Loading...
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 "async-thread.h"
26#include "ctree.h"
27
28#define WORK_DONE_BIT 0
29#define WORK_ORDER_DONE_BIT 1
30#define WORK_HIGH_PRIO_BIT 2
31
32#define NO_THRESHOLD (-1)
33#define DFT_THRESHOLD (32)
34
35struct __btrfs_workqueue {
36 struct workqueue_struct *normal_wq;
37 /* List head pointing to ordered work list */
38 struct list_head ordered_list;
39
40 /* Spinlock for ordered_list */
41 spinlock_t list_lock;
42
43 /* Thresholding related variants */
44 atomic_t pending;
45
46 /* Up limit of concurrency workers */
47 int limit_active;
48
49 /* Current number of concurrency workers */
50 int current_active;
51
52 /* Threshold to change current_active */
53 int thresh;
54 unsigned int count;
55 spinlock_t thres_lock;
56};
57
58struct btrfs_workqueue {
59 struct __btrfs_workqueue *normal;
60 struct __btrfs_workqueue *high;
61};
62
63static void normal_work_helper(struct btrfs_work *work);
64
65#define BTRFS_WORK_HELPER(name) \
66void btrfs_##name(struct work_struct *arg) \
67{ \
68 struct btrfs_work *work = container_of(arg, struct btrfs_work, \
69 normal_work); \
70 normal_work_helper(work); \
71}
72
73BTRFS_WORK_HELPER(worker_helper);
74BTRFS_WORK_HELPER(delalloc_helper);
75BTRFS_WORK_HELPER(flush_delalloc_helper);
76BTRFS_WORK_HELPER(cache_helper);
77BTRFS_WORK_HELPER(submit_helper);
78BTRFS_WORK_HELPER(fixup_helper);
79BTRFS_WORK_HELPER(endio_helper);
80BTRFS_WORK_HELPER(endio_meta_helper);
81BTRFS_WORK_HELPER(endio_meta_write_helper);
82BTRFS_WORK_HELPER(endio_raid56_helper);
83BTRFS_WORK_HELPER(endio_repair_helper);
84BTRFS_WORK_HELPER(rmw_helper);
85BTRFS_WORK_HELPER(endio_write_helper);
86BTRFS_WORK_HELPER(freespace_write_helper);
87BTRFS_WORK_HELPER(delayed_meta_helper);
88BTRFS_WORK_HELPER(readahead_helper);
89BTRFS_WORK_HELPER(qgroup_rescan_helper);
90BTRFS_WORK_HELPER(extent_refs_helper);
91BTRFS_WORK_HELPER(scrub_helper);
92BTRFS_WORK_HELPER(scrubwrc_helper);
93BTRFS_WORK_HELPER(scrubnc_helper);
94BTRFS_WORK_HELPER(scrubparity_helper);
95
96static struct __btrfs_workqueue *
97__btrfs_alloc_workqueue(const char *name, unsigned int flags, int limit_active,
98 int thresh)
99{
100 struct __btrfs_workqueue *ret = kzalloc(sizeof(*ret), GFP_KERNEL);
101
102 if (!ret)
103 return NULL;
104
105 ret->limit_active = limit_active;
106 atomic_set(&ret->pending, 0);
107 if (thresh == 0)
108 thresh = DFT_THRESHOLD;
109 /* For low threshold, disabling threshold is a better choice */
110 if (thresh < DFT_THRESHOLD) {
111 ret->current_active = limit_active;
112 ret->thresh = NO_THRESHOLD;
113 } else {
114 /*
115 * For threshold-able wq, let its concurrency grow on demand.
116 * Use minimal max_active at alloc time to reduce resource
117 * usage.
118 */
119 ret->current_active = 1;
120 ret->thresh = thresh;
121 }
122
123 if (flags & WQ_HIGHPRI)
124 ret->normal_wq = alloc_workqueue("%s-%s-high", flags,
125 ret->current_active, "btrfs",
126 name);
127 else
128 ret->normal_wq = alloc_workqueue("%s-%s", flags,
129 ret->current_active, "btrfs",
130 name);
131 if (!ret->normal_wq) {
132 kfree(ret);
133 return NULL;
134 }
135
136 INIT_LIST_HEAD(&ret->ordered_list);
137 spin_lock_init(&ret->list_lock);
138 spin_lock_init(&ret->thres_lock);
139 trace_btrfs_workqueue_alloc(ret, name, flags & WQ_HIGHPRI);
140 return ret;
141}
142
143static inline void
144__btrfs_destroy_workqueue(struct __btrfs_workqueue *wq);
145
146struct btrfs_workqueue *btrfs_alloc_workqueue(const char *name,
147 unsigned int flags,
148 int limit_active,
149 int thresh)
150{
151 struct btrfs_workqueue *ret = kzalloc(sizeof(*ret), GFP_KERNEL);
152
153 if (!ret)
154 return NULL;
155
156 ret->normal = __btrfs_alloc_workqueue(name, flags & ~WQ_HIGHPRI,
157 limit_active, thresh);
158 if (!ret->normal) {
159 kfree(ret);
160 return NULL;
161 }
162
163 if (flags & WQ_HIGHPRI) {
164 ret->high = __btrfs_alloc_workqueue(name, flags, limit_active,
165 thresh);
166 if (!ret->high) {
167 __btrfs_destroy_workqueue(ret->normal);
168 kfree(ret);
169 return NULL;
170 }
171 }
172 return ret;
173}
174
175/*
176 * Hook for threshold which will be called in btrfs_queue_work.
177 * This hook WILL be called in IRQ handler context,
178 * so workqueue_set_max_active MUST NOT be called in this hook
179 */
180static inline void thresh_queue_hook(struct __btrfs_workqueue *wq)
181{
182 if (wq->thresh == NO_THRESHOLD)
183 return;
184 atomic_inc(&wq->pending);
185}
186
187/*
188 * Hook for threshold which will be called before executing the work,
189 * This hook is called in kthread content.
190 * So workqueue_set_max_active is called here.
191 */
192static inline void thresh_exec_hook(struct __btrfs_workqueue *wq)
193{
194 int new_current_active;
195 long pending;
196 int need_change = 0;
197
198 if (wq->thresh == NO_THRESHOLD)
199 return;
200
201 atomic_dec(&wq->pending);
202 spin_lock(&wq->thres_lock);
203 /*
204 * Use wq->count to limit the calling frequency of
205 * workqueue_set_max_active.
206 */
207 wq->count++;
208 wq->count %= (wq->thresh / 4);
209 if (!wq->count)
210 goto out;
211 new_current_active = wq->current_active;
212
213 /*
214 * pending may be changed later, but it's OK since we really
215 * don't need it so accurate to calculate new_max_active.
216 */
217 pending = atomic_read(&wq->pending);
218 if (pending > wq->thresh)
219 new_current_active++;
220 if (pending < wq->thresh / 2)
221 new_current_active--;
222 new_current_active = clamp_val(new_current_active, 1, wq->limit_active);
223 if (new_current_active != wq->current_active) {
224 need_change = 1;
225 wq->current_active = new_current_active;
226 }
227out:
228 spin_unlock(&wq->thres_lock);
229
230 if (need_change) {
231 workqueue_set_max_active(wq->normal_wq, wq->current_active);
232 }
233}
234
235static void run_ordered_work(struct __btrfs_workqueue *wq)
236{
237 struct list_head *list = &wq->ordered_list;
238 struct btrfs_work *work;
239 spinlock_t *lock = &wq->list_lock;
240 unsigned long flags;
241
242 while (1) {
243 spin_lock_irqsave(lock, flags);
244 if (list_empty(list))
245 break;
246 work = list_entry(list->next, struct btrfs_work,
247 ordered_list);
248 if (!test_bit(WORK_DONE_BIT, &work->flags))
249 break;
250
251 /*
252 * we are going to call the ordered done function, but
253 * we leave the work item on the list as a barrier so
254 * that later work items that are done don't have their
255 * functions called before this one returns
256 */
257 if (test_and_set_bit(WORK_ORDER_DONE_BIT, &work->flags))
258 break;
259 trace_btrfs_ordered_sched(work);
260 spin_unlock_irqrestore(lock, flags);
261 work->ordered_func(work);
262
263 /* now take the lock again and drop our item from the list */
264 spin_lock_irqsave(lock, flags);
265 list_del(&work->ordered_list);
266 spin_unlock_irqrestore(lock, flags);
267
268 /*
269 * we don't want to call the ordered free functions
270 * with the lock held though
271 */
272 work->ordered_free(work);
273 trace_btrfs_all_work_done(work);
274 }
275 spin_unlock_irqrestore(lock, flags);
276}
277
278static void normal_work_helper(struct btrfs_work *work)
279{
280 struct __btrfs_workqueue *wq;
281 int need_order = 0;
282
283 /*
284 * We should not touch things inside work in the following cases:
285 * 1) after work->func() if it has no ordered_free
286 * Since the struct is freed in work->func().
287 * 2) after setting WORK_DONE_BIT
288 * The work may be freed in other threads almost instantly.
289 * So we save the needed things here.
290 */
291 if (work->ordered_func)
292 need_order = 1;
293 wq = work->wq;
294
295 trace_btrfs_work_sched(work);
296 thresh_exec_hook(wq);
297 work->func(work);
298 if (need_order) {
299 set_bit(WORK_DONE_BIT, &work->flags);
300 run_ordered_work(wq);
301 }
302 if (!need_order)
303 trace_btrfs_all_work_done(work);
304}
305
306void btrfs_init_work(struct btrfs_work *work, btrfs_work_func_t uniq_func,
307 btrfs_func_t func,
308 btrfs_func_t ordered_func,
309 btrfs_func_t ordered_free)
310{
311 work->func = func;
312 work->ordered_func = ordered_func;
313 work->ordered_free = ordered_free;
314 INIT_WORK(&work->normal_work, uniq_func);
315 INIT_LIST_HEAD(&work->ordered_list);
316 work->flags = 0;
317}
318
319static inline void __btrfs_queue_work(struct __btrfs_workqueue *wq,
320 struct btrfs_work *work)
321{
322 unsigned long flags;
323
324 work->wq = wq;
325 thresh_queue_hook(wq);
326 if (work->ordered_func) {
327 spin_lock_irqsave(&wq->list_lock, flags);
328 list_add_tail(&work->ordered_list, &wq->ordered_list);
329 spin_unlock_irqrestore(&wq->list_lock, flags);
330 }
331 trace_btrfs_work_queued(work);
332 queue_work(wq->normal_wq, &work->normal_work);
333}
334
335void btrfs_queue_work(struct btrfs_workqueue *wq,
336 struct btrfs_work *work)
337{
338 struct __btrfs_workqueue *dest_wq;
339
340 if (test_bit(WORK_HIGH_PRIO_BIT, &work->flags) && wq->high)
341 dest_wq = wq->high;
342 else
343 dest_wq = wq->normal;
344 __btrfs_queue_work(dest_wq, work);
345}
346
347static inline void
348__btrfs_destroy_workqueue(struct __btrfs_workqueue *wq)
349{
350 destroy_workqueue(wq->normal_wq);
351 trace_btrfs_workqueue_destroy(wq);
352 kfree(wq);
353}
354
355void btrfs_destroy_workqueue(struct btrfs_workqueue *wq)
356{
357 if (!wq)
358 return;
359 if (wq->high)
360 __btrfs_destroy_workqueue(wq->high);
361 __btrfs_destroy_workqueue(wq->normal);
362 kfree(wq);
363}
364
365void btrfs_workqueue_set_max(struct btrfs_workqueue *wq, int limit_active)
366{
367 if (!wq)
368 return;
369 wq->normal->limit_active = limit_active;
370 if (wq->high)
371 wq->high->limit_active = limit_active;
372}
373
374void btrfs_set_work_high_priority(struct btrfs_work *work)
375{
376 set_bit(WORK_HIGH_PRIO_BIT, &work->flags);
377}
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}