ceph: handle interrupted ceph_writepage()
[linux-2.6-block.git] / fs / btrfs / async-thread.c
CommitLineData
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>
8b712842 25#include "async-thread.h"
52483bc2 26#include "ctree.h"
8b712842 27
a046e9c8
QW
28#define WORK_DONE_BIT 0
29#define WORK_ORDER_DONE_BIT 1
30#define WORK_HIGH_PRIO_BIT 2
4a69a410 31
0bd9289c
QW
32#define NO_THRESHOLD (-1)
33#define DFT_THRESHOLD (32)
34
d458b054 35struct __btrfs_workqueue {
08a9ff32
QW
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;
0bd9289c
QW
42
43 /* Thresholding related variants */
44 atomic_t pending;
c6dd6ea5
QW
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 */
0bd9289c
QW
53 int thresh;
54 unsigned int count;
55 spinlock_t thres_lock;
08a9ff32
QW
56};
57
d458b054
QW
58struct btrfs_workqueue {
59 struct __btrfs_workqueue *normal;
60 struct __btrfs_workqueue *high;
1ca08976
QW
61};
62
9e0af237
LB
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);
8b110e39 83BTRFS_WORK_HELPER(endio_repair_helper);
9e0af237
LB
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);
20b2e302 94BTRFS_WORK_HELPER(scrubparity_helper);
9e0af237
LB
95
96static struct __btrfs_workqueue *
c6dd6ea5 97__btrfs_alloc_workqueue(const char *name, unsigned int flags, int limit_active,
c3a46891 98 int thresh)
1ca08976 99{
61dd5ae6 100 struct __btrfs_workqueue *ret = kzalloc(sizeof(*ret), GFP_KERNEL);
1ca08976 101
5d99a998 102 if (!ret)
1ca08976
QW
103 return NULL;
104
c6dd6ea5 105 ret->limit_active = limit_active;
0bd9289c
QW
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) {
c6dd6ea5 111 ret->current_active = limit_active;
0bd9289c
QW
112 ret->thresh = NO_THRESHOLD;
113 } else {
c6dd6ea5
QW
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;
0bd9289c
QW
120 ret->thresh = thresh;
121 }
122
1ca08976
QW
123 if (flags & WQ_HIGHPRI)
124 ret->normal_wq = alloc_workqueue("%s-%s-high", flags,
c6dd6ea5
QW
125 ret->current_active, "btrfs",
126 name);
1ca08976
QW
127 else
128 ret->normal_wq = alloc_workqueue("%s-%s", flags,
c6dd6ea5 129 ret->current_active, "btrfs",
0bd9289c 130 name);
5d99a998 131 if (!ret->normal_wq) {
1ca08976
QW
132 kfree(ret);
133 return NULL;
134 }
135
136 INIT_LIST_HEAD(&ret->ordered_list);
137 spin_lock_init(&ret->list_lock);
0bd9289c 138 spin_lock_init(&ret->thres_lock);
c3a46891 139 trace_btrfs_workqueue_alloc(ret, name, flags & WQ_HIGHPRI);
1ca08976
QW
140 return ret;
141}
142
143static inline void
d458b054 144__btrfs_destroy_workqueue(struct __btrfs_workqueue *wq);
1ca08976 145
c3a46891 146struct btrfs_workqueue *btrfs_alloc_workqueue(const char *name,
6f011058 147 unsigned int flags,
c6dd6ea5 148 int limit_active,
d458b054 149 int thresh)
08a9ff32 150{
61dd5ae6 151 struct btrfs_workqueue *ret = kzalloc(sizeof(*ret), GFP_KERNEL);
08a9ff32 152
5d99a998 153 if (!ret)
08a9ff32
QW
154 return NULL;
155
1ca08976 156 ret->normal = __btrfs_alloc_workqueue(name, flags & ~WQ_HIGHPRI,
c6dd6ea5 157 limit_active, thresh);
5d99a998 158 if (!ret->normal) {
08a9ff32
QW
159 kfree(ret);
160 return NULL;
161 }
162
1ca08976 163 if (flags & WQ_HIGHPRI) {
c6dd6ea5 164 ret->high = __btrfs_alloc_workqueue(name, flags, limit_active,
0bd9289c 165 thresh);
5d99a998 166 if (!ret->high) {
1ca08976
QW
167 __btrfs_destroy_workqueue(ret->normal);
168 kfree(ret);
169 return NULL;
170 }
171 }
08a9ff32
QW
172 return ret;
173}
174
0bd9289c
QW
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 */
d458b054 180static inline void thresh_queue_hook(struct __btrfs_workqueue *wq)
0bd9289c
QW
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 */
d458b054 192static inline void thresh_exec_hook(struct __btrfs_workqueue *wq)
0bd9289c 193{
c6dd6ea5 194 int new_current_active;
0bd9289c
QW
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;
c6dd6ea5 211 new_current_active = wq->current_active;
0bd9289c
QW
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)
c6dd6ea5 219 new_current_active++;
0bd9289c 220 if (pending < wq->thresh / 2)
c6dd6ea5
QW
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) {
0bd9289c 224 need_change = 1;
c6dd6ea5 225 wq->current_active = new_current_active;
0bd9289c
QW
226 }
227out:
228 spin_unlock(&wq->thres_lock);
229
230 if (need_change) {
c6dd6ea5 231 workqueue_set_max_active(wq->normal_wq, wq->current_active);
0bd9289c
QW
232 }
233}
234
d458b054 235static void run_ordered_work(struct __btrfs_workqueue *wq)
08a9ff32
QW
236{
237 struct list_head *list = &wq->ordered_list;
d458b054 238 struct btrfs_work *work;
08a9ff32
QW
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;
d458b054 246 work = list_entry(list->next, struct btrfs_work,
08a9ff32
QW
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;
52483bc2 259 trace_btrfs_ordered_sched(work);
08a9ff32
QW
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);
52483bc2 273 trace_btrfs_all_work_done(work);
08a9ff32
QW
274 }
275 spin_unlock_irqrestore(lock, flags);
276}
277
9e0af237 278static void normal_work_helper(struct btrfs_work *work)
08a9ff32 279{
d458b054 280 struct __btrfs_workqueue *wq;
08a9ff32
QW
281 int need_order = 0;
282
08a9ff32
QW
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
52483bc2 295 trace_btrfs_work_sched(work);
0bd9289c 296 thresh_exec_hook(wq);
08a9ff32
QW
297 work->func(work);
298 if (need_order) {
299 set_bit(WORK_DONE_BIT, &work->flags);
300 run_ordered_work(wq);
301 }
52483bc2
QW
302 if (!need_order)
303 trace_btrfs_all_work_done(work);
08a9ff32
QW
304}
305
9e0af237 306void btrfs_init_work(struct btrfs_work *work, btrfs_work_func_t uniq_func,
6db8914f
QW
307 btrfs_func_t func,
308 btrfs_func_t ordered_func,
309 btrfs_func_t ordered_free)
08a9ff32
QW
310{
311 work->func = func;
312 work->ordered_func = ordered_func;
313 work->ordered_free = ordered_free;
9e0af237 314 INIT_WORK(&work->normal_work, uniq_func);
08a9ff32
QW
315 INIT_LIST_HEAD(&work->ordered_list);
316 work->flags = 0;
317}
318
d458b054
QW
319static inline void __btrfs_queue_work(struct __btrfs_workqueue *wq,
320 struct btrfs_work *work)
08a9ff32
QW
321{
322 unsigned long flags;
323
324 work->wq = wq;
0bd9289c 325 thresh_queue_hook(wq);
08a9ff32
QW
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 }
52483bc2 331 trace_btrfs_work_queued(work);
0a95b851 332 queue_work(wq->normal_wq, &work->normal_work);
08a9ff32
QW
333}
334
d458b054
QW
335void btrfs_queue_work(struct btrfs_workqueue *wq,
336 struct btrfs_work *work)
1ca08976 337{
d458b054 338 struct __btrfs_workqueue *dest_wq;
1ca08976
QW
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
d458b054 348__btrfs_destroy_workqueue(struct __btrfs_workqueue *wq)
08a9ff32
QW
349{
350 destroy_workqueue(wq->normal_wq);
c3a46891 351 trace_btrfs_workqueue_destroy(wq);
08a9ff32
QW
352 kfree(wq);
353}
354
d458b054 355void btrfs_destroy_workqueue(struct btrfs_workqueue *wq)
1ca08976
QW
356{
357 if (!wq)
358 return;
359 if (wq->high)
360 __btrfs_destroy_workqueue(wq->high);
361 __btrfs_destroy_workqueue(wq->normal);
ef66af10 362 kfree(wq);
1ca08976
QW
363}
364
c6dd6ea5 365void btrfs_workqueue_set_max(struct btrfs_workqueue *wq, int limit_active)
08a9ff32 366{
800ee224
ST
367 if (!wq)
368 return;
c6dd6ea5 369 wq->normal->limit_active = limit_active;
1ca08976 370 if (wq->high)
c6dd6ea5 371 wq->high->limit_active = limit_active;
1ca08976
QW
372}
373
d458b054 374void btrfs_set_work_high_priority(struct btrfs_work *work)
1ca08976
QW
375{
376 set_bit(WORK_HIGH_PRIO_BIT, &work->flags);
08a9ff32 377}