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"
21 #define NO_THRESHOLD (-1)
22 #define DFT_THRESHOLD (32)
24 struct __btrfs_workqueue {
25 struct workqueue_struct *normal_wq;
27 /* File system this workqueue services */
28 struct btrfs_fs_info *fs_info;
30 /* List head pointing to ordered work list */
31 struct list_head ordered_list;
33 /* Spinlock for ordered_list */
36 /* Thresholding related variants */
39 /* Up limit of concurrency workers */
42 /* Current number of concurrency workers */
45 /* Threshold to change current_active */
48 spinlock_t thres_lock;
51 struct btrfs_workqueue {
52 struct __btrfs_workqueue *normal;
53 struct __btrfs_workqueue *high;
56 struct btrfs_fs_info *
57 btrfs_workqueue_owner(const struct __btrfs_workqueue *wq)
62 struct btrfs_fs_info *
63 btrfs_work_owner(const struct btrfs_work *work)
65 return work->wq->fs_info;
68 bool btrfs_workqueue_normal_congested(const struct btrfs_workqueue *wq)
71 * We could compare wq->normal->pending with num_online_cpus()
72 * to support "thresh == NO_THRESHOLD" case, but it requires
73 * moving up atomic_inc/dec in thresh_queue/exec_hook. Let's
74 * postpone it until someone needs the support of that case.
76 if (wq->normal->thresh == NO_THRESHOLD)
79 return atomic_read(&wq->normal->pending) > wq->normal->thresh * 2;
82 static struct __btrfs_workqueue *
83 __btrfs_alloc_workqueue(struct btrfs_fs_info *fs_info, const char *name,
84 unsigned int flags, int limit_active, int thresh)
86 struct __btrfs_workqueue *ret = kzalloc(sizeof(*ret), GFP_KERNEL);
91 ret->fs_info = fs_info;
92 ret->limit_active = limit_active;
93 atomic_set(&ret->pending, 0);
95 thresh = DFT_THRESHOLD;
96 /* For low threshold, disabling threshold is a better choice */
97 if (thresh < DFT_THRESHOLD) {
98 ret->current_active = limit_active;
99 ret->thresh = NO_THRESHOLD;
102 * For threshold-able wq, let its concurrency grow on demand.
103 * Use minimal max_active at alloc time to reduce resource
106 ret->current_active = 1;
107 ret->thresh = thresh;
110 if (flags & WQ_HIGHPRI)
111 ret->normal_wq = alloc_workqueue("btrfs-%s-high", flags,
112 ret->current_active, name);
114 ret->normal_wq = alloc_workqueue("btrfs-%s", flags,
115 ret->current_active, name);
116 if (!ret->normal_wq) {
121 INIT_LIST_HEAD(&ret->ordered_list);
122 spin_lock_init(&ret->list_lock);
123 spin_lock_init(&ret->thres_lock);
124 trace_btrfs_workqueue_alloc(ret, name, flags & WQ_HIGHPRI);
129 __btrfs_destroy_workqueue(struct __btrfs_workqueue *wq);
131 struct btrfs_workqueue *btrfs_alloc_workqueue(struct btrfs_fs_info *fs_info,
137 struct btrfs_workqueue *ret = kzalloc(sizeof(*ret), GFP_KERNEL);
142 ret->normal = __btrfs_alloc_workqueue(fs_info, name,
144 limit_active, thresh);
150 if (flags & WQ_HIGHPRI) {
151 ret->high = __btrfs_alloc_workqueue(fs_info, name, flags,
152 limit_active, thresh);
154 __btrfs_destroy_workqueue(ret->normal);
163 * Hook for threshold which will be called in btrfs_queue_work.
164 * This hook WILL be called in IRQ handler context,
165 * so workqueue_set_max_active MUST NOT be called in this hook
167 static inline void thresh_queue_hook(struct __btrfs_workqueue *wq)
169 if (wq->thresh == NO_THRESHOLD)
171 atomic_inc(&wq->pending);
175 * Hook for threshold which will be called before executing the work,
176 * This hook is called in kthread content.
177 * So workqueue_set_max_active is called here.
179 static inline void thresh_exec_hook(struct __btrfs_workqueue *wq)
181 int new_current_active;
185 if (wq->thresh == NO_THRESHOLD)
188 atomic_dec(&wq->pending);
189 spin_lock(&wq->thres_lock);
191 * Use wq->count to limit the calling frequency of
192 * workqueue_set_max_active.
195 wq->count %= (wq->thresh / 4);
198 new_current_active = wq->current_active;
201 * pending may be changed later, but it's OK since we really
202 * don't need it so accurate to calculate new_max_active.
204 pending = atomic_read(&wq->pending);
205 if (pending > wq->thresh)
206 new_current_active++;
207 if (pending < wq->thresh / 2)
208 new_current_active--;
209 new_current_active = clamp_val(new_current_active, 1, wq->limit_active);
210 if (new_current_active != wq->current_active) {
212 wq->current_active = new_current_active;
215 spin_unlock(&wq->thres_lock);
218 workqueue_set_max_active(wq->normal_wq, wq->current_active);
222 static void run_ordered_work(struct __btrfs_workqueue *wq,
223 struct btrfs_work *self)
225 struct list_head *list = &wq->ordered_list;
226 struct btrfs_work *work;
227 spinlock_t *lock = &wq->list_lock;
230 bool free_self = false;
233 spin_lock_irqsave(lock, flags);
234 if (list_empty(list))
236 work = list_entry(list->next, struct btrfs_work,
238 if (!test_bit(WORK_DONE_BIT, &work->flags))
241 * Orders all subsequent loads after reading WORK_DONE_BIT,
242 * paired with the smp_mb__before_atomic in btrfs_work_helper
243 * this guarantees that the ordered function will see all
244 * updates from ordinary work function.
249 * we are going to call the ordered done function, but
250 * we leave the work item on the list as a barrier so
251 * that later work items that are done don't have their
252 * functions called before this one returns
254 if (test_and_set_bit(WORK_ORDER_DONE_BIT, &work->flags))
256 trace_btrfs_ordered_sched(work);
257 spin_unlock_irqrestore(lock, flags);
258 work->ordered_func(work);
260 /* now take the lock again and drop our item from the list */
261 spin_lock_irqsave(lock, flags);
262 list_del(&work->ordered_list);
263 spin_unlock_irqrestore(lock, flags);
267 * This is the work item that the worker is currently
270 * The kernel workqueue code guarantees non-reentrancy
271 * of work items. I.e., if a work item with the same
272 * address and work function is queued twice, the second
273 * execution is blocked until the first one finishes. A
274 * work item may be freed and recycled with the same
275 * work function; the workqueue code assumes that the
276 * original work item cannot depend on the recycled work
277 * item in that case (see find_worker_executing_work()).
279 * Note that different types of Btrfs work can depend on
280 * each other, and one type of work on one Btrfs
281 * filesystem may even depend on the same type of work
282 * on another Btrfs filesystem via, e.g., a loop device.
283 * Therefore, we must not allow the current work item to
284 * be recycled until we are really done, otherwise we
285 * break the above assumption and can deadlock.
290 * We don't want to call the ordered free functions with
291 * the lock held though. Save the work as tag for the
292 * trace event, because the callback could free the
296 work->ordered_free(work);
297 trace_btrfs_all_work_done(wq->fs_info, wtag);
300 spin_unlock_irqrestore(lock, flags);
304 self->ordered_free(self);
305 trace_btrfs_all_work_done(wq->fs_info, wtag);
309 static void btrfs_work_helper(struct work_struct *normal_work)
311 struct btrfs_work *work = container_of(normal_work, struct btrfs_work,
313 struct __btrfs_workqueue *wq;
318 * We should not touch things inside work in the following cases:
319 * 1) after work->func() if it has no ordered_free
320 * Since the struct is freed in work->func().
321 * 2) after setting WORK_DONE_BIT
322 * The work may be freed in other threads almost instantly.
323 * So we save the needed things here.
325 if (work->ordered_func)
328 /* Safe for tracepoints in case work gets freed by the callback */
331 trace_btrfs_work_sched(work);
332 thresh_exec_hook(wq);
336 * Ensures all memory accesses done in the work function are
337 * ordered before setting the WORK_DONE_BIT. Ensuring the thread
338 * which is going to executed the ordered work sees them.
339 * Pairs with the smp_rmb in run_ordered_work.
341 smp_mb__before_atomic();
342 set_bit(WORK_DONE_BIT, &work->flags);
343 run_ordered_work(wq, work);
346 trace_btrfs_all_work_done(wq->fs_info, wtag);
349 void btrfs_init_work(struct btrfs_work *work, btrfs_func_t func,
350 btrfs_func_t ordered_func, btrfs_func_t ordered_free)
353 work->ordered_func = ordered_func;
354 work->ordered_free = ordered_free;
355 INIT_WORK(&work->normal_work, btrfs_work_helper);
356 INIT_LIST_HEAD(&work->ordered_list);
360 static inline void __btrfs_queue_work(struct __btrfs_workqueue *wq,
361 struct btrfs_work *work)
366 thresh_queue_hook(wq);
367 if (work->ordered_func) {
368 spin_lock_irqsave(&wq->list_lock, flags);
369 list_add_tail(&work->ordered_list, &wq->ordered_list);
370 spin_unlock_irqrestore(&wq->list_lock, flags);
372 trace_btrfs_work_queued(work);
373 queue_work(wq->normal_wq, &work->normal_work);
376 void btrfs_queue_work(struct btrfs_workqueue *wq,
377 struct btrfs_work *work)
379 struct __btrfs_workqueue *dest_wq;
381 if (test_bit(WORK_HIGH_PRIO_BIT, &work->flags) && wq->high)
384 dest_wq = wq->normal;
385 __btrfs_queue_work(dest_wq, work);
389 __btrfs_destroy_workqueue(struct __btrfs_workqueue *wq)
391 destroy_workqueue(wq->normal_wq);
392 trace_btrfs_workqueue_destroy(wq);
396 void btrfs_destroy_workqueue(struct btrfs_workqueue *wq)
401 __btrfs_destroy_workqueue(wq->high);
402 __btrfs_destroy_workqueue(wq->normal);
406 void btrfs_workqueue_set_max(struct btrfs_workqueue *wq, int limit_active)
410 wq->normal->limit_active = limit_active;
412 wq->high->limit_active = limit_active;
415 void btrfs_set_work_high_priority(struct btrfs_work *work)
417 set_bit(WORK_HIGH_PRIO_BIT, &work->flags);
420 void btrfs_flush_workqueue(struct btrfs_workqueue *wq)
423 flush_workqueue(wq->high->normal_wq);
425 flush_workqueue(wq->normal->normal_wq);