blob: ff0b0be92d6122402f02c039eed8e4599fdce4df [file] [log] [blame]
Chris Mason8b712842008-06-11 16:50:36 -04001/*
2 * Copyright (C) 2007 Oracle. All rights reserved.
Qu Wenruo08a9ff32014-02-28 10:46:03 +08003 * Copyright (C) 2014 Fujitsu. All rights reserved.
Chris Mason8b712842008-06-11 16:50:36 -04004 *
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>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090021#include <linux/slab.h>
Chris Mason8b712842008-06-11 16:50:36 -040022#include <linux/list.h>
23#include <linux/spinlock.h>
Chris Masonb51912c2009-02-04 09:23:24 -050024#include <linux/freezer.h>
Chris Mason8b712842008-06-11 16:50:36 -040025#include "async-thread.h"
Qu Wenruo52483bc2014-03-06 04:19:50 +000026#include "ctree.h"
Chris Mason8b712842008-06-11 16:50:36 -040027
Qu Wenruoa046e9c2014-02-28 10:46:18 +080028#define WORK_DONE_BIT 0
29#define WORK_ORDER_DONE_BIT 1
30#define WORK_HIGH_PRIO_BIT 2
Chris Mason4a69a412008-11-06 22:03:00 -050031
Qu Wenruo0bd92892014-02-28 10:46:05 +080032#define NO_THRESHOLD (-1)
33#define DFT_THRESHOLD (32)
34
Qu Wenruod458b052014-02-28 10:46:19 +080035struct __btrfs_workqueue {
Qu Wenruo08a9ff32014-02-28 10:46:03 +080036 struct workqueue_struct *normal_wq;
Jeff Mahoneycb001092016-06-09 16:22:11 -040037
38 /* File system this workqueue services */
39 struct btrfs_fs_info *fs_info;
40
Qu Wenruo08a9ff32014-02-28 10:46:03 +080041 /* List head pointing to ordered work list */
42 struct list_head ordered_list;
43
44 /* Spinlock for ordered_list */
45 spinlock_t list_lock;
Qu Wenruo0bd92892014-02-28 10:46:05 +080046
47 /* Thresholding related variants */
48 atomic_t pending;
Qu Wenruoc6dd6ea2015-08-20 09:30:39 +080049
50 /* Up limit of concurrency workers */
51 int limit_active;
52
53 /* Current number of concurrency workers */
54 int current_active;
55
56 /* Threshold to change current_active */
Qu Wenruo0bd92892014-02-28 10:46:05 +080057 int thresh;
58 unsigned int count;
59 spinlock_t thres_lock;
Qu Wenruo08a9ff32014-02-28 10:46:03 +080060};
61
Qu Wenruod458b052014-02-28 10:46:19 +080062struct btrfs_workqueue {
63 struct __btrfs_workqueue *normal;
64 struct __btrfs_workqueue *high;
Qu Wenruo1ca08972014-02-28 10:46:04 +080065};
66
Liu Bo9e0af232014-08-15 23:36:53 +080067static void normal_work_helper(struct btrfs_work *work);
68
69#define BTRFS_WORK_HELPER(name) \
70void btrfs_##name(struct work_struct *arg) \
71{ \
72 struct btrfs_work *work = container_of(arg, struct btrfs_work, \
73 normal_work); \
74 normal_work_helper(work); \
75}
76
Jeff Mahoneycb001092016-06-09 16:22:11 -040077struct btrfs_fs_info *
78btrfs_workqueue_owner(struct __btrfs_workqueue *wq)
79{
80 return wq->fs_info;
81}
82
83struct btrfs_fs_info *
84btrfs_work_owner(struct btrfs_work *work)
85{
86 return work->wq->fs_info;
87}
88
Maxim Patlasov3bac3222016-12-12 14:32:44 -080089bool btrfs_workqueue_normal_congested(struct btrfs_workqueue *wq)
90{
91 /*
92 * We could compare wq->normal->pending with num_online_cpus()
93 * to support "thresh == NO_THRESHOLD" case, but it requires
94 * moving up atomic_inc/dec in thresh_queue/exec_hook. Let's
95 * postpone it until someone needs the support of that case.
96 */
97 if (wq->normal->thresh == NO_THRESHOLD)
98 return false;
99
100 return atomic_read(&wq->normal->pending) > wq->normal->thresh * 2;
101}
102
Liu Bo9e0af232014-08-15 23:36:53 +0800103BTRFS_WORK_HELPER(worker_helper);
104BTRFS_WORK_HELPER(delalloc_helper);
105BTRFS_WORK_HELPER(flush_delalloc_helper);
106BTRFS_WORK_HELPER(cache_helper);
107BTRFS_WORK_HELPER(submit_helper);
108BTRFS_WORK_HELPER(fixup_helper);
109BTRFS_WORK_HELPER(endio_helper);
110BTRFS_WORK_HELPER(endio_meta_helper);
111BTRFS_WORK_HELPER(endio_meta_write_helper);
112BTRFS_WORK_HELPER(endio_raid56_helper);
Miao Xie8b110e32014-09-12 18:44:03 +0800113BTRFS_WORK_HELPER(endio_repair_helper);
Liu Bo9e0af232014-08-15 23:36:53 +0800114BTRFS_WORK_HELPER(rmw_helper);
115BTRFS_WORK_HELPER(endio_write_helper);
116BTRFS_WORK_HELPER(freespace_write_helper);
117BTRFS_WORK_HELPER(delayed_meta_helper);
118BTRFS_WORK_HELPER(readahead_helper);
119BTRFS_WORK_HELPER(qgroup_rescan_helper);
120BTRFS_WORK_HELPER(extent_refs_helper);
121BTRFS_WORK_HELPER(scrub_helper);
122BTRFS_WORK_HELPER(scrubwrc_helper);
123BTRFS_WORK_HELPER(scrubnc_helper);
Zhao Lei20b2e302015-06-04 20:09:15 +0800124BTRFS_WORK_HELPER(scrubparity_helper);
Liu Bo9e0af232014-08-15 23:36:53 +0800125
126static struct __btrfs_workqueue *
Jeff Mahoneycb001092016-06-09 16:22:11 -0400127__btrfs_alloc_workqueue(struct btrfs_fs_info *fs_info, const char *name,
128 unsigned int flags, int limit_active, int thresh)
Qu Wenruo08a9ff32014-02-28 10:46:03 +0800129{
David Sterba61dd5ae2015-12-01 18:04:30 +0100130 struct __btrfs_workqueue *ret = kzalloc(sizeof(*ret), GFP_KERNEL);
Qu Wenruo08a9ff32014-02-28 10:46:03 +0800131
David Sterba5d99a9982014-09-29 19:20:37 +0200132 if (!ret)
Qu Wenruo08a9ff32014-02-28 10:46:03 +0800133 return NULL;
134
Jeff Mahoneycb001092016-06-09 16:22:11 -0400135 ret->fs_info = fs_info;
Qu Wenruoc6dd6ea2015-08-20 09:30:39 +0800136 ret->limit_active = limit_active;
Qu Wenruo0bd92892014-02-28 10:46:05 +0800137 atomic_set(&ret->pending, 0);
138 if (thresh == 0)
139 thresh = DFT_THRESHOLD;
140 /* For low threshold, disabling threshold is a better choice */
141 if (thresh < DFT_THRESHOLD) {
Qu Wenruoc6dd6ea2015-08-20 09:30:39 +0800142 ret->current_active = limit_active;
Qu Wenruo0bd92892014-02-28 10:46:05 +0800143 ret->thresh = NO_THRESHOLD;
144 } else {
Qu Wenruoc6dd6ea2015-08-20 09:30:39 +0800145 /*
146 * For threshold-able wq, let its concurrency grow on demand.
147 * Use minimal max_active at alloc time to reduce resource
148 * usage.
149 */
150 ret->current_active = 1;
Qu Wenruo0bd92892014-02-28 10:46:05 +0800151 ret->thresh = thresh;
152 }
153
Qu Wenruo1ca08972014-02-28 10:46:04 +0800154 if (flags & WQ_HIGHPRI)
155 ret->normal_wq = alloc_workqueue("%s-%s-high", flags,
Qu Wenruoc6dd6ea2015-08-20 09:30:39 +0800156 ret->current_active, "btrfs",
157 name);
Qu Wenruo1ca08972014-02-28 10:46:04 +0800158 else
159 ret->normal_wq = alloc_workqueue("%s-%s", flags,
Qu Wenruoc6dd6ea2015-08-20 09:30:39 +0800160 ret->current_active, "btrfs",
Qu Wenruo0bd92892014-02-28 10:46:05 +0800161 name);
David Sterba5d99a9982014-09-29 19:20:37 +0200162 if (!ret->normal_wq) {
Qu Wenruo08a9ff32014-02-28 10:46:03 +0800163 kfree(ret);
164 return NULL;
165 }
166
167 INIT_LIST_HEAD(&ret->ordered_list);
168 spin_lock_init(&ret->list_lock);
Qu Wenruo0bd92892014-02-28 10:46:05 +0800169 spin_lock_init(&ret->thres_lock);
Qu Wenruoc3a46892014-03-12 08:05:33 +0000170 trace_btrfs_workqueue_alloc(ret, name, flags & WQ_HIGHPRI);
Qu Wenruo08a9ff32014-02-28 10:46:03 +0800171 return ret;
172}
173
Qu Wenruo1ca08972014-02-28 10:46:04 +0800174static inline void
Qu Wenruod458b052014-02-28 10:46:19 +0800175__btrfs_destroy_workqueue(struct __btrfs_workqueue *wq);
Qu Wenruo1ca08972014-02-28 10:46:04 +0800176
Jeff Mahoneycb001092016-06-09 16:22:11 -0400177struct btrfs_workqueue *btrfs_alloc_workqueue(struct btrfs_fs_info *fs_info,
178 const char *name,
David Sterba6f011052015-02-16 18:34:01 +0100179 unsigned int flags,
Qu Wenruoc6dd6ea2015-08-20 09:30:39 +0800180 int limit_active,
Qu Wenruod458b052014-02-28 10:46:19 +0800181 int thresh)
Qu Wenruo1ca08972014-02-28 10:46:04 +0800182{
David Sterba61dd5ae2015-12-01 18:04:30 +0100183 struct btrfs_workqueue *ret = kzalloc(sizeof(*ret), GFP_KERNEL);
Qu Wenruo1ca08972014-02-28 10:46:04 +0800184
David Sterba5d99a9982014-09-29 19:20:37 +0200185 if (!ret)
Qu Wenruo1ca08972014-02-28 10:46:04 +0800186 return NULL;
187
Jeff Mahoneycb001092016-06-09 16:22:11 -0400188 ret->normal = __btrfs_alloc_workqueue(fs_info, name,
189 flags & ~WQ_HIGHPRI,
Qu Wenruoc6dd6ea2015-08-20 09:30:39 +0800190 limit_active, thresh);
David Sterba5d99a9982014-09-29 19:20:37 +0200191 if (!ret->normal) {
Qu Wenruo1ca08972014-02-28 10:46:04 +0800192 kfree(ret);
193 return NULL;
194 }
195
196 if (flags & WQ_HIGHPRI) {
Jeff Mahoneycb001092016-06-09 16:22:11 -0400197 ret->high = __btrfs_alloc_workqueue(fs_info, name, flags,
198 limit_active, thresh);
David Sterba5d99a9982014-09-29 19:20:37 +0200199 if (!ret->high) {
Qu Wenruo1ca08972014-02-28 10:46:04 +0800200 __btrfs_destroy_workqueue(ret->normal);
201 kfree(ret);
202 return NULL;
203 }
204 }
205 return ret;
206}
207
Qu Wenruo0bd92892014-02-28 10:46:05 +0800208/*
209 * Hook for threshold which will be called in btrfs_queue_work.
210 * This hook WILL be called in IRQ handler context,
211 * so workqueue_set_max_active MUST NOT be called in this hook
212 */
Qu Wenruod458b052014-02-28 10:46:19 +0800213static inline void thresh_queue_hook(struct __btrfs_workqueue *wq)
Qu Wenruo0bd92892014-02-28 10:46:05 +0800214{
215 if (wq->thresh == NO_THRESHOLD)
216 return;
217 atomic_inc(&wq->pending);
218}
219
220/*
221 * Hook for threshold which will be called before executing the work,
222 * This hook is called in kthread content.
223 * So workqueue_set_max_active is called here.
224 */
Qu Wenruod458b052014-02-28 10:46:19 +0800225static inline void thresh_exec_hook(struct __btrfs_workqueue *wq)
Qu Wenruo0bd92892014-02-28 10:46:05 +0800226{
Qu Wenruoc6dd6ea2015-08-20 09:30:39 +0800227 int new_current_active;
Qu Wenruo0bd92892014-02-28 10:46:05 +0800228 long pending;
229 int need_change = 0;
230
231 if (wq->thresh == NO_THRESHOLD)
232 return;
233
234 atomic_dec(&wq->pending);
235 spin_lock(&wq->thres_lock);
236 /*
237 * Use wq->count to limit the calling frequency of
238 * workqueue_set_max_active.
239 */
240 wq->count++;
241 wq->count %= (wq->thresh / 4);
242 if (!wq->count)
243 goto out;
Qu Wenruoc6dd6ea2015-08-20 09:30:39 +0800244 new_current_active = wq->current_active;
Qu Wenruo0bd92892014-02-28 10:46:05 +0800245
246 /*
247 * pending may be changed later, but it's OK since we really
248 * don't need it so accurate to calculate new_max_active.
249 */
250 pending = atomic_read(&wq->pending);
251 if (pending > wq->thresh)
Qu Wenruoc6dd6ea2015-08-20 09:30:39 +0800252 new_current_active++;
Qu Wenruo0bd92892014-02-28 10:46:05 +0800253 if (pending < wq->thresh / 2)
Qu Wenruoc6dd6ea2015-08-20 09:30:39 +0800254 new_current_active--;
255 new_current_active = clamp_val(new_current_active, 1, wq->limit_active);
256 if (new_current_active != wq->current_active) {
Qu Wenruo0bd92892014-02-28 10:46:05 +0800257 need_change = 1;
Qu Wenruoc6dd6ea2015-08-20 09:30:39 +0800258 wq->current_active = new_current_active;
Qu Wenruo0bd92892014-02-28 10:46:05 +0800259 }
260out:
261 spin_unlock(&wq->thres_lock);
262
263 if (need_change) {
Qu Wenruoc6dd6ea2015-08-20 09:30:39 +0800264 workqueue_set_max_active(wq->normal_wq, wq->current_active);
Qu Wenruo0bd92892014-02-28 10:46:05 +0800265 }
266}
267
Qu Wenruod458b052014-02-28 10:46:19 +0800268static void run_ordered_work(struct __btrfs_workqueue *wq)
Qu Wenruo08a9ff32014-02-28 10:46:03 +0800269{
270 struct list_head *list = &wq->ordered_list;
Qu Wenruod458b052014-02-28 10:46:19 +0800271 struct btrfs_work *work;
Qu Wenruo08a9ff32014-02-28 10:46:03 +0800272 spinlock_t *lock = &wq->list_lock;
273 unsigned long flags;
274
275 while (1) {
David Sterba28dad9a2017-01-06 14:12:51 +0100276 void *wtag;
277
Qu Wenruo08a9ff32014-02-28 10:46:03 +0800278 spin_lock_irqsave(lock, flags);
279 if (list_empty(list))
280 break;
Qu Wenruod458b052014-02-28 10:46:19 +0800281 work = list_entry(list->next, struct btrfs_work,
Qu Wenruo08a9ff32014-02-28 10:46:03 +0800282 ordered_list);
283 if (!test_bit(WORK_DONE_BIT, &work->flags))
284 break;
285
286 /*
287 * we are going to call the ordered done function, but
288 * we leave the work item on the list as a barrier so
289 * that later work items that are done don't have their
290 * functions called before this one returns
291 */
292 if (test_and_set_bit(WORK_ORDER_DONE_BIT, &work->flags))
293 break;
Qu Wenruo52483bc2014-03-06 04:19:50 +0000294 trace_btrfs_ordered_sched(work);
Qu Wenruo08a9ff32014-02-28 10:46:03 +0800295 spin_unlock_irqrestore(lock, flags);
296 work->ordered_func(work);
297
298 /* now take the lock again and drop our item from the list */
299 spin_lock_irqsave(lock, flags);
300 list_del(&work->ordered_list);
301 spin_unlock_irqrestore(lock, flags);
302
303 /*
David Sterba28dad9a2017-01-06 14:12:51 +0100304 * We don't want to call the ordered free functions with the
305 * lock held though. Save the work as tag for the trace event,
306 * because the callback could free the structure.
Qu Wenruo08a9ff32014-02-28 10:46:03 +0800307 */
David Sterba28dad9a2017-01-06 14:12:51 +0100308 wtag = work;
Qu Wenruo08a9ff32014-02-28 10:46:03 +0800309 work->ordered_free(work);
David Sterba28dad9a2017-01-06 14:12:51 +0100310 trace_btrfs_all_work_done(wq->fs_info, wtag);
Qu Wenruo08a9ff32014-02-28 10:46:03 +0800311 }
312 spin_unlock_irqrestore(lock, flags);
313}
314
Liu Bo9e0af232014-08-15 23:36:53 +0800315static void normal_work_helper(struct btrfs_work *work)
Qu Wenruo08a9ff32014-02-28 10:46:03 +0800316{
Qu Wenruod458b052014-02-28 10:46:19 +0800317 struct __btrfs_workqueue *wq;
David Sterba28dad9a2017-01-06 14:12:51 +0100318 void *wtag;
Qu Wenruo08a9ff32014-02-28 10:46:03 +0800319 int need_order = 0;
320
Qu Wenruo08a9ff32014-02-28 10:46:03 +0800321 /*
322 * We should not touch things inside work in the following cases:
323 * 1) after work->func() if it has no ordered_free
324 * Since the struct is freed in work->func().
325 * 2) after setting WORK_DONE_BIT
326 * The work may be freed in other threads almost instantly.
327 * So we save the needed things here.
328 */
329 if (work->ordered_func)
330 need_order = 1;
331 wq = work->wq;
David Sterba28dad9a2017-01-06 14:12:51 +0100332 /* Safe for tracepoints in case work gets freed by the callback */
333 wtag = work;
Qu Wenruo08a9ff32014-02-28 10:46:03 +0800334
Qu Wenruo52483bc2014-03-06 04:19:50 +0000335 trace_btrfs_work_sched(work);
Qu Wenruo0bd92892014-02-28 10:46:05 +0800336 thresh_exec_hook(wq);
Qu Wenruo08a9ff32014-02-28 10:46:03 +0800337 work->func(work);
338 if (need_order) {
339 set_bit(WORK_DONE_BIT, &work->flags);
340 run_ordered_work(wq);
341 }
Qu Wenruo52483bc2014-03-06 04:19:50 +0000342 if (!need_order)
David Sterba28dad9a2017-01-06 14:12:51 +0100343 trace_btrfs_all_work_done(wq->fs_info, wtag);
Qu Wenruo08a9ff32014-02-28 10:46:03 +0800344}
345
Liu Bo9e0af232014-08-15 23:36:53 +0800346void btrfs_init_work(struct btrfs_work *work, btrfs_work_func_t uniq_func,
Qu Wenruo6db89142014-03-06 04:19:50 +0000347 btrfs_func_t func,
348 btrfs_func_t ordered_func,
349 btrfs_func_t ordered_free)
Qu Wenruo08a9ff32014-02-28 10:46:03 +0800350{
351 work->func = func;
352 work->ordered_func = ordered_func;
353 work->ordered_free = ordered_free;
Liu Bo9e0af232014-08-15 23:36:53 +0800354 INIT_WORK(&work->normal_work, uniq_func);
Qu Wenruo08a9ff32014-02-28 10:46:03 +0800355 INIT_LIST_HEAD(&work->ordered_list);
356 work->flags = 0;
357}
358
Qu Wenruod458b052014-02-28 10:46:19 +0800359static inline void __btrfs_queue_work(struct __btrfs_workqueue *wq,
360 struct btrfs_work *work)
Qu Wenruo08a9ff32014-02-28 10:46:03 +0800361{
362 unsigned long flags;
363
364 work->wq = wq;
Qu Wenruo0bd92892014-02-28 10:46:05 +0800365 thresh_queue_hook(wq);
Qu Wenruo08a9ff32014-02-28 10:46:03 +0800366 if (work->ordered_func) {
367 spin_lock_irqsave(&wq->list_lock, flags);
368 list_add_tail(&work->ordered_list, &wq->ordered_list);
369 spin_unlock_irqrestore(&wq->list_lock, flags);
370 }
Qu Wenruo52483bc2014-03-06 04:19:50 +0000371 trace_btrfs_work_queued(work);
Qu Wenruo0a95b852016-01-22 09:28:38 +0800372 queue_work(wq->normal_wq, &work->normal_work);
Qu Wenruo08a9ff32014-02-28 10:46:03 +0800373}
374
Qu Wenruod458b052014-02-28 10:46:19 +0800375void btrfs_queue_work(struct btrfs_workqueue *wq,
376 struct btrfs_work *work)
Qu Wenruo1ca08972014-02-28 10:46:04 +0800377{
Qu Wenruod458b052014-02-28 10:46:19 +0800378 struct __btrfs_workqueue *dest_wq;
Qu Wenruo1ca08972014-02-28 10:46:04 +0800379
380 if (test_bit(WORK_HIGH_PRIO_BIT, &work->flags) && wq->high)
381 dest_wq = wq->high;
382 else
383 dest_wq = wq->normal;
384 __btrfs_queue_work(dest_wq, work);
385}
386
387static inline void
Qu Wenruod458b052014-02-28 10:46:19 +0800388__btrfs_destroy_workqueue(struct __btrfs_workqueue *wq)
Qu Wenruo08a9ff32014-02-28 10:46:03 +0800389{
390 destroy_workqueue(wq->normal_wq);
Qu Wenruoc3a46892014-03-12 08:05:33 +0000391 trace_btrfs_workqueue_destroy(wq);
Qu Wenruo08a9ff32014-02-28 10:46:03 +0800392 kfree(wq);
393}
394
Qu Wenruod458b052014-02-28 10:46:19 +0800395void btrfs_destroy_workqueue(struct btrfs_workqueue *wq)
Qu Wenruo1ca08972014-02-28 10:46:04 +0800396{
397 if (!wq)
398 return;
399 if (wq->high)
400 __btrfs_destroy_workqueue(wq->high);
401 __btrfs_destroy_workqueue(wq->normal);
Filipe Mananaef66af12014-03-11 14:31:44 +0000402 kfree(wq);
Qu Wenruo1ca08972014-02-28 10:46:04 +0800403}
404
Qu Wenruoc6dd6ea2015-08-20 09:30:39 +0800405void btrfs_workqueue_set_max(struct btrfs_workqueue *wq, int limit_active)
Qu Wenruo08a9ff32014-02-28 10:46:03 +0800406{
Sergei Trofimovich800ee222014-04-07 10:55:46 +0300407 if (!wq)
408 return;
Qu Wenruoc6dd6ea2015-08-20 09:30:39 +0800409 wq->normal->limit_active = limit_active;
Qu Wenruo1ca08972014-02-28 10:46:04 +0800410 if (wq->high)
Qu Wenruoc6dd6ea2015-08-20 09:30:39 +0800411 wq->high->limit_active = limit_active;
Qu Wenruo1ca08972014-02-28 10:46:04 +0800412}
413
Qu Wenruod458b052014-02-28 10:46:19 +0800414void btrfs_set_work_high_priority(struct btrfs_work *work)
Qu Wenruo1ca08972014-02-28 10:46:04 +0800415{
416 set_bit(WORK_HIGH_PRIO_BIT, &work->flags);
Qu Wenruo08a9ff32014-02-28 10:46:03 +0800417}