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 "async-thread.h"
13#include "ctree.h"
14
15enum {
16 WORK_DONE_BIT,
17 WORK_ORDER_DONE_BIT,
18 WORK_HIGH_PRIO_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_workqueue {
52 struct __btrfs_workqueue *normal;
53 struct __btrfs_workqueue *high;
54};
55
56struct btrfs_fs_info * __pure btrfs_workqueue_owner(const struct __btrfs_workqueue *wq)
57{
58 return wq->fs_info;
59}
60
61struct btrfs_fs_info * __pure btrfs_work_owner(const struct btrfs_work *work)
62{
63 return work->wq->fs_info;
64}
65
66bool btrfs_workqueue_normal_congested(const struct btrfs_workqueue *wq)
67{
68 /*
69 * We could compare wq->normal->pending with num_online_cpus()
70 * to support "thresh == NO_THRESHOLD" case, but it requires
71 * moving up atomic_inc/dec in thresh_queue/exec_hook. Let's
72 * postpone it until someone needs the support of that case.
73 */
74 if (wq->normal->thresh == NO_THRESHOLD)
75 return false;
76
77 return atomic_read(&wq->normal->pending) > wq->normal->thresh * 2;
78}
79
80static struct __btrfs_workqueue *
81__btrfs_alloc_workqueue(struct btrfs_fs_info *fs_info, const char *name,
82 unsigned int flags, int limit_active, int thresh)
83{
84 struct __btrfs_workqueue *ret = kzalloc(sizeof(*ret), GFP_KERNEL);
85
86 if (!ret)
87 return NULL;
88
89 ret->fs_info = fs_info;
90 ret->limit_active = limit_active;
91 atomic_set(&ret->pending, 0);
92 if (thresh == 0)
93 thresh = DFT_THRESHOLD;
94 /* For low threshold, disabling threshold is a better choice */
95 if (thresh < DFT_THRESHOLD) {
96 ret->current_active = limit_active;
97 ret->thresh = NO_THRESHOLD;
98 } else {
99 /*
100 * For threshold-able wq, let its concurrency grow on demand.
101 * Use minimal max_active at alloc time to reduce resource
102 * usage.
103 */
104 ret->current_active = 1;
105 ret->thresh = thresh;
106 }
107
108 if (flags & WQ_HIGHPRI)
109 ret->normal_wq = alloc_workqueue("btrfs-%s-high", flags,
110 ret->current_active, name);
111 else
112 ret->normal_wq = alloc_workqueue("btrfs-%s", flags,
113 ret->current_active, name);
114 if (!ret->normal_wq) {
115 kfree(ret);
116 return NULL;
117 }
118
119 INIT_LIST_HEAD(&ret->ordered_list);
120 spin_lock_init(&ret->list_lock);
121 spin_lock_init(&ret->thres_lock);
122 trace_btrfs_workqueue_alloc(ret, name, flags & WQ_HIGHPRI);
123 return ret;
124}
125
126static inline void
127__btrfs_destroy_workqueue(struct __btrfs_workqueue *wq);
128
129struct btrfs_workqueue *btrfs_alloc_workqueue(struct btrfs_fs_info *fs_info,
130 const char *name,
131 unsigned int flags,
132 int limit_active,
133 int thresh)
134{
135 struct btrfs_workqueue *ret = kzalloc(sizeof(*ret), GFP_KERNEL);
136
137 if (!ret)
138 return NULL;
139
140 ret->normal = __btrfs_alloc_workqueue(fs_info, name,
141 flags & ~WQ_HIGHPRI,
142 limit_active, thresh);
143 if (!ret->normal) {
144 kfree(ret);
145 return NULL;
146 }
147
148 if (flags & WQ_HIGHPRI) {
149 ret->high = __btrfs_alloc_workqueue(fs_info, name, flags,
150 limit_active, thresh);
151 if (!ret->high) {
152 __btrfs_destroy_workqueue(ret->normal);
153 kfree(ret);
154 return NULL;
155 }
156 }
157 return ret;
158}
159
160/*
161 * Hook for threshold which will be called in btrfs_queue_work.
162 * This hook WILL be called in IRQ handler context,
163 * so workqueue_set_max_active MUST NOT be called in this hook
164 */
165static inline void thresh_queue_hook(struct __btrfs_workqueue *wq)
166{
167 if (wq->thresh == NO_THRESHOLD)
168 return;
169 atomic_inc(&wq->pending);
170}
171
172/*
173 * Hook for threshold which will be called before executing the work,
174 * This hook is called in kthread content.
175 * So workqueue_set_max_active is called here.
176 */
177static inline void thresh_exec_hook(struct __btrfs_workqueue *wq)
178{
179 int new_current_active;
180 long pending;
181 int need_change = 0;
182
183 if (wq->thresh == NO_THRESHOLD)
184 return;
185
186 atomic_dec(&wq->pending);
187 spin_lock(&wq->thres_lock);
188 /*
189 * Use wq->count to limit the calling frequency of
190 * workqueue_set_max_active.
191 */
192 wq->count++;
193 wq->count %= (wq->thresh / 4);
194 if (!wq->count)
195 goto out;
196 new_current_active = wq->current_active;
197
198 /*
199 * pending may be changed later, but it's OK since we really
200 * don't need it so accurate to calculate new_max_active.
201 */
202 pending = atomic_read(&wq->pending);
203 if (pending > wq->thresh)
204 new_current_active++;
205 if (pending < wq->thresh / 2)
206 new_current_active--;
207 new_current_active = clamp_val(new_current_active, 1, wq->limit_active);
208 if (new_current_active != wq->current_active) {
209 need_change = 1;
210 wq->current_active = new_current_active;
211 }
212out:
213 spin_unlock(&wq->thres_lock);
214
215 if (need_change) {
216 workqueue_set_max_active(wq->normal_wq, wq->current_active);
217 }
218}
219
220static void run_ordered_work(struct __btrfs_workqueue *wq,
221 struct btrfs_work *self)
222{
223 struct list_head *list = &wq->ordered_list;
224 struct btrfs_work *work;
225 spinlock_t *lock = &wq->list_lock;
226 unsigned long flags;
227 bool free_self = false;
228
229 while (1) {
230 spin_lock_irqsave(lock, flags);
231 if (list_empty(list))
232 break;
233 work = list_entry(list->next, struct btrfs_work,
234 ordered_list);
235 if (!test_bit(WORK_DONE_BIT, &work->flags))
236 break;
237
238 /*
239 * we are going to call the ordered done function, but
240 * we leave the work item on the list as a barrier so
241 * that later work items that are done don't have their
242 * functions called before this one returns
243 */
244 if (test_and_set_bit(WORK_ORDER_DONE_BIT, &work->flags))
245 break;
246 trace_btrfs_ordered_sched(work);
247 spin_unlock_irqrestore(lock, flags);
248 work->ordered_func(work);
249
250 /* now take the lock again and drop our item from the list */
251 spin_lock_irqsave(lock, flags);
252 list_del(&work->ordered_list);
253 spin_unlock_irqrestore(lock, flags);
254
255 if (work == self) {
256 /*
257 * This is the work item that the worker is currently
258 * executing.
259 *
260 * The kernel workqueue code guarantees non-reentrancy
261 * of work items. I.e., if a work item with the same
262 * address and work function is queued twice, the second
263 * execution is blocked until the first one finishes. A
264 * work item may be freed and recycled with the same
265 * work function; the workqueue code assumes that the
266 * original work item cannot depend on the recycled work
267 * item in that case (see find_worker_executing_work()).
268 *
269 * Note that different types of Btrfs work can depend on
270 * each other, and one type of work on one Btrfs
271 * filesystem may even depend on the same type of work
272 * on another Btrfs filesystem via, e.g., a loop device.
273 * Therefore, we must not allow the current work item to
274 * be recycled until we are really done, otherwise we
275 * break the above assumption and can deadlock.
276 */
277 free_self = true;
278 } else {
279 /*
280 * We don't want to call the ordered free functions with
281 * the lock held.
282 */
283 work->ordered_free(work);
284 /* NB: work must not be dereferenced past this point. */
285 trace_btrfs_all_work_done(wq->fs_info, work);
286 }
287 }
288 spin_unlock_irqrestore(lock, flags);
289
290 if (free_self) {
291 self->ordered_free(self);
292 /* NB: self must not be dereferenced past this point. */
293 trace_btrfs_all_work_done(wq->fs_info, self);
294 }
295}
296
297static void btrfs_work_helper(struct work_struct *normal_work)
298{
299 struct btrfs_work *work = container_of(normal_work, struct btrfs_work,
300 normal_work);
301 struct __btrfs_workqueue *wq;
302 int need_order = 0;
303
304 /*
305 * We should not touch things inside work in the following cases:
306 * 1) after work->func() if it has no ordered_free
307 * Since the struct is freed in work->func().
308 * 2) after setting WORK_DONE_BIT
309 * The work may be freed in other threads almost instantly.
310 * So we save the needed things here.
311 */
312 if (work->ordered_func)
313 need_order = 1;
314 wq = work->wq;
315
316 trace_btrfs_work_sched(work);
317 thresh_exec_hook(wq);
318 work->func(work);
319 if (need_order) {
320 set_bit(WORK_DONE_BIT, &work->flags);
321 run_ordered_work(wq, work);
322 } else {
323 /* NB: work must not be dereferenced past this point. */
324 trace_btrfs_all_work_done(wq->fs_info, work);
325 }
326}
327
328void btrfs_init_work(struct btrfs_work *work, btrfs_func_t func,
329 btrfs_func_t ordered_func, btrfs_func_t ordered_free)
330{
331 work->func = func;
332 work->ordered_func = ordered_func;
333 work->ordered_free = ordered_free;
334 INIT_WORK(&work->normal_work, btrfs_work_helper);
335 INIT_LIST_HEAD(&work->ordered_list);
336 work->flags = 0;
337}
338
339static inline void __btrfs_queue_work(struct __btrfs_workqueue *wq,
340 struct btrfs_work *work)
341{
342 unsigned long flags;
343
344 work->wq = wq;
345 thresh_queue_hook(wq);
346 if (work->ordered_func) {
347 spin_lock_irqsave(&wq->list_lock, flags);
348 list_add_tail(&work->ordered_list, &wq->ordered_list);
349 spin_unlock_irqrestore(&wq->list_lock, flags);
350 }
351 trace_btrfs_work_queued(work);
352 queue_work(wq->normal_wq, &work->normal_work);
353}
354
355void btrfs_queue_work(struct btrfs_workqueue *wq,
356 struct btrfs_work *work)
357{
358 struct __btrfs_workqueue *dest_wq;
359
360 if (test_bit(WORK_HIGH_PRIO_BIT, &work->flags) && wq->high)
361 dest_wq = wq->high;
362 else
363 dest_wq = wq->normal;
364 __btrfs_queue_work(dest_wq, work);
365}
366
367static inline void
368__btrfs_destroy_workqueue(struct __btrfs_workqueue *wq)
369{
370 destroy_workqueue(wq->normal_wq);
371 trace_btrfs_workqueue_destroy(wq);
372 kfree(wq);
373}
374
375void btrfs_destroy_workqueue(struct btrfs_workqueue *wq)
376{
377 if (!wq)
378 return;
379 if (wq->high)
380 __btrfs_destroy_workqueue(wq->high);
381 __btrfs_destroy_workqueue(wq->normal);
382 kfree(wq);
383}
384
385void btrfs_workqueue_set_max(struct btrfs_workqueue *wq, int limit_active)
386{
387 if (!wq)
388 return;
389 wq->normal->limit_active = limit_active;
390 if (wq->high)
391 wq->high->limit_active = limit_active;
392}
393
394void btrfs_set_work_high_priority(struct btrfs_work *work)
395{
396 set_bit(WORK_HIGH_PRIO_BIT, &work->flags);
397}
398
399void btrfs_flush_workqueue(struct btrfs_workqueue *wq)
400{
401 if (wq->high)
402 flush_workqueue(wq->high->normal_wq);
403
404 flush_workqueue(wq->normal->normal_wq);
405}
1/*
2 * Copyright (C) 2007 Oracle. All rights reserved.
3 * Copyright (C) 2014 Fujitsu. All rights reserved.
4 *
5 * This program is free software; you can redistribute it and/or
6 * modify it under the terms of the GNU General Public
7 * License v2 as published by the Free Software Foundation.
8 *
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
12 * General Public License for more details.
13 *
14 * You should have received a copy of the GNU General Public
15 * License along with this program; if not, write to the
16 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
17 * Boston, MA 021110-1307, USA.
18 */
19
20#include <linux/kthread.h>
21#include <linux/slab.h>
22#include <linux/list.h>
23#include <linux/spinlock.h>
24#include <linux/freezer.h>
25#include <linux/workqueue.h>
26#include "async-thread.h"
27#include "ctree.h"
28
29#define WORK_DONE_BIT 0
30#define WORK_ORDER_DONE_BIT 1
31#define WORK_HIGH_PRIO_BIT 2
32
33#define NO_THRESHOLD (-1)
34#define DFT_THRESHOLD (32)
35
36struct __btrfs_workqueue {
37 struct workqueue_struct *normal_wq;
38 /* List head pointing to ordered work list */
39 struct list_head ordered_list;
40
41 /* Spinlock for ordered_list */
42 spinlock_t list_lock;
43
44 /* Thresholding related variants */
45 atomic_t pending;
46 int max_active;
47 int current_max;
48 int thresh;
49 unsigned int count;
50 spinlock_t thres_lock;
51};
52
53struct btrfs_workqueue {
54 struct __btrfs_workqueue *normal;
55 struct __btrfs_workqueue *high;
56};
57
58static inline struct __btrfs_workqueue
59*__btrfs_alloc_workqueue(const char *name, int flags, int max_active,
60 int thresh)
61{
62 struct __btrfs_workqueue *ret = kzalloc(sizeof(*ret), GFP_NOFS);
63
64 if (unlikely(!ret))
65 return NULL;
66
67 ret->max_active = max_active;
68 atomic_set(&ret->pending, 0);
69 if (thresh == 0)
70 thresh = DFT_THRESHOLD;
71 /* For low threshold, disabling threshold is a better choice */
72 if (thresh < DFT_THRESHOLD) {
73 ret->current_max = max_active;
74 ret->thresh = NO_THRESHOLD;
75 } else {
76 ret->current_max = 1;
77 ret->thresh = thresh;
78 }
79
80 if (flags & WQ_HIGHPRI)
81 ret->normal_wq = alloc_workqueue("%s-%s-high", flags,
82 ret->max_active,
83 "btrfs", name);
84 else
85 ret->normal_wq = alloc_workqueue("%s-%s", flags,
86 ret->max_active, "btrfs",
87 name);
88 if (unlikely(!ret->normal_wq)) {
89 kfree(ret);
90 return NULL;
91 }
92
93 INIT_LIST_HEAD(&ret->ordered_list);
94 spin_lock_init(&ret->list_lock);
95 spin_lock_init(&ret->thres_lock);
96 trace_btrfs_workqueue_alloc(ret, name, flags & WQ_HIGHPRI);
97 return ret;
98}
99
100static inline void
101__btrfs_destroy_workqueue(struct __btrfs_workqueue *wq);
102
103struct btrfs_workqueue *btrfs_alloc_workqueue(const char *name,
104 int flags,
105 int max_active,
106 int thresh)
107{
108 struct btrfs_workqueue *ret = kzalloc(sizeof(*ret), GFP_NOFS);
109
110 if (unlikely(!ret))
111 return NULL;
112
113 ret->normal = __btrfs_alloc_workqueue(name, flags & ~WQ_HIGHPRI,
114 max_active, thresh);
115 if (unlikely(!ret->normal)) {
116 kfree(ret);
117 return NULL;
118 }
119
120 if (flags & WQ_HIGHPRI) {
121 ret->high = __btrfs_alloc_workqueue(name, flags, max_active,
122 thresh);
123 if (unlikely(!ret->high)) {
124 __btrfs_destroy_workqueue(ret->normal);
125 kfree(ret);
126 return NULL;
127 }
128 }
129 return ret;
130}
131
132/*
133 * Hook for threshold which will be called in btrfs_queue_work.
134 * This hook WILL be called in IRQ handler context,
135 * so workqueue_set_max_active MUST NOT be called in this hook
136 */
137static inline void thresh_queue_hook(struct __btrfs_workqueue *wq)
138{
139 if (wq->thresh == NO_THRESHOLD)
140 return;
141 atomic_inc(&wq->pending);
142}
143
144/*
145 * Hook for threshold which will be called before executing the work,
146 * This hook is called in kthread content.
147 * So workqueue_set_max_active is called here.
148 */
149static inline void thresh_exec_hook(struct __btrfs_workqueue *wq)
150{
151 int new_max_active;
152 long pending;
153 int need_change = 0;
154
155 if (wq->thresh == NO_THRESHOLD)
156 return;
157
158 atomic_dec(&wq->pending);
159 spin_lock(&wq->thres_lock);
160 /*
161 * Use wq->count to limit the calling frequency of
162 * workqueue_set_max_active.
163 */
164 wq->count++;
165 wq->count %= (wq->thresh / 4);
166 if (!wq->count)
167 goto out;
168 new_max_active = wq->current_max;
169
170 /*
171 * pending may be changed later, but it's OK since we really
172 * don't need it so accurate to calculate new_max_active.
173 */
174 pending = atomic_read(&wq->pending);
175 if (pending > wq->thresh)
176 new_max_active++;
177 if (pending < wq->thresh / 2)
178 new_max_active--;
179 new_max_active = clamp_val(new_max_active, 1, wq->max_active);
180 if (new_max_active != wq->current_max) {
181 need_change = 1;
182 wq->current_max = new_max_active;
183 }
184out:
185 spin_unlock(&wq->thres_lock);
186
187 if (need_change) {
188 workqueue_set_max_active(wq->normal_wq, wq->current_max);
189 }
190}
191
192static void run_ordered_work(struct __btrfs_workqueue *wq)
193{
194 struct list_head *list = &wq->ordered_list;
195 struct btrfs_work *work;
196 spinlock_t *lock = &wq->list_lock;
197 unsigned long flags;
198
199 while (1) {
200 spin_lock_irqsave(lock, flags);
201 if (list_empty(list))
202 break;
203 work = list_entry(list->next, struct btrfs_work,
204 ordered_list);
205 if (!test_bit(WORK_DONE_BIT, &work->flags))
206 break;
207
208 /*
209 * we are going to call the ordered done function, but
210 * we leave the work item on the list as a barrier so
211 * that later work items that are done don't have their
212 * functions called before this one returns
213 */
214 if (test_and_set_bit(WORK_ORDER_DONE_BIT, &work->flags))
215 break;
216 trace_btrfs_ordered_sched(work);
217 spin_unlock_irqrestore(lock, flags);
218 work->ordered_func(work);
219
220 /* now take the lock again and drop our item from the list */
221 spin_lock_irqsave(lock, flags);
222 list_del(&work->ordered_list);
223 spin_unlock_irqrestore(lock, flags);
224
225 /*
226 * we don't want to call the ordered free functions
227 * with the lock held though
228 */
229 work->ordered_free(work);
230 trace_btrfs_all_work_done(work);
231 }
232 spin_unlock_irqrestore(lock, flags);
233}
234
235static void normal_work_helper(struct work_struct *arg)
236{
237 struct btrfs_work *work;
238 struct __btrfs_workqueue *wq;
239 int need_order = 0;
240
241 work = container_of(arg, struct btrfs_work, normal_work);
242 /*
243 * We should not touch things inside work in the following cases:
244 * 1) after work->func() if it has no ordered_free
245 * Since the struct is freed in work->func().
246 * 2) after setting WORK_DONE_BIT
247 * The work may be freed in other threads almost instantly.
248 * So we save the needed things here.
249 */
250 if (work->ordered_func)
251 need_order = 1;
252 wq = work->wq;
253
254 trace_btrfs_work_sched(work);
255 thresh_exec_hook(wq);
256 work->func(work);
257 if (need_order) {
258 set_bit(WORK_DONE_BIT, &work->flags);
259 run_ordered_work(wq);
260 }
261 if (!need_order)
262 trace_btrfs_all_work_done(work);
263}
264
265void btrfs_init_work(struct btrfs_work *work,
266 btrfs_func_t func,
267 btrfs_func_t ordered_func,
268 btrfs_func_t ordered_free)
269{
270 work->func = func;
271 work->ordered_func = ordered_func;
272 work->ordered_free = ordered_free;
273 INIT_WORK(&work->normal_work, normal_work_helper);
274 INIT_LIST_HEAD(&work->ordered_list);
275 work->flags = 0;
276}
277
278static inline void __btrfs_queue_work(struct __btrfs_workqueue *wq,
279 struct btrfs_work *work)
280{
281 unsigned long flags;
282
283 work->wq = wq;
284 thresh_queue_hook(wq);
285 if (work->ordered_func) {
286 spin_lock_irqsave(&wq->list_lock, flags);
287 list_add_tail(&work->ordered_list, &wq->ordered_list);
288 spin_unlock_irqrestore(&wq->list_lock, flags);
289 }
290 queue_work(wq->normal_wq, &work->normal_work);
291 trace_btrfs_work_queued(work);
292}
293
294void btrfs_queue_work(struct btrfs_workqueue *wq,
295 struct btrfs_work *work)
296{
297 struct __btrfs_workqueue *dest_wq;
298
299 if (test_bit(WORK_HIGH_PRIO_BIT, &work->flags) && wq->high)
300 dest_wq = wq->high;
301 else
302 dest_wq = wq->normal;
303 __btrfs_queue_work(dest_wq, work);
304}
305
306static inline void
307__btrfs_destroy_workqueue(struct __btrfs_workqueue *wq)
308{
309 destroy_workqueue(wq->normal_wq);
310 trace_btrfs_workqueue_destroy(wq);
311 kfree(wq);
312}
313
314void btrfs_destroy_workqueue(struct btrfs_workqueue *wq)
315{
316 if (!wq)
317 return;
318 if (wq->high)
319 __btrfs_destroy_workqueue(wq->high);
320 __btrfs_destroy_workqueue(wq->normal);
321 kfree(wq);
322}
323
324void btrfs_workqueue_set_max(struct btrfs_workqueue *wq, int max)
325{
326 if (!wq)
327 return;
328 wq->normal->max_active = max;
329 if (wq->high)
330 wq->high->max_active = max;
331}
332
333void btrfs_set_work_high_priority(struct btrfs_work *work)
334{
335 set_bit(WORK_HIGH_PRIO_BIT, &work->flags);
336}