xref: /openbmc/linux/block/blk-throttle.c (revision 3821a065)
1 /*
2  * Interface for controlling IO bandwidth on a request queue
3  *
4  * Copyright (C) 2010 Vivek Goyal <vgoyal@redhat.com>
5  */
6 
7 #include <linux/module.h>
8 #include <linux/slab.h>
9 #include <linux/blkdev.h>
10 #include <linux/bio.h>
11 #include <linux/blktrace_api.h>
12 #include <linux/blk-cgroup.h>
13 #include "blk.h"
14 
15 /* Max dispatch from a group in 1 round */
16 static int throtl_grp_quantum = 8;
17 
18 /* Total max dispatch from all groups in one round */
19 static int throtl_quantum = 32;
20 
21 /* Throttling is performed over 100ms slice and after that slice is renewed */
22 static unsigned long throtl_slice = HZ/10;	/* 100 ms */
23 
24 static struct blkcg_policy blkcg_policy_throtl;
25 
26 /* A workqueue to queue throttle related work */
27 static struct workqueue_struct *kthrotld_workqueue;
28 
29 /*
30  * To implement hierarchical throttling, throtl_grps form a tree and bios
31  * are dispatched upwards level by level until they reach the top and get
32  * issued.  When dispatching bios from the children and local group at each
33  * level, if the bios are dispatched into a single bio_list, there's a risk
34  * of a local or child group which can queue many bios at once filling up
35  * the list starving others.
36  *
37  * To avoid such starvation, dispatched bios are queued separately
38  * according to where they came from.  When they are again dispatched to
39  * the parent, they're popped in round-robin order so that no single source
40  * hogs the dispatch window.
41  *
42  * throtl_qnode is used to keep the queued bios separated by their sources.
43  * Bios are queued to throtl_qnode which in turn is queued to
44  * throtl_service_queue and then dispatched in round-robin order.
45  *
46  * It's also used to track the reference counts on blkg's.  A qnode always
47  * belongs to a throtl_grp and gets queued on itself or the parent, so
48  * incrementing the reference of the associated throtl_grp when a qnode is
49  * queued and decrementing when dequeued is enough to keep the whole blkg
50  * tree pinned while bios are in flight.
51  */
52 struct throtl_qnode {
53 	struct list_head	node;		/* service_queue->queued[] */
54 	struct bio_list		bios;		/* queued bios */
55 	struct throtl_grp	*tg;		/* tg this qnode belongs to */
56 };
57 
58 struct throtl_service_queue {
59 	struct throtl_service_queue *parent_sq;	/* the parent service_queue */
60 
61 	/*
62 	 * Bios queued directly to this service_queue or dispatched from
63 	 * children throtl_grp's.
64 	 */
65 	struct list_head	queued[2];	/* throtl_qnode [READ/WRITE] */
66 	unsigned int		nr_queued[2];	/* number of queued bios */
67 
68 	/*
69 	 * RB tree of active children throtl_grp's, which are sorted by
70 	 * their ->disptime.
71 	 */
72 	struct rb_root		pending_tree;	/* RB tree of active tgs */
73 	struct rb_node		*first_pending;	/* first node in the tree */
74 	unsigned int		nr_pending;	/* # queued in the tree */
75 	unsigned long		first_pending_disptime;	/* disptime of the first tg */
76 	struct timer_list	pending_timer;	/* fires on first_pending_disptime */
77 };
78 
79 enum tg_state_flags {
80 	THROTL_TG_PENDING	= 1 << 0,	/* on parent's pending tree */
81 	THROTL_TG_WAS_EMPTY	= 1 << 1,	/* bio_lists[] became non-empty */
82 };
83 
84 #define rb_entry_tg(node)	rb_entry((node), struct throtl_grp, rb_node)
85 
86 struct throtl_grp {
87 	/* must be the first member */
88 	struct blkg_policy_data pd;
89 
90 	/* active throtl group service_queue member */
91 	struct rb_node rb_node;
92 
93 	/* throtl_data this group belongs to */
94 	struct throtl_data *td;
95 
96 	/* this group's service queue */
97 	struct throtl_service_queue service_queue;
98 
99 	/*
100 	 * qnode_on_self is used when bios are directly queued to this
101 	 * throtl_grp so that local bios compete fairly with bios
102 	 * dispatched from children.  qnode_on_parent is used when bios are
103 	 * dispatched from this throtl_grp into its parent and will compete
104 	 * with the sibling qnode_on_parents and the parent's
105 	 * qnode_on_self.
106 	 */
107 	struct throtl_qnode qnode_on_self[2];
108 	struct throtl_qnode qnode_on_parent[2];
109 
110 	/*
111 	 * Dispatch time in jiffies. This is the estimated time when group
112 	 * will unthrottle and is ready to dispatch more bio. It is used as
113 	 * key to sort active groups in service tree.
114 	 */
115 	unsigned long disptime;
116 
117 	unsigned int flags;
118 
119 	/* are there any throtl rules between this group and td? */
120 	bool has_rules[2];
121 
122 	/* bytes per second rate limits */
123 	uint64_t bps[2];
124 
125 	/* IOPS limits */
126 	unsigned int iops[2];
127 
128 	/* Number of bytes disptached in current slice */
129 	uint64_t bytes_disp[2];
130 	/* Number of bio's dispatched in current slice */
131 	unsigned int io_disp[2];
132 
133 	/* When did we start a new slice */
134 	unsigned long slice_start[2];
135 	unsigned long slice_end[2];
136 };
137 
138 struct throtl_data
139 {
140 	/* service tree for active throtl groups */
141 	struct throtl_service_queue service_queue;
142 
143 	struct request_queue *queue;
144 
145 	/* Total Number of queued bios on READ and WRITE lists */
146 	unsigned int nr_queued[2];
147 
148 	/*
149 	 * number of total undestroyed groups
150 	 */
151 	unsigned int nr_undestroyed_grps;
152 
153 	/* Work for dispatching throttled bios */
154 	struct work_struct dispatch_work;
155 };
156 
157 static void throtl_pending_timer_fn(unsigned long arg);
158 
159 static inline struct throtl_grp *pd_to_tg(struct blkg_policy_data *pd)
160 {
161 	return pd ? container_of(pd, struct throtl_grp, pd) : NULL;
162 }
163 
164 static inline struct throtl_grp *blkg_to_tg(struct blkcg_gq *blkg)
165 {
166 	return pd_to_tg(blkg_to_pd(blkg, &blkcg_policy_throtl));
167 }
168 
169 static inline struct blkcg_gq *tg_to_blkg(struct throtl_grp *tg)
170 {
171 	return pd_to_blkg(&tg->pd);
172 }
173 
174 /**
175  * sq_to_tg - return the throl_grp the specified service queue belongs to
176  * @sq: the throtl_service_queue of interest
177  *
178  * Return the throtl_grp @sq belongs to.  If @sq is the top-level one
179  * embedded in throtl_data, %NULL is returned.
180  */
181 static struct throtl_grp *sq_to_tg(struct throtl_service_queue *sq)
182 {
183 	if (sq && sq->parent_sq)
184 		return container_of(sq, struct throtl_grp, service_queue);
185 	else
186 		return NULL;
187 }
188 
189 /**
190  * sq_to_td - return throtl_data the specified service queue belongs to
191  * @sq: the throtl_service_queue of interest
192  *
193  * A service_queue can be embeded in either a throtl_grp or throtl_data.
194  * Determine the associated throtl_data accordingly and return it.
195  */
196 static struct throtl_data *sq_to_td(struct throtl_service_queue *sq)
197 {
198 	struct throtl_grp *tg = sq_to_tg(sq);
199 
200 	if (tg)
201 		return tg->td;
202 	else
203 		return container_of(sq, struct throtl_data, service_queue);
204 }
205 
206 /**
207  * throtl_log - log debug message via blktrace
208  * @sq: the service_queue being reported
209  * @fmt: printf format string
210  * @args: printf args
211  *
212  * The messages are prefixed with "throtl BLKG_NAME" if @sq belongs to a
213  * throtl_grp; otherwise, just "throtl".
214  *
215  * TODO: this should be made a function and name formatting should happen
216  * after testing whether blktrace is enabled.
217  */
218 #define throtl_log(sq, fmt, args...)	do {				\
219 	struct throtl_grp *__tg = sq_to_tg((sq));			\
220 	struct throtl_data *__td = sq_to_td((sq));			\
221 									\
222 	(void)__td;							\
223 	if ((__tg)) {							\
224 		char __pbuf[128];					\
225 									\
226 		blkg_path(tg_to_blkg(__tg), __pbuf, sizeof(__pbuf));	\
227 		blk_add_trace_msg(__td->queue, "throtl %s " fmt, __pbuf, ##args); \
228 	} else {							\
229 		blk_add_trace_msg(__td->queue, "throtl " fmt, ##args);	\
230 	}								\
231 } while (0)
232 
233 static void throtl_qnode_init(struct throtl_qnode *qn, struct throtl_grp *tg)
234 {
235 	INIT_LIST_HEAD(&qn->node);
236 	bio_list_init(&qn->bios);
237 	qn->tg = tg;
238 }
239 
240 /**
241  * throtl_qnode_add_bio - add a bio to a throtl_qnode and activate it
242  * @bio: bio being added
243  * @qn: qnode to add bio to
244  * @queued: the service_queue->queued[] list @qn belongs to
245  *
246  * Add @bio to @qn and put @qn on @queued if it's not already on.
247  * @qn->tg's reference count is bumped when @qn is activated.  See the
248  * comment on top of throtl_qnode definition for details.
249  */
250 static void throtl_qnode_add_bio(struct bio *bio, struct throtl_qnode *qn,
251 				 struct list_head *queued)
252 {
253 	bio_list_add(&qn->bios, bio);
254 	if (list_empty(&qn->node)) {
255 		list_add_tail(&qn->node, queued);
256 		blkg_get(tg_to_blkg(qn->tg));
257 	}
258 }
259 
260 /**
261  * throtl_peek_queued - peek the first bio on a qnode list
262  * @queued: the qnode list to peek
263  */
264 static struct bio *throtl_peek_queued(struct list_head *queued)
265 {
266 	struct throtl_qnode *qn = list_first_entry(queued, struct throtl_qnode, node);
267 	struct bio *bio;
268 
269 	if (list_empty(queued))
270 		return NULL;
271 
272 	bio = bio_list_peek(&qn->bios);
273 	WARN_ON_ONCE(!bio);
274 	return bio;
275 }
276 
277 /**
278  * throtl_pop_queued - pop the first bio form a qnode list
279  * @queued: the qnode list to pop a bio from
280  * @tg_to_put: optional out argument for throtl_grp to put
281  *
282  * Pop the first bio from the qnode list @queued.  After popping, the first
283  * qnode is removed from @queued if empty or moved to the end of @queued so
284  * that the popping order is round-robin.
285  *
286  * When the first qnode is removed, its associated throtl_grp should be put
287  * too.  If @tg_to_put is NULL, this function automatically puts it;
288  * otherwise, *@tg_to_put is set to the throtl_grp to put and the caller is
289  * responsible for putting it.
290  */
291 static struct bio *throtl_pop_queued(struct list_head *queued,
292 				     struct throtl_grp **tg_to_put)
293 {
294 	struct throtl_qnode *qn = list_first_entry(queued, struct throtl_qnode, node);
295 	struct bio *bio;
296 
297 	if (list_empty(queued))
298 		return NULL;
299 
300 	bio = bio_list_pop(&qn->bios);
301 	WARN_ON_ONCE(!bio);
302 
303 	if (bio_list_empty(&qn->bios)) {
304 		list_del_init(&qn->node);
305 		if (tg_to_put)
306 			*tg_to_put = qn->tg;
307 		else
308 			blkg_put(tg_to_blkg(qn->tg));
309 	} else {
310 		list_move_tail(&qn->node, queued);
311 	}
312 
313 	return bio;
314 }
315 
316 /* init a service_queue, assumes the caller zeroed it */
317 static void throtl_service_queue_init(struct throtl_service_queue *sq)
318 {
319 	INIT_LIST_HEAD(&sq->queued[0]);
320 	INIT_LIST_HEAD(&sq->queued[1]);
321 	sq->pending_tree = RB_ROOT;
322 	setup_timer(&sq->pending_timer, throtl_pending_timer_fn,
323 		    (unsigned long)sq);
324 }
325 
326 static struct blkg_policy_data *throtl_pd_alloc(gfp_t gfp, int node)
327 {
328 	struct throtl_grp *tg;
329 	int rw;
330 
331 	tg = kzalloc_node(sizeof(*tg), gfp, node);
332 	if (!tg)
333 		return NULL;
334 
335 	throtl_service_queue_init(&tg->service_queue);
336 
337 	for (rw = READ; rw <= WRITE; rw++) {
338 		throtl_qnode_init(&tg->qnode_on_self[rw], tg);
339 		throtl_qnode_init(&tg->qnode_on_parent[rw], tg);
340 	}
341 
342 	RB_CLEAR_NODE(&tg->rb_node);
343 	tg->bps[READ] = -1;
344 	tg->bps[WRITE] = -1;
345 	tg->iops[READ] = -1;
346 	tg->iops[WRITE] = -1;
347 
348 	return &tg->pd;
349 }
350 
351 static void throtl_pd_init(struct blkg_policy_data *pd)
352 {
353 	struct throtl_grp *tg = pd_to_tg(pd);
354 	struct blkcg_gq *blkg = tg_to_blkg(tg);
355 	struct throtl_data *td = blkg->q->td;
356 	struct throtl_service_queue *sq = &tg->service_queue;
357 
358 	/*
359 	 * If on the default hierarchy, we switch to properly hierarchical
360 	 * behavior where limits on a given throtl_grp are applied to the
361 	 * whole subtree rather than just the group itself.  e.g. If 16M
362 	 * read_bps limit is set on the root group, the whole system can't
363 	 * exceed 16M for the device.
364 	 *
365 	 * If not on the default hierarchy, the broken flat hierarchy
366 	 * behavior is retained where all throtl_grps are treated as if
367 	 * they're all separate root groups right below throtl_data.
368 	 * Limits of a group don't interact with limits of other groups
369 	 * regardless of the position of the group in the hierarchy.
370 	 */
371 	sq->parent_sq = &td->service_queue;
372 	if (cgroup_on_dfl(blkg->blkcg->css.cgroup) && blkg->parent)
373 		sq->parent_sq = &blkg_to_tg(blkg->parent)->service_queue;
374 	tg->td = td;
375 }
376 
377 /*
378  * Set has_rules[] if @tg or any of its parents have limits configured.
379  * This doesn't require walking up to the top of the hierarchy as the
380  * parent's has_rules[] is guaranteed to be correct.
381  */
382 static void tg_update_has_rules(struct throtl_grp *tg)
383 {
384 	struct throtl_grp *parent_tg = sq_to_tg(tg->service_queue.parent_sq);
385 	int rw;
386 
387 	for (rw = READ; rw <= WRITE; rw++)
388 		tg->has_rules[rw] = (parent_tg && parent_tg->has_rules[rw]) ||
389 				    (tg->bps[rw] != -1 || tg->iops[rw] != -1);
390 }
391 
392 static void throtl_pd_online(struct blkg_policy_data *pd)
393 {
394 	/*
395 	 * We don't want new groups to escape the limits of its ancestors.
396 	 * Update has_rules[] after a new group is brought online.
397 	 */
398 	tg_update_has_rules(pd_to_tg(pd));
399 }
400 
401 static void throtl_pd_free(struct blkg_policy_data *pd)
402 {
403 	struct throtl_grp *tg = pd_to_tg(pd);
404 
405 	del_timer_sync(&tg->service_queue.pending_timer);
406 	kfree(tg);
407 }
408 
409 static struct throtl_grp *
410 throtl_rb_first(struct throtl_service_queue *parent_sq)
411 {
412 	/* Service tree is empty */
413 	if (!parent_sq->nr_pending)
414 		return NULL;
415 
416 	if (!parent_sq->first_pending)
417 		parent_sq->first_pending = rb_first(&parent_sq->pending_tree);
418 
419 	if (parent_sq->first_pending)
420 		return rb_entry_tg(parent_sq->first_pending);
421 
422 	return NULL;
423 }
424 
425 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
426 {
427 	rb_erase(n, root);
428 	RB_CLEAR_NODE(n);
429 }
430 
431 static void throtl_rb_erase(struct rb_node *n,
432 			    struct throtl_service_queue *parent_sq)
433 {
434 	if (parent_sq->first_pending == n)
435 		parent_sq->first_pending = NULL;
436 	rb_erase_init(n, &parent_sq->pending_tree);
437 	--parent_sq->nr_pending;
438 }
439 
440 static void update_min_dispatch_time(struct throtl_service_queue *parent_sq)
441 {
442 	struct throtl_grp *tg;
443 
444 	tg = throtl_rb_first(parent_sq);
445 	if (!tg)
446 		return;
447 
448 	parent_sq->first_pending_disptime = tg->disptime;
449 }
450 
451 static void tg_service_queue_add(struct throtl_grp *tg)
452 {
453 	struct throtl_service_queue *parent_sq = tg->service_queue.parent_sq;
454 	struct rb_node **node = &parent_sq->pending_tree.rb_node;
455 	struct rb_node *parent = NULL;
456 	struct throtl_grp *__tg;
457 	unsigned long key = tg->disptime;
458 	int left = 1;
459 
460 	while (*node != NULL) {
461 		parent = *node;
462 		__tg = rb_entry_tg(parent);
463 
464 		if (time_before(key, __tg->disptime))
465 			node = &parent->rb_left;
466 		else {
467 			node = &parent->rb_right;
468 			left = 0;
469 		}
470 	}
471 
472 	if (left)
473 		parent_sq->first_pending = &tg->rb_node;
474 
475 	rb_link_node(&tg->rb_node, parent, node);
476 	rb_insert_color(&tg->rb_node, &parent_sq->pending_tree);
477 }
478 
479 static void __throtl_enqueue_tg(struct throtl_grp *tg)
480 {
481 	tg_service_queue_add(tg);
482 	tg->flags |= THROTL_TG_PENDING;
483 	tg->service_queue.parent_sq->nr_pending++;
484 }
485 
486 static void throtl_enqueue_tg(struct throtl_grp *tg)
487 {
488 	if (!(tg->flags & THROTL_TG_PENDING))
489 		__throtl_enqueue_tg(tg);
490 }
491 
492 static void __throtl_dequeue_tg(struct throtl_grp *tg)
493 {
494 	throtl_rb_erase(&tg->rb_node, tg->service_queue.parent_sq);
495 	tg->flags &= ~THROTL_TG_PENDING;
496 }
497 
498 static void throtl_dequeue_tg(struct throtl_grp *tg)
499 {
500 	if (tg->flags & THROTL_TG_PENDING)
501 		__throtl_dequeue_tg(tg);
502 }
503 
504 /* Call with queue lock held */
505 static void throtl_schedule_pending_timer(struct throtl_service_queue *sq,
506 					  unsigned long expires)
507 {
508 	mod_timer(&sq->pending_timer, expires);
509 	throtl_log(sq, "schedule timer. delay=%lu jiffies=%lu",
510 		   expires - jiffies, jiffies);
511 }
512 
513 /**
514  * throtl_schedule_next_dispatch - schedule the next dispatch cycle
515  * @sq: the service_queue to schedule dispatch for
516  * @force: force scheduling
517  *
518  * Arm @sq->pending_timer so that the next dispatch cycle starts on the
519  * dispatch time of the first pending child.  Returns %true if either timer
520  * is armed or there's no pending child left.  %false if the current
521  * dispatch window is still open and the caller should continue
522  * dispatching.
523  *
524  * If @force is %true, the dispatch timer is always scheduled and this
525  * function is guaranteed to return %true.  This is to be used when the
526  * caller can't dispatch itself and needs to invoke pending_timer
527  * unconditionally.  Note that forced scheduling is likely to induce short
528  * delay before dispatch starts even if @sq->first_pending_disptime is not
529  * in the future and thus shouldn't be used in hot paths.
530  */
531 static bool throtl_schedule_next_dispatch(struct throtl_service_queue *sq,
532 					  bool force)
533 {
534 	/* any pending children left? */
535 	if (!sq->nr_pending)
536 		return true;
537 
538 	update_min_dispatch_time(sq);
539 
540 	/* is the next dispatch time in the future? */
541 	if (force || time_after(sq->first_pending_disptime, jiffies)) {
542 		throtl_schedule_pending_timer(sq, sq->first_pending_disptime);
543 		return true;
544 	}
545 
546 	/* tell the caller to continue dispatching */
547 	return false;
548 }
549 
550 static inline void throtl_start_new_slice_with_credit(struct throtl_grp *tg,
551 		bool rw, unsigned long start)
552 {
553 	tg->bytes_disp[rw] = 0;
554 	tg->io_disp[rw] = 0;
555 
556 	/*
557 	 * Previous slice has expired. We must have trimmed it after last
558 	 * bio dispatch. That means since start of last slice, we never used
559 	 * that bandwidth. Do try to make use of that bandwidth while giving
560 	 * credit.
561 	 */
562 	if (time_after_eq(start, tg->slice_start[rw]))
563 		tg->slice_start[rw] = start;
564 
565 	tg->slice_end[rw] = jiffies + throtl_slice;
566 	throtl_log(&tg->service_queue,
567 		   "[%c] new slice with credit start=%lu end=%lu jiffies=%lu",
568 		   rw == READ ? 'R' : 'W', tg->slice_start[rw],
569 		   tg->slice_end[rw], jiffies);
570 }
571 
572 static inline void throtl_start_new_slice(struct throtl_grp *tg, bool rw)
573 {
574 	tg->bytes_disp[rw] = 0;
575 	tg->io_disp[rw] = 0;
576 	tg->slice_start[rw] = jiffies;
577 	tg->slice_end[rw] = jiffies + throtl_slice;
578 	throtl_log(&tg->service_queue,
579 		   "[%c] new slice start=%lu end=%lu jiffies=%lu",
580 		   rw == READ ? 'R' : 'W', tg->slice_start[rw],
581 		   tg->slice_end[rw], jiffies);
582 }
583 
584 static inline void throtl_set_slice_end(struct throtl_grp *tg, bool rw,
585 					unsigned long jiffy_end)
586 {
587 	tg->slice_end[rw] = roundup(jiffy_end, throtl_slice);
588 }
589 
590 static inline void throtl_extend_slice(struct throtl_grp *tg, bool rw,
591 				       unsigned long jiffy_end)
592 {
593 	tg->slice_end[rw] = roundup(jiffy_end, throtl_slice);
594 	throtl_log(&tg->service_queue,
595 		   "[%c] extend slice start=%lu end=%lu jiffies=%lu",
596 		   rw == READ ? 'R' : 'W', tg->slice_start[rw],
597 		   tg->slice_end[rw], jiffies);
598 }
599 
600 /* Determine if previously allocated or extended slice is complete or not */
601 static bool throtl_slice_used(struct throtl_grp *tg, bool rw)
602 {
603 	if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw]))
604 		return false;
605 
606 	return 1;
607 }
608 
609 /* Trim the used slices and adjust slice start accordingly */
610 static inline void throtl_trim_slice(struct throtl_grp *tg, bool rw)
611 {
612 	unsigned long nr_slices, time_elapsed, io_trim;
613 	u64 bytes_trim, tmp;
614 
615 	BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw]));
616 
617 	/*
618 	 * If bps are unlimited (-1), then time slice don't get
619 	 * renewed. Don't try to trim the slice if slice is used. A new
620 	 * slice will start when appropriate.
621 	 */
622 	if (throtl_slice_used(tg, rw))
623 		return;
624 
625 	/*
626 	 * A bio has been dispatched. Also adjust slice_end. It might happen
627 	 * that initially cgroup limit was very low resulting in high
628 	 * slice_end, but later limit was bumped up and bio was dispached
629 	 * sooner, then we need to reduce slice_end. A high bogus slice_end
630 	 * is bad because it does not allow new slice to start.
631 	 */
632 
633 	throtl_set_slice_end(tg, rw, jiffies + throtl_slice);
634 
635 	time_elapsed = jiffies - tg->slice_start[rw];
636 
637 	nr_slices = time_elapsed / throtl_slice;
638 
639 	if (!nr_slices)
640 		return;
641 	tmp = tg->bps[rw] * throtl_slice * nr_slices;
642 	do_div(tmp, HZ);
643 	bytes_trim = tmp;
644 
645 	io_trim = (tg->iops[rw] * throtl_slice * nr_slices)/HZ;
646 
647 	if (!bytes_trim && !io_trim)
648 		return;
649 
650 	if (tg->bytes_disp[rw] >= bytes_trim)
651 		tg->bytes_disp[rw] -= bytes_trim;
652 	else
653 		tg->bytes_disp[rw] = 0;
654 
655 	if (tg->io_disp[rw] >= io_trim)
656 		tg->io_disp[rw] -= io_trim;
657 	else
658 		tg->io_disp[rw] = 0;
659 
660 	tg->slice_start[rw] += nr_slices * throtl_slice;
661 
662 	throtl_log(&tg->service_queue,
663 		   "[%c] trim slice nr=%lu bytes=%llu io=%lu start=%lu end=%lu jiffies=%lu",
664 		   rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim,
665 		   tg->slice_start[rw], tg->slice_end[rw], jiffies);
666 }
667 
668 static bool tg_with_in_iops_limit(struct throtl_grp *tg, struct bio *bio,
669 				  unsigned long *wait)
670 {
671 	bool rw = bio_data_dir(bio);
672 	unsigned int io_allowed;
673 	unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
674 	u64 tmp;
675 
676 	jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
677 
678 	/* Slice has just started. Consider one slice interval */
679 	if (!jiffy_elapsed)
680 		jiffy_elapsed_rnd = throtl_slice;
681 
682 	jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice);
683 
684 	/*
685 	 * jiffy_elapsed_rnd should not be a big value as minimum iops can be
686 	 * 1 then at max jiffy elapsed should be equivalent of 1 second as we
687 	 * will allow dispatch after 1 second and after that slice should
688 	 * have been trimmed.
689 	 */
690 
691 	tmp = (u64)tg->iops[rw] * jiffy_elapsed_rnd;
692 	do_div(tmp, HZ);
693 
694 	if (tmp > UINT_MAX)
695 		io_allowed = UINT_MAX;
696 	else
697 		io_allowed = tmp;
698 
699 	if (tg->io_disp[rw] + 1 <= io_allowed) {
700 		if (wait)
701 			*wait = 0;
702 		return true;
703 	}
704 
705 	/* Calc approx time to dispatch */
706 	jiffy_wait = ((tg->io_disp[rw] + 1) * HZ)/tg->iops[rw] + 1;
707 
708 	if (jiffy_wait > jiffy_elapsed)
709 		jiffy_wait = jiffy_wait - jiffy_elapsed;
710 	else
711 		jiffy_wait = 1;
712 
713 	if (wait)
714 		*wait = jiffy_wait;
715 	return 0;
716 }
717 
718 static bool tg_with_in_bps_limit(struct throtl_grp *tg, struct bio *bio,
719 				 unsigned long *wait)
720 {
721 	bool rw = bio_data_dir(bio);
722 	u64 bytes_allowed, extra_bytes, tmp;
723 	unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
724 
725 	jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
726 
727 	/* Slice has just started. Consider one slice interval */
728 	if (!jiffy_elapsed)
729 		jiffy_elapsed_rnd = throtl_slice;
730 
731 	jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice);
732 
733 	tmp = tg->bps[rw] * jiffy_elapsed_rnd;
734 	do_div(tmp, HZ);
735 	bytes_allowed = tmp;
736 
737 	if (tg->bytes_disp[rw] + bio->bi_iter.bi_size <= bytes_allowed) {
738 		if (wait)
739 			*wait = 0;
740 		return true;
741 	}
742 
743 	/* Calc approx time to dispatch */
744 	extra_bytes = tg->bytes_disp[rw] + bio->bi_iter.bi_size - bytes_allowed;
745 	jiffy_wait = div64_u64(extra_bytes * HZ, tg->bps[rw]);
746 
747 	if (!jiffy_wait)
748 		jiffy_wait = 1;
749 
750 	/*
751 	 * This wait time is without taking into consideration the rounding
752 	 * up we did. Add that time also.
753 	 */
754 	jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed);
755 	if (wait)
756 		*wait = jiffy_wait;
757 	return 0;
758 }
759 
760 /*
761  * Returns whether one can dispatch a bio or not. Also returns approx number
762  * of jiffies to wait before this bio is with-in IO rate and can be dispatched
763  */
764 static bool tg_may_dispatch(struct throtl_grp *tg, struct bio *bio,
765 			    unsigned long *wait)
766 {
767 	bool rw = bio_data_dir(bio);
768 	unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0;
769 
770 	/*
771  	 * Currently whole state machine of group depends on first bio
772 	 * queued in the group bio list. So one should not be calling
773 	 * this function with a different bio if there are other bios
774 	 * queued.
775 	 */
776 	BUG_ON(tg->service_queue.nr_queued[rw] &&
777 	       bio != throtl_peek_queued(&tg->service_queue.queued[rw]));
778 
779 	/* If tg->bps = -1, then BW is unlimited */
780 	if (tg->bps[rw] == -1 && tg->iops[rw] == -1) {
781 		if (wait)
782 			*wait = 0;
783 		return true;
784 	}
785 
786 	/*
787 	 * If previous slice expired, start a new one otherwise renew/extend
788 	 * existing slice to make sure it is at least throtl_slice interval
789 	 * long since now.
790 	 */
791 	if (throtl_slice_used(tg, rw))
792 		throtl_start_new_slice(tg, rw);
793 	else {
794 		if (time_before(tg->slice_end[rw], jiffies + throtl_slice))
795 			throtl_extend_slice(tg, rw, jiffies + throtl_slice);
796 	}
797 
798 	if (tg_with_in_bps_limit(tg, bio, &bps_wait) &&
799 	    tg_with_in_iops_limit(tg, bio, &iops_wait)) {
800 		if (wait)
801 			*wait = 0;
802 		return 1;
803 	}
804 
805 	max_wait = max(bps_wait, iops_wait);
806 
807 	if (wait)
808 		*wait = max_wait;
809 
810 	if (time_before(tg->slice_end[rw], jiffies + max_wait))
811 		throtl_extend_slice(tg, rw, jiffies + max_wait);
812 
813 	return 0;
814 }
815 
816 static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio)
817 {
818 	bool rw = bio_data_dir(bio);
819 
820 	/* Charge the bio to the group */
821 	tg->bytes_disp[rw] += bio->bi_iter.bi_size;
822 	tg->io_disp[rw]++;
823 
824 	/*
825 	 * REQ_THROTTLED is used to prevent the same bio to be throttled
826 	 * more than once as a throttled bio will go through blk-throtl the
827 	 * second time when it eventually gets issued.  Set it when a bio
828 	 * is being charged to a tg.
829 	 */
830 	if (!(bio->bi_rw & REQ_THROTTLED))
831 		bio->bi_rw |= REQ_THROTTLED;
832 }
833 
834 /**
835  * throtl_add_bio_tg - add a bio to the specified throtl_grp
836  * @bio: bio to add
837  * @qn: qnode to use
838  * @tg: the target throtl_grp
839  *
840  * Add @bio to @tg's service_queue using @qn.  If @qn is not specified,
841  * tg->qnode_on_self[] is used.
842  */
843 static void throtl_add_bio_tg(struct bio *bio, struct throtl_qnode *qn,
844 			      struct throtl_grp *tg)
845 {
846 	struct throtl_service_queue *sq = &tg->service_queue;
847 	bool rw = bio_data_dir(bio);
848 
849 	if (!qn)
850 		qn = &tg->qnode_on_self[rw];
851 
852 	/*
853 	 * If @tg doesn't currently have any bios queued in the same
854 	 * direction, queueing @bio can change when @tg should be
855 	 * dispatched.  Mark that @tg was empty.  This is automatically
856 	 * cleaered on the next tg_update_disptime().
857 	 */
858 	if (!sq->nr_queued[rw])
859 		tg->flags |= THROTL_TG_WAS_EMPTY;
860 
861 	throtl_qnode_add_bio(bio, qn, &sq->queued[rw]);
862 
863 	sq->nr_queued[rw]++;
864 	throtl_enqueue_tg(tg);
865 }
866 
867 static void tg_update_disptime(struct throtl_grp *tg)
868 {
869 	struct throtl_service_queue *sq = &tg->service_queue;
870 	unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime;
871 	struct bio *bio;
872 
873 	if ((bio = throtl_peek_queued(&sq->queued[READ])))
874 		tg_may_dispatch(tg, bio, &read_wait);
875 
876 	if ((bio = throtl_peek_queued(&sq->queued[WRITE])))
877 		tg_may_dispatch(tg, bio, &write_wait);
878 
879 	min_wait = min(read_wait, write_wait);
880 	disptime = jiffies + min_wait;
881 
882 	/* Update dispatch time */
883 	throtl_dequeue_tg(tg);
884 	tg->disptime = disptime;
885 	throtl_enqueue_tg(tg);
886 
887 	/* see throtl_add_bio_tg() */
888 	tg->flags &= ~THROTL_TG_WAS_EMPTY;
889 }
890 
891 static void start_parent_slice_with_credit(struct throtl_grp *child_tg,
892 					struct throtl_grp *parent_tg, bool rw)
893 {
894 	if (throtl_slice_used(parent_tg, rw)) {
895 		throtl_start_new_slice_with_credit(parent_tg, rw,
896 				child_tg->slice_start[rw]);
897 	}
898 
899 }
900 
901 static void tg_dispatch_one_bio(struct throtl_grp *tg, bool rw)
902 {
903 	struct throtl_service_queue *sq = &tg->service_queue;
904 	struct throtl_service_queue *parent_sq = sq->parent_sq;
905 	struct throtl_grp *parent_tg = sq_to_tg(parent_sq);
906 	struct throtl_grp *tg_to_put = NULL;
907 	struct bio *bio;
908 
909 	/*
910 	 * @bio is being transferred from @tg to @parent_sq.  Popping a bio
911 	 * from @tg may put its reference and @parent_sq might end up
912 	 * getting released prematurely.  Remember the tg to put and put it
913 	 * after @bio is transferred to @parent_sq.
914 	 */
915 	bio = throtl_pop_queued(&sq->queued[rw], &tg_to_put);
916 	sq->nr_queued[rw]--;
917 
918 	throtl_charge_bio(tg, bio);
919 
920 	/*
921 	 * If our parent is another tg, we just need to transfer @bio to
922 	 * the parent using throtl_add_bio_tg().  If our parent is
923 	 * @td->service_queue, @bio is ready to be issued.  Put it on its
924 	 * bio_lists[] and decrease total number queued.  The caller is
925 	 * responsible for issuing these bios.
926 	 */
927 	if (parent_tg) {
928 		throtl_add_bio_tg(bio, &tg->qnode_on_parent[rw], parent_tg);
929 		start_parent_slice_with_credit(tg, parent_tg, rw);
930 	} else {
931 		throtl_qnode_add_bio(bio, &tg->qnode_on_parent[rw],
932 				     &parent_sq->queued[rw]);
933 		BUG_ON(tg->td->nr_queued[rw] <= 0);
934 		tg->td->nr_queued[rw]--;
935 	}
936 
937 	throtl_trim_slice(tg, rw);
938 
939 	if (tg_to_put)
940 		blkg_put(tg_to_blkg(tg_to_put));
941 }
942 
943 static int throtl_dispatch_tg(struct throtl_grp *tg)
944 {
945 	struct throtl_service_queue *sq = &tg->service_queue;
946 	unsigned int nr_reads = 0, nr_writes = 0;
947 	unsigned int max_nr_reads = throtl_grp_quantum*3/4;
948 	unsigned int max_nr_writes = throtl_grp_quantum - max_nr_reads;
949 	struct bio *bio;
950 
951 	/* Try to dispatch 75% READS and 25% WRITES */
952 
953 	while ((bio = throtl_peek_queued(&sq->queued[READ])) &&
954 	       tg_may_dispatch(tg, bio, NULL)) {
955 
956 		tg_dispatch_one_bio(tg, bio_data_dir(bio));
957 		nr_reads++;
958 
959 		if (nr_reads >= max_nr_reads)
960 			break;
961 	}
962 
963 	while ((bio = throtl_peek_queued(&sq->queued[WRITE])) &&
964 	       tg_may_dispatch(tg, bio, NULL)) {
965 
966 		tg_dispatch_one_bio(tg, bio_data_dir(bio));
967 		nr_writes++;
968 
969 		if (nr_writes >= max_nr_writes)
970 			break;
971 	}
972 
973 	return nr_reads + nr_writes;
974 }
975 
976 static int throtl_select_dispatch(struct throtl_service_queue *parent_sq)
977 {
978 	unsigned int nr_disp = 0;
979 
980 	while (1) {
981 		struct throtl_grp *tg = throtl_rb_first(parent_sq);
982 		struct throtl_service_queue *sq = &tg->service_queue;
983 
984 		if (!tg)
985 			break;
986 
987 		if (time_before(jiffies, tg->disptime))
988 			break;
989 
990 		throtl_dequeue_tg(tg);
991 
992 		nr_disp += throtl_dispatch_tg(tg);
993 
994 		if (sq->nr_queued[0] || sq->nr_queued[1])
995 			tg_update_disptime(tg);
996 
997 		if (nr_disp >= throtl_quantum)
998 			break;
999 	}
1000 
1001 	return nr_disp;
1002 }
1003 
1004 /**
1005  * throtl_pending_timer_fn - timer function for service_queue->pending_timer
1006  * @arg: the throtl_service_queue being serviced
1007  *
1008  * This timer is armed when a child throtl_grp with active bio's become
1009  * pending and queued on the service_queue's pending_tree and expires when
1010  * the first child throtl_grp should be dispatched.  This function
1011  * dispatches bio's from the children throtl_grps to the parent
1012  * service_queue.
1013  *
1014  * If the parent's parent is another throtl_grp, dispatching is propagated
1015  * by either arming its pending_timer or repeating dispatch directly.  If
1016  * the top-level service_tree is reached, throtl_data->dispatch_work is
1017  * kicked so that the ready bio's are issued.
1018  */
1019 static void throtl_pending_timer_fn(unsigned long arg)
1020 {
1021 	struct throtl_service_queue *sq = (void *)arg;
1022 	struct throtl_grp *tg = sq_to_tg(sq);
1023 	struct throtl_data *td = sq_to_td(sq);
1024 	struct request_queue *q = td->queue;
1025 	struct throtl_service_queue *parent_sq;
1026 	bool dispatched;
1027 	int ret;
1028 
1029 	spin_lock_irq(q->queue_lock);
1030 again:
1031 	parent_sq = sq->parent_sq;
1032 	dispatched = false;
1033 
1034 	while (true) {
1035 		throtl_log(sq, "dispatch nr_queued=%u read=%u write=%u",
1036 			   sq->nr_queued[READ] + sq->nr_queued[WRITE],
1037 			   sq->nr_queued[READ], sq->nr_queued[WRITE]);
1038 
1039 		ret = throtl_select_dispatch(sq);
1040 		if (ret) {
1041 			throtl_log(sq, "bios disp=%u", ret);
1042 			dispatched = true;
1043 		}
1044 
1045 		if (throtl_schedule_next_dispatch(sq, false))
1046 			break;
1047 
1048 		/* this dispatch windows is still open, relax and repeat */
1049 		spin_unlock_irq(q->queue_lock);
1050 		cpu_relax();
1051 		spin_lock_irq(q->queue_lock);
1052 	}
1053 
1054 	if (!dispatched)
1055 		goto out_unlock;
1056 
1057 	if (parent_sq) {
1058 		/* @parent_sq is another throl_grp, propagate dispatch */
1059 		if (tg->flags & THROTL_TG_WAS_EMPTY) {
1060 			tg_update_disptime(tg);
1061 			if (!throtl_schedule_next_dispatch(parent_sq, false)) {
1062 				/* window is already open, repeat dispatching */
1063 				sq = parent_sq;
1064 				tg = sq_to_tg(sq);
1065 				goto again;
1066 			}
1067 		}
1068 	} else {
1069 		/* reached the top-level, queue issueing */
1070 		queue_work(kthrotld_workqueue, &td->dispatch_work);
1071 	}
1072 out_unlock:
1073 	spin_unlock_irq(q->queue_lock);
1074 }
1075 
1076 /**
1077  * blk_throtl_dispatch_work_fn - work function for throtl_data->dispatch_work
1078  * @work: work item being executed
1079  *
1080  * This function is queued for execution when bio's reach the bio_lists[]
1081  * of throtl_data->service_queue.  Those bio's are ready and issued by this
1082  * function.
1083  */
1084 static void blk_throtl_dispatch_work_fn(struct work_struct *work)
1085 {
1086 	struct throtl_data *td = container_of(work, struct throtl_data,
1087 					      dispatch_work);
1088 	struct throtl_service_queue *td_sq = &td->service_queue;
1089 	struct request_queue *q = td->queue;
1090 	struct bio_list bio_list_on_stack;
1091 	struct bio *bio;
1092 	struct blk_plug plug;
1093 	int rw;
1094 
1095 	bio_list_init(&bio_list_on_stack);
1096 
1097 	spin_lock_irq(q->queue_lock);
1098 	for (rw = READ; rw <= WRITE; rw++)
1099 		while ((bio = throtl_pop_queued(&td_sq->queued[rw], NULL)))
1100 			bio_list_add(&bio_list_on_stack, bio);
1101 	spin_unlock_irq(q->queue_lock);
1102 
1103 	if (!bio_list_empty(&bio_list_on_stack)) {
1104 		blk_start_plug(&plug);
1105 		while((bio = bio_list_pop(&bio_list_on_stack)))
1106 			generic_make_request(bio);
1107 		blk_finish_plug(&plug);
1108 	}
1109 }
1110 
1111 static u64 tg_prfill_conf_u64(struct seq_file *sf, struct blkg_policy_data *pd,
1112 			      int off)
1113 {
1114 	struct throtl_grp *tg = pd_to_tg(pd);
1115 	u64 v = *(u64 *)((void *)tg + off);
1116 
1117 	if (v == -1)
1118 		return 0;
1119 	return __blkg_prfill_u64(sf, pd, v);
1120 }
1121 
1122 static u64 tg_prfill_conf_uint(struct seq_file *sf, struct blkg_policy_data *pd,
1123 			       int off)
1124 {
1125 	struct throtl_grp *tg = pd_to_tg(pd);
1126 	unsigned int v = *(unsigned int *)((void *)tg + off);
1127 
1128 	if (v == -1)
1129 		return 0;
1130 	return __blkg_prfill_u64(sf, pd, v);
1131 }
1132 
1133 static int tg_print_conf_u64(struct seq_file *sf, void *v)
1134 {
1135 	blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_u64,
1136 			  &blkcg_policy_throtl, seq_cft(sf)->private, false);
1137 	return 0;
1138 }
1139 
1140 static int tg_print_conf_uint(struct seq_file *sf, void *v)
1141 {
1142 	blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_uint,
1143 			  &blkcg_policy_throtl, seq_cft(sf)->private, false);
1144 	return 0;
1145 }
1146 
1147 static void tg_conf_updated(struct throtl_grp *tg)
1148 {
1149 	struct throtl_service_queue *sq = &tg->service_queue;
1150 	struct cgroup_subsys_state *pos_css;
1151 	struct blkcg_gq *blkg;
1152 
1153 	throtl_log(&tg->service_queue,
1154 		   "limit change rbps=%llu wbps=%llu riops=%u wiops=%u",
1155 		   tg->bps[READ], tg->bps[WRITE],
1156 		   tg->iops[READ], tg->iops[WRITE]);
1157 
1158 	/*
1159 	 * Update has_rules[] flags for the updated tg's subtree.  A tg is
1160 	 * considered to have rules if either the tg itself or any of its
1161 	 * ancestors has rules.  This identifies groups without any
1162 	 * restrictions in the whole hierarchy and allows them to bypass
1163 	 * blk-throttle.
1164 	 */
1165 	blkg_for_each_descendant_pre(blkg, pos_css, tg_to_blkg(tg))
1166 		tg_update_has_rules(blkg_to_tg(blkg));
1167 
1168 	/*
1169 	 * We're already holding queue_lock and know @tg is valid.  Let's
1170 	 * apply the new config directly.
1171 	 *
1172 	 * Restart the slices for both READ and WRITES. It might happen
1173 	 * that a group's limit are dropped suddenly and we don't want to
1174 	 * account recently dispatched IO with new low rate.
1175 	 */
1176 	throtl_start_new_slice(tg, 0);
1177 	throtl_start_new_slice(tg, 1);
1178 
1179 	if (tg->flags & THROTL_TG_PENDING) {
1180 		tg_update_disptime(tg);
1181 		throtl_schedule_next_dispatch(sq->parent_sq, true);
1182 	}
1183 }
1184 
1185 static ssize_t tg_set_conf(struct kernfs_open_file *of,
1186 			   char *buf, size_t nbytes, loff_t off, bool is_u64)
1187 {
1188 	struct blkcg *blkcg = css_to_blkcg(of_css(of));
1189 	struct blkg_conf_ctx ctx;
1190 	struct throtl_grp *tg;
1191 	int ret;
1192 	u64 v;
1193 
1194 	ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx);
1195 	if (ret)
1196 		return ret;
1197 
1198 	ret = -EINVAL;
1199 	if (sscanf(ctx.body, "%llu", &v) != 1)
1200 		goto out_finish;
1201 	if (!v)
1202 		v = -1;
1203 
1204 	tg = blkg_to_tg(ctx.blkg);
1205 
1206 	if (is_u64)
1207 		*(u64 *)((void *)tg + of_cft(of)->private) = v;
1208 	else
1209 		*(unsigned int *)((void *)tg + of_cft(of)->private) = v;
1210 
1211 	tg_conf_updated(tg);
1212 	ret = 0;
1213 out_finish:
1214 	blkg_conf_finish(&ctx);
1215 	return ret ?: nbytes;
1216 }
1217 
1218 static ssize_t tg_set_conf_u64(struct kernfs_open_file *of,
1219 			       char *buf, size_t nbytes, loff_t off)
1220 {
1221 	return tg_set_conf(of, buf, nbytes, off, true);
1222 }
1223 
1224 static ssize_t tg_set_conf_uint(struct kernfs_open_file *of,
1225 				char *buf, size_t nbytes, loff_t off)
1226 {
1227 	return tg_set_conf(of, buf, nbytes, off, false);
1228 }
1229 
1230 static struct cftype throtl_legacy_files[] = {
1231 	{
1232 		.name = "throttle.read_bps_device",
1233 		.private = offsetof(struct throtl_grp, bps[READ]),
1234 		.seq_show = tg_print_conf_u64,
1235 		.write = tg_set_conf_u64,
1236 	},
1237 	{
1238 		.name = "throttle.write_bps_device",
1239 		.private = offsetof(struct throtl_grp, bps[WRITE]),
1240 		.seq_show = tg_print_conf_u64,
1241 		.write = tg_set_conf_u64,
1242 	},
1243 	{
1244 		.name = "throttle.read_iops_device",
1245 		.private = offsetof(struct throtl_grp, iops[READ]),
1246 		.seq_show = tg_print_conf_uint,
1247 		.write = tg_set_conf_uint,
1248 	},
1249 	{
1250 		.name = "throttle.write_iops_device",
1251 		.private = offsetof(struct throtl_grp, iops[WRITE]),
1252 		.seq_show = tg_print_conf_uint,
1253 		.write = tg_set_conf_uint,
1254 	},
1255 	{
1256 		.name = "throttle.io_service_bytes",
1257 		.private = (unsigned long)&blkcg_policy_throtl,
1258 		.seq_show = blkg_print_stat_bytes,
1259 	},
1260 	{
1261 		.name = "throttle.io_serviced",
1262 		.private = (unsigned long)&blkcg_policy_throtl,
1263 		.seq_show = blkg_print_stat_ios,
1264 	},
1265 	{ }	/* terminate */
1266 };
1267 
1268 static u64 tg_prfill_max(struct seq_file *sf, struct blkg_policy_data *pd,
1269 			 int off)
1270 {
1271 	struct throtl_grp *tg = pd_to_tg(pd);
1272 	const char *dname = blkg_dev_name(pd->blkg);
1273 	char bufs[4][21] = { "max", "max", "max", "max" };
1274 
1275 	if (!dname)
1276 		return 0;
1277 	if (tg->bps[READ] == -1 && tg->bps[WRITE] == -1 &&
1278 	    tg->iops[READ] == -1 && tg->iops[WRITE] == -1)
1279 		return 0;
1280 
1281 	if (tg->bps[READ] != -1)
1282 		snprintf(bufs[0], sizeof(bufs[0]), "%llu", tg->bps[READ]);
1283 	if (tg->bps[WRITE] != -1)
1284 		snprintf(bufs[1], sizeof(bufs[1]), "%llu", tg->bps[WRITE]);
1285 	if (tg->iops[READ] != -1)
1286 		snprintf(bufs[2], sizeof(bufs[2]), "%u", tg->iops[READ]);
1287 	if (tg->iops[WRITE] != -1)
1288 		snprintf(bufs[3], sizeof(bufs[3]), "%u", tg->iops[WRITE]);
1289 
1290 	seq_printf(sf, "%s rbps=%s wbps=%s riops=%s wiops=%s\n",
1291 		   dname, bufs[0], bufs[1], bufs[2], bufs[3]);
1292 	return 0;
1293 }
1294 
1295 static int tg_print_max(struct seq_file *sf, void *v)
1296 {
1297 	blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_max,
1298 			  &blkcg_policy_throtl, seq_cft(sf)->private, false);
1299 	return 0;
1300 }
1301 
1302 static ssize_t tg_set_max(struct kernfs_open_file *of,
1303 			  char *buf, size_t nbytes, loff_t off)
1304 {
1305 	struct blkcg *blkcg = css_to_blkcg(of_css(of));
1306 	struct blkg_conf_ctx ctx;
1307 	struct throtl_grp *tg;
1308 	u64 v[4];
1309 	int ret;
1310 
1311 	ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx);
1312 	if (ret)
1313 		return ret;
1314 
1315 	tg = blkg_to_tg(ctx.blkg);
1316 
1317 	v[0] = tg->bps[READ];
1318 	v[1] = tg->bps[WRITE];
1319 	v[2] = tg->iops[READ];
1320 	v[3] = tg->iops[WRITE];
1321 
1322 	while (true) {
1323 		char tok[27];	/* wiops=18446744073709551616 */
1324 		char *p;
1325 		u64 val = -1;
1326 		int len;
1327 
1328 		if (sscanf(ctx.body, "%26s%n", tok, &len) != 1)
1329 			break;
1330 		if (tok[0] == '\0')
1331 			break;
1332 		ctx.body += len;
1333 
1334 		ret = -EINVAL;
1335 		p = tok;
1336 		strsep(&p, "=");
1337 		if (!p || (sscanf(p, "%llu", &val) != 1 && strcmp(p, "max")))
1338 			goto out_finish;
1339 
1340 		ret = -ERANGE;
1341 		if (!val)
1342 			goto out_finish;
1343 
1344 		ret = -EINVAL;
1345 		if (!strcmp(tok, "rbps"))
1346 			v[0] = val;
1347 		else if (!strcmp(tok, "wbps"))
1348 			v[1] = val;
1349 		else if (!strcmp(tok, "riops"))
1350 			v[2] = min_t(u64, val, UINT_MAX);
1351 		else if (!strcmp(tok, "wiops"))
1352 			v[3] = min_t(u64, val, UINT_MAX);
1353 		else
1354 			goto out_finish;
1355 	}
1356 
1357 	tg->bps[READ] = v[0];
1358 	tg->bps[WRITE] = v[1];
1359 	tg->iops[READ] = v[2];
1360 	tg->iops[WRITE] = v[3];
1361 
1362 	tg_conf_updated(tg);
1363 	ret = 0;
1364 out_finish:
1365 	blkg_conf_finish(&ctx);
1366 	return ret ?: nbytes;
1367 }
1368 
1369 static struct cftype throtl_files[] = {
1370 	{
1371 		.name = "max",
1372 		.flags = CFTYPE_NOT_ON_ROOT,
1373 		.seq_show = tg_print_max,
1374 		.write = tg_set_max,
1375 	},
1376 	{ }	/* terminate */
1377 };
1378 
1379 static void throtl_shutdown_wq(struct request_queue *q)
1380 {
1381 	struct throtl_data *td = q->td;
1382 
1383 	cancel_work_sync(&td->dispatch_work);
1384 }
1385 
1386 static struct blkcg_policy blkcg_policy_throtl = {
1387 	.dfl_cftypes		= throtl_files,
1388 	.legacy_cftypes		= throtl_legacy_files,
1389 
1390 	.pd_alloc_fn		= throtl_pd_alloc,
1391 	.pd_init_fn		= throtl_pd_init,
1392 	.pd_online_fn		= throtl_pd_online,
1393 	.pd_free_fn		= throtl_pd_free,
1394 };
1395 
1396 bool blk_throtl_bio(struct request_queue *q, struct blkcg_gq *blkg,
1397 		    struct bio *bio)
1398 {
1399 	struct throtl_qnode *qn = NULL;
1400 	struct throtl_grp *tg = blkg_to_tg(blkg ?: q->root_blkg);
1401 	struct throtl_service_queue *sq;
1402 	bool rw = bio_data_dir(bio);
1403 	bool throttled = false;
1404 
1405 	WARN_ON_ONCE(!rcu_read_lock_held());
1406 
1407 	/* see throtl_charge_bio() */
1408 	if ((bio->bi_rw & REQ_THROTTLED) || !tg->has_rules[rw])
1409 		goto out;
1410 
1411 	spin_lock_irq(q->queue_lock);
1412 
1413 	if (unlikely(blk_queue_bypass(q)))
1414 		goto out_unlock;
1415 
1416 	sq = &tg->service_queue;
1417 
1418 	while (true) {
1419 		/* throtl is FIFO - if bios are already queued, should queue */
1420 		if (sq->nr_queued[rw])
1421 			break;
1422 
1423 		/* if above limits, break to queue */
1424 		if (!tg_may_dispatch(tg, bio, NULL))
1425 			break;
1426 
1427 		/* within limits, let's charge and dispatch directly */
1428 		throtl_charge_bio(tg, bio);
1429 
1430 		/*
1431 		 * We need to trim slice even when bios are not being queued
1432 		 * otherwise it might happen that a bio is not queued for
1433 		 * a long time and slice keeps on extending and trim is not
1434 		 * called for a long time. Now if limits are reduced suddenly
1435 		 * we take into account all the IO dispatched so far at new
1436 		 * low rate and * newly queued IO gets a really long dispatch
1437 		 * time.
1438 		 *
1439 		 * So keep on trimming slice even if bio is not queued.
1440 		 */
1441 		throtl_trim_slice(tg, rw);
1442 
1443 		/*
1444 		 * @bio passed through this layer without being throttled.
1445 		 * Climb up the ladder.  If we''re already at the top, it
1446 		 * can be executed directly.
1447 		 */
1448 		qn = &tg->qnode_on_parent[rw];
1449 		sq = sq->parent_sq;
1450 		tg = sq_to_tg(sq);
1451 		if (!tg)
1452 			goto out_unlock;
1453 	}
1454 
1455 	/* out-of-limit, queue to @tg */
1456 	throtl_log(sq, "[%c] bio. bdisp=%llu sz=%u bps=%llu iodisp=%u iops=%u queued=%d/%d",
1457 		   rw == READ ? 'R' : 'W',
1458 		   tg->bytes_disp[rw], bio->bi_iter.bi_size, tg->bps[rw],
1459 		   tg->io_disp[rw], tg->iops[rw],
1460 		   sq->nr_queued[READ], sq->nr_queued[WRITE]);
1461 
1462 	bio_associate_current(bio);
1463 	tg->td->nr_queued[rw]++;
1464 	throtl_add_bio_tg(bio, qn, tg);
1465 	throttled = true;
1466 
1467 	/*
1468 	 * Update @tg's dispatch time and force schedule dispatch if @tg
1469 	 * was empty before @bio.  The forced scheduling isn't likely to
1470 	 * cause undue delay as @bio is likely to be dispatched directly if
1471 	 * its @tg's disptime is not in the future.
1472 	 */
1473 	if (tg->flags & THROTL_TG_WAS_EMPTY) {
1474 		tg_update_disptime(tg);
1475 		throtl_schedule_next_dispatch(tg->service_queue.parent_sq, true);
1476 	}
1477 
1478 out_unlock:
1479 	spin_unlock_irq(q->queue_lock);
1480 out:
1481 	/*
1482 	 * As multiple blk-throtls may stack in the same issue path, we
1483 	 * don't want bios to leave with the flag set.  Clear the flag if
1484 	 * being issued.
1485 	 */
1486 	if (!throttled)
1487 		bio->bi_rw &= ~REQ_THROTTLED;
1488 	return throttled;
1489 }
1490 
1491 /*
1492  * Dispatch all bios from all children tg's queued on @parent_sq.  On
1493  * return, @parent_sq is guaranteed to not have any active children tg's
1494  * and all bios from previously active tg's are on @parent_sq->bio_lists[].
1495  */
1496 static void tg_drain_bios(struct throtl_service_queue *parent_sq)
1497 {
1498 	struct throtl_grp *tg;
1499 
1500 	while ((tg = throtl_rb_first(parent_sq))) {
1501 		struct throtl_service_queue *sq = &tg->service_queue;
1502 		struct bio *bio;
1503 
1504 		throtl_dequeue_tg(tg);
1505 
1506 		while ((bio = throtl_peek_queued(&sq->queued[READ])))
1507 			tg_dispatch_one_bio(tg, bio_data_dir(bio));
1508 		while ((bio = throtl_peek_queued(&sq->queued[WRITE])))
1509 			tg_dispatch_one_bio(tg, bio_data_dir(bio));
1510 	}
1511 }
1512 
1513 /**
1514  * blk_throtl_drain - drain throttled bios
1515  * @q: request_queue to drain throttled bios for
1516  *
1517  * Dispatch all currently throttled bios on @q through ->make_request_fn().
1518  */
1519 void blk_throtl_drain(struct request_queue *q)
1520 	__releases(q->queue_lock) __acquires(q->queue_lock)
1521 {
1522 	struct throtl_data *td = q->td;
1523 	struct blkcg_gq *blkg;
1524 	struct cgroup_subsys_state *pos_css;
1525 	struct bio *bio;
1526 	int rw;
1527 
1528 	queue_lockdep_assert_held(q);
1529 	rcu_read_lock();
1530 
1531 	/*
1532 	 * Drain each tg while doing post-order walk on the blkg tree, so
1533 	 * that all bios are propagated to td->service_queue.  It'd be
1534 	 * better to walk service_queue tree directly but blkg walk is
1535 	 * easier.
1536 	 */
1537 	blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg)
1538 		tg_drain_bios(&blkg_to_tg(blkg)->service_queue);
1539 
1540 	/* finally, transfer bios from top-level tg's into the td */
1541 	tg_drain_bios(&td->service_queue);
1542 
1543 	rcu_read_unlock();
1544 	spin_unlock_irq(q->queue_lock);
1545 
1546 	/* all bios now should be in td->service_queue, issue them */
1547 	for (rw = READ; rw <= WRITE; rw++)
1548 		while ((bio = throtl_pop_queued(&td->service_queue.queued[rw],
1549 						NULL)))
1550 			generic_make_request(bio);
1551 
1552 	spin_lock_irq(q->queue_lock);
1553 }
1554 
1555 int blk_throtl_init(struct request_queue *q)
1556 {
1557 	struct throtl_data *td;
1558 	int ret;
1559 
1560 	td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node);
1561 	if (!td)
1562 		return -ENOMEM;
1563 
1564 	INIT_WORK(&td->dispatch_work, blk_throtl_dispatch_work_fn);
1565 	throtl_service_queue_init(&td->service_queue);
1566 
1567 	q->td = td;
1568 	td->queue = q;
1569 
1570 	/* activate policy */
1571 	ret = blkcg_activate_policy(q, &blkcg_policy_throtl);
1572 	if (ret)
1573 		kfree(td);
1574 	return ret;
1575 }
1576 
1577 void blk_throtl_exit(struct request_queue *q)
1578 {
1579 	BUG_ON(!q->td);
1580 	throtl_shutdown_wq(q);
1581 	blkcg_deactivate_policy(q, &blkcg_policy_throtl);
1582 	kfree(q->td);
1583 }
1584 
1585 static int __init throtl_init(void)
1586 {
1587 	kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0);
1588 	if (!kthrotld_workqueue)
1589 		panic("Failed to create kthrotld\n");
1590 
1591 	return blkcg_policy_register(&blkcg_policy_throtl);
1592 }
1593 
1594 module_init(throtl_init);
1595