xref: /openbmc/linux/fs/btrfs/async-thread.c (revision f9834f18)
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 "async-thread.h"
13 #include "ctree.h"
14 
15 enum {
16 	WORK_DONE_BIT,
17 	WORK_ORDER_DONE_BIT,
18 	WORK_HIGH_PRIO_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_workqueue {
52 	struct __btrfs_workqueue *normal;
53 	struct __btrfs_workqueue *high;
54 };
55 
56 struct btrfs_fs_info * __pure btrfs_workqueue_owner(const struct __btrfs_workqueue *wq)
57 {
58 	return wq->fs_info;
59 }
60 
61 struct btrfs_fs_info * __pure btrfs_work_owner(const struct btrfs_work *work)
62 {
63 	return work->wq->fs_info;
64 }
65 
66 bool btrfs_workqueue_normal_congested(const struct btrfs_workqueue *wq)
67 {
68 	/*
69 	 * We could compare wq->normal->pending with num_online_cpus()
70 	 * to support "thresh == NO_THRESHOLD" case, but it requires
71 	 * moving up atomic_inc/dec in thresh_queue/exec_hook. Let's
72 	 * postpone it until someone needs the support of that case.
73 	 */
74 	if (wq->normal->thresh == NO_THRESHOLD)
75 		return false;
76 
77 	return atomic_read(&wq->normal->pending) > wq->normal->thresh * 2;
78 }
79 
80 static struct __btrfs_workqueue *
81 __btrfs_alloc_workqueue(struct btrfs_fs_info *fs_info, const char *name,
82 			unsigned int flags, int limit_active, int thresh)
83 {
84 	struct __btrfs_workqueue *ret = kzalloc(sizeof(*ret), GFP_KERNEL);
85 
86 	if (!ret)
87 		return NULL;
88 
89 	ret->fs_info = fs_info;
90 	ret->limit_active = limit_active;
91 	atomic_set(&ret->pending, 0);
92 	if (thresh == 0)
93 		thresh = DFT_THRESHOLD;
94 	/* For low threshold, disabling threshold is a better choice */
95 	if (thresh < DFT_THRESHOLD) {
96 		ret->current_active = limit_active;
97 		ret->thresh = NO_THRESHOLD;
98 	} else {
99 		/*
100 		 * For threshold-able wq, let its concurrency grow on demand.
101 		 * Use minimal max_active at alloc time to reduce resource
102 		 * usage.
103 		 */
104 		ret->current_active = 1;
105 		ret->thresh = thresh;
106 	}
107 
108 	if (flags & WQ_HIGHPRI)
109 		ret->normal_wq = alloc_workqueue("btrfs-%s-high", flags,
110 						 ret->current_active, name);
111 	else
112 		ret->normal_wq = alloc_workqueue("btrfs-%s", flags,
113 						 ret->current_active, name);
114 	if (!ret->normal_wq) {
115 		kfree(ret);
116 		return NULL;
117 	}
118 
119 	INIT_LIST_HEAD(&ret->ordered_list);
120 	spin_lock_init(&ret->list_lock);
121 	spin_lock_init(&ret->thres_lock);
122 	trace_btrfs_workqueue_alloc(ret, name, flags & WQ_HIGHPRI);
123 	return ret;
124 }
125 
126 static inline void
127 __btrfs_destroy_workqueue(struct __btrfs_workqueue *wq);
128 
129 struct btrfs_workqueue *btrfs_alloc_workqueue(struct btrfs_fs_info *fs_info,
130 					      const char *name,
131 					      unsigned int flags,
132 					      int limit_active,
133 					      int thresh)
134 {
135 	struct btrfs_workqueue *ret = kzalloc(sizeof(*ret), GFP_KERNEL);
136 
137 	if (!ret)
138 		return NULL;
139 
140 	ret->normal = __btrfs_alloc_workqueue(fs_info, name,
141 					      flags & ~WQ_HIGHPRI,
142 					      limit_active, thresh);
143 	if (!ret->normal) {
144 		kfree(ret);
145 		return NULL;
146 	}
147 
148 	if (flags & WQ_HIGHPRI) {
149 		ret->high = __btrfs_alloc_workqueue(fs_info, name, flags,
150 						    limit_active, thresh);
151 		if (!ret->high) {
152 			__btrfs_destroy_workqueue(ret->normal);
153 			kfree(ret);
154 			return NULL;
155 		}
156 	}
157 	return ret;
158 }
159 
160 /*
161  * Hook for threshold which will be called in btrfs_queue_work.
162  * This hook WILL be called in IRQ handler context,
163  * so workqueue_set_max_active MUST NOT be called in this hook
164  */
165 static inline void thresh_queue_hook(struct __btrfs_workqueue *wq)
166 {
167 	if (wq->thresh == NO_THRESHOLD)
168 		return;
169 	atomic_inc(&wq->pending);
170 }
171 
172 /*
173  * Hook for threshold which will be called before executing the work,
174  * This hook is called in kthread content.
175  * So workqueue_set_max_active is called here.
176  */
177 static inline void thresh_exec_hook(struct __btrfs_workqueue *wq)
178 {
179 	int new_current_active;
180 	long pending;
181 	int need_change = 0;
182 
183 	if (wq->thresh == NO_THRESHOLD)
184 		return;
185 
186 	atomic_dec(&wq->pending);
187 	spin_lock(&wq->thres_lock);
188 	/*
189 	 * Use wq->count to limit the calling frequency of
190 	 * workqueue_set_max_active.
191 	 */
192 	wq->count++;
193 	wq->count %= (wq->thresh / 4);
194 	if (!wq->count)
195 		goto  out;
196 	new_current_active = wq->current_active;
197 
198 	/*
199 	 * pending may be changed later, but it's OK since we really
200 	 * don't need it so accurate to calculate new_max_active.
201 	 */
202 	pending = atomic_read(&wq->pending);
203 	if (pending > wq->thresh)
204 		new_current_active++;
205 	if (pending < wq->thresh / 2)
206 		new_current_active--;
207 	new_current_active = clamp_val(new_current_active, 1, wq->limit_active);
208 	if (new_current_active != wq->current_active)  {
209 		need_change = 1;
210 		wq->current_active = new_current_active;
211 	}
212 out:
213 	spin_unlock(&wq->thres_lock);
214 
215 	if (need_change) {
216 		workqueue_set_max_active(wq->normal_wq, wq->current_active);
217 	}
218 }
219 
220 static void run_ordered_work(struct __btrfs_workqueue *wq,
221 			     struct btrfs_work *self)
222 {
223 	struct list_head *list = &wq->ordered_list;
224 	struct btrfs_work *work;
225 	spinlock_t *lock = &wq->list_lock;
226 	unsigned long flags;
227 	bool free_self = false;
228 
229 	while (1) {
230 		spin_lock_irqsave(lock, flags);
231 		if (list_empty(list))
232 			break;
233 		work = list_entry(list->next, struct btrfs_work,
234 				  ordered_list);
235 		if (!test_bit(WORK_DONE_BIT, &work->flags))
236 			break;
237 
238 		/*
239 		 * we are going to call the ordered done function, but
240 		 * we leave the work item on the list as a barrier so
241 		 * that later work items that are done don't have their
242 		 * functions called before this one returns
243 		 */
244 		if (test_and_set_bit(WORK_ORDER_DONE_BIT, &work->flags))
245 			break;
246 		trace_btrfs_ordered_sched(work);
247 		spin_unlock_irqrestore(lock, flags);
248 		work->ordered_func(work);
249 
250 		/* now take the lock again and drop our item from the list */
251 		spin_lock_irqsave(lock, flags);
252 		list_del(&work->ordered_list);
253 		spin_unlock_irqrestore(lock, flags);
254 
255 		if (work == self) {
256 			/*
257 			 * This is the work item that the worker is currently
258 			 * executing.
259 			 *
260 			 * The kernel workqueue code guarantees non-reentrancy
261 			 * of work items. I.e., if a work item with the same
262 			 * address and work function is queued twice, the second
263 			 * execution is blocked until the first one finishes. A
264 			 * work item may be freed and recycled with the same
265 			 * work function; the workqueue code assumes that the
266 			 * original work item cannot depend on the recycled work
267 			 * item in that case (see find_worker_executing_work()).
268 			 *
269 			 * Note that different types of Btrfs work can depend on
270 			 * each other, and one type of work on one Btrfs
271 			 * filesystem may even depend on the same type of work
272 			 * on another Btrfs filesystem via, e.g., a loop device.
273 			 * Therefore, we must not allow the current work item to
274 			 * be recycled until we are really done, otherwise we
275 			 * break the above assumption and can deadlock.
276 			 */
277 			free_self = true;
278 		} else {
279 			/*
280 			 * We don't want to call the ordered free functions with
281 			 * the lock held.
282 			 */
283 			work->ordered_free(work);
284 			/* NB: work must not be dereferenced past this point. */
285 			trace_btrfs_all_work_done(wq->fs_info, work);
286 		}
287 	}
288 	spin_unlock_irqrestore(lock, flags);
289 
290 	if (free_self) {
291 		self->ordered_free(self);
292 		/* NB: self must not be dereferenced past this point. */
293 		trace_btrfs_all_work_done(wq->fs_info, self);
294 	}
295 }
296 
297 static void btrfs_work_helper(struct work_struct *normal_work)
298 {
299 	struct btrfs_work *work = container_of(normal_work, struct btrfs_work,
300 					       normal_work);
301 	struct __btrfs_workqueue *wq;
302 	int need_order = 0;
303 
304 	/*
305 	 * We should not touch things inside work in the following cases:
306 	 * 1) after work->func() if it has no ordered_free
307 	 *    Since the struct is freed in work->func().
308 	 * 2) after setting WORK_DONE_BIT
309 	 *    The work may be freed in other threads almost instantly.
310 	 * So we save the needed things here.
311 	 */
312 	if (work->ordered_func)
313 		need_order = 1;
314 	wq = work->wq;
315 
316 	trace_btrfs_work_sched(work);
317 	thresh_exec_hook(wq);
318 	work->func(work);
319 	if (need_order) {
320 		set_bit(WORK_DONE_BIT, &work->flags);
321 		run_ordered_work(wq, work);
322 	} else {
323 		/* NB: work must not be dereferenced past this point. */
324 		trace_btrfs_all_work_done(wq->fs_info, work);
325 	}
326 }
327 
328 void btrfs_init_work(struct btrfs_work *work, btrfs_func_t func,
329 		     btrfs_func_t ordered_func, btrfs_func_t ordered_free)
330 {
331 	work->func = func;
332 	work->ordered_func = ordered_func;
333 	work->ordered_free = ordered_free;
334 	INIT_WORK(&work->normal_work, btrfs_work_helper);
335 	INIT_LIST_HEAD(&work->ordered_list);
336 	work->flags = 0;
337 }
338 
339 static inline void __btrfs_queue_work(struct __btrfs_workqueue *wq,
340 				      struct btrfs_work *work)
341 {
342 	unsigned long flags;
343 
344 	work->wq = wq;
345 	thresh_queue_hook(wq);
346 	if (work->ordered_func) {
347 		spin_lock_irqsave(&wq->list_lock, flags);
348 		list_add_tail(&work->ordered_list, &wq->ordered_list);
349 		spin_unlock_irqrestore(&wq->list_lock, flags);
350 	}
351 	trace_btrfs_work_queued(work);
352 	queue_work(wq->normal_wq, &work->normal_work);
353 }
354 
355 void btrfs_queue_work(struct btrfs_workqueue *wq,
356 		      struct btrfs_work *work)
357 {
358 	struct __btrfs_workqueue *dest_wq;
359 
360 	if (test_bit(WORK_HIGH_PRIO_BIT, &work->flags) && wq->high)
361 		dest_wq = wq->high;
362 	else
363 		dest_wq = wq->normal;
364 	__btrfs_queue_work(dest_wq, work);
365 }
366 
367 static inline void
368 __btrfs_destroy_workqueue(struct __btrfs_workqueue *wq)
369 {
370 	destroy_workqueue(wq->normal_wq);
371 	trace_btrfs_workqueue_destroy(wq);
372 	kfree(wq);
373 }
374 
375 void btrfs_destroy_workqueue(struct btrfs_workqueue *wq)
376 {
377 	if (!wq)
378 		return;
379 	if (wq->high)
380 		__btrfs_destroy_workqueue(wq->high);
381 	__btrfs_destroy_workqueue(wq->normal);
382 	kfree(wq);
383 }
384 
385 void btrfs_workqueue_set_max(struct btrfs_workqueue *wq, int limit_active)
386 {
387 	if (!wq)
388 		return;
389 	wq->normal->limit_active = limit_active;
390 	if (wq->high)
391 		wq->high->limit_active = limit_active;
392 }
393 
394 void btrfs_set_work_high_priority(struct btrfs_work *work)
395 {
396 	set_bit(WORK_HIGH_PRIO_BIT, &work->flags);
397 }
398