Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (C) 2007 Oracle. All rights reserved.
4 * Copyright (C) 2014 Fujitsu. All rights reserved.
5 */
6
7#include <linux/kthread.h>
8#include <linux/slab.h>
9#include <linux/list.h>
10#include <linux/spinlock.h>
11#include <linux/freezer.h>
12#include <trace/events/btrfs.h>
13#include "async-thread.h"
14#include "ctree.h"
15
16enum {
17 WORK_DONE_BIT,
18 WORK_ORDER_DONE_BIT,
19};
20
21#define NO_THRESHOLD (-1)
22#define DFT_THRESHOLD (32)
23
24struct btrfs_workqueue {
25 struct workqueue_struct *normal_wq;
26
27 /* File system this workqueue services */
28 struct btrfs_fs_info *fs_info;
29
30 /* List head pointing to ordered work list */
31 struct list_head ordered_list;
32
33 /* Spinlock for ordered_list */
34 spinlock_t list_lock;
35
36 /* Thresholding related variants */
37 atomic_t pending;
38
39 /* Up limit of concurrency workers */
40 int limit_active;
41
42 /* Current number of concurrency workers */
43 int current_active;
44
45 /* Threshold to change current_active */
46 int thresh;
47 unsigned int count;
48 spinlock_t thres_lock;
49};
50
51struct btrfs_fs_info * __pure btrfs_workqueue_owner(const struct btrfs_workqueue *wq)
52{
53 return wq->fs_info;
54}
55
56struct btrfs_fs_info * __pure btrfs_work_owner(const struct btrfs_work *work)
57{
58 return work->wq->fs_info;
59}
60
61bool btrfs_workqueue_normal_congested(const struct btrfs_workqueue *wq)
62{
63 /*
64 * We could compare wq->pending with num_online_cpus()
65 * to support "thresh == NO_THRESHOLD" case, but it requires
66 * moving up atomic_inc/dec in thresh_queue/exec_hook. Let's
67 * postpone it until someone needs the support of that case.
68 */
69 if (wq->thresh == NO_THRESHOLD)
70 return false;
71
72 return atomic_read(&wq->pending) > wq->thresh * 2;
73}
74
75static void btrfs_init_workqueue(struct btrfs_workqueue *wq,
76 struct btrfs_fs_info *fs_info)
77{
78 wq->fs_info = fs_info;
79 atomic_set(&wq->pending, 0);
80 INIT_LIST_HEAD(&wq->ordered_list);
81 spin_lock_init(&wq->list_lock);
82 spin_lock_init(&wq->thres_lock);
83}
84
85struct btrfs_workqueue *btrfs_alloc_workqueue(struct btrfs_fs_info *fs_info,
86 const char *name, unsigned int flags,
87 int limit_active, int thresh)
88{
89 struct btrfs_workqueue *ret = kzalloc(sizeof(*ret), GFP_KERNEL);
90
91 if (!ret)
92 return NULL;
93
94 btrfs_init_workqueue(ret, fs_info);
95
96 ret->limit_active = limit_active;
97 if (thresh == 0)
98 thresh = DFT_THRESHOLD;
99 /* For low threshold, disabling threshold is a better choice */
100 if (thresh < DFT_THRESHOLD) {
101 ret->current_active = limit_active;
102 ret->thresh = NO_THRESHOLD;
103 } else {
104 /*
105 * For threshold-able wq, let its concurrency grow on demand.
106 * Use minimal max_active at alloc time to reduce resource
107 * usage.
108 */
109 ret->current_active = 1;
110 ret->thresh = thresh;
111 }
112
113 ret->normal_wq = alloc_workqueue("btrfs-%s", flags, ret->current_active,
114 name);
115 if (!ret->normal_wq) {
116 kfree(ret);
117 return NULL;
118 }
119
120 trace_btrfs_workqueue_alloc(ret, name);
121 return ret;
122}
123
124struct btrfs_workqueue *btrfs_alloc_ordered_workqueue(
125 struct btrfs_fs_info *fs_info, const char *name,
126 unsigned int flags)
127{
128 struct btrfs_workqueue *ret;
129
130 ret = kzalloc(sizeof(*ret), GFP_KERNEL);
131 if (!ret)
132 return NULL;
133
134 btrfs_init_workqueue(ret, fs_info);
135
136 /* Ordered workqueues don't allow @max_active adjustments. */
137 ret->limit_active = 1;
138 ret->current_active = 1;
139 ret->thresh = NO_THRESHOLD;
140
141 ret->normal_wq = alloc_ordered_workqueue("btrfs-%s", flags, name);
142 if (!ret->normal_wq) {
143 kfree(ret);
144 return NULL;
145 }
146
147 trace_btrfs_workqueue_alloc(ret, name);
148 return ret;
149}
150
151/*
152 * Hook for threshold which will be called in btrfs_queue_work.
153 * This hook WILL be called in IRQ handler context,
154 * so workqueue_set_max_active MUST NOT be called in this hook
155 */
156static inline void thresh_queue_hook(struct btrfs_workqueue *wq)
157{
158 if (wq->thresh == NO_THRESHOLD)
159 return;
160 atomic_inc(&wq->pending);
161}
162
163/*
164 * Hook for threshold which will be called before executing the work,
165 * This hook is called in kthread content.
166 * So workqueue_set_max_active is called here.
167 */
168static inline void thresh_exec_hook(struct btrfs_workqueue *wq)
169{
170 int new_current_active;
171 long pending;
172 int need_change = 0;
173
174 if (wq->thresh == NO_THRESHOLD)
175 return;
176
177 atomic_dec(&wq->pending);
178 spin_lock(&wq->thres_lock);
179 /*
180 * Use wq->count to limit the calling frequency of
181 * workqueue_set_max_active.
182 */
183 wq->count++;
184 wq->count %= (wq->thresh / 4);
185 if (!wq->count)
186 goto out;
187 new_current_active = wq->current_active;
188
189 /*
190 * pending may be changed later, but it's OK since we really
191 * don't need it so accurate to calculate new_max_active.
192 */
193 pending = atomic_read(&wq->pending);
194 if (pending > wq->thresh)
195 new_current_active++;
196 if (pending < wq->thresh / 2)
197 new_current_active--;
198 new_current_active = clamp_val(new_current_active, 1, wq->limit_active);
199 if (new_current_active != wq->current_active) {
200 need_change = 1;
201 wq->current_active = new_current_active;
202 }
203out:
204 spin_unlock(&wq->thres_lock);
205
206 if (need_change) {
207 workqueue_set_max_active(wq->normal_wq, wq->current_active);
208 }
209}
210
211static void run_ordered_work(struct btrfs_workqueue *wq,
212 struct btrfs_work *self)
213{
214 struct list_head *list = &wq->ordered_list;
215 struct btrfs_work *work;
216 spinlock_t *lock = &wq->list_lock;
217 unsigned long flags;
218 bool free_self = false;
219
220 while (1) {
221 spin_lock_irqsave(lock, flags);
222 if (list_empty(list))
223 break;
224 work = list_entry(list->next, struct btrfs_work,
225 ordered_list);
226 if (!test_bit(WORK_DONE_BIT, &work->flags))
227 break;
228 /*
229 * Orders all subsequent loads after reading WORK_DONE_BIT,
230 * paired with the smp_mb__before_atomic in btrfs_work_helper
231 * this guarantees that the ordered function will see all
232 * updates from ordinary work function.
233 */
234 smp_rmb();
235
236 /*
237 * we are going to call the ordered done function, but
238 * we leave the work item on the list as a barrier so
239 * that later work items that are done don't have their
240 * functions called before this one returns
241 */
242 if (test_and_set_bit(WORK_ORDER_DONE_BIT, &work->flags))
243 break;
244 trace_btrfs_ordered_sched(work);
245 spin_unlock_irqrestore(lock, flags);
246 work->ordered_func(work, false);
247
248 /* now take the lock again and drop our item from the list */
249 spin_lock_irqsave(lock, flags);
250 list_del(&work->ordered_list);
251 spin_unlock_irqrestore(lock, flags);
252
253 if (work == self) {
254 /*
255 * This is the work item that the worker is currently
256 * executing.
257 *
258 * The kernel workqueue code guarantees non-reentrancy
259 * of work items. I.e., if a work item with the same
260 * address and work function is queued twice, the second
261 * execution is blocked until the first one finishes. A
262 * work item may be freed and recycled with the same
263 * work function; the workqueue code assumes that the
264 * original work item cannot depend on the recycled work
265 * item in that case (see find_worker_executing_work()).
266 *
267 * Note that different types of Btrfs work can depend on
268 * each other, and one type of work on one Btrfs
269 * filesystem may even depend on the same type of work
270 * on another Btrfs filesystem via, e.g., a loop device.
271 * Therefore, we must not allow the current work item to
272 * be recycled until we are really done, otherwise we
273 * break the above assumption and can deadlock.
274 */
275 free_self = true;
276 } else {
277 /*
278 * We don't want to call the ordered free functions with
279 * the lock held.
280 */
281 work->ordered_func(work, true);
282 /* NB: work must not be dereferenced past this point. */
283 trace_btrfs_all_work_done(wq->fs_info, work);
284 }
285 }
286 spin_unlock_irqrestore(lock, flags);
287
288 if (free_self) {
289 self->ordered_func(self, true);
290 /* NB: self must not be dereferenced past this point. */
291 trace_btrfs_all_work_done(wq->fs_info, self);
292 }
293}
294
295static void btrfs_work_helper(struct work_struct *normal_work)
296{
297 struct btrfs_work *work = container_of(normal_work, struct btrfs_work,
298 normal_work);
299 struct btrfs_workqueue *wq = work->wq;
300 int need_order = 0;
301
302 /*
303 * We should not touch things inside work in the following cases:
304 * 1) after work->func() if it has no ordered_func(..., true) to free
305 * Since the struct is freed in work->func().
306 * 2) after setting WORK_DONE_BIT
307 * The work may be freed in other threads almost instantly.
308 * So we save the needed things here.
309 */
310 if (work->ordered_func)
311 need_order = 1;
312
313 trace_btrfs_work_sched(work);
314 thresh_exec_hook(wq);
315 work->func(work);
316 if (need_order) {
317 /*
318 * Ensures all memory accesses done in the work function are
319 * ordered before setting the WORK_DONE_BIT. Ensuring the thread
320 * which is going to executed the ordered work sees them.
321 * Pairs with the smp_rmb in run_ordered_work.
322 */
323 smp_mb__before_atomic();
324 set_bit(WORK_DONE_BIT, &work->flags);
325 run_ordered_work(wq, work);
326 } else {
327 /* NB: work must not be dereferenced past this point. */
328 trace_btrfs_all_work_done(wq->fs_info, work);
329 }
330}
331
332void btrfs_init_work(struct btrfs_work *work, btrfs_func_t func,
333 btrfs_ordered_func_t ordered_func)
334{
335 work->func = func;
336 work->ordered_func = ordered_func;
337 INIT_WORK(&work->normal_work, btrfs_work_helper);
338 INIT_LIST_HEAD(&work->ordered_list);
339 work->flags = 0;
340}
341
342void btrfs_queue_work(struct btrfs_workqueue *wq, struct btrfs_work *work)
343{
344 unsigned long flags;
345
346 work->wq = wq;
347 thresh_queue_hook(wq);
348 if (work->ordered_func) {
349 spin_lock_irqsave(&wq->list_lock, flags);
350 list_add_tail(&work->ordered_list, &wq->ordered_list);
351 spin_unlock_irqrestore(&wq->list_lock, flags);
352 }
353 trace_btrfs_work_queued(work);
354 queue_work(wq->normal_wq, &work->normal_work);
355}
356
357void btrfs_destroy_workqueue(struct btrfs_workqueue *wq)
358{
359 if (!wq)
360 return;
361 destroy_workqueue(wq->normal_wq);
362 trace_btrfs_workqueue_destroy(wq);
363 kfree(wq);
364}
365
366void btrfs_workqueue_set_max(struct btrfs_workqueue *wq, int limit_active)
367{
368 if (wq)
369 wq->limit_active = limit_active;
370}
371
372void btrfs_flush_workqueue(struct btrfs_workqueue *wq)
373{
374 flush_workqueue(wq->normal_wq);
375}
1/*
2 * Copyright (C) 2007 Oracle. All rights reserved.
3 *
4 * This program is free software; you can redistribute it and/or
5 * modify it under the terms of the GNU General Public
6 * License v2 as published by the Free Software Foundation.
7 *
8 * This program is distributed in the hope that it will be useful,
9 * but WITHOUT ANY WARRANTY; without even the implied warranty of
10 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
11 * General Public License for more details.
12 *
13 * You should have received a copy of the GNU General Public
14 * License along with this program; if not, write to the
15 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
16 * Boston, MA 021110-1307, USA.
17 */
18
19#include <linux/kthread.h>
20#include <linux/slab.h>
21#include <linux/list.h>
22#include <linux/spinlock.h>
23#include <linux/freezer.h>
24#include "async-thread.h"
25
26#define WORK_QUEUED_BIT 0
27#define WORK_DONE_BIT 1
28#define WORK_ORDER_DONE_BIT 2
29#define WORK_HIGH_PRIO_BIT 3
30
31/*
32 * container for the kthread task pointer and the list of pending work
33 * One of these is allocated per thread.
34 */
35struct btrfs_worker_thread {
36 /* pool we belong to */
37 struct btrfs_workers *workers;
38
39 /* list of struct btrfs_work that are waiting for service */
40 struct list_head pending;
41 struct list_head prio_pending;
42
43 /* list of worker threads from struct btrfs_workers */
44 struct list_head worker_list;
45
46 /* kthread */
47 struct task_struct *task;
48
49 /* number of things on the pending list */
50 atomic_t num_pending;
51
52 /* reference counter for this struct */
53 atomic_t refs;
54
55 unsigned long sequence;
56
57 /* protects the pending list. */
58 spinlock_t lock;
59
60 /* set to non-zero when this thread is already awake and kicking */
61 int working;
62
63 /* are we currently idle */
64 int idle;
65};
66
67/*
68 * btrfs_start_workers uses kthread_run, which can block waiting for memory
69 * for a very long time. It will actually throttle on page writeback,
70 * and so it may not make progress until after our btrfs worker threads
71 * process all of the pending work structs in their queue
72 *
73 * This means we can't use btrfs_start_workers from inside a btrfs worker
74 * thread that is used as part of cleaning dirty memory, which pretty much
75 * involves all of the worker threads.
76 *
77 * Instead we have a helper queue who never has more than one thread
78 * where we scheduler thread start operations. This worker_start struct
79 * is used to contain the work and hold a pointer to the queue that needs
80 * another worker.
81 */
82struct worker_start {
83 struct btrfs_work work;
84 struct btrfs_workers *queue;
85};
86
87static void start_new_worker_func(struct btrfs_work *work)
88{
89 struct worker_start *start;
90 start = container_of(work, struct worker_start, work);
91 btrfs_start_workers(start->queue, 1);
92 kfree(start);
93}
94
95static int start_new_worker(struct btrfs_workers *queue)
96{
97 struct worker_start *start;
98 int ret;
99
100 start = kzalloc(sizeof(*start), GFP_NOFS);
101 if (!start)
102 return -ENOMEM;
103
104 start->work.func = start_new_worker_func;
105 start->queue = queue;
106 ret = btrfs_queue_worker(queue->atomic_worker_start, &start->work);
107 if (ret)
108 kfree(start);
109 return ret;
110}
111
112/*
113 * helper function to move a thread onto the idle list after it
114 * has finished some requests.
115 */
116static void check_idle_worker(struct btrfs_worker_thread *worker)
117{
118 if (!worker->idle && atomic_read(&worker->num_pending) <
119 worker->workers->idle_thresh / 2) {
120 unsigned long flags;
121 spin_lock_irqsave(&worker->workers->lock, flags);
122 worker->idle = 1;
123
124 /* the list may be empty if the worker is just starting */
125 if (!list_empty(&worker->worker_list)) {
126 list_move(&worker->worker_list,
127 &worker->workers->idle_list);
128 }
129 spin_unlock_irqrestore(&worker->workers->lock, flags);
130 }
131}
132
133/*
134 * helper function to move a thread off the idle list after new
135 * pending work is added.
136 */
137static void check_busy_worker(struct btrfs_worker_thread *worker)
138{
139 if (worker->idle && atomic_read(&worker->num_pending) >=
140 worker->workers->idle_thresh) {
141 unsigned long flags;
142 spin_lock_irqsave(&worker->workers->lock, flags);
143 worker->idle = 0;
144
145 if (!list_empty(&worker->worker_list)) {
146 list_move_tail(&worker->worker_list,
147 &worker->workers->worker_list);
148 }
149 spin_unlock_irqrestore(&worker->workers->lock, flags);
150 }
151}
152
153static void check_pending_worker_creates(struct btrfs_worker_thread *worker)
154{
155 struct btrfs_workers *workers = worker->workers;
156 unsigned long flags;
157
158 rmb();
159 if (!workers->atomic_start_pending)
160 return;
161
162 spin_lock_irqsave(&workers->lock, flags);
163 if (!workers->atomic_start_pending)
164 goto out;
165
166 workers->atomic_start_pending = 0;
167 if (workers->num_workers + workers->num_workers_starting >=
168 workers->max_workers)
169 goto out;
170
171 workers->num_workers_starting += 1;
172 spin_unlock_irqrestore(&workers->lock, flags);
173 start_new_worker(workers);
174 return;
175
176out:
177 spin_unlock_irqrestore(&workers->lock, flags);
178}
179
180static noinline int run_ordered_completions(struct btrfs_workers *workers,
181 struct btrfs_work *work)
182{
183 if (!workers->ordered)
184 return 0;
185
186 set_bit(WORK_DONE_BIT, &work->flags);
187
188 spin_lock(&workers->order_lock);
189
190 while (1) {
191 if (!list_empty(&workers->prio_order_list)) {
192 work = list_entry(workers->prio_order_list.next,
193 struct btrfs_work, order_list);
194 } else if (!list_empty(&workers->order_list)) {
195 work = list_entry(workers->order_list.next,
196 struct btrfs_work, order_list);
197 } else {
198 break;
199 }
200 if (!test_bit(WORK_DONE_BIT, &work->flags))
201 break;
202
203 /* we are going to call the ordered done function, but
204 * we leave the work item on the list as a barrier so
205 * that later work items that are done don't have their
206 * functions called before this one returns
207 */
208 if (test_and_set_bit(WORK_ORDER_DONE_BIT, &work->flags))
209 break;
210
211 spin_unlock(&workers->order_lock);
212
213 work->ordered_func(work);
214
215 /* now take the lock again and call the freeing code */
216 spin_lock(&workers->order_lock);
217 list_del(&work->order_list);
218 work->ordered_free(work);
219 }
220
221 spin_unlock(&workers->order_lock);
222 return 0;
223}
224
225static void put_worker(struct btrfs_worker_thread *worker)
226{
227 if (atomic_dec_and_test(&worker->refs))
228 kfree(worker);
229}
230
231static int try_worker_shutdown(struct btrfs_worker_thread *worker)
232{
233 int freeit = 0;
234
235 spin_lock_irq(&worker->lock);
236 spin_lock(&worker->workers->lock);
237 if (worker->workers->num_workers > 1 &&
238 worker->idle &&
239 !worker->working &&
240 !list_empty(&worker->worker_list) &&
241 list_empty(&worker->prio_pending) &&
242 list_empty(&worker->pending) &&
243 atomic_read(&worker->num_pending) == 0) {
244 freeit = 1;
245 list_del_init(&worker->worker_list);
246 worker->workers->num_workers--;
247 }
248 spin_unlock(&worker->workers->lock);
249 spin_unlock_irq(&worker->lock);
250
251 if (freeit)
252 put_worker(worker);
253 return freeit;
254}
255
256static struct btrfs_work *get_next_work(struct btrfs_worker_thread *worker,
257 struct list_head *prio_head,
258 struct list_head *head)
259{
260 struct btrfs_work *work = NULL;
261 struct list_head *cur = NULL;
262
263 if(!list_empty(prio_head))
264 cur = prio_head->next;
265
266 smp_mb();
267 if (!list_empty(&worker->prio_pending))
268 goto refill;
269
270 if (!list_empty(head))
271 cur = head->next;
272
273 if (cur)
274 goto out;
275
276refill:
277 spin_lock_irq(&worker->lock);
278 list_splice_tail_init(&worker->prio_pending, prio_head);
279 list_splice_tail_init(&worker->pending, head);
280
281 if (!list_empty(prio_head))
282 cur = prio_head->next;
283 else if (!list_empty(head))
284 cur = head->next;
285 spin_unlock_irq(&worker->lock);
286
287 if (!cur)
288 goto out_fail;
289
290out:
291 work = list_entry(cur, struct btrfs_work, list);
292
293out_fail:
294 return work;
295}
296
297/*
298 * main loop for servicing work items
299 */
300static int worker_loop(void *arg)
301{
302 struct btrfs_worker_thread *worker = arg;
303 struct list_head head;
304 struct list_head prio_head;
305 struct btrfs_work *work;
306
307 INIT_LIST_HEAD(&head);
308 INIT_LIST_HEAD(&prio_head);
309
310 do {
311again:
312 while (1) {
313
314
315 work = get_next_work(worker, &prio_head, &head);
316 if (!work)
317 break;
318
319 list_del(&work->list);
320 clear_bit(WORK_QUEUED_BIT, &work->flags);
321
322 work->worker = worker;
323
324 work->func(work);
325
326 atomic_dec(&worker->num_pending);
327 /*
328 * unless this is an ordered work queue,
329 * 'work' was probably freed by func above.
330 */
331 run_ordered_completions(worker->workers, work);
332
333 check_pending_worker_creates(worker);
334
335 }
336
337 spin_lock_irq(&worker->lock);
338 check_idle_worker(worker);
339
340 if (freezing(current)) {
341 worker->working = 0;
342 spin_unlock_irq(&worker->lock);
343 refrigerator();
344 } else {
345 spin_unlock_irq(&worker->lock);
346 if (!kthread_should_stop()) {
347 cpu_relax();
348 /*
349 * we've dropped the lock, did someone else
350 * jump_in?
351 */
352 smp_mb();
353 if (!list_empty(&worker->pending) ||
354 !list_empty(&worker->prio_pending))
355 continue;
356
357 /*
358 * this short schedule allows more work to
359 * come in without the queue functions
360 * needing to go through wake_up_process()
361 *
362 * worker->working is still 1, so nobody
363 * is going to try and wake us up
364 */
365 schedule_timeout(1);
366 smp_mb();
367 if (!list_empty(&worker->pending) ||
368 !list_empty(&worker->prio_pending))
369 continue;
370
371 if (kthread_should_stop())
372 break;
373
374 /* still no more work?, sleep for real */
375 spin_lock_irq(&worker->lock);
376 set_current_state(TASK_INTERRUPTIBLE);
377 if (!list_empty(&worker->pending) ||
378 !list_empty(&worker->prio_pending)) {
379 spin_unlock_irq(&worker->lock);
380 set_current_state(TASK_RUNNING);
381 goto again;
382 }
383
384 /*
385 * this makes sure we get a wakeup when someone
386 * adds something new to the queue
387 */
388 worker->working = 0;
389 spin_unlock_irq(&worker->lock);
390
391 if (!kthread_should_stop()) {
392 schedule_timeout(HZ * 120);
393 if (!worker->working &&
394 try_worker_shutdown(worker)) {
395 return 0;
396 }
397 }
398 }
399 __set_current_state(TASK_RUNNING);
400 }
401 } while (!kthread_should_stop());
402 return 0;
403}
404
405/*
406 * this will wait for all the worker threads to shutdown
407 */
408int btrfs_stop_workers(struct btrfs_workers *workers)
409{
410 struct list_head *cur;
411 struct btrfs_worker_thread *worker;
412 int can_stop;
413
414 spin_lock_irq(&workers->lock);
415 list_splice_init(&workers->idle_list, &workers->worker_list);
416 while (!list_empty(&workers->worker_list)) {
417 cur = workers->worker_list.next;
418 worker = list_entry(cur, struct btrfs_worker_thread,
419 worker_list);
420
421 atomic_inc(&worker->refs);
422 workers->num_workers -= 1;
423 if (!list_empty(&worker->worker_list)) {
424 list_del_init(&worker->worker_list);
425 put_worker(worker);
426 can_stop = 1;
427 } else
428 can_stop = 0;
429 spin_unlock_irq(&workers->lock);
430 if (can_stop)
431 kthread_stop(worker->task);
432 spin_lock_irq(&workers->lock);
433 put_worker(worker);
434 }
435 spin_unlock_irq(&workers->lock);
436 return 0;
437}
438
439/*
440 * simple init on struct btrfs_workers
441 */
442void btrfs_init_workers(struct btrfs_workers *workers, char *name, int max,
443 struct btrfs_workers *async_helper)
444{
445 workers->num_workers = 0;
446 workers->num_workers_starting = 0;
447 INIT_LIST_HEAD(&workers->worker_list);
448 INIT_LIST_HEAD(&workers->idle_list);
449 INIT_LIST_HEAD(&workers->order_list);
450 INIT_LIST_HEAD(&workers->prio_order_list);
451 spin_lock_init(&workers->lock);
452 spin_lock_init(&workers->order_lock);
453 workers->max_workers = max;
454 workers->idle_thresh = 32;
455 workers->name = name;
456 workers->ordered = 0;
457 workers->atomic_start_pending = 0;
458 workers->atomic_worker_start = async_helper;
459}
460
461/*
462 * starts new worker threads. This does not enforce the max worker
463 * count in case you need to temporarily go past it.
464 */
465static int __btrfs_start_workers(struct btrfs_workers *workers,
466 int num_workers)
467{
468 struct btrfs_worker_thread *worker;
469 int ret = 0;
470 int i;
471
472 for (i = 0; i < num_workers; i++) {
473 worker = kzalloc(sizeof(*worker), GFP_NOFS);
474 if (!worker) {
475 ret = -ENOMEM;
476 goto fail;
477 }
478
479 INIT_LIST_HEAD(&worker->pending);
480 INIT_LIST_HEAD(&worker->prio_pending);
481 INIT_LIST_HEAD(&worker->worker_list);
482 spin_lock_init(&worker->lock);
483
484 atomic_set(&worker->num_pending, 0);
485 atomic_set(&worker->refs, 1);
486 worker->workers = workers;
487 worker->task = kthread_run(worker_loop, worker,
488 "btrfs-%s-%d", workers->name,
489 workers->num_workers + i);
490 if (IS_ERR(worker->task)) {
491 ret = PTR_ERR(worker->task);
492 kfree(worker);
493 goto fail;
494 }
495 spin_lock_irq(&workers->lock);
496 list_add_tail(&worker->worker_list, &workers->idle_list);
497 worker->idle = 1;
498 workers->num_workers++;
499 workers->num_workers_starting--;
500 WARN_ON(workers->num_workers_starting < 0);
501 spin_unlock_irq(&workers->lock);
502 }
503 return 0;
504fail:
505 btrfs_stop_workers(workers);
506 return ret;
507}
508
509int btrfs_start_workers(struct btrfs_workers *workers, int num_workers)
510{
511 spin_lock_irq(&workers->lock);
512 workers->num_workers_starting += num_workers;
513 spin_unlock_irq(&workers->lock);
514 return __btrfs_start_workers(workers, num_workers);
515}
516
517/*
518 * run through the list and find a worker thread that doesn't have a lot
519 * to do right now. This can return null if we aren't yet at the thread
520 * count limit and all of the threads are busy.
521 */
522static struct btrfs_worker_thread *next_worker(struct btrfs_workers *workers)
523{
524 struct btrfs_worker_thread *worker;
525 struct list_head *next;
526 int enforce_min;
527
528 enforce_min = (workers->num_workers + workers->num_workers_starting) <
529 workers->max_workers;
530
531 /*
532 * if we find an idle thread, don't move it to the end of the
533 * idle list. This improves the chance that the next submission
534 * will reuse the same thread, and maybe catch it while it is still
535 * working
536 */
537 if (!list_empty(&workers->idle_list)) {
538 next = workers->idle_list.next;
539 worker = list_entry(next, struct btrfs_worker_thread,
540 worker_list);
541 return worker;
542 }
543 if (enforce_min || list_empty(&workers->worker_list))
544 return NULL;
545
546 /*
547 * if we pick a busy task, move the task to the end of the list.
548 * hopefully this will keep things somewhat evenly balanced.
549 * Do the move in batches based on the sequence number. This groups
550 * requests submitted at roughly the same time onto the same worker.
551 */
552 next = workers->worker_list.next;
553 worker = list_entry(next, struct btrfs_worker_thread, worker_list);
554 worker->sequence++;
555
556 if (worker->sequence % workers->idle_thresh == 0)
557 list_move_tail(next, &workers->worker_list);
558 return worker;
559}
560
561/*
562 * selects a worker thread to take the next job. This will either find
563 * an idle worker, start a new worker up to the max count, or just return
564 * one of the existing busy workers.
565 */
566static struct btrfs_worker_thread *find_worker(struct btrfs_workers *workers)
567{
568 struct btrfs_worker_thread *worker;
569 unsigned long flags;
570 struct list_head *fallback;
571
572again:
573 spin_lock_irqsave(&workers->lock, flags);
574 worker = next_worker(workers);
575
576 if (!worker) {
577 if (workers->num_workers + workers->num_workers_starting >=
578 workers->max_workers) {
579 goto fallback;
580 } else if (workers->atomic_worker_start) {
581 workers->atomic_start_pending = 1;
582 goto fallback;
583 } else {
584 workers->num_workers_starting++;
585 spin_unlock_irqrestore(&workers->lock, flags);
586 /* we're below the limit, start another worker */
587 __btrfs_start_workers(workers, 1);
588 goto again;
589 }
590 }
591 goto found;
592
593fallback:
594 fallback = NULL;
595 /*
596 * we have failed to find any workers, just
597 * return the first one we can find.
598 */
599 if (!list_empty(&workers->worker_list))
600 fallback = workers->worker_list.next;
601 if (!list_empty(&workers->idle_list))
602 fallback = workers->idle_list.next;
603 BUG_ON(!fallback);
604 worker = list_entry(fallback,
605 struct btrfs_worker_thread, worker_list);
606found:
607 /*
608 * this makes sure the worker doesn't exit before it is placed
609 * onto a busy/idle list
610 */
611 atomic_inc(&worker->num_pending);
612 spin_unlock_irqrestore(&workers->lock, flags);
613 return worker;
614}
615
616/*
617 * btrfs_requeue_work just puts the work item back on the tail of the list
618 * it was taken from. It is intended for use with long running work functions
619 * that make some progress and want to give the cpu up for others.
620 */
621int btrfs_requeue_work(struct btrfs_work *work)
622{
623 struct btrfs_worker_thread *worker = work->worker;
624 unsigned long flags;
625 int wake = 0;
626
627 if (test_and_set_bit(WORK_QUEUED_BIT, &work->flags))
628 goto out;
629
630 spin_lock_irqsave(&worker->lock, flags);
631 if (test_bit(WORK_HIGH_PRIO_BIT, &work->flags))
632 list_add_tail(&work->list, &worker->prio_pending);
633 else
634 list_add_tail(&work->list, &worker->pending);
635 atomic_inc(&worker->num_pending);
636
637 /* by definition we're busy, take ourselves off the idle
638 * list
639 */
640 if (worker->idle) {
641 spin_lock(&worker->workers->lock);
642 worker->idle = 0;
643 list_move_tail(&worker->worker_list,
644 &worker->workers->worker_list);
645 spin_unlock(&worker->workers->lock);
646 }
647 if (!worker->working) {
648 wake = 1;
649 worker->working = 1;
650 }
651
652 if (wake)
653 wake_up_process(worker->task);
654 spin_unlock_irqrestore(&worker->lock, flags);
655out:
656
657 return 0;
658}
659
660void btrfs_set_work_high_prio(struct btrfs_work *work)
661{
662 set_bit(WORK_HIGH_PRIO_BIT, &work->flags);
663}
664
665/*
666 * places a struct btrfs_work into the pending queue of one of the kthreads
667 */
668int btrfs_queue_worker(struct btrfs_workers *workers, struct btrfs_work *work)
669{
670 struct btrfs_worker_thread *worker;
671 unsigned long flags;
672 int wake = 0;
673
674 /* don't requeue something already on a list */
675 if (test_and_set_bit(WORK_QUEUED_BIT, &work->flags))
676 goto out;
677
678 worker = find_worker(workers);
679 if (workers->ordered) {
680 /*
681 * you're not allowed to do ordered queues from an
682 * interrupt handler
683 */
684 spin_lock(&workers->order_lock);
685 if (test_bit(WORK_HIGH_PRIO_BIT, &work->flags)) {
686 list_add_tail(&work->order_list,
687 &workers->prio_order_list);
688 } else {
689 list_add_tail(&work->order_list, &workers->order_list);
690 }
691 spin_unlock(&workers->order_lock);
692 } else {
693 INIT_LIST_HEAD(&work->order_list);
694 }
695
696 spin_lock_irqsave(&worker->lock, flags);
697
698 if (test_bit(WORK_HIGH_PRIO_BIT, &work->flags))
699 list_add_tail(&work->list, &worker->prio_pending);
700 else
701 list_add_tail(&work->list, &worker->pending);
702 check_busy_worker(worker);
703
704 /*
705 * avoid calling into wake_up_process if this thread has already
706 * been kicked
707 */
708 if (!worker->working)
709 wake = 1;
710 worker->working = 1;
711
712 if (wake)
713 wake_up_process(worker->task);
714 spin_unlock_irqrestore(&worker->lock, flags);
715
716out:
717 return 0;
718}