2 * Interface for controlling IO bandwidth on a request queue
4 * Copyright (C) 2010 Vivek Goyal <vgoyal@redhat.com>
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>
15 /* Max dispatch from a group in 1 round */
16 static int throtl_grp_quantum
= 8;
18 /* Total max dispatch from all groups in one round */
19 static int throtl_quantum
= 32;
21 /* Throttling is performed over 100ms slice and after that slice is renewed */
22 static unsigned long throtl_slice
= HZ
/10; /* 100 ms */
24 static struct blkcg_policy blkcg_policy_throtl
;
26 /* A workqueue to queue throttle related work */
27 static struct workqueue_struct
*kthrotld_workqueue
;
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.
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.
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.
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.
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 */
58 struct throtl_service_queue
{
59 struct throtl_service_queue
*parent_sq
; /* the parent service_queue */
62 * Bios queued directly to this service_queue or dispatched from
63 * children throtl_grp's.
65 struct list_head queued
[2]; /* throtl_qnode [READ/WRITE] */
66 unsigned int nr_queued
[2]; /* number of queued bios */
69 * RB tree of active children throtl_grp's, which are sorted by
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 */
80 THROTL_TG_PENDING
= 1 << 0, /* on parent's pending tree */
81 THROTL_TG_WAS_EMPTY
= 1 << 1, /* bio_lists[] became non-empty */
84 #define rb_entry_tg(node) rb_entry((node), struct throtl_grp, rb_node)
87 /* must be the first member */
88 struct blkg_policy_data pd
;
90 /* active throtl group service_queue member */
91 struct rb_node rb_node
;
93 /* throtl_data this group belongs to */
94 struct throtl_data
*td
;
96 /* this group's service queue */
97 struct throtl_service_queue service_queue
;
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
107 struct throtl_qnode qnode_on_self
[2];
108 struct throtl_qnode qnode_on_parent
[2];
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.
115 unsigned long disptime
;
119 /* are there any throtl rules between this group and td? */
122 /* bytes per second rate limits */
126 unsigned int iops
[2];
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];
133 /* When did we start a new slice */
134 unsigned long slice_start
[2];
135 unsigned long slice_end
[2];
137 /* total bytes transferred */
138 struct blkg_rwstat service_bytes
;
139 /* total IOs serviced, post merge */
140 struct blkg_rwstat serviced
;
145 /* service tree for active throtl groups */
146 struct throtl_service_queue service_queue
;
148 struct request_queue
*queue
;
150 /* Total Number of queued bios on READ and WRITE lists */
151 unsigned int nr_queued
[2];
154 * number of total undestroyed groups
156 unsigned int nr_undestroyed_grps
;
158 /* Work for dispatching throttled bios */
159 struct work_struct dispatch_work
;
162 static void throtl_pending_timer_fn(unsigned long arg
);
164 static inline struct throtl_grp
*pd_to_tg(struct blkg_policy_data
*pd
)
166 return pd
? container_of(pd
, struct throtl_grp
, pd
) : NULL
;
169 static inline struct throtl_grp
*blkg_to_tg(struct blkcg_gq
*blkg
)
171 return pd_to_tg(blkg_to_pd(blkg
, &blkcg_policy_throtl
));
174 static inline struct blkcg_gq
*tg_to_blkg(struct throtl_grp
*tg
)
176 return pd_to_blkg(&tg
->pd
);
180 * sq_to_tg - return the throl_grp the specified service queue belongs to
181 * @sq: the throtl_service_queue of interest
183 * Return the throtl_grp @sq belongs to. If @sq is the top-level one
184 * embedded in throtl_data, %NULL is returned.
186 static struct throtl_grp
*sq_to_tg(struct throtl_service_queue
*sq
)
188 if (sq
&& sq
->parent_sq
)
189 return container_of(sq
, struct throtl_grp
, service_queue
);
195 * sq_to_td - return throtl_data the specified service queue belongs to
196 * @sq: the throtl_service_queue of interest
198 * A service_queue can be embeded in either a throtl_grp or throtl_data.
199 * Determine the associated throtl_data accordingly and return it.
201 static struct throtl_data
*sq_to_td(struct throtl_service_queue
*sq
)
203 struct throtl_grp
*tg
= sq_to_tg(sq
);
208 return container_of(sq
, struct throtl_data
, service_queue
);
212 * throtl_log - log debug message via blktrace
213 * @sq: the service_queue being reported
214 * @fmt: printf format string
217 * The messages are prefixed with "throtl BLKG_NAME" if @sq belongs to a
218 * throtl_grp; otherwise, just "throtl".
220 * TODO: this should be made a function and name formatting should happen
221 * after testing whether blktrace is enabled.
223 #define throtl_log(sq, fmt, args...) do { \
224 struct throtl_grp *__tg = sq_to_tg((sq)); \
225 struct throtl_data *__td = sq_to_td((sq)); \
231 blkg_path(tg_to_blkg(__tg), __pbuf, sizeof(__pbuf)); \
232 blk_add_trace_msg(__td->queue, "throtl %s " fmt, __pbuf, ##args); \
234 blk_add_trace_msg(__td->queue, "throtl " fmt, ##args); \
238 static void throtl_qnode_init(struct throtl_qnode
*qn
, struct throtl_grp
*tg
)
240 INIT_LIST_HEAD(&qn
->node
);
241 bio_list_init(&qn
->bios
);
246 * throtl_qnode_add_bio - add a bio to a throtl_qnode and activate it
247 * @bio: bio being added
248 * @qn: qnode to add bio to
249 * @queued: the service_queue->queued[] list @qn belongs to
251 * Add @bio to @qn and put @qn on @queued if it's not already on.
252 * @qn->tg's reference count is bumped when @qn is activated. See the
253 * comment on top of throtl_qnode definition for details.
255 static void throtl_qnode_add_bio(struct bio
*bio
, struct throtl_qnode
*qn
,
256 struct list_head
*queued
)
258 bio_list_add(&qn
->bios
, bio
);
259 if (list_empty(&qn
->node
)) {
260 list_add_tail(&qn
->node
, queued
);
261 blkg_get(tg_to_blkg(qn
->tg
));
266 * throtl_peek_queued - peek the first bio on a qnode list
267 * @queued: the qnode list to peek
269 static struct bio
*throtl_peek_queued(struct list_head
*queued
)
271 struct throtl_qnode
*qn
= list_first_entry(queued
, struct throtl_qnode
, node
);
274 if (list_empty(queued
))
277 bio
= bio_list_peek(&qn
->bios
);
283 * throtl_pop_queued - pop the first bio form a qnode list
284 * @queued: the qnode list to pop a bio from
285 * @tg_to_put: optional out argument for throtl_grp to put
287 * Pop the first bio from the qnode list @queued. After popping, the first
288 * qnode is removed from @queued if empty or moved to the end of @queued so
289 * that the popping order is round-robin.
291 * When the first qnode is removed, its associated throtl_grp should be put
292 * too. If @tg_to_put is NULL, this function automatically puts it;
293 * otherwise, *@tg_to_put is set to the throtl_grp to put and the caller is
294 * responsible for putting it.
296 static struct bio
*throtl_pop_queued(struct list_head
*queued
,
297 struct throtl_grp
**tg_to_put
)
299 struct throtl_qnode
*qn
= list_first_entry(queued
, struct throtl_qnode
, node
);
302 if (list_empty(queued
))
305 bio
= bio_list_pop(&qn
->bios
);
308 if (bio_list_empty(&qn
->bios
)) {
309 list_del_init(&qn
->node
);
313 blkg_put(tg_to_blkg(qn
->tg
));
315 list_move_tail(&qn
->node
, queued
);
321 /* init a service_queue, assumes the caller zeroed it */
322 static void throtl_service_queue_init(struct throtl_service_queue
*sq
)
324 INIT_LIST_HEAD(&sq
->queued
[0]);
325 INIT_LIST_HEAD(&sq
->queued
[1]);
326 sq
->pending_tree
= RB_ROOT
;
327 setup_timer(&sq
->pending_timer
, throtl_pending_timer_fn
,
331 static struct blkg_policy_data
*throtl_pd_alloc(gfp_t gfp
, int node
)
333 struct throtl_grp
*tg
;
336 tg
= kzalloc_node(sizeof(*tg
), gfp
, node
);
340 if (blkg_rwstat_init(&tg
->service_bytes
, gfp
) ||
341 blkg_rwstat_init(&tg
->serviced
, gfp
))
344 throtl_service_queue_init(&tg
->service_queue
);
346 for (rw
= READ
; rw
<= WRITE
; rw
++) {
347 throtl_qnode_init(&tg
->qnode_on_self
[rw
], tg
);
348 throtl_qnode_init(&tg
->qnode_on_parent
[rw
], tg
);
351 RB_CLEAR_NODE(&tg
->rb_node
);
355 tg
->iops
[WRITE
] = -1;
360 blkg_rwstat_exit(&tg
->serviced
);
361 blkg_rwstat_exit(&tg
->service_bytes
);
367 static void throtl_pd_init(struct blkg_policy_data
*pd
)
369 struct throtl_grp
*tg
= pd_to_tg(pd
);
370 struct blkcg_gq
*blkg
= tg_to_blkg(tg
);
371 struct throtl_data
*td
= blkg
->q
->td
;
372 struct throtl_service_queue
*sq
= &tg
->service_queue
;
375 * If on the default hierarchy, we switch to properly hierarchical
376 * behavior where limits on a given throtl_grp are applied to the
377 * whole subtree rather than just the group itself. e.g. If 16M
378 * read_bps limit is set on the root group, the whole system can't
379 * exceed 16M for the device.
381 * If not on the default hierarchy, the broken flat hierarchy
382 * behavior is retained where all throtl_grps are treated as if
383 * they're all separate root groups right below throtl_data.
384 * Limits of a group don't interact with limits of other groups
385 * regardless of the position of the group in the hierarchy.
387 sq
->parent_sq
= &td
->service_queue
;
388 if (cgroup_on_dfl(blkg
->blkcg
->css
.cgroup
) && blkg
->parent
)
389 sq
->parent_sq
= &blkg_to_tg(blkg
->parent
)->service_queue
;
394 * Set has_rules[] if @tg or any of its parents have limits configured.
395 * This doesn't require walking up to the top of the hierarchy as the
396 * parent's has_rules[] is guaranteed to be correct.
398 static void tg_update_has_rules(struct throtl_grp
*tg
)
400 struct throtl_grp
*parent_tg
= sq_to_tg(tg
->service_queue
.parent_sq
);
403 for (rw
= READ
; rw
<= WRITE
; rw
++)
404 tg
->has_rules
[rw
] = (parent_tg
&& parent_tg
->has_rules
[rw
]) ||
405 (tg
->bps
[rw
] != -1 || tg
->iops
[rw
] != -1);
408 static void throtl_pd_online(struct blkg_policy_data
*pd
)
411 * We don't want new groups to escape the limits of its ancestors.
412 * Update has_rules[] after a new group is brought online.
414 tg_update_has_rules(pd_to_tg(pd
));
417 static void throtl_pd_free(struct blkg_policy_data
*pd
)
419 struct throtl_grp
*tg
= pd_to_tg(pd
);
421 del_timer_sync(&tg
->service_queue
.pending_timer
);
422 blkg_rwstat_exit(&tg
->serviced
);
423 blkg_rwstat_exit(&tg
->service_bytes
);
427 static void throtl_pd_reset_stats(struct blkg_policy_data
*pd
)
429 struct throtl_grp
*tg
= pd_to_tg(pd
);
431 blkg_rwstat_reset(&tg
->service_bytes
);
432 blkg_rwstat_reset(&tg
->serviced
);
435 static struct throtl_grp
*
436 throtl_rb_first(struct throtl_service_queue
*parent_sq
)
438 /* Service tree is empty */
439 if (!parent_sq
->nr_pending
)
442 if (!parent_sq
->first_pending
)
443 parent_sq
->first_pending
= rb_first(&parent_sq
->pending_tree
);
445 if (parent_sq
->first_pending
)
446 return rb_entry_tg(parent_sq
->first_pending
);
451 static void rb_erase_init(struct rb_node
*n
, struct rb_root
*root
)
457 static void throtl_rb_erase(struct rb_node
*n
,
458 struct throtl_service_queue
*parent_sq
)
460 if (parent_sq
->first_pending
== n
)
461 parent_sq
->first_pending
= NULL
;
462 rb_erase_init(n
, &parent_sq
->pending_tree
);
463 --parent_sq
->nr_pending
;
466 static void update_min_dispatch_time(struct throtl_service_queue
*parent_sq
)
468 struct throtl_grp
*tg
;
470 tg
= throtl_rb_first(parent_sq
);
474 parent_sq
->first_pending_disptime
= tg
->disptime
;
477 static void tg_service_queue_add(struct throtl_grp
*tg
)
479 struct throtl_service_queue
*parent_sq
= tg
->service_queue
.parent_sq
;
480 struct rb_node
**node
= &parent_sq
->pending_tree
.rb_node
;
481 struct rb_node
*parent
= NULL
;
482 struct throtl_grp
*__tg
;
483 unsigned long key
= tg
->disptime
;
486 while (*node
!= NULL
) {
488 __tg
= rb_entry_tg(parent
);
490 if (time_before(key
, __tg
->disptime
))
491 node
= &parent
->rb_left
;
493 node
= &parent
->rb_right
;
499 parent_sq
->first_pending
= &tg
->rb_node
;
501 rb_link_node(&tg
->rb_node
, parent
, node
);
502 rb_insert_color(&tg
->rb_node
, &parent_sq
->pending_tree
);
505 static void __throtl_enqueue_tg(struct throtl_grp
*tg
)
507 tg_service_queue_add(tg
);
508 tg
->flags
|= THROTL_TG_PENDING
;
509 tg
->service_queue
.parent_sq
->nr_pending
++;
512 static void throtl_enqueue_tg(struct throtl_grp
*tg
)
514 if (!(tg
->flags
& THROTL_TG_PENDING
))
515 __throtl_enqueue_tg(tg
);
518 static void __throtl_dequeue_tg(struct throtl_grp
*tg
)
520 throtl_rb_erase(&tg
->rb_node
, tg
->service_queue
.parent_sq
);
521 tg
->flags
&= ~THROTL_TG_PENDING
;
524 static void throtl_dequeue_tg(struct throtl_grp
*tg
)
526 if (tg
->flags
& THROTL_TG_PENDING
)
527 __throtl_dequeue_tg(tg
);
530 /* Call with queue lock held */
531 static void throtl_schedule_pending_timer(struct throtl_service_queue
*sq
,
532 unsigned long expires
)
534 mod_timer(&sq
->pending_timer
, expires
);
535 throtl_log(sq
, "schedule timer. delay=%lu jiffies=%lu",
536 expires
- jiffies
, jiffies
);
540 * throtl_schedule_next_dispatch - schedule the next dispatch cycle
541 * @sq: the service_queue to schedule dispatch for
542 * @force: force scheduling
544 * Arm @sq->pending_timer so that the next dispatch cycle starts on the
545 * dispatch time of the first pending child. Returns %true if either timer
546 * is armed or there's no pending child left. %false if the current
547 * dispatch window is still open and the caller should continue
550 * If @force is %true, the dispatch timer is always scheduled and this
551 * function is guaranteed to return %true. This is to be used when the
552 * caller can't dispatch itself and needs to invoke pending_timer
553 * unconditionally. Note that forced scheduling is likely to induce short
554 * delay before dispatch starts even if @sq->first_pending_disptime is not
555 * in the future and thus shouldn't be used in hot paths.
557 static bool throtl_schedule_next_dispatch(struct throtl_service_queue
*sq
,
560 /* any pending children left? */
564 update_min_dispatch_time(sq
);
566 /* is the next dispatch time in the future? */
567 if (force
|| time_after(sq
->first_pending_disptime
, jiffies
)) {
568 throtl_schedule_pending_timer(sq
, sq
->first_pending_disptime
);
572 /* tell the caller to continue dispatching */
576 static inline void throtl_start_new_slice_with_credit(struct throtl_grp
*tg
,
577 bool rw
, unsigned long start
)
579 tg
->bytes_disp
[rw
] = 0;
583 * Previous slice has expired. We must have trimmed it after last
584 * bio dispatch. That means since start of last slice, we never used
585 * that bandwidth. Do try to make use of that bandwidth while giving
588 if (time_after_eq(start
, tg
->slice_start
[rw
]))
589 tg
->slice_start
[rw
] = start
;
591 tg
->slice_end
[rw
] = jiffies
+ throtl_slice
;
592 throtl_log(&tg
->service_queue
,
593 "[%c] new slice with credit start=%lu end=%lu jiffies=%lu",
594 rw
== READ
? 'R' : 'W', tg
->slice_start
[rw
],
595 tg
->slice_end
[rw
], jiffies
);
598 static inline void throtl_start_new_slice(struct throtl_grp
*tg
, bool rw
)
600 tg
->bytes_disp
[rw
] = 0;
602 tg
->slice_start
[rw
] = jiffies
;
603 tg
->slice_end
[rw
] = jiffies
+ throtl_slice
;
604 throtl_log(&tg
->service_queue
,
605 "[%c] new slice start=%lu end=%lu jiffies=%lu",
606 rw
== READ
? 'R' : 'W', tg
->slice_start
[rw
],
607 tg
->slice_end
[rw
], jiffies
);
610 static inline void throtl_set_slice_end(struct throtl_grp
*tg
, bool rw
,
611 unsigned long jiffy_end
)
613 tg
->slice_end
[rw
] = roundup(jiffy_end
, throtl_slice
);
616 static inline void throtl_extend_slice(struct throtl_grp
*tg
, bool rw
,
617 unsigned long jiffy_end
)
619 tg
->slice_end
[rw
] = roundup(jiffy_end
, throtl_slice
);
620 throtl_log(&tg
->service_queue
,
621 "[%c] extend slice start=%lu end=%lu jiffies=%lu",
622 rw
== READ
? 'R' : 'W', tg
->slice_start
[rw
],
623 tg
->slice_end
[rw
], jiffies
);
626 /* Determine if previously allocated or extended slice is complete or not */
627 static bool throtl_slice_used(struct throtl_grp
*tg
, bool rw
)
629 if (time_in_range(jiffies
, tg
->slice_start
[rw
], tg
->slice_end
[rw
]))
635 /* Trim the used slices and adjust slice start accordingly */
636 static inline void throtl_trim_slice(struct throtl_grp
*tg
, bool rw
)
638 unsigned long nr_slices
, time_elapsed
, io_trim
;
641 BUG_ON(time_before(tg
->slice_end
[rw
], tg
->slice_start
[rw
]));
644 * If bps are unlimited (-1), then time slice don't get
645 * renewed. Don't try to trim the slice if slice is used. A new
646 * slice will start when appropriate.
648 if (throtl_slice_used(tg
, rw
))
652 * A bio has been dispatched. Also adjust slice_end. It might happen
653 * that initially cgroup limit was very low resulting in high
654 * slice_end, but later limit was bumped up and bio was dispached
655 * sooner, then we need to reduce slice_end. A high bogus slice_end
656 * is bad because it does not allow new slice to start.
659 throtl_set_slice_end(tg
, rw
, jiffies
+ throtl_slice
);
661 time_elapsed
= jiffies
- tg
->slice_start
[rw
];
663 nr_slices
= time_elapsed
/ throtl_slice
;
667 tmp
= tg
->bps
[rw
] * throtl_slice
* nr_slices
;
671 io_trim
= (tg
->iops
[rw
] * throtl_slice
* nr_slices
)/HZ
;
673 if (!bytes_trim
&& !io_trim
)
676 if (tg
->bytes_disp
[rw
] >= bytes_trim
)
677 tg
->bytes_disp
[rw
] -= bytes_trim
;
679 tg
->bytes_disp
[rw
] = 0;
681 if (tg
->io_disp
[rw
] >= io_trim
)
682 tg
->io_disp
[rw
] -= io_trim
;
686 tg
->slice_start
[rw
] += nr_slices
* throtl_slice
;
688 throtl_log(&tg
->service_queue
,
689 "[%c] trim slice nr=%lu bytes=%llu io=%lu start=%lu end=%lu jiffies=%lu",
690 rw
== READ
? 'R' : 'W', nr_slices
, bytes_trim
, io_trim
,
691 tg
->slice_start
[rw
], tg
->slice_end
[rw
], jiffies
);
694 static bool tg_with_in_iops_limit(struct throtl_grp
*tg
, struct bio
*bio
,
697 bool rw
= bio_data_dir(bio
);
698 unsigned int io_allowed
;
699 unsigned long jiffy_elapsed
, jiffy_wait
, jiffy_elapsed_rnd
;
702 jiffy_elapsed
= jiffy_elapsed_rnd
= jiffies
- tg
->slice_start
[rw
];
704 /* Slice has just started. Consider one slice interval */
706 jiffy_elapsed_rnd
= throtl_slice
;
708 jiffy_elapsed_rnd
= roundup(jiffy_elapsed_rnd
, throtl_slice
);
711 * jiffy_elapsed_rnd should not be a big value as minimum iops can be
712 * 1 then at max jiffy elapsed should be equivalent of 1 second as we
713 * will allow dispatch after 1 second and after that slice should
717 tmp
= (u64
)tg
->iops
[rw
] * jiffy_elapsed_rnd
;
721 io_allowed
= UINT_MAX
;
725 if (tg
->io_disp
[rw
] + 1 <= io_allowed
) {
731 /* Calc approx time to dispatch */
732 jiffy_wait
= ((tg
->io_disp
[rw
] + 1) * HZ
)/tg
->iops
[rw
] + 1;
734 if (jiffy_wait
> jiffy_elapsed
)
735 jiffy_wait
= jiffy_wait
- jiffy_elapsed
;
744 static bool tg_with_in_bps_limit(struct throtl_grp
*tg
, struct bio
*bio
,
747 bool rw
= bio_data_dir(bio
);
748 u64 bytes_allowed
, extra_bytes
, tmp
;
749 unsigned long jiffy_elapsed
, jiffy_wait
, jiffy_elapsed_rnd
;
751 jiffy_elapsed
= jiffy_elapsed_rnd
= jiffies
- tg
->slice_start
[rw
];
753 /* Slice has just started. Consider one slice interval */
755 jiffy_elapsed_rnd
= throtl_slice
;
757 jiffy_elapsed_rnd
= roundup(jiffy_elapsed_rnd
, throtl_slice
);
759 tmp
= tg
->bps
[rw
] * jiffy_elapsed_rnd
;
763 if (tg
->bytes_disp
[rw
] + bio
->bi_iter
.bi_size
<= bytes_allowed
) {
769 /* Calc approx time to dispatch */
770 extra_bytes
= tg
->bytes_disp
[rw
] + bio
->bi_iter
.bi_size
- bytes_allowed
;
771 jiffy_wait
= div64_u64(extra_bytes
* HZ
, tg
->bps
[rw
]);
777 * This wait time is without taking into consideration the rounding
778 * up we did. Add that time also.
780 jiffy_wait
= jiffy_wait
+ (jiffy_elapsed_rnd
- jiffy_elapsed
);
787 * Returns whether one can dispatch a bio or not. Also returns approx number
788 * of jiffies to wait before this bio is with-in IO rate and can be dispatched
790 static bool tg_may_dispatch(struct throtl_grp
*tg
, struct bio
*bio
,
793 bool rw
= bio_data_dir(bio
);
794 unsigned long bps_wait
= 0, iops_wait
= 0, max_wait
= 0;
797 * Currently whole state machine of group depends on first bio
798 * queued in the group bio list. So one should not be calling
799 * this function with a different bio if there are other bios
802 BUG_ON(tg
->service_queue
.nr_queued
[rw
] &&
803 bio
!= throtl_peek_queued(&tg
->service_queue
.queued
[rw
]));
805 /* If tg->bps = -1, then BW is unlimited */
806 if (tg
->bps
[rw
] == -1 && tg
->iops
[rw
] == -1) {
813 * If previous slice expired, start a new one otherwise renew/extend
814 * existing slice to make sure it is at least throtl_slice interval
817 if (throtl_slice_used(tg
, rw
))
818 throtl_start_new_slice(tg
, rw
);
820 if (time_before(tg
->slice_end
[rw
], jiffies
+ throtl_slice
))
821 throtl_extend_slice(tg
, rw
, jiffies
+ throtl_slice
);
824 if (tg_with_in_bps_limit(tg
, bio
, &bps_wait
) &&
825 tg_with_in_iops_limit(tg
, bio
, &iops_wait
)) {
831 max_wait
= max(bps_wait
, iops_wait
);
836 if (time_before(tg
->slice_end
[rw
], jiffies
+ max_wait
))
837 throtl_extend_slice(tg
, rw
, jiffies
+ max_wait
);
842 static void throtl_update_dispatch_stats(struct blkcg_gq
*blkg
, u64 bytes
,
845 struct throtl_grp
*tg
= blkg_to_tg(blkg
);
849 * Disabling interrupts to provide mutual exclusion between two
850 * writes on same cpu. It probably is not needed for 64bit. Not
851 * optimizing that case yet.
853 local_irq_save(flags
);
855 blkg_rwstat_add(&tg
->serviced
, rw
, 1);
856 blkg_rwstat_add(&tg
->service_bytes
, rw
, bytes
);
858 local_irq_restore(flags
);
861 static void throtl_charge_bio(struct throtl_grp
*tg
, struct bio
*bio
)
863 bool rw
= bio_data_dir(bio
);
865 /* Charge the bio to the group */
866 tg
->bytes_disp
[rw
] += bio
->bi_iter
.bi_size
;
870 * REQ_THROTTLED is used to prevent the same bio to be throttled
871 * more than once as a throttled bio will go through blk-throtl the
872 * second time when it eventually gets issued. Set it when a bio
873 * is being charged to a tg.
875 * Dispatch stats aren't recursive and each @bio should only be
876 * accounted by the @tg it was originally associated with. Let's
877 * update the stats when setting REQ_THROTTLED for the first time
878 * which is guaranteed to be for the @bio's original tg.
880 if (!(bio
->bi_rw
& REQ_THROTTLED
)) {
881 bio
->bi_rw
|= REQ_THROTTLED
;
882 throtl_update_dispatch_stats(tg_to_blkg(tg
),
883 bio
->bi_iter
.bi_size
, bio
->bi_rw
);
888 * throtl_add_bio_tg - add a bio to the specified throtl_grp
891 * @tg: the target throtl_grp
893 * Add @bio to @tg's service_queue using @qn. If @qn is not specified,
894 * tg->qnode_on_self[] is used.
896 static void throtl_add_bio_tg(struct bio
*bio
, struct throtl_qnode
*qn
,
897 struct throtl_grp
*tg
)
899 struct throtl_service_queue
*sq
= &tg
->service_queue
;
900 bool rw
= bio_data_dir(bio
);
903 qn
= &tg
->qnode_on_self
[rw
];
906 * If @tg doesn't currently have any bios queued in the same
907 * direction, queueing @bio can change when @tg should be
908 * dispatched. Mark that @tg was empty. This is automatically
909 * cleaered on the next tg_update_disptime().
911 if (!sq
->nr_queued
[rw
])
912 tg
->flags
|= THROTL_TG_WAS_EMPTY
;
914 throtl_qnode_add_bio(bio
, qn
, &sq
->queued
[rw
]);
917 throtl_enqueue_tg(tg
);
920 static void tg_update_disptime(struct throtl_grp
*tg
)
922 struct throtl_service_queue
*sq
= &tg
->service_queue
;
923 unsigned long read_wait
= -1, write_wait
= -1, min_wait
= -1, disptime
;
926 if ((bio
= throtl_peek_queued(&sq
->queued
[READ
])))
927 tg_may_dispatch(tg
, bio
, &read_wait
);
929 if ((bio
= throtl_peek_queued(&sq
->queued
[WRITE
])))
930 tg_may_dispatch(tg
, bio
, &write_wait
);
932 min_wait
= min(read_wait
, write_wait
);
933 disptime
= jiffies
+ min_wait
;
935 /* Update dispatch time */
936 throtl_dequeue_tg(tg
);
937 tg
->disptime
= disptime
;
938 throtl_enqueue_tg(tg
);
940 /* see throtl_add_bio_tg() */
941 tg
->flags
&= ~THROTL_TG_WAS_EMPTY
;
944 static void start_parent_slice_with_credit(struct throtl_grp
*child_tg
,
945 struct throtl_grp
*parent_tg
, bool rw
)
947 if (throtl_slice_used(parent_tg
, rw
)) {
948 throtl_start_new_slice_with_credit(parent_tg
, rw
,
949 child_tg
->slice_start
[rw
]);
954 static void tg_dispatch_one_bio(struct throtl_grp
*tg
, bool rw
)
956 struct throtl_service_queue
*sq
= &tg
->service_queue
;
957 struct throtl_service_queue
*parent_sq
= sq
->parent_sq
;
958 struct throtl_grp
*parent_tg
= sq_to_tg(parent_sq
);
959 struct throtl_grp
*tg_to_put
= NULL
;
963 * @bio is being transferred from @tg to @parent_sq. Popping a bio
964 * from @tg may put its reference and @parent_sq might end up
965 * getting released prematurely. Remember the tg to put and put it
966 * after @bio is transferred to @parent_sq.
968 bio
= throtl_pop_queued(&sq
->queued
[rw
], &tg_to_put
);
971 throtl_charge_bio(tg
, bio
);
974 * If our parent is another tg, we just need to transfer @bio to
975 * the parent using throtl_add_bio_tg(). If our parent is
976 * @td->service_queue, @bio is ready to be issued. Put it on its
977 * bio_lists[] and decrease total number queued. The caller is
978 * responsible for issuing these bios.
981 throtl_add_bio_tg(bio
, &tg
->qnode_on_parent
[rw
], parent_tg
);
982 start_parent_slice_with_credit(tg
, parent_tg
, rw
);
984 throtl_qnode_add_bio(bio
, &tg
->qnode_on_parent
[rw
],
985 &parent_sq
->queued
[rw
]);
986 BUG_ON(tg
->td
->nr_queued
[rw
] <= 0);
987 tg
->td
->nr_queued
[rw
]--;
990 throtl_trim_slice(tg
, rw
);
993 blkg_put(tg_to_blkg(tg_to_put
));
996 static int throtl_dispatch_tg(struct throtl_grp
*tg
)
998 struct throtl_service_queue
*sq
= &tg
->service_queue
;
999 unsigned int nr_reads
= 0, nr_writes
= 0;
1000 unsigned int max_nr_reads
= throtl_grp_quantum
*3/4;
1001 unsigned int max_nr_writes
= throtl_grp_quantum
- max_nr_reads
;
1004 /* Try to dispatch 75% READS and 25% WRITES */
1006 while ((bio
= throtl_peek_queued(&sq
->queued
[READ
])) &&
1007 tg_may_dispatch(tg
, bio
, NULL
)) {
1009 tg_dispatch_one_bio(tg
, bio_data_dir(bio
));
1012 if (nr_reads
>= max_nr_reads
)
1016 while ((bio
= throtl_peek_queued(&sq
->queued
[WRITE
])) &&
1017 tg_may_dispatch(tg
, bio
, NULL
)) {
1019 tg_dispatch_one_bio(tg
, bio_data_dir(bio
));
1022 if (nr_writes
>= max_nr_writes
)
1026 return nr_reads
+ nr_writes
;
1029 static int throtl_select_dispatch(struct throtl_service_queue
*parent_sq
)
1031 unsigned int nr_disp
= 0;
1034 struct throtl_grp
*tg
= throtl_rb_first(parent_sq
);
1035 struct throtl_service_queue
*sq
= &tg
->service_queue
;
1040 if (time_before(jiffies
, tg
->disptime
))
1043 throtl_dequeue_tg(tg
);
1045 nr_disp
+= throtl_dispatch_tg(tg
);
1047 if (sq
->nr_queued
[0] || sq
->nr_queued
[1])
1048 tg_update_disptime(tg
);
1050 if (nr_disp
>= throtl_quantum
)
1058 * throtl_pending_timer_fn - timer function for service_queue->pending_timer
1059 * @arg: the throtl_service_queue being serviced
1061 * This timer is armed when a child throtl_grp with active bio's become
1062 * pending and queued on the service_queue's pending_tree and expires when
1063 * the first child throtl_grp should be dispatched. This function
1064 * dispatches bio's from the children throtl_grps to the parent
1067 * If the parent's parent is another throtl_grp, dispatching is propagated
1068 * by either arming its pending_timer or repeating dispatch directly. If
1069 * the top-level service_tree is reached, throtl_data->dispatch_work is
1070 * kicked so that the ready bio's are issued.
1072 static void throtl_pending_timer_fn(unsigned long arg
)
1074 struct throtl_service_queue
*sq
= (void *)arg
;
1075 struct throtl_grp
*tg
= sq_to_tg(sq
);
1076 struct throtl_data
*td
= sq_to_td(sq
);
1077 struct request_queue
*q
= td
->queue
;
1078 struct throtl_service_queue
*parent_sq
;
1082 spin_lock_irq(q
->queue_lock
);
1084 parent_sq
= sq
->parent_sq
;
1088 throtl_log(sq
, "dispatch nr_queued=%u read=%u write=%u",
1089 sq
->nr_queued
[READ
] + sq
->nr_queued
[WRITE
],
1090 sq
->nr_queued
[READ
], sq
->nr_queued
[WRITE
]);
1092 ret
= throtl_select_dispatch(sq
);
1094 throtl_log(sq
, "bios disp=%u", ret
);
1098 if (throtl_schedule_next_dispatch(sq
, false))
1101 /* this dispatch windows is still open, relax and repeat */
1102 spin_unlock_irq(q
->queue_lock
);
1104 spin_lock_irq(q
->queue_lock
);
1111 /* @parent_sq is another throl_grp, propagate dispatch */
1112 if (tg
->flags
& THROTL_TG_WAS_EMPTY
) {
1113 tg_update_disptime(tg
);
1114 if (!throtl_schedule_next_dispatch(parent_sq
, false)) {
1115 /* window is already open, repeat dispatching */
1122 /* reached the top-level, queue issueing */
1123 queue_work(kthrotld_workqueue
, &td
->dispatch_work
);
1126 spin_unlock_irq(q
->queue_lock
);
1130 * blk_throtl_dispatch_work_fn - work function for throtl_data->dispatch_work
1131 * @work: work item being executed
1133 * This function is queued for execution when bio's reach the bio_lists[]
1134 * of throtl_data->service_queue. Those bio's are ready and issued by this
1137 static void blk_throtl_dispatch_work_fn(struct work_struct
*work
)
1139 struct throtl_data
*td
= container_of(work
, struct throtl_data
,
1141 struct throtl_service_queue
*td_sq
= &td
->service_queue
;
1142 struct request_queue
*q
= td
->queue
;
1143 struct bio_list bio_list_on_stack
;
1145 struct blk_plug plug
;
1148 bio_list_init(&bio_list_on_stack
);
1150 spin_lock_irq(q
->queue_lock
);
1151 for (rw
= READ
; rw
<= WRITE
; rw
++)
1152 while ((bio
= throtl_pop_queued(&td_sq
->queued
[rw
], NULL
)))
1153 bio_list_add(&bio_list_on_stack
, bio
);
1154 spin_unlock_irq(q
->queue_lock
);
1156 if (!bio_list_empty(&bio_list_on_stack
)) {
1157 blk_start_plug(&plug
);
1158 while((bio
= bio_list_pop(&bio_list_on_stack
)))
1159 generic_make_request(bio
);
1160 blk_finish_plug(&plug
);
1164 static int tg_print_rwstat(struct seq_file
*sf
, void *v
)
1166 blkcg_print_blkgs(sf
, css_to_blkcg(seq_css(sf
)), blkg_prfill_rwstat
,
1167 &blkcg_policy_throtl
, seq_cft(sf
)->private, true);
1171 static u64
tg_prfill_conf_u64(struct seq_file
*sf
, struct blkg_policy_data
*pd
,
1174 struct throtl_grp
*tg
= pd_to_tg(pd
);
1175 u64 v
= *(u64
*)((void *)tg
+ off
);
1179 return __blkg_prfill_u64(sf
, pd
, v
);
1182 static u64
tg_prfill_conf_uint(struct seq_file
*sf
, struct blkg_policy_data
*pd
,
1185 struct throtl_grp
*tg
= pd_to_tg(pd
);
1186 unsigned int v
= *(unsigned int *)((void *)tg
+ off
);
1190 return __blkg_prfill_u64(sf
, pd
, v
);
1193 static int tg_print_conf_u64(struct seq_file
*sf
, void *v
)
1195 blkcg_print_blkgs(sf
, css_to_blkcg(seq_css(sf
)), tg_prfill_conf_u64
,
1196 &blkcg_policy_throtl
, seq_cft(sf
)->private, false);
1200 static int tg_print_conf_uint(struct seq_file
*sf
, void *v
)
1202 blkcg_print_blkgs(sf
, css_to_blkcg(seq_css(sf
)), tg_prfill_conf_uint
,
1203 &blkcg_policy_throtl
, seq_cft(sf
)->private, false);
1207 static ssize_t
tg_set_conf(struct kernfs_open_file
*of
,
1208 char *buf
, size_t nbytes
, loff_t off
, bool is_u64
)
1210 struct blkcg
*blkcg
= css_to_blkcg(of_css(of
));
1211 struct blkg_conf_ctx ctx
;
1212 struct throtl_grp
*tg
;
1213 struct throtl_service_queue
*sq
;
1214 struct blkcg_gq
*blkg
;
1215 struct cgroup_subsys_state
*pos_css
;
1218 ret
= blkg_conf_prep(blkcg
, &blkcg_policy_throtl
, buf
, &ctx
);
1222 tg
= blkg_to_tg(ctx
.blkg
);
1223 sq
= &tg
->service_queue
;
1229 *(u64
*)((void *)tg
+ of_cft(of
)->private) = ctx
.v
;
1231 *(unsigned int *)((void *)tg
+ of_cft(of
)->private) = ctx
.v
;
1233 throtl_log(&tg
->service_queue
,
1234 "limit change rbps=%llu wbps=%llu riops=%u wiops=%u",
1235 tg
->bps
[READ
], tg
->bps
[WRITE
],
1236 tg
->iops
[READ
], tg
->iops
[WRITE
]);
1239 * Update has_rules[] flags for the updated tg's subtree. A tg is
1240 * considered to have rules if either the tg itself or any of its
1241 * ancestors has rules. This identifies groups without any
1242 * restrictions in the whole hierarchy and allows them to bypass
1245 blkg_for_each_descendant_pre(blkg
, pos_css
, ctx
.blkg
)
1246 tg_update_has_rules(blkg_to_tg(blkg
));
1249 * We're already holding queue_lock and know @tg is valid. Let's
1250 * apply the new config directly.
1252 * Restart the slices for both READ and WRITES. It might happen
1253 * that a group's limit are dropped suddenly and we don't want to
1254 * account recently dispatched IO with new low rate.
1256 throtl_start_new_slice(tg
, 0);
1257 throtl_start_new_slice(tg
, 1);
1259 if (tg
->flags
& THROTL_TG_PENDING
) {
1260 tg_update_disptime(tg
);
1261 throtl_schedule_next_dispatch(sq
->parent_sq
, true);
1264 blkg_conf_finish(&ctx
);
1268 static ssize_t
tg_set_conf_u64(struct kernfs_open_file
*of
,
1269 char *buf
, size_t nbytes
, loff_t off
)
1271 return tg_set_conf(of
, buf
, nbytes
, off
, true);
1274 static ssize_t
tg_set_conf_uint(struct kernfs_open_file
*of
,
1275 char *buf
, size_t nbytes
, loff_t off
)
1277 return tg_set_conf(of
, buf
, nbytes
, off
, false);
1280 static struct cftype throtl_files
[] = {
1282 .name
= "throttle.read_bps_device",
1283 .private = offsetof(struct throtl_grp
, bps
[READ
]),
1284 .seq_show
= tg_print_conf_u64
,
1285 .write
= tg_set_conf_u64
,
1288 .name
= "throttle.write_bps_device",
1289 .private = offsetof(struct throtl_grp
, bps
[WRITE
]),
1290 .seq_show
= tg_print_conf_u64
,
1291 .write
= tg_set_conf_u64
,
1294 .name
= "throttle.read_iops_device",
1295 .private = offsetof(struct throtl_grp
, iops
[READ
]),
1296 .seq_show
= tg_print_conf_uint
,
1297 .write
= tg_set_conf_uint
,
1300 .name
= "throttle.write_iops_device",
1301 .private = offsetof(struct throtl_grp
, iops
[WRITE
]),
1302 .seq_show
= tg_print_conf_uint
,
1303 .write
= tg_set_conf_uint
,
1306 .name
= "throttle.io_service_bytes",
1307 .private = offsetof(struct throtl_grp
, service_bytes
),
1308 .seq_show
= tg_print_rwstat
,
1311 .name
= "throttle.io_serviced",
1312 .private = offsetof(struct throtl_grp
, serviced
),
1313 .seq_show
= tg_print_rwstat
,
1318 static void throtl_shutdown_wq(struct request_queue
*q
)
1320 struct throtl_data
*td
= q
->td
;
1322 cancel_work_sync(&td
->dispatch_work
);
1325 static struct blkcg_policy blkcg_policy_throtl
= {
1326 .cftypes
= throtl_files
,
1328 .pd_alloc_fn
= throtl_pd_alloc
,
1329 .pd_init_fn
= throtl_pd_init
,
1330 .pd_online_fn
= throtl_pd_online
,
1331 .pd_free_fn
= throtl_pd_free
,
1332 .pd_reset_stats_fn
= throtl_pd_reset_stats
,
1335 bool blk_throtl_bio(struct request_queue
*q
, struct blkcg_gq
*blkg
,
1338 struct throtl_qnode
*qn
= NULL
;
1339 struct throtl_grp
*tg
= blkg_to_tg(blkg
?: q
->root_blkg
);
1340 struct throtl_service_queue
*sq
;
1341 bool rw
= bio_data_dir(bio
);
1342 bool throttled
= false;
1344 WARN_ON_ONCE(!rcu_read_lock_held());
1346 /* see throtl_charge_bio() */
1347 if ((bio
->bi_rw
& REQ_THROTTLED
) || !tg
->has_rules
[rw
])
1350 spin_lock_irq(q
->queue_lock
);
1352 if (unlikely(blk_queue_bypass(q
)))
1355 sq
= &tg
->service_queue
;
1358 /* throtl is FIFO - if bios are already queued, should queue */
1359 if (sq
->nr_queued
[rw
])
1362 /* if above limits, break to queue */
1363 if (!tg_may_dispatch(tg
, bio
, NULL
))
1366 /* within limits, let's charge and dispatch directly */
1367 throtl_charge_bio(tg
, bio
);
1370 * We need to trim slice even when bios are not being queued
1371 * otherwise it might happen that a bio is not queued for
1372 * a long time and slice keeps on extending and trim is not
1373 * called for a long time. Now if limits are reduced suddenly
1374 * we take into account all the IO dispatched so far at new
1375 * low rate and * newly queued IO gets a really long dispatch
1378 * So keep on trimming slice even if bio is not queued.
1380 throtl_trim_slice(tg
, rw
);
1383 * @bio passed through this layer without being throttled.
1384 * Climb up the ladder. If we''re already at the top, it
1385 * can be executed directly.
1387 qn
= &tg
->qnode_on_parent
[rw
];
1394 /* out-of-limit, queue to @tg */
1395 throtl_log(sq
, "[%c] bio. bdisp=%llu sz=%u bps=%llu iodisp=%u iops=%u queued=%d/%d",
1396 rw
== READ
? 'R' : 'W',
1397 tg
->bytes_disp
[rw
], bio
->bi_iter
.bi_size
, tg
->bps
[rw
],
1398 tg
->io_disp
[rw
], tg
->iops
[rw
],
1399 sq
->nr_queued
[READ
], sq
->nr_queued
[WRITE
]);
1401 bio_associate_current(bio
);
1402 tg
->td
->nr_queued
[rw
]++;
1403 throtl_add_bio_tg(bio
, qn
, tg
);
1407 * Update @tg's dispatch time and force schedule dispatch if @tg
1408 * was empty before @bio. The forced scheduling isn't likely to
1409 * cause undue delay as @bio is likely to be dispatched directly if
1410 * its @tg's disptime is not in the future.
1412 if (tg
->flags
& THROTL_TG_WAS_EMPTY
) {
1413 tg_update_disptime(tg
);
1414 throtl_schedule_next_dispatch(tg
->service_queue
.parent_sq
, true);
1418 spin_unlock_irq(q
->queue_lock
);
1421 * As multiple blk-throtls may stack in the same issue path, we
1422 * don't want bios to leave with the flag set. Clear the flag if
1426 bio
->bi_rw
&= ~REQ_THROTTLED
;
1431 * Dispatch all bios from all children tg's queued on @parent_sq. On
1432 * return, @parent_sq is guaranteed to not have any active children tg's
1433 * and all bios from previously active tg's are on @parent_sq->bio_lists[].
1435 static void tg_drain_bios(struct throtl_service_queue
*parent_sq
)
1437 struct throtl_grp
*tg
;
1439 while ((tg
= throtl_rb_first(parent_sq
))) {
1440 struct throtl_service_queue
*sq
= &tg
->service_queue
;
1443 throtl_dequeue_tg(tg
);
1445 while ((bio
= throtl_peek_queued(&sq
->queued
[READ
])))
1446 tg_dispatch_one_bio(tg
, bio_data_dir(bio
));
1447 while ((bio
= throtl_peek_queued(&sq
->queued
[WRITE
])))
1448 tg_dispatch_one_bio(tg
, bio_data_dir(bio
));
1453 * blk_throtl_drain - drain throttled bios
1454 * @q: request_queue to drain throttled bios for
1456 * Dispatch all currently throttled bios on @q through ->make_request_fn().
1458 void blk_throtl_drain(struct request_queue
*q
)
1459 __releases(q
->queue_lock
) __acquires(q
->queue_lock
)
1461 struct throtl_data
*td
= q
->td
;
1462 struct blkcg_gq
*blkg
;
1463 struct cgroup_subsys_state
*pos_css
;
1467 queue_lockdep_assert_held(q
);
1471 * Drain each tg while doing post-order walk on the blkg tree, so
1472 * that all bios are propagated to td->service_queue. It'd be
1473 * better to walk service_queue tree directly but blkg walk is
1476 blkg_for_each_descendant_post(blkg
, pos_css
, td
->queue
->root_blkg
)
1477 tg_drain_bios(&blkg_to_tg(blkg
)->service_queue
);
1479 /* finally, transfer bios from top-level tg's into the td */
1480 tg_drain_bios(&td
->service_queue
);
1483 spin_unlock_irq(q
->queue_lock
);
1485 /* all bios now should be in td->service_queue, issue them */
1486 for (rw
= READ
; rw
<= WRITE
; rw
++)
1487 while ((bio
= throtl_pop_queued(&td
->service_queue
.queued
[rw
],
1489 generic_make_request(bio
);
1491 spin_lock_irq(q
->queue_lock
);
1494 int blk_throtl_init(struct request_queue
*q
)
1496 struct throtl_data
*td
;
1499 td
= kzalloc_node(sizeof(*td
), GFP_KERNEL
, q
->node
);
1503 INIT_WORK(&td
->dispatch_work
, blk_throtl_dispatch_work_fn
);
1504 throtl_service_queue_init(&td
->service_queue
);
1509 /* activate policy */
1510 ret
= blkcg_activate_policy(q
, &blkcg_policy_throtl
);
1516 void blk_throtl_exit(struct request_queue
*q
)
1519 throtl_shutdown_wq(q
);
1520 blkcg_deactivate_policy(q
, &blkcg_policy_throtl
);
1524 static int __init
throtl_init(void)
1526 kthrotld_workqueue
= alloc_workqueue("kthrotld", WQ_MEM_RECLAIM
, 0);
1527 if (!kthrotld_workqueue
)
1528 panic("Failed to create kthrotld\n");
1530 return blkcg_policy_register(&blkcg_policy_throtl
);
1533 module_init(throtl_init
);