2 * Copyright (C) 2007 Oracle. All rights reserved.
3 * Copyright (C) 2014 Fujitsu. All rights reserved.
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.
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.
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.
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"
28 #define WORK_DONE_BIT 0
29 #define WORK_ORDER_DONE_BIT 1
30 #define WORK_HIGH_PRIO_BIT 2
32 #define NO_THRESHOLD (-1)
33 #define DFT_THRESHOLD (32)
35 struct __btrfs_workqueue {
36 struct workqueue_struct *normal_wq;
38 /* File system this workqueue services */
39 struct btrfs_fs_info *fs_info;
41 /* List head pointing to ordered work list */
42 struct list_head ordered_list;
44 /* Spinlock for ordered_list */
47 /* Thresholding related variants */
50 /* Up limit of concurrency workers */
53 /* Current number of concurrency workers */
56 /* Threshold to change current_active */
59 spinlock_t thres_lock;
62 struct btrfs_workqueue {
63 struct __btrfs_workqueue *normal;
64 struct __btrfs_workqueue *high;
67 static void normal_work_helper(struct btrfs_work *work);
69 #define BTRFS_WORK_HELPER(name) \
70 void btrfs_##name(struct work_struct *arg) \
72 struct btrfs_work *work = container_of(arg, struct btrfs_work, \
74 normal_work_helper(work); \
77 struct btrfs_fs_info *
78 btrfs_workqueue_owner(const struct __btrfs_workqueue *wq)
83 struct btrfs_fs_info *
84 btrfs_work_owner(const struct btrfs_work *work)
86 return work->wq->fs_info;
89 bool btrfs_workqueue_normal_congested(const struct btrfs_workqueue *wq)
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.
97 if (wq->normal->thresh == NO_THRESHOLD)
100 return atomic_read(&wq->normal->pending) > wq->normal->thresh * 2;
103 BTRFS_WORK_HELPER(worker_helper);
104 BTRFS_WORK_HELPER(delalloc_helper);
105 BTRFS_WORK_HELPER(flush_delalloc_helper);
106 BTRFS_WORK_HELPER(cache_helper);
107 BTRFS_WORK_HELPER(submit_helper);
108 BTRFS_WORK_HELPER(fixup_helper);
109 BTRFS_WORK_HELPER(endio_helper);
110 BTRFS_WORK_HELPER(endio_meta_helper);
111 BTRFS_WORK_HELPER(endio_meta_write_helper);
112 BTRFS_WORK_HELPER(endio_raid56_helper);
113 BTRFS_WORK_HELPER(endio_repair_helper);
114 BTRFS_WORK_HELPER(rmw_helper);
115 BTRFS_WORK_HELPER(endio_write_helper);
116 BTRFS_WORK_HELPER(freespace_write_helper);
117 BTRFS_WORK_HELPER(delayed_meta_helper);
118 BTRFS_WORK_HELPER(readahead_helper);
119 BTRFS_WORK_HELPER(qgroup_rescan_helper);
120 BTRFS_WORK_HELPER(extent_refs_helper);
121 BTRFS_WORK_HELPER(scrub_helper);
122 BTRFS_WORK_HELPER(scrubwrc_helper);
123 BTRFS_WORK_HELPER(scrubnc_helper);
124 BTRFS_WORK_HELPER(scrubparity_helper);
126 static struct __btrfs_workqueue *
127 __btrfs_alloc_workqueue(struct btrfs_fs_info *fs_info, const char *name,
128 unsigned int flags, int limit_active, int thresh)
130 struct __btrfs_workqueue *ret = kzalloc(sizeof(*ret), GFP_KERNEL);
135 ret->fs_info = fs_info;
136 ret->limit_active = limit_active;
137 atomic_set(&ret->pending, 0);
139 thresh = DFT_THRESHOLD;
140 /* For low threshold, disabling threshold is a better choice */
141 if (thresh < DFT_THRESHOLD) {
142 ret->current_active = limit_active;
143 ret->thresh = NO_THRESHOLD;
146 * For threshold-able wq, let its concurrency grow on demand.
147 * Use minimal max_active at alloc time to reduce resource
150 ret->current_active = 1;
151 ret->thresh = thresh;
154 if (flags & WQ_HIGHPRI)
155 ret->normal_wq = alloc_workqueue("%s-%s-high", flags,
156 ret->current_active, "btrfs",
159 ret->normal_wq = alloc_workqueue("%s-%s", flags,
160 ret->current_active, "btrfs",
162 if (!ret->normal_wq) {
167 INIT_LIST_HEAD(&ret->ordered_list);
168 spin_lock_init(&ret->list_lock);
169 spin_lock_init(&ret->thres_lock);
170 trace_btrfs_workqueue_alloc(ret, name, flags & WQ_HIGHPRI);
175 __btrfs_destroy_workqueue(struct __btrfs_workqueue *wq);
177 struct btrfs_workqueue *btrfs_alloc_workqueue(struct btrfs_fs_info *fs_info,
183 struct btrfs_workqueue *ret = kzalloc(sizeof(*ret), GFP_KERNEL);
188 ret->normal = __btrfs_alloc_workqueue(fs_info, name,
190 limit_active, thresh);
196 if (flags & WQ_HIGHPRI) {
197 ret->high = __btrfs_alloc_workqueue(fs_info, name, flags,
198 limit_active, thresh);
200 __btrfs_destroy_workqueue(ret->normal);
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
213 static inline void thresh_queue_hook(struct __btrfs_workqueue *wq)
215 if (wq->thresh == NO_THRESHOLD)
217 atomic_inc(&wq->pending);
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.
225 static inline void thresh_exec_hook(struct __btrfs_workqueue *wq)
227 int new_current_active;
231 if (wq->thresh == NO_THRESHOLD)
234 atomic_dec(&wq->pending);
235 spin_lock(&wq->thres_lock);
237 * Use wq->count to limit the calling frequency of
238 * workqueue_set_max_active.
241 wq->count %= (wq->thresh / 4);
244 new_current_active = wq->current_active;
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.
250 pending = atomic_read(&wq->pending);
251 if (pending > wq->thresh)
252 new_current_active++;
253 if (pending < wq->thresh / 2)
254 new_current_active--;
255 new_current_active = clamp_val(new_current_active, 1, wq->limit_active);
256 if (new_current_active != wq->current_active) {
258 wq->current_active = new_current_active;
261 spin_unlock(&wq->thres_lock);
264 workqueue_set_max_active(wq->normal_wq, wq->current_active);
268 static void run_ordered_work(struct __btrfs_workqueue *wq,
269 struct btrfs_work *self)
271 struct list_head *list = &wq->ordered_list;
272 struct btrfs_work *work;
273 spinlock_t *lock = &wq->list_lock;
276 bool free_self = false;
279 spin_lock_irqsave(lock, flags);
280 if (list_empty(list))
282 work = list_entry(list->next, struct btrfs_work,
284 if (!test_bit(WORK_DONE_BIT, &work->flags))
287 * Orders all subsequent loads after reading WORK_DONE_BIT,
288 * paired with the smp_mb__before_atomic in btrfs_work_helper
289 * this guarantees that the ordered function will see all
290 * updates from ordinary work function.
295 * we are going to call the ordered done function, but
296 * we leave the work item on the list as a barrier so
297 * that later work items that are done don't have their
298 * functions called before this one returns
300 if (test_and_set_bit(WORK_ORDER_DONE_BIT, &work->flags))
302 trace_btrfs_ordered_sched(work);
303 spin_unlock_irqrestore(lock, flags);
304 work->ordered_func(work);
306 /* now take the lock again and drop our item from the list */
307 spin_lock_irqsave(lock, flags);
308 list_del(&work->ordered_list);
309 spin_unlock_irqrestore(lock, flags);
313 * This is the work item that the worker is currently
316 * The kernel workqueue code guarantees non-reentrancy
317 * of work items. I.e., if a work item with the same
318 * address and work function is queued twice, the second
319 * execution is blocked until the first one finishes. A
320 * work item may be freed and recycled with the same
321 * work function; the workqueue code assumes that the
322 * original work item cannot depend on the recycled work
323 * item in that case (see find_worker_executing_work()).
325 * Note that the work of one Btrfs filesystem may depend
326 * on the work of another Btrfs filesystem via, e.g., a
327 * loop device. Therefore, we must not allow the current
328 * work item to be recycled until we are really done,
329 * otherwise we break the above assumption and can
335 * We don't want to call the ordered free functions with
336 * the lock held though. Save the work as tag for the
337 * trace event, because the callback could free the
341 work->ordered_free(work);
342 trace_btrfs_all_work_done(wq->fs_info, wtag);
345 spin_unlock_irqrestore(lock, flags);
349 self->ordered_free(self);
350 trace_btrfs_all_work_done(wq->fs_info, wtag);
354 static void normal_work_helper(struct btrfs_work *work)
356 struct __btrfs_workqueue *wq;
361 * We should not touch things inside work in the following cases:
362 * 1) after work->func() if it has no ordered_free
363 * Since the struct is freed in work->func().
364 * 2) after setting WORK_DONE_BIT
365 * The work may be freed in other threads almost instantly.
366 * So we save the needed things here.
368 if (work->ordered_func)
371 /* Safe for tracepoints in case work gets freed by the callback */
374 trace_btrfs_work_sched(work);
375 thresh_exec_hook(wq);
379 * Ensures all memory accesses done in the work function are
380 * ordered before setting the WORK_DONE_BIT. Ensuring the thread
381 * which is going to executed the ordered work sees them.
382 * Pairs with the smp_rmb in run_ordered_work.
384 smp_mb__before_atomic();
385 set_bit(WORK_DONE_BIT, &work->flags);
386 run_ordered_work(wq, work);
389 trace_btrfs_all_work_done(wq->fs_info, wtag);
392 void btrfs_init_work(struct btrfs_work *work, btrfs_work_func_t uniq_func,
394 btrfs_func_t ordered_func,
395 btrfs_func_t ordered_free)
398 work->ordered_func = ordered_func;
399 work->ordered_free = ordered_free;
400 INIT_WORK(&work->normal_work, uniq_func);
401 INIT_LIST_HEAD(&work->ordered_list);
405 static inline void __btrfs_queue_work(struct __btrfs_workqueue *wq,
406 struct btrfs_work *work)
411 thresh_queue_hook(wq);
412 if (work->ordered_func) {
413 spin_lock_irqsave(&wq->list_lock, flags);
414 list_add_tail(&work->ordered_list, &wq->ordered_list);
415 spin_unlock_irqrestore(&wq->list_lock, flags);
417 trace_btrfs_work_queued(work);
418 queue_work(wq->normal_wq, &work->normal_work);
421 void btrfs_queue_work(struct btrfs_workqueue *wq,
422 struct btrfs_work *work)
424 struct __btrfs_workqueue *dest_wq;
426 if (test_bit(WORK_HIGH_PRIO_BIT, &work->flags) && wq->high)
429 dest_wq = wq->normal;
430 __btrfs_queue_work(dest_wq, work);
434 __btrfs_destroy_workqueue(struct __btrfs_workqueue *wq)
436 destroy_workqueue(wq->normal_wq);
437 trace_btrfs_workqueue_destroy(wq);
441 void btrfs_destroy_workqueue(struct btrfs_workqueue *wq)
446 __btrfs_destroy_workqueue(wq->high);
447 __btrfs_destroy_workqueue(wq->normal);
451 void btrfs_workqueue_set_max(struct btrfs_workqueue *wq, int limit_active)
455 wq->normal->limit_active = limit_active;
457 wq->high->limit_active = limit_active;
460 void btrfs_set_work_high_priority(struct btrfs_work *work)
462 set_bit(WORK_HIGH_PRIO_BIT, &work->flags);
465 void btrfs_flush_workqueue(struct btrfs_workqueue *wq)
468 flush_workqueue(wq->high->normal_wq);
470 flush_workqueue(wq->normal->normal_wq);