David Howells | 07fe7cb | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 1 | /* Worker thread pool for slow items, such as filesystem lookups or mkdirs |
| 2 | * |
| 3 | * Copyright (C) 2008 Red Hat, Inc. All Rights Reserved. |
| 4 | * Written by David Howells (dhowells@redhat.com) |
| 5 | * |
| 6 | * This program is free software; you can redistribute it and/or |
| 7 | * modify it under the terms of the GNU General Public Licence |
| 8 | * as published by the Free Software Foundation; either version |
| 9 | * 2 of the Licence, or (at your option) any later version. |
David Howells | 8f0aa2f | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 10 | * |
| 11 | * See Documentation/slow-work.txt |
David Howells | 07fe7cb | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 12 | */ |
| 13 | |
| 14 | #include <linux/module.h> |
| 15 | #include <linux/slow-work.h> |
| 16 | #include <linux/kthread.h> |
| 17 | #include <linux/freezer.h> |
| 18 | #include <linux/wait.h> |
David Howells | f13a48b | 2009-12-01 15:36:11 +0000 | [diff] [blame] | 19 | #include <linux/debugfs.h> |
David Howells | 8fba10a | 2009-11-19 18:10:51 +0000 | [diff] [blame] | 20 | #include "slow-work.h" |
David Howells | 3d7a641 | 2009-11-19 18:10:23 +0000 | [diff] [blame] | 21 | |
David Howells | 109d927 | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 22 | static void slow_work_cull_timeout(unsigned long); |
| 23 | static void slow_work_oom_timeout(unsigned long); |
| 24 | |
David Howells | 12e22c5 | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 25 | #ifdef CONFIG_SYSCTL |
Alexey Dobriyan | 8d65af7 | 2009-09-23 15:57:19 -0700 | [diff] [blame] | 26 | static int slow_work_min_threads_sysctl(struct ctl_table *, int, |
David Howells | 12e22c5 | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 27 | void __user *, size_t *, loff_t *); |
| 28 | |
Alexey Dobriyan | 8d65af7 | 2009-09-23 15:57:19 -0700 | [diff] [blame] | 29 | static int slow_work_max_threads_sysctl(struct ctl_table *, int , |
David Howells | 12e22c5 | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 30 | void __user *, size_t *, loff_t *); |
| 31 | #endif |
| 32 | |
David Howells | 07fe7cb | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 33 | /* |
| 34 | * The pool of threads has at least min threads in it as long as someone is |
| 35 | * using the facility, and may have as many as max. |
| 36 | * |
| 37 | * A portion of the pool may be processing very slow operations. |
| 38 | */ |
| 39 | static unsigned slow_work_min_threads = 2; |
| 40 | static unsigned slow_work_max_threads = 4; |
| 41 | static unsigned vslow_work_proportion = 50; /* % of threads that may process |
| 42 | * very slow work */ |
David Howells | 12e22c5 | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 43 | |
| 44 | #ifdef CONFIG_SYSCTL |
| 45 | static const int slow_work_min_min_threads = 2; |
David Howells | 3d7a641 | 2009-11-19 18:10:23 +0000 | [diff] [blame] | 46 | static int slow_work_max_max_threads = SLOW_WORK_THREAD_LIMIT; |
David Howells | 12e22c5 | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 47 | static const int slow_work_min_vslow = 1; |
| 48 | static const int slow_work_max_vslow = 99; |
| 49 | |
| 50 | ctl_table slow_work_sysctls[] = { |
| 51 | { |
David Howells | 12e22c5 | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 52 | .procname = "min-threads", |
| 53 | .data = &slow_work_min_threads, |
| 54 | .maxlen = sizeof(unsigned), |
| 55 | .mode = 0644, |
| 56 | .proc_handler = slow_work_min_threads_sysctl, |
| 57 | .extra1 = (void *) &slow_work_min_min_threads, |
| 58 | .extra2 = &slow_work_max_threads, |
| 59 | }, |
| 60 | { |
David Howells | 12e22c5 | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 61 | .procname = "max-threads", |
| 62 | .data = &slow_work_max_threads, |
| 63 | .maxlen = sizeof(unsigned), |
| 64 | .mode = 0644, |
| 65 | .proc_handler = slow_work_max_threads_sysctl, |
| 66 | .extra1 = &slow_work_min_threads, |
| 67 | .extra2 = (void *) &slow_work_max_max_threads, |
| 68 | }, |
| 69 | { |
David Howells | 12e22c5 | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 70 | .procname = "vslow-percentage", |
| 71 | .data = &vslow_work_proportion, |
| 72 | .maxlen = sizeof(unsigned), |
| 73 | .mode = 0644, |
Eric W. Biederman | 6d45611 | 2009-11-16 03:11:48 -0800 | [diff] [blame] | 74 | .proc_handler = proc_dointvec_minmax, |
David Howells | 12e22c5 | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 75 | .extra1 = (void *) &slow_work_min_vslow, |
| 76 | .extra2 = (void *) &slow_work_max_vslow, |
| 77 | }, |
Eric W. Biederman | 5699230 | 2009-11-05 15:38:40 -0800 | [diff] [blame] | 78 | {} |
David Howells | 12e22c5 | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 79 | }; |
| 80 | #endif |
| 81 | |
| 82 | /* |
| 83 | * The active state of the thread pool |
| 84 | */ |
David Howells | 07fe7cb | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 85 | static atomic_t slow_work_thread_count; |
| 86 | static atomic_t vslow_work_executing_count; |
| 87 | |
David Howells | 109d927 | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 88 | static bool slow_work_may_not_start_new_thread; |
| 89 | static bool slow_work_cull; /* cull a thread due to lack of activity */ |
| 90 | static DEFINE_TIMER(slow_work_cull_timer, slow_work_cull_timeout, 0, 0); |
| 91 | static DEFINE_TIMER(slow_work_oom_timer, slow_work_oom_timeout, 0, 0); |
| 92 | static struct slow_work slow_work_new_thread; /* new thread starter */ |
| 93 | |
David Howells | 07fe7cb | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 94 | /* |
David Howells | 3d7a641 | 2009-11-19 18:10:23 +0000 | [diff] [blame] | 95 | * slow work ID allocation (use slow_work_queue_lock) |
| 96 | */ |
| 97 | static DECLARE_BITMAP(slow_work_ids, SLOW_WORK_THREAD_LIMIT); |
| 98 | |
| 99 | /* |
| 100 | * Unregistration tracking to prevent put_ref() from disappearing during module |
| 101 | * unload |
| 102 | */ |
| 103 | #ifdef CONFIG_MODULES |
| 104 | static struct module *slow_work_thread_processing[SLOW_WORK_THREAD_LIMIT]; |
| 105 | static struct module *slow_work_unreg_module; |
| 106 | static struct slow_work *slow_work_unreg_work_item; |
| 107 | static DECLARE_WAIT_QUEUE_HEAD(slow_work_unreg_wq); |
| 108 | static DEFINE_MUTEX(slow_work_unreg_sync_lock); |
David Howells | fa1dae4 | 2009-12-01 13:52:08 +0000 | [diff] [blame] | 109 | |
| 110 | static void slow_work_set_thread_processing(int id, struct slow_work *work) |
| 111 | { |
| 112 | if (work) |
| 113 | slow_work_thread_processing[id] = work->owner; |
| 114 | } |
| 115 | static void slow_work_done_thread_processing(int id, struct slow_work *work) |
| 116 | { |
| 117 | struct module *module = slow_work_thread_processing[id]; |
| 118 | |
| 119 | slow_work_thread_processing[id] = NULL; |
| 120 | smp_mb(); |
| 121 | if (slow_work_unreg_work_item == work || |
| 122 | slow_work_unreg_module == module) |
| 123 | wake_up_all(&slow_work_unreg_wq); |
| 124 | } |
| 125 | static void slow_work_clear_thread_processing(int id) |
| 126 | { |
| 127 | slow_work_thread_processing[id] = NULL; |
| 128 | } |
| 129 | #else |
| 130 | static void slow_work_set_thread_processing(int id, struct slow_work *work) {} |
| 131 | static void slow_work_done_thread_processing(int id, struct slow_work *work) {} |
| 132 | static void slow_work_clear_thread_processing(int id) {} |
David Howells | 3d7a641 | 2009-11-19 18:10:23 +0000 | [diff] [blame] | 133 | #endif |
| 134 | |
| 135 | /* |
David Howells | 8fba10a | 2009-11-19 18:10:51 +0000 | [diff] [blame] | 136 | * Data for tracking currently executing items for indication through /proc |
| 137 | */ |
David Howells | f13a48b | 2009-12-01 15:36:11 +0000 | [diff] [blame] | 138 | #ifdef CONFIG_SLOW_WORK_DEBUG |
David Howells | 8fba10a | 2009-11-19 18:10:51 +0000 | [diff] [blame] | 139 | struct slow_work *slow_work_execs[SLOW_WORK_THREAD_LIMIT]; |
| 140 | pid_t slow_work_pids[SLOW_WORK_THREAD_LIMIT]; |
| 141 | DEFINE_RWLOCK(slow_work_execs_lock); |
| 142 | #endif |
| 143 | |
| 144 | /* |
David Howells | 07fe7cb | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 145 | * The queues of work items and the lock governing access to them. These are |
| 146 | * shared between all the CPUs. It doesn't make sense to have per-CPU queues |
| 147 | * as the number of threads bears no relation to the number of CPUs. |
| 148 | * |
| 149 | * There are two queues of work items: one for slow work items, and one for |
| 150 | * very slow work items. |
| 151 | */ |
David Howells | 8fba10a | 2009-11-19 18:10:51 +0000 | [diff] [blame] | 152 | LIST_HEAD(slow_work_queue); |
| 153 | LIST_HEAD(vslow_work_queue); |
| 154 | DEFINE_SPINLOCK(slow_work_queue_lock); |
David Howells | 07fe7cb | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 155 | |
| 156 | /* |
David Howells | 3bde31a | 2009-11-19 18:10:57 +0000 | [diff] [blame] | 157 | * The following are two wait queues that get pinged when a work item is placed |
| 158 | * on an empty queue. These allow work items that are hogging a thread by |
| 159 | * sleeping in a way that could be deferred to yield their thread and enqueue |
| 160 | * themselves. |
| 161 | */ |
| 162 | static DECLARE_WAIT_QUEUE_HEAD(slow_work_queue_waits_for_occupation); |
| 163 | static DECLARE_WAIT_QUEUE_HEAD(vslow_work_queue_waits_for_occupation); |
| 164 | |
| 165 | /* |
David Howells | 07fe7cb | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 166 | * The thread controls. A variable used to signal to the threads that they |
| 167 | * should exit when the queue is empty, a waitqueue used by the threads to wait |
| 168 | * for signals, and a completion set by the last thread to exit. |
| 169 | */ |
| 170 | static bool slow_work_threads_should_exit; |
| 171 | static DECLARE_WAIT_QUEUE_HEAD(slow_work_thread_wq); |
| 172 | static DECLARE_COMPLETION(slow_work_last_thread_exited); |
| 173 | |
| 174 | /* |
| 175 | * The number of users of the thread pool and its lock. Whilst this is zero we |
| 176 | * have no threads hanging around, and when this reaches zero, we wait for all |
| 177 | * active or queued work items to complete and kill all the threads we do have. |
| 178 | */ |
| 179 | static int slow_work_user_count; |
| 180 | static DEFINE_MUTEX(slow_work_user_lock); |
| 181 | |
Jens Axboe | 4d8bb2c | 2009-11-19 18:10:39 +0000 | [diff] [blame] | 182 | static inline int slow_work_get_ref(struct slow_work *work) |
| 183 | { |
| 184 | if (work->ops->get_ref) |
| 185 | return work->ops->get_ref(work); |
| 186 | |
| 187 | return 0; |
| 188 | } |
| 189 | |
| 190 | static inline void slow_work_put_ref(struct slow_work *work) |
| 191 | { |
| 192 | if (work->ops->put_ref) |
| 193 | work->ops->put_ref(work); |
| 194 | } |
| 195 | |
David Howells | 07fe7cb | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 196 | /* |
| 197 | * Calculate the maximum number of active threads in the pool that are |
| 198 | * permitted to process very slow work items. |
| 199 | * |
| 200 | * The answer is rounded up to at least 1, but may not equal or exceed the |
| 201 | * maximum number of the threads in the pool. This means we always have at |
| 202 | * least one thread that can process slow work items, and we always have at |
| 203 | * least one thread that won't get tied up doing so. |
| 204 | */ |
| 205 | static unsigned slow_work_calc_vsmax(void) |
| 206 | { |
| 207 | unsigned vsmax; |
| 208 | |
| 209 | vsmax = atomic_read(&slow_work_thread_count) * vslow_work_proportion; |
| 210 | vsmax /= 100; |
| 211 | vsmax = max(vsmax, 1U); |
| 212 | return min(vsmax, slow_work_max_threads - 1); |
| 213 | } |
| 214 | |
| 215 | /* |
| 216 | * Attempt to execute stuff queued on a slow thread. Return true if we managed |
| 217 | * it, false if there was nothing to do. |
| 218 | */ |
David Howells | 8fba10a | 2009-11-19 18:10:51 +0000 | [diff] [blame] | 219 | static noinline bool slow_work_execute(int id) |
David Howells | 07fe7cb | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 220 | { |
| 221 | struct slow_work *work = NULL; |
| 222 | unsigned vsmax; |
| 223 | bool very_slow; |
| 224 | |
| 225 | vsmax = slow_work_calc_vsmax(); |
| 226 | |
David Howells | 109d927 | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 227 | /* see if we can schedule a new thread to be started if we're not |
| 228 | * keeping up with the work */ |
| 229 | if (!waitqueue_active(&slow_work_thread_wq) && |
| 230 | (!list_empty(&slow_work_queue) || !list_empty(&vslow_work_queue)) && |
| 231 | atomic_read(&slow_work_thread_count) < slow_work_max_threads && |
| 232 | !slow_work_may_not_start_new_thread) |
| 233 | slow_work_enqueue(&slow_work_new_thread); |
| 234 | |
David Howells | 07fe7cb | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 235 | /* find something to execute */ |
| 236 | spin_lock_irq(&slow_work_queue_lock); |
| 237 | if (!list_empty(&vslow_work_queue) && |
| 238 | atomic_read(&vslow_work_executing_count) < vsmax) { |
| 239 | work = list_entry(vslow_work_queue.next, |
| 240 | struct slow_work, link); |
| 241 | if (test_and_set_bit_lock(SLOW_WORK_EXECUTING, &work->flags)) |
| 242 | BUG(); |
| 243 | list_del_init(&work->link); |
| 244 | atomic_inc(&vslow_work_executing_count); |
| 245 | very_slow = true; |
| 246 | } else if (!list_empty(&slow_work_queue)) { |
| 247 | work = list_entry(slow_work_queue.next, |
| 248 | struct slow_work, link); |
| 249 | if (test_and_set_bit_lock(SLOW_WORK_EXECUTING, &work->flags)) |
| 250 | BUG(); |
| 251 | list_del_init(&work->link); |
| 252 | very_slow = false; |
| 253 | } else { |
| 254 | very_slow = false; /* avoid the compiler warning */ |
| 255 | } |
David Howells | 3d7a641 | 2009-11-19 18:10:23 +0000 | [diff] [blame] | 256 | |
David Howells | fa1dae4 | 2009-12-01 13:52:08 +0000 | [diff] [blame] | 257 | slow_work_set_thread_processing(id, work); |
David Howells | 8fba10a | 2009-11-19 18:10:51 +0000 | [diff] [blame] | 258 | if (work) { |
| 259 | slow_work_mark_time(work); |
| 260 | slow_work_begin_exec(id, work); |
| 261 | } |
David Howells | 3d7a641 | 2009-11-19 18:10:23 +0000 | [diff] [blame] | 262 | |
David Howells | 07fe7cb | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 263 | spin_unlock_irq(&slow_work_queue_lock); |
| 264 | |
| 265 | if (!work) |
| 266 | return false; |
| 267 | |
| 268 | if (!test_and_clear_bit(SLOW_WORK_PENDING, &work->flags)) |
| 269 | BUG(); |
| 270 | |
Jens Axboe | 0160950 | 2009-11-19 18:10:43 +0000 | [diff] [blame] | 271 | /* don't execute if the work is in the process of being cancelled */ |
| 272 | if (!test_bit(SLOW_WORK_CANCELLING, &work->flags)) |
| 273 | work->ops->execute(work); |
David Howells | 07fe7cb | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 274 | |
| 275 | if (very_slow) |
| 276 | atomic_dec(&vslow_work_executing_count); |
| 277 | clear_bit_unlock(SLOW_WORK_EXECUTING, &work->flags); |
| 278 | |
Jens Axboe | 0160950 | 2009-11-19 18:10:43 +0000 | [diff] [blame] | 279 | /* wake up anyone waiting for this work to be complete */ |
| 280 | wake_up_bit(&work->flags, SLOW_WORK_EXECUTING); |
| 281 | |
David Howells | 8fba10a | 2009-11-19 18:10:51 +0000 | [diff] [blame] | 282 | slow_work_end_exec(id, work); |
| 283 | |
David Howells | 07fe7cb | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 284 | /* if someone tried to enqueue the item whilst we were executing it, |
| 285 | * then it'll be left unenqueued to avoid multiple threads trying to |
| 286 | * execute it simultaneously |
| 287 | * |
| 288 | * there is, however, a race between us testing the pending flag and |
| 289 | * getting the spinlock, and between the enqueuer setting the pending |
| 290 | * flag and getting the spinlock, so we use a deferral bit to tell us |
| 291 | * if the enqueuer got there first |
| 292 | */ |
| 293 | if (test_bit(SLOW_WORK_PENDING, &work->flags)) { |
| 294 | spin_lock_irq(&slow_work_queue_lock); |
| 295 | |
| 296 | if (!test_bit(SLOW_WORK_EXECUTING, &work->flags) && |
| 297 | test_and_clear_bit(SLOW_WORK_ENQ_DEFERRED, &work->flags)) |
| 298 | goto auto_requeue; |
| 299 | |
| 300 | spin_unlock_irq(&slow_work_queue_lock); |
| 301 | } |
| 302 | |
David Howells | 3d7a641 | 2009-11-19 18:10:23 +0000 | [diff] [blame] | 303 | /* sort out the race between module unloading and put_ref() */ |
Jens Axboe | 4d8bb2c | 2009-11-19 18:10:39 +0000 | [diff] [blame] | 304 | slow_work_put_ref(work); |
David Howells | fa1dae4 | 2009-12-01 13:52:08 +0000 | [diff] [blame] | 305 | slow_work_done_thread_processing(id, work); |
David Howells | 3d7a641 | 2009-11-19 18:10:23 +0000 | [diff] [blame] | 306 | |
David Howells | 07fe7cb | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 307 | return true; |
| 308 | |
| 309 | auto_requeue: |
| 310 | /* we must complete the enqueue operation |
| 311 | * - we transfer our ref on the item back to the appropriate queue |
| 312 | * - don't wake another thread up as we're awake already |
| 313 | */ |
David Howells | 8fba10a | 2009-11-19 18:10:51 +0000 | [diff] [blame] | 314 | slow_work_mark_time(work); |
David Howells | 07fe7cb | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 315 | if (test_bit(SLOW_WORK_VERY_SLOW, &work->flags)) |
| 316 | list_add_tail(&work->link, &vslow_work_queue); |
| 317 | else |
| 318 | list_add_tail(&work->link, &slow_work_queue); |
| 319 | spin_unlock_irq(&slow_work_queue_lock); |
David Howells | fa1dae4 | 2009-12-01 13:52:08 +0000 | [diff] [blame] | 320 | slow_work_clear_thread_processing(id); |
David Howells | 07fe7cb | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 321 | return true; |
| 322 | } |
| 323 | |
| 324 | /** |
David Howells | 3bde31a | 2009-11-19 18:10:57 +0000 | [diff] [blame] | 325 | * slow_work_sleep_till_thread_needed - Sleep till thread needed by other work |
| 326 | * work: The work item under execution that wants to sleep |
| 327 | * _timeout: Scheduler sleep timeout |
| 328 | * |
| 329 | * Allow a requeueable work item to sleep on a slow-work processor thread until |
| 330 | * that thread is needed to do some other work or the sleep is interrupted by |
| 331 | * some other event. |
| 332 | * |
| 333 | * The caller must set up a wake up event before calling this and must have set |
| 334 | * the appropriate sleep mode (such as TASK_UNINTERRUPTIBLE) and tested its own |
| 335 | * condition before calling this function as no test is made here. |
| 336 | * |
| 337 | * False is returned if there is nothing on the queue; true is returned if the |
| 338 | * work item should be requeued |
| 339 | */ |
| 340 | bool slow_work_sleep_till_thread_needed(struct slow_work *work, |
| 341 | signed long *_timeout) |
| 342 | { |
| 343 | wait_queue_head_t *wfo_wq; |
| 344 | struct list_head *queue; |
| 345 | |
| 346 | DEFINE_WAIT(wait); |
| 347 | |
| 348 | if (test_bit(SLOW_WORK_VERY_SLOW, &work->flags)) { |
| 349 | wfo_wq = &vslow_work_queue_waits_for_occupation; |
| 350 | queue = &vslow_work_queue; |
| 351 | } else { |
| 352 | wfo_wq = &slow_work_queue_waits_for_occupation; |
| 353 | queue = &slow_work_queue; |
| 354 | } |
| 355 | |
| 356 | if (!list_empty(queue)) |
| 357 | return true; |
| 358 | |
| 359 | add_wait_queue_exclusive(wfo_wq, &wait); |
| 360 | if (list_empty(queue)) |
| 361 | *_timeout = schedule_timeout(*_timeout); |
| 362 | finish_wait(wfo_wq, &wait); |
| 363 | |
| 364 | return !list_empty(queue); |
| 365 | } |
| 366 | EXPORT_SYMBOL(slow_work_sleep_till_thread_needed); |
| 367 | |
| 368 | /** |
David Howells | 07fe7cb | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 369 | * slow_work_enqueue - Schedule a slow work item for processing |
| 370 | * @work: The work item to queue |
| 371 | * |
| 372 | * Schedule a slow work item for processing. If the item is already undergoing |
| 373 | * execution, this guarantees not to re-enter the execution routine until the |
| 374 | * first execution finishes. |
| 375 | * |
| 376 | * The item is pinned by this function as it retains a reference to it, managed |
| 377 | * through the item operations. The item is unpinned once it has been |
| 378 | * executed. |
| 379 | * |
| 380 | * An item may hog the thread that is running it for a relatively large amount |
| 381 | * of time, sufficient, for example, to perform several lookup, mkdir, create |
| 382 | * and setxattr operations. It may sleep on I/O and may sleep to obtain locks. |
| 383 | * |
| 384 | * Conversely, if a number of items are awaiting processing, it may take some |
| 385 | * time before any given item is given attention. The number of threads in the |
| 386 | * pool may be increased to deal with demand, but only up to a limit. |
| 387 | * |
| 388 | * If SLOW_WORK_VERY_SLOW is set on the work item, then it will be placed in |
| 389 | * the very slow queue, from which only a portion of the threads will be |
| 390 | * allowed to pick items to execute. This ensures that very slow items won't |
| 391 | * overly block ones that are just ordinarily slow. |
| 392 | * |
Jens Axboe | 0160950 | 2009-11-19 18:10:43 +0000 | [diff] [blame] | 393 | * Returns 0 if successful, -EAGAIN if not (or -ECANCELED if cancelled work is |
| 394 | * attempted queued) |
David Howells | 07fe7cb | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 395 | */ |
| 396 | int slow_work_enqueue(struct slow_work *work) |
| 397 | { |
David Howells | 3bde31a | 2009-11-19 18:10:57 +0000 | [diff] [blame] | 398 | wait_queue_head_t *wfo_wq; |
| 399 | struct list_head *queue; |
David Howells | 07fe7cb | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 400 | unsigned long flags; |
Jens Axboe | 0160950 | 2009-11-19 18:10:43 +0000 | [diff] [blame] | 401 | int ret; |
| 402 | |
| 403 | if (test_bit(SLOW_WORK_CANCELLING, &work->flags)) |
| 404 | return -ECANCELED; |
David Howells | 07fe7cb | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 405 | |
| 406 | BUG_ON(slow_work_user_count <= 0); |
| 407 | BUG_ON(!work); |
| 408 | BUG_ON(!work->ops); |
David Howells | 07fe7cb | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 409 | |
| 410 | /* when honouring an enqueue request, we only promise that we will run |
| 411 | * the work function in the future; we do not promise to run it once |
| 412 | * per enqueue request |
| 413 | * |
| 414 | * we use the PENDING bit to merge together repeat requests without |
| 415 | * having to disable IRQs and take the spinlock, whilst still |
| 416 | * maintaining our promise |
| 417 | */ |
| 418 | if (!test_and_set_bit_lock(SLOW_WORK_PENDING, &work->flags)) { |
David Howells | 3bde31a | 2009-11-19 18:10:57 +0000 | [diff] [blame] | 419 | if (test_bit(SLOW_WORK_VERY_SLOW, &work->flags)) { |
| 420 | wfo_wq = &vslow_work_queue_waits_for_occupation; |
| 421 | queue = &vslow_work_queue; |
| 422 | } else { |
| 423 | wfo_wq = &slow_work_queue_waits_for_occupation; |
| 424 | queue = &slow_work_queue; |
| 425 | } |
| 426 | |
David Howells | 07fe7cb | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 427 | spin_lock_irqsave(&slow_work_queue_lock, flags); |
| 428 | |
Jens Axboe | 0160950 | 2009-11-19 18:10:43 +0000 | [diff] [blame] | 429 | if (unlikely(test_bit(SLOW_WORK_CANCELLING, &work->flags))) |
| 430 | goto cancelled; |
| 431 | |
David Howells | 07fe7cb | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 432 | /* we promise that we will not attempt to execute the work |
| 433 | * function in more than one thread simultaneously |
| 434 | * |
| 435 | * this, however, leaves us with a problem if we're asked to |
| 436 | * enqueue the work whilst someone is executing the work |
| 437 | * function as simply queueing the work immediately means that |
| 438 | * another thread may try executing it whilst it is already |
| 439 | * under execution |
| 440 | * |
| 441 | * to deal with this, we set the ENQ_DEFERRED bit instead of |
| 442 | * enqueueing, and the thread currently executing the work |
| 443 | * function will enqueue the work item when the work function |
| 444 | * returns and it has cleared the EXECUTING bit |
| 445 | */ |
| 446 | if (test_bit(SLOW_WORK_EXECUTING, &work->flags)) { |
| 447 | set_bit(SLOW_WORK_ENQ_DEFERRED, &work->flags); |
| 448 | } else { |
Jens Axboe | 0160950 | 2009-11-19 18:10:43 +0000 | [diff] [blame] | 449 | ret = slow_work_get_ref(work); |
| 450 | if (ret < 0) |
| 451 | goto failed; |
David Howells | 8fba10a | 2009-11-19 18:10:51 +0000 | [diff] [blame] | 452 | slow_work_mark_time(work); |
David Howells | 3bde31a | 2009-11-19 18:10:57 +0000 | [diff] [blame] | 453 | list_add_tail(&work->link, queue); |
David Howells | 07fe7cb | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 454 | wake_up(&slow_work_thread_wq); |
David Howells | 3bde31a | 2009-11-19 18:10:57 +0000 | [diff] [blame] | 455 | |
| 456 | /* if someone who could be requeued is sleeping on a |
| 457 | * thread, then ask them to yield their thread */ |
| 458 | if (work->link.prev == queue) |
| 459 | wake_up(wfo_wq); |
David Howells | 07fe7cb | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 460 | } |
| 461 | |
| 462 | spin_unlock_irqrestore(&slow_work_queue_lock, flags); |
| 463 | } |
| 464 | return 0; |
| 465 | |
Jens Axboe | 0160950 | 2009-11-19 18:10:43 +0000 | [diff] [blame] | 466 | cancelled: |
| 467 | ret = -ECANCELED; |
| 468 | failed: |
David Howells | 07fe7cb | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 469 | spin_unlock_irqrestore(&slow_work_queue_lock, flags); |
Jens Axboe | 0160950 | 2009-11-19 18:10:43 +0000 | [diff] [blame] | 470 | return ret; |
David Howells | 07fe7cb | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 471 | } |
| 472 | EXPORT_SYMBOL(slow_work_enqueue); |
| 473 | |
Jens Axboe | 0160950 | 2009-11-19 18:10:43 +0000 | [diff] [blame] | 474 | static int slow_work_wait(void *word) |
| 475 | { |
| 476 | schedule(); |
| 477 | return 0; |
| 478 | } |
| 479 | |
| 480 | /** |
| 481 | * slow_work_cancel - Cancel a slow work item |
| 482 | * @work: The work item to cancel |
| 483 | * |
| 484 | * This function will cancel a previously enqueued work item. If we cannot |
| 485 | * cancel the work item, it is guarenteed to have run when this function |
| 486 | * returns. |
| 487 | */ |
| 488 | void slow_work_cancel(struct slow_work *work) |
| 489 | { |
| 490 | bool wait = true, put = false; |
| 491 | |
| 492 | set_bit(SLOW_WORK_CANCELLING, &work->flags); |
Jens Axboe | 6b8268b | 2009-11-19 18:10:47 +0000 | [diff] [blame] | 493 | smp_mb(); |
| 494 | |
| 495 | /* if the work item is a delayed work item with an active timer, we |
| 496 | * need to wait for the timer to finish _before_ getting the spinlock, |
| 497 | * lest we deadlock against the timer routine |
| 498 | * |
| 499 | * the timer routine will leave DELAYED set if it notices the |
| 500 | * CANCELLING flag in time |
| 501 | */ |
| 502 | if (test_bit(SLOW_WORK_DELAYED, &work->flags)) { |
| 503 | struct delayed_slow_work *dwork = |
| 504 | container_of(work, struct delayed_slow_work, work); |
| 505 | del_timer_sync(&dwork->timer); |
| 506 | } |
Jens Axboe | 0160950 | 2009-11-19 18:10:43 +0000 | [diff] [blame] | 507 | |
| 508 | spin_lock_irq(&slow_work_queue_lock); |
| 509 | |
Jens Axboe | 6b8268b | 2009-11-19 18:10:47 +0000 | [diff] [blame] | 510 | if (test_bit(SLOW_WORK_DELAYED, &work->flags)) { |
| 511 | /* the timer routine aborted or never happened, so we are left |
| 512 | * holding the timer's reference on the item and should just |
| 513 | * drop the pending flag and wait for any ongoing execution to |
| 514 | * finish */ |
| 515 | struct delayed_slow_work *dwork = |
| 516 | container_of(work, struct delayed_slow_work, work); |
| 517 | |
| 518 | BUG_ON(timer_pending(&dwork->timer)); |
| 519 | BUG_ON(!list_empty(&work->link)); |
| 520 | |
| 521 | clear_bit(SLOW_WORK_DELAYED, &work->flags); |
| 522 | put = true; |
| 523 | clear_bit(SLOW_WORK_PENDING, &work->flags); |
| 524 | |
| 525 | } else if (test_bit(SLOW_WORK_PENDING, &work->flags) && |
| 526 | !list_empty(&work->link)) { |
Jens Axboe | 0160950 | 2009-11-19 18:10:43 +0000 | [diff] [blame] | 527 | /* the link in the pending queue holds a reference on the item |
| 528 | * that we will need to release */ |
| 529 | list_del_init(&work->link); |
| 530 | wait = false; |
| 531 | put = true; |
| 532 | clear_bit(SLOW_WORK_PENDING, &work->flags); |
| 533 | |
| 534 | } else if (test_and_clear_bit(SLOW_WORK_ENQ_DEFERRED, &work->flags)) { |
| 535 | /* the executor is holding our only reference on the item, so |
| 536 | * we merely need to wait for it to finish executing */ |
| 537 | clear_bit(SLOW_WORK_PENDING, &work->flags); |
| 538 | } |
| 539 | |
| 540 | spin_unlock_irq(&slow_work_queue_lock); |
| 541 | |
| 542 | /* the EXECUTING flag is set by the executor whilst the spinlock is set |
| 543 | * and before the item is dequeued - so assuming the above doesn't |
| 544 | * actually dequeue it, simply waiting for the EXECUTING flag to be |
| 545 | * released here should be sufficient */ |
| 546 | if (wait) |
| 547 | wait_on_bit(&work->flags, SLOW_WORK_EXECUTING, slow_work_wait, |
| 548 | TASK_UNINTERRUPTIBLE); |
| 549 | |
| 550 | clear_bit(SLOW_WORK_CANCELLING, &work->flags); |
| 551 | if (put) |
| 552 | slow_work_put_ref(work); |
| 553 | } |
| 554 | EXPORT_SYMBOL(slow_work_cancel); |
| 555 | |
David Howells | 07fe7cb | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 556 | /* |
Jens Axboe | 6b8268b | 2009-11-19 18:10:47 +0000 | [diff] [blame] | 557 | * Handle expiry of the delay timer, indicating that a delayed slow work item |
| 558 | * should now be queued if not cancelled |
| 559 | */ |
| 560 | static void delayed_slow_work_timer(unsigned long data) |
| 561 | { |
David Howells | 3bde31a | 2009-11-19 18:10:57 +0000 | [diff] [blame] | 562 | wait_queue_head_t *wfo_wq; |
| 563 | struct list_head *queue; |
Jens Axboe | 6b8268b | 2009-11-19 18:10:47 +0000 | [diff] [blame] | 564 | struct slow_work *work = (struct slow_work *) data; |
| 565 | unsigned long flags; |
David Howells | 3bde31a | 2009-11-19 18:10:57 +0000 | [diff] [blame] | 566 | bool queued = false, put = false, first = false; |
| 567 | |
| 568 | if (test_bit(SLOW_WORK_VERY_SLOW, &work->flags)) { |
| 569 | wfo_wq = &vslow_work_queue_waits_for_occupation; |
| 570 | queue = &vslow_work_queue; |
| 571 | } else { |
| 572 | wfo_wq = &slow_work_queue_waits_for_occupation; |
| 573 | queue = &slow_work_queue; |
| 574 | } |
Jens Axboe | 6b8268b | 2009-11-19 18:10:47 +0000 | [diff] [blame] | 575 | |
| 576 | spin_lock_irqsave(&slow_work_queue_lock, flags); |
| 577 | if (likely(!test_bit(SLOW_WORK_CANCELLING, &work->flags))) { |
| 578 | clear_bit(SLOW_WORK_DELAYED, &work->flags); |
| 579 | |
| 580 | if (test_bit(SLOW_WORK_EXECUTING, &work->flags)) { |
| 581 | /* we discard the reference the timer was holding in |
| 582 | * favour of the one the executor holds */ |
| 583 | set_bit(SLOW_WORK_ENQ_DEFERRED, &work->flags); |
| 584 | put = true; |
| 585 | } else { |
David Howells | 8fba10a | 2009-11-19 18:10:51 +0000 | [diff] [blame] | 586 | slow_work_mark_time(work); |
David Howells | 3bde31a | 2009-11-19 18:10:57 +0000 | [diff] [blame] | 587 | list_add_tail(&work->link, queue); |
Jens Axboe | 6b8268b | 2009-11-19 18:10:47 +0000 | [diff] [blame] | 588 | queued = true; |
David Howells | 3bde31a | 2009-11-19 18:10:57 +0000 | [diff] [blame] | 589 | if (work->link.prev == queue) |
| 590 | first = true; |
Jens Axboe | 6b8268b | 2009-11-19 18:10:47 +0000 | [diff] [blame] | 591 | } |
| 592 | } |
| 593 | |
| 594 | spin_unlock_irqrestore(&slow_work_queue_lock, flags); |
| 595 | if (put) |
| 596 | slow_work_put_ref(work); |
David Howells | 3bde31a | 2009-11-19 18:10:57 +0000 | [diff] [blame] | 597 | if (first) |
| 598 | wake_up(wfo_wq); |
Jens Axboe | 6b8268b | 2009-11-19 18:10:47 +0000 | [diff] [blame] | 599 | if (queued) |
| 600 | wake_up(&slow_work_thread_wq); |
| 601 | } |
| 602 | |
| 603 | /** |
| 604 | * delayed_slow_work_enqueue - Schedule a delayed slow work item for processing |
| 605 | * @dwork: The delayed work item to queue |
| 606 | * @delay: When to start executing the work, in jiffies from now |
| 607 | * |
| 608 | * This is similar to slow_work_enqueue(), but it adds a delay before the work |
| 609 | * is actually queued for processing. |
| 610 | * |
| 611 | * The item can have delayed processing requested on it whilst it is being |
| 612 | * executed. The delay will begin immediately, and if it expires before the |
| 613 | * item finishes executing, the item will be placed back on the queue when it |
| 614 | * has done executing. |
| 615 | */ |
| 616 | int delayed_slow_work_enqueue(struct delayed_slow_work *dwork, |
| 617 | unsigned long delay) |
| 618 | { |
| 619 | struct slow_work *work = &dwork->work; |
| 620 | unsigned long flags; |
| 621 | int ret; |
| 622 | |
| 623 | if (delay == 0) |
| 624 | return slow_work_enqueue(&dwork->work); |
| 625 | |
| 626 | BUG_ON(slow_work_user_count <= 0); |
| 627 | BUG_ON(!work); |
| 628 | BUG_ON(!work->ops); |
| 629 | |
| 630 | if (test_bit(SLOW_WORK_CANCELLING, &work->flags)) |
| 631 | return -ECANCELED; |
| 632 | |
| 633 | if (!test_and_set_bit_lock(SLOW_WORK_PENDING, &work->flags)) { |
| 634 | spin_lock_irqsave(&slow_work_queue_lock, flags); |
| 635 | |
| 636 | if (test_bit(SLOW_WORK_CANCELLING, &work->flags)) |
| 637 | goto cancelled; |
| 638 | |
| 639 | /* the timer holds a reference whilst it is pending */ |
Dave Airlie | 88be12c | 2010-03-29 12:01:50 +0100 | [diff] [blame] | 640 | ret = slow_work_get_ref(work); |
Jens Axboe | 6b8268b | 2009-11-19 18:10:47 +0000 | [diff] [blame] | 641 | if (ret < 0) |
| 642 | goto cant_get_ref; |
| 643 | |
| 644 | if (test_and_set_bit(SLOW_WORK_DELAYED, &work->flags)) |
| 645 | BUG(); |
| 646 | dwork->timer.expires = jiffies + delay; |
| 647 | dwork->timer.data = (unsigned long) work; |
| 648 | dwork->timer.function = delayed_slow_work_timer; |
| 649 | add_timer(&dwork->timer); |
| 650 | |
| 651 | spin_unlock_irqrestore(&slow_work_queue_lock, flags); |
| 652 | } |
| 653 | |
| 654 | return 0; |
| 655 | |
| 656 | cancelled: |
| 657 | ret = -ECANCELED; |
| 658 | cant_get_ref: |
| 659 | spin_unlock_irqrestore(&slow_work_queue_lock, flags); |
| 660 | return ret; |
| 661 | } |
| 662 | EXPORT_SYMBOL(delayed_slow_work_enqueue); |
| 663 | |
| 664 | /* |
Chris Peterson | 009789f | 2009-06-16 15:33:43 -0700 | [diff] [blame] | 665 | * Schedule a cull of the thread pool at some time in the near future |
| 666 | */ |
| 667 | static void slow_work_schedule_cull(void) |
| 668 | { |
| 669 | mod_timer(&slow_work_cull_timer, |
| 670 | round_jiffies(jiffies + SLOW_WORK_CULL_TIMEOUT)); |
| 671 | } |
| 672 | |
| 673 | /* |
David Howells | 109d927 | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 674 | * Worker thread culling algorithm |
| 675 | */ |
| 676 | static bool slow_work_cull_thread(void) |
| 677 | { |
| 678 | unsigned long flags; |
| 679 | bool do_cull = false; |
| 680 | |
| 681 | spin_lock_irqsave(&slow_work_queue_lock, flags); |
| 682 | |
| 683 | if (slow_work_cull) { |
| 684 | slow_work_cull = false; |
| 685 | |
| 686 | if (list_empty(&slow_work_queue) && |
| 687 | list_empty(&vslow_work_queue) && |
| 688 | atomic_read(&slow_work_thread_count) > |
| 689 | slow_work_min_threads) { |
Chris Peterson | 009789f | 2009-06-16 15:33:43 -0700 | [diff] [blame] | 690 | slow_work_schedule_cull(); |
David Howells | 109d927 | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 691 | do_cull = true; |
| 692 | } |
| 693 | } |
| 694 | |
| 695 | spin_unlock_irqrestore(&slow_work_queue_lock, flags); |
| 696 | return do_cull; |
| 697 | } |
| 698 | |
| 699 | /* |
David Howells | 07fe7cb | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 700 | * Determine if there is slow work available for dispatch |
| 701 | */ |
| 702 | static inline bool slow_work_available(int vsmax) |
| 703 | { |
| 704 | return !list_empty(&slow_work_queue) || |
| 705 | (!list_empty(&vslow_work_queue) && |
| 706 | atomic_read(&vslow_work_executing_count) < vsmax); |
| 707 | } |
| 708 | |
| 709 | /* |
| 710 | * Worker thread dispatcher |
| 711 | */ |
| 712 | static int slow_work_thread(void *_data) |
| 713 | { |
David Howells | 3d7a641 | 2009-11-19 18:10:23 +0000 | [diff] [blame] | 714 | int vsmax, id; |
David Howells | 07fe7cb | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 715 | |
| 716 | DEFINE_WAIT(wait); |
| 717 | |
| 718 | set_freezable(); |
| 719 | set_user_nice(current, -5); |
| 720 | |
David Howells | 3d7a641 | 2009-11-19 18:10:23 +0000 | [diff] [blame] | 721 | /* allocate ourselves an ID */ |
| 722 | spin_lock_irq(&slow_work_queue_lock); |
| 723 | id = find_first_zero_bit(slow_work_ids, SLOW_WORK_THREAD_LIMIT); |
| 724 | BUG_ON(id < 0 || id >= SLOW_WORK_THREAD_LIMIT); |
| 725 | __set_bit(id, slow_work_ids); |
David Howells | 8fba10a | 2009-11-19 18:10:51 +0000 | [diff] [blame] | 726 | slow_work_set_thread_pid(id, current->pid); |
David Howells | 3d7a641 | 2009-11-19 18:10:23 +0000 | [diff] [blame] | 727 | spin_unlock_irq(&slow_work_queue_lock); |
| 728 | |
| 729 | sprintf(current->comm, "kslowd%03u", id); |
| 730 | |
David Howells | 07fe7cb | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 731 | for (;;) { |
| 732 | vsmax = vslow_work_proportion; |
| 733 | vsmax *= atomic_read(&slow_work_thread_count); |
| 734 | vsmax /= 100; |
| 735 | |
Oleg Nesterov | b415c49 | 2009-06-11 13:12:55 +0100 | [diff] [blame] | 736 | prepare_to_wait_exclusive(&slow_work_thread_wq, &wait, |
| 737 | TASK_INTERRUPTIBLE); |
David Howells | 07fe7cb | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 738 | if (!freezing(current) && |
| 739 | !slow_work_threads_should_exit && |
David Howells | 109d927 | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 740 | !slow_work_available(vsmax) && |
| 741 | !slow_work_cull) |
David Howells | 07fe7cb | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 742 | schedule(); |
| 743 | finish_wait(&slow_work_thread_wq, &wait); |
| 744 | |
| 745 | try_to_freeze(); |
| 746 | |
| 747 | vsmax = vslow_work_proportion; |
| 748 | vsmax *= atomic_read(&slow_work_thread_count); |
| 749 | vsmax /= 100; |
| 750 | |
David Howells | 3d7a641 | 2009-11-19 18:10:23 +0000 | [diff] [blame] | 751 | if (slow_work_available(vsmax) && slow_work_execute(id)) { |
David Howells | 07fe7cb | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 752 | cond_resched(); |
David Howells | 109d927 | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 753 | if (list_empty(&slow_work_queue) && |
| 754 | list_empty(&vslow_work_queue) && |
| 755 | atomic_read(&slow_work_thread_count) > |
| 756 | slow_work_min_threads) |
Chris Peterson | 009789f | 2009-06-16 15:33:43 -0700 | [diff] [blame] | 757 | slow_work_schedule_cull(); |
David Howells | 07fe7cb | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 758 | continue; |
| 759 | } |
| 760 | |
| 761 | if (slow_work_threads_should_exit) |
| 762 | break; |
David Howells | 109d927 | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 763 | |
| 764 | if (slow_work_cull && slow_work_cull_thread()) |
| 765 | break; |
David Howells | 07fe7cb | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 766 | } |
| 767 | |
David Howells | 3d7a641 | 2009-11-19 18:10:23 +0000 | [diff] [blame] | 768 | spin_lock_irq(&slow_work_queue_lock); |
David Howells | 8fba10a | 2009-11-19 18:10:51 +0000 | [diff] [blame] | 769 | slow_work_set_thread_pid(id, 0); |
David Howells | 3d7a641 | 2009-11-19 18:10:23 +0000 | [diff] [blame] | 770 | __clear_bit(id, slow_work_ids); |
| 771 | spin_unlock_irq(&slow_work_queue_lock); |
| 772 | |
David Howells | 07fe7cb | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 773 | if (atomic_dec_and_test(&slow_work_thread_count)) |
| 774 | complete_and_exit(&slow_work_last_thread_exited, 0); |
| 775 | return 0; |
| 776 | } |
| 777 | |
David Howells | 109d927 | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 778 | /* |
| 779 | * Handle thread cull timer expiration |
| 780 | */ |
| 781 | static void slow_work_cull_timeout(unsigned long data) |
| 782 | { |
| 783 | slow_work_cull = true; |
| 784 | wake_up(&slow_work_thread_wq); |
| 785 | } |
| 786 | |
| 787 | /* |
David Howells | 109d927 | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 788 | * Start a new slow work thread |
| 789 | */ |
| 790 | static void slow_work_new_thread_execute(struct slow_work *work) |
| 791 | { |
| 792 | struct task_struct *p; |
| 793 | |
| 794 | if (slow_work_threads_should_exit) |
| 795 | return; |
| 796 | |
| 797 | if (atomic_read(&slow_work_thread_count) >= slow_work_max_threads) |
| 798 | return; |
| 799 | |
| 800 | if (!mutex_trylock(&slow_work_user_lock)) |
| 801 | return; |
| 802 | |
| 803 | slow_work_may_not_start_new_thread = true; |
| 804 | atomic_inc(&slow_work_thread_count); |
| 805 | p = kthread_run(slow_work_thread, NULL, "kslowd"); |
| 806 | if (IS_ERR(p)) { |
| 807 | printk(KERN_DEBUG "Slow work thread pool: OOM\n"); |
| 808 | if (atomic_dec_and_test(&slow_work_thread_count)) |
| 809 | BUG(); /* we're running on a slow work thread... */ |
| 810 | mod_timer(&slow_work_oom_timer, |
Chris Peterson | 009789f | 2009-06-16 15:33:43 -0700 | [diff] [blame] | 811 | round_jiffies(jiffies + SLOW_WORK_OOM_TIMEOUT)); |
David Howells | 109d927 | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 812 | } else { |
| 813 | /* ratelimit the starting of new threads */ |
| 814 | mod_timer(&slow_work_oom_timer, jiffies + 1); |
| 815 | } |
| 816 | |
| 817 | mutex_unlock(&slow_work_user_lock); |
| 818 | } |
| 819 | |
| 820 | static const struct slow_work_ops slow_work_new_thread_ops = { |
David Howells | 3d7a641 | 2009-11-19 18:10:23 +0000 | [diff] [blame] | 821 | .owner = THIS_MODULE, |
David Howells | 109d927 | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 822 | .execute = slow_work_new_thread_execute, |
David Howells | f13a48b | 2009-12-01 15:36:11 +0000 | [diff] [blame] | 823 | #ifdef CONFIG_SLOW_WORK_DEBUG |
David Howells | 8fba10a | 2009-11-19 18:10:51 +0000 | [diff] [blame] | 824 | .desc = slow_work_new_thread_desc, |
| 825 | #endif |
David Howells | 109d927 | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 826 | }; |
| 827 | |
| 828 | /* |
| 829 | * post-OOM new thread start suppression expiration |
| 830 | */ |
| 831 | static void slow_work_oom_timeout(unsigned long data) |
| 832 | { |
| 833 | slow_work_may_not_start_new_thread = false; |
| 834 | } |
| 835 | |
David Howells | 12e22c5 | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 836 | #ifdef CONFIG_SYSCTL |
| 837 | /* |
| 838 | * Handle adjustment of the minimum number of threads |
| 839 | */ |
| 840 | static int slow_work_min_threads_sysctl(struct ctl_table *table, int write, |
Alexey Dobriyan | 8d65af7 | 2009-09-23 15:57:19 -0700 | [diff] [blame] | 841 | void __user *buffer, |
David Howells | 12e22c5 | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 842 | size_t *lenp, loff_t *ppos) |
| 843 | { |
Alexey Dobriyan | 8d65af7 | 2009-09-23 15:57:19 -0700 | [diff] [blame] | 844 | int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
David Howells | 12e22c5 | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 845 | int n; |
| 846 | |
| 847 | if (ret == 0) { |
| 848 | mutex_lock(&slow_work_user_lock); |
| 849 | if (slow_work_user_count > 0) { |
| 850 | /* see if we need to start or stop threads */ |
| 851 | n = atomic_read(&slow_work_thread_count) - |
| 852 | slow_work_min_threads; |
| 853 | |
| 854 | if (n < 0 && !slow_work_may_not_start_new_thread) |
| 855 | slow_work_enqueue(&slow_work_new_thread); |
| 856 | else if (n > 0) |
Chris Peterson | 009789f | 2009-06-16 15:33:43 -0700 | [diff] [blame] | 857 | slow_work_schedule_cull(); |
David Howells | 12e22c5 | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 858 | } |
| 859 | mutex_unlock(&slow_work_user_lock); |
| 860 | } |
| 861 | |
| 862 | return ret; |
| 863 | } |
| 864 | |
| 865 | /* |
| 866 | * Handle adjustment of the maximum number of threads |
| 867 | */ |
| 868 | static int slow_work_max_threads_sysctl(struct ctl_table *table, int write, |
Alexey Dobriyan | 8d65af7 | 2009-09-23 15:57:19 -0700 | [diff] [blame] | 869 | void __user *buffer, |
David Howells | 12e22c5 | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 870 | size_t *lenp, loff_t *ppos) |
| 871 | { |
Alexey Dobriyan | 8d65af7 | 2009-09-23 15:57:19 -0700 | [diff] [blame] | 872 | int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
David Howells | 12e22c5 | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 873 | int n; |
| 874 | |
| 875 | if (ret == 0) { |
| 876 | mutex_lock(&slow_work_user_lock); |
| 877 | if (slow_work_user_count > 0) { |
| 878 | /* see if we need to stop threads */ |
| 879 | n = slow_work_max_threads - |
| 880 | atomic_read(&slow_work_thread_count); |
| 881 | |
| 882 | if (n < 0) |
Chris Peterson | 009789f | 2009-06-16 15:33:43 -0700 | [diff] [blame] | 883 | slow_work_schedule_cull(); |
David Howells | 12e22c5 | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 884 | } |
| 885 | mutex_unlock(&slow_work_user_lock); |
| 886 | } |
| 887 | |
| 888 | return ret; |
| 889 | } |
| 890 | #endif /* CONFIG_SYSCTL */ |
| 891 | |
David Howells | 07fe7cb | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 892 | /** |
| 893 | * slow_work_register_user - Register a user of the facility |
David Howells | 3d7a641 | 2009-11-19 18:10:23 +0000 | [diff] [blame] | 894 | * @module: The module about to make use of the facility |
David Howells | 07fe7cb | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 895 | * |
| 896 | * Register a user of the facility, starting up the initial threads if there |
| 897 | * aren't any other users at this point. This will return 0 if successful, or |
| 898 | * an error if not. |
| 899 | */ |
David Howells | 3d7a641 | 2009-11-19 18:10:23 +0000 | [diff] [blame] | 900 | int slow_work_register_user(struct module *module) |
David Howells | 07fe7cb | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 901 | { |
| 902 | struct task_struct *p; |
| 903 | int loop; |
| 904 | |
| 905 | mutex_lock(&slow_work_user_lock); |
| 906 | |
| 907 | if (slow_work_user_count == 0) { |
| 908 | printk(KERN_NOTICE "Slow work thread pool: Starting up\n"); |
| 909 | init_completion(&slow_work_last_thread_exited); |
| 910 | |
| 911 | slow_work_threads_should_exit = false; |
David Howells | 109d927 | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 912 | slow_work_init(&slow_work_new_thread, |
| 913 | &slow_work_new_thread_ops); |
| 914 | slow_work_may_not_start_new_thread = false; |
| 915 | slow_work_cull = false; |
David Howells | 07fe7cb | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 916 | |
| 917 | /* start the minimum number of threads */ |
| 918 | for (loop = 0; loop < slow_work_min_threads; loop++) { |
| 919 | atomic_inc(&slow_work_thread_count); |
| 920 | p = kthread_run(slow_work_thread, NULL, "kslowd"); |
| 921 | if (IS_ERR(p)) |
| 922 | goto error; |
| 923 | } |
| 924 | printk(KERN_NOTICE "Slow work thread pool: Ready\n"); |
| 925 | } |
| 926 | |
| 927 | slow_work_user_count++; |
| 928 | mutex_unlock(&slow_work_user_lock); |
| 929 | return 0; |
| 930 | |
| 931 | error: |
| 932 | if (atomic_dec_and_test(&slow_work_thread_count)) |
| 933 | complete(&slow_work_last_thread_exited); |
| 934 | if (loop > 0) { |
| 935 | printk(KERN_ERR "Slow work thread pool:" |
| 936 | " Aborting startup on ENOMEM\n"); |
| 937 | slow_work_threads_should_exit = true; |
| 938 | wake_up_all(&slow_work_thread_wq); |
| 939 | wait_for_completion(&slow_work_last_thread_exited); |
| 940 | printk(KERN_ERR "Slow work thread pool: Aborted\n"); |
| 941 | } |
| 942 | mutex_unlock(&slow_work_user_lock); |
| 943 | return PTR_ERR(p); |
| 944 | } |
| 945 | EXPORT_SYMBOL(slow_work_register_user); |
| 946 | |
David Howells | 3d7a641 | 2009-11-19 18:10:23 +0000 | [diff] [blame] | 947 | /* |
| 948 | * wait for all outstanding items from the calling module to complete |
| 949 | * - note that more items may be queued whilst we're waiting |
| 950 | */ |
| 951 | static void slow_work_wait_for_items(struct module *module) |
| 952 | { |
David Howells | fa1dae4 | 2009-12-01 13:52:08 +0000 | [diff] [blame] | 953 | #ifdef CONFIG_MODULES |
David Howells | 3d7a641 | 2009-11-19 18:10:23 +0000 | [diff] [blame] | 954 | DECLARE_WAITQUEUE(myself, current); |
| 955 | struct slow_work *work; |
| 956 | int loop; |
| 957 | |
| 958 | mutex_lock(&slow_work_unreg_sync_lock); |
| 959 | add_wait_queue(&slow_work_unreg_wq, &myself); |
| 960 | |
| 961 | for (;;) { |
| 962 | spin_lock_irq(&slow_work_queue_lock); |
| 963 | |
| 964 | /* first of all, we wait for the last queued item in each list |
| 965 | * to be processed */ |
| 966 | list_for_each_entry_reverse(work, &vslow_work_queue, link) { |
| 967 | if (work->owner == module) { |
| 968 | set_current_state(TASK_UNINTERRUPTIBLE); |
| 969 | slow_work_unreg_work_item = work; |
| 970 | goto do_wait; |
| 971 | } |
| 972 | } |
| 973 | list_for_each_entry_reverse(work, &slow_work_queue, link) { |
| 974 | if (work->owner == module) { |
| 975 | set_current_state(TASK_UNINTERRUPTIBLE); |
| 976 | slow_work_unreg_work_item = work; |
| 977 | goto do_wait; |
| 978 | } |
| 979 | } |
| 980 | |
| 981 | /* then we wait for the items being processed to finish */ |
| 982 | slow_work_unreg_module = module; |
| 983 | smp_mb(); |
| 984 | for (loop = 0; loop < SLOW_WORK_THREAD_LIMIT; loop++) { |
| 985 | if (slow_work_thread_processing[loop] == module) |
| 986 | goto do_wait; |
| 987 | } |
| 988 | spin_unlock_irq(&slow_work_queue_lock); |
| 989 | break; /* okay, we're done */ |
| 990 | |
| 991 | do_wait: |
| 992 | spin_unlock_irq(&slow_work_queue_lock); |
| 993 | schedule(); |
| 994 | slow_work_unreg_work_item = NULL; |
| 995 | slow_work_unreg_module = NULL; |
| 996 | } |
| 997 | |
| 998 | remove_wait_queue(&slow_work_unreg_wq, &myself); |
| 999 | mutex_unlock(&slow_work_unreg_sync_lock); |
David Howells | fa1dae4 | 2009-12-01 13:52:08 +0000 | [diff] [blame] | 1000 | #endif /* CONFIG_MODULES */ |
David Howells | 3d7a641 | 2009-11-19 18:10:23 +0000 | [diff] [blame] | 1001 | } |
| 1002 | |
David Howells | 07fe7cb | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 1003 | /** |
| 1004 | * slow_work_unregister_user - Unregister a user of the facility |
David Howells | 3d7a641 | 2009-11-19 18:10:23 +0000 | [diff] [blame] | 1005 | * @module: The module whose items should be cleared |
David Howells | 07fe7cb | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 1006 | * |
| 1007 | * Unregister a user of the facility, killing all the threads if this was the |
| 1008 | * last one. |
David Howells | 3d7a641 | 2009-11-19 18:10:23 +0000 | [diff] [blame] | 1009 | * |
| 1010 | * This waits for all the work items belonging to the nominated module to go |
| 1011 | * away before proceeding. |
David Howells | 07fe7cb | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 1012 | */ |
David Howells | 3d7a641 | 2009-11-19 18:10:23 +0000 | [diff] [blame] | 1013 | void slow_work_unregister_user(struct module *module) |
David Howells | 07fe7cb | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 1014 | { |
David Howells | 3d7a641 | 2009-11-19 18:10:23 +0000 | [diff] [blame] | 1015 | /* first of all, wait for all outstanding items from the calling module |
| 1016 | * to complete */ |
| 1017 | if (module) |
| 1018 | slow_work_wait_for_items(module); |
| 1019 | |
| 1020 | /* then we can actually go about shutting down the facility if need |
| 1021 | * be */ |
David Howells | 07fe7cb | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 1022 | mutex_lock(&slow_work_user_lock); |
| 1023 | |
| 1024 | BUG_ON(slow_work_user_count <= 0); |
| 1025 | |
| 1026 | slow_work_user_count--; |
| 1027 | if (slow_work_user_count == 0) { |
| 1028 | printk(KERN_NOTICE "Slow work thread pool: Shutting down\n"); |
| 1029 | slow_work_threads_should_exit = true; |
Jonathan Corbet | 418df63c | 2009-04-22 12:01:49 +0100 | [diff] [blame] | 1030 | del_timer_sync(&slow_work_cull_timer); |
| 1031 | del_timer_sync(&slow_work_oom_timer); |
David Howells | 07fe7cb | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 1032 | wake_up_all(&slow_work_thread_wq); |
| 1033 | wait_for_completion(&slow_work_last_thread_exited); |
| 1034 | printk(KERN_NOTICE "Slow work thread pool:" |
| 1035 | " Shut down complete\n"); |
| 1036 | } |
| 1037 | |
David Howells | 07fe7cb | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 1038 | mutex_unlock(&slow_work_user_lock); |
| 1039 | } |
| 1040 | EXPORT_SYMBOL(slow_work_unregister_user); |
| 1041 | |
| 1042 | /* |
| 1043 | * Initialise the slow work facility |
| 1044 | */ |
| 1045 | static int __init init_slow_work(void) |
| 1046 | { |
| 1047 | unsigned nr_cpus = num_possible_cpus(); |
| 1048 | |
David Howells | 12e22c5 | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 1049 | if (slow_work_max_threads < nr_cpus) |
David Howells | 07fe7cb | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 1050 | slow_work_max_threads = nr_cpus; |
David Howells | 12e22c5 | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 1051 | #ifdef CONFIG_SYSCTL |
| 1052 | if (slow_work_max_max_threads < nr_cpus * 2) |
| 1053 | slow_work_max_max_threads = nr_cpus * 2; |
| 1054 | #endif |
David Howells | f13a48b | 2009-12-01 15:36:11 +0000 | [diff] [blame] | 1055 | #ifdef CONFIG_SLOW_WORK_DEBUG |
| 1056 | { |
| 1057 | struct dentry *dbdir; |
| 1058 | |
| 1059 | dbdir = debugfs_create_dir("slow_work", NULL); |
| 1060 | if (dbdir && !IS_ERR(dbdir)) |
| 1061 | debugfs_create_file("runqueue", S_IFREG | 0400, dbdir, |
| 1062 | NULL, &slow_work_runqueue_fops); |
| 1063 | } |
David Howells | 8fba10a | 2009-11-19 18:10:51 +0000 | [diff] [blame] | 1064 | #endif |
David Howells | 07fe7cb | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 1065 | return 0; |
| 1066 | } |
| 1067 | |
| 1068 | subsys_initcall(init_slow_work); |