xref: /linux/fs/btrfs/async-thread.c (revision b17ef04bf3a4346d66404454d6a646343ddc9749)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Copyright (C) 2007 Oracle.  All rights reserved.
4  * Copyright (C) 2014 Fujitsu.  All rights reserved.
5  */
6 
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 <trace/events/btrfs.h>
13 #include "async-thread.h"
14 #include "ctree.h"
15 
16 enum {
17 	WORK_DONE_BIT,
18 	WORK_ORDER_DONE_BIT,
19 };
20 
21 #define NO_THRESHOLD (-1)
22 #define DFT_THRESHOLD (32)
23 
24 struct btrfs_workqueue {
25 	struct workqueue_struct *normal_wq;
26 
27 	/* File system this workqueue services */
28 	struct btrfs_fs_info *fs_info;
29 
30 	/* List head pointing to ordered work list */
31 	struct list_head ordered_list;
32 
33 	/* Spinlock for ordered_list */
34 	spinlock_t list_lock;
35 
36 	/* Thresholding related variants */
37 	atomic_t pending;
38 
39 	/* Up limit of concurrency workers */
40 	int limit_active;
41 
42 	/* Current number of concurrency workers */
43 	int current_active;
44 
45 	/* Threshold to change current_active */
46 	int thresh;
47 	unsigned int count;
48 	spinlock_t thres_lock;
49 };
50 
51 struct btrfs_fs_info * __pure btrfs_workqueue_owner(const struct btrfs_workqueue *wq)
52 {
53 	return wq->fs_info;
54 }
55 
56 struct btrfs_fs_info * __pure btrfs_work_owner(const struct btrfs_work *work)
57 {
58 	return work->wq->fs_info;
59 }
60 
61 bool btrfs_workqueue_normal_congested(const struct btrfs_workqueue *wq)
62 {
63 	/*
64 	 * We could compare wq->pending with num_online_cpus()
65 	 * to support "thresh == NO_THRESHOLD" case, but it requires
66 	 * moving up atomic_inc/dec in thresh_queue/exec_hook. Let's
67 	 * postpone it until someone needs the support of that case.
68 	 */
69 	if (wq->thresh == NO_THRESHOLD)
70 		return false;
71 
72 	return atomic_read(&wq->pending) > wq->thresh * 2;
73 }
74 
75 static void btrfs_init_workqueue(struct btrfs_workqueue *wq,
76 				 struct btrfs_fs_info *fs_info)
77 {
78 	wq->fs_info = fs_info;
79 	atomic_set(&wq->pending, 0);
80 	INIT_LIST_HEAD(&wq->ordered_list);
81 	spin_lock_init(&wq->list_lock);
82 	spin_lock_init(&wq->thres_lock);
83 }
84 
85 struct btrfs_workqueue *btrfs_alloc_workqueue(struct btrfs_fs_info *fs_info,
86 					      const char *name, unsigned int flags,
87 					      int limit_active, int thresh)
88 {
89 	struct btrfs_workqueue *ret = kzalloc(sizeof(*ret), GFP_KERNEL);
90 
91 	if (!ret)
92 		return NULL;
93 
94 	btrfs_init_workqueue(ret, fs_info);
95 
96 	ret->limit_active = limit_active;
97 	if (thresh == 0)
98 		thresh = DFT_THRESHOLD;
99 	/* For low threshold, disabling threshold is a better choice */
100 	if (thresh < DFT_THRESHOLD) {
101 		ret->current_active = limit_active;
102 		ret->thresh = NO_THRESHOLD;
103 	} else {
104 		/*
105 		 * For threshold-able wq, let its concurrency grow on demand.
106 		 * Use minimal max_active at alloc time to reduce resource
107 		 * usage.
108 		 */
109 		ret->current_active = 1;
110 		ret->thresh = thresh;
111 	}
112 
113 	ret->normal_wq = alloc_workqueue("btrfs-%s", flags, ret->current_active,
114 					 name);
115 	if (!ret->normal_wq) {
116 		kfree(ret);
117 		return NULL;
118 	}
119 
120 	trace_btrfs_workqueue_alloc(ret, name);
121 	return ret;
122 }
123 
124 struct btrfs_workqueue *btrfs_alloc_ordered_workqueue(
125 				struct btrfs_fs_info *fs_info, const char *name,
126 				unsigned int flags)
127 {
128 	struct btrfs_workqueue *ret;
129 
130 	ret = kzalloc(sizeof(*ret), GFP_KERNEL);
131 	if (!ret)
132 		return NULL;
133 
134 	btrfs_init_workqueue(ret, fs_info);
135 
136 	/* Ordered workqueues don't allow @max_active adjustments. */
137 	ret->limit_active = 1;
138 	ret->current_active = 1;
139 	ret->thresh = NO_THRESHOLD;
140 
141 	ret->normal_wq = alloc_ordered_workqueue("btrfs-%s", flags, name);
142 	if (!ret->normal_wq) {
143 		kfree(ret);
144 		return NULL;
145 	}
146 
147 	trace_btrfs_workqueue_alloc(ret, name);
148 	return ret;
149 }
150 
151 /*
152  * Hook for threshold which will be called in btrfs_queue_work.
153  * This hook WILL be called in IRQ handler context,
154  * so workqueue_set_max_active MUST NOT be called in this hook
155  */
156 static inline void thresh_queue_hook(struct btrfs_workqueue *wq)
157 {
158 	if (wq->thresh == NO_THRESHOLD)
159 		return;
160 	atomic_inc(&wq->pending);
161 }
162 
163 /*
164  * Hook for threshold which will be called before executing the work,
165  * This hook is called in kthread content.
166  * So workqueue_set_max_active is called here.
167  */
168 static inline void thresh_exec_hook(struct btrfs_workqueue *wq)
169 {
170 	int new_current_active;
171 	long pending;
172 	int need_change = 0;
173 
174 	if (wq->thresh == NO_THRESHOLD)
175 		return;
176 
177 	atomic_dec(&wq->pending);
178 	spin_lock(&wq->thres_lock);
179 	/*
180 	 * Use wq->count to limit the calling frequency of
181 	 * workqueue_set_max_active.
182 	 */
183 	wq->count++;
184 	wq->count %= (wq->thresh / 4);
185 	if (!wq->count)
186 		goto  out;
187 	new_current_active = wq->current_active;
188 
189 	/*
190 	 * pending may be changed later, but it's OK since we really
191 	 * don't need it so accurate to calculate new_max_active.
192 	 */
193 	pending = atomic_read(&wq->pending);
194 	if (pending > wq->thresh)
195 		new_current_active++;
196 	if (pending < wq->thresh / 2)
197 		new_current_active--;
198 	new_current_active = clamp_val(new_current_active, 1, wq->limit_active);
199 	if (new_current_active != wq->current_active)  {
200 		need_change = 1;
201 		wq->current_active = new_current_active;
202 	}
203 out:
204 	spin_unlock(&wq->thres_lock);
205 
206 	if (need_change) {
207 		workqueue_set_max_active(wq->normal_wq, wq->current_active);
208 	}
209 }
210 
211 static void run_ordered_work(struct btrfs_workqueue *wq,
212 			     struct btrfs_work *self)
213 {
214 	struct list_head *list = &wq->ordered_list;
215 	struct btrfs_work *work;
216 	spinlock_t *lock = &wq->list_lock;
217 	unsigned long flags;
218 	bool free_self = false;
219 
220 	while (1) {
221 		spin_lock_irqsave(lock, flags);
222 		if (list_empty(list))
223 			break;
224 		work = list_entry(list->next, struct btrfs_work,
225 				  ordered_list);
226 		if (!test_bit(WORK_DONE_BIT, &work->flags))
227 			break;
228 		/*
229 		 * Orders all subsequent loads after reading WORK_DONE_BIT,
230 		 * paired with the smp_mb__before_atomic in btrfs_work_helper
231 		 * this guarantees that the ordered function will see all
232 		 * updates from ordinary work function.
233 		 */
234 		smp_rmb();
235 
236 		/*
237 		 * we are going to call the ordered done function, but
238 		 * we leave the work item on the list as a barrier so
239 		 * that later work items that are done don't have their
240 		 * functions called before this one returns
241 		 */
242 		if (test_and_set_bit(WORK_ORDER_DONE_BIT, &work->flags))
243 			break;
244 		trace_btrfs_ordered_sched(work);
245 		spin_unlock_irqrestore(lock, flags);
246 		work->ordered_func(work, false);
247 
248 		/* now take the lock again and drop our item from the list */
249 		spin_lock_irqsave(lock, flags);
250 		list_del(&work->ordered_list);
251 		spin_unlock_irqrestore(lock, flags);
252 
253 		if (work == self) {
254 			/*
255 			 * This is the work item that the worker is currently
256 			 * executing.
257 			 *
258 			 * The kernel workqueue code guarantees non-reentrancy
259 			 * of work items. I.e., if a work item with the same
260 			 * address and work function is queued twice, the second
261 			 * execution is blocked until the first one finishes. A
262 			 * work item may be freed and recycled with the same
263 			 * work function; the workqueue code assumes that the
264 			 * original work item cannot depend on the recycled work
265 			 * item in that case (see find_worker_executing_work()).
266 			 *
267 			 * Note that different types of Btrfs work can depend on
268 			 * each other, and one type of work on one Btrfs
269 			 * filesystem may even depend on the same type of work
270 			 * on another Btrfs filesystem via, e.g., a loop device.
271 			 * Therefore, we must not allow the current work item to
272 			 * be recycled until we are really done, otherwise we
273 			 * break the above assumption and can deadlock.
274 			 */
275 			free_self = true;
276 		} else {
277 			/*
278 			 * We don't want to call the ordered free functions with
279 			 * the lock held.
280 			 */
281 			work->ordered_func(work, true);
282 			/* NB: work must not be dereferenced past this point. */
283 			trace_btrfs_all_work_done(wq->fs_info, work);
284 		}
285 	}
286 	spin_unlock_irqrestore(lock, flags);
287 
288 	if (free_self) {
289 		self->ordered_func(self, true);
290 		/* NB: self must not be dereferenced past this point. */
291 		trace_btrfs_all_work_done(wq->fs_info, self);
292 	}
293 }
294 
295 static void btrfs_work_helper(struct work_struct *normal_work)
296 {
297 	struct btrfs_work *work = container_of(normal_work, struct btrfs_work,
298 					       normal_work);
299 	struct btrfs_workqueue *wq = work->wq;
300 	int need_order = 0;
301 
302 	/*
303 	 * We should not touch things inside work in the following cases:
304 	 * 1) after work->func() if it has no ordered_func(..., true) to free
305 	 *    Since the struct is freed in work->func().
306 	 * 2) after setting WORK_DONE_BIT
307 	 *    The work may be freed in other threads almost instantly.
308 	 * So we save the needed things here.
309 	 */
310 	if (work->ordered_func)
311 		need_order = 1;
312 
313 	trace_btrfs_work_sched(work);
314 	thresh_exec_hook(wq);
315 	work->func(work);
316 	if (need_order) {
317 		/*
318 		 * Ensures all memory accesses done in the work function are
319 		 * ordered before setting the WORK_DONE_BIT. Ensuring the thread
320 		 * which is going to executed the ordered work sees them.
321 		 * Pairs with the smp_rmb in run_ordered_work.
322 		 */
323 		smp_mb__before_atomic();
324 		set_bit(WORK_DONE_BIT, &work->flags);
325 		run_ordered_work(wq, work);
326 	} else {
327 		/* NB: work must not be dereferenced past this point. */
328 		trace_btrfs_all_work_done(wq->fs_info, work);
329 	}
330 }
331 
332 void btrfs_init_work(struct btrfs_work *work, btrfs_func_t func,
333 		     btrfs_ordered_func_t ordered_func)
334 {
335 	work->func = func;
336 	work->ordered_func = ordered_func;
337 	INIT_WORK(&work->normal_work, btrfs_work_helper);
338 	INIT_LIST_HEAD(&work->ordered_list);
339 	work->flags = 0;
340 }
341 
342 void btrfs_queue_work(struct btrfs_workqueue *wq, struct btrfs_work *work)
343 {
344 	unsigned long flags;
345 
346 	work->wq = wq;
347 	thresh_queue_hook(wq);
348 	if (work->ordered_func) {
349 		spin_lock_irqsave(&wq->list_lock, flags);
350 		list_add_tail(&work->ordered_list, &wq->ordered_list);
351 		spin_unlock_irqrestore(&wq->list_lock, flags);
352 	}
353 	trace_btrfs_work_queued(work);
354 	queue_work(wq->normal_wq, &work->normal_work);
355 }
356 
357 void btrfs_destroy_workqueue(struct btrfs_workqueue *wq)
358 {
359 	if (!wq)
360 		return;
361 	destroy_workqueue(wq->normal_wq);
362 	trace_btrfs_workqueue_destroy(wq);
363 	kfree(wq);
364 }
365 
366 void btrfs_workqueue_set_max(struct btrfs_workqueue *wq, int limit_active)
367 {
368 	if (wq)
369 		wq->limit_active = limit_active;
370 }
371 
372 void btrfs_flush_workqueue(struct btrfs_workqueue *wq)
373 {
374 	flush_workqueue(wq->normal_wq);
375 }
376