xref: /openbmc/linux/fs/btrfs/async-thread.c (revision 8306d6f3)
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 };
19 
20 #define NO_THRESHOLD (-1)
21 #define DFT_THRESHOLD (32)
22 
23 struct btrfs_workqueue {
24 	struct workqueue_struct *normal_wq;
25 
26 	/* File system this workqueue services */
27 	struct btrfs_fs_info *fs_info;
28 
29 	/* List head pointing to ordered work list */
30 	struct list_head ordered_list;
31 
32 	/* Spinlock for ordered_list */
33 	spinlock_t list_lock;
34 
35 	/* Thresholding related variants */
36 	atomic_t pending;
37 
38 	/* Up limit of concurrency workers */
39 	int limit_active;
40 
41 	/* Current number of concurrency workers */
42 	int current_active;
43 
44 	/* Threshold to change current_active */
45 	int thresh;
46 	unsigned int count;
47 	spinlock_t thres_lock;
48 };
49 
50 struct btrfs_fs_info * __pure btrfs_workqueue_owner(const struct btrfs_workqueue *wq)
51 {
52 	return wq->fs_info;
53 }
54 
55 struct btrfs_fs_info * __pure btrfs_work_owner(const struct btrfs_work *work)
56 {
57 	return work->wq->fs_info;
58 }
59 
60 bool btrfs_workqueue_normal_congested(const struct btrfs_workqueue *wq)
61 {
62 	/*
63 	 * We could compare wq->pending with num_online_cpus()
64 	 * to support "thresh == NO_THRESHOLD" case, but it requires
65 	 * moving up atomic_inc/dec in thresh_queue/exec_hook. Let's
66 	 * postpone it until someone needs the support of that case.
67 	 */
68 	if (wq->thresh == NO_THRESHOLD)
69 		return false;
70 
71 	return atomic_read(&wq->pending) > wq->thresh * 2;
72 }
73 
74 static void btrfs_init_workqueue(struct btrfs_workqueue *wq,
75 				 struct btrfs_fs_info *fs_info)
76 {
77 	wq->fs_info = fs_info;
78 	atomic_set(&wq->pending, 0);
79 	INIT_LIST_HEAD(&wq->ordered_list);
80 	spin_lock_init(&wq->list_lock);
81 	spin_lock_init(&wq->thres_lock);
82 }
83 
84 struct btrfs_workqueue *btrfs_alloc_workqueue(struct btrfs_fs_info *fs_info,
85 					      const char *name, unsigned int flags,
86 					      int limit_active, int thresh)
87 {
88 	struct btrfs_workqueue *ret = kzalloc(sizeof(*ret), GFP_KERNEL);
89 
90 	if (!ret)
91 		return NULL;
92 
93 	btrfs_init_workqueue(ret, fs_info);
94 
95 	ret->limit_active = limit_active;
96 	if (thresh == 0)
97 		thresh = DFT_THRESHOLD;
98 	/* For low threshold, disabling threshold is a better choice */
99 	if (thresh < DFT_THRESHOLD) {
100 		ret->current_active = limit_active;
101 		ret->thresh = NO_THRESHOLD;
102 	} else {
103 		/*
104 		 * For threshold-able wq, let its concurrency grow on demand.
105 		 * Use minimal max_active at alloc time to reduce resource
106 		 * usage.
107 		 */
108 		ret->current_active = 1;
109 		ret->thresh = thresh;
110 	}
111 
112 	ret->normal_wq = alloc_workqueue("btrfs-%s", flags, ret->current_active,
113 					 name);
114 	if (!ret->normal_wq) {
115 		kfree(ret);
116 		return NULL;
117 	}
118 
119 	trace_btrfs_workqueue_alloc(ret, name);
120 	return ret;
121 }
122 
123 struct btrfs_workqueue *btrfs_alloc_ordered_workqueue(
124 				struct btrfs_fs_info *fs_info, const char *name,
125 				unsigned int flags)
126 {
127 	struct btrfs_workqueue *ret;
128 
129 	ret = kzalloc(sizeof(*ret), GFP_KERNEL);
130 	if (!ret)
131 		return NULL;
132 
133 	btrfs_init_workqueue(ret, fs_info);
134 
135 	/* Ordered workqueues don't allow @max_active adjustments. */
136 	ret->limit_active = 1;
137 	ret->current_active = 1;
138 	ret->thresh = NO_THRESHOLD;
139 
140 	ret->normal_wq = alloc_ordered_workqueue("btrfs-%s", flags, name);
141 	if (!ret->normal_wq) {
142 		kfree(ret);
143 		return NULL;
144 	}
145 
146 	trace_btrfs_workqueue_alloc(ret, name);
147 	return ret;
148 }
149 
150 /*
151  * Hook for threshold which will be called in btrfs_queue_work.
152  * This hook WILL be called in IRQ handler context,
153  * so workqueue_set_max_active MUST NOT be called in this hook
154  */
155 static inline void thresh_queue_hook(struct btrfs_workqueue *wq)
156 {
157 	if (wq->thresh == NO_THRESHOLD)
158 		return;
159 	atomic_inc(&wq->pending);
160 }
161 
162 /*
163  * Hook for threshold which will be called before executing the work,
164  * This hook is called in kthread content.
165  * So workqueue_set_max_active is called here.
166  */
167 static inline void thresh_exec_hook(struct btrfs_workqueue *wq)
168 {
169 	int new_current_active;
170 	long pending;
171 	int need_change = 0;
172 
173 	if (wq->thresh == NO_THRESHOLD)
174 		return;
175 
176 	atomic_dec(&wq->pending);
177 	spin_lock(&wq->thres_lock);
178 	/*
179 	 * Use wq->count to limit the calling frequency of
180 	 * workqueue_set_max_active.
181 	 */
182 	wq->count++;
183 	wq->count %= (wq->thresh / 4);
184 	if (!wq->count)
185 		goto  out;
186 	new_current_active = wq->current_active;
187 
188 	/*
189 	 * pending may be changed later, but it's OK since we really
190 	 * don't need it so accurate to calculate new_max_active.
191 	 */
192 	pending = atomic_read(&wq->pending);
193 	if (pending > wq->thresh)
194 		new_current_active++;
195 	if (pending < wq->thresh / 2)
196 		new_current_active--;
197 	new_current_active = clamp_val(new_current_active, 1, wq->limit_active);
198 	if (new_current_active != wq->current_active)  {
199 		need_change = 1;
200 		wq->current_active = new_current_active;
201 	}
202 out:
203 	spin_unlock(&wq->thres_lock);
204 
205 	if (need_change) {
206 		workqueue_set_max_active(wq->normal_wq, wq->current_active);
207 	}
208 }
209 
210 static void run_ordered_work(struct btrfs_workqueue *wq,
211 			     struct btrfs_work *self)
212 {
213 	struct list_head *list = &wq->ordered_list;
214 	struct btrfs_work *work;
215 	spinlock_t *lock = &wq->list_lock;
216 	unsigned long flags;
217 	bool free_self = false;
218 
219 	while (1) {
220 		spin_lock_irqsave(lock, flags);
221 		if (list_empty(list))
222 			break;
223 		work = list_entry(list->next, struct btrfs_work,
224 				  ordered_list);
225 		if (!test_bit(WORK_DONE_BIT, &work->flags))
226 			break;
227 		/*
228 		 * Orders all subsequent loads after reading WORK_DONE_BIT,
229 		 * paired with the smp_mb__before_atomic in btrfs_work_helper
230 		 * this guarantees that the ordered function will see all
231 		 * updates from ordinary work function.
232 		 */
233 		smp_rmb();
234 
235 		/*
236 		 * we are going to call the ordered done function, but
237 		 * we leave the work item on the list as a barrier so
238 		 * that later work items that are done don't have their
239 		 * functions called before this one returns
240 		 */
241 		if (test_and_set_bit(WORK_ORDER_DONE_BIT, &work->flags))
242 			break;
243 		trace_btrfs_ordered_sched(work);
244 		spin_unlock_irqrestore(lock, flags);
245 		work->ordered_func(work);
246 
247 		/* now take the lock again and drop our item from the list */
248 		spin_lock_irqsave(lock, flags);
249 		list_del(&work->ordered_list);
250 		spin_unlock_irqrestore(lock, flags);
251 
252 		if (work == self) {
253 			/*
254 			 * This is the work item that the worker is currently
255 			 * executing.
256 			 *
257 			 * The kernel workqueue code guarantees non-reentrancy
258 			 * of work items. I.e., if a work item with the same
259 			 * address and work function is queued twice, the second
260 			 * execution is blocked until the first one finishes. A
261 			 * work item may be freed and recycled with the same
262 			 * work function; the workqueue code assumes that the
263 			 * original work item cannot depend on the recycled work
264 			 * item in that case (see find_worker_executing_work()).
265 			 *
266 			 * Note that different types of Btrfs work can depend on
267 			 * each other, and one type of work on one Btrfs
268 			 * filesystem may even depend on the same type of work
269 			 * on another Btrfs filesystem via, e.g., a loop device.
270 			 * Therefore, we must not allow the current work item to
271 			 * be recycled until we are really done, otherwise we
272 			 * break the above assumption and can deadlock.
273 			 */
274 			free_self = true;
275 		} else {
276 			/*
277 			 * We don't want to call the ordered free functions with
278 			 * the lock held.
279 			 */
280 			work->ordered_free(work);
281 			/* NB: work must not be dereferenced past this point. */
282 			trace_btrfs_all_work_done(wq->fs_info, work);
283 		}
284 	}
285 	spin_unlock_irqrestore(lock, flags);
286 
287 	if (free_self) {
288 		self->ordered_free(self);
289 		/* NB: self must not be dereferenced past this point. */
290 		trace_btrfs_all_work_done(wq->fs_info, self);
291 	}
292 }
293 
294 static void btrfs_work_helper(struct work_struct *normal_work)
295 {
296 	struct btrfs_work *work = container_of(normal_work, struct btrfs_work,
297 					       normal_work);
298 	struct btrfs_workqueue *wq = work->wq;
299 	int need_order = 0;
300 
301 	/*
302 	 * We should not touch things inside work in the following cases:
303 	 * 1) after work->func() if it has no ordered_free
304 	 *    Since the struct is freed in work->func().
305 	 * 2) after setting WORK_DONE_BIT
306 	 *    The work may be freed in other threads almost instantly.
307 	 * So we save the needed things here.
308 	 */
309 	if (work->ordered_func)
310 		need_order = 1;
311 
312 	trace_btrfs_work_sched(work);
313 	thresh_exec_hook(wq);
314 	work->func(work);
315 	if (need_order) {
316 		/*
317 		 * Ensures all memory accesses done in the work function are
318 		 * ordered before setting the WORK_DONE_BIT. Ensuring the thread
319 		 * which is going to executed the ordered work sees them.
320 		 * Pairs with the smp_rmb in run_ordered_work.
321 		 */
322 		smp_mb__before_atomic();
323 		set_bit(WORK_DONE_BIT, &work->flags);
324 		run_ordered_work(wq, work);
325 	} else {
326 		/* NB: work must not be dereferenced past this point. */
327 		trace_btrfs_all_work_done(wq->fs_info, work);
328 	}
329 }
330 
331 void btrfs_init_work(struct btrfs_work *work, btrfs_func_t func,
332 		     btrfs_func_t ordered_func, btrfs_func_t ordered_free)
333 {
334 	work->func = func;
335 	work->ordered_func = ordered_func;
336 	work->ordered_free = ordered_free;
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