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 #define WORK_DONE_BIT 0
16 #define WORK_ORDER_DONE_BIT 1
17 #define WORK_HIGH_PRIO_BIT 2
18 
19 #define NO_THRESHOLD (-1)
20 #define DFT_THRESHOLD (32)
21 
22 struct __btrfs_workqueue {
23 	struct workqueue_struct *normal_wq;
24 
25 	/* File system this workqueue services */
26 	struct btrfs_fs_info *fs_info;
27 
28 	/* List head pointing to ordered work list */
29 	struct list_head ordered_list;
30 
31 	/* Spinlock for ordered_list */
32 	spinlock_t list_lock;
33 
34 	/* Thresholding related variants */
35 	atomic_t pending;
36 
37 	/* Up limit of concurrency workers */
38 	int limit_active;
39 
40 	/* Current number of concurrency workers */
41 	int current_active;
42 
43 	/* Threshold to change current_active */
44 	int thresh;
45 	unsigned int count;
46 	spinlock_t thres_lock;
47 };
48 
49 struct btrfs_workqueue {
50 	struct __btrfs_workqueue *normal;
51 	struct __btrfs_workqueue *high;
52 };
53 
54 static void normal_work_helper(struct btrfs_work *work);
55 
56 #define BTRFS_WORK_HELPER(name)					\
57 noinline_for_stack void btrfs_##name(struct work_struct *arg)		\
58 {									\
59 	struct btrfs_work *work = container_of(arg, struct btrfs_work,	\
60 					       normal_work);		\
61 	normal_work_helper(work);					\
62 }
63 
64 struct btrfs_fs_info *
btrfs_workqueue_owner(const struct __btrfs_workqueue * wq)65 btrfs_workqueue_owner(const struct __btrfs_workqueue *wq)
66 {
67 	return wq->fs_info;
68 }
69 
70 struct btrfs_fs_info *
btrfs_work_owner(const struct btrfs_work * work)71 btrfs_work_owner(const struct btrfs_work *work)
72 {
73 	return work->wq->fs_info;
74 }
75 
btrfs_workqueue_normal_congested(const struct btrfs_workqueue * wq)76 bool btrfs_workqueue_normal_congested(const struct btrfs_workqueue *wq)
77 {
78 	/*
79 	 * We could compare wq->normal->pending with num_online_cpus()
80 	 * to support "thresh == NO_THRESHOLD" case, but it requires
81 	 * moving up atomic_inc/dec in thresh_queue/exec_hook. Let's
82 	 * postpone it until someone needs the support of that case.
83 	 */
84 	if (wq->normal->thresh == NO_THRESHOLD)
85 		return false;
86 
87 	return atomic_read(&wq->normal->pending) > wq->normal->thresh * 2;
88 }
89 
90 BTRFS_WORK_HELPER(worker_helper);
91 BTRFS_WORK_HELPER(delalloc_helper);
92 BTRFS_WORK_HELPER(flush_delalloc_helper);
93 BTRFS_WORK_HELPER(cache_helper);
94 BTRFS_WORK_HELPER(submit_helper);
95 BTRFS_WORK_HELPER(fixup_helper);
96 BTRFS_WORK_HELPER(endio_helper);
97 BTRFS_WORK_HELPER(endio_meta_helper);
98 BTRFS_WORK_HELPER(endio_meta_write_helper);
99 BTRFS_WORK_HELPER(endio_raid56_helper);
100 BTRFS_WORK_HELPER(endio_repair_helper);
101 BTRFS_WORK_HELPER(rmw_helper);
102 BTRFS_WORK_HELPER(endio_write_helper);
103 BTRFS_WORK_HELPER(freespace_write_helper);
104 BTRFS_WORK_HELPER(delayed_meta_helper);
105 BTRFS_WORK_HELPER(readahead_helper);
106 BTRFS_WORK_HELPER(qgroup_rescan_helper);
107 BTRFS_WORK_HELPER(extent_refs_helper);
108 BTRFS_WORK_HELPER(scrub_helper);
109 BTRFS_WORK_HELPER(scrubwrc_helper);
110 BTRFS_WORK_HELPER(scrubnc_helper);
111 BTRFS_WORK_HELPER(scrubparity_helper);
112 
113 static struct __btrfs_workqueue *
__btrfs_alloc_workqueue(struct btrfs_fs_info * fs_info,const char * name,unsigned int flags,int limit_active,int thresh)114 __btrfs_alloc_workqueue(struct btrfs_fs_info *fs_info, const char *name,
115 			unsigned int flags, int limit_active, int thresh)
116 {
117 	struct __btrfs_workqueue *ret = kzalloc(sizeof(*ret), GFP_KERNEL);
118 
119 	if (!ret)
120 		return NULL;
121 
122 	ret->fs_info = fs_info;
123 	ret->limit_active = limit_active;
124 	atomic_set(&ret->pending, 0);
125 	if (thresh == 0)
126 		thresh = DFT_THRESHOLD;
127 	/* For low threshold, disabling threshold is a better choice */
128 	if (thresh < DFT_THRESHOLD) {
129 		ret->current_active = limit_active;
130 		ret->thresh = NO_THRESHOLD;
131 	} else {
132 		/*
133 		 * For threshold-able wq, let its concurrency grow on demand.
134 		 * Use minimal max_active at alloc time to reduce resource
135 		 * usage.
136 		 */
137 		ret->current_active = 1;
138 		ret->thresh = thresh;
139 	}
140 
141 	if (flags & WQ_HIGHPRI)
142 		ret->normal_wq = alloc_workqueue("%s-%s-high", flags,
143 						 ret->current_active, "btrfs",
144 						 name);
145 	else
146 		ret->normal_wq = alloc_workqueue("%s-%s", flags,
147 						 ret->current_active, "btrfs",
148 						 name);
149 	if (!ret->normal_wq) {
150 		kfree(ret);
151 		return NULL;
152 	}
153 
154 	INIT_LIST_HEAD(&ret->ordered_list);
155 	spin_lock_init(&ret->list_lock);
156 	spin_lock_init(&ret->thres_lock);
157 	trace_btrfs_workqueue_alloc(ret, name, flags & WQ_HIGHPRI);
158 	return ret;
159 }
160 
161 static inline void
162 __btrfs_destroy_workqueue(struct __btrfs_workqueue *wq);
163 
btrfs_alloc_workqueue(struct btrfs_fs_info * fs_info,const char * name,unsigned int flags,int limit_active,int thresh)164 struct btrfs_workqueue *btrfs_alloc_workqueue(struct btrfs_fs_info *fs_info,
165 					      const char *name,
166 					      unsigned int flags,
167 					      int limit_active,
168 					      int thresh)
169 {
170 	struct btrfs_workqueue *ret = kzalloc(sizeof(*ret), GFP_KERNEL);
171 
172 	if (!ret)
173 		return NULL;
174 
175 	ret->normal = __btrfs_alloc_workqueue(fs_info, name,
176 					      flags & ~WQ_HIGHPRI,
177 					      limit_active, thresh);
178 	if (!ret->normal) {
179 		kfree(ret);
180 		return NULL;
181 	}
182 
183 	if (flags & WQ_HIGHPRI) {
184 		ret->high = __btrfs_alloc_workqueue(fs_info, name, flags,
185 						    limit_active, thresh);
186 		if (!ret->high) {
187 			__btrfs_destroy_workqueue(ret->normal);
188 			kfree(ret);
189 			return NULL;
190 		}
191 	}
192 	return ret;
193 }
194 
195 /*
196  * Hook for threshold which will be called in btrfs_queue_work.
197  * This hook WILL be called in IRQ handler context,
198  * so workqueue_set_max_active MUST NOT be called in this hook
199  */
thresh_queue_hook(struct __btrfs_workqueue * wq)200 static inline void thresh_queue_hook(struct __btrfs_workqueue *wq)
201 {
202 	if (wq->thresh == NO_THRESHOLD)
203 		return;
204 	atomic_inc(&wq->pending);
205 }
206 
207 /*
208  * Hook for threshold which will be called before executing the work,
209  * This hook is called in kthread content.
210  * So workqueue_set_max_active is called here.
211  */
thresh_exec_hook(struct __btrfs_workqueue * wq)212 static inline void thresh_exec_hook(struct __btrfs_workqueue *wq)
213 {
214 	int new_current_active;
215 	long pending;
216 	int need_change = 0;
217 
218 	if (wq->thresh == NO_THRESHOLD)
219 		return;
220 
221 	atomic_dec(&wq->pending);
222 	spin_lock(&wq->thres_lock);
223 	/*
224 	 * Use wq->count to limit the calling frequency of
225 	 * workqueue_set_max_active.
226 	 */
227 	wq->count++;
228 	wq->count %= (wq->thresh / 4);
229 	if (!wq->count)
230 		goto  out;
231 	new_current_active = wq->current_active;
232 
233 	/*
234 	 * pending may be changed later, but it's OK since we really
235 	 * don't need it so accurate to calculate new_max_active.
236 	 */
237 	pending = atomic_read(&wq->pending);
238 	if (pending > wq->thresh)
239 		new_current_active++;
240 	if (pending < wq->thresh / 2)
241 		new_current_active--;
242 	new_current_active = clamp_val(new_current_active, 1, wq->limit_active);
243 	if (new_current_active != wq->current_active)  {
244 		need_change = 1;
245 		wq->current_active = new_current_active;
246 	}
247 out:
248 	spin_unlock(&wq->thres_lock);
249 
250 	if (need_change) {
251 		workqueue_set_max_active(wq->normal_wq, wq->current_active);
252 	}
253 }
254 
run_ordered_work(struct __btrfs_workqueue * wq,struct btrfs_work * self)255 static void run_ordered_work(struct __btrfs_workqueue *wq,
256 			     struct btrfs_work *self)
257 {
258 	struct list_head *list = &wq->ordered_list;
259 	struct btrfs_work *work;
260 	spinlock_t *lock = &wq->list_lock;
261 	unsigned long flags;
262 	void *wtag;
263 	bool free_self = false;
264 
265 	while (1) {
266 		spin_lock_irqsave(lock, flags);
267 		if (list_empty(list))
268 			break;
269 		work = list_entry(list->next, struct btrfs_work,
270 				  ordered_list);
271 		if (!test_bit(WORK_DONE_BIT, &work->flags))
272 			break;
273 		/*
274 		 * Orders all subsequent loads after reading WORK_DONE_BIT,
275 		 * paired with the smp_mb__before_atomic in btrfs_work_helper
276 		 * this guarantees that the ordered function will see all
277 		 * updates from ordinary work function.
278 		 */
279 		smp_rmb();
280 
281 		/*
282 		 * we are going to call the ordered done function, but
283 		 * we leave the work item on the list as a barrier so
284 		 * that later work items that are done don't have their
285 		 * functions called before this one returns
286 		 */
287 		if (test_and_set_bit(WORK_ORDER_DONE_BIT, &work->flags))
288 			break;
289 		trace_btrfs_ordered_sched(work);
290 		spin_unlock_irqrestore(lock, flags);
291 		work->ordered_func(work);
292 
293 		/* now take the lock again and drop our item from the list */
294 		spin_lock_irqsave(lock, flags);
295 		list_del(&work->ordered_list);
296 		spin_unlock_irqrestore(lock, flags);
297 
298 		if (work == self) {
299 			/*
300 			 * This is the work item that the worker is currently
301 			 * executing.
302 			 *
303 			 * The kernel workqueue code guarantees non-reentrancy
304 			 * of work items. I.e., if a work item with the same
305 			 * address and work function is queued twice, the second
306 			 * execution is blocked until the first one finishes. A
307 			 * work item may be freed and recycled with the same
308 			 * work function; the workqueue code assumes that the
309 			 * original work item cannot depend on the recycled work
310 			 * item in that case (see find_worker_executing_work()).
311 			 *
312 			 * Note that the work of one Btrfs filesystem may depend
313 			 * on the work of another Btrfs filesystem via, e.g., a
314 			 * loop device. Therefore, we must not allow the current
315 			 * work item to be recycled until we are really done,
316 			 * otherwise we break the above assumption and can
317 			 * deadlock.
318 			 */
319 			free_self = true;
320 		} else {
321 			/*
322 			 * We don't want to call the ordered free functions with
323 			 * the lock held though. Save the work as tag for the
324 			 * trace event, because the callback could free the
325 			 * structure.
326 			 */
327 			wtag = work;
328 			work->ordered_free(work);
329 			trace_btrfs_all_work_done(wq->fs_info, wtag);
330 		}
331 	}
332 	spin_unlock_irqrestore(lock, flags);
333 
334 	if (free_self) {
335 		wtag = self;
336 		self->ordered_free(self);
337 		trace_btrfs_all_work_done(wq->fs_info, wtag);
338 	}
339 }
340 
normal_work_helper(struct btrfs_work * work)341 static void normal_work_helper(struct btrfs_work *work)
342 {
343 	struct __btrfs_workqueue *wq;
344 	void *wtag;
345 	int need_order = 0;
346 
347 	/*
348 	 * We should not touch things inside work in the following cases:
349 	 * 1) after work->func() if it has no ordered_free
350 	 *    Since the struct is freed in work->func().
351 	 * 2) after setting WORK_DONE_BIT
352 	 *    The work may be freed in other threads almost instantly.
353 	 * So we save the needed things here.
354 	 */
355 	if (work->ordered_func)
356 		need_order = 1;
357 	wq = work->wq;
358 	/* Safe for tracepoints in case work gets freed by the callback */
359 	wtag = work;
360 
361 	trace_btrfs_work_sched(work);
362 	thresh_exec_hook(wq);
363 	work->func(work);
364 	if (need_order) {
365 		/*
366 		 * Ensures all memory accesses done in the work function are
367 		 * ordered before setting the WORK_DONE_BIT. Ensuring the thread
368 		 * which is going to executed the ordered work sees them.
369 		 * Pairs with the smp_rmb in run_ordered_work.
370 		 */
371 		smp_mb__before_atomic();
372 		set_bit(WORK_DONE_BIT, &work->flags);
373 		run_ordered_work(wq, work);
374 	}
375 	if (!need_order)
376 		trace_btrfs_all_work_done(wq->fs_info, wtag);
377 }
378 
btrfs_init_work(struct btrfs_work * work,btrfs_work_func_t uniq_func,btrfs_func_t func,btrfs_func_t ordered_func,btrfs_func_t ordered_free)379 void btrfs_init_work(struct btrfs_work *work, btrfs_work_func_t uniq_func,
380 		     btrfs_func_t func,
381 		     btrfs_func_t ordered_func,
382 		     btrfs_func_t ordered_free)
383 {
384 	work->func = func;
385 	work->ordered_func = ordered_func;
386 	work->ordered_free = ordered_free;
387 	INIT_WORK(&work->normal_work, uniq_func);
388 	INIT_LIST_HEAD(&work->ordered_list);
389 	work->flags = 0;
390 }
391 
__btrfs_queue_work(struct __btrfs_workqueue * wq,struct btrfs_work * work)392 static inline void __btrfs_queue_work(struct __btrfs_workqueue *wq,
393 				      struct btrfs_work *work)
394 {
395 	unsigned long flags;
396 
397 	work->wq = wq;
398 	thresh_queue_hook(wq);
399 	if (work->ordered_func) {
400 		spin_lock_irqsave(&wq->list_lock, flags);
401 		list_add_tail(&work->ordered_list, &wq->ordered_list);
402 		spin_unlock_irqrestore(&wq->list_lock, flags);
403 	}
404 	trace_btrfs_work_queued(work);
405 	queue_work(wq->normal_wq, &work->normal_work);
406 }
407 
btrfs_queue_work(struct btrfs_workqueue * wq,struct btrfs_work * work)408 void btrfs_queue_work(struct btrfs_workqueue *wq,
409 		      struct btrfs_work *work)
410 {
411 	struct __btrfs_workqueue *dest_wq;
412 
413 	if (test_bit(WORK_HIGH_PRIO_BIT, &work->flags) && wq->high)
414 		dest_wq = wq->high;
415 	else
416 		dest_wq = wq->normal;
417 	__btrfs_queue_work(dest_wq, work);
418 }
419 
420 static inline void
__btrfs_destroy_workqueue(struct __btrfs_workqueue * wq)421 __btrfs_destroy_workqueue(struct __btrfs_workqueue *wq)
422 {
423 	destroy_workqueue(wq->normal_wq);
424 	trace_btrfs_workqueue_destroy(wq);
425 	kfree(wq);
426 }
427 
btrfs_destroy_workqueue(struct btrfs_workqueue * wq)428 void btrfs_destroy_workqueue(struct btrfs_workqueue *wq)
429 {
430 	if (!wq)
431 		return;
432 	if (wq->high)
433 		__btrfs_destroy_workqueue(wq->high);
434 	__btrfs_destroy_workqueue(wq->normal);
435 	kfree(wq);
436 }
437 
btrfs_workqueue_set_max(struct btrfs_workqueue * wq,int limit_active)438 void btrfs_workqueue_set_max(struct btrfs_workqueue *wq, int limit_active)
439 {
440 	if (!wq)
441 		return;
442 	wq->normal->limit_active = limit_active;
443 	if (wq->high)
444 		wq->high->limit_active = limit_active;
445 }
446 
btrfs_set_work_high_priority(struct btrfs_work * work)447 void btrfs_set_work_high_priority(struct btrfs_work *work)
448 {
449 	set_bit(WORK_HIGH_PRIO_BIT, &work->flags);
450 }
451 
btrfs_flush_workqueue(struct btrfs_workqueue * wq)452 void btrfs_flush_workqueue(struct btrfs_workqueue *wq)
453 {
454 	if (wq->high)
455 		flush_workqueue(wq->high->normal_wq);
456 
457 	flush_workqueue(wq->normal->normal_wq);
458 }
459