1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (C) 2007 Oracle. All rights reserved.
4 * Copyright (C) 2014 Fujitsu. All rights reserved.
7 #include <linux/kthread.h>
8 #include <linux/slab.h>
9 #include <linux/list.h>
10 #include <linux/spinlock.h>
11 #include <linux/freezer.h>
12 #include "async-thread.h"
15 #define WORK_DONE_BIT 0
16 #define WORK_ORDER_DONE_BIT 1
17 #define WORK_HIGH_PRIO_BIT 2
19 #define NO_THRESHOLD (-1)
20 #define DFT_THRESHOLD (32)
22 struct __btrfs_workqueue {
23 struct workqueue_struct *normal_wq;
25 /* File system this workqueue services */
26 struct btrfs_fs_info *fs_info;
28 /* List head pointing to ordered work list */
29 struct list_head ordered_list;
31 /* Spinlock for ordered_list */
34 /* Thresholding related variants */
37 /* Up limit of concurrency workers */
40 /* Current number of concurrency workers */
43 /* Threshold to change current_active */
46 spinlock_t thres_lock;
49 struct btrfs_workqueue {
50 struct __btrfs_workqueue *normal;
51 struct __btrfs_workqueue *high;
54 static void normal_work_helper(struct btrfs_work *work);
56 #define BTRFS_WORK_HELPER(name) \
57 noinline_for_stack void btrfs_##name(struct work_struct *arg) \
59 struct btrfs_work *work = container_of(arg, struct btrfs_work, \
61 normal_work_helper(work); \
64 struct btrfs_fs_info *
65 btrfs_workqueue_owner(const struct __btrfs_workqueue *wq)
70 struct btrfs_fs_info *
71 btrfs_work_owner(const struct btrfs_work *work)
73 return work->wq->fs_info;
76 bool btrfs_workqueue_normal_congested(const struct btrfs_workqueue *wq)
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.
84 if (wq->normal->thresh == NO_THRESHOLD)
87 return atomic_read(&wq->normal->pending) > wq->normal->thresh * 2;
90 BTRFS_WORK_HELPER(worker_helper);
91 BTRFS_WORK_HELPER(delalloc_helper);
92 BTRFS_WORK_HELPER(flush_delalloc_helper);
93 BTRFS_WORK_HELPER(cache_helper);
94 BTRFS_WORK_HELPER(submit_helper);
95 BTRFS_WORK_HELPER(fixup_helper);
96 BTRFS_WORK_HELPER(endio_helper);
97 BTRFS_WORK_HELPER(endio_meta_helper);
98 BTRFS_WORK_HELPER(endio_meta_write_helper);
99 BTRFS_WORK_HELPER(endio_raid56_helper);
100 BTRFS_WORK_HELPER(endio_repair_helper);
101 BTRFS_WORK_HELPER(rmw_helper);
102 BTRFS_WORK_HELPER(endio_write_helper);
103 BTRFS_WORK_HELPER(freespace_write_helper);
104 BTRFS_WORK_HELPER(delayed_meta_helper);
105 BTRFS_WORK_HELPER(readahead_helper);
106 BTRFS_WORK_HELPER(qgroup_rescan_helper);
107 BTRFS_WORK_HELPER(extent_refs_helper);
108 BTRFS_WORK_HELPER(scrub_helper);
109 BTRFS_WORK_HELPER(scrubwrc_helper);
110 BTRFS_WORK_HELPER(scrubnc_helper);
111 BTRFS_WORK_HELPER(scrubparity_helper);
113 static struct __btrfs_workqueue *
114 __btrfs_alloc_workqueue(struct btrfs_fs_info *fs_info, const char *name,
115 unsigned int flags, int limit_active, int thresh)
117 struct __btrfs_workqueue *ret = kzalloc(sizeof(*ret), GFP_KERNEL);
122 ret->fs_info = fs_info;
123 ret->limit_active = limit_active;
124 atomic_set(&ret->pending, 0);
126 thresh = DFT_THRESHOLD;
127 /* For low threshold, disabling threshold is a better choice */
128 if (thresh < DFT_THRESHOLD) {
129 ret->current_active = limit_active;
130 ret->thresh = NO_THRESHOLD;
133 * For threshold-able wq, let its concurrency grow on demand.
134 * Use minimal max_active at alloc time to reduce resource
137 ret->current_active = 1;
138 ret->thresh = thresh;
141 if (flags & WQ_HIGHPRI)
142 ret->normal_wq = alloc_workqueue("%s-%s-high", flags,
143 ret->current_active, "btrfs",
146 ret->normal_wq = alloc_workqueue("%s-%s", flags,
147 ret->current_active, "btrfs",
149 if (!ret->normal_wq) {
154 INIT_LIST_HEAD(&ret->ordered_list);
155 spin_lock_init(&ret->list_lock);
156 spin_lock_init(&ret->thres_lock);
157 trace_btrfs_workqueue_alloc(ret, name, flags & WQ_HIGHPRI);
162 __btrfs_destroy_workqueue(struct __btrfs_workqueue *wq);
164 struct btrfs_workqueue *btrfs_alloc_workqueue(struct btrfs_fs_info *fs_info,
170 struct btrfs_workqueue *ret = kzalloc(sizeof(*ret), GFP_KERNEL);
175 ret->normal = __btrfs_alloc_workqueue(fs_info, name,
177 limit_active, thresh);
183 if (flags & WQ_HIGHPRI) {
184 ret->high = __btrfs_alloc_workqueue(fs_info, name, flags,
185 limit_active, thresh);
187 __btrfs_destroy_workqueue(ret->normal);
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
200 static inline void thresh_queue_hook(struct __btrfs_workqueue *wq)
202 if (wq->thresh == NO_THRESHOLD)
204 atomic_inc(&wq->pending);
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.
212 static inline void thresh_exec_hook(struct __btrfs_workqueue *wq)
214 int new_current_active;
218 if (wq->thresh == NO_THRESHOLD)
221 atomic_dec(&wq->pending);
222 spin_lock(&wq->thres_lock);
224 * Use wq->count to limit the calling frequency of
225 * workqueue_set_max_active.
228 wq->count %= (wq->thresh / 4);
231 new_current_active = wq->current_active;
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.
237 pending = atomic_read(&wq->pending);
238 if (pending > wq->thresh)
239 new_current_active++;
240 if (pending < wq->thresh / 2)
241 new_current_active--;
242 new_current_active = clamp_val(new_current_active, 1, wq->limit_active);
243 if (new_current_active != wq->current_active) {
245 wq->current_active = new_current_active;
248 spin_unlock(&wq->thres_lock);
251 workqueue_set_max_active(wq->normal_wq, wq->current_active);
255 static void run_ordered_work(struct __btrfs_workqueue *wq,
256 struct btrfs_work *self)
258 struct list_head *list = &wq->ordered_list;
259 struct btrfs_work *work;
260 spinlock_t *lock = &wq->list_lock;
263 bool free_self = false;
266 spin_lock_irqsave(lock, flags);
267 if (list_empty(list))
269 work = list_entry(list->next, struct btrfs_work,
271 if (!test_bit(WORK_DONE_BIT, &work->flags))
275 * we are going to call the ordered done function, but
276 * we leave the work item on the list as a barrier so
277 * that later work items that are done don't have their
278 * functions called before this one returns
280 if (test_and_set_bit(WORK_ORDER_DONE_BIT, &work->flags))
282 trace_btrfs_ordered_sched(work);
283 spin_unlock_irqrestore(lock, flags);
284 work->ordered_func(work);
286 /* now take the lock again and drop our item from the list */
287 spin_lock_irqsave(lock, flags);
288 list_del(&work->ordered_list);
289 spin_unlock_irqrestore(lock, flags);
293 * This is the work item that the worker is currently
296 * The kernel workqueue code guarantees non-reentrancy
297 * of work items. I.e., if a work item with the same
298 * address and work function is queued twice, the second
299 * execution is blocked until the first one finishes. A
300 * work item may be freed and recycled with the same
301 * work function; the workqueue code assumes that the
302 * original work item cannot depend on the recycled work
303 * item in that case (see find_worker_executing_work()).
305 * Note that the work of one Btrfs filesystem may depend
306 * on the work of another Btrfs filesystem via, e.g., a
307 * loop device. Therefore, we must not allow the current
308 * work item to be recycled until we are really done,
309 * otherwise we break the above assumption and can
315 * We don't want to call the ordered free functions with
316 * the lock held though. Save the work as tag for the
317 * trace event, because the callback could free the
321 work->ordered_free(work);
322 trace_btrfs_all_work_done(wq->fs_info, wtag);
325 spin_unlock_irqrestore(lock, flags);
329 self->ordered_free(self);
330 trace_btrfs_all_work_done(wq->fs_info, wtag);
334 static void normal_work_helper(struct btrfs_work *work)
336 struct __btrfs_workqueue *wq;
341 * We should not touch things inside work in the following cases:
342 * 1) after work->func() if it has no ordered_free
343 * Since the struct is freed in work->func().
344 * 2) after setting WORK_DONE_BIT
345 * The work may be freed in other threads almost instantly.
346 * So we save the needed things here.
348 if (work->ordered_func)
351 /* Safe for tracepoints in case work gets freed by the callback */
354 trace_btrfs_work_sched(work);
355 thresh_exec_hook(wq);
358 set_bit(WORK_DONE_BIT, &work->flags);
359 run_ordered_work(wq, work);
362 trace_btrfs_all_work_done(wq->fs_info, wtag);
365 void btrfs_init_work(struct btrfs_work *work, btrfs_work_func_t uniq_func,
367 btrfs_func_t ordered_func,
368 btrfs_func_t ordered_free)
371 work->ordered_func = ordered_func;
372 work->ordered_free = ordered_free;
373 INIT_WORK(&work->normal_work, uniq_func);
374 INIT_LIST_HEAD(&work->ordered_list);
378 static inline void __btrfs_queue_work(struct __btrfs_workqueue *wq,
379 struct btrfs_work *work)
384 thresh_queue_hook(wq);
385 if (work->ordered_func) {
386 spin_lock_irqsave(&wq->list_lock, flags);
387 list_add_tail(&work->ordered_list, &wq->ordered_list);
388 spin_unlock_irqrestore(&wq->list_lock, flags);
390 trace_btrfs_work_queued(work);
391 queue_work(wq->normal_wq, &work->normal_work);
394 void btrfs_queue_work(struct btrfs_workqueue *wq,
395 struct btrfs_work *work)
397 struct __btrfs_workqueue *dest_wq;
399 if (test_bit(WORK_HIGH_PRIO_BIT, &work->flags) && wq->high)
402 dest_wq = wq->normal;
403 __btrfs_queue_work(dest_wq, work);
407 __btrfs_destroy_workqueue(struct __btrfs_workqueue *wq)
409 destroy_workqueue(wq->normal_wq);
410 trace_btrfs_workqueue_destroy(wq);
414 void btrfs_destroy_workqueue(struct btrfs_workqueue *wq)
419 __btrfs_destroy_workqueue(wq->high);
420 __btrfs_destroy_workqueue(wq->normal);
424 void btrfs_workqueue_set_max(struct btrfs_workqueue *wq, int limit_active)
428 wq->normal->limit_active = limit_active;
430 wq->high->limit_active = limit_active;
433 void btrfs_set_work_high_priority(struct btrfs_work *work)
435 set_bit(WORK_HIGH_PRIO_BIT, &work->flags);
438 void btrfs_flush_workqueue(struct btrfs_workqueue *wq)
441 flush_workqueue(wq->high->normal_wq);
443 flush_workqueue(wq->normal->normal_wq);