blob: d522494698fa4ee50b91adfe94eec125506e62d4 [file] [log] [blame]
David Sterbac1d7c512018-04-03 19:23:33 +02001// SPDX-License-Identifier: GPL-2.0
Chris Mason8b712842008-06-11 16:50:36 -04002/*
3 * Copyright (C) 2007 Oracle. All rights reserved.
Qu Wenruo08a9ff32014-02-28 10:46:03 +08004 * Copyright (C) 2014 Fujitsu. All rights reserved.
Chris Mason8b712842008-06-11 16:50:36 -04005 */
6
7#include <linux/kthread.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +09008#include <linux/slab.h>
Chris Mason8b712842008-06-11 16:50:36 -04009#include <linux/list.h>
10#include <linux/spinlock.h>
Chris Masonb51912c2009-02-04 09:23:24 -050011#include <linux/freezer.h>
Chris Mason8b712842008-06-11 16:50:36 -040012#include "async-thread.h"
Qu Wenruo52483bc2014-03-06 04:19:50 +000013#include "ctree.h"
Chris Mason8b712842008-06-11 16:50:36 -040014
Qu Wenruoa046e9c2014-02-28 10:46:18 +080015#define WORK_DONE_BIT 0
16#define WORK_ORDER_DONE_BIT 1
17#define WORK_HIGH_PRIO_BIT 2
Chris Mason4a69a412008-11-06 22:03:00 -050018
Qu Wenruo0bd92892014-02-28 10:46:05 +080019#define NO_THRESHOLD (-1)
20#define DFT_THRESHOLD (32)
21
Qu Wenruod458b052014-02-28 10:46:19 +080022struct __btrfs_workqueue {
Qu Wenruo08a9ff32014-02-28 10:46:03 +080023 struct workqueue_struct *normal_wq;
Jeff Mahoneycb001092016-06-09 16:22:11 -040024
25 /* File system this workqueue services */
26 struct btrfs_fs_info *fs_info;
27
Qu Wenruo08a9ff32014-02-28 10:46:03 +080028 /* List head pointing to ordered work list */
29 struct list_head ordered_list;
30
31 /* Spinlock for ordered_list */
32 spinlock_t list_lock;
Qu Wenruo0bd92892014-02-28 10:46:05 +080033
34 /* Thresholding related variants */
35 atomic_t pending;
Qu Wenruoc6dd6ea2015-08-20 09:30:39 +080036
37 /* Up limit of concurrency workers */
38 int limit_active;
39
40 /* Current number of concurrency workers */
41 int current_active;
42
43 /* Threshold to change current_active */
Qu Wenruo0bd92892014-02-28 10:46:05 +080044 int thresh;
45 unsigned int count;
46 spinlock_t thres_lock;
Qu Wenruo08a9ff32014-02-28 10:46:03 +080047};
48
Qu Wenruod458b052014-02-28 10:46:19 +080049struct btrfs_workqueue {
50 struct __btrfs_workqueue *normal;
51 struct __btrfs_workqueue *high;
Qu Wenruo1ca08972014-02-28 10:46:04 +080052};
53
Liu Bo9e0af232014-08-15 23:36:53 +080054static void normal_work_helper(struct btrfs_work *work);
55
56#define BTRFS_WORK_HELPER(name) \
Liu Bo6939f662017-09-13 12:09:28 -060057noinline_for_stack void btrfs_##name(struct work_struct *arg) \
Liu Bo9e0af232014-08-15 23:36:53 +080058{ \
59 struct btrfs_work *work = container_of(arg, struct btrfs_work, \
60 normal_work); \
61 normal_work_helper(work); \
62}
63
Jeff Mahoneycb001092016-06-09 16:22:11 -040064struct btrfs_fs_info *
Jeff Mahoney9a35b632017-06-28 21:56:54 -060065btrfs_workqueue_owner(const struct __btrfs_workqueue *wq)
Jeff Mahoneycb001092016-06-09 16:22:11 -040066{
67 return wq->fs_info;
68}
69
70struct btrfs_fs_info *
Jeff Mahoney9a35b632017-06-28 21:56:54 -060071btrfs_work_owner(const struct btrfs_work *work)
Jeff Mahoneycb001092016-06-09 16:22:11 -040072{
73 return work->wq->fs_info;
74}
75
Jeff Mahoney9a35b632017-06-28 21:56:54 -060076bool btrfs_workqueue_normal_congested(const struct btrfs_workqueue *wq)
Maxim Patlasov2939e1a2016-12-12 14:32:44 -080077{
78 /*
79 * We could compare wq->normal->pending with num_online_cpus()
80 * to support "thresh == NO_THRESHOLD" case, but it requires
81 * moving up atomic_inc/dec in thresh_queue/exec_hook. Let's
82 * postpone it until someone needs the support of that case.
83 */
84 if (wq->normal->thresh == NO_THRESHOLD)
85 return false;
86
87 return atomic_read(&wq->normal->pending) > wq->normal->thresh * 2;
88}
89
Liu Bo9e0af232014-08-15 23:36:53 +080090BTRFS_WORK_HELPER(worker_helper);
91BTRFS_WORK_HELPER(delalloc_helper);
92BTRFS_WORK_HELPER(flush_delalloc_helper);
93BTRFS_WORK_HELPER(cache_helper);
94BTRFS_WORK_HELPER(submit_helper);
95BTRFS_WORK_HELPER(fixup_helper);
96BTRFS_WORK_HELPER(endio_helper);
97BTRFS_WORK_HELPER(endio_meta_helper);
98BTRFS_WORK_HELPER(endio_meta_write_helper);
99BTRFS_WORK_HELPER(endio_raid56_helper);
Miao Xie8b110e32014-09-12 18:44:03 +0800100BTRFS_WORK_HELPER(endio_repair_helper);
Liu Bo9e0af232014-08-15 23:36:53 +0800101BTRFS_WORK_HELPER(rmw_helper);
102BTRFS_WORK_HELPER(endio_write_helper);
103BTRFS_WORK_HELPER(freespace_write_helper);
104BTRFS_WORK_HELPER(delayed_meta_helper);
105BTRFS_WORK_HELPER(readahead_helper);
106BTRFS_WORK_HELPER(qgroup_rescan_helper);
107BTRFS_WORK_HELPER(extent_refs_helper);
108BTRFS_WORK_HELPER(scrub_helper);
109BTRFS_WORK_HELPER(scrubwrc_helper);
110BTRFS_WORK_HELPER(scrubnc_helper);
Zhao Lei20b2e302015-06-04 20:09:15 +0800111BTRFS_WORK_HELPER(scrubparity_helper);
Liu Bo9e0af232014-08-15 23:36:53 +0800112
113static struct __btrfs_workqueue *
Jeff Mahoneycb001092016-06-09 16:22:11 -0400114__btrfs_alloc_workqueue(struct btrfs_fs_info *fs_info, const char *name,
115 unsigned int flags, int limit_active, int thresh)
Qu Wenruo08a9ff32014-02-28 10:46:03 +0800116{
David Sterba61dd5ae2015-12-01 18:04:30 +0100117 struct __btrfs_workqueue *ret = kzalloc(sizeof(*ret), GFP_KERNEL);
Qu Wenruo08a9ff32014-02-28 10:46:03 +0800118
David Sterba5d99a9982014-09-29 19:20:37 +0200119 if (!ret)
Qu Wenruo08a9ff32014-02-28 10:46:03 +0800120 return NULL;
121
Jeff Mahoneycb001092016-06-09 16:22:11 -0400122 ret->fs_info = fs_info;
Qu Wenruoc6dd6ea2015-08-20 09:30:39 +0800123 ret->limit_active = limit_active;
Qu Wenruo0bd92892014-02-28 10:46:05 +0800124 atomic_set(&ret->pending, 0);
125 if (thresh == 0)
126 thresh = DFT_THRESHOLD;
127 /* For low threshold, disabling threshold is a better choice */
128 if (thresh < DFT_THRESHOLD) {
Qu Wenruoc6dd6ea2015-08-20 09:30:39 +0800129 ret->current_active = limit_active;
Qu Wenruo0bd92892014-02-28 10:46:05 +0800130 ret->thresh = NO_THRESHOLD;
131 } else {
Qu Wenruoc6dd6ea2015-08-20 09:30:39 +0800132 /*
133 * For threshold-able wq, let its concurrency grow on demand.
134 * Use minimal max_active at alloc time to reduce resource
135 * usage.
136 */
137 ret->current_active = 1;
Qu Wenruo0bd92892014-02-28 10:46:05 +0800138 ret->thresh = thresh;
139 }
140
Qu Wenruo1ca08972014-02-28 10:46:04 +0800141 if (flags & WQ_HIGHPRI)
142 ret->normal_wq = alloc_workqueue("%s-%s-high", flags,
Qu Wenruoc6dd6ea2015-08-20 09:30:39 +0800143 ret->current_active, "btrfs",
144 name);
Qu Wenruo1ca08972014-02-28 10:46:04 +0800145 else
146 ret->normal_wq = alloc_workqueue("%s-%s", flags,
Qu Wenruoc6dd6ea2015-08-20 09:30:39 +0800147 ret->current_active, "btrfs",
Qu Wenruo0bd92892014-02-28 10:46:05 +0800148 name);
David Sterba5d99a9982014-09-29 19:20:37 +0200149 if (!ret->normal_wq) {
Qu Wenruo08a9ff32014-02-28 10:46:03 +0800150 kfree(ret);
151 return NULL;
152 }
153
154 INIT_LIST_HEAD(&ret->ordered_list);
155 spin_lock_init(&ret->list_lock);
Qu Wenruo0bd92892014-02-28 10:46:05 +0800156 spin_lock_init(&ret->thres_lock);
Qu Wenruoc3a46892014-03-12 08:05:33 +0000157 trace_btrfs_workqueue_alloc(ret, name, flags & WQ_HIGHPRI);
Qu Wenruo08a9ff32014-02-28 10:46:03 +0800158 return ret;
159}
160
Qu Wenruo1ca08972014-02-28 10:46:04 +0800161static inline void
Qu Wenruod458b052014-02-28 10:46:19 +0800162__btrfs_destroy_workqueue(struct __btrfs_workqueue *wq);
Qu Wenruo1ca08972014-02-28 10:46:04 +0800163
Jeff Mahoneycb001092016-06-09 16:22:11 -0400164struct btrfs_workqueue *btrfs_alloc_workqueue(struct btrfs_fs_info *fs_info,
165 const char *name,
David Sterba6f011052015-02-16 18:34:01 +0100166 unsigned int flags,
Qu Wenruoc6dd6ea2015-08-20 09:30:39 +0800167 int limit_active,
Qu Wenruod458b052014-02-28 10:46:19 +0800168 int thresh)
Qu Wenruo1ca08972014-02-28 10:46:04 +0800169{
David Sterba61dd5ae2015-12-01 18:04:30 +0100170 struct btrfs_workqueue *ret = kzalloc(sizeof(*ret), GFP_KERNEL);
Qu Wenruo1ca08972014-02-28 10:46:04 +0800171
David Sterba5d99a9982014-09-29 19:20:37 +0200172 if (!ret)
Qu Wenruo1ca08972014-02-28 10:46:04 +0800173 return NULL;
174
Jeff Mahoneycb001092016-06-09 16:22:11 -0400175 ret->normal = __btrfs_alloc_workqueue(fs_info, name,
176 flags & ~WQ_HIGHPRI,
Qu Wenruoc6dd6ea2015-08-20 09:30:39 +0800177 limit_active, thresh);
David Sterba5d99a9982014-09-29 19:20:37 +0200178 if (!ret->normal) {
Qu Wenruo1ca08972014-02-28 10:46:04 +0800179 kfree(ret);
180 return NULL;
181 }
182
183 if (flags & WQ_HIGHPRI) {
Jeff Mahoneycb001092016-06-09 16:22:11 -0400184 ret->high = __btrfs_alloc_workqueue(fs_info, name, flags,
185 limit_active, thresh);
David Sterba5d99a9982014-09-29 19:20:37 +0200186 if (!ret->high) {
Qu Wenruo1ca08972014-02-28 10:46:04 +0800187 __btrfs_destroy_workqueue(ret->normal);
188 kfree(ret);
189 return NULL;
190 }
191 }
192 return ret;
193}
194
Qu Wenruo0bd92892014-02-28 10:46:05 +0800195/*
196 * Hook for threshold which will be called in btrfs_queue_work.
197 * This hook WILL be called in IRQ handler context,
198 * so workqueue_set_max_active MUST NOT be called in this hook
199 */
Qu Wenruod458b052014-02-28 10:46:19 +0800200static inline void thresh_queue_hook(struct __btrfs_workqueue *wq)
Qu Wenruo0bd92892014-02-28 10:46:05 +0800201{
202 if (wq->thresh == NO_THRESHOLD)
203 return;
204 atomic_inc(&wq->pending);
205}
206
207/*
208 * Hook for threshold which will be called before executing the work,
209 * This hook is called in kthread content.
210 * So workqueue_set_max_active is called here.
211 */
Qu Wenruod458b052014-02-28 10:46:19 +0800212static inline void thresh_exec_hook(struct __btrfs_workqueue *wq)
Qu Wenruo0bd92892014-02-28 10:46:05 +0800213{
Qu Wenruoc6dd6ea2015-08-20 09:30:39 +0800214 int new_current_active;
Qu Wenruo0bd92892014-02-28 10:46:05 +0800215 long pending;
216 int need_change = 0;
217
218 if (wq->thresh == NO_THRESHOLD)
219 return;
220
221 atomic_dec(&wq->pending);
222 spin_lock(&wq->thres_lock);
223 /*
224 * Use wq->count to limit the calling frequency of
225 * workqueue_set_max_active.
226 */
227 wq->count++;
228 wq->count %= (wq->thresh / 4);
229 if (!wq->count)
230 goto out;
Qu Wenruoc6dd6ea2015-08-20 09:30:39 +0800231 new_current_active = wq->current_active;
Qu Wenruo0bd92892014-02-28 10:46:05 +0800232
233 /*
234 * pending may be changed later, but it's OK since we really
235 * don't need it so accurate to calculate new_max_active.
236 */
237 pending = atomic_read(&wq->pending);
238 if (pending > wq->thresh)
Qu Wenruoc6dd6ea2015-08-20 09:30:39 +0800239 new_current_active++;
Qu Wenruo0bd92892014-02-28 10:46:05 +0800240 if (pending < wq->thresh / 2)
Qu Wenruoc6dd6ea2015-08-20 09:30:39 +0800241 new_current_active--;
242 new_current_active = clamp_val(new_current_active, 1, wq->limit_active);
243 if (new_current_active != wq->current_active) {
Qu Wenruo0bd92892014-02-28 10:46:05 +0800244 need_change = 1;
Qu Wenruoc6dd6ea2015-08-20 09:30:39 +0800245 wq->current_active = new_current_active;
Qu Wenruo0bd92892014-02-28 10:46:05 +0800246 }
247out:
248 spin_unlock(&wq->thres_lock);
249
250 if (need_change) {
Qu Wenruoc6dd6ea2015-08-20 09:30:39 +0800251 workqueue_set_max_active(wq->normal_wq, wq->current_active);
Qu Wenruo0bd92892014-02-28 10:46:05 +0800252 }
253}
254
Qu Wenruod458b052014-02-28 10:46:19 +0800255static void run_ordered_work(struct __btrfs_workqueue *wq)
Qu Wenruo08a9ff32014-02-28 10:46:03 +0800256{
257 struct list_head *list = &wq->ordered_list;
Qu Wenruod458b052014-02-28 10:46:19 +0800258 struct btrfs_work *work;
Qu Wenruo08a9ff32014-02-28 10:46:03 +0800259 spinlock_t *lock = &wq->list_lock;
260 unsigned long flags;
261
262 while (1) {
David Sterbaac0c7cf2017-01-06 14:12:51 +0100263 void *wtag;
264
Qu Wenruo08a9ff32014-02-28 10:46:03 +0800265 spin_lock_irqsave(lock, flags);
266 if (list_empty(list))
267 break;
Qu Wenruod458b052014-02-28 10:46:19 +0800268 work = list_entry(list->next, struct btrfs_work,
Qu Wenruo08a9ff32014-02-28 10:46:03 +0800269 ordered_list);
270 if (!test_bit(WORK_DONE_BIT, &work->flags))
271 break;
272
273 /*
274 * we are going to call the ordered done function, but
275 * we leave the work item on the list as a barrier so
276 * that later work items that are done don't have their
277 * functions called before this one returns
278 */
279 if (test_and_set_bit(WORK_ORDER_DONE_BIT, &work->flags))
280 break;
Qu Wenruo52483bc2014-03-06 04:19:50 +0000281 trace_btrfs_ordered_sched(work);
Qu Wenruo08a9ff32014-02-28 10:46:03 +0800282 spin_unlock_irqrestore(lock, flags);
283 work->ordered_func(work);
284
285 /* now take the lock again and drop our item from the list */
286 spin_lock_irqsave(lock, flags);
287 list_del(&work->ordered_list);
288 spin_unlock_irqrestore(lock, flags);
289
290 /*
David Sterbaac0c7cf2017-01-06 14:12:51 +0100291 * We don't want to call the ordered free functions with the
292 * lock held though. Save the work as tag for the trace event,
293 * because the callback could free the structure.
Qu Wenruo08a9ff32014-02-28 10:46:03 +0800294 */
David Sterbaac0c7cf2017-01-06 14:12:51 +0100295 wtag = work;
Qu Wenruo08a9ff32014-02-28 10:46:03 +0800296 work->ordered_free(work);
David Sterbaac0c7cf2017-01-06 14:12:51 +0100297 trace_btrfs_all_work_done(wq->fs_info, wtag);
Qu Wenruo08a9ff32014-02-28 10:46:03 +0800298 }
299 spin_unlock_irqrestore(lock, flags);
300}
301
Liu Bo9e0af232014-08-15 23:36:53 +0800302static void normal_work_helper(struct btrfs_work *work)
Qu Wenruo08a9ff32014-02-28 10:46:03 +0800303{
Qu Wenruod458b052014-02-28 10:46:19 +0800304 struct __btrfs_workqueue *wq;
David Sterbaac0c7cf2017-01-06 14:12:51 +0100305 void *wtag;
Qu Wenruo08a9ff32014-02-28 10:46:03 +0800306 int need_order = 0;
307
Qu Wenruo08a9ff32014-02-28 10:46:03 +0800308 /*
309 * We should not touch things inside work in the following cases:
310 * 1) after work->func() if it has no ordered_free
311 * Since the struct is freed in work->func().
312 * 2) after setting WORK_DONE_BIT
313 * The work may be freed in other threads almost instantly.
314 * So we save the needed things here.
315 */
316 if (work->ordered_func)
317 need_order = 1;
318 wq = work->wq;
David Sterbaac0c7cf2017-01-06 14:12:51 +0100319 /* Safe for tracepoints in case work gets freed by the callback */
320 wtag = work;
Qu Wenruo08a9ff32014-02-28 10:46:03 +0800321
Qu Wenruo52483bc2014-03-06 04:19:50 +0000322 trace_btrfs_work_sched(work);
Qu Wenruo0bd92892014-02-28 10:46:05 +0800323 thresh_exec_hook(wq);
Qu Wenruo08a9ff32014-02-28 10:46:03 +0800324 work->func(work);
325 if (need_order) {
326 set_bit(WORK_DONE_BIT, &work->flags);
327 run_ordered_work(wq);
328 }
Qu Wenruo52483bc2014-03-06 04:19:50 +0000329 if (!need_order)
David Sterbaac0c7cf2017-01-06 14:12:51 +0100330 trace_btrfs_all_work_done(wq->fs_info, wtag);
Qu Wenruo08a9ff32014-02-28 10:46:03 +0800331}
332
Liu Bo9e0af232014-08-15 23:36:53 +0800333void btrfs_init_work(struct btrfs_work *work, btrfs_work_func_t uniq_func,
Qu Wenruo6db89142014-03-06 04:19:50 +0000334 btrfs_func_t func,
335 btrfs_func_t ordered_func,
336 btrfs_func_t ordered_free)
Qu Wenruo08a9ff32014-02-28 10:46:03 +0800337{
338 work->func = func;
339 work->ordered_func = ordered_func;
340 work->ordered_free = ordered_free;
Liu Bo9e0af232014-08-15 23:36:53 +0800341 INIT_WORK(&work->normal_work, uniq_func);
Qu Wenruo08a9ff32014-02-28 10:46:03 +0800342 INIT_LIST_HEAD(&work->ordered_list);
343 work->flags = 0;
344}
345
Qu Wenruod458b052014-02-28 10:46:19 +0800346static inline void __btrfs_queue_work(struct __btrfs_workqueue *wq,
347 struct btrfs_work *work)
Qu Wenruo08a9ff32014-02-28 10:46:03 +0800348{
349 unsigned long flags;
350
351 work->wq = wq;
Qu Wenruo0bd92892014-02-28 10:46:05 +0800352 thresh_queue_hook(wq);
Qu Wenruo08a9ff32014-02-28 10:46:03 +0800353 if (work->ordered_func) {
354 spin_lock_irqsave(&wq->list_lock, flags);
355 list_add_tail(&work->ordered_list, &wq->ordered_list);
356 spin_unlock_irqrestore(&wq->list_lock, flags);
357 }
Qu Wenruo52483bc2014-03-06 04:19:50 +0000358 trace_btrfs_work_queued(work);
Qu Wenruo0a95b852016-01-22 09:28:38 +0800359 queue_work(wq->normal_wq, &work->normal_work);
Qu Wenruo08a9ff32014-02-28 10:46:03 +0800360}
361
Qu Wenruod458b052014-02-28 10:46:19 +0800362void btrfs_queue_work(struct btrfs_workqueue *wq,
363 struct btrfs_work *work)
Qu Wenruo1ca08972014-02-28 10:46:04 +0800364{
Qu Wenruod458b052014-02-28 10:46:19 +0800365 struct __btrfs_workqueue *dest_wq;
Qu Wenruo1ca08972014-02-28 10:46:04 +0800366
367 if (test_bit(WORK_HIGH_PRIO_BIT, &work->flags) && wq->high)
368 dest_wq = wq->high;
369 else
370 dest_wq = wq->normal;
371 __btrfs_queue_work(dest_wq, work);
372}
373
374static inline void
Qu Wenruod458b052014-02-28 10:46:19 +0800375__btrfs_destroy_workqueue(struct __btrfs_workqueue *wq)
Qu Wenruo08a9ff32014-02-28 10:46:03 +0800376{
377 destroy_workqueue(wq->normal_wq);
Qu Wenruoc3a46892014-03-12 08:05:33 +0000378 trace_btrfs_workqueue_destroy(wq);
Qu Wenruo08a9ff32014-02-28 10:46:03 +0800379 kfree(wq);
380}
381
Qu Wenruod458b052014-02-28 10:46:19 +0800382void btrfs_destroy_workqueue(struct btrfs_workqueue *wq)
Qu Wenruo1ca08972014-02-28 10:46:04 +0800383{
384 if (!wq)
385 return;
386 if (wq->high)
387 __btrfs_destroy_workqueue(wq->high);
388 __btrfs_destroy_workqueue(wq->normal);
Filipe Mananaef66af12014-03-11 14:31:44 +0000389 kfree(wq);
Qu Wenruo1ca08972014-02-28 10:46:04 +0800390}
391
Qu Wenruoc6dd6ea2015-08-20 09:30:39 +0800392void btrfs_workqueue_set_max(struct btrfs_workqueue *wq, int limit_active)
Qu Wenruo08a9ff32014-02-28 10:46:03 +0800393{
Sergei Trofimovich800ee222014-04-07 10:55:46 +0300394 if (!wq)
395 return;
Qu Wenruoc6dd6ea2015-08-20 09:30:39 +0800396 wq->normal->limit_active = limit_active;
Qu Wenruo1ca08972014-02-28 10:46:04 +0800397 if (wq->high)
Qu Wenruoc6dd6ea2015-08-20 09:30:39 +0800398 wq->high->limit_active = limit_active;
Qu Wenruo1ca08972014-02-28 10:46:04 +0800399}
400
Qu Wenruod458b052014-02-28 10:46:19 +0800401void btrfs_set_work_high_priority(struct btrfs_work *work)
Qu Wenruo1ca08972014-02-28 10:46:04 +0800402{
403 set_bit(WORK_HIGH_PRIO_BIT, &work->flags);
Qu Wenruo08a9ff32014-02-28 10:46:03 +0800404}