Commit | Line | Data |
---|---|---|
1da177e4 | 1 | /* |
c54fce6e | 2 | * kernel/workqueue.c - generic async execution with shared worker pool |
1da177e4 | 3 | * |
c54fce6e | 4 | * Copyright (C) 2002 Ingo Molnar |
1da177e4 | 5 | * |
c54fce6e TH |
6 | * Derived from the taskqueue/keventd code by: |
7 | * David Woodhouse <dwmw2@infradead.org> | |
8 | * Andrew Morton | |
9 | * Kai Petzke <wpp@marie.physik.tu-berlin.de> | |
10 | * Theodore Ts'o <tytso@mit.edu> | |
1da177e4 | 11 | * |
c54fce6e | 12 | * Made to use alloc_percpu by Christoph Lameter. |
1da177e4 | 13 | * |
c54fce6e TH |
14 | * Copyright (C) 2010 SUSE Linux Products GmbH |
15 | * Copyright (C) 2010 Tejun Heo <tj@kernel.org> | |
89ada679 | 16 | * |
c54fce6e TH |
17 | * This is the generic async execution mechanism. Work items as are |
18 | * executed in process context. The worker pool is shared and | |
b11895c4 L |
19 | * automatically managed. There are two worker pools for each CPU (one for |
20 | * normal work items and the other for high priority ones) and some extra | |
21 | * pools for workqueues which are not bound to any specific CPU - the | |
22 | * number of these backing pools is dynamic. | |
c54fce6e TH |
23 | * |
24 | * Please read Documentation/workqueue.txt for details. | |
1da177e4 LT |
25 | */ |
26 | ||
9984de1a | 27 | #include <linux/export.h> |
1da177e4 LT |
28 | #include <linux/kernel.h> |
29 | #include <linux/sched.h> | |
30 | #include <linux/init.h> | |
31 | #include <linux/signal.h> | |
32 | #include <linux/completion.h> | |
33 | #include <linux/workqueue.h> | |
34 | #include <linux/slab.h> | |
35 | #include <linux/cpu.h> | |
36 | #include <linux/notifier.h> | |
37 | #include <linux/kthread.h> | |
1fa44eca | 38 | #include <linux/hardirq.h> |
46934023 | 39 | #include <linux/mempolicy.h> |
341a5958 | 40 | #include <linux/freezer.h> |
d5abe669 PZ |
41 | #include <linux/kallsyms.h> |
42 | #include <linux/debug_locks.h> | |
4e6045f1 | 43 | #include <linux/lockdep.h> |
c34056a3 | 44 | #include <linux/idr.h> |
29c91e99 | 45 | #include <linux/jhash.h> |
42f8570f | 46 | #include <linux/hashtable.h> |
76af4d93 | 47 | #include <linux/rculist.h> |
bce90380 | 48 | #include <linux/nodemask.h> |
4c16bd32 | 49 | #include <linux/moduleparam.h> |
3d1cb205 | 50 | #include <linux/uaccess.h> |
e22bee78 | 51 | |
ea138446 | 52 | #include "workqueue_internal.h" |
1da177e4 | 53 | |
c8e55f36 | 54 | enum { |
24647570 TH |
55 | /* |
56 | * worker_pool flags | |
bc2ae0f5 | 57 | * |
24647570 | 58 | * A bound pool is either associated or disassociated with its CPU. |
bc2ae0f5 TH |
59 | * While associated (!DISASSOCIATED), all workers are bound to the |
60 | * CPU and none has %WORKER_UNBOUND set and concurrency management | |
61 | * is in effect. | |
62 | * | |
63 | * While DISASSOCIATED, the cpu may be offline and all workers have | |
64 | * %WORKER_UNBOUND set and concurrency management disabled, and may | |
24647570 | 65 | * be executing on any CPU. The pool behaves as an unbound one. |
bc2ae0f5 | 66 | * |
bc3a1afc TH |
67 | * Note that DISASSOCIATED should be flipped only while holding |
68 | * manager_mutex to avoid changing binding state while | |
24647570 | 69 | * create_worker() is in progress. |
bc2ae0f5 | 70 | */ |
24647570 | 71 | POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */ |
35b6bb63 | 72 | POOL_FREEZING = 1 << 3, /* freeze in progress */ |
db7bccf4 | 73 | |
c8e55f36 | 74 | /* worker flags */ |
c8e55f36 TH |
75 | WORKER_DIE = 1 << 1, /* die die die */ |
76 | WORKER_IDLE = 1 << 2, /* is idle */ | |
e22bee78 | 77 | WORKER_PREP = 1 << 3, /* preparing to run works */ |
fb0e7beb | 78 | WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */ |
f3421797 | 79 | WORKER_UNBOUND = 1 << 7, /* worker is unbound */ |
a9ab775b | 80 | WORKER_REBOUND = 1 << 8, /* worker was rebound */ |
e22bee78 | 81 | |
a9ab775b TH |
82 | WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE | |
83 | WORKER_UNBOUND | WORKER_REBOUND, | |
db7bccf4 | 84 | |
e34cdddb | 85 | NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */ |
4ce62e9e | 86 | |
29c91e99 | 87 | UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */ |
c8e55f36 | 88 | BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */ |
db7bccf4 | 89 | |
e22bee78 TH |
90 | MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */ |
91 | IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */ | |
92 | ||
3233cdbd TH |
93 | MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2, |
94 | /* call for help after 10ms | |
95 | (min two ticks) */ | |
e22bee78 TH |
96 | MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */ |
97 | CREATE_COOLDOWN = HZ, /* time to breath after fail */ | |
e22bee78 TH |
98 | |
99 | /* | |
100 | * Rescue workers are used only on emergencies and shared by | |
101 | * all cpus. Give -20. | |
102 | */ | |
103 | RESCUER_NICE_LEVEL = -20, | |
3270476a | 104 | HIGHPRI_NICE_LEVEL = -20, |
ecf6881f TH |
105 | |
106 | WQ_NAME_LEN = 24, | |
c8e55f36 | 107 | }; |
1da177e4 LT |
108 | |
109 | /* | |
4690c4ab TH |
110 | * Structure fields follow one of the following exclusion rules. |
111 | * | |
e41e704b TH |
112 | * I: Modifiable by initialization/destruction paths and read-only for |
113 | * everyone else. | |
4690c4ab | 114 | * |
e22bee78 TH |
115 | * P: Preemption protected. Disabling preemption is enough and should |
116 | * only be modified and accessed from the local cpu. | |
117 | * | |
d565ed63 | 118 | * L: pool->lock protected. Access with pool->lock held. |
4690c4ab | 119 | * |
d565ed63 TH |
120 | * X: During normal operation, modification requires pool->lock and should |
121 | * be done only from local cpu. Either disabling preemption on local | |
122 | * cpu or grabbing pool->lock is enough for read access. If | |
123 | * POOL_DISASSOCIATED is set, it's identical to L. | |
e22bee78 | 124 | * |
9625ab17 | 125 | * M: pool->manager_mutex protected. |
822d8405 | 126 | * |
68e13a67 | 127 | * PL: wq_pool_mutex protected. |
5bcab335 | 128 | * |
68e13a67 | 129 | * PR: wq_pool_mutex protected for writes. Sched-RCU protected for reads. |
76af4d93 | 130 | * |
3c25a55d LJ |
131 | * WQ: wq->mutex protected. |
132 | * | |
b5927605 | 133 | * WR: wq->mutex protected for writes. Sched-RCU protected for reads. |
2e109a28 TH |
134 | * |
135 | * MD: wq_mayday_lock protected. | |
1da177e4 | 136 | */ |
1da177e4 | 137 | |
2eaebdb3 | 138 | /* struct worker is defined in workqueue_internal.h */ |
c34056a3 | 139 | |
bd7bdd43 | 140 | struct worker_pool { |
d565ed63 | 141 | spinlock_t lock; /* the pool lock */ |
d84ff051 | 142 | int cpu; /* I: the associated cpu */ |
f3f90ad4 | 143 | int node; /* I: the associated node ID */ |
9daf9e67 | 144 | int id; /* I: pool ID */ |
11ebea50 | 145 | unsigned int flags; /* X: flags */ |
bd7bdd43 TH |
146 | |
147 | struct list_head worklist; /* L: list of pending works */ | |
148 | int nr_workers; /* L: total number of workers */ | |
ea1abd61 LJ |
149 | |
150 | /* nr_idle includes the ones off idle_list for rebinding */ | |
bd7bdd43 TH |
151 | int nr_idle; /* L: currently idle ones */ |
152 | ||
153 | struct list_head idle_list; /* X: list of idle workers */ | |
154 | struct timer_list idle_timer; /* L: worker idle timeout */ | |
155 | struct timer_list mayday_timer; /* L: SOS timer for workers */ | |
156 | ||
c5aa87bb | 157 | /* a workers is either on busy_hash or idle_list, or the manager */ |
c9e7cf27 TH |
158 | DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER); |
159 | /* L: hash of busy workers */ | |
160 | ||
bc3a1afc | 161 | /* see manage_workers() for details on the two manager mutexes */ |
34a06bd6 | 162 | struct mutex manager_arb; /* manager arbitration */ |
bc3a1afc | 163 | struct mutex manager_mutex; /* manager exclusion */ |
9625ab17 | 164 | struct idr worker_idr; /* M: worker IDs and iteration */ |
60f5a4bc | 165 | struct completion *detach_completion; /* all workers detached */ |
e19e397a | 166 | |
7a4e344c | 167 | struct workqueue_attrs *attrs; /* I: worker attributes */ |
68e13a67 LJ |
168 | struct hlist_node hash_node; /* PL: unbound_pool_hash node */ |
169 | int refcnt; /* PL: refcnt for unbound pools */ | |
7a4e344c | 170 | |
e19e397a TH |
171 | /* |
172 | * The current concurrency level. As it's likely to be accessed | |
173 | * from other CPUs during try_to_wake_up(), put it in a separate | |
174 | * cacheline. | |
175 | */ | |
176 | atomic_t nr_running ____cacheline_aligned_in_smp; | |
29c91e99 TH |
177 | |
178 | /* | |
179 | * Destruction of pool is sched-RCU protected to allow dereferences | |
180 | * from get_work_pool(). | |
181 | */ | |
182 | struct rcu_head rcu; | |
8b03ae3c TH |
183 | } ____cacheline_aligned_in_smp; |
184 | ||
1da177e4 | 185 | /* |
112202d9 TH |
186 | * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS |
187 | * of work_struct->data are used for flags and the remaining high bits | |
188 | * point to the pwq; thus, pwqs need to be aligned at two's power of the | |
189 | * number of flag bits. | |
1da177e4 | 190 | */ |
112202d9 | 191 | struct pool_workqueue { |
bd7bdd43 | 192 | struct worker_pool *pool; /* I: the associated pool */ |
4690c4ab | 193 | struct workqueue_struct *wq; /* I: the owning workqueue */ |
73f53c4a TH |
194 | int work_color; /* L: current color */ |
195 | int flush_color; /* L: flushing color */ | |
8864b4e5 | 196 | int refcnt; /* L: reference count */ |
73f53c4a TH |
197 | int nr_in_flight[WORK_NR_COLORS]; |
198 | /* L: nr of in_flight works */ | |
1e19ffc6 | 199 | int nr_active; /* L: nr of active works */ |
a0a1a5fd | 200 | int max_active; /* L: max active works */ |
1e19ffc6 | 201 | struct list_head delayed_works; /* L: delayed works */ |
3c25a55d | 202 | struct list_head pwqs_node; /* WR: node on wq->pwqs */ |
2e109a28 | 203 | struct list_head mayday_node; /* MD: node on wq->maydays */ |
8864b4e5 TH |
204 | |
205 | /* | |
206 | * Release of unbound pwq is punted to system_wq. See put_pwq() | |
207 | * and pwq_unbound_release_workfn() for details. pool_workqueue | |
208 | * itself is also sched-RCU protected so that the first pwq can be | |
b09f4fd3 | 209 | * determined without grabbing wq->mutex. |
8864b4e5 TH |
210 | */ |
211 | struct work_struct unbound_release_work; | |
212 | struct rcu_head rcu; | |
e904e6c2 | 213 | } __aligned(1 << WORK_STRUCT_FLAG_BITS); |
1da177e4 | 214 | |
73f53c4a TH |
215 | /* |
216 | * Structure used to wait for workqueue flush. | |
217 | */ | |
218 | struct wq_flusher { | |
3c25a55d LJ |
219 | struct list_head list; /* WQ: list of flushers */ |
220 | int flush_color; /* WQ: flush color waiting for */ | |
73f53c4a TH |
221 | struct completion done; /* flush completion */ |
222 | }; | |
223 | ||
226223ab TH |
224 | struct wq_device; |
225 | ||
1da177e4 | 226 | /* |
c5aa87bb TH |
227 | * The externally visible workqueue. It relays the issued work items to |
228 | * the appropriate worker_pool through its pool_workqueues. | |
1da177e4 LT |
229 | */ |
230 | struct workqueue_struct { | |
3c25a55d | 231 | struct list_head pwqs; /* WR: all pwqs of this wq */ |
68e13a67 | 232 | struct list_head list; /* PL: list of all workqueues */ |
73f53c4a | 233 | |
3c25a55d LJ |
234 | struct mutex mutex; /* protects this wq */ |
235 | int work_color; /* WQ: current work color */ | |
236 | int flush_color; /* WQ: current flush color */ | |
112202d9 | 237 | atomic_t nr_pwqs_to_flush; /* flush in progress */ |
3c25a55d LJ |
238 | struct wq_flusher *first_flusher; /* WQ: first flusher */ |
239 | struct list_head flusher_queue; /* WQ: flush waiters */ | |
240 | struct list_head flusher_overflow; /* WQ: flush overflow list */ | |
73f53c4a | 241 | |
2e109a28 | 242 | struct list_head maydays; /* MD: pwqs requesting rescue */ |
e22bee78 TH |
243 | struct worker *rescuer; /* I: rescue worker */ |
244 | ||
87fc741e | 245 | int nr_drainers; /* WQ: drain in progress */ |
a357fc03 | 246 | int saved_max_active; /* WQ: saved pwq max_active */ |
226223ab | 247 | |
6029a918 | 248 | struct workqueue_attrs *unbound_attrs; /* WQ: only for unbound wqs */ |
4c16bd32 | 249 | struct pool_workqueue *dfl_pwq; /* WQ: only for unbound wqs */ |
6029a918 | 250 | |
226223ab TH |
251 | #ifdef CONFIG_SYSFS |
252 | struct wq_device *wq_dev; /* I: for sysfs interface */ | |
253 | #endif | |
4e6045f1 | 254 | #ifdef CONFIG_LOCKDEP |
4690c4ab | 255 | struct lockdep_map lockdep_map; |
4e6045f1 | 256 | #endif |
ecf6881f | 257 | char name[WQ_NAME_LEN]; /* I: workqueue name */ |
2728fd2f TH |
258 | |
259 | /* hot fields used during command issue, aligned to cacheline */ | |
260 | unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */ | |
261 | struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwqs */ | |
df2d5ae4 | 262 | struct pool_workqueue __rcu *numa_pwq_tbl[]; /* FR: unbound pwqs indexed by node */ |
1da177e4 LT |
263 | }; |
264 | ||
e904e6c2 TH |
265 | static struct kmem_cache *pwq_cache; |
266 | ||
bce90380 TH |
267 | static int wq_numa_tbl_len; /* highest possible NUMA node id + 1 */ |
268 | static cpumask_var_t *wq_numa_possible_cpumask; | |
269 | /* possible CPUs of each node */ | |
270 | ||
d55262c4 TH |
271 | static bool wq_disable_numa; |
272 | module_param_named(disable_numa, wq_disable_numa, bool, 0444); | |
273 | ||
cee22a15 VK |
274 | /* see the comment above the definition of WQ_POWER_EFFICIENT */ |
275 | #ifdef CONFIG_WQ_POWER_EFFICIENT_DEFAULT | |
276 | static bool wq_power_efficient = true; | |
277 | #else | |
278 | static bool wq_power_efficient; | |
279 | #endif | |
280 | ||
281 | module_param_named(power_efficient, wq_power_efficient, bool, 0444); | |
282 | ||
bce90380 TH |
283 | static bool wq_numa_enabled; /* unbound NUMA affinity enabled */ |
284 | ||
4c16bd32 TH |
285 | /* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */ |
286 | static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf; | |
287 | ||
68e13a67 | 288 | static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */ |
2e109a28 | 289 | static DEFINE_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */ |
5bcab335 | 290 | |
68e13a67 LJ |
291 | static LIST_HEAD(workqueues); /* PL: list of all workqueues */ |
292 | static bool workqueue_freezing; /* PL: have wqs started freezing? */ | |
7d19c5ce TH |
293 | |
294 | /* the per-cpu worker pools */ | |
295 | static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], | |
296 | cpu_worker_pools); | |
297 | ||
68e13a67 | 298 | static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */ |
7d19c5ce | 299 | |
68e13a67 | 300 | /* PL: hash of all unbound pools keyed by pool->attrs */ |
29c91e99 TH |
301 | static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER); |
302 | ||
c5aa87bb | 303 | /* I: attributes used when instantiating standard unbound pools on demand */ |
29c91e99 TH |
304 | static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS]; |
305 | ||
8a2b7538 TH |
306 | /* I: attributes used when instantiating ordered pools on demand */ |
307 | static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS]; | |
308 | ||
d320c038 | 309 | struct workqueue_struct *system_wq __read_mostly; |
ad7b1f84 | 310 | EXPORT_SYMBOL(system_wq); |
044c782c | 311 | struct workqueue_struct *system_highpri_wq __read_mostly; |
1aabe902 | 312 | EXPORT_SYMBOL_GPL(system_highpri_wq); |
044c782c | 313 | struct workqueue_struct *system_long_wq __read_mostly; |
d320c038 | 314 | EXPORT_SYMBOL_GPL(system_long_wq); |
044c782c | 315 | struct workqueue_struct *system_unbound_wq __read_mostly; |
f3421797 | 316 | EXPORT_SYMBOL_GPL(system_unbound_wq); |
044c782c | 317 | struct workqueue_struct *system_freezable_wq __read_mostly; |
24d51add | 318 | EXPORT_SYMBOL_GPL(system_freezable_wq); |
0668106c VK |
319 | struct workqueue_struct *system_power_efficient_wq __read_mostly; |
320 | EXPORT_SYMBOL_GPL(system_power_efficient_wq); | |
321 | struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly; | |
322 | EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq); | |
d320c038 | 323 | |
7d19c5ce TH |
324 | static int worker_thread(void *__worker); |
325 | static void copy_workqueue_attrs(struct workqueue_attrs *to, | |
326 | const struct workqueue_attrs *from); | |
327 | ||
97bd2347 TH |
328 | #define CREATE_TRACE_POINTS |
329 | #include <trace/events/workqueue.h> | |
330 | ||
68e13a67 | 331 | #define assert_rcu_or_pool_mutex() \ |
5bcab335 | 332 | rcu_lockdep_assert(rcu_read_lock_sched_held() || \ |
68e13a67 LJ |
333 | lockdep_is_held(&wq_pool_mutex), \ |
334 | "sched RCU or wq_pool_mutex should be held") | |
5bcab335 | 335 | |
b09f4fd3 | 336 | #define assert_rcu_or_wq_mutex(wq) \ |
76af4d93 | 337 | rcu_lockdep_assert(rcu_read_lock_sched_held() || \ |
b5927605 | 338 | lockdep_is_held(&wq->mutex), \ |
b09f4fd3 | 339 | "sched RCU or wq->mutex should be held") |
76af4d93 | 340 | |
f02ae73a TH |
341 | #define for_each_cpu_worker_pool(pool, cpu) \ |
342 | for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \ | |
343 | (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \ | |
7a62c2c8 | 344 | (pool)++) |
4ce62e9e | 345 | |
17116969 TH |
346 | /** |
347 | * for_each_pool - iterate through all worker_pools in the system | |
348 | * @pool: iteration cursor | |
611c92a0 | 349 | * @pi: integer used for iteration |
fa1b54e6 | 350 | * |
68e13a67 LJ |
351 | * This must be called either with wq_pool_mutex held or sched RCU read |
352 | * locked. If the pool needs to be used beyond the locking in effect, the | |
353 | * caller is responsible for guaranteeing that the pool stays online. | |
fa1b54e6 TH |
354 | * |
355 | * The if/else clause exists only for the lockdep assertion and can be | |
356 | * ignored. | |
17116969 | 357 | */ |
611c92a0 TH |
358 | #define for_each_pool(pool, pi) \ |
359 | idr_for_each_entry(&worker_pool_idr, pool, pi) \ | |
68e13a67 | 360 | if (({ assert_rcu_or_pool_mutex(); false; })) { } \ |
fa1b54e6 | 361 | else |
17116969 | 362 | |
822d8405 TH |
363 | /** |
364 | * for_each_pool_worker - iterate through all workers of a worker_pool | |
365 | * @worker: iteration cursor | |
366 | * @wi: integer used for iteration | |
367 | * @pool: worker_pool to iterate workers of | |
368 | * | |
9625ab17 | 369 | * This must be called with @pool->manager_mutex. |
822d8405 TH |
370 | * |
371 | * The if/else clause exists only for the lockdep assertion and can be | |
372 | * ignored. | |
373 | */ | |
374 | #define for_each_pool_worker(worker, wi, pool) \ | |
375 | idr_for_each_entry(&(pool)->worker_idr, (worker), (wi)) \ | |
9625ab17 | 376 | if (({ lockdep_assert_held(&pool->manager_mutex); false; })) { } \ |
822d8405 TH |
377 | else |
378 | ||
49e3cf44 TH |
379 | /** |
380 | * for_each_pwq - iterate through all pool_workqueues of the specified workqueue | |
381 | * @pwq: iteration cursor | |
382 | * @wq: the target workqueue | |
76af4d93 | 383 | * |
b09f4fd3 | 384 | * This must be called either with wq->mutex held or sched RCU read locked. |
794b18bc TH |
385 | * If the pwq needs to be used beyond the locking in effect, the caller is |
386 | * responsible for guaranteeing that the pwq stays online. | |
76af4d93 TH |
387 | * |
388 | * The if/else clause exists only for the lockdep assertion and can be | |
389 | * ignored. | |
49e3cf44 TH |
390 | */ |
391 | #define for_each_pwq(pwq, wq) \ | |
76af4d93 | 392 | list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \ |
b09f4fd3 | 393 | if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \ |
76af4d93 | 394 | else |
f3421797 | 395 | |
dc186ad7 TG |
396 | #ifdef CONFIG_DEBUG_OBJECTS_WORK |
397 | ||
398 | static struct debug_obj_descr work_debug_descr; | |
399 | ||
99777288 SG |
400 | static void *work_debug_hint(void *addr) |
401 | { | |
402 | return ((struct work_struct *) addr)->func; | |
403 | } | |
404 | ||
dc186ad7 TG |
405 | /* |
406 | * fixup_init is called when: | |
407 | * - an active object is initialized | |
408 | */ | |
409 | static int work_fixup_init(void *addr, enum debug_obj_state state) | |
410 | { | |
411 | struct work_struct *work = addr; | |
412 | ||
413 | switch (state) { | |
414 | case ODEBUG_STATE_ACTIVE: | |
415 | cancel_work_sync(work); | |
416 | debug_object_init(work, &work_debug_descr); | |
417 | return 1; | |
418 | default: | |
419 | return 0; | |
420 | } | |
421 | } | |
422 | ||
423 | /* | |
424 | * fixup_activate is called when: | |
425 | * - an active object is activated | |
426 | * - an unknown object is activated (might be a statically initialized object) | |
427 | */ | |
428 | static int work_fixup_activate(void *addr, enum debug_obj_state state) | |
429 | { | |
430 | struct work_struct *work = addr; | |
431 | ||
432 | switch (state) { | |
433 | ||
434 | case ODEBUG_STATE_NOTAVAILABLE: | |
435 | /* | |
436 | * This is not really a fixup. The work struct was | |
437 | * statically initialized. We just make sure that it | |
438 | * is tracked in the object tracker. | |
439 | */ | |
22df02bb | 440 | if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) { |
dc186ad7 TG |
441 | debug_object_init(work, &work_debug_descr); |
442 | debug_object_activate(work, &work_debug_descr); | |
443 | return 0; | |
444 | } | |
445 | WARN_ON_ONCE(1); | |
446 | return 0; | |
447 | ||
448 | case ODEBUG_STATE_ACTIVE: | |
449 | WARN_ON(1); | |
450 | ||
451 | default: | |
452 | return 0; | |
453 | } | |
454 | } | |
455 | ||
456 | /* | |
457 | * fixup_free is called when: | |
458 | * - an active object is freed | |
459 | */ | |
460 | static int work_fixup_free(void *addr, enum debug_obj_state state) | |
461 | { | |
462 | struct work_struct *work = addr; | |
463 | ||
464 | switch (state) { | |
465 | case ODEBUG_STATE_ACTIVE: | |
466 | cancel_work_sync(work); | |
467 | debug_object_free(work, &work_debug_descr); | |
468 | return 1; | |
469 | default: | |
470 | return 0; | |
471 | } | |
472 | } | |
473 | ||
474 | static struct debug_obj_descr work_debug_descr = { | |
475 | .name = "work_struct", | |
99777288 | 476 | .debug_hint = work_debug_hint, |
dc186ad7 TG |
477 | .fixup_init = work_fixup_init, |
478 | .fixup_activate = work_fixup_activate, | |
479 | .fixup_free = work_fixup_free, | |
480 | }; | |
481 | ||
482 | static inline void debug_work_activate(struct work_struct *work) | |
483 | { | |
484 | debug_object_activate(work, &work_debug_descr); | |
485 | } | |
486 | ||
487 | static inline void debug_work_deactivate(struct work_struct *work) | |
488 | { | |
489 | debug_object_deactivate(work, &work_debug_descr); | |
490 | } | |
491 | ||
492 | void __init_work(struct work_struct *work, int onstack) | |
493 | { | |
494 | if (onstack) | |
495 | debug_object_init_on_stack(work, &work_debug_descr); | |
496 | else | |
497 | debug_object_init(work, &work_debug_descr); | |
498 | } | |
499 | EXPORT_SYMBOL_GPL(__init_work); | |
500 | ||
501 | void destroy_work_on_stack(struct work_struct *work) | |
502 | { | |
503 | debug_object_free(work, &work_debug_descr); | |
504 | } | |
505 | EXPORT_SYMBOL_GPL(destroy_work_on_stack); | |
506 | ||
ea2e64f2 TG |
507 | void destroy_delayed_work_on_stack(struct delayed_work *work) |
508 | { | |
509 | destroy_timer_on_stack(&work->timer); | |
510 | debug_object_free(&work->work, &work_debug_descr); | |
511 | } | |
512 | EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack); | |
513 | ||
dc186ad7 TG |
514 | #else |
515 | static inline void debug_work_activate(struct work_struct *work) { } | |
516 | static inline void debug_work_deactivate(struct work_struct *work) { } | |
517 | #endif | |
518 | ||
4e8b22bd LB |
519 | /** |
520 | * worker_pool_assign_id - allocate ID and assing it to @pool | |
521 | * @pool: the pool pointer of interest | |
522 | * | |
523 | * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned | |
524 | * successfully, -errno on failure. | |
525 | */ | |
9daf9e67 TH |
526 | static int worker_pool_assign_id(struct worker_pool *pool) |
527 | { | |
528 | int ret; | |
529 | ||
68e13a67 | 530 | lockdep_assert_held(&wq_pool_mutex); |
5bcab335 | 531 | |
4e8b22bd LB |
532 | ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE, |
533 | GFP_KERNEL); | |
229641a6 | 534 | if (ret >= 0) { |
e68035fb | 535 | pool->id = ret; |
229641a6 TH |
536 | return 0; |
537 | } | |
fa1b54e6 | 538 | return ret; |
7c3eed5c TH |
539 | } |
540 | ||
df2d5ae4 TH |
541 | /** |
542 | * unbound_pwq_by_node - return the unbound pool_workqueue for the given node | |
543 | * @wq: the target workqueue | |
544 | * @node: the node ID | |
545 | * | |
546 | * This must be called either with pwq_lock held or sched RCU read locked. | |
547 | * If the pwq needs to be used beyond the locking in effect, the caller is | |
548 | * responsible for guaranteeing that the pwq stays online. | |
d185af30 YB |
549 | * |
550 | * Return: The unbound pool_workqueue for @node. | |
df2d5ae4 TH |
551 | */ |
552 | static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq, | |
553 | int node) | |
554 | { | |
555 | assert_rcu_or_wq_mutex(wq); | |
556 | return rcu_dereference_raw(wq->numa_pwq_tbl[node]); | |
557 | } | |
558 | ||
73f53c4a TH |
559 | static unsigned int work_color_to_flags(int color) |
560 | { | |
561 | return color << WORK_STRUCT_COLOR_SHIFT; | |
562 | } | |
563 | ||
564 | static int get_work_color(struct work_struct *work) | |
565 | { | |
566 | return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) & | |
567 | ((1 << WORK_STRUCT_COLOR_BITS) - 1); | |
568 | } | |
569 | ||
570 | static int work_next_color(int color) | |
571 | { | |
572 | return (color + 1) % WORK_NR_COLORS; | |
573 | } | |
1da177e4 | 574 | |
14441960 | 575 | /* |
112202d9 TH |
576 | * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data |
577 | * contain the pointer to the queued pwq. Once execution starts, the flag | |
7c3eed5c | 578 | * is cleared and the high bits contain OFFQ flags and pool ID. |
7a22ad75 | 579 | * |
112202d9 TH |
580 | * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling() |
581 | * and clear_work_data() can be used to set the pwq, pool or clear | |
bbb68dfa TH |
582 | * work->data. These functions should only be called while the work is |
583 | * owned - ie. while the PENDING bit is set. | |
7a22ad75 | 584 | * |
112202d9 | 585 | * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq |
7c3eed5c | 586 | * corresponding to a work. Pool is available once the work has been |
112202d9 | 587 | * queued anywhere after initialization until it is sync canceled. pwq is |
7c3eed5c | 588 | * available only while the work item is queued. |
7a22ad75 | 589 | * |
bbb68dfa TH |
590 | * %WORK_OFFQ_CANCELING is used to mark a work item which is being |
591 | * canceled. While being canceled, a work item may have its PENDING set | |
592 | * but stay off timer and worklist for arbitrarily long and nobody should | |
593 | * try to steal the PENDING bit. | |
14441960 | 594 | */ |
7a22ad75 TH |
595 | static inline void set_work_data(struct work_struct *work, unsigned long data, |
596 | unsigned long flags) | |
365970a1 | 597 | { |
6183c009 | 598 | WARN_ON_ONCE(!work_pending(work)); |
7a22ad75 TH |
599 | atomic_long_set(&work->data, data | flags | work_static(work)); |
600 | } | |
365970a1 | 601 | |
112202d9 | 602 | static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq, |
7a22ad75 TH |
603 | unsigned long extra_flags) |
604 | { | |
112202d9 TH |
605 | set_work_data(work, (unsigned long)pwq, |
606 | WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags); | |
365970a1 DH |
607 | } |
608 | ||
4468a00f LJ |
609 | static void set_work_pool_and_keep_pending(struct work_struct *work, |
610 | int pool_id) | |
611 | { | |
612 | set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, | |
613 | WORK_STRUCT_PENDING); | |
614 | } | |
615 | ||
7c3eed5c TH |
616 | static void set_work_pool_and_clear_pending(struct work_struct *work, |
617 | int pool_id) | |
7a22ad75 | 618 | { |
23657bb1 TH |
619 | /* |
620 | * The following wmb is paired with the implied mb in | |
621 | * test_and_set_bit(PENDING) and ensures all updates to @work made | |
622 | * here are visible to and precede any updates by the next PENDING | |
623 | * owner. | |
624 | */ | |
625 | smp_wmb(); | |
7c3eed5c | 626 | set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0); |
7a22ad75 | 627 | } |
f756d5e2 | 628 | |
7a22ad75 | 629 | static void clear_work_data(struct work_struct *work) |
1da177e4 | 630 | { |
7c3eed5c TH |
631 | smp_wmb(); /* see set_work_pool_and_clear_pending() */ |
632 | set_work_data(work, WORK_STRUCT_NO_POOL, 0); | |
1da177e4 LT |
633 | } |
634 | ||
112202d9 | 635 | static struct pool_workqueue *get_work_pwq(struct work_struct *work) |
b1f4ec17 | 636 | { |
e120153d | 637 | unsigned long data = atomic_long_read(&work->data); |
7a22ad75 | 638 | |
112202d9 | 639 | if (data & WORK_STRUCT_PWQ) |
e120153d TH |
640 | return (void *)(data & WORK_STRUCT_WQ_DATA_MASK); |
641 | else | |
642 | return NULL; | |
4d707b9f ON |
643 | } |
644 | ||
7c3eed5c TH |
645 | /** |
646 | * get_work_pool - return the worker_pool a given work was associated with | |
647 | * @work: the work item of interest | |
648 | * | |
68e13a67 LJ |
649 | * Pools are created and destroyed under wq_pool_mutex, and allows read |
650 | * access under sched-RCU read lock. As such, this function should be | |
651 | * called under wq_pool_mutex or with preemption disabled. | |
fa1b54e6 TH |
652 | * |
653 | * All fields of the returned pool are accessible as long as the above | |
654 | * mentioned locking is in effect. If the returned pool needs to be used | |
655 | * beyond the critical section, the caller is responsible for ensuring the | |
656 | * returned pool is and stays online. | |
d185af30 YB |
657 | * |
658 | * Return: The worker_pool @work was last associated with. %NULL if none. | |
7c3eed5c TH |
659 | */ |
660 | static struct worker_pool *get_work_pool(struct work_struct *work) | |
365970a1 | 661 | { |
e120153d | 662 | unsigned long data = atomic_long_read(&work->data); |
7c3eed5c | 663 | int pool_id; |
7a22ad75 | 664 | |
68e13a67 | 665 | assert_rcu_or_pool_mutex(); |
fa1b54e6 | 666 | |
112202d9 TH |
667 | if (data & WORK_STRUCT_PWQ) |
668 | return ((struct pool_workqueue *) | |
7c3eed5c | 669 | (data & WORK_STRUCT_WQ_DATA_MASK))->pool; |
7a22ad75 | 670 | |
7c3eed5c TH |
671 | pool_id = data >> WORK_OFFQ_POOL_SHIFT; |
672 | if (pool_id == WORK_OFFQ_POOL_NONE) | |
7a22ad75 TH |
673 | return NULL; |
674 | ||
fa1b54e6 | 675 | return idr_find(&worker_pool_idr, pool_id); |
7c3eed5c TH |
676 | } |
677 | ||
678 | /** | |
679 | * get_work_pool_id - return the worker pool ID a given work is associated with | |
680 | * @work: the work item of interest | |
681 | * | |
d185af30 | 682 | * Return: The worker_pool ID @work was last associated with. |
7c3eed5c TH |
683 | * %WORK_OFFQ_POOL_NONE if none. |
684 | */ | |
685 | static int get_work_pool_id(struct work_struct *work) | |
686 | { | |
54d5b7d0 LJ |
687 | unsigned long data = atomic_long_read(&work->data); |
688 | ||
112202d9 TH |
689 | if (data & WORK_STRUCT_PWQ) |
690 | return ((struct pool_workqueue *) | |
54d5b7d0 | 691 | (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id; |
7c3eed5c | 692 | |
54d5b7d0 | 693 | return data >> WORK_OFFQ_POOL_SHIFT; |
7c3eed5c TH |
694 | } |
695 | ||
bbb68dfa TH |
696 | static void mark_work_canceling(struct work_struct *work) |
697 | { | |
7c3eed5c | 698 | unsigned long pool_id = get_work_pool_id(work); |
bbb68dfa | 699 | |
7c3eed5c TH |
700 | pool_id <<= WORK_OFFQ_POOL_SHIFT; |
701 | set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING); | |
bbb68dfa TH |
702 | } |
703 | ||
704 | static bool work_is_canceling(struct work_struct *work) | |
705 | { | |
706 | unsigned long data = atomic_long_read(&work->data); | |
707 | ||
112202d9 | 708 | return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING); |
bbb68dfa TH |
709 | } |
710 | ||
e22bee78 | 711 | /* |
3270476a TH |
712 | * Policy functions. These define the policies on how the global worker |
713 | * pools are managed. Unless noted otherwise, these functions assume that | |
d565ed63 | 714 | * they're being called with pool->lock held. |
e22bee78 TH |
715 | */ |
716 | ||
63d95a91 | 717 | static bool __need_more_worker(struct worker_pool *pool) |
a848e3b6 | 718 | { |
e19e397a | 719 | return !atomic_read(&pool->nr_running); |
a848e3b6 ON |
720 | } |
721 | ||
4594bf15 | 722 | /* |
e22bee78 TH |
723 | * Need to wake up a worker? Called from anything but currently |
724 | * running workers. | |
974271c4 TH |
725 | * |
726 | * Note that, because unbound workers never contribute to nr_running, this | |
706026c2 | 727 | * function will always return %true for unbound pools as long as the |
974271c4 | 728 | * worklist isn't empty. |
4594bf15 | 729 | */ |
63d95a91 | 730 | static bool need_more_worker(struct worker_pool *pool) |
365970a1 | 731 | { |
63d95a91 | 732 | return !list_empty(&pool->worklist) && __need_more_worker(pool); |
e22bee78 | 733 | } |
4594bf15 | 734 | |
e22bee78 | 735 | /* Can I start working? Called from busy but !running workers. */ |
63d95a91 | 736 | static bool may_start_working(struct worker_pool *pool) |
e22bee78 | 737 | { |
63d95a91 | 738 | return pool->nr_idle; |
e22bee78 TH |
739 | } |
740 | ||
741 | /* Do I need to keep working? Called from currently running workers. */ | |
63d95a91 | 742 | static bool keep_working(struct worker_pool *pool) |
e22bee78 | 743 | { |
e19e397a TH |
744 | return !list_empty(&pool->worklist) && |
745 | atomic_read(&pool->nr_running) <= 1; | |
e22bee78 TH |
746 | } |
747 | ||
748 | /* Do we need a new worker? Called from manager. */ | |
63d95a91 | 749 | static bool need_to_create_worker(struct worker_pool *pool) |
e22bee78 | 750 | { |
63d95a91 | 751 | return need_more_worker(pool) && !may_start_working(pool); |
e22bee78 | 752 | } |
365970a1 | 753 | |
e22bee78 | 754 | /* Do we have too many workers and should some go away? */ |
63d95a91 | 755 | static bool too_many_workers(struct worker_pool *pool) |
e22bee78 | 756 | { |
34a06bd6 | 757 | bool managing = mutex_is_locked(&pool->manager_arb); |
63d95a91 TH |
758 | int nr_idle = pool->nr_idle + managing; /* manager is considered idle */ |
759 | int nr_busy = pool->nr_workers - nr_idle; | |
e22bee78 | 760 | |
ea1abd61 LJ |
761 | /* |
762 | * nr_idle and idle_list may disagree if idle rebinding is in | |
763 | * progress. Never return %true if idle_list is empty. | |
764 | */ | |
765 | if (list_empty(&pool->idle_list)) | |
766 | return false; | |
767 | ||
e22bee78 | 768 | return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy; |
365970a1 DH |
769 | } |
770 | ||
4d707b9f | 771 | /* |
e22bee78 TH |
772 | * Wake up functions. |
773 | */ | |
774 | ||
7e11629d | 775 | /* Return the first worker. Safe with preemption disabled */ |
63d95a91 | 776 | static struct worker *first_worker(struct worker_pool *pool) |
7e11629d | 777 | { |
63d95a91 | 778 | if (unlikely(list_empty(&pool->idle_list))) |
7e11629d TH |
779 | return NULL; |
780 | ||
63d95a91 | 781 | return list_first_entry(&pool->idle_list, struct worker, entry); |
7e11629d TH |
782 | } |
783 | ||
784 | /** | |
785 | * wake_up_worker - wake up an idle worker | |
63d95a91 | 786 | * @pool: worker pool to wake worker from |
7e11629d | 787 | * |
63d95a91 | 788 | * Wake up the first idle worker of @pool. |
7e11629d TH |
789 | * |
790 | * CONTEXT: | |
d565ed63 | 791 | * spin_lock_irq(pool->lock). |
7e11629d | 792 | */ |
63d95a91 | 793 | static void wake_up_worker(struct worker_pool *pool) |
7e11629d | 794 | { |
63d95a91 | 795 | struct worker *worker = first_worker(pool); |
7e11629d TH |
796 | |
797 | if (likely(worker)) | |
798 | wake_up_process(worker->task); | |
799 | } | |
800 | ||
d302f017 | 801 | /** |
e22bee78 TH |
802 | * wq_worker_waking_up - a worker is waking up |
803 | * @task: task waking up | |
804 | * @cpu: CPU @task is waking up to | |
805 | * | |
806 | * This function is called during try_to_wake_up() when a worker is | |
807 | * being awoken. | |
808 | * | |
809 | * CONTEXT: | |
810 | * spin_lock_irq(rq->lock) | |
811 | */ | |
d84ff051 | 812 | void wq_worker_waking_up(struct task_struct *task, int cpu) |
e22bee78 TH |
813 | { |
814 | struct worker *worker = kthread_data(task); | |
815 | ||
36576000 | 816 | if (!(worker->flags & WORKER_NOT_RUNNING)) { |
ec22ca5e | 817 | WARN_ON_ONCE(worker->pool->cpu != cpu); |
e19e397a | 818 | atomic_inc(&worker->pool->nr_running); |
36576000 | 819 | } |
e22bee78 TH |
820 | } |
821 | ||
822 | /** | |
823 | * wq_worker_sleeping - a worker is going to sleep | |
824 | * @task: task going to sleep | |
825 | * @cpu: CPU in question, must be the current CPU number | |
826 | * | |
827 | * This function is called during schedule() when a busy worker is | |
828 | * going to sleep. Worker on the same cpu can be woken up by | |
829 | * returning pointer to its task. | |
830 | * | |
831 | * CONTEXT: | |
832 | * spin_lock_irq(rq->lock) | |
833 | * | |
d185af30 | 834 | * Return: |
e22bee78 TH |
835 | * Worker task on @cpu to wake up, %NULL if none. |
836 | */ | |
d84ff051 | 837 | struct task_struct *wq_worker_sleeping(struct task_struct *task, int cpu) |
e22bee78 TH |
838 | { |
839 | struct worker *worker = kthread_data(task), *to_wakeup = NULL; | |
111c225a | 840 | struct worker_pool *pool; |
e22bee78 | 841 | |
111c225a TH |
842 | /* |
843 | * Rescuers, which may not have all the fields set up like normal | |
844 | * workers, also reach here, let's not access anything before | |
845 | * checking NOT_RUNNING. | |
846 | */ | |
2d64672e | 847 | if (worker->flags & WORKER_NOT_RUNNING) |
e22bee78 TH |
848 | return NULL; |
849 | ||
111c225a | 850 | pool = worker->pool; |
111c225a | 851 | |
e22bee78 | 852 | /* this can only happen on the local cpu */ |
6183c009 TH |
853 | if (WARN_ON_ONCE(cpu != raw_smp_processor_id())) |
854 | return NULL; | |
e22bee78 TH |
855 | |
856 | /* | |
857 | * The counterpart of the following dec_and_test, implied mb, | |
858 | * worklist not empty test sequence is in insert_work(). | |
859 | * Please read comment there. | |
860 | * | |
628c78e7 TH |
861 | * NOT_RUNNING is clear. This means that we're bound to and |
862 | * running on the local cpu w/ rq lock held and preemption | |
863 | * disabled, which in turn means that none else could be | |
d565ed63 | 864 | * manipulating idle_list, so dereferencing idle_list without pool |
628c78e7 | 865 | * lock is safe. |
e22bee78 | 866 | */ |
e19e397a TH |
867 | if (atomic_dec_and_test(&pool->nr_running) && |
868 | !list_empty(&pool->worklist)) | |
63d95a91 | 869 | to_wakeup = first_worker(pool); |
e22bee78 TH |
870 | return to_wakeup ? to_wakeup->task : NULL; |
871 | } | |
872 | ||
873 | /** | |
874 | * worker_set_flags - set worker flags and adjust nr_running accordingly | |
cb444766 | 875 | * @worker: self |
d302f017 TH |
876 | * @flags: flags to set |
877 | * @wakeup: wakeup an idle worker if necessary | |
878 | * | |
e22bee78 TH |
879 | * Set @flags in @worker->flags and adjust nr_running accordingly. If |
880 | * nr_running becomes zero and @wakeup is %true, an idle worker is | |
881 | * woken up. | |
d302f017 | 882 | * |
cb444766 | 883 | * CONTEXT: |
d565ed63 | 884 | * spin_lock_irq(pool->lock) |
d302f017 TH |
885 | */ |
886 | static inline void worker_set_flags(struct worker *worker, unsigned int flags, | |
887 | bool wakeup) | |
888 | { | |
bd7bdd43 | 889 | struct worker_pool *pool = worker->pool; |
e22bee78 | 890 | |
cb444766 TH |
891 | WARN_ON_ONCE(worker->task != current); |
892 | ||
e22bee78 TH |
893 | /* |
894 | * If transitioning into NOT_RUNNING, adjust nr_running and | |
895 | * wake up an idle worker as necessary if requested by | |
896 | * @wakeup. | |
897 | */ | |
898 | if ((flags & WORKER_NOT_RUNNING) && | |
899 | !(worker->flags & WORKER_NOT_RUNNING)) { | |
e22bee78 | 900 | if (wakeup) { |
e19e397a | 901 | if (atomic_dec_and_test(&pool->nr_running) && |
bd7bdd43 | 902 | !list_empty(&pool->worklist)) |
63d95a91 | 903 | wake_up_worker(pool); |
e22bee78 | 904 | } else |
e19e397a | 905 | atomic_dec(&pool->nr_running); |
e22bee78 TH |
906 | } |
907 | ||
d302f017 TH |
908 | worker->flags |= flags; |
909 | } | |
910 | ||
911 | /** | |
e22bee78 | 912 | * worker_clr_flags - clear worker flags and adjust nr_running accordingly |
cb444766 | 913 | * @worker: self |
d302f017 TH |
914 | * @flags: flags to clear |
915 | * | |
e22bee78 | 916 | * Clear @flags in @worker->flags and adjust nr_running accordingly. |
d302f017 | 917 | * |
cb444766 | 918 | * CONTEXT: |
d565ed63 | 919 | * spin_lock_irq(pool->lock) |
d302f017 TH |
920 | */ |
921 | static inline void worker_clr_flags(struct worker *worker, unsigned int flags) | |
922 | { | |
63d95a91 | 923 | struct worker_pool *pool = worker->pool; |
e22bee78 TH |
924 | unsigned int oflags = worker->flags; |
925 | ||
cb444766 TH |
926 | WARN_ON_ONCE(worker->task != current); |
927 | ||
d302f017 | 928 | worker->flags &= ~flags; |
e22bee78 | 929 | |
42c025f3 TH |
930 | /* |
931 | * If transitioning out of NOT_RUNNING, increment nr_running. Note | |
932 | * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask | |
933 | * of multiple flags, not a single flag. | |
934 | */ | |
e22bee78 TH |
935 | if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING)) |
936 | if (!(worker->flags & WORKER_NOT_RUNNING)) | |
e19e397a | 937 | atomic_inc(&pool->nr_running); |
d302f017 TH |
938 | } |
939 | ||
8cca0eea TH |
940 | /** |
941 | * find_worker_executing_work - find worker which is executing a work | |
c9e7cf27 | 942 | * @pool: pool of interest |
8cca0eea TH |
943 | * @work: work to find worker for |
944 | * | |
c9e7cf27 TH |
945 | * Find a worker which is executing @work on @pool by searching |
946 | * @pool->busy_hash which is keyed by the address of @work. For a worker | |
a2c1c57b TH |
947 | * to match, its current execution should match the address of @work and |
948 | * its work function. This is to avoid unwanted dependency between | |
949 | * unrelated work executions through a work item being recycled while still | |
950 | * being executed. | |
951 | * | |
952 | * This is a bit tricky. A work item may be freed once its execution | |
953 | * starts and nothing prevents the freed area from being recycled for | |
954 | * another work item. If the same work item address ends up being reused | |
955 | * before the original execution finishes, workqueue will identify the | |
956 | * recycled work item as currently executing and make it wait until the | |
957 | * current execution finishes, introducing an unwanted dependency. | |
958 | * | |
c5aa87bb TH |
959 | * This function checks the work item address and work function to avoid |
960 | * false positives. Note that this isn't complete as one may construct a | |
961 | * work function which can introduce dependency onto itself through a | |
962 | * recycled work item. Well, if somebody wants to shoot oneself in the | |
963 | * foot that badly, there's only so much we can do, and if such deadlock | |
964 | * actually occurs, it should be easy to locate the culprit work function. | |
8cca0eea TH |
965 | * |
966 | * CONTEXT: | |
d565ed63 | 967 | * spin_lock_irq(pool->lock). |
8cca0eea | 968 | * |
d185af30 YB |
969 | * Return: |
970 | * Pointer to worker which is executing @work if found, %NULL | |
8cca0eea | 971 | * otherwise. |
4d707b9f | 972 | */ |
c9e7cf27 | 973 | static struct worker *find_worker_executing_work(struct worker_pool *pool, |
8cca0eea | 974 | struct work_struct *work) |
4d707b9f | 975 | { |
42f8570f | 976 | struct worker *worker; |
42f8570f | 977 | |
b67bfe0d | 978 | hash_for_each_possible(pool->busy_hash, worker, hentry, |
a2c1c57b TH |
979 | (unsigned long)work) |
980 | if (worker->current_work == work && | |
981 | worker->current_func == work->func) | |
42f8570f SL |
982 | return worker; |
983 | ||
984 | return NULL; | |
4d707b9f ON |
985 | } |
986 | ||
bf4ede01 TH |
987 | /** |
988 | * move_linked_works - move linked works to a list | |
989 | * @work: start of series of works to be scheduled | |
990 | * @head: target list to append @work to | |
991 | * @nextp: out paramter for nested worklist walking | |
992 | * | |
993 | * Schedule linked works starting from @work to @head. Work series to | |
994 | * be scheduled starts at @work and includes any consecutive work with | |
995 | * WORK_STRUCT_LINKED set in its predecessor. | |
996 | * | |
997 | * If @nextp is not NULL, it's updated to point to the next work of | |
998 | * the last scheduled work. This allows move_linked_works() to be | |
999 | * nested inside outer list_for_each_entry_safe(). | |
1000 | * | |
1001 | * CONTEXT: | |
d565ed63 | 1002 | * spin_lock_irq(pool->lock). |
bf4ede01 TH |
1003 | */ |
1004 | static void move_linked_works(struct work_struct *work, struct list_head *head, | |
1005 | struct work_struct **nextp) | |
1006 | { | |
1007 | struct work_struct *n; | |
1008 | ||
1009 | /* | |
1010 | * Linked worklist will always end before the end of the list, | |
1011 | * use NULL for list head. | |
1012 | */ | |
1013 | list_for_each_entry_safe_from(work, n, NULL, entry) { | |
1014 | list_move_tail(&work->entry, head); | |
1015 | if (!(*work_data_bits(work) & WORK_STRUCT_LINKED)) | |
1016 | break; | |
1017 | } | |
1018 | ||
1019 | /* | |
1020 | * If we're already inside safe list traversal and have moved | |
1021 | * multiple works to the scheduled queue, the next position | |
1022 | * needs to be updated. | |
1023 | */ | |
1024 | if (nextp) | |
1025 | *nextp = n; | |
1026 | } | |
1027 | ||
8864b4e5 TH |
1028 | /** |
1029 | * get_pwq - get an extra reference on the specified pool_workqueue | |
1030 | * @pwq: pool_workqueue to get | |
1031 | * | |
1032 | * Obtain an extra reference on @pwq. The caller should guarantee that | |
1033 | * @pwq has positive refcnt and be holding the matching pool->lock. | |
1034 | */ | |
1035 | static void get_pwq(struct pool_workqueue *pwq) | |
1036 | { | |
1037 | lockdep_assert_held(&pwq->pool->lock); | |
1038 | WARN_ON_ONCE(pwq->refcnt <= 0); | |
1039 | pwq->refcnt++; | |
1040 | } | |
1041 | ||
1042 | /** | |
1043 | * put_pwq - put a pool_workqueue reference | |
1044 | * @pwq: pool_workqueue to put | |
1045 | * | |
1046 | * Drop a reference of @pwq. If its refcnt reaches zero, schedule its | |
1047 | * destruction. The caller should be holding the matching pool->lock. | |
1048 | */ | |
1049 | static void put_pwq(struct pool_workqueue *pwq) | |
1050 | { | |
1051 | lockdep_assert_held(&pwq->pool->lock); | |
1052 | if (likely(--pwq->refcnt)) | |
1053 | return; | |
1054 | if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND))) | |
1055 | return; | |
1056 | /* | |
1057 | * @pwq can't be released under pool->lock, bounce to | |
1058 | * pwq_unbound_release_workfn(). This never recurses on the same | |
1059 | * pool->lock as this path is taken only for unbound workqueues and | |
1060 | * the release work item is scheduled on a per-cpu workqueue. To | |
1061 | * avoid lockdep warning, unbound pool->locks are given lockdep | |
1062 | * subclass of 1 in get_unbound_pool(). | |
1063 | */ | |
1064 | schedule_work(&pwq->unbound_release_work); | |
1065 | } | |
1066 | ||
dce90d47 TH |
1067 | /** |
1068 | * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock | |
1069 | * @pwq: pool_workqueue to put (can be %NULL) | |
1070 | * | |
1071 | * put_pwq() with locking. This function also allows %NULL @pwq. | |
1072 | */ | |
1073 | static void put_pwq_unlocked(struct pool_workqueue *pwq) | |
1074 | { | |
1075 | if (pwq) { | |
1076 | /* | |
1077 | * As both pwqs and pools are sched-RCU protected, the | |
1078 | * following lock operations are safe. | |
1079 | */ | |
1080 | spin_lock_irq(&pwq->pool->lock); | |
1081 | put_pwq(pwq); | |
1082 | spin_unlock_irq(&pwq->pool->lock); | |
1083 | } | |
1084 | } | |
1085 | ||
112202d9 | 1086 | static void pwq_activate_delayed_work(struct work_struct *work) |
bf4ede01 | 1087 | { |
112202d9 | 1088 | struct pool_workqueue *pwq = get_work_pwq(work); |
bf4ede01 TH |
1089 | |
1090 | trace_workqueue_activate_work(work); | |
112202d9 | 1091 | move_linked_works(work, &pwq->pool->worklist, NULL); |
bf4ede01 | 1092 | __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work)); |
112202d9 | 1093 | pwq->nr_active++; |
bf4ede01 TH |
1094 | } |
1095 | ||
112202d9 | 1096 | static void pwq_activate_first_delayed(struct pool_workqueue *pwq) |
3aa62497 | 1097 | { |
112202d9 | 1098 | struct work_struct *work = list_first_entry(&pwq->delayed_works, |
3aa62497 LJ |
1099 | struct work_struct, entry); |
1100 | ||
112202d9 | 1101 | pwq_activate_delayed_work(work); |
3aa62497 LJ |
1102 | } |
1103 | ||
bf4ede01 | 1104 | /** |
112202d9 TH |
1105 | * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight |
1106 | * @pwq: pwq of interest | |
bf4ede01 | 1107 | * @color: color of work which left the queue |
bf4ede01 TH |
1108 | * |
1109 | * A work either has completed or is removed from pending queue, | |
112202d9 | 1110 | * decrement nr_in_flight of its pwq and handle workqueue flushing. |
bf4ede01 TH |
1111 | * |
1112 | * CONTEXT: | |
d565ed63 | 1113 | * spin_lock_irq(pool->lock). |
bf4ede01 | 1114 | */ |
112202d9 | 1115 | static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color) |
bf4ede01 | 1116 | { |
8864b4e5 | 1117 | /* uncolored work items don't participate in flushing or nr_active */ |
bf4ede01 | 1118 | if (color == WORK_NO_COLOR) |
8864b4e5 | 1119 | goto out_put; |
bf4ede01 | 1120 | |
112202d9 | 1121 | pwq->nr_in_flight[color]--; |
bf4ede01 | 1122 | |
112202d9 TH |
1123 | pwq->nr_active--; |
1124 | if (!list_empty(&pwq->delayed_works)) { | |
b3f9f405 | 1125 | /* one down, submit a delayed one */ |
112202d9 TH |
1126 | if (pwq->nr_active < pwq->max_active) |
1127 | pwq_activate_first_delayed(pwq); | |
bf4ede01 TH |
1128 | } |
1129 | ||
1130 | /* is flush in progress and are we at the flushing tip? */ | |
112202d9 | 1131 | if (likely(pwq->flush_color != color)) |
8864b4e5 | 1132 | goto out_put; |
bf4ede01 TH |
1133 | |
1134 | /* are there still in-flight works? */ | |
112202d9 | 1135 | if (pwq->nr_in_flight[color]) |
8864b4e5 | 1136 | goto out_put; |
bf4ede01 | 1137 | |
112202d9 TH |
1138 | /* this pwq is done, clear flush_color */ |
1139 | pwq->flush_color = -1; | |
bf4ede01 TH |
1140 | |
1141 | /* | |
112202d9 | 1142 | * If this was the last pwq, wake up the first flusher. It |
bf4ede01 TH |
1143 | * will handle the rest. |
1144 | */ | |
112202d9 TH |
1145 | if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush)) |
1146 | complete(&pwq->wq->first_flusher->done); | |
8864b4e5 TH |
1147 | out_put: |
1148 | put_pwq(pwq); | |
bf4ede01 TH |
1149 | } |
1150 | ||
36e227d2 | 1151 | /** |
bbb68dfa | 1152 | * try_to_grab_pending - steal work item from worklist and disable irq |
36e227d2 TH |
1153 | * @work: work item to steal |
1154 | * @is_dwork: @work is a delayed_work | |
bbb68dfa | 1155 | * @flags: place to store irq state |
36e227d2 TH |
1156 | * |
1157 | * Try to grab PENDING bit of @work. This function can handle @work in any | |
d185af30 | 1158 | * stable state - idle, on timer or on worklist. |
36e227d2 | 1159 | * |
d185af30 | 1160 | * Return: |
36e227d2 TH |
1161 | * 1 if @work was pending and we successfully stole PENDING |
1162 | * 0 if @work was idle and we claimed PENDING | |
1163 | * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry | |
bbb68dfa TH |
1164 | * -ENOENT if someone else is canceling @work, this state may persist |
1165 | * for arbitrarily long | |
36e227d2 | 1166 | * |
d185af30 | 1167 | * Note: |
bbb68dfa | 1168 | * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting |
e0aecdd8 TH |
1169 | * interrupted while holding PENDING and @work off queue, irq must be |
1170 | * disabled on entry. This, combined with delayed_work->timer being | |
1171 | * irqsafe, ensures that we return -EAGAIN for finite short period of time. | |
bbb68dfa TH |
1172 | * |
1173 | * On successful return, >= 0, irq is disabled and the caller is | |
1174 | * responsible for releasing it using local_irq_restore(*@flags). | |
1175 | * | |
e0aecdd8 | 1176 | * This function is safe to call from any context including IRQ handler. |
bf4ede01 | 1177 | */ |
bbb68dfa TH |
1178 | static int try_to_grab_pending(struct work_struct *work, bool is_dwork, |
1179 | unsigned long *flags) | |
bf4ede01 | 1180 | { |
d565ed63 | 1181 | struct worker_pool *pool; |
112202d9 | 1182 | struct pool_workqueue *pwq; |
bf4ede01 | 1183 | |
bbb68dfa TH |
1184 | local_irq_save(*flags); |
1185 | ||
36e227d2 TH |
1186 | /* try to steal the timer if it exists */ |
1187 | if (is_dwork) { | |
1188 | struct delayed_work *dwork = to_delayed_work(work); | |
1189 | ||
e0aecdd8 TH |
1190 | /* |
1191 | * dwork->timer is irqsafe. If del_timer() fails, it's | |
1192 | * guaranteed that the timer is not queued anywhere and not | |
1193 | * running on the local CPU. | |
1194 | */ | |
36e227d2 TH |
1195 | if (likely(del_timer(&dwork->timer))) |
1196 | return 1; | |
1197 | } | |
1198 | ||
1199 | /* try to claim PENDING the normal way */ | |
bf4ede01 TH |
1200 | if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) |
1201 | return 0; | |
1202 | ||
1203 | /* | |
1204 | * The queueing is in progress, or it is already queued. Try to | |
1205 | * steal it from ->worklist without clearing WORK_STRUCT_PENDING. | |
1206 | */ | |
d565ed63 TH |
1207 | pool = get_work_pool(work); |
1208 | if (!pool) | |
bbb68dfa | 1209 | goto fail; |
bf4ede01 | 1210 | |
d565ed63 | 1211 | spin_lock(&pool->lock); |
0b3dae68 | 1212 | /* |
112202d9 TH |
1213 | * work->data is guaranteed to point to pwq only while the work |
1214 | * item is queued on pwq->wq, and both updating work->data to point | |
1215 | * to pwq on queueing and to pool on dequeueing are done under | |
1216 | * pwq->pool->lock. This in turn guarantees that, if work->data | |
1217 | * points to pwq which is associated with a locked pool, the work | |
0b3dae68 LJ |
1218 | * item is currently queued on that pool. |
1219 | */ | |
112202d9 TH |
1220 | pwq = get_work_pwq(work); |
1221 | if (pwq && pwq->pool == pool) { | |
16062836 TH |
1222 | debug_work_deactivate(work); |
1223 | ||
1224 | /* | |
1225 | * A delayed work item cannot be grabbed directly because | |
1226 | * it might have linked NO_COLOR work items which, if left | |
112202d9 | 1227 | * on the delayed_list, will confuse pwq->nr_active |
16062836 TH |
1228 | * management later on and cause stall. Make sure the work |
1229 | * item is activated before grabbing. | |
1230 | */ | |
1231 | if (*work_data_bits(work) & WORK_STRUCT_DELAYED) | |
112202d9 | 1232 | pwq_activate_delayed_work(work); |
16062836 TH |
1233 | |
1234 | list_del_init(&work->entry); | |
112202d9 | 1235 | pwq_dec_nr_in_flight(get_work_pwq(work), get_work_color(work)); |
16062836 | 1236 | |
112202d9 | 1237 | /* work->data points to pwq iff queued, point to pool */ |
16062836 TH |
1238 | set_work_pool_and_keep_pending(work, pool->id); |
1239 | ||
1240 | spin_unlock(&pool->lock); | |
1241 | return 1; | |
bf4ede01 | 1242 | } |
d565ed63 | 1243 | spin_unlock(&pool->lock); |
bbb68dfa TH |
1244 | fail: |
1245 | local_irq_restore(*flags); | |
1246 | if (work_is_canceling(work)) | |
1247 | return -ENOENT; | |
1248 | cpu_relax(); | |
36e227d2 | 1249 | return -EAGAIN; |
bf4ede01 TH |
1250 | } |
1251 | ||
4690c4ab | 1252 | /** |
706026c2 | 1253 | * insert_work - insert a work into a pool |
112202d9 | 1254 | * @pwq: pwq @work belongs to |
4690c4ab TH |
1255 | * @work: work to insert |
1256 | * @head: insertion point | |
1257 | * @extra_flags: extra WORK_STRUCT_* flags to set | |
1258 | * | |
112202d9 | 1259 | * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to |
706026c2 | 1260 | * work_struct flags. |
4690c4ab TH |
1261 | * |
1262 | * CONTEXT: | |
d565ed63 | 1263 | * spin_lock_irq(pool->lock). |
4690c4ab | 1264 | */ |
112202d9 TH |
1265 | static void insert_work(struct pool_workqueue *pwq, struct work_struct *work, |
1266 | struct list_head *head, unsigned int extra_flags) | |
b89deed3 | 1267 | { |
112202d9 | 1268 | struct worker_pool *pool = pwq->pool; |
e22bee78 | 1269 | |
4690c4ab | 1270 | /* we own @work, set data and link */ |
112202d9 | 1271 | set_work_pwq(work, pwq, extra_flags); |
1a4d9b0a | 1272 | list_add_tail(&work->entry, head); |
8864b4e5 | 1273 | get_pwq(pwq); |
e22bee78 TH |
1274 | |
1275 | /* | |
c5aa87bb TH |
1276 | * Ensure either wq_worker_sleeping() sees the above |
1277 | * list_add_tail() or we see zero nr_running to avoid workers lying | |
1278 | * around lazily while there are works to be processed. | |
e22bee78 TH |
1279 | */ |
1280 | smp_mb(); | |
1281 | ||
63d95a91 TH |
1282 | if (__need_more_worker(pool)) |
1283 | wake_up_worker(pool); | |
b89deed3 ON |
1284 | } |
1285 | ||
c8efcc25 TH |
1286 | /* |
1287 | * Test whether @work is being queued from another work executing on the | |
8d03ecfe | 1288 | * same workqueue. |
c8efcc25 TH |
1289 | */ |
1290 | static bool is_chained_work(struct workqueue_struct *wq) | |
1291 | { | |
8d03ecfe TH |
1292 | struct worker *worker; |
1293 | ||
1294 | worker = current_wq_worker(); | |
1295 | /* | |
1296 | * Return %true iff I'm a worker execuing a work item on @wq. If | |
1297 | * I'm @worker, it's safe to dereference it without locking. | |
1298 | */ | |
112202d9 | 1299 | return worker && worker->current_pwq->wq == wq; |
c8efcc25 TH |
1300 | } |
1301 | ||
d84ff051 | 1302 | static void __queue_work(int cpu, struct workqueue_struct *wq, |
1da177e4 LT |
1303 | struct work_struct *work) |
1304 | { | |
112202d9 | 1305 | struct pool_workqueue *pwq; |
c9178087 | 1306 | struct worker_pool *last_pool; |
1e19ffc6 | 1307 | struct list_head *worklist; |
8a2e8e5d | 1308 | unsigned int work_flags; |
b75cac93 | 1309 | unsigned int req_cpu = cpu; |
8930caba TH |
1310 | |
1311 | /* | |
1312 | * While a work item is PENDING && off queue, a task trying to | |
1313 | * steal the PENDING will busy-loop waiting for it to either get | |
1314 | * queued or lose PENDING. Grabbing PENDING and queueing should | |
1315 | * happen with IRQ disabled. | |
1316 | */ | |
1317 | WARN_ON_ONCE(!irqs_disabled()); | |
1da177e4 | 1318 | |
dc186ad7 | 1319 | debug_work_activate(work); |
1e19ffc6 | 1320 | |
9ef28a73 | 1321 | /* if draining, only works from the same workqueue are allowed */ |
618b01eb | 1322 | if (unlikely(wq->flags & __WQ_DRAINING) && |
c8efcc25 | 1323 | WARN_ON_ONCE(!is_chained_work(wq))) |
e41e704b | 1324 | return; |
9e8cd2f5 | 1325 | retry: |
df2d5ae4 TH |
1326 | if (req_cpu == WORK_CPU_UNBOUND) |
1327 | cpu = raw_smp_processor_id(); | |
1328 | ||
c9178087 | 1329 | /* pwq which will be used unless @work is executing elsewhere */ |
df2d5ae4 | 1330 | if (!(wq->flags & WQ_UNBOUND)) |
7fb98ea7 | 1331 | pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); |
df2d5ae4 TH |
1332 | else |
1333 | pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); | |
dbf2576e | 1334 | |
c9178087 TH |
1335 | /* |
1336 | * If @work was previously on a different pool, it might still be | |
1337 | * running there, in which case the work needs to be queued on that | |
1338 | * pool to guarantee non-reentrancy. | |
1339 | */ | |
1340 | last_pool = get_work_pool(work); | |
1341 | if (last_pool && last_pool != pwq->pool) { | |
1342 | struct worker *worker; | |
18aa9eff | 1343 | |
c9178087 | 1344 | spin_lock(&last_pool->lock); |
18aa9eff | 1345 | |
c9178087 | 1346 | worker = find_worker_executing_work(last_pool, work); |
18aa9eff | 1347 | |
c9178087 TH |
1348 | if (worker && worker->current_pwq->wq == wq) { |
1349 | pwq = worker->current_pwq; | |
8930caba | 1350 | } else { |
c9178087 TH |
1351 | /* meh... not running there, queue here */ |
1352 | spin_unlock(&last_pool->lock); | |
112202d9 | 1353 | spin_lock(&pwq->pool->lock); |
8930caba | 1354 | } |
f3421797 | 1355 | } else { |
112202d9 | 1356 | spin_lock(&pwq->pool->lock); |
502ca9d8 TH |
1357 | } |
1358 | ||
9e8cd2f5 TH |
1359 | /* |
1360 | * pwq is determined and locked. For unbound pools, we could have | |
1361 | * raced with pwq release and it could already be dead. If its | |
1362 | * refcnt is zero, repeat pwq selection. Note that pwqs never die | |
df2d5ae4 TH |
1363 | * without another pwq replacing it in the numa_pwq_tbl or while |
1364 | * work items are executing on it, so the retrying is guaranteed to | |
9e8cd2f5 TH |
1365 | * make forward-progress. |
1366 | */ | |
1367 | if (unlikely(!pwq->refcnt)) { | |
1368 | if (wq->flags & WQ_UNBOUND) { | |
1369 | spin_unlock(&pwq->pool->lock); | |
1370 | cpu_relax(); | |
1371 | goto retry; | |
1372 | } | |
1373 | /* oops */ | |
1374 | WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt", | |
1375 | wq->name, cpu); | |
1376 | } | |
1377 | ||
112202d9 TH |
1378 | /* pwq determined, queue */ |
1379 | trace_workqueue_queue_work(req_cpu, pwq, work); | |
502ca9d8 | 1380 | |
f5b2552b | 1381 | if (WARN_ON(!list_empty(&work->entry))) { |
112202d9 | 1382 | spin_unlock(&pwq->pool->lock); |
f5b2552b DC |
1383 | return; |
1384 | } | |
1e19ffc6 | 1385 | |
112202d9 TH |
1386 | pwq->nr_in_flight[pwq->work_color]++; |
1387 | work_flags = work_color_to_flags(pwq->work_color); | |
1e19ffc6 | 1388 | |
112202d9 | 1389 | if (likely(pwq->nr_active < pwq->max_active)) { |
cdadf009 | 1390 | trace_workqueue_activate_work(work); |
112202d9 TH |
1391 | pwq->nr_active++; |
1392 | worklist = &pwq->pool->worklist; | |
8a2e8e5d TH |
1393 | } else { |
1394 | work_flags |= WORK_STRUCT_DELAYED; | |
112202d9 | 1395 | worklist = &pwq->delayed_works; |
8a2e8e5d | 1396 | } |
1e19ffc6 | 1397 | |
112202d9 | 1398 | insert_work(pwq, work, worklist, work_flags); |
1e19ffc6 | 1399 | |
112202d9 | 1400 | spin_unlock(&pwq->pool->lock); |
1da177e4 LT |
1401 | } |
1402 | ||
0fcb78c2 | 1403 | /** |
c1a220e7 ZR |
1404 | * queue_work_on - queue work on specific cpu |
1405 | * @cpu: CPU number to execute work on | |
0fcb78c2 REB |
1406 | * @wq: workqueue to use |
1407 | * @work: work to queue | |
1408 | * | |
c1a220e7 ZR |
1409 | * We queue the work to a specific CPU, the caller must ensure it |
1410 | * can't go away. | |
d185af30 YB |
1411 | * |
1412 | * Return: %false if @work was already on a queue, %true otherwise. | |
1da177e4 | 1413 | */ |
d4283e93 TH |
1414 | bool queue_work_on(int cpu, struct workqueue_struct *wq, |
1415 | struct work_struct *work) | |
1da177e4 | 1416 | { |
d4283e93 | 1417 | bool ret = false; |
8930caba | 1418 | unsigned long flags; |
ef1ca236 | 1419 | |
8930caba | 1420 | local_irq_save(flags); |
c1a220e7 | 1421 | |
22df02bb | 1422 | if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { |
4690c4ab | 1423 | __queue_work(cpu, wq, work); |
d4283e93 | 1424 | ret = true; |
c1a220e7 | 1425 | } |
ef1ca236 | 1426 | |
8930caba | 1427 | local_irq_restore(flags); |
1da177e4 LT |
1428 | return ret; |
1429 | } | |
ad7b1f84 | 1430 | EXPORT_SYMBOL(queue_work_on); |
1da177e4 | 1431 | |
d8e794df | 1432 | void delayed_work_timer_fn(unsigned long __data) |
1da177e4 | 1433 | { |
52bad64d | 1434 | struct delayed_work *dwork = (struct delayed_work *)__data; |
1da177e4 | 1435 | |
e0aecdd8 | 1436 | /* should have been called from irqsafe timer with irq already off */ |
60c057bc | 1437 | __queue_work(dwork->cpu, dwork->wq, &dwork->work); |
1da177e4 | 1438 | } |
1438ade5 | 1439 | EXPORT_SYMBOL(delayed_work_timer_fn); |
1da177e4 | 1440 | |
7beb2edf TH |
1441 | static void __queue_delayed_work(int cpu, struct workqueue_struct *wq, |
1442 | struct delayed_work *dwork, unsigned long delay) | |
1da177e4 | 1443 | { |
7beb2edf TH |
1444 | struct timer_list *timer = &dwork->timer; |
1445 | struct work_struct *work = &dwork->work; | |
7beb2edf TH |
1446 | |
1447 | WARN_ON_ONCE(timer->function != delayed_work_timer_fn || | |
1448 | timer->data != (unsigned long)dwork); | |
fc4b514f TH |
1449 | WARN_ON_ONCE(timer_pending(timer)); |
1450 | WARN_ON_ONCE(!list_empty(&work->entry)); | |
7beb2edf | 1451 | |
8852aac2 TH |
1452 | /* |
1453 | * If @delay is 0, queue @dwork->work immediately. This is for | |
1454 | * both optimization and correctness. The earliest @timer can | |
1455 | * expire is on the closest next tick and delayed_work users depend | |
1456 | * on that there's no such delay when @delay is 0. | |
1457 | */ | |
1458 | if (!delay) { | |
1459 | __queue_work(cpu, wq, &dwork->work); | |
1460 | return; | |
1461 | } | |
1462 | ||
7beb2edf | 1463 | timer_stats_timer_set_start_info(&dwork->timer); |
1da177e4 | 1464 | |
60c057bc | 1465 | dwork->wq = wq; |
1265057f | 1466 | dwork->cpu = cpu; |
7beb2edf TH |
1467 | timer->expires = jiffies + delay; |
1468 | ||
1469 | if (unlikely(cpu != WORK_CPU_UNBOUND)) | |
1470 | add_timer_on(timer, cpu); | |
1471 | else | |
1472 | add_timer(timer); | |
1da177e4 LT |
1473 | } |
1474 | ||
0fcb78c2 REB |
1475 | /** |
1476 | * queue_delayed_work_on - queue work on specific CPU after delay | |
1477 | * @cpu: CPU number to execute work on | |
1478 | * @wq: workqueue to use | |
af9997e4 | 1479 | * @dwork: work to queue |
0fcb78c2 REB |
1480 | * @delay: number of jiffies to wait before queueing |
1481 | * | |
d185af30 | 1482 | * Return: %false if @work was already on a queue, %true otherwise. If |
715f1300 TH |
1483 | * @delay is zero and @dwork is idle, it will be scheduled for immediate |
1484 | * execution. | |
0fcb78c2 | 1485 | */ |
d4283e93 TH |
1486 | bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq, |
1487 | struct delayed_work *dwork, unsigned long delay) | |
7a6bc1cd | 1488 | { |
52bad64d | 1489 | struct work_struct *work = &dwork->work; |
d4283e93 | 1490 | bool ret = false; |
8930caba | 1491 | unsigned long flags; |
7a6bc1cd | 1492 | |
8930caba TH |
1493 | /* read the comment in __queue_work() */ |
1494 | local_irq_save(flags); | |
7a6bc1cd | 1495 | |
22df02bb | 1496 | if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { |
7beb2edf | 1497 | __queue_delayed_work(cpu, wq, dwork, delay); |
d4283e93 | 1498 | ret = true; |
7a6bc1cd | 1499 | } |
8a3e77cc | 1500 | |
8930caba | 1501 | local_irq_restore(flags); |
7a6bc1cd VP |
1502 | return ret; |
1503 | } | |
ad7b1f84 | 1504 | EXPORT_SYMBOL(queue_delayed_work_on); |
c7fc77f7 | 1505 | |
8376fe22 TH |
1506 | /** |
1507 | * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU | |
1508 | * @cpu: CPU number to execute work on | |
1509 | * @wq: workqueue to use | |
1510 | * @dwork: work to queue | |
1511 | * @delay: number of jiffies to wait before queueing | |
1512 | * | |
1513 | * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise, | |
1514 | * modify @dwork's timer so that it expires after @delay. If @delay is | |
1515 | * zero, @work is guaranteed to be scheduled immediately regardless of its | |
1516 | * current state. | |
1517 | * | |
d185af30 | 1518 | * Return: %false if @dwork was idle and queued, %true if @dwork was |
8376fe22 TH |
1519 | * pending and its timer was modified. |
1520 | * | |
e0aecdd8 | 1521 | * This function is safe to call from any context including IRQ handler. |
8376fe22 TH |
1522 | * See try_to_grab_pending() for details. |
1523 | */ | |
1524 | bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq, | |
1525 | struct delayed_work *dwork, unsigned long delay) | |
1526 | { | |
1527 | unsigned long flags; | |
1528 | int ret; | |
c7fc77f7 | 1529 | |
8376fe22 TH |
1530 | do { |
1531 | ret = try_to_grab_pending(&dwork->work, true, &flags); | |
1532 | } while (unlikely(ret == -EAGAIN)); | |
63bc0362 | 1533 | |
8376fe22 TH |
1534 | if (likely(ret >= 0)) { |
1535 | __queue_delayed_work(cpu, wq, dwork, delay); | |
1536 | local_irq_restore(flags); | |
7a6bc1cd | 1537 | } |
8376fe22 TH |
1538 | |
1539 | /* -ENOENT from try_to_grab_pending() becomes %true */ | |
7a6bc1cd VP |
1540 | return ret; |
1541 | } | |
8376fe22 TH |
1542 | EXPORT_SYMBOL_GPL(mod_delayed_work_on); |
1543 | ||
c8e55f36 TH |
1544 | /** |
1545 | * worker_enter_idle - enter idle state | |
1546 | * @worker: worker which is entering idle state | |
1547 | * | |
1548 | * @worker is entering idle state. Update stats and idle timer if | |
1549 | * necessary. | |
1550 | * | |
1551 | * LOCKING: | |
d565ed63 | 1552 | * spin_lock_irq(pool->lock). |
c8e55f36 TH |
1553 | */ |
1554 | static void worker_enter_idle(struct worker *worker) | |
1da177e4 | 1555 | { |
bd7bdd43 | 1556 | struct worker_pool *pool = worker->pool; |
c8e55f36 | 1557 | |
6183c009 TH |
1558 | if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) || |
1559 | WARN_ON_ONCE(!list_empty(&worker->entry) && | |
1560 | (worker->hentry.next || worker->hentry.pprev))) | |
1561 | return; | |
c8e55f36 | 1562 | |
cb444766 TH |
1563 | /* can't use worker_set_flags(), also called from start_worker() */ |
1564 | worker->flags |= WORKER_IDLE; | |
bd7bdd43 | 1565 | pool->nr_idle++; |
e22bee78 | 1566 | worker->last_active = jiffies; |
c8e55f36 TH |
1567 | |
1568 | /* idle_list is LIFO */ | |
bd7bdd43 | 1569 | list_add(&worker->entry, &pool->idle_list); |
db7bccf4 | 1570 | |
628c78e7 TH |
1571 | if (too_many_workers(pool) && !timer_pending(&pool->idle_timer)) |
1572 | mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT); | |
cb444766 | 1573 | |
544ecf31 | 1574 | /* |
706026c2 | 1575 | * Sanity check nr_running. Because wq_unbind_fn() releases |
d565ed63 | 1576 | * pool->lock between setting %WORKER_UNBOUND and zapping |
628c78e7 TH |
1577 | * nr_running, the warning may trigger spuriously. Check iff |
1578 | * unbind is not in progress. | |
544ecf31 | 1579 | */ |
24647570 | 1580 | WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && |
bd7bdd43 | 1581 | pool->nr_workers == pool->nr_idle && |
e19e397a | 1582 | atomic_read(&pool->nr_running)); |
c8e55f36 TH |
1583 | } |
1584 | ||
1585 | /** | |
1586 | * worker_leave_idle - leave idle state | |
1587 | * @worker: worker which is leaving idle state | |
1588 | * | |
1589 | * @worker is leaving idle state. Update stats. | |
1590 | * | |
1591 | * LOCKING: | |
d565ed63 | 1592 | * spin_lock_irq(pool->lock). |
c8e55f36 TH |
1593 | */ |
1594 | static void worker_leave_idle(struct worker *worker) | |
1595 | { | |
bd7bdd43 | 1596 | struct worker_pool *pool = worker->pool; |
c8e55f36 | 1597 | |
6183c009 TH |
1598 | if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE))) |
1599 | return; | |
d302f017 | 1600 | worker_clr_flags(worker, WORKER_IDLE); |
bd7bdd43 | 1601 | pool->nr_idle--; |
c8e55f36 TH |
1602 | list_del_init(&worker->entry); |
1603 | } | |
1604 | ||
e22bee78 | 1605 | /** |
f36dc67b LJ |
1606 | * worker_maybe_bind_and_lock - try to bind %current to worker_pool and lock it |
1607 | * @pool: target worker_pool | |
1608 | * | |
1609 | * Bind %current to the cpu of @pool if it is associated and lock @pool. | |
e22bee78 TH |
1610 | * |
1611 | * Works which are scheduled while the cpu is online must at least be | |
1612 | * scheduled to a worker which is bound to the cpu so that if they are | |
1613 | * flushed from cpu callbacks while cpu is going down, they are | |
1614 | * guaranteed to execute on the cpu. | |
1615 | * | |
f5faa077 | 1616 | * This function is to be used by unbound workers and rescuers to bind |
e22bee78 TH |
1617 | * themselves to the target cpu and may race with cpu going down or |
1618 | * coming online. kthread_bind() can't be used because it may put the | |
1619 | * worker to already dead cpu and set_cpus_allowed_ptr() can't be used | |
706026c2 | 1620 | * verbatim as it's best effort and blocking and pool may be |
e22bee78 TH |
1621 | * [dis]associated in the meantime. |
1622 | * | |
706026c2 | 1623 | * This function tries set_cpus_allowed() and locks pool and verifies the |
24647570 | 1624 | * binding against %POOL_DISASSOCIATED which is set during |
f2d5a0ee TH |
1625 | * %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker |
1626 | * enters idle state or fetches works without dropping lock, it can | |
1627 | * guarantee the scheduling requirement described in the first paragraph. | |
e22bee78 TH |
1628 | * |
1629 | * CONTEXT: | |
d565ed63 | 1630 | * Might sleep. Called without any lock but returns with pool->lock |
e22bee78 TH |
1631 | * held. |
1632 | * | |
d185af30 | 1633 | * Return: |
706026c2 | 1634 | * %true if the associated pool is online (@worker is successfully |
e22bee78 TH |
1635 | * bound), %false if offline. |
1636 | */ | |
f36dc67b | 1637 | static bool worker_maybe_bind_and_lock(struct worker_pool *pool) |
d565ed63 | 1638 | __acquires(&pool->lock) |
e22bee78 | 1639 | { |
e22bee78 | 1640 | while (true) { |
4e6045f1 | 1641 | /* |
e22bee78 TH |
1642 | * The following call may fail, succeed or succeed |
1643 | * without actually migrating the task to the cpu if | |
1644 | * it races with cpu hotunplug operation. Verify | |
24647570 | 1645 | * against POOL_DISASSOCIATED. |
4e6045f1 | 1646 | */ |
24647570 | 1647 | if (!(pool->flags & POOL_DISASSOCIATED)) |
7a4e344c | 1648 | set_cpus_allowed_ptr(current, pool->attrs->cpumask); |
e22bee78 | 1649 | |
d565ed63 | 1650 | spin_lock_irq(&pool->lock); |
24647570 | 1651 | if (pool->flags & POOL_DISASSOCIATED) |
e22bee78 | 1652 | return false; |
f5faa077 | 1653 | if (task_cpu(current) == pool->cpu && |
7a4e344c | 1654 | cpumask_equal(¤t->cpus_allowed, pool->attrs->cpumask)) |
e22bee78 | 1655 | return true; |
d565ed63 | 1656 | spin_unlock_irq(&pool->lock); |
e22bee78 | 1657 | |
5035b20f TH |
1658 | /* |
1659 | * We've raced with CPU hot[un]plug. Give it a breather | |
1660 | * and retry migration. cond_resched() is required here; | |
1661 | * otherwise, we might deadlock against cpu_stop trying to | |
1662 | * bring down the CPU on non-preemptive kernel. | |
1663 | */ | |
e22bee78 | 1664 | cpu_relax(); |
5035b20f | 1665 | cond_resched(); |
e22bee78 TH |
1666 | } |
1667 | } | |
1668 | ||
c34056a3 TH |
1669 | static struct worker *alloc_worker(void) |
1670 | { | |
1671 | struct worker *worker; | |
1672 | ||
1673 | worker = kzalloc(sizeof(*worker), GFP_KERNEL); | |
c8e55f36 TH |
1674 | if (worker) { |
1675 | INIT_LIST_HEAD(&worker->entry); | |
affee4b2 | 1676 | INIT_LIST_HEAD(&worker->scheduled); |
e22bee78 TH |
1677 | /* on creation a worker is in !idle && prep state */ |
1678 | worker->flags = WORKER_PREP; | |
c8e55f36 | 1679 | } |
c34056a3 TH |
1680 | return worker; |
1681 | } | |
1682 | ||
60f5a4bc LJ |
1683 | /** |
1684 | * worker_detach_from_pool() - detach a worker from its pool | |
1685 | * @worker: worker which is attached to its pool | |
1686 | * @pool: the pool @worker is attached to | |
1687 | * | |
1688 | * Undo the attaching which had been done in create_worker(). The caller | |
1689 | * worker shouldn't access to the pool after detached except it has other | |
1690 | * reference to the pool. | |
1691 | */ | |
1692 | static void worker_detach_from_pool(struct worker *worker, | |
1693 | struct worker_pool *pool) | |
1694 | { | |
1695 | struct completion *detach_completion = NULL; | |
1696 | ||
1697 | mutex_lock(&pool->manager_mutex); | |
1698 | idr_remove(&pool->worker_idr, worker->id); | |
1699 | if (idr_is_empty(&pool->worker_idr)) | |
1700 | detach_completion = pool->detach_completion; | |
1701 | mutex_unlock(&pool->manager_mutex); | |
1702 | ||
1703 | if (detach_completion) | |
1704 | complete(detach_completion); | |
1705 | } | |
1706 | ||
c34056a3 TH |
1707 | /** |
1708 | * create_worker - create a new workqueue worker | |
63d95a91 | 1709 | * @pool: pool the new worker will belong to |
c34056a3 | 1710 | * |
73eb7fe7 LJ |
1711 | * Create a new worker which is attached to @pool. The new worker must be |
1712 | * started by start_worker(). | |
c34056a3 TH |
1713 | * |
1714 | * CONTEXT: | |
1715 | * Might sleep. Does GFP_KERNEL allocations. | |
1716 | * | |
d185af30 | 1717 | * Return: |
c34056a3 TH |
1718 | * Pointer to the newly created worker. |
1719 | */ | |
bc2ae0f5 | 1720 | static struct worker *create_worker(struct worker_pool *pool) |
c34056a3 | 1721 | { |
c34056a3 | 1722 | struct worker *worker = NULL; |
f3421797 | 1723 | int id = -1; |
e3c916a4 | 1724 | char id_buf[16]; |
c34056a3 | 1725 | |
cd549687 TH |
1726 | lockdep_assert_held(&pool->manager_mutex); |
1727 | ||
822d8405 TH |
1728 | /* |
1729 | * ID is needed to determine kthread name. Allocate ID first | |
1730 | * without installing the pointer. | |
1731 | */ | |
9625ab17 | 1732 | id = idr_alloc(&pool->worker_idr, NULL, 0, 0, GFP_KERNEL); |
822d8405 TH |
1733 | if (id < 0) |
1734 | goto fail; | |
c34056a3 TH |
1735 | |
1736 | worker = alloc_worker(); | |
1737 | if (!worker) | |
1738 | goto fail; | |
1739 | ||
bd7bdd43 | 1740 | worker->pool = pool; |
c34056a3 TH |
1741 | worker->id = id; |
1742 | ||
29c91e99 | 1743 | if (pool->cpu >= 0) |
e3c916a4 TH |
1744 | snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id, |
1745 | pool->attrs->nice < 0 ? "H" : ""); | |
f3421797 | 1746 | else |
e3c916a4 TH |
1747 | snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id); |
1748 | ||
f3f90ad4 | 1749 | worker->task = kthread_create_on_node(worker_thread, worker, pool->node, |
e3c916a4 | 1750 | "kworker/%s", id_buf); |
c34056a3 TH |
1751 | if (IS_ERR(worker->task)) |
1752 | goto fail; | |
1753 | ||
91151228 ON |
1754 | set_user_nice(worker->task, pool->attrs->nice); |
1755 | ||
1756 | /* prevent userland from meddling with cpumask of workqueue workers */ | |
1757 | worker->task->flags |= PF_NO_SETAFFINITY; | |
1758 | ||
c5aa87bb TH |
1759 | /* |
1760 | * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any | |
1761 | * online CPUs. It'll be re-applied when any of the CPUs come up. | |
1762 | */ | |
7a4e344c | 1763 | set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask); |
3270476a | 1764 | |
7a4e344c TH |
1765 | /* |
1766 | * The caller is responsible for ensuring %POOL_DISASSOCIATED | |
1767 | * remains stable across this function. See the comments above the | |
1768 | * flag definition for details. | |
1769 | */ | |
1770 | if (pool->flags & POOL_DISASSOCIATED) | |
bc2ae0f5 | 1771 | worker->flags |= WORKER_UNBOUND; |
c34056a3 | 1772 | |
822d8405 | 1773 | /* successful, commit the pointer to idr */ |
822d8405 | 1774 | idr_replace(&pool->worker_idr, worker, worker->id); |
822d8405 | 1775 | |
c34056a3 | 1776 | return worker; |
822d8405 | 1777 | |
c34056a3 | 1778 | fail: |
9625ab17 | 1779 | if (id >= 0) |
822d8405 | 1780 | idr_remove(&pool->worker_idr, id); |
c34056a3 TH |
1781 | kfree(worker); |
1782 | return NULL; | |
1783 | } | |
1784 | ||
1785 | /** | |
1786 | * start_worker - start a newly created worker | |
1787 | * @worker: worker to start | |
1788 | * | |
706026c2 | 1789 | * Make the pool aware of @worker and start it. |
c34056a3 TH |
1790 | * |
1791 | * CONTEXT: | |
d565ed63 | 1792 | * spin_lock_irq(pool->lock). |
c34056a3 TH |
1793 | */ |
1794 | static void start_worker(struct worker *worker) | |
1795 | { | |
bd7bdd43 | 1796 | worker->pool->nr_workers++; |
c8e55f36 | 1797 | worker_enter_idle(worker); |
c34056a3 TH |
1798 | wake_up_process(worker->task); |
1799 | } | |
1800 | ||
ebf44d16 TH |
1801 | /** |
1802 | * create_and_start_worker - create and start a worker for a pool | |
1803 | * @pool: the target pool | |
1804 | * | |
cd549687 | 1805 | * Grab the managership of @pool and create and start a new worker for it. |
d185af30 YB |
1806 | * |
1807 | * Return: 0 on success. A negative error code otherwise. | |
ebf44d16 TH |
1808 | */ |
1809 | static int create_and_start_worker(struct worker_pool *pool) | |
1810 | { | |
1811 | struct worker *worker; | |
1812 | ||
cd549687 TH |
1813 | mutex_lock(&pool->manager_mutex); |
1814 | ||
ebf44d16 TH |
1815 | worker = create_worker(pool); |
1816 | if (worker) { | |
1817 | spin_lock_irq(&pool->lock); | |
1818 | start_worker(worker); | |
1819 | spin_unlock_irq(&pool->lock); | |
1820 | } | |
1821 | ||
cd549687 TH |
1822 | mutex_unlock(&pool->manager_mutex); |
1823 | ||
ebf44d16 TH |
1824 | return worker ? 0 : -ENOMEM; |
1825 | } | |
1826 | ||
c34056a3 TH |
1827 | /** |
1828 | * destroy_worker - destroy a workqueue worker | |
1829 | * @worker: worker to be destroyed | |
1830 | * | |
73eb7fe7 LJ |
1831 | * Destroy @worker and adjust @pool stats accordingly. The worker should |
1832 | * be idle. | |
c8e55f36 TH |
1833 | * |
1834 | * CONTEXT: | |
60f5a4bc | 1835 | * spin_lock_irq(pool->lock). |
c34056a3 TH |
1836 | */ |
1837 | static void destroy_worker(struct worker *worker) | |
1838 | { | |
bd7bdd43 | 1839 | struct worker_pool *pool = worker->pool; |
c34056a3 | 1840 | |
cd549687 TH |
1841 | lockdep_assert_held(&pool->lock); |
1842 | ||
c34056a3 | 1843 | /* sanity check frenzy */ |
6183c009 | 1844 | if (WARN_ON(worker->current_work) || |
73eb7fe7 LJ |
1845 | WARN_ON(!list_empty(&worker->scheduled)) || |
1846 | WARN_ON(!(worker->flags & WORKER_IDLE))) | |
6183c009 | 1847 | return; |
c34056a3 | 1848 | |
73eb7fe7 LJ |
1849 | pool->nr_workers--; |
1850 | pool->nr_idle--; | |
c8e55f36 TH |
1851 | |
1852 | list_del_init(&worker->entry); | |
cb444766 | 1853 | worker->flags |= WORKER_DIE; |
60f5a4bc | 1854 | wake_up_process(worker->task); |
c34056a3 TH |
1855 | } |
1856 | ||
63d95a91 | 1857 | static void idle_worker_timeout(unsigned long __pool) |
e22bee78 | 1858 | { |
63d95a91 | 1859 | struct worker_pool *pool = (void *)__pool; |
e22bee78 | 1860 | |
d565ed63 | 1861 | spin_lock_irq(&pool->lock); |
e22bee78 | 1862 | |
3347fc9f | 1863 | while (too_many_workers(pool)) { |
e22bee78 TH |
1864 | struct worker *worker; |
1865 | unsigned long expires; | |
1866 | ||
1867 | /* idle_list is kept in LIFO order, check the last one */ | |
63d95a91 | 1868 | worker = list_entry(pool->idle_list.prev, struct worker, entry); |
e22bee78 TH |
1869 | expires = worker->last_active + IDLE_WORKER_TIMEOUT; |
1870 | ||
3347fc9f | 1871 | if (time_before(jiffies, expires)) { |
63d95a91 | 1872 | mod_timer(&pool->idle_timer, expires); |
3347fc9f | 1873 | break; |
d5abe669 | 1874 | } |
3347fc9f LJ |
1875 | |
1876 | destroy_worker(worker); | |
e22bee78 TH |
1877 | } |
1878 | ||
d565ed63 | 1879 | spin_unlock_irq(&pool->lock); |
e22bee78 | 1880 | } |
d5abe669 | 1881 | |
493a1724 | 1882 | static void send_mayday(struct work_struct *work) |
e22bee78 | 1883 | { |
112202d9 TH |
1884 | struct pool_workqueue *pwq = get_work_pwq(work); |
1885 | struct workqueue_struct *wq = pwq->wq; | |
493a1724 | 1886 | |
2e109a28 | 1887 | lockdep_assert_held(&wq_mayday_lock); |
e22bee78 | 1888 | |
493008a8 | 1889 | if (!wq->rescuer) |
493a1724 | 1890 | return; |
e22bee78 TH |
1891 | |
1892 | /* mayday mayday mayday */ | |
493a1724 | 1893 | if (list_empty(&pwq->mayday_node)) { |
77668c8b LJ |
1894 | /* |
1895 | * If @pwq is for an unbound wq, its base ref may be put at | |
1896 | * any time due to an attribute change. Pin @pwq until the | |
1897 | * rescuer is done with it. | |
1898 | */ | |
1899 | get_pwq(pwq); | |
493a1724 | 1900 | list_add_tail(&pwq->mayday_node, &wq->maydays); |
e22bee78 | 1901 | wake_up_process(wq->rescuer->task); |
493a1724 | 1902 | } |
e22bee78 TH |
1903 | } |
1904 | ||
706026c2 | 1905 | static void pool_mayday_timeout(unsigned long __pool) |
e22bee78 | 1906 | { |
63d95a91 | 1907 | struct worker_pool *pool = (void *)__pool; |
e22bee78 TH |
1908 | struct work_struct *work; |
1909 | ||
2e109a28 | 1910 | spin_lock_irq(&wq_mayday_lock); /* for wq->maydays */ |
493a1724 | 1911 | spin_lock(&pool->lock); |
e22bee78 | 1912 | |
63d95a91 | 1913 | if (need_to_create_worker(pool)) { |
e22bee78 TH |
1914 | /* |
1915 | * We've been trying to create a new worker but | |
1916 | * haven't been successful. We might be hitting an | |
1917 | * allocation deadlock. Send distress signals to | |
1918 | * rescuers. | |
1919 | */ | |
63d95a91 | 1920 | list_for_each_entry(work, &pool->worklist, entry) |
e22bee78 | 1921 | send_mayday(work); |
1da177e4 | 1922 | } |
e22bee78 | 1923 | |
493a1724 | 1924 | spin_unlock(&pool->lock); |
2e109a28 | 1925 | spin_unlock_irq(&wq_mayday_lock); |
e22bee78 | 1926 | |
63d95a91 | 1927 | mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL); |
1da177e4 LT |
1928 | } |
1929 | ||
e22bee78 TH |
1930 | /** |
1931 | * maybe_create_worker - create a new worker if necessary | |
63d95a91 | 1932 | * @pool: pool to create a new worker for |
e22bee78 | 1933 | * |
63d95a91 | 1934 | * Create a new worker for @pool if necessary. @pool is guaranteed to |
e22bee78 TH |
1935 | * have at least one idle worker on return from this function. If |
1936 | * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is | |
63d95a91 | 1937 | * sent to all rescuers with works scheduled on @pool to resolve |
e22bee78 TH |
1938 | * possible allocation deadlock. |
1939 | * | |
c5aa87bb TH |
1940 | * On return, need_to_create_worker() is guaranteed to be %false and |
1941 | * may_start_working() %true. | |
e22bee78 TH |
1942 | * |
1943 | * LOCKING: | |
d565ed63 | 1944 | * spin_lock_irq(pool->lock) which may be released and regrabbed |
e22bee78 TH |
1945 | * multiple times. Does GFP_KERNEL allocations. Called only from |
1946 | * manager. | |
1947 | * | |
d185af30 | 1948 | * Return: |
c5aa87bb | 1949 | * %false if no action was taken and pool->lock stayed locked, %true |
e22bee78 TH |
1950 | * otherwise. |
1951 | */ | |
63d95a91 | 1952 | static bool maybe_create_worker(struct worker_pool *pool) |
d565ed63 TH |
1953 | __releases(&pool->lock) |
1954 | __acquires(&pool->lock) | |
1da177e4 | 1955 | { |
63d95a91 | 1956 | if (!need_to_create_worker(pool)) |
e22bee78 TH |
1957 | return false; |
1958 | restart: | |
d565ed63 | 1959 | spin_unlock_irq(&pool->lock); |
9f9c2364 | 1960 | |
e22bee78 | 1961 | /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */ |
63d95a91 | 1962 | mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT); |
e22bee78 TH |
1963 | |
1964 | while (true) { | |
1965 | struct worker *worker; | |
1966 | ||
bc2ae0f5 | 1967 | worker = create_worker(pool); |
e22bee78 | 1968 | if (worker) { |
63d95a91 | 1969 | del_timer_sync(&pool->mayday_timer); |
d565ed63 | 1970 | spin_lock_irq(&pool->lock); |
e22bee78 | 1971 | start_worker(worker); |
6183c009 TH |
1972 | if (WARN_ON_ONCE(need_to_create_worker(pool))) |
1973 | goto restart; | |
e22bee78 TH |
1974 | return true; |
1975 | } | |
1976 | ||
63d95a91 | 1977 | if (!need_to_create_worker(pool)) |
e22bee78 | 1978 | break; |
1da177e4 | 1979 | |
e22bee78 TH |
1980 | __set_current_state(TASK_INTERRUPTIBLE); |
1981 | schedule_timeout(CREATE_COOLDOWN); | |
9f9c2364 | 1982 | |
63d95a91 | 1983 | if (!need_to_create_worker(pool)) |
e22bee78 TH |
1984 | break; |
1985 | } | |
1986 | ||
63d95a91 | 1987 | del_timer_sync(&pool->mayday_timer); |
d565ed63 | 1988 | spin_lock_irq(&pool->lock); |
63d95a91 | 1989 | if (need_to_create_worker(pool)) |
e22bee78 TH |
1990 | goto restart; |
1991 | return true; | |
1992 | } | |
1993 | ||
73f53c4a | 1994 | /** |
e22bee78 TH |
1995 | * manage_workers - manage worker pool |
1996 | * @worker: self | |
73f53c4a | 1997 | * |
706026c2 | 1998 | * Assume the manager role and manage the worker pool @worker belongs |
e22bee78 | 1999 | * to. At any given time, there can be only zero or one manager per |
706026c2 | 2000 | * pool. The exclusion is handled automatically by this function. |
e22bee78 TH |
2001 | * |
2002 | * The caller can safely start processing works on false return. On | |
2003 | * true return, it's guaranteed that need_to_create_worker() is false | |
2004 | * and may_start_working() is true. | |
73f53c4a TH |
2005 | * |
2006 | * CONTEXT: | |
d565ed63 | 2007 | * spin_lock_irq(pool->lock) which may be released and regrabbed |
e22bee78 TH |
2008 | * multiple times. Does GFP_KERNEL allocations. |
2009 | * | |
d185af30 | 2010 | * Return: |
2d498db9 L |
2011 | * %false if the pool don't need management and the caller can safely start |
2012 | * processing works, %true indicates that the function released pool->lock | |
2013 | * and reacquired it to perform some management function and that the | |
2014 | * conditions that the caller verified while holding the lock before | |
2015 | * calling the function might no longer be true. | |
73f53c4a | 2016 | */ |
e22bee78 | 2017 | static bool manage_workers(struct worker *worker) |
73f53c4a | 2018 | { |
63d95a91 | 2019 | struct worker_pool *pool = worker->pool; |
e22bee78 | 2020 | bool ret = false; |
73f53c4a | 2021 | |
bc3a1afc TH |
2022 | /* |
2023 | * Managership is governed by two mutexes - manager_arb and | |
2024 | * manager_mutex. manager_arb handles arbitration of manager role. | |
2025 | * Anyone who successfully grabs manager_arb wins the arbitration | |
2026 | * and becomes the manager. mutex_trylock() on pool->manager_arb | |
2027 | * failure while holding pool->lock reliably indicates that someone | |
2028 | * else is managing the pool and the worker which failed trylock | |
2029 | * can proceed to executing work items. This means that anyone | |
2030 | * grabbing manager_arb is responsible for actually performing | |
2031 | * manager duties. If manager_arb is grabbed and released without | |
2032 | * actual management, the pool may stall indefinitely. | |
2033 | * | |
2034 | * manager_mutex is used for exclusion of actual management | |
2035 | * operations. The holder of manager_mutex can be sure that none | |
2036 | * of management operations, including creation and destruction of | |
2037 | * workers, won't take place until the mutex is released. Because | |
2038 | * manager_mutex doesn't interfere with manager role arbitration, | |
2039 | * it is guaranteed that the pool's management, while may be | |
2040 | * delayed, won't be disturbed by someone else grabbing | |
2041 | * manager_mutex. | |
2042 | */ | |
34a06bd6 | 2043 | if (!mutex_trylock(&pool->manager_arb)) |
e22bee78 | 2044 | return ret; |
1e19ffc6 | 2045 | |
ee378aa4 | 2046 | /* |
bc3a1afc TH |
2047 | * With manager arbitration won, manager_mutex would be free in |
2048 | * most cases. trylock first without dropping @pool->lock. | |
ee378aa4 | 2049 | */ |
bc3a1afc | 2050 | if (unlikely(!mutex_trylock(&pool->manager_mutex))) { |
d565ed63 | 2051 | spin_unlock_irq(&pool->lock); |
bc3a1afc | 2052 | mutex_lock(&pool->manager_mutex); |
8f174b11 | 2053 | spin_lock_irq(&pool->lock); |
ee378aa4 LJ |
2054 | ret = true; |
2055 | } | |
73f53c4a | 2056 | |
63d95a91 | 2057 | ret |= maybe_create_worker(pool); |
e22bee78 | 2058 | |
bc3a1afc | 2059 | mutex_unlock(&pool->manager_mutex); |
34a06bd6 | 2060 | mutex_unlock(&pool->manager_arb); |
e22bee78 | 2061 | return ret; |
73f53c4a TH |
2062 | } |
2063 | ||
a62428c0 TH |
2064 | /** |
2065 | * process_one_work - process single work | |
c34056a3 | 2066 | * @worker: self |
a62428c0 TH |
2067 | * @work: work to process |
2068 | * | |
2069 | * Process @work. This function contains all the logics necessary to | |
2070 | * process a single work including synchronization against and | |
2071 | * interaction with other workers on the same cpu, queueing and | |
2072 | * flushing. As long as context requirement is met, any worker can | |
2073 | * call this function to process a work. | |
2074 | * | |
2075 | * CONTEXT: | |
d565ed63 | 2076 | * spin_lock_irq(pool->lock) which is released and regrabbed. |
a62428c0 | 2077 | */ |
c34056a3 | 2078 | static void process_one_work(struct worker *worker, struct work_struct *work) |
d565ed63 TH |
2079 | __releases(&pool->lock) |
2080 | __acquires(&pool->lock) | |
a62428c0 | 2081 | { |
112202d9 | 2082 | struct pool_workqueue *pwq = get_work_pwq(work); |
bd7bdd43 | 2083 | struct worker_pool *pool = worker->pool; |
112202d9 | 2084 | bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE; |
73f53c4a | 2085 | int work_color; |
7e11629d | 2086 | struct worker *collision; |
a62428c0 TH |
2087 | #ifdef CONFIG_LOCKDEP |
2088 | /* | |
2089 | * It is permissible to free the struct work_struct from | |
2090 | * inside the function that is called from it, this we need to | |
2091 | * take into account for lockdep too. To avoid bogus "held | |
2092 | * lock freed" warnings as well as problems when looking into | |
2093 | * work->lockdep_map, make a copy and use that here. | |
2094 | */ | |
4d82a1de PZ |
2095 | struct lockdep_map lockdep_map; |
2096 | ||
2097 | lockdep_copy_map(&lockdep_map, &work->lockdep_map); | |
a62428c0 | 2098 | #endif |
6fec10a1 TH |
2099 | /* |
2100 | * Ensure we're on the correct CPU. DISASSOCIATED test is | |
2101 | * necessary to avoid spurious warnings from rescuers servicing the | |
24647570 | 2102 | * unbound or a disassociated pool. |
6fec10a1 | 2103 | */ |
5f7dabfd | 2104 | WARN_ON_ONCE(!(worker->flags & WORKER_UNBOUND) && |
24647570 | 2105 | !(pool->flags & POOL_DISASSOCIATED) && |
ec22ca5e | 2106 | raw_smp_processor_id() != pool->cpu); |
25511a47 | 2107 | |
7e11629d TH |
2108 | /* |
2109 | * A single work shouldn't be executed concurrently by | |
2110 | * multiple workers on a single cpu. Check whether anyone is | |
2111 | * already processing the work. If so, defer the work to the | |
2112 | * currently executing one. | |
2113 | */ | |
c9e7cf27 | 2114 | collision = find_worker_executing_work(pool, work); |
7e11629d TH |
2115 | if (unlikely(collision)) { |
2116 | move_linked_works(work, &collision->scheduled, NULL); | |
2117 | return; | |
2118 | } | |
2119 | ||
8930caba | 2120 | /* claim and dequeue */ |
a62428c0 | 2121 | debug_work_deactivate(work); |
c9e7cf27 | 2122 | hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work); |
c34056a3 | 2123 | worker->current_work = work; |
a2c1c57b | 2124 | worker->current_func = work->func; |
112202d9 | 2125 | worker->current_pwq = pwq; |
73f53c4a | 2126 | work_color = get_work_color(work); |
7a22ad75 | 2127 | |
a62428c0 TH |
2128 | list_del_init(&work->entry); |
2129 | ||
fb0e7beb TH |
2130 | /* |
2131 | * CPU intensive works don't participate in concurrency | |
2132 | * management. They're the scheduler's responsibility. | |
2133 | */ | |
2134 | if (unlikely(cpu_intensive)) | |
2135 | worker_set_flags(worker, WORKER_CPU_INTENSIVE, true); | |
2136 | ||
974271c4 | 2137 | /* |
d565ed63 | 2138 | * Unbound pool isn't concurrency managed and work items should be |
974271c4 TH |
2139 | * executed ASAP. Wake up another worker if necessary. |
2140 | */ | |
63d95a91 TH |
2141 | if ((worker->flags & WORKER_UNBOUND) && need_more_worker(pool)) |
2142 | wake_up_worker(pool); | |
974271c4 | 2143 | |
8930caba | 2144 | /* |
7c3eed5c | 2145 | * Record the last pool and clear PENDING which should be the last |
d565ed63 | 2146 | * update to @work. Also, do this inside @pool->lock so that |
23657bb1 TH |
2147 | * PENDING and queued state changes happen together while IRQ is |
2148 | * disabled. | |
8930caba | 2149 | */ |
7c3eed5c | 2150 | set_work_pool_and_clear_pending(work, pool->id); |
a62428c0 | 2151 | |
d565ed63 | 2152 | spin_unlock_irq(&pool->lock); |
a62428c0 | 2153 | |
112202d9 | 2154 | lock_map_acquire_read(&pwq->wq->lockdep_map); |
a62428c0 | 2155 | lock_map_acquire(&lockdep_map); |
e36c886a | 2156 | trace_workqueue_execute_start(work); |
a2c1c57b | 2157 | worker->current_func(work); |
e36c886a AV |
2158 | /* |
2159 | * While we must be careful to not use "work" after this, the trace | |
2160 | * point will only record its address. | |
2161 | */ | |
2162 | trace_workqueue_execute_end(work); | |
a62428c0 | 2163 | lock_map_release(&lockdep_map); |
112202d9 | 2164 | lock_map_release(&pwq->wq->lockdep_map); |
a62428c0 TH |
2165 | |
2166 | if (unlikely(in_atomic() || lockdep_depth(current) > 0)) { | |
044c782c VI |
2167 | pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n" |
2168 | " last function: %pf\n", | |
a2c1c57b TH |
2169 | current->comm, preempt_count(), task_pid_nr(current), |
2170 | worker->current_func); | |
a62428c0 TH |
2171 | debug_show_held_locks(current); |
2172 | dump_stack(); | |
2173 | } | |
2174 | ||
b22ce278 TH |
2175 | /* |
2176 | * The following prevents a kworker from hogging CPU on !PREEMPT | |
2177 | * kernels, where a requeueing work item waiting for something to | |
2178 | * happen could deadlock with stop_machine as such work item could | |
2179 | * indefinitely requeue itself while all other CPUs are trapped in | |
2180 | * stop_machine. | |
2181 | */ | |
2182 | cond_resched(); | |
2183 | ||
d565ed63 | 2184 | spin_lock_irq(&pool->lock); |
a62428c0 | 2185 | |
fb0e7beb TH |
2186 | /* clear cpu intensive status */ |
2187 | if (unlikely(cpu_intensive)) | |
2188 | worker_clr_flags(worker, WORKER_CPU_INTENSIVE); | |
2189 | ||
a62428c0 | 2190 | /* we're done with it, release */ |
42f8570f | 2191 | hash_del(&worker->hentry); |
c34056a3 | 2192 | worker->current_work = NULL; |
a2c1c57b | 2193 | worker->current_func = NULL; |
112202d9 | 2194 | worker->current_pwq = NULL; |
3d1cb205 | 2195 | worker->desc_valid = false; |
112202d9 | 2196 | pwq_dec_nr_in_flight(pwq, work_color); |
a62428c0 TH |
2197 | } |
2198 | ||
affee4b2 TH |
2199 | /** |
2200 | * process_scheduled_works - process scheduled works | |
2201 | * @worker: self | |
2202 | * | |
2203 | * Process all scheduled works. Please note that the scheduled list | |
2204 | * may change while processing a work, so this function repeatedly | |
2205 | * fetches a work from the top and executes it. | |
2206 | * | |
2207 | * CONTEXT: | |
d565ed63 | 2208 | * spin_lock_irq(pool->lock) which may be released and regrabbed |
affee4b2 TH |
2209 | * multiple times. |
2210 | */ | |
2211 | static void process_scheduled_works(struct worker *worker) | |
1da177e4 | 2212 | { |
affee4b2 TH |
2213 | while (!list_empty(&worker->scheduled)) { |
2214 | struct work_struct *work = list_first_entry(&worker->scheduled, | |
1da177e4 | 2215 | struct work_struct, entry); |
c34056a3 | 2216 | process_one_work(worker, work); |
1da177e4 | 2217 | } |
1da177e4 LT |
2218 | } |
2219 | ||
4690c4ab TH |
2220 | /** |
2221 | * worker_thread - the worker thread function | |
c34056a3 | 2222 | * @__worker: self |
4690c4ab | 2223 | * |
c5aa87bb TH |
2224 | * The worker thread function. All workers belong to a worker_pool - |
2225 | * either a per-cpu one or dynamic unbound one. These workers process all | |
2226 | * work items regardless of their specific target workqueue. The only | |
2227 | * exception is work items which belong to workqueues with a rescuer which | |
2228 | * will be explained in rescuer_thread(). | |
d185af30 YB |
2229 | * |
2230 | * Return: 0 | |
4690c4ab | 2231 | */ |
c34056a3 | 2232 | static int worker_thread(void *__worker) |
1da177e4 | 2233 | { |
c34056a3 | 2234 | struct worker *worker = __worker; |
bd7bdd43 | 2235 | struct worker_pool *pool = worker->pool; |
1da177e4 | 2236 | |
e22bee78 TH |
2237 | /* tell the scheduler that this is a workqueue worker */ |
2238 | worker->task->flags |= PF_WQ_WORKER; | |
c8e55f36 | 2239 | woke_up: |
d565ed63 | 2240 | spin_lock_irq(&pool->lock); |
1da177e4 | 2241 | |
a9ab775b TH |
2242 | /* am I supposed to die? */ |
2243 | if (unlikely(worker->flags & WORKER_DIE)) { | |
d565ed63 | 2244 | spin_unlock_irq(&pool->lock); |
a9ab775b TH |
2245 | WARN_ON_ONCE(!list_empty(&worker->entry)); |
2246 | worker->task->flags &= ~PF_WQ_WORKER; | |
60f5a4bc LJ |
2247 | |
2248 | set_task_comm(worker->task, "kworker/dying"); | |
2249 | worker_detach_from_pool(worker, pool); | |
2250 | kfree(worker); | |
a9ab775b | 2251 | return 0; |
c8e55f36 | 2252 | } |
affee4b2 | 2253 | |
c8e55f36 | 2254 | worker_leave_idle(worker); |
db7bccf4 | 2255 | recheck: |
e22bee78 | 2256 | /* no more worker necessary? */ |
63d95a91 | 2257 | if (!need_more_worker(pool)) |
e22bee78 TH |
2258 | goto sleep; |
2259 | ||
2260 | /* do we need to manage? */ | |
63d95a91 | 2261 | if (unlikely(!may_start_working(pool)) && manage_workers(worker)) |
e22bee78 TH |
2262 | goto recheck; |
2263 | ||
c8e55f36 TH |
2264 | /* |
2265 | * ->scheduled list can only be filled while a worker is | |
2266 | * preparing to process a work or actually processing it. | |
2267 | * Make sure nobody diddled with it while I was sleeping. | |
2268 | */ | |
6183c009 | 2269 | WARN_ON_ONCE(!list_empty(&worker->scheduled)); |
c8e55f36 | 2270 | |
e22bee78 | 2271 | /* |
a9ab775b TH |
2272 | * Finish PREP stage. We're guaranteed to have at least one idle |
2273 | * worker or that someone else has already assumed the manager | |
2274 | * role. This is where @worker starts participating in concurrency | |
2275 | * management if applicable and concurrency management is restored | |
2276 | * after being rebound. See rebind_workers() for details. | |
e22bee78 | 2277 | */ |
a9ab775b | 2278 | worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND); |
e22bee78 TH |
2279 | |
2280 | do { | |
c8e55f36 | 2281 | struct work_struct *work = |
bd7bdd43 | 2282 | list_first_entry(&pool->worklist, |
c8e55f36 TH |
2283 | struct work_struct, entry); |
2284 | ||
2285 | if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) { | |
2286 | /* optimization path, not strictly necessary */ | |
2287 | process_one_work(worker, work); | |
2288 | if (unlikely(!list_empty(&worker->scheduled))) | |
affee4b2 | 2289 | process_scheduled_works(worker); |
c8e55f36 TH |
2290 | } else { |
2291 | move_linked_works(work, &worker->scheduled, NULL); | |
2292 | process_scheduled_works(worker); | |
affee4b2 | 2293 | } |
63d95a91 | 2294 | } while (keep_working(pool)); |
e22bee78 TH |
2295 | |
2296 | worker_set_flags(worker, WORKER_PREP, false); | |
d313dd85 | 2297 | sleep: |
c8e55f36 | 2298 | /* |
d565ed63 TH |
2299 | * pool->lock is held and there's no work to process and no need to |
2300 | * manage, sleep. Workers are woken up only while holding | |
2301 | * pool->lock or from local cpu, so setting the current state | |
2302 | * before releasing pool->lock is enough to prevent losing any | |
2303 | * event. | |
c8e55f36 TH |
2304 | */ |
2305 | worker_enter_idle(worker); | |
2306 | __set_current_state(TASK_INTERRUPTIBLE); | |
d565ed63 | 2307 | spin_unlock_irq(&pool->lock); |
c8e55f36 TH |
2308 | schedule(); |
2309 | goto woke_up; | |
1da177e4 LT |
2310 | } |
2311 | ||
e22bee78 TH |
2312 | /** |
2313 | * rescuer_thread - the rescuer thread function | |
111c225a | 2314 | * @__rescuer: self |
e22bee78 TH |
2315 | * |
2316 | * Workqueue rescuer thread function. There's one rescuer for each | |
493008a8 | 2317 | * workqueue which has WQ_MEM_RECLAIM set. |
e22bee78 | 2318 | * |
706026c2 | 2319 | * Regular work processing on a pool may block trying to create a new |
e22bee78 TH |
2320 | * worker which uses GFP_KERNEL allocation which has slight chance of |
2321 | * developing into deadlock if some works currently on the same queue | |
2322 | * need to be processed to satisfy the GFP_KERNEL allocation. This is | |
2323 | * the problem rescuer solves. | |
2324 | * | |
706026c2 TH |
2325 | * When such condition is possible, the pool summons rescuers of all |
2326 | * workqueues which have works queued on the pool and let them process | |
e22bee78 TH |
2327 | * those works so that forward progress can be guaranteed. |
2328 | * | |
2329 | * This should happen rarely. | |
d185af30 YB |
2330 | * |
2331 | * Return: 0 | |
e22bee78 | 2332 | */ |
111c225a | 2333 | static int rescuer_thread(void *__rescuer) |
e22bee78 | 2334 | { |
111c225a TH |
2335 | struct worker *rescuer = __rescuer; |
2336 | struct workqueue_struct *wq = rescuer->rescue_wq; | |
e22bee78 | 2337 | struct list_head *scheduled = &rescuer->scheduled; |
4d595b86 | 2338 | bool should_stop; |
e22bee78 TH |
2339 | |
2340 | set_user_nice(current, RESCUER_NICE_LEVEL); | |
111c225a TH |
2341 | |
2342 | /* | |
2343 | * Mark rescuer as worker too. As WORKER_PREP is never cleared, it | |
2344 | * doesn't participate in concurrency management. | |
2345 | */ | |
2346 | rescuer->task->flags |= PF_WQ_WORKER; | |
e22bee78 TH |
2347 | repeat: |
2348 | set_current_state(TASK_INTERRUPTIBLE); | |
2349 | ||
4d595b86 LJ |
2350 | /* |
2351 | * By the time the rescuer is requested to stop, the workqueue | |
2352 | * shouldn't have any work pending, but @wq->maydays may still have | |
2353 | * pwq(s) queued. This can happen by non-rescuer workers consuming | |
2354 | * all the work items before the rescuer got to them. Go through | |
2355 | * @wq->maydays processing before acting on should_stop so that the | |
2356 | * list is always empty on exit. | |
2357 | */ | |
2358 | should_stop = kthread_should_stop(); | |
e22bee78 | 2359 | |
493a1724 | 2360 | /* see whether any pwq is asking for help */ |
2e109a28 | 2361 | spin_lock_irq(&wq_mayday_lock); |
493a1724 TH |
2362 | |
2363 | while (!list_empty(&wq->maydays)) { | |
2364 | struct pool_workqueue *pwq = list_first_entry(&wq->maydays, | |
2365 | struct pool_workqueue, mayday_node); | |
112202d9 | 2366 | struct worker_pool *pool = pwq->pool; |
e22bee78 TH |
2367 | struct work_struct *work, *n; |
2368 | ||
2369 | __set_current_state(TASK_RUNNING); | |
493a1724 TH |
2370 | list_del_init(&pwq->mayday_node); |
2371 | ||
2e109a28 | 2372 | spin_unlock_irq(&wq_mayday_lock); |
e22bee78 TH |
2373 | |
2374 | /* migrate to the target cpu if possible */ | |
f36dc67b | 2375 | worker_maybe_bind_and_lock(pool); |
b3104104 | 2376 | rescuer->pool = pool; |
e22bee78 TH |
2377 | |
2378 | /* | |
2379 | * Slurp in all works issued via this workqueue and | |
2380 | * process'em. | |
2381 | */ | |
6183c009 | 2382 | WARN_ON_ONCE(!list_empty(&rescuer->scheduled)); |
bd7bdd43 | 2383 | list_for_each_entry_safe(work, n, &pool->worklist, entry) |
112202d9 | 2384 | if (get_work_pwq(work) == pwq) |
e22bee78 TH |
2385 | move_linked_works(work, scheduled, &n); |
2386 | ||
2387 | process_scheduled_works(rescuer); | |
7576958a | 2388 | |
77668c8b LJ |
2389 | /* |
2390 | * Put the reference grabbed by send_mayday(). @pool won't | |
2391 | * go away while we're holding its lock. | |
2392 | */ | |
2393 | put_pwq(pwq); | |
2394 | ||
7576958a | 2395 | /* |
d565ed63 | 2396 | * Leave this pool. If keep_working() is %true, notify a |
7576958a TH |
2397 | * regular worker; otherwise, we end up with 0 concurrency |
2398 | * and stalling the execution. | |
2399 | */ | |
63d95a91 TH |
2400 | if (keep_working(pool)) |
2401 | wake_up_worker(pool); | |
7576958a | 2402 | |
b3104104 | 2403 | rescuer->pool = NULL; |
493a1724 | 2404 | spin_unlock(&pool->lock); |
2e109a28 | 2405 | spin_lock(&wq_mayday_lock); |
e22bee78 TH |
2406 | } |
2407 | ||
2e109a28 | 2408 | spin_unlock_irq(&wq_mayday_lock); |
493a1724 | 2409 | |
4d595b86 LJ |
2410 | if (should_stop) { |
2411 | __set_current_state(TASK_RUNNING); | |
2412 | rescuer->task->flags &= ~PF_WQ_WORKER; | |
2413 | return 0; | |
2414 | } | |
2415 | ||
111c225a TH |
2416 | /* rescuers should never participate in concurrency management */ |
2417 | WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING)); | |
e22bee78 TH |
2418 | schedule(); |
2419 | goto repeat; | |
1da177e4 LT |
2420 | } |
2421 | ||
fc2e4d70 ON |
2422 | struct wq_barrier { |
2423 | struct work_struct work; | |
2424 | struct completion done; | |
2425 | }; | |
2426 | ||
2427 | static void wq_barrier_func(struct work_struct *work) | |
2428 | { | |
2429 | struct wq_barrier *barr = container_of(work, struct wq_barrier, work); | |
2430 | complete(&barr->done); | |
2431 | } | |
2432 | ||
4690c4ab TH |
2433 | /** |
2434 | * insert_wq_barrier - insert a barrier work | |
112202d9 | 2435 | * @pwq: pwq to insert barrier into |
4690c4ab | 2436 | * @barr: wq_barrier to insert |
affee4b2 TH |
2437 | * @target: target work to attach @barr to |
2438 | * @worker: worker currently executing @target, NULL if @target is not executing | |
4690c4ab | 2439 | * |
affee4b2 TH |
2440 | * @barr is linked to @target such that @barr is completed only after |
2441 | * @target finishes execution. Please note that the ordering | |
2442 | * guarantee is observed only with respect to @target and on the local | |
2443 | * cpu. | |
2444 | * | |
2445 | * Currently, a queued barrier can't be canceled. This is because | |
2446 | * try_to_grab_pending() can't determine whether the work to be | |
2447 | * grabbed is at the head of the queue and thus can't clear LINKED | |
2448 | * flag of the previous work while there must be a valid next work | |
2449 | * after a work with LINKED flag set. | |
2450 | * | |
2451 | * Note that when @worker is non-NULL, @target may be modified | |
112202d9 | 2452 | * underneath us, so we can't reliably determine pwq from @target. |
4690c4ab TH |
2453 | * |
2454 | * CONTEXT: | |
d565ed63 | 2455 | * spin_lock_irq(pool->lock). |
4690c4ab | 2456 | */ |
112202d9 | 2457 | static void insert_wq_barrier(struct pool_workqueue *pwq, |
affee4b2 TH |
2458 | struct wq_barrier *barr, |
2459 | struct work_struct *target, struct worker *worker) | |
fc2e4d70 | 2460 | { |
affee4b2 TH |
2461 | struct list_head *head; |
2462 | unsigned int linked = 0; | |
2463 | ||
dc186ad7 | 2464 | /* |
d565ed63 | 2465 | * debugobject calls are safe here even with pool->lock locked |
dc186ad7 TG |
2466 | * as we know for sure that this will not trigger any of the |
2467 | * checks and call back into the fixup functions where we | |
2468 | * might deadlock. | |
2469 | */ | |
ca1cab37 | 2470 | INIT_WORK_ONSTACK(&barr->work, wq_barrier_func); |
22df02bb | 2471 | __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work)); |
fc2e4d70 | 2472 | init_completion(&barr->done); |
83c22520 | 2473 | |
affee4b2 TH |
2474 | /* |
2475 | * If @target is currently being executed, schedule the | |
2476 | * barrier to the worker; otherwise, put it after @target. | |
2477 | */ | |
2478 | if (worker) | |
2479 | head = worker->scheduled.next; | |
2480 | else { | |
2481 | unsigned long *bits = work_data_bits(target); | |
2482 | ||
2483 | head = target->entry.next; | |
2484 | /* there can already be other linked works, inherit and set */ | |
2485 | linked = *bits & WORK_STRUCT_LINKED; | |
2486 | __set_bit(WORK_STRUCT_LINKED_BIT, bits); | |
2487 | } | |
2488 | ||
dc186ad7 | 2489 | debug_work_activate(&barr->work); |
112202d9 | 2490 | insert_work(pwq, &barr->work, head, |
affee4b2 | 2491 | work_color_to_flags(WORK_NO_COLOR) | linked); |
fc2e4d70 ON |
2492 | } |
2493 | ||
73f53c4a | 2494 | /** |
112202d9 | 2495 | * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing |
73f53c4a TH |
2496 | * @wq: workqueue being flushed |
2497 | * @flush_color: new flush color, < 0 for no-op | |
2498 | * @work_color: new work color, < 0 for no-op | |
2499 | * | |
112202d9 | 2500 | * Prepare pwqs for workqueue flushing. |
73f53c4a | 2501 | * |
112202d9 TH |
2502 | * If @flush_color is non-negative, flush_color on all pwqs should be |
2503 | * -1. If no pwq has in-flight commands at the specified color, all | |
2504 | * pwq->flush_color's stay at -1 and %false is returned. If any pwq | |
2505 | * has in flight commands, its pwq->flush_color is set to | |
2506 | * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq | |
73f53c4a TH |
2507 | * wakeup logic is armed and %true is returned. |
2508 | * | |
2509 | * The caller should have initialized @wq->first_flusher prior to | |
2510 | * calling this function with non-negative @flush_color. If | |
2511 | * @flush_color is negative, no flush color update is done and %false | |
2512 | * is returned. | |
2513 | * | |
112202d9 | 2514 | * If @work_color is non-negative, all pwqs should have the same |
73f53c4a TH |
2515 | * work_color which is previous to @work_color and all will be |
2516 | * advanced to @work_color. | |
2517 | * | |
2518 | * CONTEXT: | |
3c25a55d | 2519 | * mutex_lock(wq->mutex). |
73f53c4a | 2520 | * |
d185af30 | 2521 | * Return: |
73f53c4a TH |
2522 | * %true if @flush_color >= 0 and there's something to flush. %false |
2523 | * otherwise. | |
2524 | */ | |
112202d9 | 2525 | static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq, |
73f53c4a | 2526 | int flush_color, int work_color) |
1da177e4 | 2527 | { |
73f53c4a | 2528 | bool wait = false; |
49e3cf44 | 2529 | struct pool_workqueue *pwq; |
1da177e4 | 2530 | |
73f53c4a | 2531 | if (flush_color >= 0) { |
6183c009 | 2532 | WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush)); |
112202d9 | 2533 | atomic_set(&wq->nr_pwqs_to_flush, 1); |
1da177e4 | 2534 | } |
2355b70f | 2535 | |
49e3cf44 | 2536 | for_each_pwq(pwq, wq) { |
112202d9 | 2537 | struct worker_pool *pool = pwq->pool; |
fc2e4d70 | 2538 | |
b09f4fd3 | 2539 | spin_lock_irq(&pool->lock); |
83c22520 | 2540 | |
73f53c4a | 2541 | if (flush_color >= 0) { |
6183c009 | 2542 | WARN_ON_ONCE(pwq->flush_color != -1); |
fc2e4d70 | 2543 | |
112202d9 TH |
2544 | if (pwq->nr_in_flight[flush_color]) { |
2545 | pwq->flush_color = flush_color; | |
2546 | atomic_inc(&wq->nr_pwqs_to_flush); | |
73f53c4a TH |
2547 | wait = true; |
2548 | } | |
2549 | } | |
1da177e4 | 2550 | |
73f53c4a | 2551 | if (work_color >= 0) { |
6183c009 | 2552 | WARN_ON_ONCE(work_color != work_next_color(pwq->work_color)); |
112202d9 | 2553 | pwq->work_color = work_color; |
73f53c4a | 2554 | } |
1da177e4 | 2555 | |
b09f4fd3 | 2556 | spin_unlock_irq(&pool->lock); |
1da177e4 | 2557 | } |
2355b70f | 2558 | |
112202d9 | 2559 | if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush)) |
73f53c4a | 2560 | complete(&wq->first_flusher->done); |
14441960 | 2561 | |
73f53c4a | 2562 | return wait; |
1da177e4 LT |
2563 | } |
2564 | ||
0fcb78c2 | 2565 | /** |
1da177e4 | 2566 | * flush_workqueue - ensure that any scheduled work has run to completion. |
0fcb78c2 | 2567 | * @wq: workqueue to flush |
1da177e4 | 2568 | * |
c5aa87bb TH |
2569 | * This function sleeps until all work items which were queued on entry |
2570 | * have finished execution, but it is not livelocked by new incoming ones. | |
1da177e4 | 2571 | */ |
7ad5b3a5 | 2572 | void flush_workqueue(struct workqueue_struct *wq) |
1da177e4 | 2573 | { |
73f53c4a TH |
2574 | struct wq_flusher this_flusher = { |
2575 | .list = LIST_HEAD_INIT(this_flusher.list), | |
2576 | .flush_color = -1, | |
2577 | .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done), | |
2578 | }; | |
2579 | int next_color; | |
1da177e4 | 2580 | |
3295f0ef IM |
2581 | lock_map_acquire(&wq->lockdep_map); |
2582 | lock_map_release(&wq->lockdep_map); | |
73f53c4a | 2583 | |
3c25a55d | 2584 | mutex_lock(&wq->mutex); |
73f53c4a TH |
2585 | |
2586 | /* | |
2587 | * Start-to-wait phase | |
2588 | */ | |
2589 | next_color = work_next_color(wq->work_color); | |
2590 | ||
2591 | if (next_color != wq->flush_color) { | |
2592 | /* | |
2593 | * Color space is not full. The current work_color | |
2594 | * becomes our flush_color and work_color is advanced | |
2595 | * by one. | |
2596 | */ | |
6183c009 | 2597 | WARN_ON_ONCE(!list_empty(&wq->flusher_overflow)); |
73f53c4a TH |
2598 | this_flusher.flush_color = wq->work_color; |
2599 | wq->work_color = next_color; | |
2600 | ||
2601 | if (!wq->first_flusher) { | |
2602 | /* no flush in progress, become the first flusher */ | |
6183c009 | 2603 | WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); |
73f53c4a TH |
2604 | |
2605 | wq->first_flusher = &this_flusher; | |
2606 | ||
112202d9 | 2607 | if (!flush_workqueue_prep_pwqs(wq, wq->flush_color, |
73f53c4a TH |
2608 | wq->work_color)) { |
2609 | /* nothing to flush, done */ | |
2610 | wq->flush_color = next_color; | |
2611 | wq->first_flusher = NULL; | |
2612 | goto out_unlock; | |
2613 | } | |
2614 | } else { | |
2615 | /* wait in queue */ | |
6183c009 | 2616 | WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color); |
73f53c4a | 2617 | list_add_tail(&this_flusher.list, &wq->flusher_queue); |
112202d9 | 2618 | flush_workqueue_prep_pwqs(wq, -1, wq->work_color); |
73f53c4a TH |
2619 | } |
2620 | } else { | |
2621 | /* | |
2622 | * Oops, color space is full, wait on overflow queue. | |
2623 | * The next flush completion will assign us | |
2624 | * flush_color and transfer to flusher_queue. | |
2625 | */ | |
2626 | list_add_tail(&this_flusher.list, &wq->flusher_overflow); | |
2627 | } | |
2628 | ||
3c25a55d | 2629 | mutex_unlock(&wq->mutex); |
73f53c4a TH |
2630 | |
2631 | wait_for_completion(&this_flusher.done); | |
2632 | ||
2633 | /* | |
2634 | * Wake-up-and-cascade phase | |
2635 | * | |
2636 | * First flushers are responsible for cascading flushes and | |
2637 | * handling overflow. Non-first flushers can simply return. | |
2638 | */ | |
2639 | if (wq->first_flusher != &this_flusher) | |
2640 | return; | |
2641 | ||
3c25a55d | 2642 | mutex_lock(&wq->mutex); |
73f53c4a | 2643 | |
4ce48b37 TH |
2644 | /* we might have raced, check again with mutex held */ |
2645 | if (wq->first_flusher != &this_flusher) | |
2646 | goto out_unlock; | |
2647 | ||
73f53c4a TH |
2648 | wq->first_flusher = NULL; |
2649 | ||
6183c009 TH |
2650 | WARN_ON_ONCE(!list_empty(&this_flusher.list)); |
2651 | WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); | |
73f53c4a TH |
2652 | |
2653 | while (true) { | |
2654 | struct wq_flusher *next, *tmp; | |
2655 | ||
2656 | /* complete all the flushers sharing the current flush color */ | |
2657 | list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) { | |
2658 | if (next->flush_color != wq->flush_color) | |
2659 | break; | |
2660 | list_del_init(&next->list); | |
2661 | complete(&next->done); | |
2662 | } | |
2663 | ||
6183c009 TH |
2664 | WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) && |
2665 | wq->flush_color != work_next_color(wq->work_color)); | |
73f53c4a TH |
2666 | |
2667 | /* this flush_color is finished, advance by one */ | |
2668 | wq->flush_color = work_next_color(wq->flush_color); | |
2669 | ||
2670 | /* one color has been freed, handle overflow queue */ | |
2671 | if (!list_empty(&wq->flusher_overflow)) { | |
2672 | /* | |
2673 | * Assign the same color to all overflowed | |
2674 | * flushers, advance work_color and append to | |
2675 | * flusher_queue. This is the start-to-wait | |
2676 | * phase for these overflowed flushers. | |
2677 | */ | |
2678 | list_for_each_entry(tmp, &wq->flusher_overflow, list) | |
2679 | tmp->flush_color = wq->work_color; | |
2680 | ||
2681 | wq->work_color = work_next_color(wq->work_color); | |
2682 | ||
2683 | list_splice_tail_init(&wq->flusher_overflow, | |
2684 | &wq->flusher_queue); | |
112202d9 | 2685 | flush_workqueue_prep_pwqs(wq, -1, wq->work_color); |
73f53c4a TH |
2686 | } |
2687 | ||
2688 | if (list_empty(&wq->flusher_queue)) { | |
6183c009 | 2689 | WARN_ON_ONCE(wq->flush_color != wq->work_color); |
73f53c4a TH |
2690 | break; |
2691 | } | |
2692 | ||
2693 | /* | |
2694 | * Need to flush more colors. Make the next flusher | |
112202d9 | 2695 | * the new first flusher and arm pwqs. |
73f53c4a | 2696 | */ |
6183c009 TH |
2697 | WARN_ON_ONCE(wq->flush_color == wq->work_color); |
2698 | WARN_ON_ONCE(wq->flush_color != next->flush_color); | |
73f53c4a TH |
2699 | |
2700 | list_del_init(&next->list); | |
2701 | wq->first_flusher = next; | |
2702 | ||
112202d9 | 2703 | if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1)) |
73f53c4a TH |
2704 | break; |
2705 | ||
2706 | /* | |
2707 | * Meh... this color is already done, clear first | |
2708 | * flusher and repeat cascading. | |
2709 | */ | |
2710 | wq->first_flusher = NULL; | |
2711 | } | |
2712 | ||
2713 | out_unlock: | |
3c25a55d | 2714 | mutex_unlock(&wq->mutex); |
1da177e4 | 2715 | } |
ae90dd5d | 2716 | EXPORT_SYMBOL_GPL(flush_workqueue); |
1da177e4 | 2717 | |
9c5a2ba7 TH |
2718 | /** |
2719 | * drain_workqueue - drain a workqueue | |
2720 | * @wq: workqueue to drain | |
2721 | * | |
2722 | * Wait until the workqueue becomes empty. While draining is in progress, | |
2723 | * only chain queueing is allowed. IOW, only currently pending or running | |
2724 | * work items on @wq can queue further work items on it. @wq is flushed | |
2725 | * repeatedly until it becomes empty. The number of flushing is detemined | |
2726 | * by the depth of chaining and should be relatively short. Whine if it | |
2727 | * takes too long. | |
2728 | */ | |
2729 | void drain_workqueue(struct workqueue_struct *wq) | |
2730 | { | |
2731 | unsigned int flush_cnt = 0; | |
49e3cf44 | 2732 | struct pool_workqueue *pwq; |
9c5a2ba7 TH |
2733 | |
2734 | /* | |
2735 | * __queue_work() needs to test whether there are drainers, is much | |
2736 | * hotter than drain_workqueue() and already looks at @wq->flags. | |
618b01eb | 2737 | * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers. |
9c5a2ba7 | 2738 | */ |
87fc741e | 2739 | mutex_lock(&wq->mutex); |
9c5a2ba7 | 2740 | if (!wq->nr_drainers++) |
618b01eb | 2741 | wq->flags |= __WQ_DRAINING; |
87fc741e | 2742 | mutex_unlock(&wq->mutex); |
9c5a2ba7 TH |
2743 | reflush: |
2744 | flush_workqueue(wq); | |
2745 | ||
b09f4fd3 | 2746 | mutex_lock(&wq->mutex); |
76af4d93 | 2747 | |
49e3cf44 | 2748 | for_each_pwq(pwq, wq) { |
fa2563e4 | 2749 | bool drained; |
9c5a2ba7 | 2750 | |
b09f4fd3 | 2751 | spin_lock_irq(&pwq->pool->lock); |
112202d9 | 2752 | drained = !pwq->nr_active && list_empty(&pwq->delayed_works); |
b09f4fd3 | 2753 | spin_unlock_irq(&pwq->pool->lock); |
fa2563e4 TT |
2754 | |
2755 | if (drained) | |
9c5a2ba7 TH |
2756 | continue; |
2757 | ||
2758 | if (++flush_cnt == 10 || | |
2759 | (flush_cnt % 100 == 0 && flush_cnt <= 1000)) | |
c5aa87bb | 2760 | pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n", |
044c782c | 2761 | wq->name, flush_cnt); |
76af4d93 | 2762 | |
b09f4fd3 | 2763 | mutex_unlock(&wq->mutex); |
9c5a2ba7 TH |
2764 | goto reflush; |
2765 | } | |
2766 | ||
9c5a2ba7 | 2767 | if (!--wq->nr_drainers) |
618b01eb | 2768 | wq->flags &= ~__WQ_DRAINING; |
87fc741e | 2769 | mutex_unlock(&wq->mutex); |
9c5a2ba7 TH |
2770 | } |
2771 | EXPORT_SYMBOL_GPL(drain_workqueue); | |
2772 | ||
606a5020 | 2773 | static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr) |
db700897 | 2774 | { |
affee4b2 | 2775 | struct worker *worker = NULL; |
c9e7cf27 | 2776 | struct worker_pool *pool; |
112202d9 | 2777 | struct pool_workqueue *pwq; |
db700897 ON |
2778 | |
2779 | might_sleep(); | |
fa1b54e6 TH |
2780 | |
2781 | local_irq_disable(); | |
c9e7cf27 | 2782 | pool = get_work_pool(work); |
fa1b54e6 TH |
2783 | if (!pool) { |
2784 | local_irq_enable(); | |
baf59022 | 2785 | return false; |
fa1b54e6 | 2786 | } |
db700897 | 2787 | |
fa1b54e6 | 2788 | spin_lock(&pool->lock); |
0b3dae68 | 2789 | /* see the comment in try_to_grab_pending() with the same code */ |
112202d9 TH |
2790 | pwq = get_work_pwq(work); |
2791 | if (pwq) { | |
2792 | if (unlikely(pwq->pool != pool)) | |
4690c4ab | 2793 | goto already_gone; |
606a5020 | 2794 | } else { |
c9e7cf27 | 2795 | worker = find_worker_executing_work(pool, work); |
affee4b2 | 2796 | if (!worker) |
4690c4ab | 2797 | goto already_gone; |
112202d9 | 2798 | pwq = worker->current_pwq; |
606a5020 | 2799 | } |
db700897 | 2800 | |
112202d9 | 2801 | insert_wq_barrier(pwq, barr, work, worker); |
d565ed63 | 2802 | spin_unlock_irq(&pool->lock); |
7a22ad75 | 2803 | |
e159489b TH |
2804 | /* |
2805 | * If @max_active is 1 or rescuer is in use, flushing another work | |
2806 | * item on the same workqueue may lead to deadlock. Make sure the | |
2807 | * flusher is not running on the same workqueue by verifying write | |
2808 | * access. | |
2809 | */ | |
493008a8 | 2810 | if (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer) |
112202d9 | 2811 | lock_map_acquire(&pwq->wq->lockdep_map); |
e159489b | 2812 | else |
112202d9 TH |
2813 | lock_map_acquire_read(&pwq->wq->lockdep_map); |
2814 | lock_map_release(&pwq->wq->lockdep_map); | |
e159489b | 2815 | |
401a8d04 | 2816 | return true; |
4690c4ab | 2817 | already_gone: |
d565ed63 | 2818 | spin_unlock_irq(&pool->lock); |
401a8d04 | 2819 | return false; |
db700897 | 2820 | } |
baf59022 TH |
2821 | |
2822 | /** | |
2823 | * flush_work - wait for a work to finish executing the last queueing instance | |
2824 | * @work: the work to flush | |
2825 | * | |
606a5020 TH |
2826 | * Wait until @work has finished execution. @work is guaranteed to be idle |
2827 | * on return if it hasn't been requeued since flush started. | |
baf59022 | 2828 | * |
d185af30 | 2829 | * Return: |
baf59022 TH |
2830 | * %true if flush_work() waited for the work to finish execution, |
2831 | * %false if it was already idle. | |
2832 | */ | |
2833 | bool flush_work(struct work_struct *work) | |
2834 | { | |
12997d1a BH |
2835 | struct wq_barrier barr; |
2836 | ||
0976dfc1 SB |
2837 | lock_map_acquire(&work->lockdep_map); |
2838 | lock_map_release(&work->lockdep_map); | |
2839 | ||
12997d1a BH |
2840 | if (start_flush_work(work, &barr)) { |
2841 | wait_for_completion(&barr.done); | |
2842 | destroy_work_on_stack(&barr.work); | |
2843 | return true; | |
2844 | } else { | |
2845 | return false; | |
2846 | } | |
6e84d644 | 2847 | } |
606a5020 | 2848 | EXPORT_SYMBOL_GPL(flush_work); |
6e84d644 | 2849 | |
36e227d2 | 2850 | static bool __cancel_work_timer(struct work_struct *work, bool is_dwork) |
1f1f642e | 2851 | { |
bbb68dfa | 2852 | unsigned long flags; |
1f1f642e ON |
2853 | int ret; |
2854 | ||
2855 | do { | |
bbb68dfa TH |
2856 | ret = try_to_grab_pending(work, is_dwork, &flags); |
2857 | /* | |
2858 | * If someone else is canceling, wait for the same event it | |
2859 | * would be waiting for before retrying. | |
2860 | */ | |
2861 | if (unlikely(ret == -ENOENT)) | |
606a5020 | 2862 | flush_work(work); |
1f1f642e ON |
2863 | } while (unlikely(ret < 0)); |
2864 | ||
bbb68dfa TH |
2865 | /* tell other tasks trying to grab @work to back off */ |
2866 | mark_work_canceling(work); | |
2867 | local_irq_restore(flags); | |
2868 | ||
606a5020 | 2869 | flush_work(work); |
7a22ad75 | 2870 | clear_work_data(work); |
1f1f642e ON |
2871 | return ret; |
2872 | } | |
2873 | ||
6e84d644 | 2874 | /** |
401a8d04 TH |
2875 | * cancel_work_sync - cancel a work and wait for it to finish |
2876 | * @work: the work to cancel | |
6e84d644 | 2877 | * |
401a8d04 TH |
2878 | * Cancel @work and wait for its execution to finish. This function |
2879 | * can be used even if the work re-queues itself or migrates to | |
2880 | * another workqueue. On return from this function, @work is | |
2881 | * guaranteed to be not pending or executing on any CPU. | |
1f1f642e | 2882 | * |
401a8d04 TH |
2883 | * cancel_work_sync(&delayed_work->work) must not be used for |
2884 | * delayed_work's. Use cancel_delayed_work_sync() instead. | |
6e84d644 | 2885 | * |
401a8d04 | 2886 | * The caller must ensure that the workqueue on which @work was last |
6e84d644 | 2887 | * queued can't be destroyed before this function returns. |
401a8d04 | 2888 | * |
d185af30 | 2889 | * Return: |
401a8d04 | 2890 | * %true if @work was pending, %false otherwise. |
6e84d644 | 2891 | */ |
401a8d04 | 2892 | bool cancel_work_sync(struct work_struct *work) |
6e84d644 | 2893 | { |
36e227d2 | 2894 | return __cancel_work_timer(work, false); |
b89deed3 | 2895 | } |
28e53bdd | 2896 | EXPORT_SYMBOL_GPL(cancel_work_sync); |
b89deed3 | 2897 | |
6e84d644 | 2898 | /** |
401a8d04 TH |
2899 | * flush_delayed_work - wait for a dwork to finish executing the last queueing |
2900 | * @dwork: the delayed work to flush | |
6e84d644 | 2901 | * |
401a8d04 TH |
2902 | * Delayed timer is cancelled and the pending work is queued for |
2903 | * immediate execution. Like flush_work(), this function only | |
2904 | * considers the last queueing instance of @dwork. | |
1f1f642e | 2905 | * |
d185af30 | 2906 | * Return: |
401a8d04 TH |
2907 | * %true if flush_work() waited for the work to finish execution, |
2908 | * %false if it was already idle. | |
6e84d644 | 2909 | */ |
401a8d04 TH |
2910 | bool flush_delayed_work(struct delayed_work *dwork) |
2911 | { | |
8930caba | 2912 | local_irq_disable(); |
401a8d04 | 2913 | if (del_timer_sync(&dwork->timer)) |
60c057bc | 2914 | __queue_work(dwork->cpu, dwork->wq, &dwork->work); |
8930caba | 2915 | local_irq_enable(); |
401a8d04 TH |
2916 | return flush_work(&dwork->work); |
2917 | } | |
2918 | EXPORT_SYMBOL(flush_delayed_work); | |
2919 | ||
09383498 | 2920 | /** |
57b30ae7 TH |
2921 | * cancel_delayed_work - cancel a delayed work |
2922 | * @dwork: delayed_work to cancel | |
09383498 | 2923 | * |
d185af30 YB |
2924 | * Kill off a pending delayed_work. |
2925 | * | |
2926 | * Return: %true if @dwork was pending and canceled; %false if it wasn't | |
2927 | * pending. | |
2928 | * | |
2929 | * Note: | |
2930 | * The work callback function may still be running on return, unless | |
2931 | * it returns %true and the work doesn't re-arm itself. Explicitly flush or | |
2932 | * use cancel_delayed_work_sync() to wait on it. | |
09383498 | 2933 | * |
57b30ae7 | 2934 | * This function is safe to call from any context including IRQ handler. |
09383498 | 2935 | */ |
57b30ae7 | 2936 | bool cancel_delayed_work(struct delayed_work *dwork) |
09383498 | 2937 | { |
57b30ae7 TH |
2938 | unsigned long flags; |
2939 | int ret; | |
2940 | ||
2941 | do { | |
2942 | ret = try_to_grab_pending(&dwork->work, true, &flags); | |
2943 | } while (unlikely(ret == -EAGAIN)); | |
2944 | ||
2945 | if (unlikely(ret < 0)) | |
2946 | return false; | |
2947 | ||
7c3eed5c TH |
2948 | set_work_pool_and_clear_pending(&dwork->work, |
2949 | get_work_pool_id(&dwork->work)); | |
57b30ae7 | 2950 | local_irq_restore(flags); |
c0158ca6 | 2951 | return ret; |
09383498 | 2952 | } |
57b30ae7 | 2953 | EXPORT_SYMBOL(cancel_delayed_work); |
09383498 | 2954 | |
401a8d04 TH |
2955 | /** |
2956 | * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish | |
2957 | * @dwork: the delayed work cancel | |
2958 | * | |
2959 | * This is cancel_work_sync() for delayed works. | |
2960 | * | |
d185af30 | 2961 | * Return: |
401a8d04 TH |
2962 | * %true if @dwork was pending, %false otherwise. |
2963 | */ | |
2964 | bool cancel_delayed_work_sync(struct delayed_work *dwork) | |
6e84d644 | 2965 | { |
36e227d2 | 2966 | return __cancel_work_timer(&dwork->work, true); |
6e84d644 | 2967 | } |
f5a421a4 | 2968 | EXPORT_SYMBOL(cancel_delayed_work_sync); |
1da177e4 | 2969 | |
b6136773 | 2970 | /** |
31ddd871 | 2971 | * schedule_on_each_cpu - execute a function synchronously on each online CPU |
b6136773 | 2972 | * @func: the function to call |
b6136773 | 2973 | * |
31ddd871 TH |
2974 | * schedule_on_each_cpu() executes @func on each online CPU using the |
2975 | * system workqueue and blocks until all CPUs have completed. | |
b6136773 | 2976 | * schedule_on_each_cpu() is very slow. |
31ddd871 | 2977 | * |
d185af30 | 2978 | * Return: |
31ddd871 | 2979 | * 0 on success, -errno on failure. |
b6136773 | 2980 | */ |
65f27f38 | 2981 | int schedule_on_each_cpu(work_func_t func) |
15316ba8 CL |
2982 | { |
2983 | int cpu; | |
38f51568 | 2984 | struct work_struct __percpu *works; |
15316ba8 | 2985 | |
b6136773 AM |
2986 | works = alloc_percpu(struct work_struct); |
2987 | if (!works) | |
15316ba8 | 2988 | return -ENOMEM; |
b6136773 | 2989 | |
93981800 TH |
2990 | get_online_cpus(); |
2991 | ||
15316ba8 | 2992 | for_each_online_cpu(cpu) { |
9bfb1839 IM |
2993 | struct work_struct *work = per_cpu_ptr(works, cpu); |
2994 | ||
2995 | INIT_WORK(work, func); | |
b71ab8c2 | 2996 | schedule_work_on(cpu, work); |
65a64464 | 2997 | } |
93981800 TH |
2998 | |
2999 | for_each_online_cpu(cpu) | |
3000 | flush_work(per_cpu_ptr(works, cpu)); | |
3001 | ||
95402b38 | 3002 | put_online_cpus(); |
b6136773 | 3003 | free_percpu(works); |
15316ba8 CL |
3004 | return 0; |
3005 | } | |
3006 | ||
eef6a7d5 AS |
3007 | /** |
3008 | * flush_scheduled_work - ensure that any scheduled work has run to completion. | |
3009 | * | |
3010 | * Forces execution of the kernel-global workqueue and blocks until its | |
3011 | * completion. | |
3012 | * | |
3013 | * Think twice before calling this function! It's very easy to get into | |
3014 | * trouble if you don't take great care. Either of the following situations | |
3015 | * will lead to deadlock: | |
3016 | * | |
3017 | * One of the work items currently on the workqueue needs to acquire | |
3018 | * a lock held by your code or its caller. | |
3019 | * | |
3020 | * Your code is running in the context of a work routine. | |
3021 | * | |
3022 | * They will be detected by lockdep when they occur, but the first might not | |
3023 | * occur very often. It depends on what work items are on the workqueue and | |
3024 | * what locks they need, which you have no control over. | |
3025 | * | |
3026 | * In most situations flushing the entire workqueue is overkill; you merely | |
3027 | * need to know that a particular work item isn't queued and isn't running. | |
3028 | * In such cases you should use cancel_delayed_work_sync() or | |
3029 | * cancel_work_sync() instead. | |
3030 | */ | |
1da177e4 LT |
3031 | void flush_scheduled_work(void) |
3032 | { | |
d320c038 | 3033 | flush_workqueue(system_wq); |
1da177e4 | 3034 | } |
ae90dd5d | 3035 | EXPORT_SYMBOL(flush_scheduled_work); |
1da177e4 | 3036 | |
1fa44eca JB |
3037 | /** |
3038 | * execute_in_process_context - reliably execute the routine with user context | |
3039 | * @fn: the function to execute | |
1fa44eca JB |
3040 | * @ew: guaranteed storage for the execute work structure (must |
3041 | * be available when the work executes) | |
3042 | * | |
3043 | * Executes the function immediately if process context is available, | |
3044 | * otherwise schedules the function for delayed execution. | |
3045 | * | |
d185af30 | 3046 | * Return: 0 - function was executed |
1fa44eca JB |
3047 | * 1 - function was scheduled for execution |
3048 | */ | |
65f27f38 | 3049 | int execute_in_process_context(work_func_t fn, struct execute_work *ew) |
1fa44eca JB |
3050 | { |
3051 | if (!in_interrupt()) { | |
65f27f38 | 3052 | fn(&ew->work); |
1fa44eca JB |
3053 | return 0; |
3054 | } | |
3055 | ||
65f27f38 | 3056 | INIT_WORK(&ew->work, fn); |
1fa44eca JB |
3057 | schedule_work(&ew->work); |
3058 | ||
3059 | return 1; | |
3060 | } | |
3061 | EXPORT_SYMBOL_GPL(execute_in_process_context); | |
3062 | ||
226223ab TH |
3063 | #ifdef CONFIG_SYSFS |
3064 | /* | |
3065 | * Workqueues with WQ_SYSFS flag set is visible to userland via | |
3066 | * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the | |
3067 | * following attributes. | |
3068 | * | |
3069 | * per_cpu RO bool : whether the workqueue is per-cpu or unbound | |
3070 | * max_active RW int : maximum number of in-flight work items | |
3071 | * | |
3072 | * Unbound workqueues have the following extra attributes. | |
3073 | * | |
3074 | * id RO int : the associated pool ID | |
3075 | * nice RW int : nice value of the workers | |
3076 | * cpumask RW mask : bitmask of allowed CPUs for the workers | |
3077 | */ | |
3078 | struct wq_device { | |
3079 | struct workqueue_struct *wq; | |
3080 | struct device dev; | |
3081 | }; | |
3082 | ||
3083 | static struct workqueue_struct *dev_to_wq(struct device *dev) | |
3084 | { | |
3085 | struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); | |
3086 | ||
3087 | return wq_dev->wq; | |
3088 | } | |
3089 | ||
1a6661da GKH |
3090 | static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr, |
3091 | char *buf) | |
226223ab TH |
3092 | { |
3093 | struct workqueue_struct *wq = dev_to_wq(dev); | |
3094 | ||
3095 | return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND)); | |
3096 | } | |
1a6661da | 3097 | static DEVICE_ATTR_RO(per_cpu); |
226223ab | 3098 | |
1a6661da GKH |
3099 | static ssize_t max_active_show(struct device *dev, |
3100 | struct device_attribute *attr, char *buf) | |
226223ab TH |
3101 | { |
3102 | struct workqueue_struct *wq = dev_to_wq(dev); | |
3103 | ||
3104 | return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active); | |
3105 | } | |
3106 | ||
1a6661da GKH |
3107 | static ssize_t max_active_store(struct device *dev, |
3108 | struct device_attribute *attr, const char *buf, | |
3109 | size_t count) | |
226223ab TH |
3110 | { |
3111 | struct workqueue_struct *wq = dev_to_wq(dev); | |
3112 | int val; | |
3113 | ||
3114 | if (sscanf(buf, "%d", &val) != 1 || val <= 0) | |
3115 | return -EINVAL; | |
3116 | ||
3117 | workqueue_set_max_active(wq, val); | |
3118 | return count; | |
3119 | } | |
1a6661da | 3120 | static DEVICE_ATTR_RW(max_active); |
226223ab | 3121 | |
1a6661da GKH |
3122 | static struct attribute *wq_sysfs_attrs[] = { |
3123 | &dev_attr_per_cpu.attr, | |
3124 | &dev_attr_max_active.attr, | |
3125 | NULL, | |
226223ab | 3126 | }; |
1a6661da | 3127 | ATTRIBUTE_GROUPS(wq_sysfs); |
226223ab | 3128 | |
d55262c4 TH |
3129 | static ssize_t wq_pool_ids_show(struct device *dev, |
3130 | struct device_attribute *attr, char *buf) | |
226223ab TH |
3131 | { |
3132 | struct workqueue_struct *wq = dev_to_wq(dev); | |
d55262c4 TH |
3133 | const char *delim = ""; |
3134 | int node, written = 0; | |
226223ab TH |
3135 | |
3136 | rcu_read_lock_sched(); | |
d55262c4 TH |
3137 | for_each_node(node) { |
3138 | written += scnprintf(buf + written, PAGE_SIZE - written, | |
3139 | "%s%d:%d", delim, node, | |
3140 | unbound_pwq_by_node(wq, node)->pool->id); | |
3141 | delim = " "; | |
3142 | } | |
3143 | written += scnprintf(buf + written, PAGE_SIZE - written, "\n"); | |
226223ab TH |
3144 | rcu_read_unlock_sched(); |
3145 | ||
3146 | return written; | |
3147 | } | |
3148 | ||
3149 | static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr, | |
3150 | char *buf) | |
3151 | { | |
3152 | struct workqueue_struct *wq = dev_to_wq(dev); | |
3153 | int written; | |
3154 | ||
6029a918 TH |
3155 | mutex_lock(&wq->mutex); |
3156 | written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice); | |
3157 | mutex_unlock(&wq->mutex); | |
226223ab TH |
3158 | |
3159 | return written; | |
3160 | } | |
3161 | ||
3162 | /* prepare workqueue_attrs for sysfs store operations */ | |
3163 | static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq) | |
3164 | { | |
3165 | struct workqueue_attrs *attrs; | |
3166 | ||
3167 | attrs = alloc_workqueue_attrs(GFP_KERNEL); | |
3168 | if (!attrs) | |
3169 | return NULL; | |
3170 | ||
6029a918 TH |
3171 | mutex_lock(&wq->mutex); |
3172 | copy_workqueue_attrs(attrs, wq->unbound_attrs); | |
3173 | mutex_unlock(&wq->mutex); | |
226223ab TH |
3174 | return attrs; |
3175 | } | |
3176 | ||
3177 | static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr, | |
3178 | const char *buf, size_t count) | |
3179 | { | |
3180 | struct workqueue_struct *wq = dev_to_wq(dev); | |
3181 | struct workqueue_attrs *attrs; | |
3182 | int ret; | |
3183 | ||
3184 | attrs = wq_sysfs_prep_attrs(wq); | |
3185 | if (!attrs) | |
3186 | return -ENOMEM; | |
3187 | ||
3188 | if (sscanf(buf, "%d", &attrs->nice) == 1 && | |
14481842 | 3189 | attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE) |
226223ab TH |
3190 | ret = apply_workqueue_attrs(wq, attrs); |
3191 | else | |
3192 | ret = -EINVAL; | |
3193 | ||
3194 | free_workqueue_attrs(attrs); | |
3195 | return ret ?: count; | |
3196 | } | |
3197 | ||
3198 | static ssize_t wq_cpumask_show(struct device *dev, | |
3199 | struct device_attribute *attr, char *buf) | |
3200 | { | |
3201 | struct workqueue_struct *wq = dev_to_wq(dev); | |
3202 | int written; | |
3203 | ||
6029a918 TH |
3204 | mutex_lock(&wq->mutex); |
3205 | written = cpumask_scnprintf(buf, PAGE_SIZE, wq->unbound_attrs->cpumask); | |
3206 | mutex_unlock(&wq->mutex); | |
226223ab TH |
3207 | |
3208 | written += scnprintf(buf + written, PAGE_SIZE - written, "\n"); | |
3209 | return written; | |
3210 | } | |
3211 | ||
3212 | static ssize_t wq_cpumask_store(struct device *dev, | |
3213 | struct device_attribute *attr, | |
3214 | const char *buf, size_t count) | |
3215 | { | |
3216 | struct workqueue_struct *wq = dev_to_wq(dev); | |
3217 | struct workqueue_attrs *attrs; | |
3218 | int ret; | |
3219 | ||
3220 | attrs = wq_sysfs_prep_attrs(wq); | |
3221 | if (!attrs) | |
3222 | return -ENOMEM; | |
3223 | ||
3224 | ret = cpumask_parse(buf, attrs->cpumask); | |
3225 | if (!ret) | |
3226 | ret = apply_workqueue_attrs(wq, attrs); | |
3227 | ||
3228 | free_workqueue_attrs(attrs); | |
3229 | return ret ?: count; | |
3230 | } | |
3231 | ||
d55262c4 TH |
3232 | static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr, |
3233 | char *buf) | |
3234 | { | |
3235 | struct workqueue_struct *wq = dev_to_wq(dev); | |
3236 | int written; | |
3237 | ||
3238 | mutex_lock(&wq->mutex); | |
3239 | written = scnprintf(buf, PAGE_SIZE, "%d\n", | |
3240 | !wq->unbound_attrs->no_numa); | |
3241 | mutex_unlock(&wq->mutex); | |
3242 | ||
3243 | return written; | |
3244 | } | |
3245 | ||
3246 | static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr, | |
3247 | const char *buf, size_t count) | |
3248 | { | |
3249 | struct workqueue_struct *wq = dev_to_wq(dev); | |
3250 | struct workqueue_attrs *attrs; | |
3251 | int v, ret; | |
3252 | ||
3253 | attrs = wq_sysfs_prep_attrs(wq); | |
3254 | if (!attrs) | |
3255 | return -ENOMEM; | |
3256 | ||
3257 | ret = -EINVAL; | |
3258 | if (sscanf(buf, "%d", &v) == 1) { | |
3259 | attrs->no_numa = !v; | |
3260 | ret = apply_workqueue_attrs(wq, attrs); | |
3261 | } | |
3262 | ||
3263 | free_workqueue_attrs(attrs); | |
3264 | return ret ?: count; | |
3265 | } | |
3266 | ||
226223ab | 3267 | static struct device_attribute wq_sysfs_unbound_attrs[] = { |
d55262c4 | 3268 | __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL), |
226223ab TH |
3269 | __ATTR(nice, 0644, wq_nice_show, wq_nice_store), |
3270 | __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store), | |
d55262c4 | 3271 | __ATTR(numa, 0644, wq_numa_show, wq_numa_store), |
226223ab TH |
3272 | __ATTR_NULL, |
3273 | }; | |
3274 | ||
3275 | static struct bus_type wq_subsys = { | |
3276 | .name = "workqueue", | |
1a6661da | 3277 | .dev_groups = wq_sysfs_groups, |
226223ab TH |
3278 | }; |
3279 | ||
3280 | static int __init wq_sysfs_init(void) | |
3281 | { | |
3282 | return subsys_virtual_register(&wq_subsys, NULL); | |
3283 | } | |
3284 | core_initcall(wq_sysfs_init); | |
3285 | ||
3286 | static void wq_device_release(struct device *dev) | |
3287 | { | |
3288 | struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); | |
3289 | ||
3290 | kfree(wq_dev); | |
3291 | } | |
3292 | ||
3293 | /** | |
3294 | * workqueue_sysfs_register - make a workqueue visible in sysfs | |
3295 | * @wq: the workqueue to register | |
3296 | * | |
3297 | * Expose @wq in sysfs under /sys/bus/workqueue/devices. | |
3298 | * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set | |
3299 | * which is the preferred method. | |
3300 | * | |
3301 | * Workqueue user should use this function directly iff it wants to apply | |
3302 | * workqueue_attrs before making the workqueue visible in sysfs; otherwise, | |
3303 | * apply_workqueue_attrs() may race against userland updating the | |
3304 | * attributes. | |
3305 | * | |
d185af30 | 3306 | * Return: 0 on success, -errno on failure. |
226223ab TH |
3307 | */ |
3308 | int workqueue_sysfs_register(struct workqueue_struct *wq) | |
3309 | { | |
3310 | struct wq_device *wq_dev; | |
3311 | int ret; | |
3312 | ||
3313 | /* | |
3314 | * Adjusting max_active or creating new pwqs by applyting | |
3315 | * attributes breaks ordering guarantee. Disallow exposing ordered | |
3316 | * workqueues. | |
3317 | */ | |
3318 | if (WARN_ON(wq->flags & __WQ_ORDERED)) | |
3319 | return -EINVAL; | |
3320 | ||
3321 | wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL); | |
3322 | if (!wq_dev) | |
3323 | return -ENOMEM; | |
3324 | ||
3325 | wq_dev->wq = wq; | |
3326 | wq_dev->dev.bus = &wq_subsys; | |
3327 | wq_dev->dev.init_name = wq->name; | |
3328 | wq_dev->dev.release = wq_device_release; | |
3329 | ||
3330 | /* | |
3331 | * unbound_attrs are created separately. Suppress uevent until | |
3332 | * everything is ready. | |
3333 | */ | |
3334 | dev_set_uevent_suppress(&wq_dev->dev, true); | |
3335 | ||
3336 | ret = device_register(&wq_dev->dev); | |
3337 | if (ret) { | |
3338 | kfree(wq_dev); | |
3339 | wq->wq_dev = NULL; | |
3340 | return ret; | |
3341 | } | |
3342 | ||
3343 | if (wq->flags & WQ_UNBOUND) { | |
3344 | struct device_attribute *attr; | |
3345 | ||
3346 | for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) { | |
3347 | ret = device_create_file(&wq_dev->dev, attr); | |
3348 | if (ret) { | |
3349 | device_unregister(&wq_dev->dev); | |
3350 | wq->wq_dev = NULL; | |
3351 | return ret; | |
3352 | } | |
3353 | } | |
3354 | } | |
3355 | ||
3356 | kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD); | |
3357 | return 0; | |
3358 | } | |
3359 | ||
3360 | /** | |
3361 | * workqueue_sysfs_unregister - undo workqueue_sysfs_register() | |
3362 | * @wq: the workqueue to unregister | |
3363 | * | |
3364 | * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister. | |
3365 | */ | |
3366 | static void workqueue_sysfs_unregister(struct workqueue_struct *wq) | |
3367 | { | |
3368 | struct wq_device *wq_dev = wq->wq_dev; | |
3369 | ||
3370 | if (!wq->wq_dev) | |
3371 | return; | |
3372 | ||
3373 | wq->wq_dev = NULL; | |
3374 | device_unregister(&wq_dev->dev); | |
3375 | } | |
3376 | #else /* CONFIG_SYSFS */ | |
3377 | static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { } | |
3378 | #endif /* CONFIG_SYSFS */ | |
3379 | ||
7a4e344c TH |
3380 | /** |
3381 | * free_workqueue_attrs - free a workqueue_attrs | |
3382 | * @attrs: workqueue_attrs to free | |
3383 | * | |
3384 | * Undo alloc_workqueue_attrs(). | |
3385 | */ | |
3386 | void free_workqueue_attrs(struct workqueue_attrs *attrs) | |
3387 | { | |
3388 | if (attrs) { | |
3389 | free_cpumask_var(attrs->cpumask); | |
3390 | kfree(attrs); | |
3391 | } | |
3392 | } | |
3393 | ||
3394 | /** | |
3395 | * alloc_workqueue_attrs - allocate a workqueue_attrs | |
3396 | * @gfp_mask: allocation mask to use | |
3397 | * | |
3398 | * Allocate a new workqueue_attrs, initialize with default settings and | |
d185af30 YB |
3399 | * return it. |
3400 | * | |
3401 | * Return: The allocated new workqueue_attr on success. %NULL on failure. | |
7a4e344c TH |
3402 | */ |
3403 | struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask) | |
3404 | { | |
3405 | struct workqueue_attrs *attrs; | |
3406 | ||
3407 | attrs = kzalloc(sizeof(*attrs), gfp_mask); | |
3408 | if (!attrs) | |
3409 | goto fail; | |
3410 | if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask)) | |
3411 | goto fail; | |
3412 | ||
13e2e556 | 3413 | cpumask_copy(attrs->cpumask, cpu_possible_mask); |
7a4e344c TH |
3414 | return attrs; |
3415 | fail: | |
3416 | free_workqueue_attrs(attrs); | |
3417 | return NULL; | |
3418 | } | |
3419 | ||
29c91e99 TH |
3420 | static void copy_workqueue_attrs(struct workqueue_attrs *to, |
3421 | const struct workqueue_attrs *from) | |
3422 | { | |
3423 | to->nice = from->nice; | |
3424 | cpumask_copy(to->cpumask, from->cpumask); | |
2865a8fb SL |
3425 | /* |
3426 | * Unlike hash and equality test, this function doesn't ignore | |
3427 | * ->no_numa as it is used for both pool and wq attrs. Instead, | |
3428 | * get_unbound_pool() explicitly clears ->no_numa after copying. | |
3429 | */ | |
3430 | to->no_numa = from->no_numa; | |
29c91e99 TH |
3431 | } |
3432 | ||
29c91e99 TH |
3433 | /* hash value of the content of @attr */ |
3434 | static u32 wqattrs_hash(const struct workqueue_attrs *attrs) | |
3435 | { | |
3436 | u32 hash = 0; | |
3437 | ||
3438 | hash = jhash_1word(attrs->nice, hash); | |
13e2e556 TH |
3439 | hash = jhash(cpumask_bits(attrs->cpumask), |
3440 | BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash); | |
29c91e99 TH |
3441 | return hash; |
3442 | } | |
3443 | ||
3444 | /* content equality test */ | |
3445 | static bool wqattrs_equal(const struct workqueue_attrs *a, | |
3446 | const struct workqueue_attrs *b) | |
3447 | { | |
3448 | if (a->nice != b->nice) | |
3449 | return false; | |
3450 | if (!cpumask_equal(a->cpumask, b->cpumask)) | |
3451 | return false; | |
3452 | return true; | |
3453 | } | |
3454 | ||
7a4e344c TH |
3455 | /** |
3456 | * init_worker_pool - initialize a newly zalloc'd worker_pool | |
3457 | * @pool: worker_pool to initialize | |
3458 | * | |
3459 | * Initiailize a newly zalloc'd @pool. It also allocates @pool->attrs. | |
d185af30 YB |
3460 | * |
3461 | * Return: 0 on success, -errno on failure. Even on failure, all fields | |
29c91e99 TH |
3462 | * inside @pool proper are initialized and put_unbound_pool() can be called |
3463 | * on @pool safely to release it. | |
7a4e344c TH |
3464 | */ |
3465 | static int init_worker_pool(struct worker_pool *pool) | |
4e1a1f9a TH |
3466 | { |
3467 | spin_lock_init(&pool->lock); | |
29c91e99 TH |
3468 | pool->id = -1; |
3469 | pool->cpu = -1; | |
f3f90ad4 | 3470 | pool->node = NUMA_NO_NODE; |
4e1a1f9a TH |
3471 | pool->flags |= POOL_DISASSOCIATED; |
3472 | INIT_LIST_HEAD(&pool->worklist); | |
3473 | INIT_LIST_HEAD(&pool->idle_list); | |
3474 | hash_init(pool->busy_hash); | |
3475 | ||
3476 | init_timer_deferrable(&pool->idle_timer); | |
3477 | pool->idle_timer.function = idle_worker_timeout; | |
3478 | pool->idle_timer.data = (unsigned long)pool; | |
3479 | ||
3480 | setup_timer(&pool->mayday_timer, pool_mayday_timeout, | |
3481 | (unsigned long)pool); | |
3482 | ||
3483 | mutex_init(&pool->manager_arb); | |
bc3a1afc | 3484 | mutex_init(&pool->manager_mutex); |
822d8405 | 3485 | idr_init(&pool->worker_idr); |
7a4e344c | 3486 | |
29c91e99 TH |
3487 | INIT_HLIST_NODE(&pool->hash_node); |
3488 | pool->refcnt = 1; | |
3489 | ||
3490 | /* shouldn't fail above this point */ | |
7a4e344c TH |
3491 | pool->attrs = alloc_workqueue_attrs(GFP_KERNEL); |
3492 | if (!pool->attrs) | |
3493 | return -ENOMEM; | |
3494 | return 0; | |
4e1a1f9a TH |
3495 | } |
3496 | ||
29c91e99 TH |
3497 | static void rcu_free_pool(struct rcu_head *rcu) |
3498 | { | |
3499 | struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu); | |
3500 | ||
822d8405 | 3501 | idr_destroy(&pool->worker_idr); |
29c91e99 TH |
3502 | free_workqueue_attrs(pool->attrs); |
3503 | kfree(pool); | |
3504 | } | |
3505 | ||
3506 | /** | |
3507 | * put_unbound_pool - put a worker_pool | |
3508 | * @pool: worker_pool to put | |
3509 | * | |
3510 | * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU | |
c5aa87bb TH |
3511 | * safe manner. get_unbound_pool() calls this function on its failure path |
3512 | * and this function should be able to release pools which went through, | |
3513 | * successfully or not, init_worker_pool(). | |
a892cacc TH |
3514 | * |
3515 | * Should be called with wq_pool_mutex held. | |
29c91e99 TH |
3516 | */ |
3517 | static void put_unbound_pool(struct worker_pool *pool) | |
3518 | { | |
60f5a4bc | 3519 | DECLARE_COMPLETION_ONSTACK(detach_completion); |
29c91e99 TH |
3520 | struct worker *worker; |
3521 | ||
a892cacc TH |
3522 | lockdep_assert_held(&wq_pool_mutex); |
3523 | ||
3524 | if (--pool->refcnt) | |
29c91e99 | 3525 | return; |
29c91e99 TH |
3526 | |
3527 | /* sanity checks */ | |
3528 | if (WARN_ON(!(pool->flags & POOL_DISASSOCIATED)) || | |
a892cacc | 3529 | WARN_ON(!list_empty(&pool->worklist))) |
29c91e99 | 3530 | return; |
29c91e99 TH |
3531 | |
3532 | /* release id and unhash */ | |
3533 | if (pool->id >= 0) | |
3534 | idr_remove(&worker_pool_idr, pool->id); | |
3535 | hash_del(&pool->hash_node); | |
3536 | ||
c5aa87bb TH |
3537 | /* |
3538 | * Become the manager and destroy all workers. Grabbing | |
3539 | * manager_arb prevents @pool's workers from blocking on | |
3540 | * manager_mutex. | |
3541 | */ | |
29c91e99 | 3542 | mutex_lock(&pool->manager_arb); |
29c91e99 | 3543 | |
60f5a4bc | 3544 | spin_lock_irq(&pool->lock); |
29c91e99 TH |
3545 | while ((worker = first_worker(pool))) |
3546 | destroy_worker(worker); | |
3547 | WARN_ON(pool->nr_workers || pool->nr_idle); | |
29c91e99 | 3548 | spin_unlock_irq(&pool->lock); |
60f5a4bc LJ |
3549 | |
3550 | mutex_lock(&pool->manager_mutex); | |
3551 | if (!idr_is_empty(&pool->worker_idr)) | |
3552 | pool->detach_completion = &detach_completion; | |
cd549687 | 3553 | mutex_unlock(&pool->manager_mutex); |
60f5a4bc LJ |
3554 | |
3555 | if (pool->detach_completion) | |
3556 | wait_for_completion(pool->detach_completion); | |
3557 | ||
29c91e99 TH |
3558 | mutex_unlock(&pool->manager_arb); |
3559 | ||
3560 | /* shut down the timers */ | |
3561 | del_timer_sync(&pool->idle_timer); | |
3562 | del_timer_sync(&pool->mayday_timer); | |
3563 | ||
3564 | /* sched-RCU protected to allow dereferences from get_work_pool() */ | |
3565 | call_rcu_sched(&pool->rcu, rcu_free_pool); | |
3566 | } | |
3567 | ||
3568 | /** | |
3569 | * get_unbound_pool - get a worker_pool with the specified attributes | |
3570 | * @attrs: the attributes of the worker_pool to get | |
3571 | * | |
3572 | * Obtain a worker_pool which has the same attributes as @attrs, bump the | |
3573 | * reference count and return it. If there already is a matching | |
3574 | * worker_pool, it will be used; otherwise, this function attempts to | |
d185af30 | 3575 | * create a new one. |
a892cacc TH |
3576 | * |
3577 | * Should be called with wq_pool_mutex held. | |
d185af30 YB |
3578 | * |
3579 | * Return: On success, a worker_pool with the same attributes as @attrs. | |
3580 | * On failure, %NULL. | |
29c91e99 TH |
3581 | */ |
3582 | static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs) | |
3583 | { | |
29c91e99 TH |
3584 | u32 hash = wqattrs_hash(attrs); |
3585 | struct worker_pool *pool; | |
f3f90ad4 | 3586 | int node; |
29c91e99 | 3587 | |
a892cacc | 3588 | lockdep_assert_held(&wq_pool_mutex); |
29c91e99 TH |
3589 | |
3590 | /* do we already have a matching pool? */ | |
29c91e99 TH |
3591 | hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) { |
3592 | if (wqattrs_equal(pool->attrs, attrs)) { | |
3593 | pool->refcnt++; | |
3594 | goto out_unlock; | |
3595 | } | |
3596 | } | |
29c91e99 TH |
3597 | |
3598 | /* nope, create a new one */ | |
3599 | pool = kzalloc(sizeof(*pool), GFP_KERNEL); | |
3600 | if (!pool || init_worker_pool(pool) < 0) | |
3601 | goto fail; | |
3602 | ||
12ee4fc6 LJ |
3603 | if (workqueue_freezing) |
3604 | pool->flags |= POOL_FREEZING; | |
3605 | ||
8864b4e5 | 3606 | lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */ |
29c91e99 TH |
3607 | copy_workqueue_attrs(pool->attrs, attrs); |
3608 | ||
2865a8fb SL |
3609 | /* |
3610 | * no_numa isn't a worker_pool attribute, always clear it. See | |
3611 | * 'struct workqueue_attrs' comments for detail. | |
3612 | */ | |
3613 | pool->attrs->no_numa = false; | |
3614 | ||
f3f90ad4 TH |
3615 | /* if cpumask is contained inside a NUMA node, we belong to that node */ |
3616 | if (wq_numa_enabled) { | |
3617 | for_each_node(node) { | |
3618 | if (cpumask_subset(pool->attrs->cpumask, | |
3619 | wq_numa_possible_cpumask[node])) { | |
3620 | pool->node = node; | |
3621 | break; | |
3622 | } | |
3623 | } | |
3624 | } | |
3625 | ||
29c91e99 TH |
3626 | if (worker_pool_assign_id(pool) < 0) |
3627 | goto fail; | |
3628 | ||
3629 | /* create and start the initial worker */ | |
ebf44d16 | 3630 | if (create_and_start_worker(pool) < 0) |
29c91e99 TH |
3631 | goto fail; |
3632 | ||
29c91e99 | 3633 | /* install */ |
29c91e99 TH |
3634 | hash_add(unbound_pool_hash, &pool->hash_node, hash); |
3635 | out_unlock: | |
29c91e99 TH |
3636 | return pool; |
3637 | fail: | |
29c91e99 TH |
3638 | if (pool) |
3639 | put_unbound_pool(pool); | |
3640 | return NULL; | |
3641 | } | |
3642 | ||
8864b4e5 TH |
3643 | static void rcu_free_pwq(struct rcu_head *rcu) |
3644 | { | |
3645 | kmem_cache_free(pwq_cache, | |
3646 | container_of(rcu, struct pool_workqueue, rcu)); | |
3647 | } | |
3648 | ||
3649 | /* | |
3650 | * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt | |
3651 | * and needs to be destroyed. | |
3652 | */ | |
3653 | static void pwq_unbound_release_workfn(struct work_struct *work) | |
3654 | { | |
3655 | struct pool_workqueue *pwq = container_of(work, struct pool_workqueue, | |
3656 | unbound_release_work); | |
3657 | struct workqueue_struct *wq = pwq->wq; | |
3658 | struct worker_pool *pool = pwq->pool; | |
bc0caf09 | 3659 | bool is_last; |
8864b4e5 TH |
3660 | |
3661 | if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND))) | |
3662 | return; | |
3663 | ||
75ccf595 | 3664 | /* |
3c25a55d | 3665 | * Unlink @pwq. Synchronization against wq->mutex isn't strictly |
75ccf595 TH |
3666 | * necessary on release but do it anyway. It's easier to verify |
3667 | * and consistent with the linking path. | |
3668 | */ | |
3c25a55d | 3669 | mutex_lock(&wq->mutex); |
8864b4e5 | 3670 | list_del_rcu(&pwq->pwqs_node); |
bc0caf09 | 3671 | is_last = list_empty(&wq->pwqs); |
3c25a55d | 3672 | mutex_unlock(&wq->mutex); |
8864b4e5 | 3673 | |
a892cacc | 3674 | mutex_lock(&wq_pool_mutex); |
8864b4e5 | 3675 | put_unbound_pool(pool); |
a892cacc TH |
3676 | mutex_unlock(&wq_pool_mutex); |
3677 | ||
8864b4e5 TH |
3678 | call_rcu_sched(&pwq->rcu, rcu_free_pwq); |
3679 | ||
3680 | /* | |
3681 | * If we're the last pwq going away, @wq is already dead and no one | |
3682 | * is gonna access it anymore. Free it. | |
3683 | */ | |
6029a918 TH |
3684 | if (is_last) { |
3685 | free_workqueue_attrs(wq->unbound_attrs); | |
8864b4e5 | 3686 | kfree(wq); |
6029a918 | 3687 | } |
8864b4e5 TH |
3688 | } |
3689 | ||
0fbd95aa | 3690 | /** |
699ce097 | 3691 | * pwq_adjust_max_active - update a pwq's max_active to the current setting |
0fbd95aa | 3692 | * @pwq: target pool_workqueue |
0fbd95aa | 3693 | * |
699ce097 TH |
3694 | * If @pwq isn't freezing, set @pwq->max_active to the associated |
3695 | * workqueue's saved_max_active and activate delayed work items | |
3696 | * accordingly. If @pwq is freezing, clear @pwq->max_active to zero. | |
0fbd95aa | 3697 | */ |
699ce097 | 3698 | static void pwq_adjust_max_active(struct pool_workqueue *pwq) |
0fbd95aa | 3699 | { |
699ce097 TH |
3700 | struct workqueue_struct *wq = pwq->wq; |
3701 | bool freezable = wq->flags & WQ_FREEZABLE; | |
3702 | ||
3703 | /* for @wq->saved_max_active */ | |
a357fc03 | 3704 | lockdep_assert_held(&wq->mutex); |
699ce097 TH |
3705 | |
3706 | /* fast exit for non-freezable wqs */ | |
3707 | if (!freezable && pwq->max_active == wq->saved_max_active) | |
3708 | return; | |
3709 | ||
a357fc03 | 3710 | spin_lock_irq(&pwq->pool->lock); |
699ce097 TH |
3711 | |
3712 | if (!freezable || !(pwq->pool->flags & POOL_FREEZING)) { | |
3713 | pwq->max_active = wq->saved_max_active; | |
0fbd95aa | 3714 | |
699ce097 TH |
3715 | while (!list_empty(&pwq->delayed_works) && |
3716 | pwq->nr_active < pwq->max_active) | |
3717 | pwq_activate_first_delayed(pwq); | |
951a078a LJ |
3718 | |
3719 | /* | |
3720 | * Need to kick a worker after thawed or an unbound wq's | |
3721 | * max_active is bumped. It's a slow path. Do it always. | |
3722 | */ | |
3723 | wake_up_worker(pwq->pool); | |
699ce097 TH |
3724 | } else { |
3725 | pwq->max_active = 0; | |
3726 | } | |
3727 | ||
a357fc03 | 3728 | spin_unlock_irq(&pwq->pool->lock); |
0fbd95aa TH |
3729 | } |
3730 | ||
e50aba9a | 3731 | /* initialize newly alloced @pwq which is associated with @wq and @pool */ |
f147f29e TH |
3732 | static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq, |
3733 | struct worker_pool *pool) | |
d2c1d404 TH |
3734 | { |
3735 | BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK); | |
3736 | ||
e50aba9a TH |
3737 | memset(pwq, 0, sizeof(*pwq)); |
3738 | ||
d2c1d404 TH |
3739 | pwq->pool = pool; |
3740 | pwq->wq = wq; | |
3741 | pwq->flush_color = -1; | |
8864b4e5 | 3742 | pwq->refcnt = 1; |
d2c1d404 | 3743 | INIT_LIST_HEAD(&pwq->delayed_works); |
1befcf30 | 3744 | INIT_LIST_HEAD(&pwq->pwqs_node); |
d2c1d404 | 3745 | INIT_LIST_HEAD(&pwq->mayday_node); |
8864b4e5 | 3746 | INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn); |
f147f29e | 3747 | } |
d2c1d404 | 3748 | |
f147f29e | 3749 | /* sync @pwq with the current state of its associated wq and link it */ |
1befcf30 | 3750 | static void link_pwq(struct pool_workqueue *pwq) |
f147f29e TH |
3751 | { |
3752 | struct workqueue_struct *wq = pwq->wq; | |
3753 | ||
3754 | lockdep_assert_held(&wq->mutex); | |
75ccf595 | 3755 | |
1befcf30 TH |
3756 | /* may be called multiple times, ignore if already linked */ |
3757 | if (!list_empty(&pwq->pwqs_node)) | |
3758 | return; | |
3759 | ||
983ca25e TH |
3760 | /* |
3761 | * Set the matching work_color. This is synchronized with | |
3c25a55d | 3762 | * wq->mutex to avoid confusing flush_workqueue(). |
983ca25e | 3763 | */ |
75ccf595 | 3764 | pwq->work_color = wq->work_color; |
983ca25e TH |
3765 | |
3766 | /* sync max_active to the current setting */ | |
3767 | pwq_adjust_max_active(pwq); | |
3768 | ||
3769 | /* link in @pwq */ | |
9e8cd2f5 | 3770 | list_add_rcu(&pwq->pwqs_node, &wq->pwqs); |
f147f29e | 3771 | } |
a357fc03 | 3772 | |
f147f29e TH |
3773 | /* obtain a pool matching @attr and create a pwq associating the pool and @wq */ |
3774 | static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq, | |
3775 | const struct workqueue_attrs *attrs) | |
3776 | { | |
3777 | struct worker_pool *pool; | |
3778 | struct pool_workqueue *pwq; | |
3779 | ||
3780 | lockdep_assert_held(&wq_pool_mutex); | |
3781 | ||
3782 | pool = get_unbound_pool(attrs); | |
3783 | if (!pool) | |
3784 | return NULL; | |
3785 | ||
e50aba9a | 3786 | pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node); |
f147f29e TH |
3787 | if (!pwq) { |
3788 | put_unbound_pool(pool); | |
3789 | return NULL; | |
df2d5ae4 | 3790 | } |
6029a918 | 3791 | |
f147f29e TH |
3792 | init_pwq(pwq, wq, pool); |
3793 | return pwq; | |
d2c1d404 TH |
3794 | } |
3795 | ||
4c16bd32 TH |
3796 | /* undo alloc_unbound_pwq(), used only in the error path */ |
3797 | static void free_unbound_pwq(struct pool_workqueue *pwq) | |
3798 | { | |
3799 | lockdep_assert_held(&wq_pool_mutex); | |
3800 | ||
3801 | if (pwq) { | |
3802 | put_unbound_pool(pwq->pool); | |
cece95df | 3803 | kmem_cache_free(pwq_cache, pwq); |
4c16bd32 TH |
3804 | } |
3805 | } | |
3806 | ||
3807 | /** | |
3808 | * wq_calc_node_mask - calculate a wq_attrs' cpumask for the specified node | |
3809 | * @attrs: the wq_attrs of interest | |
3810 | * @node: the target NUMA node | |
3811 | * @cpu_going_down: if >= 0, the CPU to consider as offline | |
3812 | * @cpumask: outarg, the resulting cpumask | |
3813 | * | |
3814 | * Calculate the cpumask a workqueue with @attrs should use on @node. If | |
3815 | * @cpu_going_down is >= 0, that cpu is considered offline during | |
d185af30 | 3816 | * calculation. The result is stored in @cpumask. |
4c16bd32 TH |
3817 | * |
3818 | * If NUMA affinity is not enabled, @attrs->cpumask is always used. If | |
3819 | * enabled and @node has online CPUs requested by @attrs, the returned | |
3820 | * cpumask is the intersection of the possible CPUs of @node and | |
3821 | * @attrs->cpumask. | |
3822 | * | |
3823 | * The caller is responsible for ensuring that the cpumask of @node stays | |
3824 | * stable. | |
d185af30 YB |
3825 | * |
3826 | * Return: %true if the resulting @cpumask is different from @attrs->cpumask, | |
3827 | * %false if equal. | |
4c16bd32 TH |
3828 | */ |
3829 | static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node, | |
3830 | int cpu_going_down, cpumask_t *cpumask) | |
3831 | { | |
d55262c4 | 3832 | if (!wq_numa_enabled || attrs->no_numa) |
4c16bd32 TH |
3833 | goto use_dfl; |
3834 | ||
3835 | /* does @node have any online CPUs @attrs wants? */ | |
3836 | cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask); | |
3837 | if (cpu_going_down >= 0) | |
3838 | cpumask_clear_cpu(cpu_going_down, cpumask); | |
3839 | ||
3840 | if (cpumask_empty(cpumask)) | |
3841 | goto use_dfl; | |
3842 | ||
3843 | /* yeap, return possible CPUs in @node that @attrs wants */ | |
3844 | cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]); | |
3845 | return !cpumask_equal(cpumask, attrs->cpumask); | |
3846 | ||
3847 | use_dfl: | |
3848 | cpumask_copy(cpumask, attrs->cpumask); | |
3849 | return false; | |
3850 | } | |
3851 | ||
1befcf30 TH |
3852 | /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */ |
3853 | static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq, | |
3854 | int node, | |
3855 | struct pool_workqueue *pwq) | |
3856 | { | |
3857 | struct pool_workqueue *old_pwq; | |
3858 | ||
3859 | lockdep_assert_held(&wq->mutex); | |
3860 | ||
3861 | /* link_pwq() can handle duplicate calls */ | |
3862 | link_pwq(pwq); | |
3863 | ||
3864 | old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); | |
3865 | rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq); | |
3866 | return old_pwq; | |
3867 | } | |
3868 | ||
9e8cd2f5 TH |
3869 | /** |
3870 | * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue | |
3871 | * @wq: the target workqueue | |
3872 | * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs() | |
3873 | * | |
4c16bd32 TH |
3874 | * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA |
3875 | * machines, this function maps a separate pwq to each NUMA node with | |
3876 | * possibles CPUs in @attrs->cpumask so that work items are affine to the | |
3877 | * NUMA node it was issued on. Older pwqs are released as in-flight work | |
3878 | * items finish. Note that a work item which repeatedly requeues itself | |
3879 | * back-to-back will stay on its current pwq. | |
9e8cd2f5 | 3880 | * |
d185af30 YB |
3881 | * Performs GFP_KERNEL allocations. |
3882 | * | |
3883 | * Return: 0 on success and -errno on failure. | |
9e8cd2f5 TH |
3884 | */ |
3885 | int apply_workqueue_attrs(struct workqueue_struct *wq, | |
3886 | const struct workqueue_attrs *attrs) | |
3887 | { | |
4c16bd32 TH |
3888 | struct workqueue_attrs *new_attrs, *tmp_attrs; |
3889 | struct pool_workqueue **pwq_tbl, *dfl_pwq; | |
f147f29e | 3890 | int node, ret; |
9e8cd2f5 | 3891 | |
8719dcea | 3892 | /* only unbound workqueues can change attributes */ |
9e8cd2f5 TH |
3893 | if (WARN_ON(!(wq->flags & WQ_UNBOUND))) |
3894 | return -EINVAL; | |
3895 | ||
8719dcea TH |
3896 | /* creating multiple pwqs breaks ordering guarantee */ |
3897 | if (WARN_ON((wq->flags & __WQ_ORDERED) && !list_empty(&wq->pwqs))) | |
3898 | return -EINVAL; | |
3899 | ||
4c16bd32 | 3900 | pwq_tbl = kzalloc(wq_numa_tbl_len * sizeof(pwq_tbl[0]), GFP_KERNEL); |
13e2e556 | 3901 | new_attrs = alloc_workqueue_attrs(GFP_KERNEL); |
4c16bd32 TH |
3902 | tmp_attrs = alloc_workqueue_attrs(GFP_KERNEL); |
3903 | if (!pwq_tbl || !new_attrs || !tmp_attrs) | |
13e2e556 TH |
3904 | goto enomem; |
3905 | ||
4c16bd32 | 3906 | /* make a copy of @attrs and sanitize it */ |
13e2e556 TH |
3907 | copy_workqueue_attrs(new_attrs, attrs); |
3908 | cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask); | |
3909 | ||
4c16bd32 TH |
3910 | /* |
3911 | * We may create multiple pwqs with differing cpumasks. Make a | |
3912 | * copy of @new_attrs which will be modified and used to obtain | |
3913 | * pools. | |
3914 | */ | |
3915 | copy_workqueue_attrs(tmp_attrs, new_attrs); | |
3916 | ||
3917 | /* | |
3918 | * CPUs should stay stable across pwq creations and installations. | |
3919 | * Pin CPUs, determine the target cpumask for each node and create | |
3920 | * pwqs accordingly. | |
3921 | */ | |
3922 | get_online_cpus(); | |
3923 | ||
a892cacc | 3924 | mutex_lock(&wq_pool_mutex); |
4c16bd32 TH |
3925 | |
3926 | /* | |
3927 | * If something goes wrong during CPU up/down, we'll fall back to | |
3928 | * the default pwq covering whole @attrs->cpumask. Always create | |
3929 | * it even if we don't use it immediately. | |
3930 | */ | |
3931 | dfl_pwq = alloc_unbound_pwq(wq, new_attrs); | |
3932 | if (!dfl_pwq) | |
3933 | goto enomem_pwq; | |
3934 | ||
3935 | for_each_node(node) { | |
3936 | if (wq_calc_node_cpumask(attrs, node, -1, tmp_attrs->cpumask)) { | |
3937 | pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs); | |
3938 | if (!pwq_tbl[node]) | |
3939 | goto enomem_pwq; | |
3940 | } else { | |
3941 | dfl_pwq->refcnt++; | |
3942 | pwq_tbl[node] = dfl_pwq; | |
3943 | } | |
3944 | } | |
3945 | ||
f147f29e | 3946 | mutex_unlock(&wq_pool_mutex); |
9e8cd2f5 | 3947 | |
4c16bd32 | 3948 | /* all pwqs have been created successfully, let's install'em */ |
f147f29e | 3949 | mutex_lock(&wq->mutex); |
a892cacc | 3950 | |
f147f29e | 3951 | copy_workqueue_attrs(wq->unbound_attrs, new_attrs); |
4c16bd32 TH |
3952 | |
3953 | /* save the previous pwq and install the new one */ | |
f147f29e | 3954 | for_each_node(node) |
4c16bd32 TH |
3955 | pwq_tbl[node] = numa_pwq_tbl_install(wq, node, pwq_tbl[node]); |
3956 | ||
3957 | /* @dfl_pwq might not have been used, ensure it's linked */ | |
3958 | link_pwq(dfl_pwq); | |
3959 | swap(wq->dfl_pwq, dfl_pwq); | |
f147f29e TH |
3960 | |
3961 | mutex_unlock(&wq->mutex); | |
9e8cd2f5 | 3962 | |
4c16bd32 TH |
3963 | /* put the old pwqs */ |
3964 | for_each_node(node) | |
3965 | put_pwq_unlocked(pwq_tbl[node]); | |
3966 | put_pwq_unlocked(dfl_pwq); | |
3967 | ||
3968 | put_online_cpus(); | |
4862125b TH |
3969 | ret = 0; |
3970 | /* fall through */ | |
3971 | out_free: | |
4c16bd32 | 3972 | free_workqueue_attrs(tmp_attrs); |
4862125b | 3973 | free_workqueue_attrs(new_attrs); |
4c16bd32 | 3974 | kfree(pwq_tbl); |
4862125b | 3975 | return ret; |
13e2e556 | 3976 | |
4c16bd32 TH |
3977 | enomem_pwq: |
3978 | free_unbound_pwq(dfl_pwq); | |
3979 | for_each_node(node) | |
3980 | if (pwq_tbl && pwq_tbl[node] != dfl_pwq) | |
3981 | free_unbound_pwq(pwq_tbl[node]); | |
3982 | mutex_unlock(&wq_pool_mutex); | |
3983 | put_online_cpus(); | |
13e2e556 | 3984 | enomem: |
4862125b TH |
3985 | ret = -ENOMEM; |
3986 | goto out_free; | |
9e8cd2f5 TH |
3987 | } |
3988 | ||
4c16bd32 TH |
3989 | /** |
3990 | * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug | |
3991 | * @wq: the target workqueue | |
3992 | * @cpu: the CPU coming up or going down | |
3993 | * @online: whether @cpu is coming up or going down | |
3994 | * | |
3995 | * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and | |
3996 | * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of | |
3997 | * @wq accordingly. | |
3998 | * | |
3999 | * If NUMA affinity can't be adjusted due to memory allocation failure, it | |
4000 | * falls back to @wq->dfl_pwq which may not be optimal but is always | |
4001 | * correct. | |
4002 | * | |
4003 | * Note that when the last allowed CPU of a NUMA node goes offline for a | |
4004 | * workqueue with a cpumask spanning multiple nodes, the workers which were | |
4005 | * already executing the work items for the workqueue will lose their CPU | |
4006 | * affinity and may execute on any CPU. This is similar to how per-cpu | |
4007 | * workqueues behave on CPU_DOWN. If a workqueue user wants strict | |
4008 | * affinity, it's the user's responsibility to flush the work item from | |
4009 | * CPU_DOWN_PREPARE. | |
4010 | */ | |
4011 | static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu, | |
4012 | bool online) | |
4013 | { | |
4014 | int node = cpu_to_node(cpu); | |
4015 | int cpu_off = online ? -1 : cpu; | |
4016 | struct pool_workqueue *old_pwq = NULL, *pwq; | |
4017 | struct workqueue_attrs *target_attrs; | |
4018 | cpumask_t *cpumask; | |
4019 | ||
4020 | lockdep_assert_held(&wq_pool_mutex); | |
4021 | ||
4022 | if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND)) | |
4023 | return; | |
4024 | ||
4025 | /* | |
4026 | * We don't wanna alloc/free wq_attrs for each wq for each CPU. | |
4027 | * Let's use a preallocated one. The following buf is protected by | |
4028 | * CPU hotplug exclusion. | |
4029 | */ | |
4030 | target_attrs = wq_update_unbound_numa_attrs_buf; | |
4031 | cpumask = target_attrs->cpumask; | |
4032 | ||
4033 | mutex_lock(&wq->mutex); | |
d55262c4 TH |
4034 | if (wq->unbound_attrs->no_numa) |
4035 | goto out_unlock; | |
4c16bd32 TH |
4036 | |
4037 | copy_workqueue_attrs(target_attrs, wq->unbound_attrs); | |
4038 | pwq = unbound_pwq_by_node(wq, node); | |
4039 | ||
4040 | /* | |
4041 | * Let's determine what needs to be done. If the target cpumask is | |
4042 | * different from wq's, we need to compare it to @pwq's and create | |
4043 | * a new one if they don't match. If the target cpumask equals | |
534a3fbb | 4044 | * wq's, the default pwq should be used. |
4c16bd32 TH |
4045 | */ |
4046 | if (wq_calc_node_cpumask(wq->unbound_attrs, node, cpu_off, cpumask)) { | |
4047 | if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask)) | |
4048 | goto out_unlock; | |
4049 | } else { | |
534a3fbb | 4050 | goto use_dfl_pwq; |
4c16bd32 TH |
4051 | } |
4052 | ||
4053 | mutex_unlock(&wq->mutex); | |
4054 | ||
4055 | /* create a new pwq */ | |
4056 | pwq = alloc_unbound_pwq(wq, target_attrs); | |
4057 | if (!pwq) { | |
2d916033 FF |
4058 | pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n", |
4059 | wq->name); | |
77f300b1 DY |
4060 | mutex_lock(&wq->mutex); |
4061 | goto use_dfl_pwq; | |
4c16bd32 TH |
4062 | } |
4063 | ||
4064 | /* | |
4065 | * Install the new pwq. As this function is called only from CPU | |
4066 | * hotplug callbacks and applying a new attrs is wrapped with | |
4067 | * get/put_online_cpus(), @wq->unbound_attrs couldn't have changed | |
4068 | * inbetween. | |
4069 | */ | |
4070 | mutex_lock(&wq->mutex); | |
4071 | old_pwq = numa_pwq_tbl_install(wq, node, pwq); | |
4072 | goto out_unlock; | |
4073 | ||
4074 | use_dfl_pwq: | |
4075 | spin_lock_irq(&wq->dfl_pwq->pool->lock); | |
4076 | get_pwq(wq->dfl_pwq); | |
4077 | spin_unlock_irq(&wq->dfl_pwq->pool->lock); | |
4078 | old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq); | |
4079 | out_unlock: | |
4080 | mutex_unlock(&wq->mutex); | |
4081 | put_pwq_unlocked(old_pwq); | |
4082 | } | |
4083 | ||
30cdf249 | 4084 | static int alloc_and_link_pwqs(struct workqueue_struct *wq) |
0f900049 | 4085 | { |
49e3cf44 | 4086 | bool highpri = wq->flags & WQ_HIGHPRI; |
8a2b7538 | 4087 | int cpu, ret; |
30cdf249 TH |
4088 | |
4089 | if (!(wq->flags & WQ_UNBOUND)) { | |
420c0ddb TH |
4090 | wq->cpu_pwqs = alloc_percpu(struct pool_workqueue); |
4091 | if (!wq->cpu_pwqs) | |
30cdf249 TH |
4092 | return -ENOMEM; |
4093 | ||
4094 | for_each_possible_cpu(cpu) { | |
7fb98ea7 TH |
4095 | struct pool_workqueue *pwq = |
4096 | per_cpu_ptr(wq->cpu_pwqs, cpu); | |
7a62c2c8 | 4097 | struct worker_pool *cpu_pools = |
f02ae73a | 4098 | per_cpu(cpu_worker_pools, cpu); |
f3421797 | 4099 | |
f147f29e TH |
4100 | init_pwq(pwq, wq, &cpu_pools[highpri]); |
4101 | ||
4102 | mutex_lock(&wq->mutex); | |
1befcf30 | 4103 | link_pwq(pwq); |
f147f29e | 4104 | mutex_unlock(&wq->mutex); |
30cdf249 | 4105 | } |
9e8cd2f5 | 4106 | return 0; |
8a2b7538 TH |
4107 | } else if (wq->flags & __WQ_ORDERED) { |
4108 | ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]); | |
4109 | /* there should only be single pwq for ordering guarantee */ | |
4110 | WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node || | |
4111 | wq->pwqs.prev != &wq->dfl_pwq->pwqs_node), | |
4112 | "ordering guarantee broken for workqueue %s\n", wq->name); | |
4113 | return ret; | |
30cdf249 | 4114 | } else { |
9e8cd2f5 | 4115 | return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]); |
30cdf249 | 4116 | } |
0f900049 TH |
4117 | } |
4118 | ||
f3421797 TH |
4119 | static int wq_clamp_max_active(int max_active, unsigned int flags, |
4120 | const char *name) | |
b71ab8c2 | 4121 | { |
f3421797 TH |
4122 | int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE; |
4123 | ||
4124 | if (max_active < 1 || max_active > lim) | |
044c782c VI |
4125 | pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n", |
4126 | max_active, name, 1, lim); | |
b71ab8c2 | 4127 | |
f3421797 | 4128 | return clamp_val(max_active, 1, lim); |
b71ab8c2 TH |
4129 | } |
4130 | ||
b196be89 | 4131 | struct workqueue_struct *__alloc_workqueue_key(const char *fmt, |
d320c038 TH |
4132 | unsigned int flags, |
4133 | int max_active, | |
4134 | struct lock_class_key *key, | |
b196be89 | 4135 | const char *lock_name, ...) |
1da177e4 | 4136 | { |
df2d5ae4 | 4137 | size_t tbl_size = 0; |
ecf6881f | 4138 | va_list args; |
1da177e4 | 4139 | struct workqueue_struct *wq; |
49e3cf44 | 4140 | struct pool_workqueue *pwq; |
b196be89 | 4141 | |
cee22a15 VK |
4142 | /* see the comment above the definition of WQ_POWER_EFFICIENT */ |
4143 | if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient) | |
4144 | flags |= WQ_UNBOUND; | |
4145 | ||
ecf6881f | 4146 | /* allocate wq and format name */ |
df2d5ae4 TH |
4147 | if (flags & WQ_UNBOUND) |
4148 | tbl_size = wq_numa_tbl_len * sizeof(wq->numa_pwq_tbl[0]); | |
4149 | ||
4150 | wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL); | |
b196be89 | 4151 | if (!wq) |
d2c1d404 | 4152 | return NULL; |
b196be89 | 4153 | |
6029a918 TH |
4154 | if (flags & WQ_UNBOUND) { |
4155 | wq->unbound_attrs = alloc_workqueue_attrs(GFP_KERNEL); | |
4156 | if (!wq->unbound_attrs) | |
4157 | goto err_free_wq; | |
4158 | } | |
4159 | ||
ecf6881f TH |
4160 | va_start(args, lock_name); |
4161 | vsnprintf(wq->name, sizeof(wq->name), fmt, args); | |
b196be89 | 4162 | va_end(args); |
1da177e4 | 4163 | |
d320c038 | 4164 | max_active = max_active ?: WQ_DFL_ACTIVE; |
b196be89 | 4165 | max_active = wq_clamp_max_active(max_active, flags, wq->name); |
3af24433 | 4166 | |
b196be89 | 4167 | /* init wq */ |
97e37d7b | 4168 | wq->flags = flags; |
a0a1a5fd | 4169 | wq->saved_max_active = max_active; |
3c25a55d | 4170 | mutex_init(&wq->mutex); |
112202d9 | 4171 | atomic_set(&wq->nr_pwqs_to_flush, 0); |
30cdf249 | 4172 | INIT_LIST_HEAD(&wq->pwqs); |
73f53c4a TH |
4173 | INIT_LIST_HEAD(&wq->flusher_queue); |
4174 | INIT_LIST_HEAD(&wq->flusher_overflow); | |
493a1724 | 4175 | INIT_LIST_HEAD(&wq->maydays); |
502ca9d8 | 4176 | |
eb13ba87 | 4177 | lockdep_init_map(&wq->lockdep_map, lock_name, key, 0); |
cce1a165 | 4178 | INIT_LIST_HEAD(&wq->list); |
3af24433 | 4179 | |
30cdf249 | 4180 | if (alloc_and_link_pwqs(wq) < 0) |
d2c1d404 | 4181 | goto err_free_wq; |
1537663f | 4182 | |
493008a8 TH |
4183 | /* |
4184 | * Workqueues which may be used during memory reclaim should | |
4185 | * have a rescuer to guarantee forward progress. | |
4186 | */ | |
4187 | if (flags & WQ_MEM_RECLAIM) { | |
e22bee78 TH |
4188 | struct worker *rescuer; |
4189 | ||
d2c1d404 | 4190 | rescuer = alloc_worker(); |
e22bee78 | 4191 | if (!rescuer) |
d2c1d404 | 4192 | goto err_destroy; |
e22bee78 | 4193 | |
111c225a TH |
4194 | rescuer->rescue_wq = wq; |
4195 | rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", | |
b196be89 | 4196 | wq->name); |
d2c1d404 TH |
4197 | if (IS_ERR(rescuer->task)) { |
4198 | kfree(rescuer); | |
4199 | goto err_destroy; | |
4200 | } | |
e22bee78 | 4201 | |
d2c1d404 | 4202 | wq->rescuer = rescuer; |
14a40ffc | 4203 | rescuer->task->flags |= PF_NO_SETAFFINITY; |
e22bee78 | 4204 | wake_up_process(rescuer->task); |
3af24433 ON |
4205 | } |
4206 | ||
226223ab TH |
4207 | if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq)) |
4208 | goto err_destroy; | |
4209 | ||
a0a1a5fd | 4210 | /* |
68e13a67 LJ |
4211 | * wq_pool_mutex protects global freeze state and workqueues list. |
4212 | * Grab it, adjust max_active and add the new @wq to workqueues | |
4213 | * list. | |
a0a1a5fd | 4214 | */ |
68e13a67 | 4215 | mutex_lock(&wq_pool_mutex); |
a0a1a5fd | 4216 | |
a357fc03 | 4217 | mutex_lock(&wq->mutex); |
699ce097 TH |
4218 | for_each_pwq(pwq, wq) |
4219 | pwq_adjust_max_active(pwq); | |
a357fc03 | 4220 | mutex_unlock(&wq->mutex); |
a0a1a5fd | 4221 | |
1537663f | 4222 | list_add(&wq->list, &workqueues); |
a0a1a5fd | 4223 | |
68e13a67 | 4224 | mutex_unlock(&wq_pool_mutex); |
1537663f | 4225 | |
3af24433 | 4226 | return wq; |
d2c1d404 TH |
4227 | |
4228 | err_free_wq: | |
6029a918 | 4229 | free_workqueue_attrs(wq->unbound_attrs); |
d2c1d404 TH |
4230 | kfree(wq); |
4231 | return NULL; | |
4232 | err_destroy: | |
4233 | destroy_workqueue(wq); | |
4690c4ab | 4234 | return NULL; |
3af24433 | 4235 | } |
d320c038 | 4236 | EXPORT_SYMBOL_GPL(__alloc_workqueue_key); |
1da177e4 | 4237 | |
3af24433 ON |
4238 | /** |
4239 | * destroy_workqueue - safely terminate a workqueue | |
4240 | * @wq: target workqueue | |
4241 | * | |
4242 | * Safely destroy a workqueue. All work currently pending will be done first. | |
4243 | */ | |
4244 | void destroy_workqueue(struct workqueue_struct *wq) | |
4245 | { | |
49e3cf44 | 4246 | struct pool_workqueue *pwq; |
4c16bd32 | 4247 | int node; |
3af24433 | 4248 | |
9c5a2ba7 TH |
4249 | /* drain it before proceeding with destruction */ |
4250 | drain_workqueue(wq); | |
c8efcc25 | 4251 | |
6183c009 | 4252 | /* sanity checks */ |
b09f4fd3 | 4253 | mutex_lock(&wq->mutex); |
49e3cf44 | 4254 | for_each_pwq(pwq, wq) { |
6183c009 TH |
4255 | int i; |
4256 | ||
76af4d93 TH |
4257 | for (i = 0; i < WORK_NR_COLORS; i++) { |
4258 | if (WARN_ON(pwq->nr_in_flight[i])) { | |
b09f4fd3 | 4259 | mutex_unlock(&wq->mutex); |
6183c009 | 4260 | return; |
76af4d93 TH |
4261 | } |
4262 | } | |
4263 | ||
5c529597 | 4264 | if (WARN_ON((pwq != wq->dfl_pwq) && (pwq->refcnt > 1)) || |
8864b4e5 | 4265 | WARN_ON(pwq->nr_active) || |
76af4d93 | 4266 | WARN_ON(!list_empty(&pwq->delayed_works))) { |
b09f4fd3 | 4267 | mutex_unlock(&wq->mutex); |
6183c009 | 4268 | return; |
76af4d93 | 4269 | } |
6183c009 | 4270 | } |
b09f4fd3 | 4271 | mutex_unlock(&wq->mutex); |
6183c009 | 4272 | |
a0a1a5fd TH |
4273 | /* |
4274 | * wq list is used to freeze wq, remove from list after | |
4275 | * flushing is complete in case freeze races us. | |
4276 | */ | |
68e13a67 | 4277 | mutex_lock(&wq_pool_mutex); |
d2c1d404 | 4278 | list_del_init(&wq->list); |
68e13a67 | 4279 | mutex_unlock(&wq_pool_mutex); |
3af24433 | 4280 | |
226223ab TH |
4281 | workqueue_sysfs_unregister(wq); |
4282 | ||
493008a8 | 4283 | if (wq->rescuer) { |
e22bee78 | 4284 | kthread_stop(wq->rescuer->task); |
8d9df9f0 | 4285 | kfree(wq->rescuer); |
493008a8 | 4286 | wq->rescuer = NULL; |
e22bee78 TH |
4287 | } |
4288 | ||
8864b4e5 TH |
4289 | if (!(wq->flags & WQ_UNBOUND)) { |
4290 | /* | |
4291 | * The base ref is never dropped on per-cpu pwqs. Directly | |
4292 | * free the pwqs and wq. | |
4293 | */ | |
4294 | free_percpu(wq->cpu_pwqs); | |
4295 | kfree(wq); | |
4296 | } else { | |
4297 | /* | |
4298 | * We're the sole accessor of @wq at this point. Directly | |
4c16bd32 TH |
4299 | * access numa_pwq_tbl[] and dfl_pwq to put the base refs. |
4300 | * @wq will be freed when the last pwq is released. | |
8864b4e5 | 4301 | */ |
4c16bd32 TH |
4302 | for_each_node(node) { |
4303 | pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); | |
4304 | RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL); | |
4305 | put_pwq_unlocked(pwq); | |
4306 | } | |
4307 | ||
4308 | /* | |
4309 | * Put dfl_pwq. @wq may be freed any time after dfl_pwq is | |
4310 | * put. Don't access it afterwards. | |
4311 | */ | |
4312 | pwq = wq->dfl_pwq; | |
4313 | wq->dfl_pwq = NULL; | |
dce90d47 | 4314 | put_pwq_unlocked(pwq); |
29c91e99 | 4315 | } |
3af24433 ON |
4316 | } |
4317 | EXPORT_SYMBOL_GPL(destroy_workqueue); | |
4318 | ||
dcd989cb TH |
4319 | /** |
4320 | * workqueue_set_max_active - adjust max_active of a workqueue | |
4321 | * @wq: target workqueue | |
4322 | * @max_active: new max_active value. | |
4323 | * | |
4324 | * Set max_active of @wq to @max_active. | |
4325 | * | |
4326 | * CONTEXT: | |
4327 | * Don't call from IRQ context. | |
4328 | */ | |
4329 | void workqueue_set_max_active(struct workqueue_struct *wq, int max_active) | |
4330 | { | |
49e3cf44 | 4331 | struct pool_workqueue *pwq; |
dcd989cb | 4332 | |
8719dcea TH |
4333 | /* disallow meddling with max_active for ordered workqueues */ |
4334 | if (WARN_ON(wq->flags & __WQ_ORDERED)) | |
4335 | return; | |
4336 | ||
f3421797 | 4337 | max_active = wq_clamp_max_active(max_active, wq->flags, wq->name); |
dcd989cb | 4338 | |
a357fc03 | 4339 | mutex_lock(&wq->mutex); |
dcd989cb TH |
4340 | |
4341 | wq->saved_max_active = max_active; | |
4342 | ||
699ce097 TH |
4343 | for_each_pwq(pwq, wq) |
4344 | pwq_adjust_max_active(pwq); | |
93981800 | 4345 | |
a357fc03 | 4346 | mutex_unlock(&wq->mutex); |
15316ba8 | 4347 | } |
dcd989cb | 4348 | EXPORT_SYMBOL_GPL(workqueue_set_max_active); |
15316ba8 | 4349 | |
e6267616 TH |
4350 | /** |
4351 | * current_is_workqueue_rescuer - is %current workqueue rescuer? | |
4352 | * | |
4353 | * Determine whether %current is a workqueue rescuer. Can be used from | |
4354 | * work functions to determine whether it's being run off the rescuer task. | |
d185af30 YB |
4355 | * |
4356 | * Return: %true if %current is a workqueue rescuer. %false otherwise. | |
e6267616 TH |
4357 | */ |
4358 | bool current_is_workqueue_rescuer(void) | |
4359 | { | |
4360 | struct worker *worker = current_wq_worker(); | |
4361 | ||
6a092dfd | 4362 | return worker && worker->rescue_wq; |
e6267616 TH |
4363 | } |
4364 | ||
eef6a7d5 | 4365 | /** |
dcd989cb TH |
4366 | * workqueue_congested - test whether a workqueue is congested |
4367 | * @cpu: CPU in question | |
4368 | * @wq: target workqueue | |
eef6a7d5 | 4369 | * |
dcd989cb TH |
4370 | * Test whether @wq's cpu workqueue for @cpu is congested. There is |
4371 | * no synchronization around this function and the test result is | |
4372 | * unreliable and only useful as advisory hints or for debugging. | |
eef6a7d5 | 4373 | * |
d3251859 TH |
4374 | * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU. |
4375 | * Note that both per-cpu and unbound workqueues may be associated with | |
4376 | * multiple pool_workqueues which have separate congested states. A | |
4377 | * workqueue being congested on one CPU doesn't mean the workqueue is also | |
4378 | * contested on other CPUs / NUMA nodes. | |
4379 | * | |
d185af30 | 4380 | * Return: |
dcd989cb | 4381 | * %true if congested, %false otherwise. |
eef6a7d5 | 4382 | */ |
d84ff051 | 4383 | bool workqueue_congested(int cpu, struct workqueue_struct *wq) |
1da177e4 | 4384 | { |
7fb98ea7 | 4385 | struct pool_workqueue *pwq; |
76af4d93 TH |
4386 | bool ret; |
4387 | ||
88109453 | 4388 | rcu_read_lock_sched(); |
7fb98ea7 | 4389 | |
d3251859 TH |
4390 | if (cpu == WORK_CPU_UNBOUND) |
4391 | cpu = smp_processor_id(); | |
4392 | ||
7fb98ea7 TH |
4393 | if (!(wq->flags & WQ_UNBOUND)) |
4394 | pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); | |
4395 | else | |
df2d5ae4 | 4396 | pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); |
dcd989cb | 4397 | |
76af4d93 | 4398 | ret = !list_empty(&pwq->delayed_works); |
88109453 | 4399 | rcu_read_unlock_sched(); |
76af4d93 TH |
4400 | |
4401 | return ret; | |
1da177e4 | 4402 | } |
dcd989cb | 4403 | EXPORT_SYMBOL_GPL(workqueue_congested); |
1da177e4 | 4404 | |
dcd989cb TH |
4405 | /** |
4406 | * work_busy - test whether a work is currently pending or running | |
4407 | * @work: the work to be tested | |
4408 | * | |
4409 | * Test whether @work is currently pending or running. There is no | |
4410 | * synchronization around this function and the test result is | |
4411 | * unreliable and only useful as advisory hints or for debugging. | |
dcd989cb | 4412 | * |
d185af30 | 4413 | * Return: |
dcd989cb TH |
4414 | * OR'd bitmask of WORK_BUSY_* bits. |
4415 | */ | |
4416 | unsigned int work_busy(struct work_struct *work) | |
1da177e4 | 4417 | { |
fa1b54e6 | 4418 | struct worker_pool *pool; |
dcd989cb TH |
4419 | unsigned long flags; |
4420 | unsigned int ret = 0; | |
1da177e4 | 4421 | |
dcd989cb TH |
4422 | if (work_pending(work)) |
4423 | ret |= WORK_BUSY_PENDING; | |
1da177e4 | 4424 | |
fa1b54e6 TH |
4425 | local_irq_save(flags); |
4426 | pool = get_work_pool(work); | |
038366c5 | 4427 | if (pool) { |
fa1b54e6 | 4428 | spin_lock(&pool->lock); |
038366c5 LJ |
4429 | if (find_worker_executing_work(pool, work)) |
4430 | ret |= WORK_BUSY_RUNNING; | |
fa1b54e6 | 4431 | spin_unlock(&pool->lock); |
038366c5 | 4432 | } |
fa1b54e6 | 4433 | local_irq_restore(flags); |
1da177e4 | 4434 | |
dcd989cb | 4435 | return ret; |
1da177e4 | 4436 | } |
dcd989cb | 4437 | EXPORT_SYMBOL_GPL(work_busy); |
1da177e4 | 4438 | |
3d1cb205 TH |
4439 | /** |
4440 | * set_worker_desc - set description for the current work item | |
4441 | * @fmt: printf-style format string | |
4442 | * @...: arguments for the format string | |
4443 | * | |
4444 | * This function can be called by a running work function to describe what | |
4445 | * the work item is about. If the worker task gets dumped, this | |
4446 | * information will be printed out together to help debugging. The | |
4447 | * description can be at most WORKER_DESC_LEN including the trailing '\0'. | |
4448 | */ | |
4449 | void set_worker_desc(const char *fmt, ...) | |
4450 | { | |
4451 | struct worker *worker = current_wq_worker(); | |
4452 | va_list args; | |
4453 | ||
4454 | if (worker) { | |
4455 | va_start(args, fmt); | |
4456 | vsnprintf(worker->desc, sizeof(worker->desc), fmt, args); | |
4457 | va_end(args); | |
4458 | worker->desc_valid = true; | |
4459 | } | |
4460 | } | |
4461 | ||
4462 | /** | |
4463 | * print_worker_info - print out worker information and description | |
4464 | * @log_lvl: the log level to use when printing | |
4465 | * @task: target task | |
4466 | * | |
4467 | * If @task is a worker and currently executing a work item, print out the | |
4468 | * name of the workqueue being serviced and worker description set with | |
4469 | * set_worker_desc() by the currently executing work item. | |
4470 | * | |
4471 | * This function can be safely called on any task as long as the | |
4472 | * task_struct itself is accessible. While safe, this function isn't | |
4473 | * synchronized and may print out mixups or garbages of limited length. | |
4474 | */ | |
4475 | void print_worker_info(const char *log_lvl, struct task_struct *task) | |
4476 | { | |
4477 | work_func_t *fn = NULL; | |
4478 | char name[WQ_NAME_LEN] = { }; | |
4479 | char desc[WORKER_DESC_LEN] = { }; | |
4480 | struct pool_workqueue *pwq = NULL; | |
4481 | struct workqueue_struct *wq = NULL; | |
4482 | bool desc_valid = false; | |
4483 | struct worker *worker; | |
4484 | ||
4485 | if (!(task->flags & PF_WQ_WORKER)) | |
4486 | return; | |
4487 | ||
4488 | /* | |
4489 | * This function is called without any synchronization and @task | |
4490 | * could be in any state. Be careful with dereferences. | |
4491 | */ | |
4492 | worker = probe_kthread_data(task); | |
4493 | ||
4494 | /* | |
4495 | * Carefully copy the associated workqueue's workfn and name. Keep | |
4496 | * the original last '\0' in case the original contains garbage. | |
4497 | */ | |
4498 | probe_kernel_read(&fn, &worker->current_func, sizeof(fn)); | |
4499 | probe_kernel_read(&pwq, &worker->current_pwq, sizeof(pwq)); | |
4500 | probe_kernel_read(&wq, &pwq->wq, sizeof(wq)); | |
4501 | probe_kernel_read(name, wq->name, sizeof(name) - 1); | |
4502 | ||
4503 | /* copy worker description */ | |
4504 | probe_kernel_read(&desc_valid, &worker->desc_valid, sizeof(desc_valid)); | |
4505 | if (desc_valid) | |
4506 | probe_kernel_read(desc, worker->desc, sizeof(desc) - 1); | |
4507 | ||
4508 | if (fn || name[0] || desc[0]) { | |
2d916033 | 4509 | pr_info("%sWorkqueue: %s %pf", log_lvl, name, fn); |
3d1cb205 TH |
4510 | if (desc[0]) |
4511 | pr_cont(" (%s)", desc); | |
4512 | pr_cont("\n"); | |
4513 | } | |
4514 | } | |
4515 | ||
db7bccf4 TH |
4516 | /* |
4517 | * CPU hotplug. | |
4518 | * | |
e22bee78 | 4519 | * There are two challenges in supporting CPU hotplug. Firstly, there |
112202d9 | 4520 | * are a lot of assumptions on strong associations among work, pwq and |
706026c2 | 4521 | * pool which make migrating pending and scheduled works very |
e22bee78 | 4522 | * difficult to implement without impacting hot paths. Secondly, |
94cf58bb | 4523 | * worker pools serve mix of short, long and very long running works making |
e22bee78 TH |
4524 | * blocked draining impractical. |
4525 | * | |
24647570 | 4526 | * This is solved by allowing the pools to be disassociated from the CPU |
628c78e7 TH |
4527 | * running as an unbound one and allowing it to be reattached later if the |
4528 | * cpu comes back online. | |
db7bccf4 | 4529 | */ |
1da177e4 | 4530 | |
706026c2 | 4531 | static void wq_unbind_fn(struct work_struct *work) |
3af24433 | 4532 | { |
38db41d9 | 4533 | int cpu = smp_processor_id(); |
4ce62e9e | 4534 | struct worker_pool *pool; |
db7bccf4 | 4535 | struct worker *worker; |
a9ab775b | 4536 | int wi; |
3af24433 | 4537 | |
f02ae73a | 4538 | for_each_cpu_worker_pool(pool, cpu) { |
6183c009 | 4539 | WARN_ON_ONCE(cpu != smp_processor_id()); |
db7bccf4 | 4540 | |
bc3a1afc | 4541 | mutex_lock(&pool->manager_mutex); |
94cf58bb | 4542 | spin_lock_irq(&pool->lock); |
3af24433 | 4543 | |
94cf58bb | 4544 | /* |
bc3a1afc | 4545 | * We've blocked all manager operations. Make all workers |
94cf58bb TH |
4546 | * unbound and set DISASSOCIATED. Before this, all workers |
4547 | * except for the ones which are still executing works from | |
4548 | * before the last CPU down must be on the cpu. After | |
4549 | * this, they may become diasporas. | |
4550 | */ | |
a9ab775b | 4551 | for_each_pool_worker(worker, wi, pool) |
c9e7cf27 | 4552 | worker->flags |= WORKER_UNBOUND; |
06ba38a9 | 4553 | |
24647570 | 4554 | pool->flags |= POOL_DISASSOCIATED; |
f2d5a0ee | 4555 | |
94cf58bb | 4556 | spin_unlock_irq(&pool->lock); |
bc3a1afc | 4557 | mutex_unlock(&pool->manager_mutex); |
628c78e7 | 4558 | |
eb283428 LJ |
4559 | /* |
4560 | * Call schedule() so that we cross rq->lock and thus can | |
4561 | * guarantee sched callbacks see the %WORKER_UNBOUND flag. | |
4562 | * This is necessary as scheduler callbacks may be invoked | |
4563 | * from other cpus. | |
4564 | */ | |
4565 | schedule(); | |
06ba38a9 | 4566 | |
eb283428 LJ |
4567 | /* |
4568 | * Sched callbacks are disabled now. Zap nr_running. | |
4569 | * After this, nr_running stays zero and need_more_worker() | |
4570 | * and keep_working() are always true as long as the | |
4571 | * worklist is not empty. This pool now behaves as an | |
4572 | * unbound (in terms of concurrency management) pool which | |
4573 | * are served by workers tied to the pool. | |
4574 | */ | |
e19e397a | 4575 | atomic_set(&pool->nr_running, 0); |
eb283428 LJ |
4576 | |
4577 | /* | |
4578 | * With concurrency management just turned off, a busy | |
4579 | * worker blocking could lead to lengthy stalls. Kick off | |
4580 | * unbound chain execution of currently pending work items. | |
4581 | */ | |
4582 | spin_lock_irq(&pool->lock); | |
4583 | wake_up_worker(pool); | |
4584 | spin_unlock_irq(&pool->lock); | |
4585 | } | |
3af24433 | 4586 | } |
3af24433 | 4587 | |
bd7c089e TH |
4588 | /** |
4589 | * rebind_workers - rebind all workers of a pool to the associated CPU | |
4590 | * @pool: pool of interest | |
4591 | * | |
a9ab775b | 4592 | * @pool->cpu is coming online. Rebind all workers to the CPU. |
bd7c089e TH |
4593 | */ |
4594 | static void rebind_workers(struct worker_pool *pool) | |
4595 | { | |
a9ab775b TH |
4596 | struct worker *worker; |
4597 | int wi; | |
bd7c089e TH |
4598 | |
4599 | lockdep_assert_held(&pool->manager_mutex); | |
bd7c089e | 4600 | |
a9ab775b TH |
4601 | /* |
4602 | * Restore CPU affinity of all workers. As all idle workers should | |
4603 | * be on the run-queue of the associated CPU before any local | |
4604 | * wake-ups for concurrency management happen, restore CPU affinty | |
4605 | * of all workers first and then clear UNBOUND. As we're called | |
4606 | * from CPU_ONLINE, the following shouldn't fail. | |
4607 | */ | |
4608 | for_each_pool_worker(worker, wi, pool) | |
4609 | WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, | |
4610 | pool->attrs->cpumask) < 0); | |
bd7c089e | 4611 | |
a9ab775b | 4612 | spin_lock_irq(&pool->lock); |
bd7c089e | 4613 | |
a9ab775b TH |
4614 | for_each_pool_worker(worker, wi, pool) { |
4615 | unsigned int worker_flags = worker->flags; | |
bd7c089e TH |
4616 | |
4617 | /* | |
a9ab775b TH |
4618 | * A bound idle worker should actually be on the runqueue |
4619 | * of the associated CPU for local wake-ups targeting it to | |
4620 | * work. Kick all idle workers so that they migrate to the | |
4621 | * associated CPU. Doing this in the same loop as | |
4622 | * replacing UNBOUND with REBOUND is safe as no worker will | |
4623 | * be bound before @pool->lock is released. | |
bd7c089e | 4624 | */ |
a9ab775b TH |
4625 | if (worker_flags & WORKER_IDLE) |
4626 | wake_up_process(worker->task); | |
bd7c089e | 4627 | |
a9ab775b TH |
4628 | /* |
4629 | * We want to clear UNBOUND but can't directly call | |
4630 | * worker_clr_flags() or adjust nr_running. Atomically | |
4631 | * replace UNBOUND with another NOT_RUNNING flag REBOUND. | |
4632 | * @worker will clear REBOUND using worker_clr_flags() when | |
4633 | * it initiates the next execution cycle thus restoring | |
4634 | * concurrency management. Note that when or whether | |
4635 | * @worker clears REBOUND doesn't affect correctness. | |
4636 | * | |
4637 | * ACCESS_ONCE() is necessary because @worker->flags may be | |
4638 | * tested without holding any lock in | |
4639 | * wq_worker_waking_up(). Without it, NOT_RUNNING test may | |
4640 | * fail incorrectly leading to premature concurrency | |
4641 | * management operations. | |
4642 | */ | |
4643 | WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND)); | |
4644 | worker_flags |= WORKER_REBOUND; | |
4645 | worker_flags &= ~WORKER_UNBOUND; | |
4646 | ACCESS_ONCE(worker->flags) = worker_flags; | |
bd7c089e | 4647 | } |
a9ab775b TH |
4648 | |
4649 | spin_unlock_irq(&pool->lock); | |
bd7c089e TH |
4650 | } |
4651 | ||
7dbc725e TH |
4652 | /** |
4653 | * restore_unbound_workers_cpumask - restore cpumask of unbound workers | |
4654 | * @pool: unbound pool of interest | |
4655 | * @cpu: the CPU which is coming up | |
4656 | * | |
4657 | * An unbound pool may end up with a cpumask which doesn't have any online | |
4658 | * CPUs. When a worker of such pool get scheduled, the scheduler resets | |
4659 | * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any | |
4660 | * online CPU before, cpus_allowed of all its workers should be restored. | |
4661 | */ | |
4662 | static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu) | |
4663 | { | |
4664 | static cpumask_t cpumask; | |
4665 | struct worker *worker; | |
4666 | int wi; | |
4667 | ||
4668 | lockdep_assert_held(&pool->manager_mutex); | |
4669 | ||
4670 | /* is @cpu allowed for @pool? */ | |
4671 | if (!cpumask_test_cpu(cpu, pool->attrs->cpumask)) | |
4672 | return; | |
4673 | ||
4674 | /* is @cpu the only online CPU? */ | |
4675 | cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask); | |
4676 | if (cpumask_weight(&cpumask) != 1) | |
4677 | return; | |
4678 | ||
4679 | /* as we're called from CPU_ONLINE, the following shouldn't fail */ | |
4680 | for_each_pool_worker(worker, wi, pool) | |
4681 | WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, | |
4682 | pool->attrs->cpumask) < 0); | |
4683 | } | |
4684 | ||
8db25e78 TH |
4685 | /* |
4686 | * Workqueues should be brought up before normal priority CPU notifiers. | |
4687 | * This will be registered high priority CPU notifier. | |
4688 | */ | |
0db0628d | 4689 | static int workqueue_cpu_up_callback(struct notifier_block *nfb, |
8db25e78 TH |
4690 | unsigned long action, |
4691 | void *hcpu) | |
3af24433 | 4692 | { |
d84ff051 | 4693 | int cpu = (unsigned long)hcpu; |
4ce62e9e | 4694 | struct worker_pool *pool; |
4c16bd32 | 4695 | struct workqueue_struct *wq; |
7dbc725e | 4696 | int pi; |
3ce63377 | 4697 | |
8db25e78 | 4698 | switch (action & ~CPU_TASKS_FROZEN) { |
3af24433 | 4699 | case CPU_UP_PREPARE: |
f02ae73a | 4700 | for_each_cpu_worker_pool(pool, cpu) { |
3ce63377 TH |
4701 | if (pool->nr_workers) |
4702 | continue; | |
ebf44d16 | 4703 | if (create_and_start_worker(pool) < 0) |
3ce63377 | 4704 | return NOTIFY_BAD; |
3af24433 | 4705 | } |
8db25e78 | 4706 | break; |
3af24433 | 4707 | |
db7bccf4 TH |
4708 | case CPU_DOWN_FAILED: |
4709 | case CPU_ONLINE: | |
68e13a67 | 4710 | mutex_lock(&wq_pool_mutex); |
7dbc725e TH |
4711 | |
4712 | for_each_pool(pool, pi) { | |
bc3a1afc | 4713 | mutex_lock(&pool->manager_mutex); |
94cf58bb | 4714 | |
7dbc725e TH |
4715 | if (pool->cpu == cpu) { |
4716 | spin_lock_irq(&pool->lock); | |
4717 | pool->flags &= ~POOL_DISASSOCIATED; | |
4718 | spin_unlock_irq(&pool->lock); | |
a9ab775b | 4719 | |
7dbc725e TH |
4720 | rebind_workers(pool); |
4721 | } else if (pool->cpu < 0) { | |
4722 | restore_unbound_workers_cpumask(pool, cpu); | |
4723 | } | |
94cf58bb | 4724 | |
bc3a1afc | 4725 | mutex_unlock(&pool->manager_mutex); |
94cf58bb | 4726 | } |
7dbc725e | 4727 | |
4c16bd32 TH |
4728 | /* update NUMA affinity of unbound workqueues */ |
4729 | list_for_each_entry(wq, &workqueues, list) | |
4730 | wq_update_unbound_numa(wq, cpu, true); | |
4731 | ||
68e13a67 | 4732 | mutex_unlock(&wq_pool_mutex); |
db7bccf4 | 4733 | break; |
00dfcaf7 | 4734 | } |
65758202 TH |
4735 | return NOTIFY_OK; |
4736 | } | |
4737 | ||
4738 | /* | |
4739 | * Workqueues should be brought down after normal priority CPU notifiers. | |
4740 | * This will be registered as low priority CPU notifier. | |
4741 | */ | |
0db0628d | 4742 | static int workqueue_cpu_down_callback(struct notifier_block *nfb, |
65758202 TH |
4743 | unsigned long action, |
4744 | void *hcpu) | |
4745 | { | |
d84ff051 | 4746 | int cpu = (unsigned long)hcpu; |
8db25e78 | 4747 | struct work_struct unbind_work; |
4c16bd32 | 4748 | struct workqueue_struct *wq; |
8db25e78 | 4749 | |
65758202 TH |
4750 | switch (action & ~CPU_TASKS_FROZEN) { |
4751 | case CPU_DOWN_PREPARE: | |
4c16bd32 | 4752 | /* unbinding per-cpu workers should happen on the local CPU */ |
706026c2 | 4753 | INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn); |
7635d2fd | 4754 | queue_work_on(cpu, system_highpri_wq, &unbind_work); |
4c16bd32 TH |
4755 | |
4756 | /* update NUMA affinity of unbound workqueues */ | |
4757 | mutex_lock(&wq_pool_mutex); | |
4758 | list_for_each_entry(wq, &workqueues, list) | |
4759 | wq_update_unbound_numa(wq, cpu, false); | |
4760 | mutex_unlock(&wq_pool_mutex); | |
4761 | ||
4762 | /* wait for per-cpu unbinding to finish */ | |
8db25e78 | 4763 | flush_work(&unbind_work); |
440a1136 | 4764 | destroy_work_on_stack(&unbind_work); |
8db25e78 | 4765 | break; |
65758202 TH |
4766 | } |
4767 | return NOTIFY_OK; | |
4768 | } | |
4769 | ||
2d3854a3 | 4770 | #ifdef CONFIG_SMP |
8ccad40d | 4771 | |
2d3854a3 | 4772 | struct work_for_cpu { |
ed48ece2 | 4773 | struct work_struct work; |
2d3854a3 RR |
4774 | long (*fn)(void *); |
4775 | void *arg; | |
4776 | long ret; | |
4777 | }; | |
4778 | ||
ed48ece2 | 4779 | static void work_for_cpu_fn(struct work_struct *work) |
2d3854a3 | 4780 | { |
ed48ece2 TH |
4781 | struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work); |
4782 | ||
2d3854a3 RR |
4783 | wfc->ret = wfc->fn(wfc->arg); |
4784 | } | |
4785 | ||
4786 | /** | |
4787 | * work_on_cpu - run a function in user context on a particular cpu | |
4788 | * @cpu: the cpu to run on | |
4789 | * @fn: the function to run | |
4790 | * @arg: the function arg | |
4791 | * | |
31ad9081 | 4792 | * It is up to the caller to ensure that the cpu doesn't go offline. |
6b44003e | 4793 | * The caller must not hold any locks which would prevent @fn from completing. |
d185af30 YB |
4794 | * |
4795 | * Return: The value @fn returns. | |
2d3854a3 | 4796 | */ |
d84ff051 | 4797 | long work_on_cpu(int cpu, long (*fn)(void *), void *arg) |
2d3854a3 | 4798 | { |
ed48ece2 | 4799 | struct work_for_cpu wfc = { .fn = fn, .arg = arg }; |
6b44003e | 4800 | |
ed48ece2 TH |
4801 | INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn); |
4802 | schedule_work_on(cpu, &wfc.work); | |
12997d1a | 4803 | flush_work(&wfc.work); |
440a1136 | 4804 | destroy_work_on_stack(&wfc.work); |
2d3854a3 RR |
4805 | return wfc.ret; |
4806 | } | |
4807 | EXPORT_SYMBOL_GPL(work_on_cpu); | |
4808 | #endif /* CONFIG_SMP */ | |
4809 | ||
a0a1a5fd TH |
4810 | #ifdef CONFIG_FREEZER |
4811 | ||
4812 | /** | |
4813 | * freeze_workqueues_begin - begin freezing workqueues | |
4814 | * | |
58a69cb4 | 4815 | * Start freezing workqueues. After this function returns, all freezable |
c5aa87bb | 4816 | * workqueues will queue new works to their delayed_works list instead of |
706026c2 | 4817 | * pool->worklist. |
a0a1a5fd TH |
4818 | * |
4819 | * CONTEXT: | |
a357fc03 | 4820 | * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. |
a0a1a5fd TH |
4821 | */ |
4822 | void freeze_workqueues_begin(void) | |
4823 | { | |
17116969 | 4824 | struct worker_pool *pool; |
24b8a847 TH |
4825 | struct workqueue_struct *wq; |
4826 | struct pool_workqueue *pwq; | |
611c92a0 | 4827 | int pi; |
a0a1a5fd | 4828 | |
68e13a67 | 4829 | mutex_lock(&wq_pool_mutex); |
a0a1a5fd | 4830 | |
6183c009 | 4831 | WARN_ON_ONCE(workqueue_freezing); |
a0a1a5fd TH |
4832 | workqueue_freezing = true; |
4833 | ||
24b8a847 | 4834 | /* set FREEZING */ |
611c92a0 | 4835 | for_each_pool(pool, pi) { |
5bcab335 | 4836 | spin_lock_irq(&pool->lock); |
17116969 TH |
4837 | WARN_ON_ONCE(pool->flags & POOL_FREEZING); |
4838 | pool->flags |= POOL_FREEZING; | |
5bcab335 | 4839 | spin_unlock_irq(&pool->lock); |
24b8a847 | 4840 | } |
a0a1a5fd | 4841 | |
24b8a847 | 4842 | list_for_each_entry(wq, &workqueues, list) { |
a357fc03 | 4843 | mutex_lock(&wq->mutex); |
699ce097 TH |
4844 | for_each_pwq(pwq, wq) |
4845 | pwq_adjust_max_active(pwq); | |
a357fc03 | 4846 | mutex_unlock(&wq->mutex); |
a0a1a5fd | 4847 | } |
5bcab335 | 4848 | |
68e13a67 | 4849 | mutex_unlock(&wq_pool_mutex); |
a0a1a5fd TH |
4850 | } |
4851 | ||
4852 | /** | |
58a69cb4 | 4853 | * freeze_workqueues_busy - are freezable workqueues still busy? |
a0a1a5fd TH |
4854 | * |
4855 | * Check whether freezing is complete. This function must be called | |
4856 | * between freeze_workqueues_begin() and thaw_workqueues(). | |
4857 | * | |
4858 | * CONTEXT: | |
68e13a67 | 4859 | * Grabs and releases wq_pool_mutex. |
a0a1a5fd | 4860 | * |
d185af30 | 4861 | * Return: |
58a69cb4 TH |
4862 | * %true if some freezable workqueues are still busy. %false if freezing |
4863 | * is complete. | |
a0a1a5fd TH |
4864 | */ |
4865 | bool freeze_workqueues_busy(void) | |
4866 | { | |
a0a1a5fd | 4867 | bool busy = false; |
24b8a847 TH |
4868 | struct workqueue_struct *wq; |
4869 | struct pool_workqueue *pwq; | |
a0a1a5fd | 4870 | |
68e13a67 | 4871 | mutex_lock(&wq_pool_mutex); |
a0a1a5fd | 4872 | |
6183c009 | 4873 | WARN_ON_ONCE(!workqueue_freezing); |
a0a1a5fd | 4874 | |
24b8a847 TH |
4875 | list_for_each_entry(wq, &workqueues, list) { |
4876 | if (!(wq->flags & WQ_FREEZABLE)) | |
4877 | continue; | |
a0a1a5fd TH |
4878 | /* |
4879 | * nr_active is monotonically decreasing. It's safe | |
4880 | * to peek without lock. | |
4881 | */ | |
88109453 | 4882 | rcu_read_lock_sched(); |
24b8a847 | 4883 | for_each_pwq(pwq, wq) { |
6183c009 | 4884 | WARN_ON_ONCE(pwq->nr_active < 0); |
112202d9 | 4885 | if (pwq->nr_active) { |
a0a1a5fd | 4886 | busy = true; |
88109453 | 4887 | rcu_read_unlock_sched(); |
a0a1a5fd TH |
4888 | goto out_unlock; |
4889 | } | |
4890 | } | |
88109453 | 4891 | rcu_read_unlock_sched(); |
a0a1a5fd TH |
4892 | } |
4893 | out_unlock: | |
68e13a67 | 4894 | mutex_unlock(&wq_pool_mutex); |
a0a1a5fd TH |
4895 | return busy; |
4896 | } | |
4897 | ||
4898 | /** | |
4899 | * thaw_workqueues - thaw workqueues | |
4900 | * | |
4901 | * Thaw workqueues. Normal queueing is restored and all collected | |
706026c2 | 4902 | * frozen works are transferred to their respective pool worklists. |
a0a1a5fd TH |
4903 | * |
4904 | * CONTEXT: | |
a357fc03 | 4905 | * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. |
a0a1a5fd TH |
4906 | */ |
4907 | void thaw_workqueues(void) | |
4908 | { | |
24b8a847 TH |
4909 | struct workqueue_struct *wq; |
4910 | struct pool_workqueue *pwq; | |
4911 | struct worker_pool *pool; | |
611c92a0 | 4912 | int pi; |
a0a1a5fd | 4913 | |
68e13a67 | 4914 | mutex_lock(&wq_pool_mutex); |
a0a1a5fd TH |
4915 | |
4916 | if (!workqueue_freezing) | |
4917 | goto out_unlock; | |
4918 | ||
24b8a847 | 4919 | /* clear FREEZING */ |
611c92a0 | 4920 | for_each_pool(pool, pi) { |
5bcab335 | 4921 | spin_lock_irq(&pool->lock); |
24b8a847 TH |
4922 | WARN_ON_ONCE(!(pool->flags & POOL_FREEZING)); |
4923 | pool->flags &= ~POOL_FREEZING; | |
5bcab335 | 4924 | spin_unlock_irq(&pool->lock); |
24b8a847 | 4925 | } |
8b03ae3c | 4926 | |
24b8a847 TH |
4927 | /* restore max_active and repopulate worklist */ |
4928 | list_for_each_entry(wq, &workqueues, list) { | |
a357fc03 | 4929 | mutex_lock(&wq->mutex); |
699ce097 TH |
4930 | for_each_pwq(pwq, wq) |
4931 | pwq_adjust_max_active(pwq); | |
a357fc03 | 4932 | mutex_unlock(&wq->mutex); |
a0a1a5fd TH |
4933 | } |
4934 | ||
4935 | workqueue_freezing = false; | |
4936 | out_unlock: | |
68e13a67 | 4937 | mutex_unlock(&wq_pool_mutex); |
a0a1a5fd TH |
4938 | } |
4939 | #endif /* CONFIG_FREEZER */ | |
4940 | ||
bce90380 TH |
4941 | static void __init wq_numa_init(void) |
4942 | { | |
4943 | cpumask_var_t *tbl; | |
4944 | int node, cpu; | |
4945 | ||
4946 | /* determine NUMA pwq table len - highest node id + 1 */ | |
4947 | for_each_node(node) | |
4948 | wq_numa_tbl_len = max(wq_numa_tbl_len, node + 1); | |
4949 | ||
4950 | if (num_possible_nodes() <= 1) | |
4951 | return; | |
4952 | ||
d55262c4 TH |
4953 | if (wq_disable_numa) { |
4954 | pr_info("workqueue: NUMA affinity support disabled\n"); | |
4955 | return; | |
4956 | } | |
4957 | ||
4c16bd32 TH |
4958 | wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(GFP_KERNEL); |
4959 | BUG_ON(!wq_update_unbound_numa_attrs_buf); | |
4960 | ||
bce90380 TH |
4961 | /* |
4962 | * We want masks of possible CPUs of each node which isn't readily | |
4963 | * available. Build one from cpu_to_node() which should have been | |
4964 | * fully initialized by now. | |
4965 | */ | |
4966 | tbl = kzalloc(wq_numa_tbl_len * sizeof(tbl[0]), GFP_KERNEL); | |
4967 | BUG_ON(!tbl); | |
4968 | ||
4969 | for_each_node(node) | |
1be0c25d TH |
4970 | BUG_ON(!alloc_cpumask_var_node(&tbl[node], GFP_KERNEL, |
4971 | node_online(node) ? node : NUMA_NO_NODE)); | |
bce90380 TH |
4972 | |
4973 | for_each_possible_cpu(cpu) { | |
4974 | node = cpu_to_node(cpu); | |
4975 | if (WARN_ON(node == NUMA_NO_NODE)) { | |
4976 | pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu); | |
4977 | /* happens iff arch is bonkers, let's just proceed */ | |
4978 | return; | |
4979 | } | |
4980 | cpumask_set_cpu(cpu, tbl[node]); | |
4981 | } | |
4982 | ||
4983 | wq_numa_possible_cpumask = tbl; | |
4984 | wq_numa_enabled = true; | |
4985 | } | |
4986 | ||
6ee0578b | 4987 | static int __init init_workqueues(void) |
1da177e4 | 4988 | { |
7a4e344c TH |
4989 | int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL }; |
4990 | int i, cpu; | |
c34056a3 | 4991 | |
e904e6c2 TH |
4992 | WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long)); |
4993 | ||
4994 | pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC); | |
4995 | ||
65758202 | 4996 | cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP); |
a5b4e57d | 4997 | hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN); |
8b03ae3c | 4998 | |
bce90380 TH |
4999 | wq_numa_init(); |
5000 | ||
706026c2 | 5001 | /* initialize CPU pools */ |
29c91e99 | 5002 | for_each_possible_cpu(cpu) { |
4ce62e9e | 5003 | struct worker_pool *pool; |
8b03ae3c | 5004 | |
7a4e344c | 5005 | i = 0; |
f02ae73a | 5006 | for_each_cpu_worker_pool(pool, cpu) { |
7a4e344c | 5007 | BUG_ON(init_worker_pool(pool)); |
ec22ca5e | 5008 | pool->cpu = cpu; |
29c91e99 | 5009 | cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu)); |
7a4e344c | 5010 | pool->attrs->nice = std_nice[i++]; |
f3f90ad4 | 5011 | pool->node = cpu_to_node(cpu); |
7a4e344c | 5012 | |
9daf9e67 | 5013 | /* alloc pool ID */ |
68e13a67 | 5014 | mutex_lock(&wq_pool_mutex); |
9daf9e67 | 5015 | BUG_ON(worker_pool_assign_id(pool)); |
68e13a67 | 5016 | mutex_unlock(&wq_pool_mutex); |
4ce62e9e | 5017 | } |
8b03ae3c TH |
5018 | } |
5019 | ||
e22bee78 | 5020 | /* create the initial worker */ |
29c91e99 | 5021 | for_each_online_cpu(cpu) { |
4ce62e9e | 5022 | struct worker_pool *pool; |
e22bee78 | 5023 | |
f02ae73a | 5024 | for_each_cpu_worker_pool(pool, cpu) { |
29c91e99 | 5025 | pool->flags &= ~POOL_DISASSOCIATED; |
ebf44d16 | 5026 | BUG_ON(create_and_start_worker(pool) < 0); |
4ce62e9e | 5027 | } |
e22bee78 TH |
5028 | } |
5029 | ||
8a2b7538 | 5030 | /* create default unbound and ordered wq attrs */ |
29c91e99 TH |
5031 | for (i = 0; i < NR_STD_WORKER_POOLS; i++) { |
5032 | struct workqueue_attrs *attrs; | |
5033 | ||
5034 | BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL))); | |
29c91e99 | 5035 | attrs->nice = std_nice[i]; |
29c91e99 | 5036 | unbound_std_wq_attrs[i] = attrs; |
8a2b7538 TH |
5037 | |
5038 | /* | |
5039 | * An ordered wq should have only one pwq as ordering is | |
5040 | * guaranteed by max_active which is enforced by pwqs. | |
5041 | * Turn off NUMA so that dfl_pwq is used for all nodes. | |
5042 | */ | |
5043 | BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL))); | |
5044 | attrs->nice = std_nice[i]; | |
5045 | attrs->no_numa = true; | |
5046 | ordered_wq_attrs[i] = attrs; | |
29c91e99 TH |
5047 | } |
5048 | ||
d320c038 | 5049 | system_wq = alloc_workqueue("events", 0, 0); |
1aabe902 | 5050 | system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0); |
d320c038 | 5051 | system_long_wq = alloc_workqueue("events_long", 0, 0); |
f3421797 TH |
5052 | system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND, |
5053 | WQ_UNBOUND_MAX_ACTIVE); | |
24d51add TH |
5054 | system_freezable_wq = alloc_workqueue("events_freezable", |
5055 | WQ_FREEZABLE, 0); | |
0668106c VK |
5056 | system_power_efficient_wq = alloc_workqueue("events_power_efficient", |
5057 | WQ_POWER_EFFICIENT, 0); | |
5058 | system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient", | |
5059 | WQ_FREEZABLE | WQ_POWER_EFFICIENT, | |
5060 | 0); | |
1aabe902 | 5061 | BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq || |
0668106c VK |
5062 | !system_unbound_wq || !system_freezable_wq || |
5063 | !system_power_efficient_wq || | |
5064 | !system_freezable_power_efficient_wq); | |
6ee0578b | 5065 | return 0; |
1da177e4 | 5066 | } |
6ee0578b | 5067 | early_initcall(init_workqueues); |