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 "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 /* A workqueue to queue throttle related work */
25 static struct workqueue_struct
*kthrotld_workqueue
;
26 static void throtl_schedule_delayed_work(struct throtl_data
*td
,
29 struct throtl_rb_root
{
33 unsigned long min_disptime
;
36 #define THROTL_RB_ROOT (struct throtl_rb_root) { .rb = RB_ROOT, .left = NULL, \
37 .count = 0, .min_disptime = 0}
39 #define rb_entry_tg(node) rb_entry((node), struct throtl_grp, rb_node)
42 /* List of throtl groups on the request queue*/
43 struct hlist_node tg_node
;
45 /* active throtl group service_tree member */
46 struct rb_node rb_node
;
49 * Dispatch time in jiffies. This is the estimated time when group
50 * will unthrottle and is ready to dispatch more bio. It is used as
51 * key to sort active groups in service tree.
53 unsigned long disptime
;
55 struct blkio_group blkg
;
59 /* Two lists for READ and WRITE */
60 struct bio_list bio_lists
[2];
62 /* Number of queued bios on READ and WRITE lists */
63 unsigned int nr_queued
[2];
65 /* bytes per second rate limits */
71 /* Number of bytes disptached in current slice */
72 uint64_t bytes_disp
[2];
73 /* Number of bio's dispatched in current slice */
74 unsigned int io_disp
[2];
76 /* When did we start a new slice */
77 unsigned long slice_start
[2];
78 unsigned long slice_end
[2];
80 /* Some throttle limits got updated for the group */
83 struct rcu_head rcu_head
;
88 /* List of throtl groups */
89 struct hlist_head tg_list
;
91 /* service tree for active throtl groups */
92 struct throtl_rb_root tg_service_tree
;
94 struct throtl_grp
*root_tg
;
95 struct request_queue
*queue
;
97 /* Total Number of queued bios on READ and WRITE lists */
98 unsigned int nr_queued
[2];
101 * number of total undestroyed groups
103 unsigned int nr_undestroyed_grps
;
105 /* Work for dispatching throttled bios */
106 struct delayed_work throtl_work
;
111 enum tg_state_flags
{
112 THROTL_TG_FLAG_on_rr
= 0, /* on round-robin busy list */
115 #define THROTL_TG_FNS(name) \
116 static inline void throtl_mark_tg_##name(struct throtl_grp *tg) \
118 (tg)->flags |= (1 << THROTL_TG_FLAG_##name); \
120 static inline void throtl_clear_tg_##name(struct throtl_grp *tg) \
122 (tg)->flags &= ~(1 << THROTL_TG_FLAG_##name); \
124 static inline int throtl_tg_##name(const struct throtl_grp *tg) \
126 return ((tg)->flags & (1 << THROTL_TG_FLAG_##name)) != 0; \
129 THROTL_TG_FNS(on_rr
);
131 #define throtl_log_tg(td, tg, fmt, args...) \
132 blk_add_trace_msg((td)->queue, "throtl %s " fmt, \
133 blkg_path(&(tg)->blkg), ##args); \
135 #define throtl_log(td, fmt, args...) \
136 blk_add_trace_msg((td)->queue, "throtl " fmt, ##args)
138 static inline struct throtl_grp
*tg_of_blkg(struct blkio_group
*blkg
)
141 return container_of(blkg
, struct throtl_grp
, blkg
);
146 static inline unsigned int total_nr_queued(struct throtl_data
*td
)
148 return td
->nr_queued
[0] + td
->nr_queued
[1];
151 static inline struct throtl_grp
*throtl_ref_get_tg(struct throtl_grp
*tg
)
153 atomic_inc(&tg
->ref
);
157 static void throtl_free_tg(struct rcu_head
*head
)
159 struct throtl_grp
*tg
;
161 tg
= container_of(head
, struct throtl_grp
, rcu_head
);
162 free_percpu(tg
->blkg
.stats_cpu
);
166 static void throtl_put_tg(struct throtl_grp
*tg
)
168 BUG_ON(atomic_read(&tg
->ref
) <= 0);
169 if (!atomic_dec_and_test(&tg
->ref
))
172 /* release the extra blkcg reference this blkg has been holding */
173 css_put(&tg
->blkg
.blkcg
->css
);
176 * A group is freed in rcu manner. But having an rcu lock does not
177 * mean that one can access all the fields of blkg and assume these
178 * are valid. For example, don't try to follow throtl_data and
179 * request queue links.
181 * Having a reference to blkg under an rcu allows acess to only
182 * values local to groups like group stats and group rate limits
184 call_rcu(&tg
->rcu_head
, throtl_free_tg
);
187 static struct blkio_group
*throtl_alloc_blkio_group(struct request_queue
*q
,
188 struct blkio_cgroup
*blkcg
)
190 struct throtl_grp
*tg
;
192 tg
= kzalloc_node(sizeof(*tg
), GFP_ATOMIC
, q
->node
);
196 INIT_HLIST_NODE(&tg
->tg_node
);
197 RB_CLEAR_NODE(&tg
->rb_node
);
198 bio_list_init(&tg
->bio_lists
[0]);
199 bio_list_init(&tg
->bio_lists
[1]);
200 tg
->limits_changed
= false;
205 tg
->iops
[WRITE
] = -1;
208 * Take the initial reference that will be released on destroy
209 * This can be thought of a joint reference by cgroup and
210 * request queue which will be dropped by either request queue
211 * exit or cgroup deletion path depending on who is exiting first.
213 atomic_set(&tg
->ref
, 1);
218 static void throtl_link_blkio_group(struct request_queue
*q
,
219 struct blkio_group
*blkg
)
221 struct throtl_data
*td
= q
->td
;
222 struct throtl_grp
*tg
= tg_of_blkg(blkg
);
224 hlist_add_head(&tg
->tg_node
, &td
->tg_list
);
225 td
->nr_undestroyed_grps
++;
229 throtl_grp
*throtl_lookup_tg(struct throtl_data
*td
, struct blkio_cgroup
*blkcg
)
232 * This is the common case when there are no blkio cgroups.
233 * Avoid lookup in this case
235 if (blkcg
== &blkio_root_cgroup
)
238 return tg_of_blkg(blkg_lookup(blkcg
, td
->queue
, BLKIO_POLICY_THROTL
));
241 static struct throtl_grp
*throtl_lookup_create_tg(struct throtl_data
*td
,
242 struct blkio_cgroup
*blkcg
)
244 struct request_queue
*q
= td
->queue
;
245 struct throtl_grp
*tg
= NULL
;
248 * This is the common case when there are no blkio cgroups.
249 * Avoid lookup in this case
251 if (blkcg
== &blkio_root_cgroup
) {
254 struct blkio_group
*blkg
;
256 blkg
= blkg_lookup_create(blkcg
, q
, BLKIO_POLICY_THROTL
, false);
258 /* if %NULL and @q is alive, fall back to root_tg */
260 tg
= tg_of_blkg(blkg
);
261 else if (!blk_queue_dead(q
))
268 static struct throtl_grp
*throtl_rb_first(struct throtl_rb_root
*root
)
270 /* Service tree is empty */
275 root
->left
= rb_first(&root
->rb
);
278 return rb_entry_tg(root
->left
);
283 static void rb_erase_init(struct rb_node
*n
, struct rb_root
*root
)
289 static void throtl_rb_erase(struct rb_node
*n
, struct throtl_rb_root
*root
)
293 rb_erase_init(n
, &root
->rb
);
297 static void update_min_dispatch_time(struct throtl_rb_root
*st
)
299 struct throtl_grp
*tg
;
301 tg
= throtl_rb_first(st
);
305 st
->min_disptime
= tg
->disptime
;
309 tg_service_tree_add(struct throtl_rb_root
*st
, struct throtl_grp
*tg
)
311 struct rb_node
**node
= &st
->rb
.rb_node
;
312 struct rb_node
*parent
= NULL
;
313 struct throtl_grp
*__tg
;
314 unsigned long key
= tg
->disptime
;
317 while (*node
!= NULL
) {
319 __tg
= rb_entry_tg(parent
);
321 if (time_before(key
, __tg
->disptime
))
322 node
= &parent
->rb_left
;
324 node
= &parent
->rb_right
;
330 st
->left
= &tg
->rb_node
;
332 rb_link_node(&tg
->rb_node
, parent
, node
);
333 rb_insert_color(&tg
->rb_node
, &st
->rb
);
336 static void __throtl_enqueue_tg(struct throtl_data
*td
, struct throtl_grp
*tg
)
338 struct throtl_rb_root
*st
= &td
->tg_service_tree
;
340 tg_service_tree_add(st
, tg
);
341 throtl_mark_tg_on_rr(tg
);
345 static void throtl_enqueue_tg(struct throtl_data
*td
, struct throtl_grp
*tg
)
347 if (!throtl_tg_on_rr(tg
))
348 __throtl_enqueue_tg(td
, tg
);
351 static void __throtl_dequeue_tg(struct throtl_data
*td
, struct throtl_grp
*tg
)
353 throtl_rb_erase(&tg
->rb_node
, &td
->tg_service_tree
);
354 throtl_clear_tg_on_rr(tg
);
357 static void throtl_dequeue_tg(struct throtl_data
*td
, struct throtl_grp
*tg
)
359 if (throtl_tg_on_rr(tg
))
360 __throtl_dequeue_tg(td
, tg
);
363 static void throtl_schedule_next_dispatch(struct throtl_data
*td
)
365 struct throtl_rb_root
*st
= &td
->tg_service_tree
;
368 * If there are more bios pending, schedule more work.
370 if (!total_nr_queued(td
))
375 update_min_dispatch_time(st
);
377 if (time_before_eq(st
->min_disptime
, jiffies
))
378 throtl_schedule_delayed_work(td
, 0);
380 throtl_schedule_delayed_work(td
, (st
->min_disptime
- jiffies
));
384 throtl_start_new_slice(struct throtl_data
*td
, struct throtl_grp
*tg
, bool rw
)
386 tg
->bytes_disp
[rw
] = 0;
388 tg
->slice_start
[rw
] = jiffies
;
389 tg
->slice_end
[rw
] = jiffies
+ throtl_slice
;
390 throtl_log_tg(td
, tg
, "[%c] new slice start=%lu end=%lu jiffies=%lu",
391 rw
== READ
? 'R' : 'W', tg
->slice_start
[rw
],
392 tg
->slice_end
[rw
], jiffies
);
395 static inline void throtl_set_slice_end(struct throtl_data
*td
,
396 struct throtl_grp
*tg
, bool rw
, unsigned long jiffy_end
)
398 tg
->slice_end
[rw
] = roundup(jiffy_end
, throtl_slice
);
401 static inline void throtl_extend_slice(struct throtl_data
*td
,
402 struct throtl_grp
*tg
, bool rw
, unsigned long jiffy_end
)
404 tg
->slice_end
[rw
] = roundup(jiffy_end
, throtl_slice
);
405 throtl_log_tg(td
, tg
, "[%c] extend slice start=%lu end=%lu jiffies=%lu",
406 rw
== READ
? 'R' : 'W', tg
->slice_start
[rw
],
407 tg
->slice_end
[rw
], jiffies
);
410 /* Determine if previously allocated or extended slice is complete or not */
412 throtl_slice_used(struct throtl_data
*td
, struct throtl_grp
*tg
, bool rw
)
414 if (time_in_range(jiffies
, tg
->slice_start
[rw
], tg
->slice_end
[rw
]))
420 /* Trim the used slices and adjust slice start accordingly */
422 throtl_trim_slice(struct throtl_data
*td
, struct throtl_grp
*tg
, bool rw
)
424 unsigned long nr_slices
, time_elapsed
, io_trim
;
427 BUG_ON(time_before(tg
->slice_end
[rw
], tg
->slice_start
[rw
]));
430 * If bps are unlimited (-1), then time slice don't get
431 * renewed. Don't try to trim the slice if slice is used. A new
432 * slice will start when appropriate.
434 if (throtl_slice_used(td
, tg
, rw
))
438 * A bio has been dispatched. Also adjust slice_end. It might happen
439 * that initially cgroup limit was very low resulting in high
440 * slice_end, but later limit was bumped up and bio was dispached
441 * sooner, then we need to reduce slice_end. A high bogus slice_end
442 * is bad because it does not allow new slice to start.
445 throtl_set_slice_end(td
, tg
, rw
, jiffies
+ throtl_slice
);
447 time_elapsed
= jiffies
- tg
->slice_start
[rw
];
449 nr_slices
= time_elapsed
/ throtl_slice
;
453 tmp
= tg
->bps
[rw
] * throtl_slice
* nr_slices
;
457 io_trim
= (tg
->iops
[rw
] * throtl_slice
* nr_slices
)/HZ
;
459 if (!bytes_trim
&& !io_trim
)
462 if (tg
->bytes_disp
[rw
] >= bytes_trim
)
463 tg
->bytes_disp
[rw
] -= bytes_trim
;
465 tg
->bytes_disp
[rw
] = 0;
467 if (tg
->io_disp
[rw
] >= io_trim
)
468 tg
->io_disp
[rw
] -= io_trim
;
472 tg
->slice_start
[rw
] += nr_slices
* throtl_slice
;
474 throtl_log_tg(td
, tg
, "[%c] trim slice nr=%lu bytes=%llu io=%lu"
475 " start=%lu end=%lu jiffies=%lu",
476 rw
== READ
? 'R' : 'W', nr_slices
, bytes_trim
, io_trim
,
477 tg
->slice_start
[rw
], tg
->slice_end
[rw
], jiffies
);
480 static bool tg_with_in_iops_limit(struct throtl_data
*td
, struct throtl_grp
*tg
,
481 struct bio
*bio
, unsigned long *wait
)
483 bool rw
= bio_data_dir(bio
);
484 unsigned int io_allowed
;
485 unsigned long jiffy_elapsed
, jiffy_wait
, jiffy_elapsed_rnd
;
488 jiffy_elapsed
= jiffy_elapsed_rnd
= jiffies
- tg
->slice_start
[rw
];
490 /* Slice has just started. Consider one slice interval */
492 jiffy_elapsed_rnd
= throtl_slice
;
494 jiffy_elapsed_rnd
= roundup(jiffy_elapsed_rnd
, throtl_slice
);
497 * jiffy_elapsed_rnd should not be a big value as minimum iops can be
498 * 1 then at max jiffy elapsed should be equivalent of 1 second as we
499 * will allow dispatch after 1 second and after that slice should
503 tmp
= (u64
)tg
->iops
[rw
] * jiffy_elapsed_rnd
;
507 io_allowed
= UINT_MAX
;
511 if (tg
->io_disp
[rw
] + 1 <= io_allowed
) {
517 /* Calc approx time to dispatch */
518 jiffy_wait
= ((tg
->io_disp
[rw
] + 1) * HZ
)/tg
->iops
[rw
] + 1;
520 if (jiffy_wait
> jiffy_elapsed
)
521 jiffy_wait
= jiffy_wait
- jiffy_elapsed
;
530 static bool tg_with_in_bps_limit(struct throtl_data
*td
, struct throtl_grp
*tg
,
531 struct bio
*bio
, unsigned long *wait
)
533 bool rw
= bio_data_dir(bio
);
534 u64 bytes_allowed
, extra_bytes
, tmp
;
535 unsigned long jiffy_elapsed
, jiffy_wait
, jiffy_elapsed_rnd
;
537 jiffy_elapsed
= jiffy_elapsed_rnd
= jiffies
- tg
->slice_start
[rw
];
539 /* Slice has just started. Consider one slice interval */
541 jiffy_elapsed_rnd
= throtl_slice
;
543 jiffy_elapsed_rnd
= roundup(jiffy_elapsed_rnd
, throtl_slice
);
545 tmp
= tg
->bps
[rw
] * jiffy_elapsed_rnd
;
549 if (tg
->bytes_disp
[rw
] + bio
->bi_size
<= bytes_allowed
) {
555 /* Calc approx time to dispatch */
556 extra_bytes
= tg
->bytes_disp
[rw
] + bio
->bi_size
- bytes_allowed
;
557 jiffy_wait
= div64_u64(extra_bytes
* HZ
, tg
->bps
[rw
]);
563 * This wait time is without taking into consideration the rounding
564 * up we did. Add that time also.
566 jiffy_wait
= jiffy_wait
+ (jiffy_elapsed_rnd
- jiffy_elapsed
);
572 static bool tg_no_rule_group(struct throtl_grp
*tg
, bool rw
) {
573 if (tg
->bps
[rw
] == -1 && tg
->iops
[rw
] == -1)
579 * Returns whether one can dispatch a bio or not. Also returns approx number
580 * of jiffies to wait before this bio is with-in IO rate and can be dispatched
582 static bool tg_may_dispatch(struct throtl_data
*td
, struct throtl_grp
*tg
,
583 struct bio
*bio
, unsigned long *wait
)
585 bool rw
= bio_data_dir(bio
);
586 unsigned long bps_wait
= 0, iops_wait
= 0, max_wait
= 0;
589 * Currently whole state machine of group depends on first bio
590 * queued in the group bio list. So one should not be calling
591 * this function with a different bio if there are other bios
594 BUG_ON(tg
->nr_queued
[rw
] && bio
!= bio_list_peek(&tg
->bio_lists
[rw
]));
596 /* If tg->bps = -1, then BW is unlimited */
597 if (tg
->bps
[rw
] == -1 && tg
->iops
[rw
] == -1) {
604 * If previous slice expired, start a new one otherwise renew/extend
605 * existing slice to make sure it is at least throtl_slice interval
608 if (throtl_slice_used(td
, tg
, rw
))
609 throtl_start_new_slice(td
, tg
, rw
);
611 if (time_before(tg
->slice_end
[rw
], jiffies
+ throtl_slice
))
612 throtl_extend_slice(td
, tg
, rw
, jiffies
+ throtl_slice
);
615 if (tg_with_in_bps_limit(td
, tg
, bio
, &bps_wait
)
616 && tg_with_in_iops_limit(td
, tg
, bio
, &iops_wait
)) {
622 max_wait
= max(bps_wait
, iops_wait
);
627 if (time_before(tg
->slice_end
[rw
], jiffies
+ max_wait
))
628 throtl_extend_slice(td
, tg
, rw
, jiffies
+ max_wait
);
633 static void throtl_charge_bio(struct throtl_grp
*tg
, struct bio
*bio
)
635 bool rw
= bio_data_dir(bio
);
636 bool sync
= rw_is_sync(bio
->bi_rw
);
638 /* Charge the bio to the group */
639 tg
->bytes_disp
[rw
] += bio
->bi_size
;
642 blkiocg_update_dispatch_stats(&tg
->blkg
, bio
->bi_size
, rw
, sync
);
645 static void throtl_add_bio_tg(struct throtl_data
*td
, struct throtl_grp
*tg
,
648 bool rw
= bio_data_dir(bio
);
650 bio_list_add(&tg
->bio_lists
[rw
], bio
);
651 /* Take a bio reference on tg */
652 throtl_ref_get_tg(tg
);
655 throtl_enqueue_tg(td
, tg
);
658 static void tg_update_disptime(struct throtl_data
*td
, struct throtl_grp
*tg
)
660 unsigned long read_wait
= -1, write_wait
= -1, min_wait
= -1, disptime
;
663 if ((bio
= bio_list_peek(&tg
->bio_lists
[READ
])))
664 tg_may_dispatch(td
, tg
, bio
, &read_wait
);
666 if ((bio
= bio_list_peek(&tg
->bio_lists
[WRITE
])))
667 tg_may_dispatch(td
, tg
, bio
, &write_wait
);
669 min_wait
= min(read_wait
, write_wait
);
670 disptime
= jiffies
+ min_wait
;
672 /* Update dispatch time */
673 throtl_dequeue_tg(td
, tg
);
674 tg
->disptime
= disptime
;
675 throtl_enqueue_tg(td
, tg
);
678 static void tg_dispatch_one_bio(struct throtl_data
*td
, struct throtl_grp
*tg
,
679 bool rw
, struct bio_list
*bl
)
683 bio
= bio_list_pop(&tg
->bio_lists
[rw
]);
685 /* Drop bio reference on tg */
688 BUG_ON(td
->nr_queued
[rw
] <= 0);
691 throtl_charge_bio(tg
, bio
);
692 bio_list_add(bl
, bio
);
693 bio
->bi_rw
|= REQ_THROTTLED
;
695 throtl_trim_slice(td
, tg
, rw
);
698 static int throtl_dispatch_tg(struct throtl_data
*td
, struct throtl_grp
*tg
,
701 unsigned int nr_reads
= 0, nr_writes
= 0;
702 unsigned int max_nr_reads
= throtl_grp_quantum
*3/4;
703 unsigned int max_nr_writes
= throtl_grp_quantum
- max_nr_reads
;
706 /* Try to dispatch 75% READS and 25% WRITES */
708 while ((bio
= bio_list_peek(&tg
->bio_lists
[READ
]))
709 && tg_may_dispatch(td
, tg
, bio
, NULL
)) {
711 tg_dispatch_one_bio(td
, tg
, bio_data_dir(bio
), bl
);
714 if (nr_reads
>= max_nr_reads
)
718 while ((bio
= bio_list_peek(&tg
->bio_lists
[WRITE
]))
719 && tg_may_dispatch(td
, tg
, bio
, NULL
)) {
721 tg_dispatch_one_bio(td
, tg
, bio_data_dir(bio
), bl
);
724 if (nr_writes
>= max_nr_writes
)
728 return nr_reads
+ nr_writes
;
731 static int throtl_select_dispatch(struct throtl_data
*td
, struct bio_list
*bl
)
733 unsigned int nr_disp
= 0;
734 struct throtl_grp
*tg
;
735 struct throtl_rb_root
*st
= &td
->tg_service_tree
;
738 tg
= throtl_rb_first(st
);
743 if (time_before(jiffies
, tg
->disptime
))
746 throtl_dequeue_tg(td
, tg
);
748 nr_disp
+= throtl_dispatch_tg(td
, tg
, bl
);
750 if (tg
->nr_queued
[0] || tg
->nr_queued
[1]) {
751 tg_update_disptime(td
, tg
);
752 throtl_enqueue_tg(td
, tg
);
755 if (nr_disp
>= throtl_quantum
)
762 static void throtl_process_limit_change(struct throtl_data
*td
)
764 struct throtl_grp
*tg
;
765 struct hlist_node
*pos
, *n
;
767 if (!td
->limits_changed
)
770 xchg(&td
->limits_changed
, false);
772 throtl_log(td
, "limits changed");
774 hlist_for_each_entry_safe(tg
, pos
, n
, &td
->tg_list
, tg_node
) {
775 if (!tg
->limits_changed
)
778 if (!xchg(&tg
->limits_changed
, false))
781 throtl_log_tg(td
, tg
, "limit change rbps=%llu wbps=%llu"
782 " riops=%u wiops=%u", tg
->bps
[READ
], tg
->bps
[WRITE
],
783 tg
->iops
[READ
], tg
->iops
[WRITE
]);
786 * Restart the slices for both READ and WRITES. It
787 * might happen that a group's limit are dropped
788 * suddenly and we don't want to account recently
789 * dispatched IO with new low rate
791 throtl_start_new_slice(td
, tg
, 0);
792 throtl_start_new_slice(td
, tg
, 1);
794 if (throtl_tg_on_rr(tg
))
795 tg_update_disptime(td
, tg
);
799 /* Dispatch throttled bios. Should be called without queue lock held. */
800 static int throtl_dispatch(struct request_queue
*q
)
802 struct throtl_data
*td
= q
->td
;
803 unsigned int nr_disp
= 0;
804 struct bio_list bio_list_on_stack
;
806 struct blk_plug plug
;
808 spin_lock_irq(q
->queue_lock
);
810 throtl_process_limit_change(td
);
812 if (!total_nr_queued(td
))
815 bio_list_init(&bio_list_on_stack
);
817 throtl_log(td
, "dispatch nr_queued=%u read=%u write=%u",
818 total_nr_queued(td
), td
->nr_queued
[READ
],
819 td
->nr_queued
[WRITE
]);
821 nr_disp
= throtl_select_dispatch(td
, &bio_list_on_stack
);
824 throtl_log(td
, "bios disp=%u", nr_disp
);
826 throtl_schedule_next_dispatch(td
);
828 spin_unlock_irq(q
->queue_lock
);
831 * If we dispatched some requests, unplug the queue to make sure
835 blk_start_plug(&plug
);
836 while((bio
= bio_list_pop(&bio_list_on_stack
)))
837 generic_make_request(bio
);
838 blk_finish_plug(&plug
);
843 void blk_throtl_work(struct work_struct
*work
)
845 struct throtl_data
*td
= container_of(work
, struct throtl_data
,
847 struct request_queue
*q
= td
->queue
;
852 /* Call with queue lock held */
854 throtl_schedule_delayed_work(struct throtl_data
*td
, unsigned long delay
)
857 struct delayed_work
*dwork
= &td
->throtl_work
;
859 /* schedule work if limits changed even if no bio is queued */
860 if (total_nr_queued(td
) || td
->limits_changed
) {
862 * We might have a work scheduled to be executed in future.
863 * Cancel that and schedule a new one.
865 __cancel_delayed_work(dwork
);
866 queue_delayed_work(kthrotld_workqueue
, dwork
, delay
);
867 throtl_log(td
, "schedule work. delay=%lu jiffies=%lu",
873 throtl_destroy_tg(struct throtl_data
*td
, struct throtl_grp
*tg
)
875 /* Something wrong if we are trying to remove same group twice */
876 BUG_ON(hlist_unhashed(&tg
->tg_node
));
878 hlist_del_init(&tg
->tg_node
);
881 * Put the reference taken at the time of creation so that when all
882 * queues are gone, group can be destroyed.
885 td
->nr_undestroyed_grps
--;
888 static bool throtl_release_tgs(struct throtl_data
*td
, bool release_root
)
890 struct hlist_node
*pos
, *n
;
891 struct throtl_grp
*tg
;
894 hlist_for_each_entry_safe(tg
, pos
, n
, &td
->tg_list
, tg_node
) {
896 if (!release_root
&& tg
== td
->root_tg
)
900 * If cgroup removal path got to blk_group first and removed
901 * it from cgroup list, then it will take care of destroying
904 if (!blkiocg_del_blkio_group(&tg
->blkg
))
905 throtl_destroy_tg(td
, tg
);
913 * Blk cgroup controller notification saying that blkio_group object is being
914 * delinked as associated cgroup object is going away. That also means that
915 * no new IO will come in this group. So get rid of this group as soon as
916 * any pending IO in the group is finished.
918 * This function is called under rcu_read_lock(). @q is the rcu protected
919 * pointer. That means @q is a valid request_queue pointer as long as we
922 * @q was fetched from blkio_group under blkio_cgroup->lock. That means
923 * it should not be NULL as even if queue was going away, cgroup deltion
924 * path got to it first.
926 void throtl_unlink_blkio_group(struct request_queue
*q
,
927 struct blkio_group
*blkg
)
931 spin_lock_irqsave(q
->queue_lock
, flags
);
932 throtl_destroy_tg(q
->td
, tg_of_blkg(blkg
));
933 spin_unlock_irqrestore(q
->queue_lock
, flags
);
936 static bool throtl_clear_queue(struct request_queue
*q
)
938 lockdep_assert_held(q
->queue_lock
);
941 * Clear tgs but leave the root one alone. This is necessary
942 * because root_tg is expected to be persistent and safe because
943 * blk-throtl can never be disabled while @q is alive. This is a
944 * kludge to prepare for unified blkg. This whole function will be
947 return throtl_release_tgs(q
->td
, false);
950 static void throtl_update_blkio_group_common(struct throtl_data
*td
,
951 struct throtl_grp
*tg
)
953 xchg(&tg
->limits_changed
, true);
954 xchg(&td
->limits_changed
, true);
955 /* Schedule a work now to process the limit change */
956 throtl_schedule_delayed_work(td
, 0);
960 * For all update functions, @q should be a valid pointer because these
961 * update functions are called under blkcg_lock, that means, blkg is
962 * valid and in turn @q is valid. queue exit path can not race because
965 * Can not take queue lock in update functions as queue lock under blkcg_lock
966 * is not allowed. Under other paths we take blkcg_lock under queue_lock.
968 static void throtl_update_blkio_group_read_bps(struct request_queue
*q
,
969 struct blkio_group
*blkg
, u64 read_bps
)
971 struct throtl_grp
*tg
= tg_of_blkg(blkg
);
973 tg
->bps
[READ
] = read_bps
;
974 throtl_update_blkio_group_common(q
->td
, tg
);
977 static void throtl_update_blkio_group_write_bps(struct request_queue
*q
,
978 struct blkio_group
*blkg
, u64 write_bps
)
980 struct throtl_grp
*tg
= tg_of_blkg(blkg
);
982 tg
->bps
[WRITE
] = write_bps
;
983 throtl_update_blkio_group_common(q
->td
, tg
);
986 static void throtl_update_blkio_group_read_iops(struct request_queue
*q
,
987 struct blkio_group
*blkg
, unsigned int read_iops
)
989 struct throtl_grp
*tg
= tg_of_blkg(blkg
);
991 tg
->iops
[READ
] = read_iops
;
992 throtl_update_blkio_group_common(q
->td
, tg
);
995 static void throtl_update_blkio_group_write_iops(struct request_queue
*q
,
996 struct blkio_group
*blkg
, unsigned int write_iops
)
998 struct throtl_grp
*tg
= tg_of_blkg(blkg
);
1000 tg
->iops
[WRITE
] = write_iops
;
1001 throtl_update_blkio_group_common(q
->td
, tg
);
1004 static void throtl_shutdown_wq(struct request_queue
*q
)
1006 struct throtl_data
*td
= q
->td
;
1008 cancel_delayed_work_sync(&td
->throtl_work
);
1011 static struct blkio_policy_type blkio_policy_throtl
= {
1013 .blkio_alloc_group_fn
= throtl_alloc_blkio_group
,
1014 .blkio_link_group_fn
= throtl_link_blkio_group
,
1015 .blkio_unlink_group_fn
= throtl_unlink_blkio_group
,
1016 .blkio_clear_queue_fn
= throtl_clear_queue
,
1017 .blkio_update_group_read_bps_fn
=
1018 throtl_update_blkio_group_read_bps
,
1019 .blkio_update_group_write_bps_fn
=
1020 throtl_update_blkio_group_write_bps
,
1021 .blkio_update_group_read_iops_fn
=
1022 throtl_update_blkio_group_read_iops
,
1023 .blkio_update_group_write_iops_fn
=
1024 throtl_update_blkio_group_write_iops
,
1026 .plid
= BLKIO_POLICY_THROTL
,
1029 bool blk_throtl_bio(struct request_queue
*q
, struct bio
*bio
)
1031 struct throtl_data
*td
= q
->td
;
1032 struct throtl_grp
*tg
;
1033 bool rw
= bio_data_dir(bio
), update_disptime
= true;
1034 struct blkio_cgroup
*blkcg
;
1035 bool throttled
= false;
1037 if (bio
->bi_rw
& REQ_THROTTLED
) {
1038 bio
->bi_rw
&= ~REQ_THROTTLED
;
1043 * A throtl_grp pointer retrieved under rcu can be used to access
1044 * basic fields like stats and io rates. If a group has no rules,
1045 * just update the dispatch stats in lockless manner and return.
1048 blkcg
= task_blkio_cgroup(current
);
1049 tg
= throtl_lookup_tg(td
, blkcg
);
1051 if (tg_no_rule_group(tg
, rw
)) {
1052 blkiocg_update_dispatch_stats(&tg
->blkg
, bio
->bi_size
,
1053 rw
, rw_is_sync(bio
->bi_rw
));
1054 goto out_unlock_rcu
;
1059 * Either group has not been allocated yet or it is not an unlimited
1062 spin_lock_irq(q
->queue_lock
);
1063 tg
= throtl_lookup_create_tg(td
, blkcg
);
1067 if (tg
->nr_queued
[rw
]) {
1069 * There is already another bio queued in same dir. No
1070 * need to update dispatch time.
1072 update_disptime
= false;
1077 /* Bio is with-in rate limit of group */
1078 if (tg_may_dispatch(td
, tg
, bio
, NULL
)) {
1079 throtl_charge_bio(tg
, bio
);
1082 * We need to trim slice even when bios are not being queued
1083 * otherwise it might happen that a bio is not queued for
1084 * a long time and slice keeps on extending and trim is not
1085 * called for a long time. Now if limits are reduced suddenly
1086 * we take into account all the IO dispatched so far at new
1087 * low rate and * newly queued IO gets a really long dispatch
1090 * So keep on trimming slice even if bio is not queued.
1092 throtl_trim_slice(td
, tg
, rw
);
1097 throtl_log_tg(td
, tg
, "[%c] bio. bdisp=%llu sz=%u bps=%llu"
1098 " iodisp=%u iops=%u queued=%d/%d",
1099 rw
== READ
? 'R' : 'W',
1100 tg
->bytes_disp
[rw
], bio
->bi_size
, tg
->bps
[rw
],
1101 tg
->io_disp
[rw
], tg
->iops
[rw
],
1102 tg
->nr_queued
[READ
], tg
->nr_queued
[WRITE
]);
1104 throtl_add_bio_tg(q
->td
, tg
, bio
);
1107 if (update_disptime
) {
1108 tg_update_disptime(td
, tg
);
1109 throtl_schedule_next_dispatch(td
);
1113 spin_unlock_irq(q
->queue_lock
);
1121 * blk_throtl_drain - drain throttled bios
1122 * @q: request_queue to drain throttled bios for
1124 * Dispatch all currently throttled bios on @q through ->make_request_fn().
1126 void blk_throtl_drain(struct request_queue
*q
)
1127 __releases(q
->queue_lock
) __acquires(q
->queue_lock
)
1129 struct throtl_data
*td
= q
->td
;
1130 struct throtl_rb_root
*st
= &td
->tg_service_tree
;
1131 struct throtl_grp
*tg
;
1135 WARN_ON_ONCE(!queue_is_locked(q
));
1139 while ((tg
= throtl_rb_first(st
))) {
1140 throtl_dequeue_tg(td
, tg
);
1142 while ((bio
= bio_list_peek(&tg
->bio_lists
[READ
])))
1143 tg_dispatch_one_bio(td
, tg
, bio_data_dir(bio
), &bl
);
1144 while ((bio
= bio_list_peek(&tg
->bio_lists
[WRITE
])))
1145 tg_dispatch_one_bio(td
, tg
, bio_data_dir(bio
), &bl
);
1147 spin_unlock_irq(q
->queue_lock
);
1149 while ((bio
= bio_list_pop(&bl
)))
1150 generic_make_request(bio
);
1152 spin_lock_irq(q
->queue_lock
);
1155 int blk_throtl_init(struct request_queue
*q
)
1157 struct throtl_data
*td
;
1158 struct blkio_group
*blkg
;
1160 td
= kzalloc_node(sizeof(*td
), GFP_KERNEL
, q
->node
);
1164 INIT_HLIST_HEAD(&td
->tg_list
);
1165 td
->tg_service_tree
= THROTL_RB_ROOT
;
1166 td
->limits_changed
= false;
1167 INIT_DELAYED_WORK(&td
->throtl_work
, blk_throtl_work
);
1172 /* alloc and init root group. */
1174 spin_lock_irq(q
->queue_lock
);
1176 blkg
= blkg_lookup_create(&blkio_root_cgroup
, q
, BLKIO_POLICY_THROTL
,
1179 td
->root_tg
= tg_of_blkg(blkg
);
1181 spin_unlock_irq(q
->queue_lock
);
1191 void blk_throtl_exit(struct request_queue
*q
)
1193 struct throtl_data
*td
= q
->td
;
1198 throtl_shutdown_wq(q
);
1200 spin_lock_irq(q
->queue_lock
);
1201 throtl_release_tgs(td
, true);
1203 /* If there are other groups */
1204 if (td
->nr_undestroyed_grps
> 0)
1207 spin_unlock_irq(q
->queue_lock
);
1210 * Wait for tg->blkg->q accessors to exit their grace periods.
1211 * Do this wait only if there are other undestroyed groups out
1212 * there (other than root group). This can happen if cgroup deletion
1213 * path claimed the responsibility of cleaning up a group before
1214 * queue cleanup code get to the group.
1216 * Do not call synchronize_rcu() unconditionally as there are drivers
1217 * which create/delete request queue hundreds of times during scan/boot
1218 * and synchronize_rcu() can take significant time and slow down boot.
1224 * Just being safe to make sure after previous flush if some body did
1225 * update limits through cgroup and another work got queued, cancel
1228 throtl_shutdown_wq(q
);
1231 void blk_throtl_release(struct request_queue
*q
)
1236 static int __init
throtl_init(void)
1238 kthrotld_workqueue
= alloc_workqueue("kthrotld", WQ_MEM_RECLAIM
, 0);
1239 if (!kthrotld_workqueue
)
1240 panic("Failed to create kthrotld\n");
1242 blkio_policy_register(&blkio_policy_throtl
);
1246 module_init(throtl_init
);