Commit | Line | Data |
---|---|---|
8b712842 CM |
1 | /* |
2 | * Copyright (C) 2007 Oracle. All rights reserved. | |
08a9ff32 | 3 | * Copyright (C) 2014 Fujitsu. All rights reserved. |
8b712842 CM |
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> | |
5a0e3ad6 | 21 | #include <linux/slab.h> |
8b712842 CM |
22 | #include <linux/list.h> |
23 | #include <linux/spinlock.h> | |
b51912c9 | 24 | #include <linux/freezer.h> |
08a9ff32 | 25 | #include <linux/workqueue.h> |
8b712842 | 26 | #include "async-thread.h" |
52483bc2 | 27 | #include "ctree.h" |
8b712842 | 28 | |
a046e9c8 QW |
29 | #define WORK_DONE_BIT 0 |
30 | #define WORK_ORDER_DONE_BIT 1 | |
31 | #define WORK_HIGH_PRIO_BIT 2 | |
4a69a410 | 32 | |
0bd9289c QW |
33 | #define NO_THRESHOLD (-1) |
34 | #define DFT_THRESHOLD (32) | |
35 | ||
d458b054 | 36 | struct __btrfs_workqueue { |
08a9ff32 QW |
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; | |
0bd9289c QW |
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; | |
08a9ff32 QW |
51 | }; |
52 | ||
d458b054 QW |
53 | struct btrfs_workqueue { |
54 | struct __btrfs_workqueue *normal; | |
55 | struct __btrfs_workqueue *high; | |
1ca08976 QW |
56 | }; |
57 | ||
d458b054 | 58 | static inline struct __btrfs_workqueue |
0bd9289c | 59 | *__btrfs_alloc_workqueue(char *name, int flags, int max_active, int thresh) |
1ca08976 | 60 | { |
d458b054 | 61 | struct __btrfs_workqueue *ret = kzalloc(sizeof(*ret), GFP_NOFS); |
1ca08976 QW |
62 | |
63 | if (unlikely(!ret)) | |
64 | return NULL; | |
65 | ||
0bd9289c QW |
66 | ret->max_active = max_active; |
67 | atomic_set(&ret->pending, 0); | |
68 | if (thresh == 0) | |
69 | thresh = DFT_THRESHOLD; | |
70 | /* For low threshold, disabling threshold is a better choice */ | |
71 | if (thresh < DFT_THRESHOLD) { | |
72 | ret->current_max = max_active; | |
73 | ret->thresh = NO_THRESHOLD; | |
74 | } else { | |
75 | ret->current_max = 1; | |
76 | ret->thresh = thresh; | |
77 | } | |
78 | ||
1ca08976 QW |
79 | if (flags & WQ_HIGHPRI) |
80 | ret->normal_wq = alloc_workqueue("%s-%s-high", flags, | |
0bd9289c QW |
81 | ret->max_active, |
82 | "btrfs", name); | |
1ca08976 QW |
83 | else |
84 | ret->normal_wq = alloc_workqueue("%s-%s", flags, | |
0bd9289c QW |
85 | ret->max_active, "btrfs", |
86 | name); | |
1ca08976 QW |
87 | if (unlikely(!ret->normal_wq)) { |
88 | kfree(ret); | |
89 | return NULL; | |
90 | } | |
91 | ||
92 | INIT_LIST_HEAD(&ret->ordered_list); | |
93 | spin_lock_init(&ret->list_lock); | |
0bd9289c | 94 | spin_lock_init(&ret->thres_lock); |
1ca08976 QW |
95 | return ret; |
96 | } | |
97 | ||
98 | static inline void | |
d458b054 | 99 | __btrfs_destroy_workqueue(struct __btrfs_workqueue *wq); |
1ca08976 | 100 | |
d458b054 QW |
101 | struct btrfs_workqueue *btrfs_alloc_workqueue(char *name, |
102 | int flags, | |
103 | int max_active, | |
104 | int thresh) | |
08a9ff32 | 105 | { |
d458b054 | 106 | struct btrfs_workqueue *ret = kzalloc(sizeof(*ret), GFP_NOFS); |
08a9ff32 QW |
107 | |
108 | if (unlikely(!ret)) | |
109 | return NULL; | |
110 | ||
1ca08976 | 111 | ret->normal = __btrfs_alloc_workqueue(name, flags & ~WQ_HIGHPRI, |
0bd9289c | 112 | max_active, thresh); |
1ca08976 | 113 | if (unlikely(!ret->normal)) { |
08a9ff32 QW |
114 | kfree(ret); |
115 | return NULL; | |
116 | } | |
117 | ||
1ca08976 | 118 | if (flags & WQ_HIGHPRI) { |
0bd9289c QW |
119 | ret->high = __btrfs_alloc_workqueue(name, flags, max_active, |
120 | thresh); | |
1ca08976 QW |
121 | if (unlikely(!ret->high)) { |
122 | __btrfs_destroy_workqueue(ret->normal); | |
123 | kfree(ret); | |
124 | return NULL; | |
125 | } | |
126 | } | |
08a9ff32 QW |
127 | return ret; |
128 | } | |
129 | ||
0bd9289c QW |
130 | /* |
131 | * Hook for threshold which will be called in btrfs_queue_work. | |
132 | * This hook WILL be called in IRQ handler context, | |
133 | * so workqueue_set_max_active MUST NOT be called in this hook | |
134 | */ | |
d458b054 | 135 | static inline void thresh_queue_hook(struct __btrfs_workqueue *wq) |
0bd9289c QW |
136 | { |
137 | if (wq->thresh == NO_THRESHOLD) | |
138 | return; | |
139 | atomic_inc(&wq->pending); | |
140 | } | |
141 | ||
142 | /* | |
143 | * Hook for threshold which will be called before executing the work, | |
144 | * This hook is called in kthread content. | |
145 | * So workqueue_set_max_active is called here. | |
146 | */ | |
d458b054 | 147 | static inline void thresh_exec_hook(struct __btrfs_workqueue *wq) |
0bd9289c QW |
148 | { |
149 | int new_max_active; | |
150 | long pending; | |
151 | int need_change = 0; | |
152 | ||
153 | if (wq->thresh == NO_THRESHOLD) | |
154 | return; | |
155 | ||
156 | atomic_dec(&wq->pending); | |
157 | spin_lock(&wq->thres_lock); | |
158 | /* | |
159 | * Use wq->count to limit the calling frequency of | |
160 | * workqueue_set_max_active. | |
161 | */ | |
162 | wq->count++; | |
163 | wq->count %= (wq->thresh / 4); | |
164 | if (!wq->count) | |
165 | goto out; | |
166 | new_max_active = wq->current_max; | |
167 | ||
168 | /* | |
169 | * pending may be changed later, but it's OK since we really | |
170 | * don't need it so accurate to calculate new_max_active. | |
171 | */ | |
172 | pending = atomic_read(&wq->pending); | |
173 | if (pending > wq->thresh) | |
174 | new_max_active++; | |
175 | if (pending < wq->thresh / 2) | |
176 | new_max_active--; | |
177 | new_max_active = clamp_val(new_max_active, 1, wq->max_active); | |
178 | if (new_max_active != wq->current_max) { | |
179 | need_change = 1; | |
180 | wq->current_max = new_max_active; | |
181 | } | |
182 | out: | |
183 | spin_unlock(&wq->thres_lock); | |
184 | ||
185 | if (need_change) { | |
186 | workqueue_set_max_active(wq->normal_wq, wq->current_max); | |
187 | } | |
188 | } | |
189 | ||
d458b054 | 190 | static void run_ordered_work(struct __btrfs_workqueue *wq) |
08a9ff32 QW |
191 | { |
192 | struct list_head *list = &wq->ordered_list; | |
d458b054 | 193 | struct btrfs_work *work; |
08a9ff32 QW |
194 | spinlock_t *lock = &wq->list_lock; |
195 | unsigned long flags; | |
196 | ||
197 | while (1) { | |
198 | spin_lock_irqsave(lock, flags); | |
199 | if (list_empty(list)) | |
200 | break; | |
d458b054 | 201 | work = list_entry(list->next, struct btrfs_work, |
08a9ff32 QW |
202 | ordered_list); |
203 | if (!test_bit(WORK_DONE_BIT, &work->flags)) | |
204 | break; | |
205 | ||
206 | /* | |
207 | * we are going to call the ordered done function, but | |
208 | * we leave the work item on the list as a barrier so | |
209 | * that later work items that are done don't have their | |
210 | * functions called before this one returns | |
211 | */ | |
212 | if (test_and_set_bit(WORK_ORDER_DONE_BIT, &work->flags)) | |
213 | break; | |
52483bc2 | 214 | trace_btrfs_ordered_sched(work); |
08a9ff32 QW |
215 | spin_unlock_irqrestore(lock, flags); |
216 | work->ordered_func(work); | |
217 | ||
218 | /* now take the lock again and drop our item from the list */ | |
219 | spin_lock_irqsave(lock, flags); | |
220 | list_del(&work->ordered_list); | |
221 | spin_unlock_irqrestore(lock, flags); | |
222 | ||
223 | /* | |
224 | * we don't want to call the ordered free functions | |
225 | * with the lock held though | |
226 | */ | |
227 | work->ordered_free(work); | |
52483bc2 | 228 | trace_btrfs_all_work_done(work); |
08a9ff32 QW |
229 | } |
230 | spin_unlock_irqrestore(lock, flags); | |
231 | } | |
232 | ||
233 | static void normal_work_helper(struct work_struct *arg) | |
234 | { | |
d458b054 QW |
235 | struct btrfs_work *work; |
236 | struct __btrfs_workqueue *wq; | |
08a9ff32 QW |
237 | int need_order = 0; |
238 | ||
d458b054 | 239 | work = container_of(arg, struct btrfs_work, normal_work); |
08a9ff32 QW |
240 | /* |
241 | * We should not touch things inside work in the following cases: | |
242 | * 1) after work->func() if it has no ordered_free | |
243 | * Since the struct is freed in work->func(). | |
244 | * 2) after setting WORK_DONE_BIT | |
245 | * The work may be freed in other threads almost instantly. | |
246 | * So we save the needed things here. | |
247 | */ | |
248 | if (work->ordered_func) | |
249 | need_order = 1; | |
250 | wq = work->wq; | |
251 | ||
52483bc2 | 252 | trace_btrfs_work_sched(work); |
0bd9289c | 253 | thresh_exec_hook(wq); |
08a9ff32 QW |
254 | work->func(work); |
255 | if (need_order) { | |
256 | set_bit(WORK_DONE_BIT, &work->flags); | |
257 | run_ordered_work(wq); | |
258 | } | |
52483bc2 QW |
259 | if (!need_order) |
260 | trace_btrfs_all_work_done(work); | |
08a9ff32 QW |
261 | } |
262 | ||
d458b054 | 263 | void btrfs_init_work(struct btrfs_work *work, |
6db8914f QW |
264 | btrfs_func_t func, |
265 | btrfs_func_t ordered_func, | |
266 | btrfs_func_t ordered_free) | |
08a9ff32 QW |
267 | { |
268 | work->func = func; | |
269 | work->ordered_func = ordered_func; | |
270 | work->ordered_free = ordered_free; | |
271 | INIT_WORK(&work->normal_work, normal_work_helper); | |
272 | INIT_LIST_HEAD(&work->ordered_list); | |
273 | work->flags = 0; | |
274 | } | |
275 | ||
d458b054 QW |
276 | static inline void __btrfs_queue_work(struct __btrfs_workqueue *wq, |
277 | struct btrfs_work *work) | |
08a9ff32 QW |
278 | { |
279 | unsigned long flags; | |
280 | ||
281 | work->wq = wq; | |
0bd9289c | 282 | thresh_queue_hook(wq); |
08a9ff32 QW |
283 | if (work->ordered_func) { |
284 | spin_lock_irqsave(&wq->list_lock, flags); | |
285 | list_add_tail(&work->ordered_list, &wq->ordered_list); | |
286 | spin_unlock_irqrestore(&wq->list_lock, flags); | |
287 | } | |
288 | queue_work(wq->normal_wq, &work->normal_work); | |
52483bc2 | 289 | trace_btrfs_work_queued(work); |
08a9ff32 QW |
290 | } |
291 | ||
d458b054 QW |
292 | void btrfs_queue_work(struct btrfs_workqueue *wq, |
293 | struct btrfs_work *work) | |
1ca08976 | 294 | { |
d458b054 | 295 | struct __btrfs_workqueue *dest_wq; |
1ca08976 QW |
296 | |
297 | if (test_bit(WORK_HIGH_PRIO_BIT, &work->flags) && wq->high) | |
298 | dest_wq = wq->high; | |
299 | else | |
300 | dest_wq = wq->normal; | |
301 | __btrfs_queue_work(dest_wq, work); | |
302 | } | |
303 | ||
304 | static inline void | |
d458b054 | 305 | __btrfs_destroy_workqueue(struct __btrfs_workqueue *wq) |
08a9ff32 QW |
306 | { |
307 | destroy_workqueue(wq->normal_wq); | |
308 | kfree(wq); | |
309 | } | |
310 | ||
d458b054 | 311 | void btrfs_destroy_workqueue(struct btrfs_workqueue *wq) |
1ca08976 QW |
312 | { |
313 | if (!wq) | |
314 | return; | |
315 | if (wq->high) | |
316 | __btrfs_destroy_workqueue(wq->high); | |
317 | __btrfs_destroy_workqueue(wq->normal); | |
318 | } | |
319 | ||
d458b054 | 320 | void btrfs_workqueue_set_max(struct btrfs_workqueue *wq, int max) |
08a9ff32 | 321 | { |
0bd9289c | 322 | wq->normal->max_active = max; |
1ca08976 | 323 | if (wq->high) |
0bd9289c | 324 | wq->high->max_active = max; |
1ca08976 QW |
325 | } |
326 | ||
d458b054 | 327 | void btrfs_set_work_high_priority(struct btrfs_work *work) |
1ca08976 QW |
328 | { |
329 | set_bit(WORK_HIGH_PRIO_BIT, &work->flags); | |
08a9ff32 | 330 | } |