xref: /openbmc/linux/fs/btrfs/async-thread.c (revision 9a29f5fc)
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 struct btrfs_workqueue *btrfs_alloc_workqueue(struct btrfs_fs_info *fs_info,
75 					      const char *name, unsigned int flags,
76 					      int limit_active, int thresh)
77 {
78 	struct btrfs_workqueue *ret = kzalloc(sizeof(*ret), GFP_KERNEL);
79 
80 	if (!ret)
81 		return NULL;
82 
83 	ret->fs_info = fs_info;
84 	ret->limit_active = limit_active;
85 	atomic_set(&ret->pending, 0);
86 	if (thresh == 0)
87 		thresh = DFT_THRESHOLD;
88 	/* For low threshold, disabling threshold is a better choice */
89 	if (thresh < DFT_THRESHOLD) {
90 		ret->current_active = limit_active;
91 		ret->thresh = NO_THRESHOLD;
92 	} else {
93 		/*
94 		 * For threshold-able wq, let its concurrency grow on demand.
95 		 * Use minimal max_active at alloc time to reduce resource
96 		 * usage.
97 		 */
98 		ret->current_active = 1;
99 		ret->thresh = thresh;
100 	}
101 
102 	ret->normal_wq = alloc_workqueue("btrfs-%s", flags, ret->current_active,
103 					 name);
104 	if (!ret->normal_wq) {
105 		kfree(ret);
106 		return NULL;
107 	}
108 
109 	INIT_LIST_HEAD(&ret->ordered_list);
110 	spin_lock_init(&ret->list_lock);
111 	spin_lock_init(&ret->thres_lock);
112 	trace_btrfs_workqueue_alloc(ret, name);
113 	return ret;
114 }
115 
116 /*
117  * Hook for threshold which will be called in btrfs_queue_work.
118  * This hook WILL be called in IRQ handler context,
119  * so workqueue_set_max_active MUST NOT be called in this hook
120  */
121 static inline void thresh_queue_hook(struct btrfs_workqueue *wq)
122 {
123 	if (wq->thresh == NO_THRESHOLD)
124 		return;
125 	atomic_inc(&wq->pending);
126 }
127 
128 /*
129  * Hook for threshold which will be called before executing the work,
130  * This hook is called in kthread content.
131  * So workqueue_set_max_active is called here.
132  */
133 static inline void thresh_exec_hook(struct btrfs_workqueue *wq)
134 {
135 	int new_current_active;
136 	long pending;
137 	int need_change = 0;
138 
139 	if (wq->thresh == NO_THRESHOLD)
140 		return;
141 
142 	atomic_dec(&wq->pending);
143 	spin_lock(&wq->thres_lock);
144 	/*
145 	 * Use wq->count to limit the calling frequency of
146 	 * workqueue_set_max_active.
147 	 */
148 	wq->count++;
149 	wq->count %= (wq->thresh / 4);
150 	if (!wq->count)
151 		goto  out;
152 	new_current_active = wq->current_active;
153 
154 	/*
155 	 * pending may be changed later, but it's OK since we really
156 	 * don't need it so accurate to calculate new_max_active.
157 	 */
158 	pending = atomic_read(&wq->pending);
159 	if (pending > wq->thresh)
160 		new_current_active++;
161 	if (pending < wq->thresh / 2)
162 		new_current_active--;
163 	new_current_active = clamp_val(new_current_active, 1, wq->limit_active);
164 	if (new_current_active != wq->current_active)  {
165 		need_change = 1;
166 		wq->current_active = new_current_active;
167 	}
168 out:
169 	spin_unlock(&wq->thres_lock);
170 
171 	if (need_change) {
172 		workqueue_set_max_active(wq->normal_wq, wq->current_active);
173 	}
174 }
175 
176 static void run_ordered_work(struct btrfs_workqueue *wq,
177 			     struct btrfs_work *self)
178 {
179 	struct list_head *list = &wq->ordered_list;
180 	struct btrfs_work *work;
181 	spinlock_t *lock = &wq->list_lock;
182 	unsigned long flags;
183 	bool free_self = false;
184 
185 	while (1) {
186 		spin_lock_irqsave(lock, flags);
187 		if (list_empty(list))
188 			break;
189 		work = list_entry(list->next, struct btrfs_work,
190 				  ordered_list);
191 		if (!test_bit(WORK_DONE_BIT, &work->flags))
192 			break;
193 		/*
194 		 * Orders all subsequent loads after reading WORK_DONE_BIT,
195 		 * paired with the smp_mb__before_atomic in btrfs_work_helper
196 		 * this guarantees that the ordered function will see all
197 		 * updates from ordinary work function.
198 		 */
199 		smp_rmb();
200 
201 		/*
202 		 * we are going to call the ordered done function, but
203 		 * we leave the work item on the list as a barrier so
204 		 * that later work items that are done don't have their
205 		 * functions called before this one returns
206 		 */
207 		if (test_and_set_bit(WORK_ORDER_DONE_BIT, &work->flags))
208 			break;
209 		trace_btrfs_ordered_sched(work);
210 		spin_unlock_irqrestore(lock, flags);
211 		work->ordered_func(work);
212 
213 		/* now take the lock again and drop our item from the list */
214 		spin_lock_irqsave(lock, flags);
215 		list_del(&work->ordered_list);
216 		spin_unlock_irqrestore(lock, flags);
217 
218 		if (work == self) {
219 			/*
220 			 * This is the work item that the worker is currently
221 			 * executing.
222 			 *
223 			 * The kernel workqueue code guarantees non-reentrancy
224 			 * of work items. I.e., if a work item with the same
225 			 * address and work function is queued twice, the second
226 			 * execution is blocked until the first one finishes. A
227 			 * work item may be freed and recycled with the same
228 			 * work function; the workqueue code assumes that the
229 			 * original work item cannot depend on the recycled work
230 			 * item in that case (see find_worker_executing_work()).
231 			 *
232 			 * Note that different types of Btrfs work can depend on
233 			 * each other, and one type of work on one Btrfs
234 			 * filesystem may even depend on the same type of work
235 			 * on another Btrfs filesystem via, e.g., a loop device.
236 			 * Therefore, we must not allow the current work item to
237 			 * be recycled until we are really done, otherwise we
238 			 * break the above assumption and can deadlock.
239 			 */
240 			free_self = true;
241 		} else {
242 			/*
243 			 * We don't want to call the ordered free functions with
244 			 * the lock held.
245 			 */
246 			work->ordered_free(work);
247 			/* NB: work must not be dereferenced past this point. */
248 			trace_btrfs_all_work_done(wq->fs_info, work);
249 		}
250 	}
251 	spin_unlock_irqrestore(lock, flags);
252 
253 	if (free_self) {
254 		self->ordered_free(self);
255 		/* NB: self must not be dereferenced past this point. */
256 		trace_btrfs_all_work_done(wq->fs_info, self);
257 	}
258 }
259 
260 static void btrfs_work_helper(struct work_struct *normal_work)
261 {
262 	struct btrfs_work *work = container_of(normal_work, struct btrfs_work,
263 					       normal_work);
264 	struct btrfs_workqueue *wq = work->wq;
265 	int need_order = 0;
266 
267 	/*
268 	 * We should not touch things inside work in the following cases:
269 	 * 1) after work->func() if it has no ordered_free
270 	 *    Since the struct is freed in work->func().
271 	 * 2) after setting WORK_DONE_BIT
272 	 *    The work may be freed in other threads almost instantly.
273 	 * So we save the needed things here.
274 	 */
275 	if (work->ordered_func)
276 		need_order = 1;
277 
278 	trace_btrfs_work_sched(work);
279 	thresh_exec_hook(wq);
280 	work->func(work);
281 	if (need_order) {
282 		/*
283 		 * Ensures all memory accesses done in the work function are
284 		 * ordered before setting the WORK_DONE_BIT. Ensuring the thread
285 		 * which is going to executed the ordered work sees them.
286 		 * Pairs with the smp_rmb in run_ordered_work.
287 		 */
288 		smp_mb__before_atomic();
289 		set_bit(WORK_DONE_BIT, &work->flags);
290 		run_ordered_work(wq, work);
291 	} else {
292 		/* NB: work must not be dereferenced past this point. */
293 		trace_btrfs_all_work_done(wq->fs_info, work);
294 	}
295 }
296 
297 void btrfs_init_work(struct btrfs_work *work, btrfs_func_t func,
298 		     btrfs_func_t ordered_func, btrfs_func_t ordered_free)
299 {
300 	work->func = func;
301 	work->ordered_func = ordered_func;
302 	work->ordered_free = ordered_free;
303 	INIT_WORK(&work->normal_work, btrfs_work_helper);
304 	INIT_LIST_HEAD(&work->ordered_list);
305 	work->flags = 0;
306 }
307 
308 void btrfs_queue_work(struct btrfs_workqueue *wq, struct btrfs_work *work)
309 {
310 	unsigned long flags;
311 
312 	work->wq = wq;
313 	thresh_queue_hook(wq);
314 	if (work->ordered_func) {
315 		spin_lock_irqsave(&wq->list_lock, flags);
316 		list_add_tail(&work->ordered_list, &wq->ordered_list);
317 		spin_unlock_irqrestore(&wq->list_lock, flags);
318 	}
319 	trace_btrfs_work_queued(work);
320 	queue_work(wq->normal_wq, &work->normal_work);
321 }
322 
323 void btrfs_destroy_workqueue(struct btrfs_workqueue *wq)
324 {
325 	if (!wq)
326 		return;
327 	destroy_workqueue(wq->normal_wq);
328 	trace_btrfs_workqueue_destroy(wq);
329 	kfree(wq);
330 }
331 
332 void btrfs_workqueue_set_max(struct btrfs_workqueue *wq, int limit_active)
333 {
334 	if (wq)
335 		wq->limit_active = limit_active;
336 }
337 
338 void btrfs_flush_workqueue(struct btrfs_workqueue *wq)
339 {
340 	flush_workqueue(wq->normal_wq);
341 }
342