xref: /openbmc/linux/fs/btrfs/async-thread.c (revision 41e4b7dc)
1 /*
2  * Copyright (C) 2007 Oracle.  All rights reserved.
3  * Copyright (C) 2014 Fujitsu.  All rights reserved.
4  *
5  * This program is free software; you can redistribute it and/or
6  * modify it under the terms of the GNU General Public
7  * License v2 as published by the Free Software Foundation.
8  *
9  * This program is distributed in the hope that it will be useful,
10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
12  * General Public License for more details.
13  *
14  * You should have received a copy of the GNU General Public
15  * License along with this program; if not, write to the
16  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
17  * Boston, MA 021110-1307, USA.
18  */
19 
20 #include <linux/kthread.h>
21 #include <linux/slab.h>
22 #include <linux/list.h>
23 #include <linux/spinlock.h>
24 #include <linux/freezer.h>
25 #include "async-thread.h"
26 #include "ctree.h"
27 
28 #define WORK_DONE_BIT 0
29 #define WORK_ORDER_DONE_BIT 1
30 #define WORK_HIGH_PRIO_BIT 2
31 
32 #define NO_THRESHOLD (-1)
33 #define DFT_THRESHOLD (32)
34 
35 struct __btrfs_workqueue {
36 	struct workqueue_struct *normal_wq;
37 
38 	/* File system this workqueue services */
39 	struct btrfs_fs_info *fs_info;
40 
41 	/* List head pointing to ordered work list */
42 	struct list_head ordered_list;
43 
44 	/* Spinlock for ordered_list */
45 	spinlock_t list_lock;
46 
47 	/* Thresholding related variants */
48 	atomic_t pending;
49 
50 	/* Up limit of concurrency workers */
51 	int limit_active;
52 
53 	/* Current number of concurrency workers */
54 	int current_active;
55 
56 	/* Threshold to change current_active */
57 	int thresh;
58 	unsigned int count;
59 	spinlock_t thres_lock;
60 };
61 
62 struct btrfs_workqueue {
63 	struct __btrfs_workqueue *normal;
64 	struct __btrfs_workqueue *high;
65 };
66 
67 static void normal_work_helper(struct btrfs_work *work);
68 
69 #define BTRFS_WORK_HELPER(name)					\
70 void btrfs_##name(struct work_struct *arg)				\
71 {									\
72 	struct btrfs_work *work = container_of(arg, struct btrfs_work,	\
73 					       normal_work);		\
74 	normal_work_helper(work);					\
75 }
76 
77 struct btrfs_fs_info *
78 btrfs_workqueue_owner(const struct __btrfs_workqueue *wq)
79 {
80 	return wq->fs_info;
81 }
82 
83 struct btrfs_fs_info *
84 btrfs_work_owner(const struct btrfs_work *work)
85 {
86 	return work->wq->fs_info;
87 }
88 
89 bool btrfs_workqueue_normal_congested(const struct btrfs_workqueue *wq)
90 {
91 	/*
92 	 * We could compare wq->normal->pending with num_online_cpus()
93 	 * to support "thresh == NO_THRESHOLD" case, but it requires
94 	 * moving up atomic_inc/dec in thresh_queue/exec_hook. Let's
95 	 * postpone it until someone needs the support of that case.
96 	 */
97 	if (wq->normal->thresh == NO_THRESHOLD)
98 		return false;
99 
100 	return atomic_read(&wq->normal->pending) > wq->normal->thresh * 2;
101 }
102 
103 BTRFS_WORK_HELPER(worker_helper);
104 BTRFS_WORK_HELPER(delalloc_helper);
105 BTRFS_WORK_HELPER(flush_delalloc_helper);
106 BTRFS_WORK_HELPER(cache_helper);
107 BTRFS_WORK_HELPER(submit_helper);
108 BTRFS_WORK_HELPER(fixup_helper);
109 BTRFS_WORK_HELPER(endio_helper);
110 BTRFS_WORK_HELPER(endio_meta_helper);
111 BTRFS_WORK_HELPER(endio_meta_write_helper);
112 BTRFS_WORK_HELPER(endio_raid56_helper);
113 BTRFS_WORK_HELPER(endio_repair_helper);
114 BTRFS_WORK_HELPER(rmw_helper);
115 BTRFS_WORK_HELPER(endio_write_helper);
116 BTRFS_WORK_HELPER(freespace_write_helper);
117 BTRFS_WORK_HELPER(delayed_meta_helper);
118 BTRFS_WORK_HELPER(readahead_helper);
119 BTRFS_WORK_HELPER(qgroup_rescan_helper);
120 BTRFS_WORK_HELPER(extent_refs_helper);
121 BTRFS_WORK_HELPER(scrub_helper);
122 BTRFS_WORK_HELPER(scrubwrc_helper);
123 BTRFS_WORK_HELPER(scrubnc_helper);
124 BTRFS_WORK_HELPER(scrubparity_helper);
125 
126 static struct __btrfs_workqueue *
127 __btrfs_alloc_workqueue(struct btrfs_fs_info *fs_info, const char *name,
128 			unsigned int flags, int limit_active, int thresh)
129 {
130 	struct __btrfs_workqueue *ret = kzalloc(sizeof(*ret), GFP_KERNEL);
131 
132 	if (!ret)
133 		return NULL;
134 
135 	ret->fs_info = fs_info;
136 	ret->limit_active = limit_active;
137 	atomic_set(&ret->pending, 0);
138 	if (thresh == 0)
139 		thresh = DFT_THRESHOLD;
140 	/* For low threshold, disabling threshold is a better choice */
141 	if (thresh < DFT_THRESHOLD) {
142 		ret->current_active = limit_active;
143 		ret->thresh = NO_THRESHOLD;
144 	} else {
145 		/*
146 		 * For threshold-able wq, let its concurrency grow on demand.
147 		 * Use minimal max_active at alloc time to reduce resource
148 		 * usage.
149 		 */
150 		ret->current_active = 1;
151 		ret->thresh = thresh;
152 	}
153 
154 	if (flags & WQ_HIGHPRI)
155 		ret->normal_wq = alloc_workqueue("%s-%s-high", flags,
156 						 ret->current_active, "btrfs",
157 						 name);
158 	else
159 		ret->normal_wq = alloc_workqueue("%s-%s", flags,
160 						 ret->current_active, "btrfs",
161 						 name);
162 	if (!ret->normal_wq) {
163 		kfree(ret);
164 		return NULL;
165 	}
166 
167 	INIT_LIST_HEAD(&ret->ordered_list);
168 	spin_lock_init(&ret->list_lock);
169 	spin_lock_init(&ret->thres_lock);
170 	trace_btrfs_workqueue_alloc(ret, name, flags & WQ_HIGHPRI);
171 	return ret;
172 }
173 
174 static inline void
175 __btrfs_destroy_workqueue(struct __btrfs_workqueue *wq);
176 
177 struct btrfs_workqueue *btrfs_alloc_workqueue(struct btrfs_fs_info *fs_info,
178 					      const char *name,
179 					      unsigned int flags,
180 					      int limit_active,
181 					      int thresh)
182 {
183 	struct btrfs_workqueue *ret = kzalloc(sizeof(*ret), GFP_KERNEL);
184 
185 	if (!ret)
186 		return NULL;
187 
188 	ret->normal = __btrfs_alloc_workqueue(fs_info, name,
189 					      flags & ~WQ_HIGHPRI,
190 					      limit_active, thresh);
191 	if (!ret->normal) {
192 		kfree(ret);
193 		return NULL;
194 	}
195 
196 	if (flags & WQ_HIGHPRI) {
197 		ret->high = __btrfs_alloc_workqueue(fs_info, name, flags,
198 						    limit_active, thresh);
199 		if (!ret->high) {
200 			__btrfs_destroy_workqueue(ret->normal);
201 			kfree(ret);
202 			return NULL;
203 		}
204 	}
205 	return ret;
206 }
207 
208 /*
209  * Hook for threshold which will be called in btrfs_queue_work.
210  * This hook WILL be called in IRQ handler context,
211  * so workqueue_set_max_active MUST NOT be called in this hook
212  */
213 static inline void thresh_queue_hook(struct __btrfs_workqueue *wq)
214 {
215 	if (wq->thresh == NO_THRESHOLD)
216 		return;
217 	atomic_inc(&wq->pending);
218 }
219 
220 /*
221  * Hook for threshold which will be called before executing the work,
222  * This hook is called in kthread content.
223  * So workqueue_set_max_active is called here.
224  */
225 static inline void thresh_exec_hook(struct __btrfs_workqueue *wq)
226 {
227 	int new_current_active;
228 	long pending;
229 	int need_change = 0;
230 
231 	if (wq->thresh == NO_THRESHOLD)
232 		return;
233 
234 	atomic_dec(&wq->pending);
235 	spin_lock(&wq->thres_lock);
236 	/*
237 	 * Use wq->count to limit the calling frequency of
238 	 * workqueue_set_max_active.
239 	 */
240 	wq->count++;
241 	wq->count %= (wq->thresh / 4);
242 	if (!wq->count)
243 		goto  out;
244 	new_current_active = wq->current_active;
245 
246 	/*
247 	 * pending may be changed later, but it's OK since we really
248 	 * don't need it so accurate to calculate new_max_active.
249 	 */
250 	pending = atomic_read(&wq->pending);
251 	if (pending > wq->thresh)
252 		new_current_active++;
253 	if (pending < wq->thresh / 2)
254 		new_current_active--;
255 	new_current_active = clamp_val(new_current_active, 1, wq->limit_active);
256 	if (new_current_active != wq->current_active)  {
257 		need_change = 1;
258 		wq->current_active = new_current_active;
259 	}
260 out:
261 	spin_unlock(&wq->thres_lock);
262 
263 	if (need_change) {
264 		workqueue_set_max_active(wq->normal_wq, wq->current_active);
265 	}
266 }
267 
268 static void run_ordered_work(struct __btrfs_workqueue *wq)
269 {
270 	struct list_head *list = &wq->ordered_list;
271 	struct btrfs_work *work;
272 	spinlock_t *lock = &wq->list_lock;
273 	unsigned long flags;
274 
275 	while (1) {
276 		void *wtag;
277 
278 		spin_lock_irqsave(lock, flags);
279 		if (list_empty(list))
280 			break;
281 		work = list_entry(list->next, struct btrfs_work,
282 				  ordered_list);
283 		if (!test_bit(WORK_DONE_BIT, &work->flags))
284 			break;
285 
286 		/*
287 		 * we are going to call the ordered done function, but
288 		 * we leave the work item on the list as a barrier so
289 		 * that later work items that are done don't have their
290 		 * functions called before this one returns
291 		 */
292 		if (test_and_set_bit(WORK_ORDER_DONE_BIT, &work->flags))
293 			break;
294 		trace_btrfs_ordered_sched(work);
295 		spin_unlock_irqrestore(lock, flags);
296 		work->ordered_func(work);
297 
298 		/* now take the lock again and drop our item from the list */
299 		spin_lock_irqsave(lock, flags);
300 		list_del(&work->ordered_list);
301 		spin_unlock_irqrestore(lock, flags);
302 
303 		/*
304 		 * We don't want to call the ordered free functions with the
305 		 * lock held though. Save the work as tag for the trace event,
306 		 * because the callback could free the structure.
307 		 */
308 		wtag = work;
309 		work->ordered_free(work);
310 		trace_btrfs_all_work_done(wq->fs_info, wtag);
311 	}
312 	spin_unlock_irqrestore(lock, flags);
313 }
314 
315 static void normal_work_helper(struct btrfs_work *work)
316 {
317 	struct __btrfs_workqueue *wq;
318 	void *wtag;
319 	int need_order = 0;
320 
321 	/*
322 	 * We should not touch things inside work in the following cases:
323 	 * 1) after work->func() if it has no ordered_free
324 	 *    Since the struct is freed in work->func().
325 	 * 2) after setting WORK_DONE_BIT
326 	 *    The work may be freed in other threads almost instantly.
327 	 * So we save the needed things here.
328 	 */
329 	if (work->ordered_func)
330 		need_order = 1;
331 	wq = work->wq;
332 	/* Safe for tracepoints in case work gets freed by the callback */
333 	wtag = work;
334 
335 	trace_btrfs_work_sched(work);
336 	thresh_exec_hook(wq);
337 	work->func(work);
338 	if (need_order) {
339 		set_bit(WORK_DONE_BIT, &work->flags);
340 		run_ordered_work(wq);
341 	}
342 	if (!need_order)
343 		trace_btrfs_all_work_done(wq->fs_info, wtag);
344 }
345 
346 void btrfs_init_work(struct btrfs_work *work, btrfs_work_func_t uniq_func,
347 		     btrfs_func_t func,
348 		     btrfs_func_t ordered_func,
349 		     btrfs_func_t ordered_free)
350 {
351 	work->func = func;
352 	work->ordered_func = ordered_func;
353 	work->ordered_free = ordered_free;
354 	INIT_WORK(&work->normal_work, uniq_func);
355 	INIT_LIST_HEAD(&work->ordered_list);
356 	work->flags = 0;
357 }
358 
359 static inline void __btrfs_queue_work(struct __btrfs_workqueue *wq,
360 				      struct btrfs_work *work)
361 {
362 	unsigned long flags;
363 
364 	work->wq = wq;
365 	thresh_queue_hook(wq);
366 	if (work->ordered_func) {
367 		spin_lock_irqsave(&wq->list_lock, flags);
368 		list_add_tail(&work->ordered_list, &wq->ordered_list);
369 		spin_unlock_irqrestore(&wq->list_lock, flags);
370 	}
371 	trace_btrfs_work_queued(work);
372 	queue_work(wq->normal_wq, &work->normal_work);
373 }
374 
375 void btrfs_queue_work(struct btrfs_workqueue *wq,
376 		      struct btrfs_work *work)
377 {
378 	struct __btrfs_workqueue *dest_wq;
379 
380 	if (test_bit(WORK_HIGH_PRIO_BIT, &work->flags) && wq->high)
381 		dest_wq = wq->high;
382 	else
383 		dest_wq = wq->normal;
384 	__btrfs_queue_work(dest_wq, work);
385 }
386 
387 static inline void
388 __btrfs_destroy_workqueue(struct __btrfs_workqueue *wq)
389 {
390 	destroy_workqueue(wq->normal_wq);
391 	trace_btrfs_workqueue_destroy(wq);
392 	kfree(wq);
393 }
394 
395 void btrfs_destroy_workqueue(struct btrfs_workqueue *wq)
396 {
397 	if (!wq)
398 		return;
399 	if (wq->high)
400 		__btrfs_destroy_workqueue(wq->high);
401 	__btrfs_destroy_workqueue(wq->normal);
402 	kfree(wq);
403 }
404 
405 void btrfs_workqueue_set_max(struct btrfs_workqueue *wq, int limit_active)
406 {
407 	if (!wq)
408 		return;
409 	wq->normal->limit_active = limit_active;
410 	if (wq->high)
411 		wq->high->limit_active = limit_active;
412 }
413 
414 void btrfs_set_work_high_priority(struct btrfs_work *work)
415 {
416 	set_bit(WORK_HIGH_PRIO_BIT, &work->flags);
417 }
418