hint, |
| 503 | .is_static_object = work_is_static_object, |
| 504 | .fixup_init = work_fixup_init, |
| 505 | .fixup_free = work_fixup_free, |
| 506 | }; |
| 507 | |
| 508 | static inline void debug_work_activate(struct work_struct *work) |
| 509 | { |
| 510 | debug_object_activate(work, &work_debug_descr); |
| 511 | } |
| 512 | |
| 513 | static inline void debug_work_deactivate(struct work_struct *work) |
| 514 | { |
| 515 | debug_object_deactivate(work, &work_debug_descr); |
| 516 | } |
| 517 | |
| 518 | void __init_work(struct work_struct *work, int onstack) |
| 519 | { |
| 520 | if (onstack) |
| 521 | debug_object_init_on_stack(work, &work_debug_descr); |
| 522 | else |
| 523 | debug_object_init(work, &work_debug_descr); |
| 524 | } |
| 525 | EXPORT_SYMBOL_GPL(__init_work); |
| 526 | |
| 527 | void destroy_work_on_stack(struct work_struct *work) |
| 528 | { |
| 529 | debug_object_free(work, &work_debug_descr); |
| 530 | } |
| 531 | EXPORT_SYMBOL_GPL(destroy_work_on_stack); |
| 532 | |
| 533 | void destroy_delayed_work_on_stack(struct delayed_work *work) |
| 534 | { |
| 535 | destroy_timer_on_stack(&work->timer); |
| 536 | debug_object_free(&work->work, &work_debug_descr); |
| 537 | } |
| 538 | EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack); |
| 539 | |
| 540 | #else |
| 541 | static inline void debug_work_activate(struct work_struct *work) { } |
| 542 | static inline void debug_work_deactivate(struct work_struct *work) { } |
| 543 | #endif |
| 544 | |
| 545 | /** |
| 546 | * worker_pool_assign_id - allocate ID and assign it to @pool |
| 547 | * @pool: the pool pointer of interest |
| 548 | * |
| 549 | * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned |
| 550 | * successfully, -errno on failure. |
| 551 | */ |
| 552 | static int worker_pool_assign_id(struct worker_pool *pool) |
| 553 | { |
| 554 | int ret; |
| 555 | |
| 556 | lockdep_assert_held(&wq_pool_mutex); |
| 557 | |
| 558 | ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE, |
| 559 | GFP_KERNEL); |
| 560 | if (ret >= 0) { |
| 561 | pool->id = ret; |
| 562 | return 0; |
| 563 | } |
| 564 | return ret; |
| 565 | } |
| 566 | |
| 567 | /** |
| 568 | * unbound_pwq_by_node - return the unbound pool_workqueue for the given node |
| 569 | * @wq: the target workqueue |
| 570 | * @node: the node ID |
| 571 | * |
| 572 | * This must be called with any of wq_pool_mutex, wq->mutex or RCU |
| 573 | * read locked. |
| 574 | * If the pwq needs to be used beyond the locking in effect, the caller is |
| 575 | * responsible for guaranteeing that the pwq stays online. |
| 576 | * |
| 577 | * Return: The unbound pool_workqueue for @node. |
| 578 | */ |
| 579 | static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq, |
| 580 | int node) |
| 581 | { |
| 582 | assert_rcu_or_wq_mutex_or_pool_mutex(wq); |
| 583 | |
| 584 | /* |
| 585 | * XXX: @node can be NUMA_NO_NODE if CPU goes offline while a |
| 586 | * delayed item is pending. The plan is to keep CPU -> NODE |
| 587 | * mapping valid and stable across CPU on/offlines. Once that |
| 588 | * happens, this workaround can be removed. |
| 589 | */ |
| 590 | if (unlikely(node == NUMA_NO_NODE)) |
| 591 | return wq->dfl_pwq; |
| 592 | |
| 593 | return rcu_dereference_raw(wq->numa_pwq_tbl[node]); |
| 594 | } |
| 595 | |
| 596 | static unsigned int work_color_to_flags(int color) |
| 597 | { |
| 598 | return color << WORK_STRUCT_COLOR_SHIFT; |
| 599 | } |
| 600 | |
| 601 | static int get_work_color(unsigned long work_data) |
| 602 | { |
| 603 | return (work_data >> WORK_STRUCT_COLOR_SHIFT) & |
| 604 | ((1 << WORK_STRUCT_COLOR_BITS) - 1); |
| 605 | } |
| 606 | |
| 607 | static int work_next_color(int color) |
| 608 | { |
| 609 | return (color + 1) % WORK_NR_COLORS; |
| 610 | } |
| 611 | |
| 612 | /* |
| 613 | * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data |
| 614 | * contain the pointer to the queued pwq. Once execution starts, the flag |
| 615 | * is cleared and the high bits contain OFFQ flags and pool ID. |
| 616 | * |
| 617 | * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling() |
| 618 | * and clear_work_data() can be used to set the pwq, pool or clear |
| 619 | * work->data. These functions should only be called while the work is |
| 620 | * owned - ie. while the PENDING bit is set. |
| 621 | * |
| 622 | * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq |
| 623 | * corresponding to a work. Pool is available once the work has been |
| 624 | * queued anywhere after initialization until it is sync canceled. pwq is |
| 625 | * available only while the work item is queued. |
| 626 | * |
| 627 | * %WORK_OFFQ_CANCELING is used to mark a work item which is being |
| 628 | * canceled. While being canceled, a work item may have its PENDING set |
| 629 | * but stay off timer and worklist for arbitrarily long and nobody should |
| 630 | * try to steal the PENDING bit. |
| 631 | */ |
| 632 | static inline void set_work_data(struct work_struct *work, unsigned long data, |
| 633 | unsigned long flags) |
| 634 | { |
| 635 | WARN_ON_ONCE(!work_pending(work)); |
| 636 | atomic_long_set(&work->data, data | flags | work_static(work)); |
| 637 | } |
| 638 | |
| 639 | static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq, |
| 640 | unsigned long extra_flags) |
| 641 | { |
| 642 | set_work_data(work, (unsigned long)pwq, |
| 643 | WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags); |
| 644 | } |
| 645 | |
| 646 | static void set_work_pool_and_keep_pending(struct work_struct *work, |
| 647 | int pool_id) |
| 648 | { |
| 649 | set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, |
| 650 | WORK_STRUCT_PENDING); |
| 651 | } |
| 652 | |
| 653 | static void set_work_pool_and_clear_pending(struct work_struct *work, |
| 654 | int pool_id) |
| 655 | { |
| 656 | /* |
| 657 | * The following wmb is paired with the implied mb in |
| 658 | * test_and_set_bit(PENDING) and ensures all updates to @work made |
| 659 | * here are visible to and precede any updates by the next PENDING |
| 660 | * owner. |
| 661 | */ |
| 662 | smp_wmb(); |
| 663 | set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0); |
| 664 | /* |
| 665 | * The following mb guarantees that previous clear of a PENDING bit |
| 666 | * will not be reordered with any speculative LOADS or STORES from |
| 667 | * work->current_func, which is executed afterwards. This possible |
| 668 | * reordering can lead to a missed execution on attempt to queue |
| 669 | * the same @work. E.g. consider this case: |
| 670 | * |
| 671 | * CPU#0 CPU#1 |
| 672 | * ---------------------------- -------------------------------- |
| 673 | * |
| 674 | * 1 STORE event_indicated |
| 675 | * 2 queue_work_on() { |
| 676 | * 3 test_and_set_bit(PENDING) |
| 677 | * 4 } set_..._and_clear_pending() { |
| 678 | * 5 set_work_data() # clear bit |
| 679 | * 6 smp_mb() |
| 680 | * 7 work->current_func() { |
| 681 | * 8 LOAD event_indicated |
| 682 | * } |
| 683 | * |
| 684 | * Without an explicit full barrier speculative LOAD on line 8 can |
| 685 | * be executed before CPU#0 does STORE on line 1. If that happens, |
| 686 | * CPU#0 observes the PENDING bit is still set and new execution of |
| 687 | * a @work is not queued in a hope, that CPU#1 will eventually |
| 688 | * finish the queued @work. Meanwhile CPU#1 does not see |
| 689 | * event_indicated is set, because speculative LOAD was executed |
| 690 | * before actual STORE. |
| 691 | */ |
| 692 | smp_mb(); |
| 693 | } |
| 694 | |
| 695 | static void clear_work_data(struct work_struct *work) |
| 696 | { |
| 697 | smp_wmb(); /* see set_work_pool_and_clear_pending() */ |
| 698 | set_work_data(work, WORK_STRUCT_NO_POOL, 0); |
| 699 | } |
| 700 | |
| 701 | static struct pool_workqueue *get_work_pwq(struct work_struct *work) |
| 702 | { |
| 703 | unsigned long data = atomic_long_read(&work->data); |
| 704 | |
| 705 | if (data & WORK_STRUCT_PWQ) |
| 706 | return (void *)(data & WORK_STRUCT_WQ_DATA_MASK); |
| 707 | else |
| 708 | return NULL; |
| 709 | } |
| 710 | |
| 711 | /** |
| 712 | * get_work_pool - return the worker_pool a given work was associated with |
| 713 | * @work: the work item of interest |
| 714 | * |
| 715 | * Pools are created and destroyed under wq_pool_mutex, and allows read |
| 716 | * access under RCU read lock. As such, this function should be |
| 717 | * called under wq_pool_mutex or inside of a rcu_read_lock() region. |
| 718 | * |
| 719 | * All fields of the returned pool are accessible as long as the above |
| 720 | * mentioned locking is in effect. If the returned pool needs to be used |
| 721 | * beyond the critical section, the caller is responsible for ensuring the |
| 722 | * returned pool is and stays online. |
| 723 | * |
| 724 | * Return: The worker_pool @work was last associated with. %NULL if none. |
| 725 | */ |
| 726 | static struct worker_pool *get_work_pool(struct work_struct *work) |
| 727 | { |
| 728 | unsigned long data = atomic_long_read(&work->data); |
| 729 | int pool_id; |
| 730 | |
| 731 | assert_rcu_or_pool_mutex(); |
| 732 | |
| 733 | if (data & WORK_STRUCT_PWQ) |
| 734 | return ((struct pool_workqueue *) |
| 735 | (data & WORK_STRUCT_WQ_DATA_MASK))->pool; |
| 736 | |
| 737 | pool_id = data >> WORK_OFFQ_POOL_SHIFT; |
| 738 | if (pool_id == WORK_OFFQ_POOL_NONE) |
| 739 | return NULL; |
| 740 | |
| 741 | return idr_find(&worker_pool_idr, pool_id); |
| 742 | } |
| 743 | |
| 744 | /** |
| 745 | * get_work_pool_id - return the worker pool ID a given work is associated with |
| 746 | * @work: the work item of interest |
| 747 | * |
| 748 | * Return: The worker_pool ID @work was last associated with. |
| 749 | * %WORK_OFFQ_POOL_NONE if none. |
| 750 | */ |
| 751 | static int get_work_pool_id(struct work_struct *work) |
| 752 | { |
| 753 | unsigned long data = atomic_long_read(&work->data); |
| 754 | |
| 755 | if (data & WORK_STRUCT_PWQ) |
| 756 | return ((struct pool_workqueue *) |
| 757 | (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id; |
| 758 | |
| 759 | return data >> WORK_OFFQ_POOL_SHIFT; |
| 760 | } |
| 761 | |
| 762 | static void mark_work_canceling(struct work_struct *work) |
| 763 | { |
| 764 | unsigned long pool_id = get_work_pool_id(work); |
| 765 | |
| 766 | pool_id <<= WORK_OFFQ_POOL_SHIFT; |
| 767 | set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING); |
| 768 | } |
| 769 | |
| 770 | static bool work_is_canceling(struct work_struct *work) |
| 771 | { |
| 772 | unsigned long data = atomic_long_read(&work->data); |
| 773 | |
| 774 | return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING); |
| 775 | } |
| 776 | |
| 777 | /* |
| 778 | * Policy functions. These define the policies on how the global worker |
| 779 | * pools are managed. Unless noted otherwise, these functions assume that |
| 780 | * they're being called with pool->lock held. |
| 781 | */ |
| 782 | |
| 783 | static bool __need_more_worker(struct worker_pool *pool) |
| 784 | { |
| 785 | return !pool->nr_running; |
| 786 | } |
| 787 | |
| 788 | /* |
| 789 | * Need to wake up a worker? Called from anything but currently |
| 790 | * running workers. |
| 791 | * |
| 792 | * Note that, because unbound workers never contribute to nr_running, this |
| 793 | * function will always return %true for unbound pools as long as the |
| 794 | * worklist isn't empty. |
| 795 | */ |
| 796 | static bool need_more_worker(struct worker_pool *pool) |
| 797 | { |
| 798 | return !list_empty(&pool->worklist) && __need_more_worker(pool); |
| 799 | } |
| 800 | |
| 801 | /* Can I start working? Called from busy but !running workers. */ |
| 802 | static bool may_start_working(struct worker_pool *pool) |
| 803 | { |
| 804 | return pool->nr_idle; |
| 805 | } |
| 806 | |
| 807 | /* Do I need to keep working? Called from currently running workers. */ |
| 808 | static bool keep_working(struct worker_pool *pool) |
| 809 | { |
| 810 | return !list_empty(&pool->worklist) && (pool->nr_running <= 1); |
| 811 | } |
| 812 | |
| 813 | /* Do we need a new worker? Called from manager. */ |
| 814 | static bool need_to_create_worker(struct worker_pool *pool) |
| 815 | { |
| 816 | return need_more_worker(pool) && !may_start_working(pool); |
| 817 | } |
| 818 | |
| 819 | /* Do we have too many workers and should some go away? */ |
| 820 | static bool too_many_workers(struct worker_pool *pool) |
| 821 | { |
| 822 | bool managing = pool->flags & POOL_MANAGER_ACTIVE; |
| 823 | int nr_idle = pool->nr_idle + managing; /* manager is considered idle */ |
| 824 | int nr_busy = pool->nr_workers - nr_idle; |
| 825 | |
| 826 | return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy; |
| 827 | } |
| 828 | |
| 829 | /* |
| 830 | * Wake up functions. |
| 831 | */ |
| 832 | |
| 833 | /* Return the first idle worker. Called with pool->lock held. */ |
| 834 | static struct worker *first_idle_worker(struct worker_pool *pool) |
| 835 | { |
| 836 | if (unlikely(list_empty(&pool->idle_list))) |
| 837 | return NULL; |
| 838 | |
| 839 | return list_first_entry(&pool->idle_list, struct worker, entry); |
| 840 | } |
| 841 | |
| 842 | /** |
| 843 | * wake_up_worker - wake up an idle worker |
| 844 | * @pool: worker pool to wake worker from |
| 845 | * |
| 846 | * Wake up the first idle worker of @pool. |
| 847 | * |
| 848 | * CONTEXT: |
| 849 | * raw_spin_lock_irq(pool->lock). |
| 850 | */ |
| 851 | static void wake_up_worker(struct worker_pool *pool) |
| 852 | { |
| 853 | struct worker *worker = first_idle_worker(pool); |
| 854 | |
| 855 | if (likely(worker)) |
| 856 | wake_up_process(worker->task); |
| 857 | } |
| 858 | |
| 859 | /** |
| 860 | * wq_worker_running - a worker is running again |
| 861 | * @task: task waking up |
| 862 | * |
| 863 | * This function is called when a worker returns from schedule() |
| 864 | */ |
| 865 | void wq_worker_running(struct task_struct *task) |
| 866 | { |
| 867 | struct worker *worker = kthread_data(task); |
| 868 | |
| 869 | if (!worker->sleeping) |
| 870 | return; |
| 871 | |
| 872 | /* |
| 873 | * If preempted by unbind_workers() between the WORKER_NOT_RUNNING check |
| 874 | * and the nr_running increment below, we may ruin the nr_running reset |
| 875 | * and leave with an unexpected pool->nr_running == 1 on the newly unbound |
| 876 | * pool. Protect against such race. |
| 877 | */ |
| 878 | preempt_disable(); |
| 879 | if (!(worker->flags & WORKER_NOT_RUNNING)) |
| 880 | worker->pool->nr_running++; |
| 881 | preempt_enable(); |
| 882 | worker->sleeping = 0; |
| 883 | } |
| 884 | |
| 885 | /** |
| 886 | * wq_worker_sleeping - a worker is going to sleep |
| 887 | * @task: task going to sleep |
| 888 | * |
| 889 | * This function is called from schedule() when a busy worker is |
| 890 | * going to sleep. |
| 891 | */ |
| 892 | void wq_worker_sleeping(struct task_struct *task) |
| 893 | { |
| 894 | struct worker *worker = kthread_data(task); |
| 895 | struct worker_pool *pool; |
| 896 | |
| 897 | /* |
| 898 | * Rescuers, which may not have all the fields set up like normal |
| 899 | * workers, also reach here, let's not access anything before |
| 900 | * checking NOT_RUNNING. |
| 901 | */ |
| 902 | if (worker->flags & WORKER_NOT_RUNNING) |
| 903 | return; |
| 904 | |
| 905 | pool = worker->pool; |
| 906 | |
| 907 | /* Return if preempted before wq_worker_running() was reached */ |
| 908 | if (worker->sleeping) |
| 909 | return; |
| 910 | |
| 911 | worker->sleeping = 1; |
| 912 | raw_spin_lock_irq(&pool->lock); |
| 913 | |
| 914 | /* |
| 915 | * Recheck in case unbind_workers() preempted us. We don't |
| 916 | * want to decrement nr_running after the worker is unbound |
| 917 | * and nr_running has been reset. |
| 918 | */ |
| 919 | if (worker->flags & WORKER_NOT_RUNNING) { |
| 920 | raw_spin_unlock_irq(&pool->lock); |
| 921 | return; |
| 922 | } |
| 923 | |
| 924 | pool->nr_running--; |
| 925 | if (need_more_worker(pool)) |
| 926 | wake_up_worker(pool); |
| 927 | raw_spin_unlock_irq(&pool->lock); |
| 928 | } |
| 929 | |
| 930 | /** |
| 931 | * wq_worker_last_func - retrieve worker's last work function |
| 932 | * @task: Task to retrieve last work function of. |
| 933 | * |
| 934 | * Determine the last function a worker executed. This is called from |
| 935 | * the scheduler to get a worker's last known identity. |
| 936 | * |
| 937 | * CONTEXT: |
| 938 | * raw_spin_lock_irq(rq->lock) |
| 939 | * |
| 940 | * This function is called during schedule() when a kworker is going |
| 941 | * to sleep. It's used by psi to identify aggregation workers during |
| 942 | * dequeuing, to allow periodic aggregation to shut-off when that |
| 943 | * worker is the last task in the system or cgroup to go to sleep. |
| 944 | * |
| 945 | * As this function doesn't involve any workqueue-related locking, it |
| 946 | * only returns stable values when called from inside the scheduler's |
| 947 | * queuing and dequeuing paths, when @task, which must be a kworker, |
| 948 | * is guaranteed to not be processing any works. |
| 949 | * |
| 950 | * Return: |
| 951 | * The last work function %current executed as a worker, NULL if it |
| 952 | * hasn't executed any work yet. |
| 953 | */ |
| 954 | work_func_t wq_worker_last_func(struct task_struct *task) |
| 955 | { |
| 956 | struct worker *worker = kthread_data(task); |
| 957 | |
| 958 | return worker->last_func; |
| 959 | } |
| 960 | |
| 961 | /** |
| 962 | * worker_set_flags - set worker flags and adjust nr_running accordingly |
| 963 | * @worker: self |
| 964 | * @flags: flags to set |
| 965 | * |
| 966 | * Set @flags in @worker->flags and adjust nr_running accordingly. |
| 967 | * |
| 968 | * CONTEXT: |
| 969 | * raw_spin_lock_irq(pool->lock) |
| 970 | */ |
| 971 | static inline void worker_set_flags(struct worker *worker, unsigned int flags) |
| 972 | { |
| 973 | struct worker_pool *pool = worker->pool; |
| 974 | |
| 975 | WARN_ON_ONCE(worker->task != current); |
| 976 | |
| 977 | /* If transitioning into NOT_RUNNING, adjust nr_running. */ |
| 978 | if ((flags & WORKER_NOT_RUNNING) && |
| 979 | !(worker->flags & WORKER_NOT_RUNNING)) { |
| 980 | pool->nr_running--; |
| 981 | } |
| 982 | |
| 983 | worker->flags |= flags; |
| 984 | } |
| 985 | |
| 986 | /** |
| 987 | * worker_clr_flags - clear worker flags and adjust nr_running accordingly |
| 988 | * @worker: self |
| 989 | * @flags: flags to clear |
| 990 | * |
| 991 | * Clear @flags in @worker->flags and adjust nr_running accordingly. |
| 992 | * |
| 993 | * CONTEXT: |
| 994 | * raw_spin_lock_irq(pool->lock) |
| 995 | */ |
| 996 | static inline void worker_clr_flags(struct worker *worker, unsigned int flags) |
| 997 | { |
| 998 | struct worker_pool *pool = worker->pool; |
| 999 | unsigned int oflags = worker->flags; |
| 1000 | |
| 1001 | WARN_ON_ONCE(worker->task != current); |
| 1002 | |
| 1003 | worker->flags &= ~flags; |
| 1004 | |
| 1005 | /* |
| 1006 | * If transitioning out of NOT_RUNNING, increment nr_running. Note |
| 1007 | * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask |
| 1008 | * of multiple flags, not a single flag. |
| 1009 | */ |
| 1010 | if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING)) |
| 1011 | if (!(worker->flags & WORKER_NOT_RUNNING)) |
| 1012 | pool->nr_running++; |
| 1013 | } |
| 1014 | |
| 1015 | /** |
| 1016 | * find_worker_executing_work - find worker which is executing a work |
| 1017 | * @pool: pool of interest |
| 1018 | * @work: work to find worker for |
| 1019 | * |
| 1020 | * Find a worker which is executing @work on @pool by searching |
| 1021 | * @pool->busy_hash which is keyed by the address of @work. For a worker |
| 1022 | * to match, its current execution should match the address of @work and |
| 1023 | * its work function. This is to avoid unwanted dependency between |
| 1024 | * unrelated work executions through a work item being recycled while still |
| 1025 | * being executed. |
| 1026 | * |
| 1027 | * This is a bit tricky. A work item may be freed once its execution |
| 1028 | * starts and nothing prevents the freed area from being recycled for |
| 1029 | * another work item. If the same work item address ends up being reused |
| 1030 | * before the original execution finishes, workqueue will identify the |
| 1031 | * recycled work item as currently executing and make it wait until the |
| 1032 | * current execution finishes, introducing an unwanted dependency. |
| 1033 | * |
| 1034 | * This function checks the work item address and work function to avoid |
| 1035 | * false positives. Note that this isn't complete as one may construct a |
| 1036 | * work function which can introduce dependency onto itself through a |
| 1037 | * recycled work item. Well, if somebody wants to shoot oneself in the |
| 1038 | * foot that badly, there's only so much we can do, and if such deadlock |
| 1039 | * actually occurs, it should be easy to locate the culprit work function. |
| 1040 | * |
| 1041 | * CONTEXT: |
| 1042 | * raw_spin_lock_irq(pool->lock). |
| 1043 | * |
| 1044 | * Return: |
| 1045 | * Pointer to worker which is executing @work if found, %NULL |
| 1046 | * otherwise. |
| 1047 | */ |
| 1048 | static struct worker *find_worker_executing_work(struct worker_pool *pool, |
| 1049 | struct work_struct *work) |
| 1050 | { |
| 1051 | struct worker *worker; |
| 1052 | |
| 1053 | hash_for_each_possible(pool->busy_hash, worker, hentry, |
| 1054 | (unsigned long)work) |
| 1055 | if (worker->current_work == work && |
| 1056 | worker->current_func == work->func) |
| 1057 | return worker; |
| 1058 | |
| 1059 | return NULL; |
| 1060 | } |
| 1061 | |
| 1062 | /** |
| 1063 | * move_linked_works - move linked works to a list |
| 1064 | * @work: start of series of works to be scheduled |
| 1065 | * @head: target list to append @work to |
| 1066 | * @nextp: out parameter for nested worklist walking |
| 1067 | * |
| 1068 | * Schedule linked works starting from @work to @head. Work series to |
| 1069 | * be scheduled starts at @work and includes any consecutive work with |
| 1070 | * WORK_STRUCT_LINKED set in its predecessor. |
| 1071 | * |
| 1072 | * If @nextp is not NULL, it's updated to point to the next work of |
| 1073 | * the last scheduled work. This allows move_linked_works() to be |
| 1074 | * nested inside outer list_for_each_entry_safe(). |
| 1075 | * |
| 1076 | * CONTEXT: |
| 1077 | * raw_spin_lock_irq(pool->lock). |
| 1078 | */ |
| 1079 | static void move_linked_works(struct work_struct *work, struct list_head *head, |
| 1080 | struct work_struct **nextp) |
| 1081 | { |
| 1082 | struct work_struct *n; |
| 1083 | |
| 1084 | /* |
| 1085 | * Linked worklist will always end before the end of the list, |
| 1086 | * use NULL for list head. |
| 1087 | */ |
| 1088 | list_for_each_entry_safe_from(work, n, NULL, entry) { |
| 1089 | list_move_tail(&work->entry, head); |
| 1090 | if (!(*work_data_bits(work) & WORK_STRUCT_LINKED)) |
| 1091 | break; |
| 1092 | } |
| 1093 | |
| 1094 | /* |
| 1095 | * If we're already inside safe list traversal and have moved |
| 1096 | * multiple works to the scheduled queue, the next position |
| 1097 | * needs to be updated. |
| 1098 | */ |
| 1099 | if (nextp) |
| 1100 | *nextp = n; |
| 1101 | } |
| 1102 | |
| 1103 | /** |
| 1104 | * get_pwq - get an extra reference on the specified pool_workqueue |
| 1105 | * @pwq: pool_workqueue to get |
| 1106 | * |
| 1107 | * Obtain an extra reference on @pwq. The caller should guarantee that |
| 1108 | * @pwq has positive refcnt and be holding the matching pool->lock. |
| 1109 | */ |
| 1110 | static void get_pwq(struct pool_workqueue *pwq) |
| 1111 | { |
| 1112 | lockdep_assert_held(&pwq->pool->lock); |
| 1113 | WARN_ON_ONCE(pwq->refcnt <= 0); |
| 1114 | pwq->refcnt++; |
| 1115 | } |
| 1116 | |
| 1117 | /** |
| 1118 | * put_pwq - put a pool_workqueue reference |
| 1119 | * @pwq: pool_workqueue to put |
| 1120 | * |
| 1121 | * Drop a reference of @pwq. If its refcnt reaches zero, schedule its |
| 1122 | * destruction. The caller should be holding the matching pool->lock. |
| 1123 | */ |
| 1124 | static void put_pwq(struct pool_workqueue *pwq) |
| 1125 | { |
| 1126 | lockdep_assert_held(&pwq->pool->lock); |
| 1127 | if (likely(--pwq->refcnt)) |
| 1128 | return; |
| 1129 | if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND))) |
| 1130 | return; |
| 1131 | /* |
| 1132 | * @pwq can't be released under pool->lock, bounce to |
| 1133 | * pwq_unbound_release_workfn(). This never recurses on the same |
| 1134 | * pool->lock as this path is taken only for unbound workqueues and |
| 1135 | * the release work item is scheduled on a per-cpu workqueue. To |
| 1136 | * avoid lockdep warning, unbound pool->locks are given lockdep |
| 1137 | * subclass of 1 in get_unbound_pool(). |
| 1138 | */ |
| 1139 | schedule_work(&pwq->unbound_release_work); |
| 1140 | } |
| 1141 | |
| 1142 | /** |
| 1143 | * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock |
| 1144 | * @pwq: pool_workqueue to put (can be %NULL) |
| 1145 | * |
| 1146 | * put_pwq() with locking. This function also allows %NULL @pwq. |
| 1147 | */ |
| 1148 | static void put_pwq_unlocked(struct pool_workqueue *pwq) |
| 1149 | { |
| 1150 | if (pwq) { |
| 1151 | /* |
| 1152 | * As both pwqs and pools are RCU protected, the |
| 1153 | * following lock operations are safe. |
| 1154 | */ |
| 1155 | raw_spin_lock_irq(&pwq->pool->lock); |
| 1156 | put_pwq(pwq); |
| 1157 | raw_spin_unlock_irq(&pwq->pool->lock); |
| 1158 | } |
| 1159 | } |
| 1160 | |
| 1161 | static void pwq_activate_inactive_work(struct work_struct *work) |
| 1162 | { |
| 1163 | struct pool_workqueue *pwq = get_work_pwq(work); |
| 1164 | |
| 1165 | trace_workqueue_activate_work(work); |
| 1166 | if (list_empty(&pwq->pool->worklist)) |
| 1167 | pwq->pool->watchdog_ts = jiffies; |
| 1168 | move_linked_works(work, &pwq->pool->worklist, NULL); |
| 1169 | __clear_bit(WORK_STRUCT_INACTIVE_BIT, work_data_bits(work)); |
| 1170 | pwq->nr_active++; |
| 1171 | } |
| 1172 | |
| 1173 | static void pwq_activate_first_inactive(struct pool_workqueue *pwq) |
| 1174 | { |
| 1175 | struct work_struct *work = list_first_entry(&pwq->inactive_works, |
| 1176 | struct work_struct, entry); |
| 1177 | |
| 1178 | pwq_activate_inactive_work(work); |
| 1179 | } |
| 1180 | |
| 1181 | /** |
| 1182 | * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight |
| 1183 | * @pwq: pwq of interest |
| 1184 | * @work_data: work_data of work which left the queue |
| 1185 | * |
| 1186 | * A work either has completed or is removed from pending queue, |
| 1187 | * decrement nr_in_flight of its pwq and handle workqueue flushing. |
| 1188 | * |
| 1189 | * CONTEXT: |
| 1190 | * raw_spin_lock_irq(pool->lock). |
| 1191 | */ |
| 1192 | static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, unsigned long work_data) |
| 1193 | { |
| 1194 | int color = get_work_color(work_data); |
| 1195 | |
| 1196 | if (!(work_data & WORK_STRUCT_INACTIVE)) { |
| 1197 | pwq->nr_active--; |
| 1198 | if (!list_empty(&pwq->inactive_works)) { |
| 1199 | /* one down, submit an inactive one */ |
| 1200 | if (pwq->nr_active < pwq->max_active) |
| 1201 | pwq_activate_first_inactive(pwq); |
| 1202 | } |
| 1203 | } |
| 1204 | |
| 1205 | pwq->nr_in_flight[color]--; |
| 1206 | |
| 1207 | /* is flush in progress and are we at the flushing tip? */ |
| 1208 | if (likely(pwq->flush_color != color)) |
| 1209 | goto out_put; |
| 1210 | |
| 1211 | /* are there still in-flight works? */ |
| 1212 | if (pwq->nr_in_flight[color]) |
| 1213 | goto out_put; |
| 1214 | |
| 1215 | /* this pwq is done, clear flush_color */ |
| 1216 | pwq->flush_color = -1; |
| 1217 | |
| 1218 | /* |
| 1219 | * If this was the last pwq, wake up the first flusher. It |
| 1220 | * will handle the rest. |
| 1221 | */ |
| 1222 | if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush)) |
| 1223 | complete(&pwq->wq->first_flusher->done); |
| 1224 | out_put: |
| 1225 | put_pwq(pwq); |
| 1226 | } |
| 1227 | |
| 1228 | /** |
| 1229 | * try_to_grab_pending - steal work item from worklist and disable irq |
| 1230 | * @work: work item to steal |
| 1231 | * @is_dwork: @work is a delayed_work |
| 1232 | * @flags: place to store irq state |
| 1233 | * |
| 1234 | * Try to grab PENDING bit of @work. This function can handle @work in any |
| 1235 | * stable state - idle, on timer or on worklist. |
| 1236 | * |
| 1237 | * Return: |
| 1238 | * |
| 1239 | * ======== ================================================================ |
| 1240 | * 1 if @work was pending and we successfully stole PENDING |
| 1241 | * 0 if @work was idle and we claimed PENDING |
| 1242 | * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry |
| 1243 | * -ENOENT if someone else is canceling @work, this state may persist |
| 1244 | * for arbitrarily long |
| 1245 | * ======== ================================================================ |
| 1246 | * |
| 1247 | * Note: |
| 1248 | * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting |
| 1249 | * interrupted while holding PENDING and @work off queue, irq must be |
| 1250 | * disabled on entry. This, combined with delayed_work->timer being |
| 1251 | * irqsafe, ensures that we return -EAGAIN for finite short period of time. |
| 1252 | * |
| 1253 | * On successful return, >= 0, irq is disabled and the caller is |
| 1254 | * responsible for releasing it using local_irq_restore(*@flags). |
| 1255 | * |
| 1256 | * This function is safe to call from any context including IRQ handler. |
| 1257 | */ |
| 1258 | static int try_to_grab_pending(struct work_struct *work, bool is_dwork, |
| 1259 | unsigned long *flags) |
| 1260 | { |
| 1261 | struct worker_pool *pool; |
| 1262 | struct pool_workqueue *pwq; |
| 1263 | |
| 1264 | local_irq_save(*flags); |
| 1265 | |
| 1266 | /* try to steal the timer if it exists */ |
| 1267 | if (is_dwork) { |
| 1268 | struct delayed_work *dwork = to_delayed_work(work); |
| 1269 | |
| 1270 | /* |
| 1271 | * dwork->timer is irqsafe. If del_timer() fails, it's |
| 1272 | * guaranteed that the timer is not queued anywhere and not |
| 1273 | * running on the local CPU. |
| 1274 | */ |
| 1275 | if (likely(del_timer(&dwork->timer))) |
| 1276 | return 1; |
| 1277 | } |
| 1278 | |
| 1279 | /* try to claim PENDING the normal way */ |
| 1280 | if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) |
| 1281 | return 0; |
| 1282 | |
| 1283 | rcu_read_lock(); |
| 1284 | /* |
| 1285 | * The queueing is in progress, or it is already queued. Try to |
| 1286 | * steal it from ->worklist without clearing WORK_STRUCT_PENDING. |
| 1287 | */ |
| 1288 | pool = get_work_pool(work); |
| 1289 | if (!pool) |
| 1290 | goto fail; |
| 1291 | |
| 1292 | raw_spin_lock(&pool->lock); |
| 1293 | /* |
| 1294 | * work->data is guaranteed to point to pwq only while the work |
| 1295 | * item is queued on pwq->wq, and both updating work->data to point |
| 1296 | * to pwq on queueing and to pool on dequeueing are done under |
| 1297 | * pwq->pool->lock. This in turn guarantees that, if work->data |
| 1298 | * points to pwq which is associated with a locked pool, the work |
| 1299 | * item is currently queued on that pool. |
| 1300 | */ |
| 1301 | pwq = get_work_pwq(work); |
| 1302 | if (pwq && pwq->pool == pool) { |
| 1303 | debug_work_deactivate(work); |
| 1304 | |
| 1305 | /* |
| 1306 | * A cancelable inactive work item must be in the |
| 1307 | * pwq->inactive_works since a queued barrier can't be |
| 1308 | * canceled (see the comments in insert_wq_barrier()). |
| 1309 | * |
| 1310 | * An inactive work item cannot be grabbed directly because |
| 1311 | * it might have linked barrier work items which, if left |
| 1312 | * on the inactive_works list, will confuse pwq->nr_active |
| 1313 | * management later on and cause stall. Make sure the work |
| 1314 | * item is activated before grabbing. |
| 1315 | */ |
| 1316 | if (*work_data_bits(work) & WORK_STRUCT_INACTIVE) |
| 1317 | pwq_activate_inactive_work(work); |
| 1318 | |
| 1319 | list_del_init(&work->entry); |
| 1320 | pwq_dec_nr_in_flight(pwq, *work_data_bits(work)); |
| 1321 | |
| 1322 | /* work->data points to pwq iff queued, point to pool */ |
| 1323 | set_work_pool_and_keep_pending(work, pool->id); |
| 1324 | |
| 1325 | raw_spin_unlock(&pool->lock); |
| 1326 | rcu_read_unlock(); |
| 1327 | return 1; |
| 1328 | } |
| 1329 | raw_spin_unlock(&pool->lock); |
| 1330 | fail: |
| 1331 | rcu_read_unlock(); |
| 1332 | local_irq_restore(*flags); |
| 1333 | if (work_is_canceling(work)) |
| 1334 | return -ENOENT; |
| 1335 | cpu_relax(); |
| 1336 | return -EAGAIN; |
| 1337 | } |
| 1338 | |
| 1339 | /** |
| 1340 | * insert_work - insert a work into a pool |
| 1341 | * @pwq: pwq @work belongs to |
| 1342 | * @work: work to insert |
| 1343 | * @head: insertion point |
| 1344 | * @extra_flags: extra WORK_STRUCT_* flags to set |
| 1345 | * |
| 1346 | * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to |
| 1347 | * work_struct flags. |
| 1348 | * |
| 1349 | * CONTEXT: |
| 1350 | * raw_spin_lock_irq(pool->lock). |
| 1351 | */ |
| 1352 | static void insert_work(struct pool_workqueue *pwq, struct work_struct *work, |
| 1353 | struct list_head *head, unsigned int extra_flags) |
| 1354 | { |
| 1355 | struct worker_pool *pool = pwq->pool; |
| 1356 | |
| 1357 | /* record the work call stack in order to print it in KASAN reports */ |
| 1358 | kasan_record_aux_stack_noalloc(work); |
| 1359 | |
| 1360 | /* we own @work, set data and link */ |
| 1361 | set_work_pwq(work, pwq, extra_flags); |
| 1362 | list_add_tail(&work->entry, head); |
| 1363 | get_pwq(pwq); |
| 1364 | |
| 1365 | if (__need_more_worker(pool)) |
| 1366 | wake_up_worker(pool); |
| 1367 | } |
| 1368 | |
| 1369 | /* |
| 1370 | * Test whether @work is being queued from another work executing on the |
| 1371 | * same workqueue. |
| 1372 | */ |
| 1373 | static bool is_chained_work(struct workqueue_struct *wq) |
| 1374 | { |
| 1375 | struct worker *worker; |
| 1376 | |
| 1377 | worker = current_wq_worker(); |
| 1378 | /* |
| 1379 | * Return %true iff I'm a worker executing a work item on @wq. If |
| 1380 | * I'm @worker, it's safe to dereference it without locking. |
| 1381 | */ |
| 1382 | return worker && worker->current_pwq->wq == wq; |
| 1383 | } |
| 1384 | |
| 1385 | /* |
| 1386 | * When queueing an unbound work item to a wq, prefer local CPU if allowed |
| 1387 | * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to |
| 1388 | * avoid perturbing sensitive tasks. |
| 1389 | */ |
| 1390 | static int wq_select_unbound_cpu(int cpu) |
| 1391 | { |
| 1392 | static bool printed_dbg_warning; |
| 1393 | int new_cpu; |
| 1394 | |
| 1395 | if (likely(!wq_debug_force_rr_cpu)) { |
| 1396 | if (cpumask_test_cpu(cpu, wq_unbound_cpumask)) |
| 1397 | return cpu; |
| 1398 | } else if (!printed_dbg_warning) { |
| 1399 | pr_warn("workqueue: round-robin CPU selection forced, expect performance impact\n"); |
| 1400 | printed_dbg_warning = true; |
| 1401 | } |
| 1402 | |
| 1403 | if (cpumask_empty(wq_unbound_cpumask)) |
| 1404 | return cpu; |
| 1405 | |
| 1406 | new_cpu = __this_cpu_read(wq_rr_cpu_last); |
| 1407 | new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask); |
| 1408 | if (unlikely(new_cpu >= nr_cpu_ids)) { |
| 1409 | new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask); |
| 1410 | if (unlikely(new_cpu >= nr_cpu_ids)) |
| 1411 | return cpu; |
| 1412 | } |
| 1413 | __this_cpu_write(wq_rr_cpu_last, new_cpu); |
| 1414 | |
| 1415 | return new_cpu; |
| 1416 | } |
| 1417 | |
| 1418 | static void __queue_work(int cpu, struct workqueue_struct *wq, |
| 1419 | struct work_struct *work) |
| 1420 | { |
| 1421 | struct pool_workqueue *pwq; |
| 1422 | struct worker_pool *last_pool; |
| 1423 | struct list_head *worklist; |
| 1424 | unsigned int work_flags; |
| 1425 | unsigned int req_cpu = cpu; |
| 1426 | |
| 1427 | /* |
| 1428 | * While a work item is PENDING && off queue, a task trying to |
| 1429 | * steal the PENDING will busy-loop waiting for it to either get |
| 1430 | * queued or lose PENDING. Grabbing PENDING and queueing should |
| 1431 | * happen with IRQ disabled. |
| 1432 | */ |
| 1433 | lockdep_assert_irqs_disabled(); |
| 1434 | |
| 1435 | |
| 1436 | /* if draining, only works from the same workqueue are allowed */ |
| 1437 | if (unlikely(wq->flags & __WQ_DRAINING) && |
| 1438 | WARN_ON_ONCE(!is_chained_work(wq))) |
| 1439 | return; |
| 1440 | rcu_read_lock(); |
| 1441 | retry: |
| 1442 | /* pwq which will be used unless @work is executing elsewhere */ |
| 1443 | if (wq->flags & WQ_UNBOUND) { |
| 1444 | if (req_cpu == WORK_CPU_UNBOUND) |
| 1445 | cpu = wq_select_unbound_cpu(raw_smp_processor_id()); |
| 1446 | pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); |
| 1447 | } else { |
| 1448 | if (req_cpu == WORK_CPU_UNBOUND) |
| 1449 | cpu = raw_smp_processor_id(); |
| 1450 | pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); |
| 1451 | } |
| 1452 | |
| 1453 | /* |
| 1454 | * If @work was previously on a different pool, it might still be |
| 1455 | * running there, in which case the work needs to be queued on that |
| 1456 | * pool to guarantee non-reentrancy. |
| 1457 | */ |
| 1458 | last_pool = get_work_pool(work); |
| 1459 | if (last_pool && last_pool != pwq->pool) { |
| 1460 | struct worker *worker; |
| 1461 | |
| 1462 | raw_spin_lock(&last_pool->lock); |
| 1463 | |
| 1464 | worker = find_worker_executing_work(last_pool, work); |
| 1465 | |
| 1466 | if (worker && worker->current_pwq->wq == wq) { |
| 1467 | pwq = worker->current_pwq; |
| 1468 | } else { |
| 1469 | /* meh... not running there, queue here */ |
| 1470 | raw_spin_unlock(&last_pool->lock); |
| 1471 | raw_spin_lock(&pwq->pool->lock); |
| 1472 | } |
| 1473 | } else { |
| 1474 | raw_spin_lock(&pwq->pool->lock); |
| 1475 | } |
| 1476 | |
| 1477 | /* |
| 1478 | * pwq is determined and locked. For unbound pools, we could have |
| 1479 | * raced with pwq release and it could already be dead. If its |
| 1480 | * refcnt is zero, repeat pwq selection. Note that pwqs never die |
| 1481 | * without another pwq replacing it in the numa_pwq_tbl or while |
| 1482 | * work items are executing on it, so the retrying is guaranteed to |
| 1483 | * make forward-progress. |
| 1484 | */ |
| 1485 | if (unlikely(!pwq->refcnt)) { |
| 1486 | if (wq->flags & WQ_UNBOUND) { |
| 1487 | raw_spin_unlock(&pwq->pool->lock); |
| 1488 | cpu_relax(); |
| 1489 | goto retry; |
| 1490 | } |
| 1491 | /* oops */ |
| 1492 | WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt", |
| 1493 | wq->name, cpu); |
| 1494 | } |
| 1495 | |
| 1496 | /* pwq determined, queue */ |
| 1497 | trace_workqueue_queue_work(req_cpu, pwq, work); |
| 1498 | |
| 1499 | if (WARN_ON(!list_empty(&work->entry))) |
| 1500 | goto out; |
| 1501 | |
| 1502 | pwq->nr_in_flight[pwq->work_color]++; |
| 1503 | work_flags = work_color_to_flags(pwq->work_color); |
| 1504 | |
| 1505 | if (likely(pwq->nr_active < pwq->max_active)) { |
| 1506 | trace_workqueue_activate_work(work); |
| 1507 | pwq->nr_active++; |
| 1508 | worklist = &pwq->pool->worklist; |
| 1509 | if (list_empty(worklist)) |
| 1510 | pwq->pool->watchdog_ts = jiffies; |
| 1511 | } else { |
| 1512 | work_flags |= WORK_STRUCT_INACTIVE; |
| 1513 | worklist = &pwq->inactive_works; |
| 1514 | } |
| 1515 | |
| 1516 | debug_work_activate(work); |
| 1517 | insert_work(pwq, work, worklist, work_flags); |
| 1518 | |
| 1519 | out: |
| 1520 | raw_spin_unlock(&pwq->pool->lock); |
| 1521 | rcu_read_unlock(); |
| 1522 | } |
| 1523 | |
| 1524 | /** |
| 1525 | * queue_work_on - queue work on specific cpu |
| 1526 | * @cpu: CPU number to execute work on |
| 1527 | * @wq: workqueue to use |
| 1528 | * @work: work to queue |
| 1529 | * |
| 1530 | * We queue the work to a specific CPU, the caller must ensure it |
| 1531 | * can't go away. Callers that fail to ensure that the specified |
| 1532 | * CPU cannot go away will execute on a randomly chosen CPU. |
| 1533 | * |
| 1534 | * Return: %false if @work was already on a queue, %true otherwise. |
| 1535 | */ |
| 1536 | bool queue_work_on(int cpu, struct workqueue_struct *wq, |
| 1537 | struct work_struct *work) |
| 1538 | { |
| 1539 | bool ret = false; |
| 1540 | unsigned long flags; |
| 1541 | |
| 1542 | local_irq_save(flags); |
| 1543 | |
| 1544 | if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { |
| 1545 | __queue_work(cpu, wq, work); |
| 1546 | ret = true; |
| 1547 | } |
| 1548 | |
| 1549 | local_irq_restore(flags); |
| 1550 | return ret; |
| 1551 | } |
| 1552 | EXPORT_SYMBOL(queue_work_on); |
| 1553 | |
| 1554 | /** |
| 1555 | * workqueue_select_cpu_near - Select a CPU based on NUMA node |
| 1556 | * @node: NUMA node ID that we want to select a CPU from |
| 1557 | * |
| 1558 | * This function will attempt to find a "random" cpu available on a given |
| 1559 | * node. If there are no CPUs available on the given node it will return |
| 1560 | * WORK_CPU_UNBOUND indicating that we should just schedule to any |
| 1561 | * available CPU if we need to schedule this work. |
| 1562 | */ |
| 1563 | static int workqueue_select_cpu_near(int node) |
| 1564 | { |
| 1565 | int cpu; |
| 1566 | |
| 1567 | /* No point in doing this if NUMA isn't enabled for workqueues */ |
| 1568 | if (!wq_numa_enabled) |
| 1569 | return WORK_CPU_UNBOUND; |
| 1570 | |
| 1571 | /* Delay binding to CPU if node is not valid or online */ |
| 1572 | if (node < 0 || node >= MAX_NUMNODES || !node_online(node)) |
| 1573 | return WORK_CPU_UNBOUND; |
| 1574 | |
| 1575 | /* Use local node/cpu if we are already there */ |
| 1576 | cpu = raw_smp_processor_id(); |
| 1577 | if (node == cpu_to_node(cpu)) |
| 1578 | return cpu; |
| 1579 | |
| 1580 | /* Use "random" otherwise know as "first" online CPU of node */ |
| 1581 | cpu = cpumask_any_and(cpumask_of_node(node), cpu_online_mask); |
| 1582 | |
| 1583 | /* If CPU is valid return that, otherwise just defer */ |
| 1584 | return cpu < nr_cpu_ids ? cpu : WORK_CPU_UNBOUND; |
| 1585 | } |
| 1586 | |
| 1587 | /** |
| 1588 | * queue_work_node - queue work on a "random" cpu for a given NUMA node |
| 1589 | * @node: NUMA node that we are targeting the work for |
| 1590 | * @wq: workqueue to use |
| 1591 | * @work: work to queue |
| 1592 | * |
| 1593 | * We queue the work to a "random" CPU within a given NUMA node. The basic |
| 1594 | * idea here is to provide a way to somehow associate work with a given |
| 1595 | * NUMA node. |
| 1596 | * |
| 1597 | * This function will only make a best effort attempt at getting this onto |
| 1598 | * the right NUMA node. If no node is requested or the requested node is |
| 1599 | * offline then we just fall back to standard queue_work behavior. |
| 1600 | * |
| 1601 | * Currently the "random" CPU ends up being the first available CPU in the |
| 1602 | * intersection of cpu_online_mask and the cpumask of the node, unless we |
| 1603 | * are running on the node. In that case we just use the current CPU. |
| 1604 | * |
| 1605 | * Return: %false if @work was already on a queue, %true otherwise. |
| 1606 | */ |
| 1607 | bool queue_work_node(int node, struct workqueue_struct *wq, |
| 1608 | struct work_struct *work) |
| 1609 | { |
| 1610 | unsigned long flags; |
| 1611 | bool ret = false; |
| 1612 | |
| 1613 | /* |
| 1614 | * This current implementation is specific to unbound workqueues. |
| 1615 | * Specifically we only return the first available CPU for a given |
| 1616 | * node instead of cycling through individual CPUs within the node. |
| 1617 | * |
| 1618 | * If this is used with a per-cpu workqueue then the logic in |
| 1619 | * workqueue_select_cpu_near would need to be updated to allow for |
| 1620 | * some round robin type logic. |
| 1621 | */ |
| 1622 | WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)); |
| 1623 | |
| 1624 | local_irq_save(flags); |
| 1625 | |
| 1626 | if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { |
| 1627 | int cpu = workqueue_select_cpu_near(node); |
| 1628 | |
| 1629 | __queue_work(cpu, wq, work); |
| 1630 | ret = true; |
| 1631 | } |
| 1632 | |
| 1633 | local_irq_restore(flags); |
| 1634 | return ret; |
| 1635 | } |
| 1636 | EXPORT_SYMBOL_GPL(queue_work_node); |
| 1637 | |
| 1638 | void delayed_work_timer_fn(struct timer_list *t) |
| 1639 | { |
| 1640 | struct delayed_work *dwork = from_timer(dwork, t, timer); |
| 1641 | |
| 1642 | /* should have been called from irqsafe timer with irq already off */ |
| 1643 | __queue_work(dwork->cpu, dwork->wq, &dwork->work); |
| 1644 | } |
| 1645 | EXPORT_SYMBOL(delayed_work_timer_fn); |
| 1646 | |
| 1647 | static void __queue_delayed_work(int cpu, struct workqueue_struct *wq, |
| 1648 | struct delayed_work *dwork, unsigned long delay) |
| 1649 | { |
| 1650 | struct timer_list *timer = &dwork->timer; |
| 1651 | struct work_struct *work = &dwork->work; |
| 1652 | |
| 1653 | WARN_ON_ONCE(!wq); |
| 1654 | WARN_ON_FUNCTION_MISMATCH(timer->function, delayed_work_timer_fn); |
| 1655 | WARN_ON_ONCE(timer_pending(timer)); |
| 1656 | WARN_ON_ONCE(!list_empty(&work->entry)); |
| 1657 | |
| 1658 | /* |
| 1659 | * If @delay is 0, queue @dwork->work immediately. This is for |
| 1660 | * both optimization and correctness. The earliest @timer can |
| 1661 | * expire is on the closest next tick and delayed_work users depend |
| 1662 | * on that there's no such delay when @delay is 0. |
| 1663 | */ |
| 1664 | if (!delay) { |
| 1665 | __queue_work(cpu, wq, &dwork->work); |
| 1666 | return; |
| 1667 | } |
| 1668 | |
| 1669 | dwork->wq = wq; |
| 1670 | dwork->cpu = cpu; |
| 1671 | timer->expires = jiffies + delay; |
| 1672 | |
| 1673 | if (unlikely(cpu != WORK_CPU_UNBOUND)) |
| 1674 | add_timer_on(timer, cpu); |
| 1675 | else |
| 1676 | add_timer(timer); |
| 1677 | } |
| 1678 | |
| 1679 | /** |
| 1680 | * queue_delayed_work_on - queue work on specific CPU after delay |
| 1681 | * @cpu: CPU number to execute work on |
| 1682 | * @wq: workqueue to use |
| 1683 | * @dwork: work to queue |
| 1684 | * @delay: number of jiffies to wait before queueing |
| 1685 | * |
| 1686 | * Return: %false if @work was already on a queue, %true otherwise. If |
| 1687 | * @delay is zero and @dwork is idle, it will be scheduled for immediate |
| 1688 | * execution. |
| 1689 | */ |
| 1690 | bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq, |
| 1691 | struct delayed_work *dwork, unsigned long delay) |
| 1692 | { |
| 1693 | struct work_struct *work = &dwork->work; |
| 1694 | bool ret = false; |
| 1695 | unsigned long flags; |
| 1696 | |
| 1697 | /* read the comment in __queue_work() */ |
| 1698 | local_irq_save(flags); |
| 1699 | |
| 1700 | if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { |
| 1701 | __queue_delayed_work(cpu, wq, dwork, delay); |
| 1702 | ret = true; |
| 1703 | } |
| 1704 | |
| 1705 | local_irq_restore(flags); |
| 1706 | return ret; |
| 1707 | } |
| 1708 | EXPORT_SYMBOL(queue_delayed_work_on); |
| 1709 | |
| 1710 | /** |
| 1711 | * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU |
| 1712 | * @cpu: CPU number to execute work on |
| 1713 | * @wq: workqueue to use |
| 1714 | * @dwork: work to queue |
| 1715 | * @delay: number of jiffies to wait before queueing |
| 1716 | * |
| 1717 | * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise, |
| 1718 | * modify @dwork's timer so that it expires after @delay. If @delay is |
| 1719 | * zero, @work is guaranteed to be scheduled immediately regardless of its |
| 1720 | * current state. |
| 1721 | * |
| 1722 | * Return: %false if @dwork was idle and queued, %true if @dwork was |
| 1723 | * pending and its timer was modified. |
| 1724 | * |
| 1725 | * This function is safe to call from any context including IRQ handler. |
| 1726 | * See try_to_grab_pending() for details. |
| 1727 | */ |
| 1728 | bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq, |
| 1729 | struct delayed_work *dwork, unsigned long delay) |
| 1730 | { |
| 1731 | unsigned long flags; |
| 1732 | int ret; |
| 1733 | |
| 1734 | do { |
| 1735 | ret = try_to_grab_pending(&dwork->work, true, &flags); |
| 1736 | } while (unlikely(ret == -EAGAIN)); |
| 1737 | |
| 1738 | if (likely(ret >= 0)) { |
| 1739 | __queue_delayed_work(cpu, wq, dwork, delay); |
| 1740 | local_irq_restore(flags); |
| 1741 | } |
| 1742 | |
| 1743 | /* -ENOENT from try_to_grab_pending() becomes %true */ |
| 1744 | return ret; |
| 1745 | } |
| 1746 | EXPORT_SYMBOL_GPL(mod_delayed_work_on); |
| 1747 | |
| 1748 | static void rcu_work_rcufn(struct rcu_head *rcu) |
| 1749 | { |
| 1750 | struct rcu_work *rwork = container_of(rcu, struct rcu_work, rcu); |
| 1751 | |
| 1752 | /* read the comment in __queue_work() */ |
| 1753 | local_irq_disable(); |
| 1754 | __queue_work(WORK_CPU_UNBOUND, rwork->wq, &rwork->work); |
| 1755 | local_irq_enable(); |
| 1756 | } |
| 1757 | |
| 1758 | /** |
| 1759 | * queue_rcu_work - queue work after a RCU grace period |
| 1760 | * @wq: workqueue to use |
| 1761 | * @rwork: work to queue |
| 1762 | * |
| 1763 | * Return: %false if @rwork was already pending, %true otherwise. Note |
| 1764 | * that a full RCU grace period is guaranteed only after a %true return. |
| 1765 | * While @rwork is guaranteed to be executed after a %false return, the |
| 1766 | * execution may happen before a full RCU grace period has passed. |
| 1767 | */ |
| 1768 | bool queue_rcu_work(struct workqueue_struct *wq, struct rcu_work *rwork) |
| 1769 | { |
| 1770 | struct work_struct *work = &rwork->work; |
| 1771 | |
| 1772 | if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { |
| 1773 | rwork->wq = wq; |
| 1774 | call_rcu(&rwork->rcu, rcu_work_rcufn); |
| 1775 | return true; |
| 1776 | } |
| 1777 | |
| 1778 | return false; |
| 1779 | } |
| 1780 | EXPORT_SYMBOL(queue_rcu_work); |
| 1781 | |
| 1782 | /** |
| 1783 | * worker_enter_idle - enter idle state |
| 1784 | * @worker: worker which is entering idle state |
| 1785 | * |
| 1786 | * @worker is entering idle state. Update stats and idle timer if |
| 1787 | * necessary. |
| 1788 | * |
| 1789 | * LOCKING: |
| 1790 | * raw_spin_lock_irq(pool->lock). |
| 1791 | */ |
| 1792 | static void worker_enter_idle(struct worker *worker) |
| 1793 | { |
| 1794 | struct worker_pool *pool = worker->pool; |
| 1795 | |
| 1796 | if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) || |
| 1797 | WARN_ON_ONCE(!list_empty(&worker->entry) && |
| 1798 | (worker->hentry.next || worker->hentry.pprev))) |
| 1799 | return; |
| 1800 | |
| 1801 | /* can't use worker_set_flags(), also called from create_worker() */ |
| 1802 | worker->flags |= WORKER_IDLE; |
| 1803 | pool->nr_idle++; |
| 1804 | worker->last_active = jiffies; |
| 1805 | |
| 1806 | /* idle_list is LIFO */ |
| 1807 | list_add(&worker->entry, &pool->idle_list); |
| 1808 | |
| 1809 | if (too_many_workers(pool) && !timer_pending(&pool->idle_timer)) |
| 1810 | mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT); |
| 1811 | |
| 1812 | /* Sanity check nr_running. */ |
| 1813 | WARN_ON_ONCE(pool->nr_workers == pool->nr_idle && pool->nr_running); |
| 1814 | } |
| 1815 | |
| 1816 | /** |
| 1817 | * worker_leave_idle - leave idle state |
| 1818 | * @worker: worker which is leaving idle state |
| 1819 | * |
| 1820 | * @worker is leaving idle state. Update stats. |
| 1821 | * |
| 1822 | * LOCKING: |
| 1823 | * raw_spin_lock_irq(pool->lock). |
| 1824 | */ |
| 1825 | static void worker_leave_idle(struct worker *worker) |
| 1826 | { |
| 1827 | struct worker_pool *pool = worker->pool; |
| 1828 | |
| 1829 | if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE))) |
| 1830 | return; |
| 1831 | worker_clr_flags(worker, WORKER_IDLE); |
| 1832 | pool->nr_idle--; |
| 1833 | list_del_init(&worker->entry); |
| 1834 | } |
| 1835 | |
| 1836 | static struct worker *alloc_worker(int node) |
| 1837 | { |
| 1838 | struct worker *worker; |
| 1839 | |
| 1840 | worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node); |
| 1841 | if (worker) { |
| 1842 | INIT_LIST_HEAD(&worker->entry); |
| 1843 | INIT_LIST_HEAD(&worker->scheduled); |
| 1844 | INIT_LIST_HEAD(&worker->node); |
| 1845 | /* on creation a worker is in !idle && prep state */ |
| 1846 | worker->flags = WORKER_PREP; |
| 1847 | } |
| 1848 | return worker; |
| 1849 | } |
| 1850 | |
| 1851 | /** |
| 1852 | * worker_attach_to_pool() - attach a worker to a pool |
| 1853 | * @worker: worker to be attached |
| 1854 | * @pool: the target pool |
| 1855 | * |
| 1856 | * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and |
| 1857 | * cpu-binding of @worker are kept coordinated with the pool across |
| 1858 | * cpu-[un]hotplugs. |
| 1859 | */ |
| 1860 | static void worker_attach_to_pool(struct worker *worker, |
| 1861 | struct worker_pool *pool) |
| 1862 | { |
| 1863 | mutex_lock(&wq_pool_attach_mutex); |
| 1864 | |
| 1865 | /* |
| 1866 | * The wq_pool_attach_mutex ensures %POOL_DISASSOCIATED remains |
| 1867 | * stable across this function. See the comments above the flag |
| 1868 | * definition for details. |
| 1869 | */ |
| 1870 | if (pool->flags & POOL_DISASSOCIATED) |
| 1871 | worker->flags |= WORKER_UNBOUND; |
| 1872 | else |
| 1873 | kthread_set_per_cpu(worker->task, pool->cpu); |
| 1874 | |
| 1875 | if (worker->rescue_wq) |
| 1876 | set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask); |
| 1877 | |
| 1878 | list_add_tail(&worker->node, &pool->workers); |
| 1879 | worker->pool = pool; |
| 1880 | |
| 1881 | mutex_unlock(&wq_pool_attach_mutex); |
| 1882 | } |
| 1883 | |
| 1884 | /** |
| 1885 | * worker_detach_from_pool() - detach a worker from its pool |
| 1886 | * @worker: worker which is attached to its pool |
| 1887 | * |
| 1888 | * Undo the attaching which had been done in worker_attach_to_pool(). The |
| 1889 | * caller worker shouldn't access to the pool after detached except it has |
| 1890 | * other reference to the pool. |
| 1891 | */ |
| 1892 | static void worker_detach_from_pool(struct worker *worker) |
| 1893 | { |
| 1894 | struct worker_pool *pool = worker->pool; |
| 1895 | struct completion *detach_completion = NULL; |
| 1896 | |
| 1897 | mutex_lock(&wq_pool_attach_mutex); |
| 1898 | |
| 1899 | kthread_set_per_cpu(worker->task, -1); |
| 1900 | list_del(&worker->node); |
| 1901 | worker->pool = NULL; |
| 1902 | |
| 1903 | if (list_empty(&pool->workers)) |
| 1904 | detach_completion = pool->detach_completion; |
| 1905 | mutex_unlock(&wq_pool_attach_mutex); |
| 1906 | |
| 1907 | /* clear leftover flags without pool->lock after it is detached */ |
| 1908 | worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND); |
| 1909 | |
| 1910 | if (detach_completion) |
| 1911 | complete(detach_completion); |
| 1912 | } |
| 1913 | |
| 1914 | /** |
| 1915 | * create_worker - create a new workqueue worker |
| 1916 | * @pool: pool the new worker will belong to |
| 1917 | * |
| 1918 | * Create and start a new worker which is attached to @pool. |
| 1919 | * |
| 1920 | * CONTEXT: |
| 1921 | * Might sleep. Does GFP_KERNEL allocations. |
| 1922 | * |
| 1923 | * Return: |
| 1924 | * Pointer to the newly created worker. |
| 1925 | */ |
| 1926 | static struct worker *create_worker(struct worker_pool *pool) |
| 1927 | { |
| 1928 | struct worker *worker; |
| 1929 | int id; |
| 1930 | char id_buf[16]; |
| 1931 | |
| 1932 | /* ID is needed to determine kthread name */ |
| 1933 | id = ida_alloc(&pool->worker_ida, GFP_KERNEL); |
| 1934 | if (id < 0) |
| 1935 | return NULL; |
| 1936 | |
| 1937 | worker = alloc_worker(pool->node); |
| 1938 | if (!worker) |
| 1939 | goto fail; |
| 1940 | |
| 1941 | worker->id = id; |
| 1942 | |
| 1943 | if (pool->cpu >= 0) |
| 1944 | snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id, |
| 1945 | pool->attrs->nice < 0 ? "H" : ""); |
| 1946 | else |
| 1947 | snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id); |
| 1948 | |
| 1949 | worker->task = kthread_create_on_node(worker_thread, worker, pool->node, |
| 1950 | "kworker/%s", id_buf); |
| 1951 | if (IS_ERR(worker->task)) |
| 1952 | goto fail; |
| 1953 | |
| 1954 | set_user_nice(worker->task, pool->attrs->nice); |
| 1955 | kthread_bind_mask(worker->task, pool->attrs->cpumask); |
| 1956 | |
| 1957 | /* successful, attach the worker to the pool */ |
| 1958 | worker_attach_to_pool(worker, pool); |
| 1959 | |
| 1960 | /* start the newly created worker */ |
| 1961 | raw_spin_lock_irq(&pool->lock); |
| 1962 | worker->pool->nr_workers++; |
| 1963 | worker_enter_idle(worker); |
| 1964 | wake_up_process(worker->task); |
| 1965 | raw_spin_unlock_irq(&pool->lock); |
| 1966 | |
| 1967 | return worker; |
| 1968 | |
| 1969 | fail: |
| 1970 | ida_free(&pool->worker_ida, id); |
| 1971 | kfree(worker); |
| 1972 | return NULL; |
| 1973 | } |
| 1974 | |
| 1975 | /** |
| 1976 | * destroy_worker - destroy a workqueue worker |
| 1977 | * @worker: worker to be destroyed |
| 1978 | * |
| 1979 | * Destroy @worker and adjust @pool stats accordingly. The worker should |
| 1980 | * be idle. |
| 1981 | * |
| 1982 | * CONTEXT: |
| 1983 | * raw_spin_lock_irq(pool->lock). |
| 1984 | */ |
| 1985 | static void destroy_worker(struct worker *worker) |
| 1986 | { |
| 1987 | struct worker_pool *pool = worker->pool; |
| 1988 | |
| 1989 | lockdep_assert_held(&pool->lock); |
| 1990 | |
| 1991 | /* sanity check frenzy */ |
| 1992 | if (WARN_ON(worker->current_work) || |
| 1993 | WARN_ON(!list_empty(&worker->scheduled)) || |
| 1994 | WARN_ON(!(worker->flags & WORKER_IDLE))) |
| 1995 | return; |
| 1996 | |
| 1997 | pool->nr_workers--; |
| 1998 | pool->nr_idle--; |
| 1999 | |
| 2000 | list_del_init(&worker->entry); |
| 2001 | worker->flags |= WORKER_DIE; |
| 2002 | wake_up_process(worker->task); |
| 2003 | } |
| 2004 | |
| 2005 | static void idle_worker_timeout(struct timer_list *t) |
| 2006 | { |
| 2007 | struct worker_pool *pool = from_timer(pool, t, idle_timer); |
| 2008 | |
| 2009 | raw_spin_lock_irq(&pool->lock); |
| 2010 | |
| 2011 | while (too_many_workers(pool)) { |
| 2012 | struct worker *worker; |
| 2013 | unsigned long expires; |
| 2014 | |
| 2015 | /* idle_list is kept in LIFO order, check the last one */ |
| 2016 | worker = list_entry(pool->idle_list.prev, struct worker, entry); |
| 2017 | expires = worker->last_active + IDLE_WORKER_TIMEOUT; |
| 2018 | |
| 2019 | if (time_before(jiffies, expires)) { |
| 2020 | mod_timer(&pool->idle_timer, expires); |
| 2021 | break; |
| 2022 | } |
| 2023 | |
| 2024 | destroy_worker(worker); |
| 2025 | } |
| 2026 | |
| 2027 | raw_spin_unlock_irq(&pool->lock); |
| 2028 | } |
| 2029 | |
| 2030 | static void send_mayday(struct work_struct *work) |
| 2031 | { |
| 2032 | struct pool_workqueue *pwq = get_work_pwq(work); |
| 2033 | struct workqueue_struct *wq = pwq->wq; |
| 2034 | |
| 2035 | lockdep_assert_held(&wq_mayday_lock); |
| 2036 | |
| 2037 | if (!wq->rescuer) |
| 2038 | return; |
| 2039 | |
| 2040 | /* mayday mayday mayday */ |
| 2041 | if (list_empty(&pwq->mayday_node)) { |
| 2042 | /* |
| 2043 | * If @pwq is for an unbound wq, its base ref may be put at |
| 2044 | * any time due to an attribute change. Pin @pwq until the |
| 2045 | * rescuer is done with it. |
| 2046 | */ |
| 2047 | get_pwq(pwq); |
| 2048 | list_add_tail(&pwq->mayday_node, &wq->maydays); |
| 2049 | wake_up_process(wq->rescuer->task); |
| 2050 | } |
| 2051 | } |
| 2052 | |
| 2053 | static void pool_mayday_timeout(struct timer_list *t) |
| 2054 | { |
| 2055 | struct worker_pool *pool = from_timer(pool, t, mayday_timer); |
| 2056 | struct work_struct *work; |
| 2057 | |
| 2058 | raw_spin_lock_irq(&pool->lock); |
| 2059 | raw_spin_lock(&wq_mayday_lock); /* for wq->maydays */ |
| 2060 | |
| 2061 | if (need_to_create_worker(pool)) { |
| 2062 | /* |
| 2063 | * We've been trying to create a new worker but |
| 2064 | * haven't been successful. We might be hitting an |
| 2065 | * allocation deadlock. Send distress signals to |
| 2066 | * rescuers. |
| 2067 | */ |
| 2068 | list_for_each_entry(work, &pool->worklist, entry) |
| 2069 | send_mayday(work); |
| 2070 | } |
| 2071 | |
| 2072 | raw_spin_unlock(&wq_mayday_lock); |
| 2073 | raw_spin_unlock_irq(&pool->lock); |
| 2074 | |
| 2075 | mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL); |
| 2076 | } |
| 2077 | |
| 2078 | /** |
| 2079 | * maybe_create_worker - create a new worker if necessary |
| 2080 | * @pool: pool to create a new worker for |
| 2081 | * |
| 2082 | * Create a new worker for @pool if necessary. @pool is guaranteed to |
| 2083 | * have at least one idle worker on return from this function. If |
| 2084 | * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is |
| 2085 | * sent to all rescuers with works scheduled on @pool to resolve |
| 2086 | * possible allocation deadlock. |
| 2087 | * |
| 2088 | * On return, need_to_create_worker() is guaranteed to be %false and |
| 2089 | * may_start_working() %true. |
| 2090 | * |
| 2091 | * LOCKING: |
| 2092 | * raw_spin_lock_irq(pool->lock) which may be released and regrabbed |
| 2093 | * multiple times. Does GFP_KERNEL allocations. Called only from |
| 2094 | * manager. |
| 2095 | */ |
| 2096 | static void maybe_create_worker(struct worker_pool *pool) |
| 2097 | __releases(&pool->lock) |
| 2098 | __acquires(&pool->lock) |
| 2099 | { |
| 2100 | restart: |
| 2101 | raw_spin_unlock_irq(&pool->lock); |
| 2102 | |
| 2103 | /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */ |
| 2104 | mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT); |
| 2105 | |
| 2106 | while (true) { |
| 2107 | if (create_worker(pool) || !need_to_create_worker(pool)) |
| 2108 | break; |
| 2109 | |
| 2110 | schedule_timeout_interruptible(CREATE_COOLDOWN); |
| 2111 | |
| 2112 | if (!need_to_create_worker(pool)) |
| 2113 | break; |
| 2114 | } |
| 2115 | |
| 2116 | del_timer_sync(&pool->mayday_timer); |
| 2117 | raw_spin_lock_irq(&pool->lock); |
| 2118 | /* |
| 2119 | * This is necessary even after a new worker was just successfully |
| 2120 | * created as @pool->lock was dropped and the new worker might have |
| 2121 | * already become busy. |
| 2122 | */ |
| 2123 | if (need_to_create_worker(pool)) |
| 2124 | goto restart; |
| 2125 | } |
| 2126 | |
| 2127 | /** |
| 2128 | * manage_workers - manage worker pool |
| 2129 | * @worker: self |
| 2130 | * |
| 2131 | * Assume the manager role and manage the worker pool @worker belongs |
| 2132 | * to. At any given time, there can be only zero or one manager per |
| 2133 | * pool. The exclusion is handled automatically by this function. |
| 2134 | * |
| 2135 | * The caller can safely start processing works on false return. On |
| 2136 | * true return, it's guaranteed that need_to_create_worker() is false |
| 2137 | * and may_start_working() is true. |
| 2138 | * |
| 2139 | * CONTEXT: |
| 2140 | * raw_spin_lock_irq(pool->lock) which may be released and regrabbed |
| 2141 | * multiple times. Does GFP_KERNEL allocations. |
| 2142 | * |
| 2143 | * Return: |
| 2144 | * %false if the pool doesn't need management and the caller can safely |
| 2145 | * start processing works, %true if management function was performed and |
| 2146 | * the conditions that the caller verified before calling the function may |
| 2147 | * no longer be true. |
| 2148 | */ |
| 2149 | static bool manage_workers(struct worker *worker) |
| 2150 | { |
| 2151 | struct worker_pool *pool = worker->pool; |
| 2152 | |
| 2153 | if (pool->flags & POOL_MANAGER_ACTIVE) |
| 2154 | return false; |
| 2155 | |
| 2156 | pool->flags |= POOL_MANAGER_ACTIVE; |
| 2157 | pool->manager = worker; |
| 2158 | |
| 2159 | maybe_create_worker(pool); |
| 2160 | |
| 2161 | pool->manager = NULL; |
| 2162 | pool->flags &= ~POOL_MANAGER_ACTIVE; |
| 2163 | rcuwait_wake_up(&manager_wait); |
| 2164 | return true; |
| 2165 | } |
| 2166 | |
| 2167 | /** |
| 2168 | * process_one_work - process single work |
| 2169 | * @worker: self |
| 2170 | * @work: work to process |
| 2171 | * |
| 2172 | * Process @work. This function contains all the logics necessary to |
| 2173 | * process a single work including synchronization against and |
| 2174 | * interaction with other workers on the same cpu, queueing and |
| 2175 | * flushing. As long as context requirement is met, any worker can |
| 2176 | * call this function to process a work. |
| 2177 | * |
| 2178 | * CONTEXT: |
| 2179 | * raw_spin_lock_irq(pool->lock) which is released and regrabbed. |
| 2180 | */ |
| 2181 | static void process_one_work(struct worker *worker, struct work_struct *work) |
| 2182 | __releases(&pool->lock) |
| 2183 | __acquires(&pool->lock) |
| 2184 | { |
| 2185 | struct pool_workqueue *pwq = get_work_pwq(work); |
| 2186 | struct worker_pool *pool = worker->pool; |
| 2187 | bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE; |
| 2188 | unsigned long work_data; |
| 2189 | struct worker *collision; |
| 2190 | #ifdef CONFIG_LOCKDEP |
| 2191 | /* |
| 2192 | * It is permissible to free the struct work_struct from |
| 2193 | * inside the function that is called from it, this we need to |
| 2194 | * take into account for lockdep too. To avoid bogus "held |
| 2195 | * lock freed" warnings as well as problems when looking into |
| 2196 | * work->lockdep_map, make a copy and use that here. |
| 2197 | */ |
| 2198 | struct lockdep_map lockdep_map; |
| 2199 | |
| 2200 | lockdep_copy_map(&lockdep_map, &work->lockdep_map); |
| 2201 | #endif |
| 2202 | /* ensure we're on the correct CPU */ |
| 2203 | WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && |
| 2204 | raw_smp_processor_id() != pool->cpu); |
| 2205 | |
| 2206 | /* |
| 2207 | * A single work shouldn't be executed concurrently by |
| 2208 | * multiple workers on a single cpu. Check whether anyone is |
| 2209 | * already processing the work. If so, defer the work to the |
| 2210 | * currently executing one. |
| 2211 | */ |
| 2212 | collision = find_worker_executing_work(pool, work); |
| 2213 | if (unlikely(collision)) { |
| 2214 | move_linked_works(work, &collision->scheduled, NULL); |
| 2215 | return; |
| 2216 | } |
| 2217 | |
| 2218 | /* claim and dequeue */ |
| 2219 | debug_work_deactivate(work); |
| 2220 | hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work); |
| 2221 | worker->current_work = work; |
| 2222 | worker->current_func = work->func; |
| 2223 | worker->current_pwq = pwq; |
| 2224 | work_data = *work_data_bits(work); |
| 2225 | worker->current_color = get_work_color(work_data); |
| 2226 | |
| 2227 | /* |
| 2228 | * Record wq name for cmdline and debug reporting, may get |
| 2229 | * overridden through set_worker_desc(). |
| 2230 | */ |
| 2231 | strscpy(worker->desc, pwq->wq->name, WORKER_DESC_LEN); |
| 2232 | |
| 2233 | list_del_init(&work->entry); |
| 2234 | |
| 2235 | /* |
| 2236 | * CPU intensive works don't participate in concurrency management. |
| 2237 | * They're the scheduler's responsibility. This takes @worker out |
| 2238 | * of concurrency management and the next code block will chain |
| 2239 | * execution of the pending work items. |
| 2240 | */ |
| 2241 | if (unlikely(cpu_intensive)) |
| 2242 | worker_set_flags(worker, WORKER_CPU_INTENSIVE); |
| 2243 | |
| 2244 | /* |
| 2245 | * Wake up another worker if necessary. The condition is always |
| 2246 | * false for normal per-cpu workers since nr_running would always |
| 2247 | * be >= 1 at this point. This is used to chain execution of the |
| 2248 | * pending work items for WORKER_NOT_RUNNING workers such as the |
| 2249 | * UNBOUND and CPU_INTENSIVE ones. |
| 2250 | */ |
| 2251 | if (need_more_worker(pool)) |
| 2252 | wake_up_worker(pool); |
| 2253 | |
| 2254 | /* |
| 2255 | * Record the last pool and clear PENDING which should be the last |
| 2256 | * update to @work. Also, do this inside @pool->lock so that |
| 2257 | * PENDING and queued state changes happen together while IRQ is |
| 2258 | * disabled. |
| 2259 | */ |
| 2260 | set_work_pool_and_clear_pending(work, pool->id); |
| 2261 | |
| 2262 | raw_spin_unlock_irq(&pool->lock); |
| 2263 | |
| 2264 | lock_map_acquire(&pwq->wq->lockdep_map); |
| 2265 | lock_map_acquire(&lockdep_map); |
| 2266 | /* |
| 2267 | * Strictly speaking we should mark the invariant state without holding |
| 2268 | * any locks, that is, before these two lock_map_acquire()'s. |
| 2269 | * |
| 2270 | * However, that would result in: |
| 2271 | * |
| 2272 | * A(W1) |
| 2273 | * WFC(C) |
| 2274 | * A(W1) |
| 2275 | * C(C) |
| 2276 | * |
| 2277 | * Which would create W1->C->W1 dependencies, even though there is no |
| 2278 | * actual deadlock possible. There are two solutions, using a |
| 2279 | * read-recursive acquire on the work(queue) 'locks', but this will then |
| 2280 | * hit the lockdep limitation on recursive locks, or simply discard |
| 2281 | * these locks. |
| 2282 | * |
| 2283 | * AFAICT there is no possible deadlock scenario between the |
| 2284 | * flush_work() and complete() primitives (except for single-threaded |
| 2285 | * workqueues), so hiding them isn't a problem. |
| 2286 | */ |
| 2287 | lockdep_invariant_state(true); |
| 2288 | trace_workqueue_execute_start(work); |
| 2289 | worker->current_func(work); |
| 2290 | /* |
| 2291 | * While we must be careful to not use "work" after this, the trace |
| 2292 | * point will only record its address. |
| 2293 | */ |
| 2294 | trace_workqueue_execute_end(work, worker->current_func); |
| 2295 | lock_map_release(&lockdep_map); |
| 2296 | lock_map_release(&pwq->wq->lockdep_map); |
| 2297 | |
| 2298 | if (unlikely(in_atomic() || lockdep_depth(current) > 0)) { |
| 2299 | pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n" |
| 2300 | " last function: %ps\n", |
| 2301 | current->comm, preempt_count(), task_pid_nr(current), |
| 2302 | worker->current_func); |
| 2303 | debug_show_held_locks(current); |
| 2304 | dump_stack(); |
| 2305 | } |
| 2306 | |
| 2307 | /* |
| 2308 | * The following prevents a kworker from hogging CPU on !PREEMPTION |
| 2309 | * kernels, where a requeueing work item waiting for something to |
| 2310 | * happen could deadlock with stop_machine as such work item could |
| 2311 | * indefinitely requeue itself while all other CPUs are trapped in |
| 2312 | * stop_machine. At the same time, report a quiescent RCU state so |
| 2313 | * the same condition doesn't freeze RCU. |
| 2314 | */ |
| 2315 | cond_resched(); |
| 2316 | |
| 2317 | raw_spin_lock_irq(&pool->lock); |
| 2318 | |
| 2319 | /* clear cpu intensive status */ |
| 2320 | if (unlikely(cpu_intensive)) |
| 2321 | worker_clr_flags(worker, WORKER_CPU_INTENSIVE); |
| 2322 | |
| 2323 | /* tag the worker for identification in schedule() */ |
| 2324 | worker->last_func = worker->current_func; |
| 2325 | |
| 2326 | /* we're done with it, release */ |
| 2327 | hash_del(&worker->hentry); |
| 2328 | worker->current_work = NULL; |
| 2329 | worker->current_func = NULL; |
| 2330 | worker->current_pwq = NULL; |
| 2331 | worker->current_color = INT_MAX; |
| 2332 | pwq_dec_nr_in_flight(pwq, work_data); |
| 2333 | } |
| 2334 | |
| 2335 | /** |
| 2336 | * process_scheduled_works - process scheduled works |
| 2337 | * @worker: self |
| 2338 | * |
| 2339 | * Process all scheduled works. Please note that the scheduled list |
| 2340 | * may change while processing a work, so this function repeatedly |
| 2341 | * fetches a work from the top and executes it. |
| 2342 | * |
| 2343 | * CONTEXT: |
| 2344 | * raw_spin_lock_irq(pool->lock) which may be released and regrabbed |
| 2345 | * multiple times. |
| 2346 | */ |
| 2347 | static void process_scheduled_works(struct worker *worker) |
| 2348 | { |
| 2349 | while (!list_empty(&worker->scheduled)) { |
| 2350 | struct work_struct *work = list_first_entry(&worker->scheduled, |
| 2351 | struct work_struct, entry); |
| 2352 | process_one_work(worker, work); |
| 2353 | } |
| 2354 | } |
| 2355 | |
| 2356 | static void set_pf_worker(bool val) |
| 2357 | { |
| 2358 | mutex_lock(&wq_pool_attach_mutex); |
| 2359 | if (val) |
| 2360 | current->flags |= PF_WQ_WORKER; |
| 2361 | else |
| 2362 | current->flags &= ~PF_WQ_WORKER; |
| 2363 | mutex_unlock(&wq_pool_attach_mutex); |
| 2364 | } |
| 2365 | |
| 2366 | /** |
| 2367 | * worker_thread - the worker thread function |
| 2368 | * @__worker: self |
| 2369 | * |
| 2370 | * The worker thread function. All workers belong to a worker_pool - |
| 2371 | * either a per-cpu one or dynamic unbound one. These workers process all |
| 2372 | * work items regardless of their specific target workqueue. The only |
| 2373 | * exception is work items which belong to workqueues with a rescuer which |
| 2374 | * will be explained in rescuer_thread(). |
| 2375 | * |
| 2376 | * Return: 0 |
| 2377 | */ |
| 2378 | static int worker_thread(void *__worker) |
| 2379 | { |
| 2380 | struct worker *worker = __worker; |
| 2381 | struct worker_pool *pool = worker->pool; |
| 2382 | |
| 2383 | /* tell the scheduler that this is a workqueue worker */ |
| 2384 | set_pf_worker(true); |
| 2385 | woke_up: |
| 2386 | raw_spin_lock_irq(&pool->lock); |
| 2387 | |
| 2388 | /* am I supposed to die? */ |
| 2389 | if (unlikely(worker->flags & WORKER_DIE)) { |
| 2390 | raw_spin_unlock_irq(&pool->lock); |
| 2391 | WARN_ON_ONCE(!list_empty(&worker->entry)); |
| 2392 | set_pf_worker(false); |
| 2393 | |
| 2394 | set_task_comm(worker->task, "kworker/dying"); |
| 2395 | ida_free(&pool->worker_ida, worker->id); |
| 2396 | worker_detach_from_pool(worker); |
| 2397 | kfree(worker); |
| 2398 | return 0; |
| 2399 | } |
| 2400 | |
| 2401 | worker_leave_idle(worker); |
| 2402 | recheck: |
| 2403 | /* no more worker necessary? */ |
| 2404 | if (!need_more_worker(pool)) |
| 2405 | goto sleep; |
| 2406 | |
| 2407 | /* do we need to manage? */ |
| 2408 | if (unlikely(!may_start_working(pool)) && manage_workers(worker)) |
| 2409 | goto recheck; |
| 2410 | |
| 2411 | /* |
| 2412 | * ->scheduled list can only be filled while a worker is |
| 2413 | * preparing to process a work or actually processing it. |
| 2414 | * Make sure nobody diddled with it while I was sleeping. |
| 2415 | */ |
| 2416 | WARN_ON_ONCE(!list_empty(&worker->scheduled)); |
| 2417 | |
| 2418 | /* |
| 2419 | * Finish PREP stage. We're guaranteed to have at least one idle |
| 2420 | * worker or that someone else has already assumed the manager |
| 2421 | * role. This is where @worker starts participating in concurrency |
| 2422 | * management if applicable and concurrency management is restored |
| 2423 | * after being rebound. See rebind_workers() for details. |
| 2424 | */ |
| 2425 | worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND); |
| 2426 | |
| 2427 | do { |
| 2428 | struct work_struct *work = |
| 2429 | list_first_entry(&pool->worklist, |
| 2430 | struct work_struct, entry); |
| 2431 | |
| 2432 | pool->watchdog_ts = jiffies; |
| 2433 | |
| 2434 | if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) { |
| 2435 | /* optimization path, not strictly necessary */ |
| 2436 | process_one_work(worker, work); |
| 2437 | if (unlikely(!list_empty(&worker->scheduled))) |
| 2438 | process_scheduled_works(worker); |
| 2439 | } else { |
| 2440 | move_linked_works(work, &worker->scheduled, NULL); |
| 2441 | process_scheduled_works(worker); |
| 2442 | } |
| 2443 | } while (keep_working(pool)); |
| 2444 | |
| 2445 | worker_set_flags(worker, WORKER_PREP); |
| 2446 | sleep: |
| 2447 | /* |
| 2448 | * pool->lock is held and there's no work to process and no need to |
| 2449 | * manage, sleep. Workers are woken up only while holding |
| 2450 | * pool->lock or from local cpu, so setting the current state |
| 2451 | * before releasing pool->lock is enough to prevent losing any |
| 2452 | * event. |
| 2453 | */ |
| 2454 | worker_enter_idle(worker); |
| 2455 | __set_current_state(TASK_IDLE); |
| 2456 | raw_spin_unlock_irq(&pool->lock); |
| 2457 | schedule(); |
| 2458 | goto woke_up; |
| 2459 | } |
| 2460 | |
| 2461 | /** |
| 2462 | * rescuer_thread - the rescuer thread function |
| 2463 | * @__rescuer: self |
| 2464 | * |
| 2465 | * Workqueue rescuer thread function. There's one rescuer for each |
| 2466 | * workqueue which has WQ_MEM_RECLAIM set. |
| 2467 | * |
| 2468 | * Regular work processing on a pool may block trying to create a new |
| 2469 | * worker which uses GFP_KERNEL allocation which has slight chance of |
| 2470 | * developing into deadlock if some works currently on the same queue |
| 2471 | * need to be processed to satisfy the GFP_KERNEL allocation. This is |
| 2472 | * the problem rescuer solves. |
| 2473 | * |
| 2474 | * When such condition is possible, the pool summons rescuers of all |
| 2475 | * workqueues which have works queued on the pool and let them process |
| 2476 | * those works so that forward progress can be guaranteed. |
| 2477 | * |
| 2478 | * This should happen rarely. |
| 2479 | * |
| 2480 | * Return: 0 |
| 2481 | */ |
| 2482 | static int rescuer_thread(void *__rescuer) |
| 2483 | { |
| 2484 | struct worker *rescuer = __rescuer; |
| 2485 | struct workqueue_struct *wq = rescuer->rescue_wq; |
| 2486 | struct list_head *scheduled = &rescuer->scheduled; |
| 2487 | bool should_stop; |
| 2488 | |
| 2489 | set_user_nice(current, RESCUER_NICE_LEVEL); |
| 2490 | |
| 2491 | /* |
| 2492 | * Mark rescuer as worker too. As WORKER_PREP is never cleared, it |
| 2493 | * doesn't participate in concurrency management. |
| 2494 | */ |
| 2495 | set_pf_worker(true); |
| 2496 | repeat: |
| 2497 | set_current_state(TASK_IDLE); |
| 2498 | |
| 2499 | /* |
| 2500 | * By the time the rescuer is requested to stop, the workqueue |
| 2501 | * shouldn't have any work pending, but @wq->maydays may still have |
| 2502 | * pwq(s) queued. This can happen by non-rescuer workers consuming |
| 2503 | * all the work items before the rescuer got to them. Go through |
| 2504 | * @wq->maydays processing before acting on should_stop so that the |
| 2505 | * list is always empty on exit. |
| 2506 | */ |
| 2507 | should_stop = kthread_should_stop(); |
| 2508 | |
| 2509 | /* see whether any pwq is asking for help */ |
| 2510 | raw_spin_lock_irq(&wq_mayday_lock); |
| 2511 | |
| 2512 | while (!list_empty(&wq->maydays)) { |
| 2513 | struct pool_workqueue *pwq = list_first_entry(&wq->maydays, |
| 2514 | struct pool_workqueue, mayday_node); |
| 2515 | struct worker_pool *pool = pwq->pool; |
| 2516 | struct work_struct *work, *n; |
| 2517 | bool first = true; |
| 2518 | |
| 2519 | __set_current_state(TASK_RUNNING); |
| 2520 | list_del_init(&pwq->mayday_node); |
| 2521 | |
| 2522 | raw_spin_unlock_irq(&wq_mayday_lock); |
| 2523 | |
| 2524 | worker_attach_to_pool(rescuer, pool); |
| 2525 | |
| 2526 | raw_spin_lock_irq(&pool->lock); |
| 2527 | |
| 2528 | /* |
| 2529 | * Slurp in all works issued via this workqueue and |
| 2530 | * process'em. |
| 2531 | */ |
| 2532 | WARN_ON_ONCE(!list_empty(scheduled)); |
| 2533 | list_for_each_entry_safe(work, n, &pool->worklist, entry) { |
| 2534 | if (get_work_pwq(work) == pwq) { |
| 2535 | if (first) |
| 2536 | pool->watchdog_ts = jiffies; |
| 2537 | move_linked_works(work, scheduled, &n); |
| 2538 | } |
| 2539 | first = false; |
| 2540 | } |
| 2541 | |
| 2542 | if (!list_empty(scheduled)) { |
| 2543 | process_scheduled_works(rescuer); |
| 2544 | |
| 2545 | /* |
| 2546 | * The above execution of rescued work items could |
| 2547 | * have created more to rescue through |
| 2548 | * pwq_activate_first_inactive() or chained |
| 2549 | * queueing. Let's put @pwq back on mayday list so |
| 2550 | * that such back-to-back work items, which may be |
| 2551 | * being used to relieve memory pressure, don't |
| 2552 | * incur MAYDAY_INTERVAL delay inbetween. |
| 2553 | */ |
| 2554 | if (pwq->nr_active && need_to_create_worker(pool)) { |
| 2555 | raw_spin_lock(&wq_mayday_lock); |
| 2556 | /* |
| 2557 | * Queue iff we aren't racing destruction |
| 2558 | * and somebody else hasn't queued it already. |
| 2559 | */ |
| 2560 | if (wq->rescuer && list_empty(&pwq->mayday_node)) { |
| 2561 | get_pwq(pwq); |
| 2562 | list_add_tail(&pwq->mayday_node, &wq->maydays); |
| 2563 | } |
| 2564 | raw_spin_unlock(&wq_mayday_lock); |
| 2565 | } |
| 2566 | } |
| 2567 | |
| 2568 | /* |
| 2569 | * Put the reference grabbed by send_mayday(). @pool won't |
| 2570 | * go away while we're still attached to it. |
| 2571 | */ |
| 2572 | put_pwq(pwq); |
| 2573 | |
| 2574 | /* |
| 2575 | * Leave this pool. If need_more_worker() is %true, notify a |
| 2576 | * regular worker; otherwise, we end up with 0 concurrency |
| 2577 | * and stalling the execution. |
| 2578 | */ |
| 2579 | if (need_more_worker(pool)) |
| 2580 | wake_up_worker(pool); |
| 2581 | |
| 2582 | raw_spin_unlock_irq(&pool->lock); |
| 2583 | |
| 2584 | worker_detach_from_pool(rescuer); |
| 2585 | |
| 2586 | raw_spin_lock_irq(&wq_mayday_lock); |
| 2587 | } |
| 2588 | |
| 2589 | raw_spin_unlock_irq(&wq_mayday_lock); |
| 2590 | |
| 2591 | if (should_stop) { |
| 2592 | __set_current_state(TASK_RUNNING); |
| 2593 | set_pf_worker(false); |
| 2594 | return 0; |
| 2595 | } |
| 2596 | |
| 2597 | /* rescuers should never participate in concurrency management */ |
| 2598 | WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING)); |
| 2599 | schedule(); |
| 2600 | goto repeat; |
| 2601 | } |
| 2602 | |
| 2603 | /** |
| 2604 | * check_flush_dependency - check for flush dependency sanity |
| 2605 | * @target_wq: workqueue being flushed |
| 2606 | * @target_work: work item being flushed (NULL for workqueue flushes) |
| 2607 | * |
| 2608 | * %current is trying to flush the whole @target_wq or @target_work on it. |
| 2609 | * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not |
| 2610 | * reclaiming memory or running on a workqueue which doesn't have |
| 2611 | * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to |
| 2612 | * a deadlock. |
| 2613 | */ |
| 2614 | static void check_flush_dependency(struct workqueue_struct *target_wq, |
| 2615 | struct work_struct *target_work) |
| 2616 | { |
| 2617 | work_func_t target_func = target_work ? target_work->func : NULL; |
| 2618 | struct worker *worker; |
| 2619 | |
| 2620 | if (target_wq->flags & WQ_MEM_RECLAIM) |
| 2621 | return; |
| 2622 | |
| 2623 | worker = current_wq_worker(); |
| 2624 | |
| 2625 | WARN_ONCE(current->flags & PF_MEMALLOC, |
| 2626 | "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%ps", |
| 2627 | current->pid, current->comm, target_wq->name, target_func); |
| 2628 | WARN_ONCE(worker && ((worker->current_pwq->wq->flags & |
| 2629 | (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM), |
| 2630 | "workqueue: WQ_MEM_RECLAIM %s:%ps is flushing !WQ_MEM_RECLAIM %s:%ps", |
| 2631 | worker->current_pwq->wq->name, worker->current_func, |
| 2632 | target_wq->name, target_func); |
| 2633 | } |
| 2634 | |
| 2635 | struct wq_barrier { |
| 2636 | struct work_struct work; |
| 2637 | struct completion done; |
| 2638 | struct task_struct *task; /* purely informational */ |
| 2639 | }; |
| 2640 | |
| 2641 | static void wq_barrier_func(struct work_struct *work) |
| 2642 | { |
| 2643 | struct wq_barrier *barr = container_of(work, struct wq_barrier, work); |
| 2644 | complete(&barr->done); |
| 2645 | } |
| 2646 | |
| 2647 | /** |
| 2648 | * insert_wq_barrier - insert a barrier work |
| 2649 | * @pwq: pwq to insert barrier into |
| 2650 | * @barr: wq_barrier to insert |
| 2651 | * @target: target work to attach @barr to |
| 2652 | * @worker: worker currently executing @target, NULL if @target is not executing |
| 2653 | * |
| 2654 | * @barr is linked to @target such that @barr is completed only after |
| 2655 | * @target finishes execution. Please note that the ordering |
| 2656 | * guarantee is observed only with respect to @target and on the local |
| 2657 | * cpu. |
| 2658 | * |
| 2659 | * Currently, a queued barrier can't be canceled. This is because |
| 2660 | * try_to_grab_pending() can't determine whether the work to be |
| 2661 | * grabbed is at the head of the queue and thus can't clear LINKED |
| 2662 | * flag of the previous work while there must be a valid next work |
| 2663 | * after a work with LINKED flag set. |
| 2664 | * |
| 2665 | * Note that when @worker is non-NULL, @target may be modified |
| 2666 | * underneath us, so we can't reliably determine pwq from @target. |
| 2667 | * |
| 2668 | * CONTEXT: |
| 2669 | * raw_spin_lock_irq(pool->lock). |
| 2670 | */ |
| 2671 | static void insert_wq_barrier(struct pool_workqueue *pwq, |
| 2672 | struct wq_barrier *barr, |
| 2673 | struct work_struct *target, struct worker *worker) |
| 2674 | { |
| 2675 | unsigned int work_flags = 0; |
| 2676 | unsigned int work_color; |
| 2677 | struct list_head *head; |
| 2678 | |
| 2679 | /* |
| 2680 | * debugobject calls are safe here even with pool->lock locked |
| 2681 | * as we know for sure that this will not trigger any of the |
| 2682 | * checks and call back into the fixup functions where we |
| 2683 | * might deadlock. |
| 2684 | */ |
| 2685 | INIT_WORK_ONSTACK(&barr->work, wq_barrier_func); |
| 2686 | __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work)); |
| 2687 | |
| 2688 | init_completion_map(&barr->done, &target->lockdep_map); |
| 2689 | |
| 2690 | barr->task = current; |
| 2691 | |
| 2692 | /* The barrier work item does not participate in pwq->nr_active. */ |
| 2693 | work_flags |= WORK_STRUCT_INACTIVE; |
| 2694 | |
| 2695 | /* |
| 2696 | * If @target is currently being executed, schedule the |
| 2697 | * barrier to the worker; otherwise, put it after @target. |
| 2698 | */ |
| 2699 | if (worker) { |
| 2700 | head = worker->scheduled.next; |
| 2701 | work_color = worker->current_color; |
| 2702 | } else { |
| 2703 | unsigned long *bits = work_data_bits(target); |
| 2704 | |
| 2705 | head = target->entry.next; |
| 2706 | /* there can already be other linked works, inherit and set */ |
| 2707 | work_flags |= *bits & WORK_STRUCT_LINKED; |
| 2708 | work_color = get_work_color(*bits); |
| 2709 | __set_bit(WORK_STRUCT_LINKED_BIT, bits); |
| 2710 | } |
| 2711 | |
| 2712 | pwq->nr_in_flight[work_color]++; |
| 2713 | work_flags |= work_color_to_flags(work_color); |
| 2714 | |
| 2715 | debug_work_activate(&barr->work); |
| 2716 | insert_work(pwq, &barr->work, head, work_flags); |
| 2717 | } |
| 2718 | |
| 2719 | /** |
| 2720 | * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing |
| 2721 | * @wq: workqueue being flushed |
| 2722 | * @flush_color: new flush color, < 0 for no-op |
| 2723 | * @work_color: new work color, < 0 for no-op |
| 2724 | * |
| 2725 | * Prepare pwqs for workqueue flushing. |
| 2726 | * |
| 2727 | * If @flush_color is non-negative, flush_color on all pwqs should be |
| 2728 | * -1. If no pwq has in-flight commands at the specified color, all |
| 2729 | * pwq->flush_color's stay at -1 and %false is returned. If any pwq |
| 2730 | * has in flight commands, its pwq->flush_color is set to |
| 2731 | * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq |
| 2732 | * wakeup logic is armed and %true is returned. |
| 2733 | * |
| 2734 | * The caller should have initialized @wq->first_flusher prior to |
| 2735 | * calling this function with non-negative @flush_color. If |
| 2736 | * @flush_color is negative, no flush color update is done and %false |
| 2737 | * is returned. |
| 2738 | * |
| 2739 | * If @work_color is non-negative, all pwqs should have the same |
| 2740 | * work_color which is previous to @work_color and all will be |
| 2741 | * advanced to @work_color. |
| 2742 | * |
| 2743 | * CONTEXT: |
| 2744 | * mutex_lock(wq->mutex). |
| 2745 | * |
| 2746 | * Return: |
| 2747 | * %true if @flush_color >= 0 and there's something to flush. %false |
| 2748 | * otherwise. |
| 2749 | */ |
| 2750 | static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq, |
| 2751 | int flush_color, int work_color) |
| 2752 | { |
| 2753 | bool wait = false; |
| 2754 | struct pool_workqueue *pwq; |
| 2755 | |
| 2756 | if (flush_color >= 0) { |
| 2757 | WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush)); |
| 2758 | atomic_set(&wq->nr_pwqs_to_flush, 1); |
| 2759 | } |
| 2760 | |
| 2761 | for_each_pwq(pwq, wq) { |
| 2762 | struct worker_pool *pool = pwq->pool; |
| 2763 | |
| 2764 | raw_spin_lock_irq(&pool->lock); |
| 2765 | |
| 2766 | if (flush_color >= 0) { |
| 2767 | WARN_ON_ONCE(pwq->flush_color != -1); |
| 2768 | |
| 2769 | if (pwq->nr_in_flight[flush_color]) { |
| 2770 | pwq->flush_color = flush_color; |
| 2771 | atomic_inc(&wq->nr_pwqs_to_flush); |
| 2772 | wait = true; |
| 2773 | } |
| 2774 | } |
| 2775 | |
| 2776 | if (work_color >= 0) { |
| 2777 | WARN_ON_ONCE(work_color != work_next_color(pwq->work_color)); |
| 2778 | pwq->work_color = work_color; |
| 2779 | } |
| 2780 | |
| 2781 | raw_spin_unlock_irq(&pool->lock); |
| 2782 | } |
| 2783 | |
| 2784 | if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush)) |
| 2785 | complete(&wq->first_flusher->done); |
| 2786 | |
| 2787 | return wait; |
| 2788 | } |
| 2789 | |
| 2790 | /** |
| 2791 | * __flush_workqueue - ensure that any scheduled work has run to completion. |
| 2792 | * @wq: workqueue to flush |
| 2793 | * |
| 2794 | * This function sleeps until all work items which were queued on entry |
| 2795 | * have finished execution, but it is not livelocked by new incoming ones. |
| 2796 | */ |
| 2797 | void __flush_workqueue(struct workqueue_struct *wq) |
| 2798 | { |
| 2799 | struct wq_flusher this_flusher = { |
| 2800 | .list = LIST_HEAD_INIT(this_flusher.list), |
| 2801 | .flush_color = -1, |
| 2802 | .done = COMPLETION_INITIALIZER_ONSTACK_MAP(this_flusher.done, wq->lockdep_map), |
| 2803 | }; |
| 2804 | int next_color; |
| 2805 | |
| 2806 | if (WARN_ON(!wq_online)) |
| 2807 | return; |
| 2808 | |
| 2809 | lock_map_acquire(&wq->lockdep_map); |
| 2810 | lock_map_release(&wq->lockdep_map); |
| 2811 | |
| 2812 | mutex_lock(&wq->mutex); |
| 2813 | |
| 2814 | /* |
| 2815 | * Start-to-wait phase |
| 2816 | */ |
| 2817 | next_color = work_next_color(wq->work_color); |
| 2818 | |
| 2819 | if (next_color != wq->flush_color) { |
| 2820 | /* |
| 2821 | * Color space is not full. The current work_color |
| 2822 | * becomes our flush_color and work_color is advanced |
| 2823 | * by one. |
| 2824 | */ |
| 2825 | WARN_ON_ONCE(!list_empty(&wq->flusher_overflow)); |
| 2826 | this_flusher.flush_color = wq->work_color; |
| 2827 | wq->work_color = next_color; |
| 2828 | |
| 2829 | if (!wq->first_flusher) { |
| 2830 | /* no flush in progress, become the first flusher */ |
| 2831 | WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); |
| 2832 | |
| 2833 | wq->first_flusher = &this_flusher; |
| 2834 | |
| 2835 | if (!flush_workqueue_prep_pwqs(wq, wq->flush_color, |
| 2836 | wq->work_color)) { |
| 2837 | /* nothing to flush, done */ |
| 2838 | wq->flush_color = next_color; |
| 2839 | wq->first_flusher = NULL; |
| 2840 | goto out_unlock; |
| 2841 | } |
| 2842 | } else { |
| 2843 | /* wait in queue */ |
| 2844 | WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color); |
| 2845 | list_add_tail(&this_flusher.list, &wq->flusher_queue); |
| 2846 | flush_workqueue_prep_pwqs(wq, -1, wq->work_color); |
| 2847 | } |
| 2848 | } else { |
| 2849 | /* |
| 2850 | * Oops, color space is full, wait on overflow queue. |
| 2851 | * The next flush completion will assign us |
| 2852 | * flush_color and transfer to flusher_queue. |
| 2853 | */ |
| 2854 | list_add_tail(&this_flusher.list, &wq->flusher_overflow); |
| 2855 | } |
| 2856 | |
| 2857 | check_flush_dependency(wq, NULL); |
| 2858 | |
| 2859 | mutex_unlock(&wq->mutex); |
| 2860 | |
| 2861 | wait_for_completion(&this_flusher.done); |
| 2862 | |
| 2863 | /* |
| 2864 | * Wake-up-and-cascade phase |
| 2865 | * |
| 2866 | * First flushers are responsible for cascading flushes and |
| 2867 | * handling overflow. Non-first flushers can simply return. |
| 2868 | */ |
| 2869 | if (READ_ONCE(wq->first_flusher) != &this_flusher) |
| 2870 | return; |
| 2871 | |
| 2872 | mutex_lock(&wq->mutex); |
| 2873 | |
| 2874 | /* we might have raced, check again with mutex held */ |
| 2875 | if (wq->first_flusher != &this_flusher) |
| 2876 | goto out_unlock; |
| 2877 | |
| 2878 | WRITE_ONCE(wq->first_flusher, NULL); |
| 2879 | |
| 2880 | WARN_ON_ONCE(!list_empty(&this_flusher.list)); |
| 2881 | WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); |
| 2882 | |
| 2883 | while (true) { |
| 2884 | struct wq_flusher *next, *tmp; |
| 2885 | |
| 2886 | /* complete all the flushers sharing the current flush color */ |
| 2887 | list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) { |
| 2888 | if (next->flush_color != wq->flush_color) |
| 2889 | break; |
| 2890 | list_del_init(&next->list); |
| 2891 | complete(&next->done); |
| 2892 | } |
| 2893 | |
| 2894 | WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) && |
| 2895 | wq->flush_color != work_next_color(wq->work_color)); |
| 2896 | |
| 2897 | /* this flush_color is finished, advance by one */ |
| 2898 | wq->flush_color = work_next_color(wq->flush_color); |
| 2899 | |
| 2900 | /* one color has been freed, handle overflow queue */ |
| 2901 | if (!list_empty(&wq->flusher_overflow)) { |
| 2902 | /* |
| 2903 | * Assign the same color to all overflowed |
| 2904 | * flushers, advance work_color and append to |
| 2905 | * flusher_queue. This is the start-to-wait |
| 2906 | * phase for these overflowed flushers. |
| 2907 | */ |
| 2908 | list_for_each_entry(tmp, &wq->flusher_overflow, list) |
| 2909 | tmp->flush_color = wq->work_color; |
| 2910 | |
| 2911 | wq->work_color = work_next_color(wq->work_color); |
| 2912 | |
| 2913 | list_splice_tail_init(&wq->flusher_overflow, |
| 2914 | &wq->flusher_queue); |
| 2915 | flush_workqueue_prep_pwqs(wq, -1, wq->work_color); |
| 2916 | } |
| 2917 | |
| 2918 | if (list_empty(&wq->flusher_queue)) { |
| 2919 | WARN_ON_ONCE(wq->flush_color != wq->work_color); |
| 2920 | break; |
| 2921 | } |
| 2922 | |
| 2923 | /* |
| 2924 | * Need to flush more colors. Make the next flusher |
| 2925 | * the new first flusher and arm pwqs. |
| 2926 | */ |
| 2927 | WARN_ON_ONCE(wq->flush_color == wq->work_color); |
| 2928 | WARN_ON_ONCE(wq->flush_color != next->flush_color); |
| 2929 | |
| 2930 | list_del_init(&next->list); |
| 2931 | wq->first_flusher = next; |
| 2932 | |
| 2933 | if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1)) |
| 2934 | break; |
| 2935 | |
| 2936 | /* |
| 2937 | * Meh... this color is already done, clear first |
| 2938 | * flusher and repeat cascading. |
| 2939 | */ |
| 2940 | wq->first_flusher = NULL; |
| 2941 | } |
| 2942 | |
| 2943 | out_unlock: |
| 2944 | mutex_unlock(&wq->mutex); |
| 2945 | } |
| 2946 | EXPORT_SYMBOL(__flush_workqueue); |
| 2947 | |
| 2948 | /** |
| 2949 | * drain_workqueue - drain a workqueue |
| 2950 | * @wq: workqueue to drain |
| 2951 | * |
| 2952 | * Wait until the workqueue becomes empty. While draining is in progress, |
| 2953 | * only chain queueing is allowed. IOW, only currently pending or running |
| 2954 | * work items on @wq can queue further work items on it. @wq is flushed |
| 2955 | * repeatedly until it becomes empty. The number of flushing is determined |
| 2956 | * by the depth of chaining and should be relatively short. Whine if it |
| 2957 | * takes too long. |
| 2958 | */ |
| 2959 | void drain_workqueue(struct workqueue_struct *wq) |
| 2960 | { |
| 2961 | unsigned int flush_cnt = 0; |
| 2962 | struct pool_workqueue *pwq; |
| 2963 | |
| 2964 | /* |
| 2965 | * __queue_work() needs to test whether there are drainers, is much |
| 2966 | * hotter than drain_workqueue() and already looks at @wq->flags. |
| 2967 | * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers. |
| 2968 | */ |
| 2969 | mutex_lock(&wq->mutex); |
| 2970 | if (!wq->nr_drainers++) |
| 2971 | wq->flags |= __WQ_DRAINING; |
| 2972 | mutex_unlock(&wq->mutex); |
| 2973 | reflush: |
| 2974 | __flush_workqueue(wq); |
| 2975 | |
| 2976 | mutex_lock(&wq->mutex); |
| 2977 | |
| 2978 | for_each_pwq(pwq, wq) { |
| 2979 | bool drained; |
| 2980 | |
| 2981 | raw_spin_lock_irq(&pwq->pool->lock); |
| 2982 | drained = !pwq->nr_active && list_empty(&pwq->inactive_works); |
| 2983 | raw_spin_unlock_irq(&pwq->pool->lock); |
| 2984 | |
| 2985 | if (drained) |
| 2986 | continue; |
| 2987 | |
| 2988 | if (++flush_cnt == 10 || |
| 2989 | (flush_cnt % 100 == 0 && flush_cnt <= 1000)) |
| 2990 | pr_warn("workqueue %s: %s() isn't complete after %u tries\n", |
| 2991 | wq->name, __func__, flush_cnt); |
| 2992 | |
| 2993 | mutex_unlock(&wq->mutex); |
| 2994 | goto reflush; |
| 2995 | } |
| 2996 | |
| 2997 | if (!--wq->nr_drainers) |
| 2998 | wq->flags &= ~__WQ_DRAINING; |
| 2999 | mutex_unlock(&wq->mutex); |
| 3000 | } |
| 3001 | EXPORT_SYMBOL_GPL(drain_workqueue); |
| 3002 | |
| 3003 | static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr, |
| 3004 | bool from_cancel) |
| 3005 | { |
| 3006 | struct worker *worker = NULL; |
| 3007 | struct worker_pool *pool; |
| 3008 | struct pool_workqueue *pwq; |
| 3009 | |
| 3010 | might_sleep(); |
| 3011 | |
| 3012 | rcu_read_lock(); |
| 3013 | pool = get_work_pool(work); |
| 3014 | if (!pool) { |
| 3015 | rcu_read_unlock(); |
| 3016 | return false; |
| 3017 | } |
| 3018 | |
| 3019 | raw_spin_lock_irq(&pool->lock); |
| 3020 | /* see the comment in try_to_grab_pending() with the same code */ |
| 3021 | pwq = get_work_pwq(work); |
| 3022 | if (pwq) { |
| 3023 | if (unlikely(pwq->pool != pool)) |
| 3024 | goto already_gone; |
| 3025 | } else { |
| 3026 | worker = find_worker_executing_work(pool, work); |
| 3027 | if (!worker) |
| 3028 | goto already_gone; |
| 3029 | pwq = worker->current_pwq; |
| 3030 | } |
| 3031 | |
| 3032 | check_flush_dependency(pwq->wq, work); |
| 3033 | |
| 3034 | insert_wq_barrier(pwq, barr, work, worker); |
| 3035 | raw_spin_unlock_irq(&pool->lock); |
| 3036 | |
| 3037 | /* |
| 3038 | * Force a lock recursion deadlock when using flush_work() inside a |
| 3039 | * single-threaded or rescuer equipped workqueue. |
| 3040 | * |
| 3041 | * For single threaded workqueues the deadlock happens when the work |
| 3042 | * is after the work issuing the flush_work(). For rescuer equipped |
| 3043 | * workqueues the deadlock happens when the rescuer stalls, blocking |
| 3044 | * forward progress. |
| 3045 | */ |
| 3046 | if (!from_cancel && |
| 3047 | (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)) { |
| 3048 | lock_map_acquire(&pwq->wq->lockdep_map); |
| 3049 | lock_map_release(&pwq->wq->lockdep_map); |
| 3050 | } |
| 3051 | rcu_read_unlock(); |
| 3052 | return true; |
| 3053 | already_gone: |
| 3054 | raw_spin_unlock_irq(&pool->lock); |
| 3055 | rcu_read_unlock(); |
| 3056 | return false; |
| 3057 | } |
| 3058 | |
| 3059 | static bool __flush_work(struct work_struct *work, bool from_cancel) |
| 3060 | { |
| 3061 | struct wq_barrier barr; |
| 3062 | |
| 3063 | if (WARN_ON(!wq_online)) |
| 3064 | return false; |
| 3065 | |
| 3066 | if (WARN_ON(!work->func)) |
| 3067 | return false; |
| 3068 | |
| 3069 | if (!from_cancel) { |
| 3070 | lock_map_acquire(&work->lockdep_map); |
| 3071 | lock_map_release(&work->lockdep_map); |
| 3072 | } |
| 3073 | |
| 3074 | if (start_flush_work(work, &barr, from_cancel)) { |
| 3075 | wait_for_completion(&barr.done); |
| 3076 | destroy_work_on_stack(&barr.work); |
| 3077 | return true; |
| 3078 | } else { |
| 3079 | return false; |
| 3080 | } |
| 3081 | } |
| 3082 | |
| 3083 | /** |
| 3084 | * flush_work - wait for a work to finish executing the last queueing instance |
| 3085 | * @work: the work to flush |
| 3086 | * |
| 3087 | * Wait until @work has finished execution. @work is guaranteed to be idle |
| 3088 | * on return if it hasn't been requeued since flush started. |
| 3089 | * |
| 3090 | * Return: |
| 3091 | * %true if flush_work() waited for the work to finish execution, |
| 3092 | * %false if it was already idle. |
| 3093 | */ |
| 3094 | bool flush_work(struct work_struct *work) |
| 3095 | { |
| 3096 | return __flush_work(work, false); |
| 3097 | } |
| 3098 | EXPORT_SYMBOL_GPL(flush_work); |
| 3099 | |
| 3100 | struct cwt_wait { |
| 3101 | wait_queue_entry_t wait; |
| 3102 | struct work_struct *work; |
| 3103 | }; |
| 3104 | |
| 3105 | static int cwt_wakefn(wait_queue_entry_t *wait, unsigned mode, int sync, void *key) |
| 3106 | { |
| 3107 | struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait); |
| 3108 | |
| 3109 | if (cwait->work != key) |
| 3110 | return 0; |
| 3111 | return autoremove_wake_function(wait, mode, sync, key); |
| 3112 | } |
| 3113 | |
| 3114 | static bool __cancel_work_timer(struct work_struct *work, bool is_dwork) |
| 3115 | { |
| 3116 | static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq); |
| 3117 | unsigned long flags; |
| 3118 | int ret; |
| 3119 | |
| 3120 | do { |
| 3121 | ret = try_to_grab_pending(work, is_dwork, &flags); |
| 3122 | /* |
| 3123 | * If someone else is already canceling, wait for it to |
| 3124 | * finish. flush_work() doesn't work for PREEMPT_NONE |
| 3125 | * because we may get scheduled between @work's completion |
| 3126 | * and the other canceling task resuming and clearing |
| 3127 | * CANCELING - flush_work() will return false immediately |
| 3128 | * as @work is no longer busy, try_to_grab_pending() will |
| 3129 | * return -ENOENT as @work is still being canceled and the |
| 3130 | * other canceling task won't be able to clear CANCELING as |
| 3131 | * we're hogging the CPU. |
| 3132 | * |
| 3133 | * Let's wait for completion using a waitqueue. As this |
| 3134 | * may lead to the thundering herd problem, use a custom |
| 3135 | * wake function which matches @work along with exclusive |
| 3136 | * wait and wakeup. |
| 3137 | */ |
| 3138 | if (unlikely(ret == -ENOENT)) { |
| 3139 | struct cwt_wait cwait; |
| 3140 | |
| 3141 | init_wait(&cwait.wait); |
| 3142 | cwait.wait.func = cwt_wakefn; |
| 3143 | cwait.work = work; |
| 3144 | |
| 3145 | prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait, |
| 3146 | TASK_UNINTERRUPTIBLE); |
| 3147 | if (work_is_canceling(work)) |
| 3148 | schedule(); |
| 3149 | finish_wait(&cancel_waitq, &cwait.wait); |
| 3150 | } |
| 3151 | } while (unlikely(ret < 0)); |
| 3152 | |
| 3153 | /* tell other tasks trying to grab @work to back off */ |
| 3154 | mark_work_canceling(work); |
| 3155 | local_irq_restore(flags); |
| 3156 | |
| 3157 | /* |
| 3158 | * This allows canceling during early boot. We know that @work |
| 3159 | * isn't executing. |
| 3160 | */ |
| 3161 | if (wq_online) |
| 3162 | __flush_work(work, true); |
| 3163 | |
| 3164 | clear_work_data(work); |
| 3165 | |
| 3166 | /* |
| 3167 | * Paired with prepare_to_wait() above so that either |
| 3168 | * waitqueue_active() is visible here or !work_is_canceling() is |
| 3169 | * visible there. |
| 3170 | */ |
| 3171 | smp_mb(); |
| 3172 | if (waitqueue_active(&cancel_waitq)) |
| 3173 | __wake_up(&cancel_waitq, TASK_NORMAL, 1, work); |
| 3174 | |
| 3175 | return ret; |
| 3176 | } |
| 3177 | |
| 3178 | /** |
| 3179 | * cancel_work_sync - cancel a work and wait for it to finish |
| 3180 | * @work: the work to cancel |
| 3181 | * |
| 3182 | * Cancel @work and wait for its execution to finish. This function |
| 3183 | * can be used even if the work re-queues itself or migrates to |
| 3184 | * another workqueue. On return from this function, @work is |
| 3185 | * guaranteed to be not pending or executing on any CPU. |
| 3186 | * |
| 3187 | * cancel_work_sync(&delayed_work->work) must not be used for |
| 3188 | * delayed_work's. Use cancel_delayed_work_sync() instead. |
| 3189 | * |
| 3190 | * The caller must ensure that the workqueue on which @work was last |
| 3191 | * queued can't be destroyed before this function returns. |
| 3192 | * |
| 3193 | * Return: |
| 3194 | * %true if @work was pending, %false otherwise. |
| 3195 | */ |
| 3196 | bool cancel_work_sync(struct work_struct *work) |
| 3197 | { |
| 3198 | return __cancel_work_timer(work, false); |
| 3199 | } |
| 3200 | EXPORT_SYMBOL_GPL(cancel_work_sync); |
| 3201 | |
| 3202 | /** |
| 3203 | * flush_delayed_work - wait for a dwork to finish executing the last queueing |
| 3204 | * @dwork: the delayed work to flush |
| 3205 | * |
| 3206 | * Delayed timer is cancelled and the pending work is queued for |
| 3207 | * immediate execution. Like flush_work(), this function only |
| 3208 | * considers the last queueing instance of @dwork. |
| 3209 | * |
| 3210 | * Return: |
| 3211 | * %true if flush_work() waited for the work to finish execution, |
| 3212 | * %false if it was already idle. |
| 3213 | */ |
| 3214 | bool flush_delayed_work(struct delayed_work *dwork) |
| 3215 | { |
| 3216 | local_irq_disable(); |
| 3217 | if (del_timer_sync(&dwork->timer)) |
| 3218 | __queue_work(dwork->cpu, dwork->wq, &dwork->work); |
| 3219 | local_irq_enable(); |
| 3220 | return flush_work(&dwork->work); |
| 3221 | } |
| 3222 | EXPORT_SYMBOL(flush_delayed_work); |
| 3223 | |
| 3224 | /** |
| 3225 | * flush_rcu_work - wait for a rwork to finish executing the last queueing |
| 3226 | * @rwork: the rcu work to flush |
| 3227 | * |
| 3228 | * Return: |
| 3229 | * %true if flush_rcu_work() waited for the work to finish execution, |
| 3230 | * %false if it was already idle. |
| 3231 | */ |
| 3232 | bool flush_rcu_work(struct rcu_work *rwork) |
| 3233 | { |
| 3234 | if (test_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&rwork->work))) { |
| 3235 | rcu_barrier(); |
| 3236 | flush_work(&rwork->work); |
| 3237 | return true; |
| 3238 | } else { |
| 3239 | return flush_work(&rwork->work); |
| 3240 | } |
| 3241 | } |
| 3242 | EXPORT_SYMBOL(flush_rcu_work); |
| 3243 | |
| 3244 | static bool __cancel_work(struct work_struct *work, bool is_dwork) |
| 3245 | { |
| 3246 | unsigned long flags; |
| 3247 | int ret; |
| 3248 | |
| 3249 | do { |
| 3250 | ret = try_to_grab_pending(work, is_dwork, &flags); |
| 3251 | } while (unlikely(ret == -EAGAIN)); |
| 3252 | |
| 3253 | if (unlikely(ret < 0)) |
| 3254 | return false; |
| 3255 | |
| 3256 | set_work_pool_and_clear_pending(work, get_work_pool_id(work)); |
| 3257 | local_irq_restore(flags); |
| 3258 | return ret; |
| 3259 | } |
| 3260 | |
| 3261 | /* |
| 3262 | * See cancel_delayed_work() |
| 3263 | */ |
| 3264 | bool cancel_work(struct work_struct *work) |
| 3265 | { |
| 3266 | return __cancel_work(work, false); |
| 3267 | } |
| 3268 | EXPORT_SYMBOL(cancel_work); |
| 3269 | |
| 3270 | /** |
| 3271 | * cancel_delayed_work - cancel a delayed work |
| 3272 | * @dwork: delayed_work to cancel |
| 3273 | * |
| 3274 | * Kill off a pending delayed_work. |
| 3275 | * |
| 3276 | * Return: %true if @dwork was pending and canceled; %false if it wasn't |
| 3277 | * pending. |
| 3278 | * |
| 3279 | * Note: |
| 3280 | * The work callback function may still be running on return, unless |
| 3281 | * it returns %true and the work doesn't re-arm itself. Explicitly flush or |
| 3282 | * use cancel_delayed_work_sync() to wait on it. |
| 3283 | * |
| 3284 | * This function is safe to call from any context including IRQ handler. |
| 3285 | */ |
| 3286 | bool cancel_delayed_work(struct delayed_work *dwork) |
| 3287 | { |
| 3288 | return __cancel_work(&dwork->work, true); |
| 3289 | } |
| 3290 | EXPORT_SYMBOL(cancel_delayed_work); |
| 3291 | |
| 3292 | /** |
| 3293 | * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish |
| 3294 | * @dwork: the delayed work cancel |
| 3295 | * |
| 3296 | * This is cancel_work_sync() for delayed works. |
| 3297 | * |
| 3298 | * Return: |
| 3299 | * %true if @dwork was pending, %false otherwise. |
| 3300 | */ |
| 3301 | bool cancel_delayed_work_sync(struct delayed_work *dwork) |
| 3302 | { |
| 3303 | return __cancel_work_timer(&dwork->work, true); |
| 3304 | } |
| 3305 | EXPORT_SYMBOL(cancel_delayed_work_sync); |
| 3306 | |
| 3307 | /** |
| 3308 | * schedule_on_each_cpu - execute a function synchronously on each online CPU |
| 3309 | * @func: the function to call |
| 3310 | * |
| 3311 | * schedule_on_each_cpu() executes @func on each online CPU using the |
| 3312 | * system workqueue and blocks until all CPUs have completed. |
| 3313 | * schedule_on_each_cpu() is very slow. |
| 3314 | * |
| 3315 | * Return: |
| 3316 | * 0 on success, -errno on failure. |
| 3317 | */ |
| 3318 | int schedule_on_each_cpu(work_func_t func) |
| 3319 | { |
| 3320 | int cpu; |
| 3321 | struct work_struct __percpu *works; |
| 3322 | |
| 3323 | works = alloc_percpu(struct work_struct); |
| 3324 | if (!works) |
| 3325 | return -ENOMEM; |
| 3326 | |
| 3327 | cpus_read_lock(); |
| 3328 | |
| 3329 | for_each_online_cpu(cpu) { |
| 3330 | struct work_struct *work = per_cpu_ptr(works, cpu); |
| 3331 | |
| 3332 | INIT_WORK(work, func); |
| 3333 | schedule_work_on(cpu, work); |
| 3334 | } |
| 3335 | |
| 3336 | for_each_online_cpu(cpu) |
| 3337 | flush_work(per_cpu_ptr(works, cpu)); |
| 3338 | |
| 3339 | cpus_read_unlock(); |
| 3340 | free_percpu(works); |
| 3341 | return 0; |
| 3342 | } |
| 3343 | |
| 3344 | /** |
| 3345 | * execute_in_process_context - reliably execute the routine with user context |
| 3346 | * @fn: the function to execute |
| 3347 | * @ew: guaranteed storage for the execute work structure (must |
| 3348 | * be available when the work executes) |
| 3349 | * |
| 3350 | * Executes the function immediately if process context is available, |
| 3351 | * otherwise schedules the function for delayed execution. |
| 3352 | * |
| 3353 | * Return: 0 - function was executed |
| 3354 | * 1 - function was scheduled for execution |
| 3355 | */ |
| 3356 | int execute_in_process_context(work_func_t fn, struct execute_work *ew) |
| 3357 | { |
| 3358 | if (!in_interrupt()) { |
| 3359 | fn(&ew->work); |
| 3360 | return 0; |
| 3361 | } |
| 3362 | |
| 3363 | INIT_WORK(&ew->work, fn); |
| 3364 | schedule_work(&ew->work); |
| 3365 | |
| 3366 | return 1; |
| 3367 | } |
| 3368 | EXPORT_SYMBOL_GPL(execute_in_process_context); |
| 3369 | |
| 3370 | /** |
| 3371 | * free_workqueue_attrs - free a workqueue_attrs |
| 3372 | * @attrs: workqueue_attrs to free |
| 3373 | * |
| 3374 | * Undo alloc_workqueue_attrs(). |
| 3375 | */ |
| 3376 | void free_workqueue_attrs(struct workqueue_attrs *attrs) |
| 3377 | { |
| 3378 | if (attrs) { |
| 3379 | free_cpumask_var(attrs->cpumask); |
| 3380 | kfree(attrs); |
| 3381 | } |
| 3382 | } |
| 3383 | |
| 3384 | /** |
| 3385 | * alloc_workqueue_attrs - allocate a workqueue_attrs |
| 3386 | * |
| 3387 | * Allocate a new workqueue_attrs, initialize with default settings and |
| 3388 | * return it. |
| 3389 | * |
| 3390 | * Return: The allocated new workqueue_attr on success. %NULL on failure. |
| 3391 | */ |
| 3392 | struct workqueue_attrs *alloc_workqueue_attrs(void) |
| 3393 | { |
| 3394 | struct workqueue_attrs *attrs; |
| 3395 | |
| 3396 | attrs = kzalloc(sizeof(*attrs), GFP_KERNEL); |
| 3397 | if (!attrs) |
| 3398 | goto fail; |
| 3399 | if (!alloc_cpumask_var(&attrs->cpumask, GFP_KERNEL)) |
| 3400 | goto fail; |
| 3401 | |
| 3402 | cpumask_copy(attrs->cpumask, cpu_possible_mask); |
| 3403 | return attrs; |
| 3404 | fail: |
| 3405 | free_workqueue_attrs(attrs); |
| 3406 | return NULL; |
| 3407 | } |
| 3408 | |
| 3409 | static void copy_workqueue_attrs(struct workqueue_attrs *to, |
| 3410 | const struct workqueue_attrs *from) |
| 3411 | { |
| 3412 | to->nice = from->nice; |
| 3413 | cpumask_copy(to->cpumask, from->cpumask); |
| 3414 | /* |
| 3415 | * Unlike hash and equality test, this function doesn't ignore |
| 3416 | * ->no_numa as it is used for both pool and wq attrs. Instead, |
| 3417 | * get_unbound_pool() explicitly clears ->no_numa after copying. |
| 3418 | */ |
| 3419 | to->no_numa = from->no_numa; |
| 3420 | } |
| 3421 | |
| 3422 | /* hash value of the content of @attr */ |
| 3423 | static u32 wqattrs_hash(const struct workqueue_attrs *attrs) |
| 3424 | { |
| 3425 | u32 hash = 0; |
| 3426 | |
| 3427 | hash = jhash_1word(attrs->nice, hash); |
| 3428 | hash = jhash(cpumask_bits(attrs->cpumask), |
| 3429 | BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash); |
| 3430 | return hash; |
| 3431 | } |
| 3432 | |
| 3433 | /* content equality test */ |
| 3434 | static bool wqattrs_equal(const struct workqueue_attrs *a, |
| 3435 | const struct workqueue_attrs *b) |
| 3436 | { |
| 3437 | if (a->nice != b->nice) |
| 3438 | return false; |
| 3439 | if (!cpumask_equal(a->cpumask, b->cpumask)) |
| 3440 | return false; |
| 3441 | return true; |
| 3442 | } |
| 3443 | |
| 3444 | /** |
| 3445 | * init_worker_pool - initialize a newly zalloc'd worker_pool |
| 3446 | * @pool: worker_pool to initialize |
| 3447 | * |
| 3448 | * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs. |
| 3449 | * |
| 3450 | * Return: 0 on success, -errno on failure. Even on failure, all fields |
| 3451 | * inside @pool proper are initialized and put_unbound_pool() can be called |
| 3452 | * on @pool safely to release it. |
| 3453 | */ |
| 3454 | static int init_worker_pool(struct worker_pool *pool) |
| 3455 | { |
| 3456 | raw_spin_lock_init(&pool->lock); |
| 3457 | pool->id = -1; |
| 3458 | pool->cpu = -1; |
| 3459 | pool->node = NUMA_NO_NODE; |
| 3460 | pool->flags |= POOL_DISASSOCIATED; |
| 3461 | pool->watchdog_ts = jiffies; |
| 3462 | INIT_LIST_HEAD(&pool->worklist); |
| 3463 | INIT_LIST_HEAD(&pool->idle_list); |
| 3464 | hash_init(pool->busy_hash); |
| 3465 | |
| 3466 | timer_setup(&pool->idle_timer, idle_worker_timeout, TIMER_DEFERRABLE); |
| 3467 | |
| 3468 | timer_setup(&pool->mayday_timer, pool_mayday_timeout, 0); |
| 3469 | |
| 3470 | INIT_LIST_HEAD(&pool->workers); |
| 3471 | |
| 3472 | ida_init(&pool->worker_ida); |
| 3473 | INIT_HLIST_NODE(&pool->hash_node); |
| 3474 | pool->refcnt = 1; |
| 3475 | |
| 3476 | /* shouldn't fail above this point */ |
| 3477 | pool->attrs = alloc_workqueue_attrs(); |
| 3478 | if (!pool->attrs) |
| 3479 | return -ENOMEM; |
| 3480 | return 0; |
| 3481 | } |
| 3482 | |
| 3483 | #ifdef CONFIG_LOCKDEP |
| 3484 | static void wq_init_lockdep(struct workqueue_struct *wq) |
| 3485 | { |
| 3486 | char *lock_name; |
| 3487 | |
| 3488 | lockdep_register_key(&wq->key); |
| 3489 | lock_name = kasprintf(GFP_KERNEL, "%s%s", "(wq_completion)", wq->name); |
| 3490 | if (!lock_name) |
| 3491 | lock_name = wq->name; |
| 3492 | |
| 3493 | wq->lock_name = lock_name; |
| 3494 | lockdep_init_map(&wq->lockdep_map, lock_name, &wq->key, 0); |
| 3495 | } |
| 3496 | |
| 3497 | static void wq_unregister_lockdep(struct workqueue_struct *wq) |
| 3498 | { |
| 3499 | lockdep_unregister_key(&wq->key); |
| 3500 | } |
| 3501 | |
| 3502 | static void wq_free_lockdep(struct workqueue_struct *wq) |
| 3503 | { |
| 3504 | if (wq->lock_name != wq->name) |
| 3505 | kfree(wq->lock_name); |
| 3506 | } |
| 3507 | #else |
| 3508 | static void wq_init_lockdep(struct workqueue_struct *wq) |
| 3509 | { |
| 3510 | } |
| 3511 | |
| 3512 | static void wq_unregister_lockdep(struct workqueue_struct *wq) |
| 3513 | { |
| 3514 | } |
| 3515 | |
| 3516 | static void wq_free_lockdep(struct workqueue_struct *wq) |
| 3517 | { |
| 3518 | } |
| 3519 | #endif |
| 3520 | |
| 3521 | static void rcu_free_wq(struct rcu_head *rcu) |
| 3522 | { |
| 3523 | struct workqueue_struct *wq = |
| 3524 | container_of(rcu, struct workqueue_struct, rcu); |
| 3525 | |
| 3526 | wq_free_lockdep(wq); |
| 3527 | |
| 3528 | if (!(wq->flags & WQ_UNBOUND)) |
| 3529 | free_percpu(wq->cpu_pwqs); |
| 3530 | else |
| 3531 | free_workqueue_attrs(wq->unbound_attrs); |
| 3532 | |
| 3533 | kfree(wq); |
| 3534 | } |
| 3535 | |
| 3536 | static void rcu_free_pool(struct rcu_head *rcu) |
| 3537 | { |
| 3538 | struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu); |
| 3539 | |
| 3540 | ida_destroy(&pool->worker_ida); |
| 3541 | free_workqueue_attrs(pool->attrs); |
| 3542 | kfree(pool); |
| 3543 | } |
| 3544 | |
| 3545 | /* This returns with the lock held on success (pool manager is inactive). */ |
| 3546 | static bool wq_manager_inactive(struct worker_pool *pool) |
| 3547 | { |
| 3548 | raw_spin_lock_irq(&pool->lock); |
| 3549 | |
| 3550 | if (pool->flags & POOL_MANAGER_ACTIVE) { |
| 3551 | raw_spin_unlock_irq(&pool->lock); |
| 3552 | return false; |
| 3553 | } |
| 3554 | return true; |
| 3555 | } |
| 3556 | |
| 3557 | /** |
| 3558 | * put_unbound_pool - put a worker_pool |
| 3559 | * @pool: worker_pool to put |
| 3560 | * |
| 3561 | * Put @pool. If its refcnt reaches zero, it gets destroyed in RCU |
| 3562 | * safe manner. get_unbound_pool() calls this function on its failure path |
| 3563 | * and this function should be able to release pools which went through, |
| 3564 | * successfully or not, init_worker_pool(). |
| 3565 | * |
| 3566 | * Should be called with wq_pool_mutex held. |
| 3567 | */ |
| 3568 | static void put_unbound_pool(struct worker_pool *pool) |
| 3569 | { |
| 3570 | DECLARE_COMPLETION_ONSTACK(detach_completion); |
| 3571 | struct worker *worker; |
| 3572 | |
| 3573 | lockdep_assert_held(&wq_pool_mutex); |
| 3574 | |
| 3575 | if (--pool->refcnt) |
| 3576 | return; |
| 3577 | |
| 3578 | /* sanity checks */ |
| 3579 | if (WARN_ON(!(pool->cpu < 0)) || |
| 3580 | WARN_ON(!list_empty(&pool->worklist))) |
| 3581 | return; |
| 3582 | |
| 3583 | /* release id and unhash */ |
| 3584 | if (pool->id >= 0) |
| 3585 | idr_remove(&worker_pool_idr, pool->id); |
| 3586 | hash_del(&pool->hash_node); |
| 3587 | |
| 3588 | /* |
| 3589 | * Become the manager and destroy all workers. This prevents |
| 3590 | * @pool's workers from blocking on attach_mutex. We're the last |
| 3591 | * manager and @pool gets freed with the flag set. |
| 3592 | * Because of how wq_manager_inactive() works, we will hold the |
| 3593 | * spinlock after a successful wait. |
| 3594 | */ |
| 3595 | rcuwait_wait_event(&manager_wait, wq_manager_inactive(pool), |
| 3596 | TASK_UNINTERRUPTIBLE); |
| 3597 | pool->flags |= POOL_MANAGER_ACTIVE; |
| 3598 | |
| 3599 | while ((worker = first_idle_worker(pool))) |
| 3600 | destroy_worker(worker); |
| 3601 | WARN_ON(pool->nr_workers || pool->nr_idle); |
| 3602 | raw_spin_unlock_irq(&pool->lock); |
| 3603 | |
| 3604 | mutex_lock(&wq_pool_attach_mutex); |
| 3605 | if (!list_empty(&pool->workers)) |
| 3606 | pool->detach_completion = &detach_completion; |
| 3607 | mutex_unlock(&wq_pool_attach_mutex); |
| 3608 | |
| 3609 | if (pool->detach_completion) |
| 3610 | wait_for_completion(pool->detach_completion); |
| 3611 | |
| 3612 | /* shut down the timers */ |
| 3613 | del_timer_sync(&pool->idle_timer); |
| 3614 | del_timer_sync(&pool->mayday_timer); |
| 3615 | |
| 3616 | /* RCU protected to allow dereferences from get_work_pool() */ |
| 3617 | call_rcu(&pool->rcu, rcu_free_pool); |
| 3618 | } |
| 3619 | |
| 3620 | /** |
| 3621 | * get_unbound_pool - get a worker_pool with the specified attributes |
| 3622 | * @attrs: the attributes of the worker_pool to get |
| 3623 | * |
| 3624 | * Obtain a worker_pool which has the same attributes as @attrs, bump the |
| 3625 | * reference count and return it. If there already is a matching |
| 3626 | * worker_pool, it will be used; otherwise, this function attempts to |
| 3627 | * create a new one. |
| 3628 | * |
| 3629 | * Should be called with wq_pool_mutex held. |
| 3630 | * |
| 3631 | * Return: On success, a worker_pool with the same attributes as @attrs. |
| 3632 | * On failure, %NULL. |
| 3633 | */ |
| 3634 | static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs) |
| 3635 | { |
| 3636 | u32 hash = wqattrs_hash(attrs); |
| 3637 | struct worker_pool *pool; |
| 3638 | int node; |
| 3639 | int target_node = NUMA_NO_NODE; |
| 3640 | |
| 3641 | lockdep_assert_held(&wq_pool_mutex); |
| 3642 | |
| 3643 | /* do we already have a matching pool? */ |
| 3644 | hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) { |
| 3645 | if (wqattrs_equal(pool->attrs, attrs)) { |
| 3646 | pool->refcnt++; |
| 3647 | return pool; |
| 3648 | } |
| 3649 | } |
| 3650 | |
| 3651 | /* if cpumask is contained inside a NUMA node, we belong to that node */ |
| 3652 | if (wq_numa_enabled) { |
| 3653 | for_each_node(node) { |
| 3654 | if (cpumask_subset(attrs->cpumask, |
| 3655 | wq_numa_possible_cpumask[node])) { |
| 3656 | target_node = node; |
| 3657 | break; |
| 3658 | } |
| 3659 | } |
| 3660 | } |
| 3661 | |
| 3662 | /* nope, create a new one */ |
| 3663 | pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, target_node); |
| 3664 | if (!pool || init_worker_pool(pool) < 0) |
| 3665 | goto fail; |
| 3666 | |
| 3667 | lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */ |
| 3668 | copy_workqueue_attrs(pool->attrs, attrs); |
| 3669 | pool->node = target_node; |
| 3670 | |
| 3671 | /* |
| 3672 | * no_numa isn't a worker_pool attribute, always clear it. See |
| 3673 | * 'struct workqueue_attrs' comments for detail. |
| 3674 | */ |
| 3675 | pool->attrs->no_numa = false; |
| 3676 | |
| 3677 | if (worker_pool_assign_id(pool) < 0) |
| 3678 | goto fail; |
| 3679 | |
| 3680 | /* create and start the initial worker */ |
| 3681 | if (wq_online && !create_worker(pool)) |
| 3682 | goto fail; |
| 3683 | |
| 3684 | /* install */ |
| 3685 | hash_add(unbound_pool_hash, &pool->hash_node, hash); |
| 3686 | |
| 3687 | return pool; |
| 3688 | fail: |
| 3689 | if (pool) |
| 3690 | put_unbound_pool(pool); |
| 3691 | return NULL; |
| 3692 | } |
| 3693 | |
| 3694 | static void rcu_free_pwq(struct rcu_head *rcu) |
| 3695 | { |
| 3696 | kmem_cache_free(pwq_cache, |
| 3697 | container_of(rcu, struct pool_workqueue, rcu)); |
| 3698 | } |
| 3699 | |
| 3700 | /* |
| 3701 | * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt |
| 3702 | * and needs to be destroyed. |
| 3703 | */ |
| 3704 | static void pwq_unbound_release_workfn(struct work_struct *work) |
| 3705 | { |
| 3706 | struct pool_workqueue *pwq = container_of(work, struct pool_workqueue, |
| 3707 | unbound_release_work); |
| 3708 | struct workqueue_struct *wq = pwq->wq; |
| 3709 | struct worker_pool *pool = pwq->pool; |
| 3710 | bool is_last = false; |
| 3711 | |
| 3712 | /* |
| 3713 | * when @pwq is not linked, it doesn't hold any reference to the |
| 3714 | * @wq, and @wq is invalid to access. |
| 3715 | */ |
| 3716 | if (!list_empty(&pwq->pwqs_node)) { |
| 3717 | if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND))) |
| 3718 | return; |
| 3719 | |
| 3720 | mutex_lock(&wq->mutex); |
| 3721 | list_del_rcu(&pwq->pwqs_node); |
| 3722 | is_last = list_empty(&wq->pwqs); |
| 3723 | mutex_unlock(&wq->mutex); |
| 3724 | } |
| 3725 | |
| 3726 | mutex_lock(&wq_pool_mutex); |
| 3727 | put_unbound_pool(pool); |
| 3728 | mutex_unlock(&wq_pool_mutex); |
| 3729 | |
| 3730 | call_rcu(&pwq->rcu, rcu_free_pwq); |
| 3731 | |
| 3732 | /* |
| 3733 | * If we're the last pwq going away, @wq is already dead and no one |
| 3734 | * is gonna access it anymore. Schedule RCU free. |
| 3735 | */ |
| 3736 | if (is_last) { |
| 3737 | wq_unregister_lockdep(wq); |
| 3738 | call_rcu(&wq->rcu, rcu_free_wq); |
| 3739 | } |
| 3740 | } |
| 3741 | |
| 3742 | /** |
| 3743 | * pwq_adjust_max_active - update a pwq's max_active to the current setting |
| 3744 | * @pwq: target pool_workqueue |
| 3745 | * |
| 3746 | * If @pwq isn't freezing, set @pwq->max_active to the associated |
| 3747 | * workqueue's saved_max_active and activate inactive work items |
| 3748 | * accordingly. If @pwq is freezing, clear @pwq->max_active to zero. |
| 3749 | */ |
| 3750 | static void pwq_adjust_max_active(struct pool_workqueue *pwq) |
| 3751 | { |
| 3752 | struct workqueue_struct *wq = pwq->wq; |
| 3753 | bool freezable = wq->flags & WQ_FREEZABLE; |
| 3754 | unsigned long flags; |
| 3755 | |
| 3756 | /* for @wq->saved_max_active */ |
| 3757 | lockdep_assert_held(&wq->mutex); |
| 3758 | |
| 3759 | /* fast exit for non-freezable wqs */ |
| 3760 | if (!freezable && pwq->max_active == wq->saved_max_active) |
| 3761 | return; |
| 3762 | |
| 3763 | /* this function can be called during early boot w/ irq disabled */ |
| 3764 | raw_spin_lock_irqsave(&pwq->pool->lock, flags); |
| 3765 | |
| 3766 | /* |
| 3767 | * During [un]freezing, the caller is responsible for ensuring that |
| 3768 | * this function is called at least once after @workqueue_freezing |
| 3769 | * is updated and visible. |
| 3770 | */ |
| 3771 | if (!freezable || !workqueue_freezing) { |
| 3772 | bool kick = false; |
| 3773 | |
| 3774 | pwq->max_active = wq->saved_max_active; |
| 3775 | |
| 3776 | while (!list_empty(&pwq->inactive_works) && |
| 3777 | pwq->nr_active < pwq->max_active) { |
| 3778 | pwq_activate_first_inactive(pwq); |
| 3779 | kick = true; |
| 3780 | } |
| 3781 | |
| 3782 | /* |
| 3783 | * Need to kick a worker after thawed or an unbound wq's |
| 3784 | * max_active is bumped. In realtime scenarios, always kicking a |
| 3785 | * worker will cause interference on the isolated cpu cores, so |
| 3786 | * let's kick iff work items were activated. |
| 3787 | */ |
| 3788 | if (kick) |
| 3789 | wake_up_worker(pwq->pool); |
| 3790 | } else { |
| 3791 | pwq->max_active = 0; |
| 3792 | } |
| 3793 | |
| 3794 | raw_spin_unlock_irqrestore(&pwq->pool->lock, flags); |
| 3795 | } |
| 3796 | |
| 3797 | /* initialize newly allocated @pwq which is associated with @wq and @pool */ |
| 3798 | static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq, |
| 3799 | struct worker_pool *pool) |
| 3800 | { |
| 3801 | BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK); |
| 3802 | |
| 3803 | memset(pwq, 0, sizeof(*pwq)); |
| 3804 | |
| 3805 | pwq->pool = pool; |
| 3806 | pwq->wq = wq; |
| 3807 | pwq->flush_color = -1; |
| 3808 | pwq->refcnt = 1; |
| 3809 | INIT_LIST_HEAD(&pwq->inactive_works); |
| 3810 | INIT_LIST_HEAD(&pwq->pwqs_node); |
| 3811 | INIT_LIST_HEAD(&pwq->mayday_node); |
| 3812 | INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn); |
| 3813 | } |
| 3814 | |
| 3815 | /* sync @pwq with the current state of its associated wq and link it */ |
| 3816 | static void link_pwq(struct pool_workqueue *pwq) |
| 3817 | { |
| 3818 | struct workqueue_struct *wq = pwq->wq; |
| 3819 | |
| 3820 | lockdep_assert_held(&wq->mutex); |
| 3821 | |
| 3822 | /* may be called multiple times, ignore if already linked */ |
| 3823 | if (!list_empty(&pwq->pwqs_node)) |
| 3824 | return; |
| 3825 | |
| 3826 | /* set the matching work_color */ |
| 3827 | pwq->work_color = wq->work_color; |
| 3828 | |
| 3829 | /* sync max_active to the current setting */ |
| 3830 | pwq_adjust_max_active(pwq); |
| 3831 | |
| 3832 | /* link in @pwq */ |
| 3833 | list_add_rcu(&pwq->pwqs_node, &wq->pwqs); |
| 3834 | } |
| 3835 | |
| 3836 | /* obtain a pool matching @attr and create a pwq associating the pool and @wq */ |
| 3837 | static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq, |
| 3838 | const struct workqueue_attrs *attrs) |
| 3839 | { |
| 3840 | struct worker_pool *pool; |
| 3841 | struct pool_workqueue *pwq; |
| 3842 | |
| 3843 | lockdep_assert_held(&wq_pool_mutex); |
| 3844 | |
| 3845 | pool = get_unbound_pool(attrs); |
| 3846 | if (!pool) |
| 3847 | return NULL; |
| 3848 | |
| 3849 | pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node); |
| 3850 | if (!pwq) { |
| 3851 | put_unbound_pool(pool); |
| 3852 | return NULL; |
| 3853 | } |
| 3854 | |
| 3855 | init_pwq(pwq, wq, pool); |
| 3856 | return pwq; |
| 3857 | } |
| 3858 | |
| 3859 | /** |
| 3860 | * wq_calc_node_cpumask - calculate a wq_attrs' cpumask for the specified node |
| 3861 | * @attrs: the wq_attrs of the default pwq of the target workqueue |
| 3862 | * @node: the target NUMA node |
| 3863 | * @cpu_going_down: if >= 0, the CPU to consider as offline |
| 3864 | * @cpumask: outarg, the resulting cpumask |
| 3865 | * |
| 3866 | * Calculate the cpumask a workqueue with @attrs should use on @node. If |
| 3867 | * @cpu_going_down is >= 0, that cpu is considered offline during |
| 3868 | * calculation. The result is stored in @cpumask. |
| 3869 | * |
| 3870 | * If NUMA affinity is not enabled, @attrs->cpumask is always used. If |
| 3871 | * enabled and @node has online CPUs requested by @attrs, the returned |
| 3872 | * cpumask is the intersection of the possible CPUs of @node and |
| 3873 | * @attrs->cpumask. |
| 3874 | * |
| 3875 | * The caller is responsible for ensuring that the cpumask of @node stays |
| 3876 | * stable. |
| 3877 | * |
| 3878 | * Return: %true if the resulting @cpumask is different from @attrs->cpumask, |
| 3879 | * %false if equal. |
| 3880 | */ |
| 3881 | static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node, |
| 3882 | int cpu_going_down, cpumask_t *cpumask) |
| 3883 | { |
| 3884 | if (!wq_numa_enabled || attrs->no_numa) |
| 3885 | goto use_dfl; |
| 3886 | |
| 3887 | /* does @node have any online CPUs @attrs wants? */ |
| 3888 | cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask); |
| 3889 | if (cpu_going_down >= 0) |
| 3890 | cpumask_clear_cpu(cpu_going_down, cpumask); |
| 3891 | |
| 3892 | if (cpumask_empty(cpumask)) |
| 3893 | goto use_dfl; |
| 3894 | |
| 3895 | /* yeap, return possible CPUs in @node that @attrs wants */ |
| 3896 | cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]); |
| 3897 | |
| 3898 | if (cpumask_empty(cpumask)) { |
| 3899 | pr_warn_once("WARNING: workqueue cpumask: online intersect > " |
| 3900 | "possible intersect\n"); |
| 3901 | return false; |
| 3902 | } |
| 3903 | |
| 3904 | return !cpumask_equal(cpumask, attrs->cpumask); |
| 3905 | |
| 3906 | use_dfl: |
| 3907 | cpumask_copy(cpumask, attrs->cpumask); |
| 3908 | return false; |
| 3909 | } |
| 3910 | |
| 3911 | /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */ |
| 3912 | static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq, |
| 3913 | int node, |
| 3914 | struct pool_workqueue *pwq) |
| 3915 | { |
| 3916 | struct pool_workqueue *old_pwq; |
| 3917 | |
| 3918 | lockdep_assert_held(&wq_pool_mutex); |
| 3919 | lockdep_assert_held(&wq->mutex); |
| 3920 | |
| 3921 | /* link_pwq() can handle duplicate calls */ |
| 3922 | link_pwq(pwq); |
| 3923 | |
| 3924 | old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); |
| 3925 | rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq); |
| 3926 | return old_pwq; |
| 3927 | } |
| 3928 | |
| 3929 | /* context to store the prepared attrs & pwqs before applying */ |
| 3930 | struct apply_wqattrs_ctx { |
| 3931 | struct workqueue_struct *wq; /* target workqueue */ |
| 3932 | struct workqueue_attrs *attrs; /* attrs to apply */ |
| 3933 | struct list_head list; /* queued for batching commit */ |
| 3934 | struct pool_workqueue *dfl_pwq; |
| 3935 | struct pool_workqueue *pwq_tbl[]; |
| 3936 | }; |
| 3937 | |
| 3938 | /* free the resources after success or abort */ |
| 3939 | static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx) |
| 3940 | { |
| 3941 | if (ctx) { |
| 3942 | int node; |
| 3943 | |
| 3944 | for_each_node(node) |
| 3945 | put_pwq_unlocked(ctx->pwq_tbl[node]); |
| 3946 | put_pwq_unlocked(ctx->dfl_pwq); |
| 3947 | |
| 3948 | free_workqueue_attrs(ctx->attrs); |
| 3949 | |
| 3950 | kfree(ctx); |
| 3951 | } |
| 3952 | } |
| 3953 | |
| 3954 | /* allocate the attrs and pwqs for later installation */ |
| 3955 | static struct apply_wqattrs_ctx * |
| 3956 | apply_wqattrs_prepare(struct workqueue_struct *wq, |
| 3957 | const struct workqueue_attrs *attrs) |
| 3958 | { |
| 3959 | struct apply_wqattrs_ctx *ctx; |
| 3960 | struct workqueue_attrs *new_attrs, *tmp_attrs; |
| 3961 | int node; |
| 3962 | |
| 3963 | lockdep_assert_held(&wq_pool_mutex); |
| 3964 | |
| 3965 | ctx = kzalloc(struct_size(ctx, pwq_tbl, nr_node_ids), GFP_KERNEL); |
| 3966 | |
| 3967 | new_attrs = alloc_workqueue_attrs(); |
| 3968 | tmp_attrs = alloc_workqueue_attrs(); |
| 3969 | if (!ctx || !new_attrs || !tmp_attrs) |
| 3970 | goto out_free; |
| 3971 | |
| 3972 | /* |
| 3973 | * Calculate the attrs of the default pwq. |
| 3974 | * If the user configured cpumask doesn't overlap with the |
| 3975 | * wq_unbound_cpumask, we fallback to the wq_unbound_cpumask. |
| 3976 | */ |
| 3977 | copy_workqueue_attrs(new_attrs, attrs); |
| 3978 | cpumask_and(new_attrs->cpumask, new_attrs->cpumask, wq_unbound_cpumask); |
| 3979 | if (unlikely(cpumask_empty(new_attrs->cpumask))) |
| 3980 | cpumask_copy(new_attrs->cpumask, wq_unbound_cpumask); |
| 3981 | |
| 3982 | /* |
| 3983 | * We may create multiple pwqs with differing cpumasks. Make a |
| 3984 | * copy of @new_attrs which will be modified and used to obtain |
| 3985 | * pools. |
| 3986 | */ |
| 3987 | copy_workqueue_attrs(tmp_attrs, new_attrs); |
| 3988 | |
| 3989 | /* |
| 3990 | * If something goes wrong during CPU up/down, we'll fall back to |
| 3991 | * the default pwq covering whole @attrs->cpumask. Always create |
| 3992 | * it even if we don't use it immediately. |
| 3993 | */ |
| 3994 | ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs); |
| 3995 | if (!ctx->dfl_pwq) |
| 3996 | goto out_free; |
| 3997 | |
| 3998 | for_each_node(node) { |
| 3999 | if (wq_calc_node_cpumask(new_attrs, node, -1, tmp_attrs->cpumask)) { |
| 4000 | ctx->pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs); |
| 4001 | if (!ctx->pwq_tbl[node]) |
| 4002 | goto out_free; |
| 4003 | } else { |
| 4004 | ctx->dfl_pwq->refcnt++; |
| 4005 | ctx->pwq_tbl[node] = ctx->dfl_pwq; |
| 4006 | } |
| 4007 | } |
| 4008 | |
| 4009 | /* save the user configured attrs and sanitize it. */ |
| 4010 | copy_workqueue_attrs(new_attrs, attrs); |
| 4011 | cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask); |
| 4012 | ctx->attrs = new_attrs; |
| 4013 | |
| 4014 | ctx->wq = wq; |
| 4015 | free_workqueue_attrs(tmp_attrs); |
| 4016 | return ctx; |
| 4017 | |
| 4018 | out_free: |
| 4019 | free_workqueue_attrs(tmp_attrs); |
| 4020 | free_workqueue_attrs(new_attrs); |
| 4021 | apply_wqattrs_cleanup(ctx); |
| 4022 | return NULL; |
| 4023 | } |
| 4024 | |
| 4025 | /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */ |
| 4026 | static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx) |
| 4027 | { |
| 4028 | int node; |
| 4029 | |
| 4030 | /* all pwqs have been created successfully, let's install'em */ |
| 4031 | mutex_lock(&ctx->wq->mutex); |
| 4032 | |
| 4033 | copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs); |
| 4034 | |
| 4035 | /* save the previous pwq and install the new one */ |
| 4036 | for_each_node(node) |
| 4037 | ctx->pwq_tbl[node] = numa_pwq_tbl_install(ctx->wq, node, |
| 4038 | ctx->pwq_tbl[node]); |
| 4039 | |
| 4040 | /* @dfl_pwq might not have been used, ensure it's linked */ |
| 4041 | link_pwq(ctx->dfl_pwq); |
| 4042 | swap(ctx->wq->dfl_pwq, ctx->dfl_pwq); |
| 4043 | |
| 4044 | mutex_unlock(&ctx->wq->mutex); |
| 4045 | } |
| 4046 | |
| 4047 | static void apply_wqattrs_lock(void) |
| 4048 | { |
| 4049 | /* CPUs should stay stable across pwq creations and installations */ |
| 4050 | cpus_read_lock(); |
| 4051 | mutex_lock(&wq_pool_mutex); |
| 4052 | } |
| 4053 | |
| 4054 | static void apply_wqattrs_unlock(void) |
| 4055 | { |
| 4056 | mutex_unlock(&wq_pool_mutex); |
| 4057 | cpus_read_unlock(); |
| 4058 | } |
| 4059 | |
| 4060 | static int apply_workqueue_attrs_locked(struct workqueue_struct *wq, |
| 4061 | const struct workqueue_attrs *attrs) |
| 4062 | { |
| 4063 | struct apply_wqattrs_ctx *ctx; |
| 4064 | |
| 4065 | /* only unbound workqueues can change attributes */ |
| 4066 | if (WARN_ON(!(wq->flags & WQ_UNBOUND))) |
| 4067 | return -EINVAL; |
| 4068 | |
| 4069 | /* creating multiple pwqs breaks ordering guarantee */ |
| 4070 | if (!list_empty(&wq->pwqs)) { |
| 4071 | if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) |
| 4072 | return -EINVAL; |
| 4073 | |
| 4074 | wq->flags &= ~__WQ_ORDERED; |
| 4075 | } |
| 4076 | |
| 4077 | ctx = apply_wqattrs_prepare(wq, attrs); |
| 4078 | if (!ctx) |
| 4079 | return -ENOMEM; |
| 4080 | |
| 4081 | /* the ctx has been prepared successfully, let's commit it */ |
| 4082 | apply_wqattrs_commit(ctx); |
| 4083 | apply_wqattrs_cleanup(ctx); |
| 4084 | |
| 4085 | return 0; |
| 4086 | } |
| 4087 | |
| 4088 | /** |
| 4089 | * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue |
| 4090 | * @wq: the target workqueue |
| 4091 | * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs() |
| 4092 | * |
| 4093 | * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA |
| 4094 | * machines, this function maps a separate pwq to each NUMA node with |
| 4095 | * possibles CPUs in @attrs->cpumask so that work items are affine to the |
| 4096 | * NUMA node it was issued on. Older pwqs are released as in-flight work |
| 4097 | * items finish. Note that a work item which repeatedly requeues itself |
| 4098 | * back-to-back will stay on its current pwq. |
| 4099 | * |
| 4100 | * Performs GFP_KERNEL allocations. |
| 4101 | * |
| 4102 | * Assumes caller has CPU hotplug read exclusion, i.e. cpus_read_lock(). |
| 4103 | * |
| 4104 | * Return: 0 on success and -errno on failure. |
| 4105 | */ |
| 4106 | int apply_workqueue_attrs(struct workqueue_struct *wq, |
| 4107 | const struct workqueue_attrs *attrs) |
| 4108 | { |
| 4109 | int ret; |
| 4110 | |
| 4111 | lockdep_assert_cpus_held(); |
| 4112 | |
| 4113 | mutex_lock(&wq_pool_mutex); |
| 4114 | ret = apply_workqueue_attrs_locked(wq, attrs); |
| 4115 | mutex_unlock(&wq_pool_mutex); |
| 4116 | |
| 4117 | return ret; |
| 4118 | } |
| 4119 | |
| 4120 | /** |
| 4121 | * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug |
| 4122 | * @wq: the target workqueue |
| 4123 | * @cpu: the CPU coming up or going down |
| 4124 | * @online: whether @cpu is coming up or going down |
| 4125 | * |
| 4126 | * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and |
| 4127 | * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of |
| 4128 | * @wq accordingly. |
| 4129 | * |
| 4130 | * If NUMA affinity can't be adjusted due to memory allocation failure, it |
| 4131 | * falls back to @wq->dfl_pwq which may not be optimal but is always |
| 4132 | * correct. |
| 4133 | * |
| 4134 | * Note that when the last allowed CPU of a NUMA node goes offline for a |
| 4135 | * workqueue with a cpumask spanning multiple nodes, the workers which were |
| 4136 | * already executing the work items for the workqueue will lose their CPU |
| 4137 | * affinity and may execute on any CPU. This is similar to how per-cpu |
| 4138 | * workqueues behave on CPU_DOWN. If a workqueue user wants strict |
| 4139 | * affinity, it's the user's responsibility to flush the work item from |
| 4140 | * CPU_DOWN_PREPARE. |
| 4141 | */ |
| 4142 | static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu, |
| 4143 | bool online) |
| 4144 | { |
| 4145 | int node = cpu_to_node(cpu); |
| 4146 | int cpu_off = online ? -1 : cpu; |
| 4147 | struct pool_workqueue *old_pwq = NULL, *pwq; |
| 4148 | struct workqueue_attrs *target_attrs; |
| 4149 | cpumask_t *cpumask; |
| 4150 | |
| 4151 | lockdep_assert_held(&wq_pool_mutex); |
| 4152 | |
| 4153 | if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND) || |
| 4154 | wq->unbound_attrs->no_numa) |
| 4155 | return; |
| 4156 | |
| 4157 | /* |
| 4158 | * We don't wanna alloc/free wq_attrs for each wq for each CPU. |
| 4159 | * Let's use a preallocated one. The following buf is protected by |
| 4160 | * CPU hotplug exclusion. |
| 4161 | */ |
| 4162 | target_attrs = wq_update_unbound_numa_attrs_buf; |
| 4163 | cpumask = target_attrs->cpumask; |
| 4164 | |
| 4165 | copy_workqueue_attrs(target_attrs, wq->unbound_attrs); |
| 4166 | pwq = unbound_pwq_by_node(wq, node); |
| 4167 | |
| 4168 | /* |
| 4169 | * Let's determine what needs to be done. If the target cpumask is |
| 4170 | * different from the default pwq's, we need to compare it to @pwq's |
| 4171 | * and create a new one if they don't match. If the target cpumask |
| 4172 | * equals the default pwq's, the default pwq should be used. |
| 4173 | */ |
| 4174 | if (wq_calc_node_cpumask(wq->dfl_pwq->pool->attrs, node, cpu_off, cpumask)) { |
| 4175 | if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask)) |
| 4176 | return; |
| 4177 | } else { |
| 4178 | goto use_dfl_pwq; |
| 4179 | } |
| 4180 | |
| 4181 | /* create a new pwq */ |
| 4182 | pwq = alloc_unbound_pwq(wq, target_attrs); |
| 4183 | if (!pwq) { |
| 4184 | pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n", |
| 4185 | wq->name); |
| 4186 | goto use_dfl_pwq; |
| 4187 | } |
| 4188 | |
| 4189 | /* Install the new pwq. */ |
| 4190 | mutex_lock(&wq->mutex); |
| 4191 | old_pwq = numa_pwq_tbl_install(wq, node, pwq); |
| 4192 | goto out_unlock; |
| 4193 | |
| 4194 | use_dfl_pwq: |
| 4195 | mutex_lock(&wq->mutex); |
| 4196 | raw_spin_lock_irq(&wq->dfl_pwq->pool->lock); |
| 4197 | get_pwq(wq->dfl_pwq); |
| 4198 | raw_spin_unlock_irq(&wq->dfl_pwq->pool->lock); |
| 4199 | old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq); |
| 4200 | out_unlock: |
| 4201 | mutex_unlock(&wq->mutex); |
| 4202 | put_pwq_unlocked(old_pwq); |
| 4203 | } |
| 4204 | |
| 4205 | static int alloc_and_link_pwqs(struct workqueue_struct *wq) |
| 4206 | { |
| 4207 | bool highpri = wq->flags & WQ_HIGHPRI; |
| 4208 | int cpu, ret; |
| 4209 | |
| 4210 | if (!(wq->flags & WQ_UNBOUND)) { |
| 4211 | wq->cpu_pwqs = alloc_percpu(struct pool_workqueue); |
| 4212 | if (!wq->cpu_pwqs) |
| 4213 | return -ENOMEM; |
| 4214 | |
| 4215 | for_each_possible_cpu(cpu) { |
| 4216 | struct pool_workqueue *pwq = |
| 4217 | per_cpu_ptr(wq->cpu_pwqs, cpu); |
| 4218 | struct worker_pool *cpu_pools = |
| 4219 | per_cpu(cpu_worker_pools, cpu); |
| 4220 | |
| 4221 | init_pwq(pwq, wq, &cpu_pools[highpri]); |
| 4222 | |
| 4223 | mutex_lock(&wq->mutex); |
| 4224 | link_pwq(pwq); |
| 4225 | mutex_unlock(&wq->mutex); |
| 4226 | } |
| 4227 | return 0; |
| 4228 | } |
| 4229 | |
| 4230 | cpus_read_lock(); |
| 4231 | if (wq->flags & __WQ_ORDERED) { |
| 4232 | ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]); |
| 4233 | /* there should only be single pwq for ordering guarantee */ |
| 4234 | WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node || |
| 4235 | wq->pwqs.prev != &wq->dfl_pwq->pwqs_node), |
| 4236 | "ordering guarantee broken for workqueue %s\n", wq->name); |
| 4237 | } else { |
| 4238 | ret = apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]); |
| 4239 | } |
| 4240 | cpus_read_unlock(); |
| 4241 | |
| 4242 | return ret; |
| 4243 | } |
| 4244 | |
| 4245 | static int wq_clamp_max_active(int max_active, unsigned int flags, |
| 4246 | const char *name) |
| 4247 | { |
| 4248 | int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE; |
| 4249 | |
| 4250 | if (max_active < 1 || max_active > lim) |
| 4251 | pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n", |
| 4252 | max_active, name, 1, lim); |
| 4253 | |
| 4254 | return clamp_val(max_active, 1, lim); |
| 4255 | } |
| 4256 | |
| 4257 | /* |
| 4258 | * Workqueues which may be used during memory reclaim should have a rescuer |
| 4259 | * to guarantee forward progress. |
| 4260 | */ |
| 4261 | static int init_rescuer(struct workqueue_struct *wq) |
| 4262 | { |
| 4263 | struct worker *rescuer; |
| 4264 | int ret; |
| 4265 | |
| 4266 | if (!(wq->flags & WQ_MEM_RECLAIM)) |
| 4267 | return 0; |
| 4268 | |
| 4269 | rescuer = alloc_worker(NUMA_NO_NODE); |
| 4270 | if (!rescuer) |
| 4271 | return -ENOMEM; |
| 4272 | |
| 4273 | rescuer->rescue_wq = wq; |
| 4274 | rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", wq->name); |
| 4275 | if (IS_ERR(rescuer->task)) { |
| 4276 | ret = PTR_ERR(rescuer->task); |
| 4277 | kfree(rescuer); |
| 4278 | return ret; |
| 4279 | } |
| 4280 | |
| 4281 | wq->rescuer = rescuer; |
| 4282 | kthread_bind_mask(rescuer->task, cpu_possible_mask); |
| 4283 | wake_up_process(rescuer->task); |
| 4284 | |
| 4285 | return 0; |
| 4286 | } |
| 4287 | |
| 4288 | __printf(1, 4) |
| 4289 | struct workqueue_struct *alloc_workqueue(const char *fmt, |
| 4290 | unsigned int flags, |
| 4291 | int max_active, ...) |
| 4292 | { |
| 4293 | size_t tbl_size = 0; |
| 4294 | va_list args; |
| 4295 | struct workqueue_struct *wq; |
| 4296 | struct pool_workqueue *pwq; |
| 4297 | |
| 4298 | /* |
| 4299 | * Unbound && max_active == 1 used to imply ordered, which is no |
| 4300 | * longer the case on NUMA machines due to per-node pools. While |
| 4301 | * alloc_ordered_workqueue() is the right way to create an ordered |
| 4302 | * workqueue, keep the previous behavior to avoid subtle breakages |
| 4303 | * on NUMA. |
| 4304 | */ |
| 4305 | if ((flags & WQ_UNBOUND) && max_active == 1) |
| 4306 | flags |= __WQ_ORDERED; |
| 4307 | |
| 4308 | /* see the comment above the definition of WQ_POWER_EFFICIENT */ |
| 4309 | if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient) |
| 4310 | flags |= WQ_UNBOUND; |
| 4311 | |
| 4312 | /* allocate wq and format name */ |
| 4313 | if (flags & WQ_UNBOUND) |
| 4314 | tbl_size = nr_node_ids * sizeof(wq->numa_pwq_tbl[0]); |
| 4315 | |
| 4316 | wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL); |
| 4317 | if (!wq) |
| 4318 | return NULL; |
| 4319 | |
| 4320 | if (flags & WQ_UNBOUND) { |
| 4321 | wq->unbound_attrs = alloc_workqueue_attrs(); |
| 4322 | if (!wq->unbound_attrs) |
| 4323 | goto err_free_wq; |
| 4324 | } |
| 4325 | |
| 4326 | va_start(args, max_active); |
| 4327 | vsnprintf(wq->name, sizeof(wq->name), fmt, args); |
| 4328 | va_end(args); |
| 4329 | |
| 4330 | max_active = max_active ?: WQ_DFL_ACTIVE; |
| 4331 | max_active = wq_clamp_max_active(max_active, flags, wq->name); |
| 4332 | |
| 4333 | /* init wq */ |
| 4334 | wq->flags = flags; |
| 4335 | wq->saved_max_active = max_active; |
| 4336 | mutex_init(&wq->mutex); |
| 4337 | atomic_set(&wq->nr_pwqs_to_flush, 0); |
| 4338 | INIT_LIST_HEAD(&wq->pwqs); |
| 4339 | INIT_LIST_HEAD(&wq->flusher_queue); |
| 4340 | INIT_LIST_HEAD(&wq->flusher_overflow); |
| 4341 | INIT_LIST_HEAD(&wq->maydays); |
| 4342 | |
| 4343 | wq_init_lockdep(wq); |
| 4344 | INIT_LIST_HEAD(&wq->list); |
| 4345 | |
| 4346 | if (alloc_and_link_pwqs(wq) < 0) |
| 4347 | goto err_unreg_lockdep; |
| 4348 | |
| 4349 | if (wq_online && init_rescuer(wq) < 0) |
| 4350 | goto err_destroy; |
| 4351 | |
| 4352 | if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq)) |
| 4353 | goto err_destroy; |
| 4354 | |
| 4355 | /* |
| 4356 | * wq_pool_mutex protects global freeze state and workqueues list. |
| 4357 | * Grab it, adjust max_active and add the new @wq to workqueues |
| 4358 | * list. |
| 4359 | */ |
| 4360 | mutex_lock(&wq_pool_mutex); |
| 4361 | |
| 4362 | mutex_lock(&wq->mutex); |
| 4363 | for_each_pwq(pwq, wq) |
| 4364 | pwq_adjust_max_active(pwq); |
| 4365 | mutex_unlock(&wq->mutex); |
| 4366 | |
| 4367 | list_add_tail_rcu(&wq->list, &workqueues); |
| 4368 | |
| 4369 | mutex_unlock(&wq_pool_mutex); |
| 4370 | |
| 4371 | return wq; |
| 4372 | |
| 4373 | err_unreg_lockdep: |
| 4374 | wq_unregister_lockdep(wq); |
| 4375 | wq_free_lockdep(wq); |
| 4376 | err_free_wq: |
| 4377 | free_workqueue_attrs(wq->unbound_attrs); |
| 4378 | kfree(wq); |
| 4379 | return NULL; |
| 4380 | err_destroy: |
| 4381 | destroy_workqueue(wq); |
| 4382 | return NULL; |
| 4383 | } |
| 4384 | EXPORT_SYMBOL_GPL(alloc_workqueue); |
| 4385 | |
| 4386 | static bool pwq_busy(struct pool_workqueue *pwq) |
| 4387 | { |
| 4388 | int i; |
| 4389 | |
| 4390 | for (i = 0; i < WORK_NR_COLORS; i++) |
| 4391 | if (pwq->nr_in_flight[i]) |
| 4392 | return true; |
| 4393 | |
| 4394 | if ((pwq != pwq->wq->dfl_pwq) && (pwq->refcnt > 1)) |
| 4395 | return true; |
| 4396 | if (pwq->nr_active || !list_empty(&pwq->inactive_works)) |
| 4397 | return true; |
| 4398 | |
| 4399 | return false; |
| 4400 | } |
| 4401 | |
| 4402 | /** |
| 4403 | * destroy_workqueue - safely terminate a workqueue |
| 4404 | * @wq: target workqueue |
| 4405 | * |
| 4406 | * Safely destroy a workqueue. All work currently pending will be done first. |
| 4407 | */ |
| 4408 | void destroy_workqueue(struct workqueue_struct *wq) |
| 4409 | { |
| 4410 | struct pool_workqueue *pwq; |
| 4411 | int node; |
| 4412 | |
| 4413 | /* |
| 4414 | * Remove it from sysfs first so that sanity check failure doesn't |
| 4415 | * lead to sysfs name conflicts. |
| 4416 | */ |
| 4417 | workqueue_sysfs_unregister(wq); |
| 4418 | |
| 4419 | /* drain it before proceeding with destruction */ |
| 4420 | drain_workqueue(wq); |
| 4421 | |
| 4422 | /* kill rescuer, if sanity checks fail, leave it w/o rescuer */ |
| 4423 | if (wq->rescuer) { |
| 4424 | struct worker *rescuer = wq->rescuer; |
| 4425 | |
| 4426 | /* this prevents new queueing */ |
| 4427 | raw_spin_lock_irq(&wq_mayday_lock); |
| 4428 | wq->rescuer = NULL; |
| 4429 | raw_spin_unlock_irq(&wq_mayday_lock); |
| 4430 | |
| 4431 | /* rescuer will empty maydays list before exiting */ |
| 4432 | kthread_stop(rescuer->task); |
| 4433 | kfree(rescuer); |
| 4434 | } |
| 4435 | |
| 4436 | /* |
| 4437 | * Sanity checks - grab all the locks so that we wait for all |
| 4438 | * in-flight operations which may do put_pwq(). |
| 4439 | */ |
| 4440 | mutex_lock(&wq_pool_mutex); |
| 4441 | mutex_lock(&wq->mutex); |
| 4442 | for_each_pwq(pwq, wq) { |
| 4443 | raw_spin_lock_irq(&pwq->pool->lock); |
| 4444 | if (WARN_ON(pwq_busy(pwq))) { |
| 4445 | pr_warn("%s: %s has the following busy pwq\n", |
| 4446 | __func__, wq->name); |
| 4447 | show_pwq(pwq); |
| 4448 | raw_spin_unlock_irq(&pwq->pool->lock); |
| 4449 | mutex_unlock(&wq->mutex); |
| 4450 | mutex_unlock(&wq_pool_mutex); |
| 4451 | show_one_workqueue(wq); |
| 4452 | return; |
| 4453 | } |
| 4454 | raw_spin_unlock_irq(&pwq->pool->lock); |
| 4455 | } |
| 4456 | mutex_unlock(&wq->mutex); |
| 4457 | |
| 4458 | /* |
| 4459 | * wq list is used to freeze wq, remove from list after |
| 4460 | * flushing is complete in case freeze races us. |
| 4461 | */ |
| 4462 | list_del_rcu(&wq->list); |
| 4463 | mutex_unlock(&wq_pool_mutex); |
| 4464 | |
| 4465 | if (!(wq->flags & WQ_UNBOUND)) { |
| 4466 | wq_unregister_lockdep(wq); |
| 4467 | /* |
| 4468 | * The base ref is never dropped on per-cpu pwqs. Directly |
| 4469 | * schedule RCU free. |
| 4470 | */ |
| 4471 | call_rcu(&wq->rcu, rcu_free_wq); |
| 4472 | } else { |
| 4473 | /* |
| 4474 | * We're the sole accessor of @wq at this point. Directly |
| 4475 | * access numa_pwq_tbl[] and dfl_pwq to put the base refs. |
| 4476 | * @wq will be freed when the last pwq is released. |
| 4477 | */ |
| 4478 | for_each_node(node) { |
| 4479 | pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); |
| 4480 | RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL); |
| 4481 | put_pwq_unlocked(pwq); |
| 4482 | } |
| 4483 | |
| 4484 | /* |
| 4485 | * Put dfl_pwq. @wq may be freed any time after dfl_pwq is |
| 4486 | * put. Don't access it afterwards. |
| 4487 | */ |
| 4488 | pwq = wq->dfl_pwq; |
| 4489 | wq->dfl_pwq = NULL; |
| 4490 | put_pwq_unlocked(pwq); |
| 4491 | } |
| 4492 | } |
| 4493 | EXPORT_SYMBOL_GPL(destroy_workqueue); |
| 4494 | |
| 4495 | /** |
| 4496 | * workqueue_set_max_active - adjust max_active of a workqueue |
| 4497 | * @wq: target workqueue |
| 4498 | * @max_active: new max_active value. |
| 4499 | * |
| 4500 | * Set max_active of @wq to @max_active. |
| 4501 | * |
| 4502 | * CONTEXT: |
| 4503 | * Don't call from IRQ context. |
| 4504 | */ |
| 4505 | void workqueue_set_max_active(struct workqueue_struct *wq, int max_active) |
| 4506 | { |
| 4507 | struct pool_workqueue *pwq; |
| 4508 | |
| 4509 | /* disallow meddling with max_active for ordered workqueues */ |
| 4510 | if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) |
| 4511 | return; |
| 4512 | |
| 4513 | max_active = wq_clamp_max_active(max_active, wq->flags, wq->name); |
| 4514 | |
| 4515 | mutex_lock(&wq->mutex); |
| 4516 | |
| 4517 | wq->flags &= ~__WQ_ORDERED; |
| 4518 | wq->saved_max_active = max_active; |
| 4519 | |
| 4520 | for_each_pwq(pwq, wq) |
| 4521 | pwq_adjust_max_active(pwq); |
| 4522 | |
| 4523 | mutex_unlock(&wq->mutex); |
| 4524 | } |
| 4525 | EXPORT_SYMBOL_GPL(workqueue_set_max_active); |
| 4526 | |
| 4527 | /** |
| 4528 | * current_work - retrieve %current task's work struct |
| 4529 | * |
| 4530 | * Determine if %current task is a workqueue worker and what it's working on. |
| 4531 | * Useful to find out the context that the %current task is running in. |
| 4532 | * |
| 4533 | * Return: work struct if %current task is a workqueue worker, %NULL otherwise. |
| 4534 | */ |
| 4535 | struct work_struct *current_work(void) |
| 4536 | { |
| 4537 | struct worker *worker = current_wq_worker(); |
| 4538 | |
| 4539 | return worker ? worker->current_work : NULL; |
| 4540 | } |
| 4541 | EXPORT_SYMBOL(current_work); |
| 4542 | |
| 4543 | /** |
| 4544 | * current_is_workqueue_rescuer - is %current workqueue rescuer? |
| 4545 | * |
| 4546 | * Determine whether %current is a workqueue rescuer. Can be used from |
| 4547 | * work functions to determine whether it's being run off the rescuer task. |
| 4548 | * |
| 4549 | * Return: %true if %current is a workqueue rescuer. %false otherwise. |
| 4550 | */ |
| 4551 | bool current_is_workqueue_rescuer(void) |
| 4552 | { |
| 4553 | struct worker *worker = current_wq_worker(); |
| 4554 | |
| 4555 | return worker && worker->rescue_wq; |
| 4556 | } |
| 4557 | |
| 4558 | /** |
| 4559 | * workqueue_congested - test whether a workqueue is congested |
| 4560 | * @cpu: CPU in question |
| 4561 | * @wq: target workqueue |
| 4562 | * |
| 4563 | * Test whether @wq's cpu workqueue for @cpu is congested. There is |
| 4564 | * no synchronization around this function and the test result is |
| 4565 | * unreliable and only useful as advisory hints or for debugging. |
| 4566 | * |
| 4567 | * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU. |
| 4568 | * Note that both per-cpu and unbound workqueues may be associated with |
| 4569 | * multiple pool_workqueues which have separate congested states. A |
| 4570 | * workqueue being congested on one CPU doesn't mean the workqueue is also |
| 4571 | * contested on other CPUs / NUMA nodes. |
| 4572 | * |
| 4573 | * Return: |
| 4574 | * %true if congested, %false otherwise. |
| 4575 | */ |
| 4576 | bool workqueue_congested(int cpu, struct workqueue_struct *wq) |
| 4577 | { |
| 4578 | struct pool_workqueue *pwq; |
| 4579 | bool ret; |
| 4580 | |
| 4581 | rcu_read_lock(); |
| 4582 | preempt_disable(); |
| 4583 | |
| 4584 | if (cpu == WORK_CPU_UNBOUND) |
| 4585 | cpu = smp_processor_id(); |
| 4586 | |
| 4587 | if (!(wq->flags & WQ_UNBOUND)) |
| 4588 | pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); |
| 4589 | else |
| 4590 | pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); |
| 4591 | |
| 4592 | ret = !list_empty(&pwq->inactive_works); |
| 4593 | preempt_enable(); |
| 4594 | rcu_read_unlock(); |
| 4595 | |
| 4596 | return ret; |
| 4597 | } |
| 4598 | EXPORT_SYMBOL_GPL(workqueue_congested); |
| 4599 | |
| 4600 | /** |
| 4601 | * work_busy - test whether a work is currently pending or running |
| 4602 | * @work: the work to be tested |
| 4603 | * |
| 4604 | * Test whether @work is currently pending or running. There is no |
| 4605 | * synchronization around this function and the test result is |
| 4606 | * unreliable and only useful as advisory hints or for debugging. |
| 4607 | * |
| 4608 | * Return: |
| 4609 | * OR'd bitmask of WORK_BUSY_* bits. |
| 4610 | */ |
| 4611 | unsigned int work_busy(struct work_struct *work) |
| 4612 | { |
| 4613 | struct worker_pool *pool; |
| 4614 | unsigned long flags; |
| 4615 | unsigned int ret = 0; |
| 4616 | |
| 4617 | if (work_pending(work)) |
| 4618 | ret |= WORK_BUSY_PENDING; |
| 4619 | |
| 4620 | rcu_read_lock(); |
| 4621 | pool = get_work_pool(work); |
| 4622 | if (pool) { |
| 4623 | raw_spin_lock_irqsave(&pool->lock, flags); |
| 4624 | if (find_worker_executing_work(pool, work)) |
| 4625 | ret |= WORK_BUSY_RUNNING; |
| 4626 | raw_spin_unlock_irqrestore(&pool->lock, flags); |
| 4627 | } |
| 4628 | rcu_read_unlock(); |
| 4629 | |
| 4630 | return ret; |
| 4631 | } |
| 4632 | EXPORT_SYMBOL_GPL(work_busy); |
| 4633 | |
| 4634 | /** |
| 4635 | * set_worker_desc - set description for the current work item |
| 4636 | * @fmt: printf-style format string |
| 4637 | * @...: arguments for the format string |
| 4638 | * |
| 4639 | * This function can be called by a running work function to describe what |
| 4640 | * the work item is about. If the worker task gets dumped, this |
| 4641 | * information will be printed out together to help debugging. The |
| 4642 | * description can be at most WORKER_DESC_LEN including the trailing '\0'. |
| 4643 | */ |
| 4644 | void set_worker_desc(const char *fmt, ...) |
| 4645 | { |
| 4646 | struct worker *worker = current_wq_worker(); |
| 4647 | va_list args; |
| 4648 | |
| 4649 | if (worker) { |
| 4650 | va_start(args, fmt); |
| 4651 | vsnprintf(worker->desc, sizeof(worker->desc), fmt, args); |
| 4652 | va_end(args); |
| 4653 | } |
| 4654 | } |
| 4655 | EXPORT_SYMBOL_GPL(set_worker_desc); |
| 4656 | |
| 4657 | /** |
| 4658 | * print_worker_info - print out worker information and description |
| 4659 | * @log_lvl: the log level to use when printing |
| 4660 | * @task: target task |
| 4661 | * |
| 4662 | * If @task is a worker and currently executing a work item, print out the |
| 4663 | * name of the workqueue being serviced and worker description set with |
| 4664 | * set_worker_desc() by the currently executing work item. |
| 4665 | * |
| 4666 | * This function can be safely called on any task as long as the |
| 4667 | * task_struct itself is accessible. While safe, this function isn't |
| 4668 | * synchronized and may print out mixups or garbages of limited length. |
| 4669 | */ |
| 4670 | void print_worker_info(const char *log_lvl, struct task_struct *task) |
| 4671 | { |
| 4672 | work_func_t *fn = NULL; |
| 4673 | char name[WQ_NAME_LEN] = { }; |
| 4674 | char desc[WORKER_DESC_LEN] = { }; |
| 4675 | struct pool_workqueue *pwq = NULL; |
| 4676 | struct workqueue_struct *wq = NULL; |
| 4677 | struct worker *worker; |
| 4678 | |
| 4679 | if (!(task->flags & PF_WQ_WORKER)) |
| 4680 | return; |
| 4681 | |
| 4682 | /* |
| 4683 | * This function is called without any synchronization and @task |
| 4684 | * could be in any state. Be careful with dereferences. |
| 4685 | */ |
| 4686 | worker = kthread_probe_data(task); |
| 4687 | |
| 4688 | /* |
| 4689 | * Carefully copy the associated workqueue's workfn, name and desc. |
| 4690 | * Keep the original last '\0' in case the original is garbage. |
| 4691 | */ |
| 4692 | copy_from_kernel_nofault(&fn, &worker->current_func, sizeof(fn)); |
| 4693 | copy_from_kernel_nofault(&pwq, &worker->current_pwq, sizeof(pwq)); |
| 4694 | copy_from_kernel_nofault(&wq, &pwq->wq, sizeof(wq)); |
| 4695 | copy_from_kernel_nofault(name, wq->name, sizeof(name) - 1); |
| 4696 | copy_from_kernel_nofault(desc, worker->desc, sizeof(desc) - 1); |
| 4697 | |
| 4698 | if (fn || name[0] || desc[0]) { |
| 4699 | printk("%sWorkqueue: %s %ps", log_lvl, name, fn); |
| 4700 | if (strcmp(name, desc)) |
| 4701 | pr_cont(" (%s)", desc); |
| 4702 | pr_cont("\n"); |
| 4703 | } |
| 4704 | } |
| 4705 | |
| 4706 | static void pr_cont_pool_info(struct worker_pool *pool) |
| 4707 | { |
| 4708 | pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask); |
| 4709 | if (pool->node != NUMA_NO_NODE) |
| 4710 | pr_cont(" node=%d", pool->node); |
| 4711 | pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice); |
| 4712 | } |
| 4713 | |
| 4714 | static void pr_cont_work(bool comma, struct work_struct *work) |
| 4715 | { |
| 4716 | if (work->func == wq_barrier_func) { |
| 4717 | struct wq_barrier *barr; |
| 4718 | |
| 4719 | barr = container_of(work, struct wq_barrier, work); |
| 4720 | |
| 4721 | pr_cont("%s BAR(%d)", comma ? "," : "", |
| 4722 | task_pid_nr(barr->task)); |
| 4723 | } else { |
| 4724 | pr_cont("%s %ps", comma ? "," : "", work->func); |
| 4725 | } |
| 4726 | } |
| 4727 | |
| 4728 | static void show_pwq(struct pool_workqueue *pwq) |
| 4729 | { |
| 4730 | struct worker_pool *pool = pwq->pool; |
| 4731 | struct work_struct *work; |
| 4732 | struct worker *worker; |
| 4733 | bool has_in_flight = false, has_pending = false; |
| 4734 | int bkt; |
| 4735 | |
| 4736 | pr_info(" pwq %d:", pool->id); |
| 4737 | pr_cont_pool_info(pool); |
| 4738 | |
| 4739 | pr_cont(" active=%d/%d refcnt=%d%s\n", |
| 4740 | pwq->nr_active, pwq->max_active, pwq->refcnt, |
| 4741 | !list_empty(&pwq->mayday_node) ? " MAYDAY" : ""); |
| 4742 | |
| 4743 | hash_for_each(pool->busy_hash, bkt, worker, hentry) { |
| 4744 | if (worker->current_pwq == pwq) { |
| 4745 | has_in_flight = true; |
| 4746 | break; |
| 4747 | } |
| 4748 | } |
| 4749 | if (has_in_flight) { |
| 4750 | bool comma = false; |
| 4751 | |
| 4752 | pr_info(" in-flight:"); |
| 4753 | hash_for_each(pool->busy_hash, bkt, worker, hentry) { |
| 4754 | if (worker->current_pwq != pwq) |
| 4755 | continue; |
| 4756 | |
| 4757 | pr_cont("%s %d%s:%ps", comma ? "," : "", |
| 4758 | task_pid_nr(worker->task), |
| 4759 | worker->rescue_wq ? "(RESCUER)" : "", |
| 4760 | worker->current_func); |
| 4761 | list_for_each_entry(work, &worker->scheduled, entry) |
| 4762 | pr_cont_work(false, work); |
| 4763 | comma = true; |
| 4764 | } |
| 4765 | pr_cont("\n"); |
| 4766 | } |
| 4767 | |
| 4768 | list_for_each_entry(work, &pool->worklist, entry) { |
| 4769 | if (get_work_pwq(work) == pwq) { |
| 4770 | has_pending = true; |
| 4771 | break; |
| 4772 | } |
| 4773 | } |
| 4774 | if (has_pending) { |
| 4775 | bool comma = false; |
| 4776 | |
| 4777 | pr_info(" pending:"); |
| 4778 | list_for_each_entry(work, &pool->worklist, entry) { |
| 4779 | if (get_work_pwq(work) != pwq) |
| 4780 | continue; |
| 4781 | |
| 4782 | pr_cont_work(comma, work); |
| 4783 | comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); |
| 4784 | } |
| 4785 | pr_cont("\n"); |
| 4786 | } |
| 4787 | |
| 4788 | if (!list_empty(&pwq->inactive_works)) { |
| 4789 | bool comma = false; |
| 4790 | |
| 4791 | pr_info(" inactive:"); |
| 4792 | list_for_each_entry(work, &pwq->inactive_works, entry) { |
| 4793 | pr_cont_work(comma, work); |
| 4794 | comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); |
| 4795 | } |
| 4796 | pr_cont("\n"); |
| 4797 | } |
| 4798 | } |
| 4799 | |
| 4800 | /** |
| 4801 | * show_one_workqueue - dump state of specified workqueue |
| 4802 | * @wq: workqueue whose state will be printed |
| 4803 | */ |
| 4804 | void show_one_workqueue(struct workqueue_struct *wq) |
| 4805 | { |
| 4806 | struct pool_workqueue *pwq; |
| 4807 | bool idle = true; |
| 4808 | unsigned long flags; |
| 4809 | |
| 4810 | for_each_pwq(pwq, wq) { |
| 4811 | if (pwq->nr_active || !list_empty(&pwq->inactive_works)) { |
| 4812 | idle = false; |
| 4813 | break; |
| 4814 | } |
| 4815 | } |
| 4816 | if (idle) /* Nothing to print for idle workqueue */ |
| 4817 | return; |
| 4818 | |
| 4819 | pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags); |
| 4820 | |
| 4821 | for_each_pwq(pwq, wq) { |
| 4822 | raw_spin_lock_irqsave(&pwq->pool->lock, flags); |
| 4823 | if (pwq->nr_active || !list_empty(&pwq->inactive_works)) { |
| 4824 | /* |
| 4825 | * Defer printing to avoid deadlocks in console |
| 4826 | * drivers that queue work while holding locks |
| 4827 | * also taken in their write paths. |
| 4828 | */ |
| 4829 | printk_deferred_enter(); |
| 4830 | show_pwq(pwq); |
| 4831 | printk_deferred_exit(); |
| 4832 | } |
| 4833 | raw_spin_unlock_irqrestore(&pwq->pool->lock, flags); |
| 4834 | /* |
| 4835 | * We could be printing a lot from atomic context, e.g. |
| 4836 | * sysrq-t -> show_all_workqueues(). Avoid triggering |
| 4837 | * hard lockup. |
| 4838 | */ |
| 4839 | touch_nmi_watchdog(); |
| 4840 | } |
| 4841 | |
| 4842 | } |
| 4843 | |
| 4844 | /** |
| 4845 | * show_one_worker_pool - dump state of specified worker pool |
| 4846 | * @pool: worker pool whose state will be printed |
| 4847 | */ |
| 4848 | static void show_one_worker_pool(struct worker_pool *pool) |
| 4849 | { |
| 4850 | struct worker *worker; |
| 4851 | bool first = true; |
| 4852 | unsigned long flags; |
| 4853 | |
| 4854 | raw_spin_lock_irqsave(&pool->lock, flags); |
| 4855 | if (pool->nr_workers == pool->nr_idle) |
| 4856 | goto next_pool; |
| 4857 | /* |
| 4858 | * Defer printing to avoid deadlocks in console drivers that |
| 4859 | * queue work while holding locks also taken in their write |
| 4860 | * paths. |
| 4861 | */ |
| 4862 | printk_deferred_enter(); |
| 4863 | pr_info("pool %d:", pool->id); |
| 4864 | pr_cont_pool_info(pool); |
| 4865 | pr_cont(" hung=%us workers=%d", |
| 4866 | jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000, |
| 4867 | pool->nr_workers); |
| 4868 | if (pool->manager) |
| 4869 | pr_cont(" manager: %d", |
| 4870 | task_pid_nr(pool->manager->task)); |
| 4871 | list_for_each_entry(worker, &pool->idle_list, entry) { |
| 4872 | pr_cont(" %s%d", first ? "idle: " : "", |
| 4873 | task_pid_nr(worker->task)); |
| 4874 | first = false; |
| 4875 | } |
| 4876 | pr_cont("\n"); |
| 4877 | printk_deferred_exit(); |
| 4878 | next_pool: |
| 4879 | raw_spin_unlock_irqrestore(&pool->lock, flags); |
| 4880 | /* |
| 4881 | * We could be printing a lot from atomic context, e.g. |
| 4882 | * sysrq-t -> show_all_workqueues(). Avoid triggering |
| 4883 | * hard lockup. |
| 4884 | */ |
| 4885 | touch_nmi_watchdog(); |
| 4886 | |
| 4887 | } |
| 4888 | |
| 4889 | /** |
| 4890 | * show_all_workqueues - dump workqueue state |
| 4891 | * |
| 4892 | * Called from a sysrq handler or try_to_freeze_tasks() and prints out |
| 4893 | * all busy workqueues and pools. |
| 4894 | */ |
| 4895 | void show_all_workqueues(void) |
| 4896 | { |
| 4897 | struct workqueue_struct *wq; |
| 4898 | struct worker_pool *pool; |
| 4899 | int pi; |
| 4900 | |
| 4901 | rcu_read_lock(); |
| 4902 | |
| 4903 | pr_info("Showing busy workqueues and worker pools:\n"); |
| 4904 | |
| 4905 | list_for_each_entry_rcu(wq, &workqueues, list) |
| 4906 | show_one_workqueue(wq); |
| 4907 | |
| 4908 | for_each_pool(pool, pi) |
| 4909 | show_one_worker_pool(pool); |
| 4910 | |
| 4911 | rcu_read_unlock(); |
| 4912 | } |
| 4913 | |
| 4914 | /* used to show worker information through /proc/PID/{comm,stat,status} */ |
| 4915 | void wq_worker_comm(char *buf, size_t size, struct task_struct *task) |
| 4916 | { |
| 4917 | int off; |
| 4918 | |
| 4919 | /* always show the actual comm */ |
| 4920 | off = strscpy(buf, task->comm, size); |
| 4921 | if (off < 0) |
| 4922 | return; |
| 4923 | |
| 4924 | /* stabilize PF_WQ_WORKER and worker pool association */ |
| 4925 | mutex_lock(&wq_pool_attach_mutex); |
| 4926 | |
| 4927 | if (task->flags & PF_WQ_WORKER) { |
| 4928 | struct worker *worker = kthread_data(task); |
| 4929 | struct worker_pool *pool = worker->pool; |
| 4930 | |
| 4931 | if (pool) { |
| 4932 | raw_spin_lock_irq(&pool->lock); |
| 4933 | /* |
| 4934 | * ->desc tracks information (wq name or |
| 4935 | * set_worker_desc()) for the latest execution. If |
| 4936 | * current, prepend '+', otherwise '-'. |
| 4937 | */ |
| 4938 | if (worker->desc[0] != '\0') { |
| 4939 | if (worker->current_work) |
| 4940 | scnprintf(buf + off, size - off, "+%s", |
| 4941 | worker->desc); |
| 4942 | else |
| 4943 | scnprintf(buf + off, size - off, "-%s", |
| 4944 | worker->desc); |
| 4945 | } |
| 4946 | raw_spin_unlock_irq(&pool->lock); |
| 4947 | } |
| 4948 | } |
| 4949 | |
| 4950 | mutex_unlock(&wq_pool_attach_mutex); |
| 4951 | } |
| 4952 | |
| 4953 | #ifdef CONFIG_SMP |
| 4954 | |
| 4955 | /* |
| 4956 | * CPU hotplug. |
| 4957 | * |
| 4958 | * There are two challenges in supporting CPU hotplug. Firstly, there |
| 4959 | * are a lot of assumptions on strong associations among work, pwq and |
| 4960 | * pool which make migrating pending and scheduled works very |
| 4961 | * difficult to implement without impacting hot paths. Secondly, |
| 4962 | * worker pools serve mix of short, long and very long running works making |
| 4963 | * blocked draining impractical. |
| 4964 | * |
| 4965 | * This is solved by allowing the pools to be disassociated from the CPU |
| 4966 | * running as an unbound one and allowing it to be reattached later if the |
| 4967 | * cpu comes back online. |
| 4968 | */ |
| 4969 | |
| 4970 | static void unbind_workers(int cpu) |
| 4971 | { |
| 4972 | struct worker_pool *pool; |
| 4973 | struct worker *worker; |
| 4974 | |
| 4975 | for_each_cpu_worker_pool(pool, cpu) { |
| 4976 | mutex_lock(&wq_pool_attach_mutex); |
| 4977 | raw_spin_lock_irq(&pool->lock); |
| 4978 | |
| 4979 | /* |
| 4980 | * We've blocked all attach/detach operations. Make all workers |
| 4981 | * unbound and set DISASSOCIATED. Before this, all workers |
| 4982 | * must be on the cpu. After this, they may become diasporas. |
| 4983 | * And the preemption disabled section in their sched callbacks |
| 4984 | * are guaranteed to see WORKER_UNBOUND since the code here |
| 4985 | * is on the same cpu. |
| 4986 | */ |
| 4987 | for_each_pool_worker(worker, pool) |
| 4988 | worker->flags |= WORKER_UNBOUND; |
| 4989 | |
| 4990 | pool->flags |= POOL_DISASSOCIATED; |
| 4991 | |
| 4992 | /* |
| 4993 | * The handling of nr_running in sched callbacks are disabled |
| 4994 | * now. Zap nr_running. After this, nr_running stays zero and |
| 4995 | * need_more_worker() and keep_working() are always true as |
| 4996 | * long as the worklist is not empty. This pool now behaves as |
| 4997 | * an unbound (in terms of concurrency management) pool which |
| 4998 | * are served by workers tied to the pool. |
| 4999 | */ |
| 5000 | pool->nr_running = 0; |
| 5001 | |
| 5002 | /* |
| 5003 | * With concurrency management just turned off, a busy |
| 5004 | * worker blocking could lead to lengthy stalls. Kick off |
| 5005 | * unbound chain execution of currently pending work items. |
| 5006 | */ |
| 5007 | wake_up_worker(pool); |
| 5008 | |
| 5009 | raw_spin_unlock_irq(&pool->lock); |
| 5010 | |
| 5011 | for_each_pool_worker(worker, pool) { |
| 5012 | kthread_set_per_cpu(worker->task, -1); |
| 5013 | if (cpumask_intersects(wq_unbound_cpumask, cpu_active_mask)) |
| 5014 | WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, wq_unbound_cpumask) < 0); |
| 5015 | else |
| 5016 | WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, cpu_possible_mask) < 0); |
| 5017 | } |
| 5018 | |
| 5019 | mutex_unlock(&wq_pool_attach_mutex); |
| 5020 | } |
| 5021 | } |
| 5022 | |
| 5023 | /** |
| 5024 | * rebind_workers - rebind all workers of a pool to the associated CPU |
| 5025 | * @pool: pool of interest |
| 5026 | * |
| 5027 | * @pool->cpu is coming online. Rebind all workers to the CPU. |
| 5028 | */ |
| 5029 | static void rebind_workers(struct worker_pool *pool) |
| 5030 | { |
| 5031 | struct worker *worker; |
| 5032 | |
| 5033 | lockdep_assert_held(&wq_pool_attach_mutex); |
| 5034 | |
| 5035 | /* |
| 5036 | * Restore CPU affinity of all workers. As all idle workers should |
| 5037 | * be on the run-queue of the associated CPU before any local |
| 5038 | * wake-ups for concurrency management happen, restore CPU affinity |
| 5039 | * of all workers first and then clear UNBOUND. As we're called |
| 5040 | * from CPU_ONLINE, the following shouldn't fail. |
| 5041 | */ |
| 5042 | for_each_pool_worker(worker, pool) { |
| 5043 | kthread_set_per_cpu(worker->task, pool->cpu); |
| 5044 | WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, |
| 5045 | pool->attrs->cpumask) < 0); |
| 5046 | } |
| 5047 | |
| 5048 | raw_spin_lock_irq(&pool->lock); |
| 5049 | |
| 5050 | pool->flags &= ~POOL_DISASSOCIATED; |
| 5051 | |
| 5052 | for_each_pool_worker(worker, pool) { |
| 5053 | unsigned int worker_flags = worker->flags; |
| 5054 | |
| 5055 | /* |
| 5056 | * We want to clear UNBOUND but can't directly call |
| 5057 | * worker_clr_flags() or adjust nr_running. Atomically |
| 5058 | * replace UNBOUND with another NOT_RUNNING flag REBOUND. |
| 5059 | * @worker will clear REBOUND using worker_clr_flags() when |
| 5060 | * it initiates the next execution cycle thus restoring |
| 5061 | * concurrency management. Note that when or whether |
| 5062 | * @worker clears REBOUND doesn't affect correctness. |
| 5063 | * |
| 5064 | * WRITE_ONCE() is necessary because @worker->flags may be |
| 5065 | * tested without holding any lock in |
| 5066 | * wq_worker_running(). Without it, NOT_RUNNING test may |
| 5067 | * fail incorrectly leading to premature concurrency |
| 5068 | * management operations. |
| 5069 | */ |
| 5070 | WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND)); |
| 5071 | worker_flags |= WORKER_REBOUND; |
| 5072 | worker_flags &= ~WORKER_UNBOUND; |
| 5073 | WRITE_ONCE(worker->flags, worker_flags); |
| 5074 | } |
| 5075 | |
| 5076 | raw_spin_unlock_irq(&pool->lock); |
| 5077 | } |
| 5078 | |
| 5079 | /** |
| 5080 | * restore_unbound_workers_cpumask - restore cpumask of unbound workers |
| 5081 | * @pool: unbound pool of interest |
| 5082 | * @cpu: the CPU which is coming up |
| 5083 | * |
| 5084 | * An unbound pool may end up with a cpumask which doesn't have any online |
| 5085 | * CPUs. When a worker of such pool get scheduled, the scheduler resets |
| 5086 | * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any |
| 5087 | * online CPU before, cpus_allowed of all its workers should be restored. |
| 5088 | */ |
| 5089 | static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu) |
| 5090 | { |
| 5091 | static cpumask_t cpumask; |
| 5092 | struct worker *worker; |
| 5093 | |
| 5094 | lockdep_assert_held(&wq_pool_attach_mutex); |
| 5095 | |
| 5096 | /* is @cpu allowed for @pool? */ |
| 5097 | if (!cpumask_test_cpu(cpu, pool->attrs->cpumask)) |
| 5098 | return; |
| 5099 | |
| 5100 | cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask); |
| 5101 | |
| 5102 | /* as we're called from CPU_ONLINE, the following shouldn't fail */ |
| 5103 | for_each_pool_worker(worker, pool) |
| 5104 | WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0); |
| 5105 | } |
| 5106 | |
| 5107 | int workqueue_prepare_cpu(unsigned int cpu) |
| 5108 | { |
| 5109 | struct worker_pool *pool; |
| 5110 | |
| 5111 | for_each_cpu_worker_pool(pool, cpu) { |
| 5112 | if (pool->nr_workers) |
| 5113 | continue; |
| 5114 | if (!create_worker(pool)) |
| 5115 | return -ENOMEM; |
| 5116 | } |
| 5117 | return 0; |
| 5118 | } |
| 5119 | |
| 5120 | int workqueue_online_cpu(unsigned int cpu) |
| 5121 | { |
| 5122 | struct worker_pool *pool; |
| 5123 | struct workqueue_struct *wq; |
| 5124 | int pi; |
| 5125 | |
| 5126 | mutex_lock(&wq_pool_mutex); |
| 5127 | |
| 5128 | for_each_pool(pool, pi) { |
| 5129 | mutex_lock(&wq_pool_attach_mutex); |
| 5130 | |
| 5131 | if (pool->cpu == cpu) |
| 5132 | rebind_workers(pool); |
| 5133 | else if (pool->cpu < 0) |
| 5134 | restore_unbound_workers_cpumask(pool, cpu); |
| 5135 | |
| 5136 | mutex_unlock(&wq_pool_attach_mutex); |
| 5137 | } |
| 5138 | |
| 5139 | /* update NUMA affinity of unbound workqueues */ |
| 5140 | list_for_each_entry(wq, &workqueues, list) |
| 5141 | wq_update_unbound_numa(wq, cpu, true); |
| 5142 | |
| 5143 | mutex_unlock(&wq_pool_mutex); |
| 5144 | return 0; |
| 5145 | } |
| 5146 | |
| 5147 | int workqueue_offline_cpu(unsigned int cpu) |
| 5148 | { |
| 5149 | struct workqueue_struct *wq; |
| 5150 | |
| 5151 | /* unbinding per-cpu workers should happen on the local CPU */ |
| 5152 | if (WARN_ON(cpu != smp_processor_id())) |
| 5153 | return -1; |
| 5154 | |
| 5155 | unbind_workers(cpu); |
| 5156 | |
| 5157 | /* update NUMA affinity of unbound workqueues */ |
| 5158 | mutex_lock(&wq_pool_mutex); |
| 5159 | list_for_each_entry(wq, &workqueues, list) |
| 5160 | wq_update_unbound_numa(wq, cpu, false); |
| 5161 | mutex_unlock(&wq_pool_mutex); |
| 5162 | |
| 5163 | return 0; |
| 5164 | } |
| 5165 | |
| 5166 | struct work_for_cpu { |
| 5167 | struct work_struct work; |
| 5168 | long (*fn)(void *); |
| 5169 | void *arg; |
| 5170 | long ret; |
| 5171 | }; |
| 5172 | |
| 5173 | static void work_for_cpu_fn(struct work_struct *work) |
| 5174 | { |
| 5175 | struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work); |
| 5176 | |
| 5177 | wfc->ret = wfc->fn(wfc->arg); |
| 5178 | } |
| 5179 | |
| 5180 | /** |
| 5181 | * work_on_cpu - run a function in thread context on a particular cpu |
| 5182 | * @cpu: the cpu to run on |
| 5183 | * @fn: the function to run |
| 5184 | * @arg: the function arg |
| 5185 | * |
| 5186 | * It is up to the caller to ensure that the cpu doesn't go offline. |
| 5187 | * The caller must not hold any locks which would prevent @fn from completing. |
| 5188 | * |
| 5189 | * Return: The value @fn returns. |
| 5190 | */ |
| 5191 | long work_on_cpu(int cpu, long (*fn)(void *), void *arg) |
| 5192 | { |
| 5193 | struct work_for_cpu wfc = { .fn = fn, .arg = arg }; |
| 5194 | |
| 5195 | INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn); |
| 5196 | schedule_work_on(cpu, &wfc.work); |
| 5197 | flush_work(&wfc.work); |
| 5198 | destroy_work_on_stack(&wfc.work); |
| 5199 | return wfc.ret; |
| 5200 | } |
| 5201 | EXPORT_SYMBOL_GPL(work_on_cpu); |
| 5202 | |
| 5203 | /** |
| 5204 | * work_on_cpu_safe - run a function in thread context on a particular cpu |
| 5205 | * @cpu: the cpu to run on |
| 5206 | * @fn: the function to run |
| 5207 | * @arg: the function argument |
| 5208 | * |
| 5209 | * Disables CPU hotplug and calls work_on_cpu(). The caller must not hold |
| 5210 | * any locks which would prevent @fn from completing. |
| 5211 | * |
| 5212 | * Return: The value @fn returns. |
| 5213 | */ |
| 5214 | long work_on_cpu_safe(int cpu, long (*fn)(void *), void *arg) |
| 5215 | { |
| 5216 | long ret = -ENODEV; |
| 5217 | |
| 5218 | cpus_read_lock(); |
| 5219 | if (cpu_online(cpu)) |
| 5220 | ret = work_on_cpu(cpu, fn, arg); |
| 5221 | cpus_read_unlock(); |
| 5222 | return ret; |
| 5223 | } |
| 5224 | EXPORT_SYMBOL_GPL(work_on_cpu_safe); |
| 5225 | #endif /* CONFIG_SMP */ |
| 5226 | |
| 5227 | #ifdef CONFIG_FREEZER |
| 5228 | |
| 5229 | /** |
| 5230 | * freeze_workqueues_begin - begin freezing workqueues |
| 5231 | * |
| 5232 | * Start freezing workqueues. After this function returns, all freezable |
| 5233 | * workqueues will queue new works to their inactive_works list instead of |
| 5234 | * pool->worklist. |
| 5235 | * |
| 5236 | * CONTEXT: |
| 5237 | * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. |
| 5238 | */ |
| 5239 | void freeze_workqueues_begin(void) |
| 5240 | { |
| 5241 | struct workqueue_struct *wq; |
| 5242 | struct pool_workqueue *pwq; |
| 5243 | |
| 5244 | mutex_lock(&wq_pool_mutex); |
| 5245 | |
| 5246 | WARN_ON_ONCE(workqueue_freezing); |
| 5247 | workqueue_freezing = true; |
| 5248 | |
| 5249 | list_for_each_entry(wq, &workqueues, list) { |
| 5250 | mutex_lock(&wq->mutex); |
| 5251 | for_each_pwq(pwq, wq) |
| 5252 | pwq_adjust_max_active(pwq); |
| 5253 | mutex_unlock(&wq->mutex); |
| 5254 | } |
| 5255 | |
| 5256 | mutex_unlock(&wq_pool_mutex); |
| 5257 | } |
| 5258 | |
| 5259 | /** |
| 5260 | * freeze_workqueues_busy - are freezable workqueues still busy? |
| 5261 | * |
| 5262 | * Check whether freezing is complete. This function must be called |
| 5263 | * between freeze_workqueues_begin() and thaw_workqueues(). |
| 5264 | * |
| 5265 | * CONTEXT: |
| 5266 | * Grabs and releases wq_pool_mutex. |
| 5267 | * |
| 5268 | * Return: |
| 5269 | * %true if some freezable workqueues are still busy. %false if freezing |
| 5270 | * is complete. |
| 5271 | */ |
| 5272 | bool freeze_workqueues_busy(void) |
| 5273 | { |
| 5274 | bool busy = false; |
| 5275 | struct workqueue_struct *wq; |
| 5276 | struct pool_workqueue *pwq; |
| 5277 | |
| 5278 | mutex_lock(&wq_pool_mutex); |
| 5279 | |
| 5280 | WARN_ON_ONCE(!workqueue_freezing); |
| 5281 | |
| 5282 | list_for_each_entry(wq, &workqueues, list) { |
| 5283 | if (!(wq->flags & WQ_FREEZABLE)) |
| 5284 | continue; |
| 5285 | /* |
| 5286 | * nr_active is monotonically decreasing. It's safe |
| 5287 | * to peek without lock. |
| 5288 | */ |
| 5289 | rcu_read_lock(); |
| 5290 | for_each_pwq(pwq, wq) { |
| 5291 | WARN_ON_ONCE(pwq->nr_active < 0); |
| 5292 | if (pwq->nr_active) { |
| 5293 | busy = true; |
| 5294 | rcu_read_unlock(); |
| 5295 | goto out_unlock; |
| 5296 | } |
| 5297 | } |
| 5298 | rcu_read_unlock(); |
| 5299 | } |
| 5300 | out_unlock: |
| 5301 | mutex_unlock(&wq_pool_mutex); |
| 5302 | return busy; |
| 5303 | } |
| 5304 | |
| 5305 | /** |
| 5306 | * thaw_workqueues - thaw workqueues |
| 5307 | * |
| 5308 | * Thaw workqueues. Normal queueing is restored and all collected |
| 5309 | * frozen works are transferred to their respective pool worklists. |
| 5310 | * |
| 5311 | * CONTEXT: |
| 5312 | * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. |
| 5313 | */ |
| 5314 | void thaw_workqueues(void) |
| 5315 | { |
| 5316 | struct workqueue_struct *wq; |
| 5317 | struct pool_workqueue *pwq; |
| 5318 | |
| 5319 | mutex_lock(&wq_pool_mutex); |
| 5320 | |
| 5321 | if (!workqueue_freezing) |
| 5322 | goto out_unlock; |
| 5323 | |
| 5324 | workqueue_freezing = false; |
| 5325 | |
| 5326 | /* restore max_active and repopulate worklist */ |
| 5327 | list_for_each_entry(wq, &workqueues, list) { |
| 5328 | mutex_lock(&wq->mutex); |
| 5329 | for_each_pwq(pwq, wq) |
| 5330 | pwq_adjust_max_active(pwq); |
| 5331 | mutex_unlock(&wq->mutex); |
| 5332 | } |
| 5333 | |
| 5334 | out_unlock: |
| 5335 | mutex_unlock(&wq_pool_mutex); |
| 5336 | } |
| 5337 | #endif /* CONFIG_FREEZER */ |
| 5338 | |
| 5339 | static int workqueue_apply_unbound_cpumask(void) |
| 5340 | { |
| 5341 | LIST_HEAD(ctxs); |
| 5342 | int ret = 0; |
| 5343 | struct workqueue_struct *wq; |
| 5344 | struct apply_wqattrs_ctx *ctx, *n; |
| 5345 | |
| 5346 | lockdep_assert_held(&wq_pool_mutex); |
| 5347 | |
| 5348 | list_for_each_entry(wq, &workqueues, list) { |
| 5349 | if (!(wq->flags & WQ_UNBOUND)) |
| 5350 | continue; |
| 5351 | /* creating multiple pwqs breaks ordering guarantee */ |
| 5352 | if (wq->flags & __WQ_ORDERED) |
| 5353 | continue; |
| 5354 | |
| 5355 | ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs); |
| 5356 | if (!ctx) { |
| 5357 | ret = -ENOMEM; |
| 5358 | break; |
| 5359 | } |
| 5360 | |
| 5361 | list_add_tail(&ctx->list, &ctxs); |
| 5362 | } |
| 5363 | |
| 5364 | list_for_each_entry_safe(ctx, n, &ctxs, list) { |
| 5365 | if (!ret) |
| 5366 | apply_wqattrs_commit(ctx); |
| 5367 | apply_wqattrs_cleanup(ctx); |
| 5368 | } |
| 5369 | |
| 5370 | return ret; |
| 5371 | } |
| 5372 | |
| 5373 | /** |
| 5374 | * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask |
| 5375 | * @cpumask: the cpumask to set |
| 5376 | * |
| 5377 | * The low-level workqueues cpumask is a global cpumask that limits |
| 5378 | * the affinity of all unbound workqueues. This function check the @cpumask |
| 5379 | * and apply it to all unbound workqueues and updates all pwqs of them. |
| 5380 | * |
| 5381 | * Return: 0 - Success |
| 5382 | * -EINVAL - Invalid @cpumask |
| 5383 | * -ENOMEM - Failed to allocate memory for attrs or pwqs. |
| 5384 | */ |
| 5385 | int workqueue_set_unbound_cpumask(cpumask_var_t cpumask) |
| 5386 | { |
| 5387 | int ret = -EINVAL; |
| 5388 | cpumask_var_t saved_cpumask; |
| 5389 | |
| 5390 | /* |
| 5391 | * Not excluding isolated cpus on purpose. |
| 5392 | * If the user wishes to include them, we allow that. |
| 5393 | */ |
| 5394 | cpumask_and(cpumask, cpumask, cpu_possible_mask); |
| 5395 | if (!cpumask_empty(cpumask)) { |
| 5396 | apply_wqattrs_lock(); |
| 5397 | if (cpumask_equal(cpumask, wq_unbound_cpumask)) { |
| 5398 | ret = 0; |
| 5399 | goto out_unlock; |
| 5400 | } |
| 5401 | |
| 5402 | if (!zalloc_cpumask_var(&saved_cpumask, GFP_KERNEL)) { |
| 5403 | ret = -ENOMEM; |
| 5404 | goto out_unlock; |
| 5405 | } |
| 5406 | |
| 5407 | /* save the old wq_unbound_cpumask. */ |
| 5408 | cpumask_copy(saved_cpumask, wq_unbound_cpumask); |
| 5409 | |
| 5410 | /* update wq_unbound_cpumask at first and apply it to wqs. */ |
| 5411 | cpumask_copy(wq_unbound_cpumask, cpumask); |
| 5412 | ret = workqueue_apply_unbound_cpumask(); |
| 5413 | |
| 5414 | /* restore the wq_unbound_cpumask when failed. */ |
| 5415 | if (ret < 0) |
| 5416 | cpumask_copy(wq_unbound_cpumask, saved_cpumask); |
| 5417 | |
| 5418 | free_cpumask_var(saved_cpumask); |
| 5419 | out_unlock: |
| 5420 | apply_wqattrs_unlock(); |
| 5421 | } |
| 5422 | |
| 5423 | return ret; |
| 5424 | } |
| 5425 | |
| 5426 | #ifdef CONFIG_SYSFS |
| 5427 | /* |
| 5428 | * Workqueues with WQ_SYSFS flag set is visible to userland via |
| 5429 | * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the |
| 5430 | * following attributes. |
| 5431 | * |
| 5432 | * per_cpu RO bool : whether the workqueue is per-cpu or unbound |
| 5433 | * max_active RW int : maximum number of in-flight work items |
| 5434 | * |
| 5435 | * Unbound workqueues have the following extra attributes. |
| 5436 | * |
| 5437 | * pool_ids RO int : the associated pool IDs for each node |
| 5438 | * nice RW int : nice value of the workers |
| 5439 | * cpumask RW mask : bitmask of allowed CPUs for the workers |
| 5440 | * numa RW bool : whether enable NUMA affinity |
| 5441 | */ |
| 5442 | struct wq_device { |
| 5443 | struct workqueue_struct *wq; |
| 5444 | struct device dev; |
| 5445 | }; |
| 5446 | |
| 5447 | static struct workqueue_struct *dev_to_wq(struct device *dev) |
| 5448 | { |
| 5449 | struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); |
| 5450 | |
| 5451 | return wq_dev->wq; |
| 5452 | } |
| 5453 | |
| 5454 | static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr, |
| 5455 | char *buf) |
| 5456 | { |
| 5457 | struct workqueue_struct *wq = dev_to_wq(dev); |
| 5458 | |
| 5459 | return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND)); |
| 5460 | } |
| 5461 | static DEVICE_ATTR_RO(per_cpu); |
| 5462 | |
| 5463 | static ssize_t max_active_show(struct device *dev, |
| 5464 | struct device_attribute *attr, char *buf) |
| 5465 | { |
| 5466 | struct workqueue_struct *wq = dev_to_wq(dev); |
| 5467 | |
| 5468 | return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active); |
| 5469 | } |
| 5470 | |
| 5471 | static ssize_t max_active_store(struct device *dev, |
| 5472 | struct device_attribute *attr, const char *buf, |
| 5473 | size_t count) |
| 5474 | { |
| 5475 | struct workqueue_struct *wq = dev_to_wq(dev); |
| 5476 | int val; |
| 5477 | |
| 5478 | if (sscanf(buf, "%d", &val) != 1 || val <= 0) |
| 5479 | return -EINVAL; |
| 5480 | |
| 5481 | workqueue_set_max_active(wq, val); |
| 5482 | return count; |
| 5483 | } |
| 5484 | static DEVICE_ATTR_RW(max_active); |
| 5485 | |
| 5486 | static struct attribute *wq_sysfs_attrs[] = { |
| 5487 | &dev_attr_per_cpu.attr, |
| 5488 | &dev_attr_max_active.attr, |
| 5489 | NULL, |
| 5490 | }; |
| 5491 | ATTRIBUTE_GROUPS(wq_sysfs); |
| 5492 | |
| 5493 | static ssize_t wq_pool_ids_show(struct device *dev, |
| 5494 | struct device_attribute *attr, char *buf) |
| 5495 | { |
| 5496 | struct workqueue_struct *wq = dev_to_wq(dev); |
| 5497 | const char *delim = ""; |
| 5498 | int node, written = 0; |
| 5499 | |
| 5500 | cpus_read_lock(); |
| 5501 | rcu_read_lock(); |
| 5502 | for_each_node(node) { |
| 5503 | written += scnprintf(buf + written, PAGE_SIZE - written, |
| 5504 | "%s%d:%d", delim, node, |
| 5505 | unbound_pwq_by_node(wq, node)->pool->id); |
| 5506 | delim = " "; |
| 5507 | } |
| 5508 | written += scnprintf(buf + written, PAGE_SIZE - written, "\n"); |
| 5509 | rcu_read_unlock(); |
| 5510 | cpus_read_unlock(); |
| 5511 | |
| 5512 | return written; |
| 5513 | } |
| 5514 | |
| 5515 | static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr, |
| 5516 | char *buf) |
| 5517 | { |
| 5518 | struct workqueue_struct *wq = dev_to_wq(dev); |
| 5519 | int written; |
| 5520 | |
| 5521 | mutex_lock(&wq->mutex); |
| 5522 | written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice); |
| 5523 | mutex_unlock(&wq->mutex); |
| 5524 | |
| 5525 | return written; |
| 5526 | } |
| 5527 | |
| 5528 | /* prepare workqueue_attrs for sysfs store operations */ |
| 5529 | static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq) |
| 5530 | { |
| 5531 | struct workqueue_attrs *attrs; |
| 5532 | |
| 5533 | lockdep_assert_held(&wq_pool_mutex); |
| 5534 | |
| 5535 | attrs = alloc_workqueue_attrs(); |
| 5536 | if (!attrs) |
| 5537 | return NULL; |
| 5538 | |
| 5539 | copy_workqueue_attrs(attrs, wq->unbound_attrs); |
| 5540 | return attrs; |
| 5541 | } |
| 5542 | |
| 5543 | static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr, |
| 5544 | const char *buf, size_t count) |
| 5545 | { |
| 5546 | struct workqueue_struct *wq = dev_to_wq(dev); |
| 5547 | struct workqueue_attrs *attrs; |
| 5548 | int ret = -ENOMEM; |
| 5549 | |
| 5550 | apply_wqattrs_lock(); |
| 5551 | |
| 5552 | attrs = wq_sysfs_prep_attrs(wq); |
| 5553 | if (!attrs) |
| 5554 | goto out_unlock; |
| 5555 | |
| 5556 | if (sscanf(buf, "%d", &attrs->nice) == 1 && |
| 5557 | attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE) |
| 5558 | ret = apply_workqueue_attrs_locked(wq, attrs); |
| 5559 | else |
| 5560 | ret = -EINVAL; |
| 5561 | |
| 5562 | out_unlock: |
| 5563 | apply_wqattrs_unlock(); |
| 5564 | free_workqueue_attrs(attrs); |
| 5565 | return ret ?: count; |
| 5566 | } |
| 5567 | |
| 5568 | static ssize_t wq_cpumask_show(struct device *dev, |
| 5569 | struct device_attribute *attr, char *buf) |
| 5570 | { |
| 5571 | struct workqueue_struct *wq = dev_to_wq(dev); |
| 5572 | int written; |
| 5573 | |
| 5574 | mutex_lock(&wq->mutex); |
| 5575 | written = scnprintf(buf, PAGE_SIZE, "%*pb\n", |
| 5576 | cpumask_pr_args(wq->unbound_attrs->cpumask)); |
| 5577 | mutex_unlock(&wq->mutex); |
| 5578 | return written; |
| 5579 | } |
| 5580 | |
| 5581 | static ssize_t wq_cpumask_store(struct device *dev, |
| 5582 | struct device_attribute *attr, |
| 5583 | const char *buf, size_t count) |
| 5584 | { |
| 5585 | struct workqueue_struct *wq = dev_to_wq(dev); |
| 5586 | struct workqueue_attrs *attrs; |
| 5587 | int ret = -ENOMEM; |
| 5588 | |
| 5589 | apply_wqattrs_lock(); |
| 5590 | |
| 5591 | attrs = wq_sysfs_prep_attrs(wq); |
| 5592 | if (!attrs) |
| 5593 | goto out_unlock; |
| 5594 | |
| 5595 | ret = cpumask_parse(buf, attrs->cpumask); |
| 5596 | if (!ret) |
| 5597 | ret = apply_workqueue_attrs_locked(wq, attrs); |
| 5598 | |
| 5599 | out_unlock: |
| 5600 | apply_wqattrs_unlock(); |
| 5601 | free_workqueue_attrs(attrs); |
| 5602 | return ret ?: count; |
| 5603 | } |
| 5604 | |
| 5605 | static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr, |
| 5606 | char *buf) |
| 5607 | { |
| 5608 | struct workqueue_struct *wq = dev_to_wq(dev); |
| 5609 | int written; |
| 5610 | |
| 5611 | mutex_lock(&wq->mutex); |
| 5612 | written = scnprintf(buf, PAGE_SIZE, "%d\n", |
| 5613 | !wq->unbound_attrs->no_numa); |
| 5614 | mutex_unlock(&wq->mutex); |
| 5615 | |
| 5616 | return written; |
| 5617 | } |
| 5618 | |
| 5619 | static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr, |
| 5620 | const char *buf, size_t count) |
| 5621 | { |
| 5622 | struct workqueue_struct *wq = dev_to_wq(dev); |
| 5623 | struct workqueue_attrs *attrs; |
| 5624 | int v, ret = -ENOMEM; |
| 5625 | |
| 5626 | apply_wqattrs_lock(); |
| 5627 | |
| 5628 | attrs = wq_sysfs_prep_attrs(wq); |
| 5629 | if (!attrs) |
| 5630 | goto out_unlock; |
| 5631 | |
| 5632 | ret = -EINVAL; |
| 5633 | if (sscanf(buf, "%d", &v) == 1) { |
| 5634 | attrs->no_numa = !v; |
| 5635 | ret = apply_workqueue_attrs_locked(wq, attrs); |
| 5636 | } |
| 5637 | |
| 5638 | out_unlock: |
| 5639 | apply_wqattrs_unlock(); |
| 5640 | free_workqueue_attrs(attrs); |
| 5641 | return ret ?: count; |
| 5642 | } |
| 5643 | |
| 5644 | static struct device_attribute wq_sysfs_unbound_attrs[] = { |
| 5645 | __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL), |
| 5646 | __ATTR(nice, 0644, wq_nice_show, wq_nice_store), |
| 5647 | __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store), |
| 5648 | __ATTR(numa, 0644, wq_numa_show, wq_numa_store), |
| 5649 | __ATTR_NULL, |
| 5650 | }; |
| 5651 | |
| 5652 | static struct bus_type wq_subsys = { |
| 5653 | .name = "workqueue", |
| 5654 | .dev_groups = wq_sysfs_groups, |
| 5655 | }; |
| 5656 | |
| 5657 | static ssize_t wq_unbound_cpumask_show(struct device *dev, |
| 5658 | struct device_attribute *attr, char *buf) |
| 5659 | { |
| 5660 | int written; |
| 5661 | |
| 5662 | mutex_lock(&wq_pool_mutex); |
| 5663 | written = scnprintf(buf, PAGE_SIZE, "%*pb\n", |
| 5664 | cpumask_pr_args(wq_unbound_cpumask)); |
| 5665 | mutex_unlock(&wq_pool_mutex); |
| 5666 | |
| 5667 | return written; |
| 5668 | } |
| 5669 | |
| 5670 | static ssize_t wq_unbound_cpumask_store(struct device *dev, |
| 5671 | struct device_attribute *attr, const char *buf, size_t count) |
| 5672 | { |
| 5673 | cpumask_var_t cpumask; |
| 5674 | int ret; |
| 5675 | |
| 5676 | if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL)) |
| 5677 | return -ENOMEM; |
| 5678 | |
| 5679 | ret = cpumask_parse(buf, cpumask); |
| 5680 | if (!ret) |
| 5681 | ret = workqueue_set_unbound_cpumask(cpumask); |
| 5682 | |
| 5683 | free_cpumask_var(cpumask); |
| 5684 | return ret ? ret : count; |
| 5685 | } |
| 5686 | |
| 5687 | static struct device_attribute wq_sysfs_cpumask_attr = |
| 5688 | __ATTR(cpumask, 0644, wq_unbound_cpumask_show, |
| 5689 | wq_unbound_cpumask_store); |
| 5690 | |
| 5691 | static int __init wq_sysfs_init(void) |
| 5692 | { |
| 5693 | int err; |
| 5694 | |
| 5695 | err = subsys_virtual_register(&wq_subsys, NULL); |
| 5696 | if (err) |
| 5697 | return err; |
| 5698 | |
| 5699 | return device_create_file(wq_subsys.dev_root, &wq_sysfs_cpumask_attr); |
| 5700 | } |
| 5701 | core_initcall(wq_sysfs_init); |
| 5702 | |
| 5703 | static void wq_device_release(struct device *dev) |
| 5704 | { |
| 5705 | struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); |
| 5706 | |
| 5707 | kfree(wq_dev); |
| 5708 | } |
| 5709 | |
| 5710 | /** |
| 5711 | * workqueue_sysfs_register - make a workqueue visible in sysfs |
| 5712 | * @wq: the workqueue to register |
| 5713 | * |
| 5714 | * Expose @wq in sysfs under /sys/bus/workqueue/devices. |
| 5715 | * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set |
| 5716 | * which is the preferred method. |
| 5717 | * |
| 5718 | * Workqueue user should use this function directly iff it wants to apply |
| 5719 | * workqueue_attrs before making the workqueue visible in sysfs; otherwise, |
| 5720 | * apply_workqueue_attrs() may race against userland updating the |
| 5721 | * attributes. |
| 5722 | * |
| 5723 | * Return: 0 on success, -errno on failure. |
| 5724 | */ |
| 5725 | int workqueue_sysfs_register(struct workqueue_struct *wq) |
| 5726 | { |
| 5727 | struct wq_device *wq_dev; |
| 5728 | int ret; |
| 5729 | |
| 5730 | /* |
| 5731 | * Adjusting max_active or creating new pwqs by applying |
| 5732 | * attributes breaks ordering guarantee. Disallow exposing ordered |
| 5733 | * workqueues. |
| 5734 | */ |
| 5735 | if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) |
| 5736 | return -EINVAL; |
| 5737 | |
| 5738 | wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL); |
| 5739 | if (!wq_dev) |
| 5740 | return -ENOMEM; |
| 5741 | |
| 5742 | wq_dev->wq = wq; |
| 5743 | wq_dev->dev.bus = &wq_subsys; |
| 5744 | wq_dev->dev.release = wq_device_release; |
| 5745 | dev_set_name(&wq_dev->dev, "%s", wq->name); |
| 5746 | |
| 5747 | /* |
| 5748 | * unbound_attrs are created separately. Suppress uevent until |
| 5749 | * everything is ready. |
| 5750 | */ |
| 5751 | dev_set_uevent_suppress(&wq_dev->dev, true); |
| 5752 | |
| 5753 | ret = device_register(&wq_dev->dev); |
| 5754 | if (ret) { |
| 5755 | put_device(&wq_dev->dev); |
| 5756 | wq->wq_dev = NULL; |
| 5757 | return ret; |
| 5758 | } |
| 5759 | |
| 5760 | if (wq->flags & WQ_UNBOUND) { |
| 5761 | struct device_attribute *attr; |
| 5762 | |
| 5763 | for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) { |
| 5764 | ret = device_create_file(&wq_dev->dev, attr); |
| 5765 | if (ret) { |
| 5766 | device_unregister(&wq_dev->dev); |
| 5767 | wq->wq_dev = NULL; |
| 5768 | return ret; |
| 5769 | } |
| 5770 | } |
| 5771 | } |
| 5772 | |
| 5773 | dev_set_uevent_suppress(&wq_dev->dev, false); |
| 5774 | kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD); |
| 5775 | return 0; |
| 5776 | } |
| 5777 | |
| 5778 | /** |
| 5779 | * workqueue_sysfs_unregister - undo workqueue_sysfs_register() |
| 5780 | * @wq: the workqueue to unregister |
| 5781 | * |
| 5782 | * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister. |
| 5783 | */ |
| 5784 | static void workqueue_sysfs_unregister(struct workqueue_struct *wq) |
| 5785 | { |
| 5786 | struct wq_device *wq_dev = wq->wq_dev; |
| 5787 | |
| 5788 | if (!wq->wq_dev) |
| 5789 | return; |
| 5790 | |
| 5791 | wq->wq_dev = NULL; |
| 5792 | device_unregister(&wq_dev->dev); |
| 5793 | } |
| 5794 | #else /* CONFIG_SYSFS */ |
| 5795 | static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { } |
| 5796 | #endif /* CONFIG_SYSFS */ |
| 5797 | |
| 5798 | /* |
| 5799 | * Workqueue watchdog. |
| 5800 | * |
| 5801 | * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal |
| 5802 | * flush dependency, a concurrency managed work item which stays RUNNING |
| 5803 | * indefinitely. Workqueue stalls can be very difficult to debug as the |
| 5804 | * usual warning mechanisms don't trigger and internal workqueue state is |
| 5805 | * largely opaque. |
| 5806 | * |
| 5807 | * Workqueue watchdog monitors all worker pools periodically and dumps |
| 5808 | * state if some pools failed to make forward progress for a while where |
| 5809 | * forward progress is defined as the first item on ->worklist changing. |
| 5810 | * |
| 5811 | * This mechanism is controlled through the kernel parameter |
| 5812 | * "workqueue.watchdog_thresh" which can be updated at runtime through the |
| 5813 | * corresponding sysfs parameter file. |
| 5814 | */ |
| 5815 | #ifdef CONFIG_WQ_WATCHDOG |
| 5816 | |
| 5817 | static unsigned long wq_watchdog_thresh = 30; |
| 5818 | static struct timer_list wq_watchdog_timer; |
| 5819 | |
| 5820 | static unsigned long wq_watchdog_touched = INITIAL_JIFFIES; |
| 5821 | static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES; |
| 5822 | |
| 5823 | static void wq_watchdog_reset_touched(void) |
| 5824 | { |
| 5825 | int cpu; |
| 5826 | |
| 5827 | wq_watchdog_touched = jiffies; |
| 5828 | for_each_possible_cpu(cpu) |
| 5829 | per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies; |
| 5830 | } |
| 5831 | |
| 5832 | static void wq_watchdog_timer_fn(struct timer_list *unused) |
| 5833 | { |
| 5834 | unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ; |
| 5835 | bool lockup_detected = false; |
| 5836 | unsigned long now = jiffies; |
| 5837 | struct worker_pool *pool; |
| 5838 | int pi; |
| 5839 | |
| 5840 | if (!thresh) |
| 5841 | return; |
| 5842 | |
| 5843 | rcu_read_lock(); |
| 5844 | |
| 5845 | for_each_pool(pool, pi) { |
| 5846 | unsigned long pool_ts, touched, ts; |
| 5847 | |
| 5848 | if (list_empty(&pool->worklist)) |
| 5849 | continue; |
| 5850 | |
| 5851 | /* |
| 5852 | * If a virtual machine is stopped by the host it can look to |
| 5853 | * the watchdog like a stall. |
| 5854 | */ |
| 5855 | kvm_check_and_clear_guest_paused(); |
| 5856 | |
| 5857 | /* get the latest of pool and touched timestamps */ |
| 5858 | if (pool->cpu >= 0) |
| 5859 | touched = READ_ONCE(per_cpu(wq_watchdog_touched_cpu, pool->cpu)); |
| 5860 | else |
| 5861 | touched = READ_ONCE(wq_watchdog_touched); |
| 5862 | pool_ts = READ_ONCE(pool->watchdog_ts); |
| 5863 | |
| 5864 | if (time_after(pool_ts, touched)) |
| 5865 | ts = pool_ts; |
| 5866 | else |
| 5867 | ts = touched; |
| 5868 | |
| 5869 | /* did we stall? */ |
| 5870 | if (time_after(now, ts + thresh)) { |
| 5871 | lockup_detected = true; |
| 5872 | pr_emerg("BUG: workqueue lockup - pool"); |
| 5873 | pr_cont_pool_info(pool); |
| 5874 | pr_cont(" stuck for %us!\n", |
| 5875 | jiffies_to_msecs(now - pool_ts) / 1000); |
| 5876 | } |
| 5877 | } |
| 5878 | |
| 5879 | rcu_read_unlock(); |
| 5880 | |
| 5881 | if (lockup_detected) |
| 5882 | show_all_workqueues(); |
| 5883 | |
| 5884 | wq_watchdog_reset_touched(); |
| 5885 | mod_timer(&wq_watchdog_timer, jiffies + thresh); |
| 5886 | } |
| 5887 | |
| 5888 | notrace void wq_watchdog_touch(int cpu) |
| 5889 | { |
| 5890 | if (cpu >= 0) |
| 5891 | per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies; |
| 5892 | |
| 5893 | wq_watchdog_touched = jiffies; |
| 5894 | } |
| 5895 | |
| 5896 | static void wq_watchdog_set_thresh(unsigned long thresh) |
| 5897 | { |
| 5898 | wq_watchdog_thresh = 0; |
| 5899 | del_timer_sync(&wq_watchdog_timer); |
| 5900 | |
| 5901 | if (thresh) { |
| 5902 | wq_watchdog_thresh = thresh; |
| 5903 | wq_watchdog_reset_touched(); |
| 5904 | mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ); |
| 5905 | } |
| 5906 | } |
| 5907 | |
| 5908 | static int wq_watchdog_param_set_thresh(const char *val, |
| 5909 | const struct kernel_param *kp) |
| 5910 | { |
| 5911 | unsigned long thresh; |
| 5912 | int ret; |
| 5913 | |
| 5914 | ret = kstrtoul(val, 0, &thresh); |
| 5915 | if (ret) |
| 5916 | return ret; |
| 5917 | |
| 5918 | if (system_wq) |
| 5919 | wq_watchdog_set_thresh(thresh); |
| 5920 | else |
| 5921 | wq_watchdog_thresh = thresh; |
| 5922 | |
| 5923 | return 0; |
| 5924 | } |
| 5925 | |
| 5926 | static const struct kernel_param_ops wq_watchdog_thresh_ops = { |
| 5927 | .set = wq_watchdog_param_set_thresh, |
| 5928 | .get = param_get_ulong, |
| 5929 | }; |
| 5930 | |
| 5931 | module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh, |
| 5932 | 0644); |
| 5933 | |
| 5934 | static void wq_watchdog_init(void) |
| 5935 | { |
| 5936 | timer_setup(&wq_watchdog_timer, wq_watchdog_timer_fn, TIMER_DEFERRABLE); |
| 5937 | wq_watchdog_set_thresh(wq_watchdog_thresh); |
| 5938 | } |
| 5939 | |
| 5940 | #else /* CONFIG_WQ_WATCHDOG */ |
| 5941 | |
| 5942 | static inline void wq_watchdog_init(void) { } |
| 5943 | |
| 5944 | #endif /* CONFIG_WQ_WATCHDOG */ |
| 5945 | |
| 5946 | static void __init wq_numa_init(void) |
| 5947 | { |
| 5948 | cpumask_var_t *tbl; |
| 5949 | int node, cpu; |
| 5950 | |
| 5951 | if (num_possible_nodes() <= 1) |
| 5952 | return; |
| 5953 | |
| 5954 | if (wq_disable_numa) { |
| 5955 | pr_info("workqueue: NUMA affinity support disabled\n"); |
| 5956 | return; |
| 5957 | } |
| 5958 | |
| 5959 | for_each_possible_cpu(cpu) { |
| 5960 | if (WARN_ON(cpu_to_node(cpu) == NUMA_NO_NODE)) { |
| 5961 | pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu); |
| 5962 | return; |
| 5963 | } |
| 5964 | } |
| 5965 | |
| 5966 | wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(); |
| 5967 | BUG_ON(!wq_update_unbound_numa_attrs_buf); |
| 5968 | |
| 5969 | /* |
| 5970 | * We want masks of possible CPUs of each node which isn't readily |
| 5971 | * available. Build one from cpu_to_node() which should have been |
| 5972 | * fully initialized by now. |
| 5973 | */ |
| 5974 | tbl = kcalloc(nr_node_ids, sizeof(tbl[0]), GFP_KERNEL); |
| 5975 | BUG_ON(!tbl); |
| 5976 | |
| 5977 | for_each_node(node) |
| 5978 | BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL, |
| 5979 | node_online(node) ? node : NUMA_NO_NODE)); |
| 5980 | |
| 5981 | for_each_possible_cpu(cpu) { |
| 5982 | node = cpu_to_node(cpu); |
| 5983 | cpumask_set_cpu(cpu, tbl[node]); |
| 5984 | } |
| 5985 | |
| 5986 | wq_numa_possible_cpumask = tbl; |
| 5987 | wq_numa_enabled = true; |
| 5988 | } |
| 5989 | |
| 5990 | /** |
| 5991 | * workqueue_init_early - early init for workqueue subsystem |
| 5992 | * |
| 5993 | * This is the first half of two-staged workqueue subsystem initialization |
| 5994 | * and invoked as soon as the bare basics - memory allocation, cpumasks and |
| 5995 | * idr are up. It sets up all the data structures and system workqueues |
| 5996 | * and allows early boot code to create workqueues and queue/cancel work |
| 5997 | * items. Actual work item execution starts only after kthreads can be |
| 5998 | * created and scheduled right before early initcalls. |
| 5999 | */ |
| 6000 | void __init workqueue_init_early(void) |
| 6001 | { |
| 6002 | int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL }; |
| 6003 | int i, cpu; |
| 6004 | |
| 6005 | BUILD_BUG_ON(__alignof__(struct pool_workqueue) < __alignof__(long long)); |
| 6006 | |
| 6007 | BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL)); |
| 6008 | cpumask_copy(wq_unbound_cpumask, housekeeping_cpumask(HK_TYPE_WQ)); |
| 6009 | cpumask_and(wq_unbound_cpumask, wq_unbound_cpumask, housekeeping_cpumask(HK_TYPE_DOMAIN)); |
| 6010 | |
| 6011 | pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC); |
| 6012 | |
| 6013 | /* initialize CPU pools */ |
| 6014 | for_each_possible_cpu(cpu) { |
| 6015 | struct worker_pool *pool; |
| 6016 | |
| 6017 | i = 0; |
| 6018 | for_each_cpu_worker_pool(pool, cpu) { |
| 6019 | BUG_ON(init_worker_pool(pool)); |
| 6020 | pool->cpu = cpu; |
| 6021 | cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu)); |
| 6022 | pool->attrs->nice = std_nice[i++]; |
| 6023 | pool->node = cpu_to_node(cpu); |
| 6024 | |
| 6025 | /* alloc pool ID */ |
| 6026 | mutex_lock(&wq_pool_mutex); |
| 6027 | BUG_ON(worker_pool_assign_id(pool)); |
| 6028 | mutex_unlock(&wq_pool_mutex); |
| 6029 | } |
| 6030 | } |
| 6031 | |
| 6032 | /* create default unbound and ordered wq attrs */ |
| 6033 | for (i = 0; i < NR_STD_WORKER_POOLS; i++) { |
| 6034 | struct workqueue_attrs *attrs; |
| 6035 | |
| 6036 | BUG_ON(!(attrs = alloc_workqueue_attrs())); |
| 6037 | attrs->nice = std_nice[i]; |
| 6038 | unbound_std_wq_attrs[i] = attrs; |
| 6039 | |
| 6040 | /* |
| 6041 | * An ordered wq should have only one pwq as ordering is |
| 6042 | * guaranteed by max_active which is enforced by pwqs. |
| 6043 | * Turn off NUMA so that dfl_pwq is used for all nodes. |
| 6044 | */ |
| 6045 | BUG_ON(!(attrs = alloc_workqueue_attrs())); |
| 6046 | attrs->nice = std_nice[i]; |
| 6047 | attrs->no_numa = true; |
| 6048 | ordered_wq_attrs[i] = attrs; |
| 6049 | } |
| 6050 | |
| 6051 | system_wq = alloc_workqueue("events", 0, 0); |
| 6052 | system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0); |
| 6053 | system_long_wq = alloc_workqueue("events_long", 0, 0); |
| 6054 | system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND, |
| 6055 | WQ_UNBOUND_MAX_ACTIVE); |
| 6056 | system_freezable_wq = alloc_workqueue("events_freezable", |
| 6057 | WQ_FREEZABLE, 0); |
| 6058 | system_power_efficient_wq = alloc_workqueue("events_power_efficient", |
| 6059 | WQ_POWER_EFFICIENT, 0); |
| 6060 | system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient", |
| 6061 | WQ_FREEZABLE | WQ_POWER_EFFICIENT, |
| 6062 | 0); |
| 6063 | BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq || |
| 6064 | !system_unbound_wq || !system_freezable_wq || |
| 6065 | !system_power_efficient_wq || |
| 6066 | !system_freezable_power_efficient_wq); |
| 6067 | } |
| 6068 | |
| 6069 | /** |
| 6070 | * workqueue_init - bring workqueue subsystem fully online |
| 6071 | * |
| 6072 | * This is the latter half of two-staged workqueue subsystem initialization |
| 6073 | * and invoked as soon as kthreads can be created and scheduled. |
| 6074 | * Workqueues have been created and work items queued on them, but there |
| 6075 | * are no kworkers executing the work items yet. Populate the worker pools |
| 6076 | * with the initial workers and enable future kworker creations. |
| 6077 | */ |
| 6078 | void __init workqueue_init(void) |
| 6079 | { |
| 6080 | struct workqueue_struct *wq; |
| 6081 | struct worker_pool *pool; |
| 6082 | int cpu, bkt; |
| 6083 | |
| 6084 | /* |
| 6085 | * It'd be simpler to initialize NUMA in workqueue_init_early() but |
| 6086 | * CPU to node mapping may not be available that early on some |
| 6087 | * archs such as power and arm64. As per-cpu pools created |
| 6088 | * previously could be missing node hint and unbound pools NUMA |
| 6089 | * affinity, fix them up. |
| 6090 | * |
| 6091 | * Also, while iterating workqueues, create rescuers if requested. |
| 6092 | */ |
| 6093 | wq_numa_init(); |
| 6094 | |
| 6095 | mutex_lock(&wq_pool_mutex); |
| 6096 | |
| 6097 | for_each_possible_cpu(cpu) { |
| 6098 | for_each_cpu_worker_pool(pool, cpu) { |
| 6099 | pool->node = cpu_to_node(cpu); |
| 6100 | } |
| 6101 | } |
| 6102 | |
| 6103 | list_for_each_entry(wq, &workqueues, list) { |
| 6104 | wq_update_unbound_numa(wq, smp_processor_id(), true); |
| 6105 | WARN(init_rescuer(wq), |
| 6106 | "workqueue: failed to create early rescuer for %s", |
| 6107 | wq->name); |
| 6108 | } |
| 6109 | |
| 6110 | mutex_unlock(&wq_pool_mutex); |
| 6111 | |
| 6112 | /* create the initial workers */ |
| 6113 | for_each_online_cpu(cpu) { |
| 6114 | for_each_cpu_worker_pool(pool, cpu) { |
| 6115 | pool->flags &= ~POOL_DISASSOCIATED; |
| 6116 | BUG_ON(!create_worker(pool)); |
| 6117 | } |
| 6118 | } |
| 6119 | |
| 6120 | hash_for_each(unbound_pool_hash, bkt, pool, hash_node) |
| 6121 | BUG_ON(!create_worker(pool)); |
| 6122 | |
| 6123 | wq_online = true; |
| 6124 | wq_watchdog_init(); |
| 6125 | } |
| 6126 | |
| 6127 | /* |
| 6128 | * Despite the naming, this is a no-op function which is here only for avoiding |
| 6129 | * link error. Since compile-time warning may fail to catch, we will need to |
| 6130 | * emit run-time warning from __flush_workqueue(). |
| 6131 | */ |
| 6132 | void __warn_flushing_systemwide_wq(void) { } |
| 6133 | EXPORT_SYMBOL(__warn_flushing_systemwide_wq); |