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 static struct blkcg_policy blkcg_policy_throtl
;
26 /* A workqueue to queue throttle related work */
27 static struct workqueue_struct
*kthrotld_workqueue
;
28 static void throtl_schedule_delayed_work(struct throtl_data
*td
,
31 struct throtl_rb_root
{
35 unsigned long min_disptime
;
38 #define THROTL_RB_ROOT (struct throtl_rb_root) { .rb = RB_ROOT, .left = NULL, \
39 .count = 0, .min_disptime = 0}
41 #define rb_entry_tg(node) rb_entry((node), struct throtl_grp, rb_node)
43 /* Per-cpu group stats */
45 /* total bytes transferred */
46 struct blkg_rwstat service_bytes
;
47 /* total IOs serviced, post merge */
48 struct blkg_rwstat serviced
;
52 /* must be the first member */
53 struct blkg_policy_data pd
;
55 /* active throtl group service_tree member */
56 struct rb_node rb_node
;
59 * Dispatch time in jiffies. This is the estimated time when group
60 * will unthrottle and is ready to dispatch more bio. It is used as
61 * key to sort active groups in service tree.
63 unsigned long disptime
;
67 /* Two lists for READ and WRITE */
68 struct bio_list bio_lists
[2];
70 /* Number of queued bios on READ and WRITE lists */
71 unsigned int nr_queued
[2];
73 /* bytes per second rate limits */
79 /* Number of bytes disptached in current slice */
80 uint64_t bytes_disp
[2];
81 /* Number of bio's dispatched in current slice */
82 unsigned int io_disp
[2];
84 /* When did we start a new slice */
85 unsigned long slice_start
[2];
86 unsigned long slice_end
[2];
88 /* Some throttle limits got updated for the group */
91 /* Per cpu stats pointer */
92 struct tg_stats_cpu __percpu
*stats_cpu
;
94 /* List of tgs waiting for per cpu stats memory to be allocated */
95 struct list_head stats_alloc_node
;
100 /* service tree for active throtl groups */
101 struct throtl_rb_root tg_service_tree
;
103 struct request_queue
*queue
;
105 /* Total Number of queued bios on READ and WRITE lists */
106 unsigned int nr_queued
[2];
109 * number of total undestroyed groups
111 unsigned int nr_undestroyed_grps
;
113 /* Work for dispatching throttled bios */
114 struct delayed_work throtl_work
;
119 /* list and work item to allocate percpu group stats */
120 static DEFINE_SPINLOCK(tg_stats_alloc_lock
);
121 static LIST_HEAD(tg_stats_alloc_list
);
123 static void tg_stats_alloc_fn(struct work_struct
*);
124 static DECLARE_DELAYED_WORK(tg_stats_alloc_work
, tg_stats_alloc_fn
);
126 static inline struct throtl_grp
*pd_to_tg(struct blkg_policy_data
*pd
)
128 return pd
? container_of(pd
, struct throtl_grp
, pd
) : NULL
;
131 static inline struct throtl_grp
*blkg_to_tg(struct blkcg_gq
*blkg
)
133 return pd_to_tg(blkg_to_pd(blkg
, &blkcg_policy_throtl
));
136 static inline struct blkcg_gq
*tg_to_blkg(struct throtl_grp
*tg
)
138 return pd_to_blkg(&tg
->pd
);
141 static inline struct throtl_grp
*td_root_tg(struct throtl_data
*td
)
143 return blkg_to_tg(td
->queue
->root_blkg
);
146 enum tg_state_flags
{
147 THROTL_TG_FLAG_on_rr
= 0, /* on round-robin busy list */
150 #define THROTL_TG_FNS(name) \
151 static inline void throtl_mark_tg_##name(struct throtl_grp *tg) \
153 (tg)->flags |= (1 << THROTL_TG_FLAG_##name); \
155 static inline void throtl_clear_tg_##name(struct throtl_grp *tg) \
157 (tg)->flags &= ~(1 << THROTL_TG_FLAG_##name); \
159 static inline int throtl_tg_##name(const struct throtl_grp *tg) \
161 return ((tg)->flags & (1 << THROTL_TG_FLAG_##name)) != 0; \
164 THROTL_TG_FNS(on_rr
);
166 #define throtl_log_tg(td, tg, fmt, args...) do { \
169 blkg_path(tg_to_blkg(tg), __pbuf, sizeof(__pbuf)); \
170 blk_add_trace_msg((td)->queue, "throtl %s " fmt, __pbuf, ##args); \
173 #define throtl_log(td, fmt, args...) \
174 blk_add_trace_msg((td)->queue, "throtl " fmt, ##args)
176 static inline unsigned int total_nr_queued(struct throtl_data
*td
)
178 return td
->nr_queued
[0] + td
->nr_queued
[1];
182 * Worker for allocating per cpu stat for tgs. This is scheduled on the
183 * system_nrt_wq once there are some groups on the alloc_list waiting for
186 static void tg_stats_alloc_fn(struct work_struct
*work
)
188 static struct tg_stats_cpu
*stats_cpu
; /* this fn is non-reentrant */
189 struct delayed_work
*dwork
= to_delayed_work(work
);
194 stats_cpu
= alloc_percpu(struct tg_stats_cpu
);
196 /* allocation failed, try again after some time */
197 queue_delayed_work(system_nrt_wq
, dwork
,
198 msecs_to_jiffies(10));
203 spin_lock_irq(&tg_stats_alloc_lock
);
205 if (!list_empty(&tg_stats_alloc_list
)) {
206 struct throtl_grp
*tg
= list_first_entry(&tg_stats_alloc_list
,
209 swap(tg
->stats_cpu
, stats_cpu
);
210 list_del_init(&tg
->stats_alloc_node
);
213 empty
= list_empty(&tg_stats_alloc_list
);
214 spin_unlock_irq(&tg_stats_alloc_lock
);
219 static void throtl_pd_init(struct blkcg_gq
*blkg
)
221 struct throtl_grp
*tg
= blkg_to_tg(blkg
);
223 RB_CLEAR_NODE(&tg
->rb_node
);
224 bio_list_init(&tg
->bio_lists
[0]);
225 bio_list_init(&tg
->bio_lists
[1]);
226 tg
->limits_changed
= false;
231 tg
->iops
[WRITE
] = -1;
234 * Ugh... We need to perform per-cpu allocation for tg->stats_cpu
235 * but percpu allocator can't be called from IO path. Queue tg on
236 * tg_stats_alloc_list and allocate from work item.
238 spin_lock(&tg_stats_alloc_lock
);
239 list_add(&tg
->stats_alloc_node
, &tg_stats_alloc_list
);
240 queue_delayed_work(system_nrt_wq
, &tg_stats_alloc_work
, 0);
241 spin_unlock(&tg_stats_alloc_lock
);
244 static void throtl_pd_exit(struct blkcg_gq
*blkg
)
246 struct throtl_grp
*tg
= blkg_to_tg(blkg
);
248 spin_lock(&tg_stats_alloc_lock
);
249 list_del_init(&tg
->stats_alloc_node
);
250 spin_unlock(&tg_stats_alloc_lock
);
252 free_percpu(tg
->stats_cpu
);
255 static void throtl_pd_reset_stats(struct blkcg_gq
*blkg
)
257 struct throtl_grp
*tg
= blkg_to_tg(blkg
);
260 if (tg
->stats_cpu
== NULL
)
263 for_each_possible_cpu(cpu
) {
264 struct tg_stats_cpu
*sc
= per_cpu_ptr(tg
->stats_cpu
, cpu
);
266 blkg_rwstat_reset(&sc
->service_bytes
);
267 blkg_rwstat_reset(&sc
->serviced
);
271 static struct throtl_grp
*throtl_lookup_tg(struct throtl_data
*td
,
275 * This is the common case when there are no blkcgs. Avoid lookup
278 if (blkcg
== &blkcg_root
)
279 return td_root_tg(td
);
281 return blkg_to_tg(blkg_lookup(blkcg
, td
->queue
));
284 static struct throtl_grp
*throtl_lookup_create_tg(struct throtl_data
*td
,
287 struct request_queue
*q
= td
->queue
;
288 struct throtl_grp
*tg
= NULL
;
291 * This is the common case when there are no blkcgs. Avoid lookup
294 if (blkcg
== &blkcg_root
) {
297 struct blkcg_gq
*blkg
;
299 blkg
= blkg_lookup_create(blkcg
, q
);
301 /* if %NULL and @q is alive, fall back to root_tg */
303 tg
= blkg_to_tg(blkg
);
304 else if (!blk_queue_dead(q
))
311 static struct throtl_grp
*throtl_rb_first(struct throtl_rb_root
*root
)
313 /* Service tree is empty */
318 root
->left
= rb_first(&root
->rb
);
321 return rb_entry_tg(root
->left
);
326 static void rb_erase_init(struct rb_node
*n
, struct rb_root
*root
)
332 static void throtl_rb_erase(struct rb_node
*n
, struct throtl_rb_root
*root
)
336 rb_erase_init(n
, &root
->rb
);
340 static void update_min_dispatch_time(struct throtl_rb_root
*st
)
342 struct throtl_grp
*tg
;
344 tg
= throtl_rb_first(st
);
348 st
->min_disptime
= tg
->disptime
;
352 tg_service_tree_add(struct throtl_rb_root
*st
, struct throtl_grp
*tg
)
354 struct rb_node
**node
= &st
->rb
.rb_node
;
355 struct rb_node
*parent
= NULL
;
356 struct throtl_grp
*__tg
;
357 unsigned long key
= tg
->disptime
;
360 while (*node
!= NULL
) {
362 __tg
= rb_entry_tg(parent
);
364 if (time_before(key
, __tg
->disptime
))
365 node
= &parent
->rb_left
;
367 node
= &parent
->rb_right
;
373 st
->left
= &tg
->rb_node
;
375 rb_link_node(&tg
->rb_node
, parent
, node
);
376 rb_insert_color(&tg
->rb_node
, &st
->rb
);
379 static void __throtl_enqueue_tg(struct throtl_data
*td
, struct throtl_grp
*tg
)
381 struct throtl_rb_root
*st
= &td
->tg_service_tree
;
383 tg_service_tree_add(st
, tg
);
384 throtl_mark_tg_on_rr(tg
);
388 static void throtl_enqueue_tg(struct throtl_data
*td
, struct throtl_grp
*tg
)
390 if (!throtl_tg_on_rr(tg
))
391 __throtl_enqueue_tg(td
, tg
);
394 static void __throtl_dequeue_tg(struct throtl_data
*td
, struct throtl_grp
*tg
)
396 throtl_rb_erase(&tg
->rb_node
, &td
->tg_service_tree
);
397 throtl_clear_tg_on_rr(tg
);
400 static void throtl_dequeue_tg(struct throtl_data
*td
, struct throtl_grp
*tg
)
402 if (throtl_tg_on_rr(tg
))
403 __throtl_dequeue_tg(td
, tg
);
406 static void throtl_schedule_next_dispatch(struct throtl_data
*td
)
408 struct throtl_rb_root
*st
= &td
->tg_service_tree
;
411 * If there are more bios pending, schedule more work.
413 if (!total_nr_queued(td
))
418 update_min_dispatch_time(st
);
420 if (time_before_eq(st
->min_disptime
, jiffies
))
421 throtl_schedule_delayed_work(td
, 0);
423 throtl_schedule_delayed_work(td
, (st
->min_disptime
- jiffies
));
427 throtl_start_new_slice(struct throtl_data
*td
, struct throtl_grp
*tg
, bool rw
)
429 tg
->bytes_disp
[rw
] = 0;
431 tg
->slice_start
[rw
] = jiffies
;
432 tg
->slice_end
[rw
] = jiffies
+ throtl_slice
;
433 throtl_log_tg(td
, tg
, "[%c] new slice start=%lu end=%lu jiffies=%lu",
434 rw
== READ
? 'R' : 'W', tg
->slice_start
[rw
],
435 tg
->slice_end
[rw
], jiffies
);
438 static inline void throtl_set_slice_end(struct throtl_data
*td
,
439 struct throtl_grp
*tg
, bool rw
, unsigned long jiffy_end
)
441 tg
->slice_end
[rw
] = roundup(jiffy_end
, throtl_slice
);
444 static inline void throtl_extend_slice(struct throtl_data
*td
,
445 struct throtl_grp
*tg
, bool rw
, unsigned long jiffy_end
)
447 tg
->slice_end
[rw
] = roundup(jiffy_end
, throtl_slice
);
448 throtl_log_tg(td
, tg
, "[%c] extend slice start=%lu end=%lu jiffies=%lu",
449 rw
== READ
? 'R' : 'W', tg
->slice_start
[rw
],
450 tg
->slice_end
[rw
], jiffies
);
453 /* Determine if previously allocated or extended slice is complete or not */
455 throtl_slice_used(struct throtl_data
*td
, struct throtl_grp
*tg
, bool rw
)
457 if (time_in_range(jiffies
, tg
->slice_start
[rw
], tg
->slice_end
[rw
]))
463 /* Trim the used slices and adjust slice start accordingly */
465 throtl_trim_slice(struct throtl_data
*td
, struct throtl_grp
*tg
, bool rw
)
467 unsigned long nr_slices
, time_elapsed
, io_trim
;
470 BUG_ON(time_before(tg
->slice_end
[rw
], tg
->slice_start
[rw
]));
473 * If bps are unlimited (-1), then time slice don't get
474 * renewed. Don't try to trim the slice if slice is used. A new
475 * slice will start when appropriate.
477 if (throtl_slice_used(td
, tg
, rw
))
481 * A bio has been dispatched. Also adjust slice_end. It might happen
482 * that initially cgroup limit was very low resulting in high
483 * slice_end, but later limit was bumped up and bio was dispached
484 * sooner, then we need to reduce slice_end. A high bogus slice_end
485 * is bad because it does not allow new slice to start.
488 throtl_set_slice_end(td
, tg
, rw
, jiffies
+ throtl_slice
);
490 time_elapsed
= jiffies
- tg
->slice_start
[rw
];
492 nr_slices
= time_elapsed
/ throtl_slice
;
496 tmp
= tg
->bps
[rw
] * throtl_slice
* nr_slices
;
500 io_trim
= (tg
->iops
[rw
] * throtl_slice
* nr_slices
)/HZ
;
502 if (!bytes_trim
&& !io_trim
)
505 if (tg
->bytes_disp
[rw
] >= bytes_trim
)
506 tg
->bytes_disp
[rw
] -= bytes_trim
;
508 tg
->bytes_disp
[rw
] = 0;
510 if (tg
->io_disp
[rw
] >= io_trim
)
511 tg
->io_disp
[rw
] -= io_trim
;
515 tg
->slice_start
[rw
] += nr_slices
* throtl_slice
;
517 throtl_log_tg(td
, tg
, "[%c] trim slice nr=%lu bytes=%llu io=%lu"
518 " start=%lu end=%lu jiffies=%lu",
519 rw
== READ
? 'R' : 'W', nr_slices
, bytes_trim
, io_trim
,
520 tg
->slice_start
[rw
], tg
->slice_end
[rw
], jiffies
);
523 static bool tg_with_in_iops_limit(struct throtl_data
*td
, struct throtl_grp
*tg
,
524 struct bio
*bio
, unsigned long *wait
)
526 bool rw
= bio_data_dir(bio
);
527 unsigned int io_allowed
;
528 unsigned long jiffy_elapsed
, jiffy_wait
, jiffy_elapsed_rnd
;
531 jiffy_elapsed
= jiffy_elapsed_rnd
= jiffies
- tg
->slice_start
[rw
];
533 /* Slice has just started. Consider one slice interval */
535 jiffy_elapsed_rnd
= throtl_slice
;
537 jiffy_elapsed_rnd
= roundup(jiffy_elapsed_rnd
, throtl_slice
);
540 * jiffy_elapsed_rnd should not be a big value as minimum iops can be
541 * 1 then at max jiffy elapsed should be equivalent of 1 second as we
542 * will allow dispatch after 1 second and after that slice should
546 tmp
= (u64
)tg
->iops
[rw
] * jiffy_elapsed_rnd
;
550 io_allowed
= UINT_MAX
;
554 if (tg
->io_disp
[rw
] + 1 <= io_allowed
) {
560 /* Calc approx time to dispatch */
561 jiffy_wait
= ((tg
->io_disp
[rw
] + 1) * HZ
)/tg
->iops
[rw
] + 1;
563 if (jiffy_wait
> jiffy_elapsed
)
564 jiffy_wait
= jiffy_wait
- jiffy_elapsed
;
573 static bool tg_with_in_bps_limit(struct throtl_data
*td
, struct throtl_grp
*tg
,
574 struct bio
*bio
, unsigned long *wait
)
576 bool rw
= bio_data_dir(bio
);
577 u64 bytes_allowed
, extra_bytes
, tmp
;
578 unsigned long jiffy_elapsed
, jiffy_wait
, jiffy_elapsed_rnd
;
580 jiffy_elapsed
= jiffy_elapsed_rnd
= jiffies
- tg
->slice_start
[rw
];
582 /* Slice has just started. Consider one slice interval */
584 jiffy_elapsed_rnd
= throtl_slice
;
586 jiffy_elapsed_rnd
= roundup(jiffy_elapsed_rnd
, throtl_slice
);
588 tmp
= tg
->bps
[rw
] * jiffy_elapsed_rnd
;
592 if (tg
->bytes_disp
[rw
] + bio
->bi_size
<= bytes_allowed
) {
598 /* Calc approx time to dispatch */
599 extra_bytes
= tg
->bytes_disp
[rw
] + bio
->bi_size
- bytes_allowed
;
600 jiffy_wait
= div64_u64(extra_bytes
* HZ
, tg
->bps
[rw
]);
606 * This wait time is without taking into consideration the rounding
607 * up we did. Add that time also.
609 jiffy_wait
= jiffy_wait
+ (jiffy_elapsed_rnd
- jiffy_elapsed
);
615 static bool tg_no_rule_group(struct throtl_grp
*tg
, bool rw
) {
616 if (tg
->bps
[rw
] == -1 && tg
->iops
[rw
] == -1)
622 * Returns whether one can dispatch a bio or not. Also returns approx number
623 * of jiffies to wait before this bio is with-in IO rate and can be dispatched
625 static bool tg_may_dispatch(struct throtl_data
*td
, struct throtl_grp
*tg
,
626 struct bio
*bio
, unsigned long *wait
)
628 bool rw
= bio_data_dir(bio
);
629 unsigned long bps_wait
= 0, iops_wait
= 0, max_wait
= 0;
632 * Currently whole state machine of group depends on first bio
633 * queued in the group bio list. So one should not be calling
634 * this function with a different bio if there are other bios
637 BUG_ON(tg
->nr_queued
[rw
] && bio
!= bio_list_peek(&tg
->bio_lists
[rw
]));
639 /* If tg->bps = -1, then BW is unlimited */
640 if (tg
->bps
[rw
] == -1 && tg
->iops
[rw
] == -1) {
647 * If previous slice expired, start a new one otherwise renew/extend
648 * existing slice to make sure it is at least throtl_slice interval
651 if (throtl_slice_used(td
, tg
, rw
))
652 throtl_start_new_slice(td
, tg
, rw
);
654 if (time_before(tg
->slice_end
[rw
], jiffies
+ throtl_slice
))
655 throtl_extend_slice(td
, tg
, rw
, jiffies
+ throtl_slice
);
658 if (tg_with_in_bps_limit(td
, tg
, bio
, &bps_wait
)
659 && tg_with_in_iops_limit(td
, tg
, bio
, &iops_wait
)) {
665 max_wait
= max(bps_wait
, iops_wait
);
670 if (time_before(tg
->slice_end
[rw
], jiffies
+ max_wait
))
671 throtl_extend_slice(td
, tg
, rw
, jiffies
+ max_wait
);
676 static void throtl_update_dispatch_stats(struct blkcg_gq
*blkg
, u64 bytes
,
679 struct throtl_grp
*tg
= blkg_to_tg(blkg
);
680 struct tg_stats_cpu
*stats_cpu
;
683 /* If per cpu stats are not allocated yet, don't do any accounting. */
684 if (tg
->stats_cpu
== NULL
)
688 * Disabling interrupts to provide mutual exclusion between two
689 * writes on same cpu. It probably is not needed for 64bit. Not
690 * optimizing that case yet.
692 local_irq_save(flags
);
694 stats_cpu
= this_cpu_ptr(tg
->stats_cpu
);
696 blkg_rwstat_add(&stats_cpu
->serviced
, rw
, 1);
697 blkg_rwstat_add(&stats_cpu
->service_bytes
, rw
, bytes
);
699 local_irq_restore(flags
);
702 static void throtl_charge_bio(struct throtl_grp
*tg
, struct bio
*bio
)
704 bool rw
= bio_data_dir(bio
);
706 /* Charge the bio to the group */
707 tg
->bytes_disp
[rw
] += bio
->bi_size
;
710 throtl_update_dispatch_stats(tg_to_blkg(tg
), bio
->bi_size
, bio
->bi_rw
);
713 static void throtl_add_bio_tg(struct throtl_data
*td
, struct throtl_grp
*tg
,
716 bool rw
= bio_data_dir(bio
);
718 bio_list_add(&tg
->bio_lists
[rw
], bio
);
719 /* Take a bio reference on tg */
720 blkg_get(tg_to_blkg(tg
));
723 throtl_enqueue_tg(td
, tg
);
726 static void tg_update_disptime(struct throtl_data
*td
, struct throtl_grp
*tg
)
728 unsigned long read_wait
= -1, write_wait
= -1, min_wait
= -1, disptime
;
731 if ((bio
= bio_list_peek(&tg
->bio_lists
[READ
])))
732 tg_may_dispatch(td
, tg
, bio
, &read_wait
);
734 if ((bio
= bio_list_peek(&tg
->bio_lists
[WRITE
])))
735 tg_may_dispatch(td
, tg
, bio
, &write_wait
);
737 min_wait
= min(read_wait
, write_wait
);
738 disptime
= jiffies
+ min_wait
;
740 /* Update dispatch time */
741 throtl_dequeue_tg(td
, tg
);
742 tg
->disptime
= disptime
;
743 throtl_enqueue_tg(td
, tg
);
746 static void tg_dispatch_one_bio(struct throtl_data
*td
, struct throtl_grp
*tg
,
747 bool rw
, struct bio_list
*bl
)
751 bio
= bio_list_pop(&tg
->bio_lists
[rw
]);
753 /* Drop bio reference on blkg */
754 blkg_put(tg_to_blkg(tg
));
756 BUG_ON(td
->nr_queued
[rw
] <= 0);
759 throtl_charge_bio(tg
, bio
);
760 bio_list_add(bl
, bio
);
761 bio
->bi_rw
|= REQ_THROTTLED
;
763 throtl_trim_slice(td
, tg
, rw
);
766 static int throtl_dispatch_tg(struct throtl_data
*td
, struct throtl_grp
*tg
,
769 unsigned int nr_reads
= 0, nr_writes
= 0;
770 unsigned int max_nr_reads
= throtl_grp_quantum
*3/4;
771 unsigned int max_nr_writes
= throtl_grp_quantum
- max_nr_reads
;
774 /* Try to dispatch 75% READS and 25% WRITES */
776 while ((bio
= bio_list_peek(&tg
->bio_lists
[READ
]))
777 && tg_may_dispatch(td
, tg
, bio
, NULL
)) {
779 tg_dispatch_one_bio(td
, tg
, bio_data_dir(bio
), bl
);
782 if (nr_reads
>= max_nr_reads
)
786 while ((bio
= bio_list_peek(&tg
->bio_lists
[WRITE
]))
787 && tg_may_dispatch(td
, tg
, bio
, NULL
)) {
789 tg_dispatch_one_bio(td
, tg
, bio_data_dir(bio
), bl
);
792 if (nr_writes
>= max_nr_writes
)
796 return nr_reads
+ nr_writes
;
799 static int throtl_select_dispatch(struct throtl_data
*td
, struct bio_list
*bl
)
801 unsigned int nr_disp
= 0;
802 struct throtl_grp
*tg
;
803 struct throtl_rb_root
*st
= &td
->tg_service_tree
;
806 tg
= throtl_rb_first(st
);
811 if (time_before(jiffies
, tg
->disptime
))
814 throtl_dequeue_tg(td
, tg
);
816 nr_disp
+= throtl_dispatch_tg(td
, tg
, bl
);
818 if (tg
->nr_queued
[0] || tg
->nr_queued
[1]) {
819 tg_update_disptime(td
, tg
);
820 throtl_enqueue_tg(td
, tg
);
823 if (nr_disp
>= throtl_quantum
)
830 static void throtl_process_limit_change(struct throtl_data
*td
)
832 struct request_queue
*q
= td
->queue
;
833 struct blkcg_gq
*blkg
, *n
;
835 if (!td
->limits_changed
)
838 xchg(&td
->limits_changed
, false);
840 throtl_log(td
, "limits changed");
842 list_for_each_entry_safe(blkg
, n
, &q
->blkg_list
, q_node
) {
843 struct throtl_grp
*tg
= blkg_to_tg(blkg
);
845 if (!tg
->limits_changed
)
848 if (!xchg(&tg
->limits_changed
, false))
851 throtl_log_tg(td
, tg
, "limit change rbps=%llu wbps=%llu"
852 " riops=%u wiops=%u", tg
->bps
[READ
], tg
->bps
[WRITE
],
853 tg
->iops
[READ
], tg
->iops
[WRITE
]);
856 * Restart the slices for both READ and WRITES. It
857 * might happen that a group's limit are dropped
858 * suddenly and we don't want to account recently
859 * dispatched IO with new low rate
861 throtl_start_new_slice(td
, tg
, 0);
862 throtl_start_new_slice(td
, tg
, 1);
864 if (throtl_tg_on_rr(tg
))
865 tg_update_disptime(td
, tg
);
869 /* Dispatch throttled bios. Should be called without queue lock held. */
870 static int throtl_dispatch(struct request_queue
*q
)
872 struct throtl_data
*td
= q
->td
;
873 unsigned int nr_disp
= 0;
874 struct bio_list bio_list_on_stack
;
876 struct blk_plug plug
;
878 spin_lock_irq(q
->queue_lock
);
880 throtl_process_limit_change(td
);
882 if (!total_nr_queued(td
))
885 bio_list_init(&bio_list_on_stack
);
887 throtl_log(td
, "dispatch nr_queued=%u read=%u write=%u",
888 total_nr_queued(td
), td
->nr_queued
[READ
],
889 td
->nr_queued
[WRITE
]);
891 nr_disp
= throtl_select_dispatch(td
, &bio_list_on_stack
);
894 throtl_log(td
, "bios disp=%u", nr_disp
);
896 throtl_schedule_next_dispatch(td
);
898 spin_unlock_irq(q
->queue_lock
);
901 * If we dispatched some requests, unplug the queue to make sure
905 blk_start_plug(&plug
);
906 while((bio
= bio_list_pop(&bio_list_on_stack
)))
907 generic_make_request(bio
);
908 blk_finish_plug(&plug
);
913 void blk_throtl_work(struct work_struct
*work
)
915 struct throtl_data
*td
= container_of(work
, struct throtl_data
,
917 struct request_queue
*q
= td
->queue
;
922 /* Call with queue lock held */
924 throtl_schedule_delayed_work(struct throtl_data
*td
, unsigned long delay
)
927 struct delayed_work
*dwork
= &td
->throtl_work
;
929 /* schedule work if limits changed even if no bio is queued */
930 if (total_nr_queued(td
) || td
->limits_changed
) {
932 * We might have a work scheduled to be executed in future.
933 * Cancel that and schedule a new one.
935 __cancel_delayed_work(dwork
);
936 queue_delayed_work(kthrotld_workqueue
, dwork
, delay
);
937 throtl_log(td
, "schedule work. delay=%lu jiffies=%lu",
942 static u64
tg_prfill_cpu_rwstat(struct seq_file
*sf
,
943 struct blkg_policy_data
*pd
, int off
)
945 struct throtl_grp
*tg
= pd_to_tg(pd
);
946 struct blkg_rwstat rwstat
= { }, tmp
;
949 for_each_possible_cpu(cpu
) {
950 struct tg_stats_cpu
*sc
= per_cpu_ptr(tg
->stats_cpu
, cpu
);
952 tmp
= blkg_rwstat_read((void *)sc
+ off
);
953 for (i
= 0; i
< BLKG_RWSTAT_NR
; i
++)
954 rwstat
.cnt
[i
] += tmp
.cnt
[i
];
957 return __blkg_prfill_rwstat(sf
, pd
, &rwstat
);
960 static int tg_print_cpu_rwstat(struct cgroup
*cgrp
, struct cftype
*cft
,
963 struct blkcg
*blkcg
= cgroup_to_blkcg(cgrp
);
965 blkcg_print_blkgs(sf
, blkcg
, tg_prfill_cpu_rwstat
, &blkcg_policy_throtl
,
970 static u64
tg_prfill_conf_u64(struct seq_file
*sf
, struct blkg_policy_data
*pd
,
973 struct throtl_grp
*tg
= pd_to_tg(pd
);
974 u64 v
= *(u64
*)((void *)tg
+ off
);
978 return __blkg_prfill_u64(sf
, pd
, v
);
981 static u64
tg_prfill_conf_uint(struct seq_file
*sf
, struct blkg_policy_data
*pd
,
984 struct throtl_grp
*tg
= pd_to_tg(pd
);
985 unsigned int v
= *(unsigned int *)((void *)tg
+ off
);
989 return __blkg_prfill_u64(sf
, pd
, v
);
992 static int tg_print_conf_u64(struct cgroup
*cgrp
, struct cftype
*cft
,
995 blkcg_print_blkgs(sf
, cgroup_to_blkcg(cgrp
), tg_prfill_conf_u64
,
996 &blkcg_policy_throtl
, cft
->private, false);
1000 static int tg_print_conf_uint(struct cgroup
*cgrp
, struct cftype
*cft
,
1001 struct seq_file
*sf
)
1003 blkcg_print_blkgs(sf
, cgroup_to_blkcg(cgrp
), tg_prfill_conf_uint
,
1004 &blkcg_policy_throtl
, cft
->private, false);
1008 static int tg_set_conf(struct cgroup
*cgrp
, struct cftype
*cft
, const char *buf
,
1011 struct blkcg
*blkcg
= cgroup_to_blkcg(cgrp
);
1012 struct blkg_conf_ctx ctx
;
1013 struct throtl_grp
*tg
;
1014 struct throtl_data
*td
;
1017 ret
= blkg_conf_prep(blkcg
, &blkcg_policy_throtl
, buf
, &ctx
);
1021 tg
= blkg_to_tg(ctx
.blkg
);
1022 td
= ctx
.blkg
->q
->td
;
1028 *(u64
*)((void *)tg
+ cft
->private) = ctx
.v
;
1030 *(unsigned int *)((void *)tg
+ cft
->private) = ctx
.v
;
1032 /* XXX: we don't need the following deferred processing */
1033 xchg(&tg
->limits_changed
, true);
1034 xchg(&td
->limits_changed
, true);
1035 throtl_schedule_delayed_work(td
, 0);
1037 blkg_conf_finish(&ctx
);
1041 static int tg_set_conf_u64(struct cgroup
*cgrp
, struct cftype
*cft
,
1044 return tg_set_conf(cgrp
, cft
, buf
, true);
1047 static int tg_set_conf_uint(struct cgroup
*cgrp
, struct cftype
*cft
,
1050 return tg_set_conf(cgrp
, cft
, buf
, false);
1053 static struct cftype throtl_files
[] = {
1055 .name
= "throttle.read_bps_device",
1056 .private = offsetof(struct throtl_grp
, bps
[READ
]),
1057 .read_seq_string
= tg_print_conf_u64
,
1058 .write_string
= tg_set_conf_u64
,
1059 .max_write_len
= 256,
1062 .name
= "throttle.write_bps_device",
1063 .private = offsetof(struct throtl_grp
, bps
[WRITE
]),
1064 .read_seq_string
= tg_print_conf_u64
,
1065 .write_string
= tg_set_conf_u64
,
1066 .max_write_len
= 256,
1069 .name
= "throttle.read_iops_device",
1070 .private = offsetof(struct throtl_grp
, iops
[READ
]),
1071 .read_seq_string
= tg_print_conf_uint
,
1072 .write_string
= tg_set_conf_uint
,
1073 .max_write_len
= 256,
1076 .name
= "throttle.write_iops_device",
1077 .private = offsetof(struct throtl_grp
, iops
[WRITE
]),
1078 .read_seq_string
= tg_print_conf_uint
,
1079 .write_string
= tg_set_conf_uint
,
1080 .max_write_len
= 256,
1083 .name
= "throttle.io_service_bytes",
1084 .private = offsetof(struct tg_stats_cpu
, service_bytes
),
1085 .read_seq_string
= tg_print_cpu_rwstat
,
1088 .name
= "throttle.io_serviced",
1089 .private = offsetof(struct tg_stats_cpu
, serviced
),
1090 .read_seq_string
= tg_print_cpu_rwstat
,
1095 static void throtl_shutdown_wq(struct request_queue
*q
)
1097 struct throtl_data
*td
= q
->td
;
1099 cancel_delayed_work_sync(&td
->throtl_work
);
1102 static struct blkcg_policy blkcg_policy_throtl
= {
1104 .pd_init_fn
= throtl_pd_init
,
1105 .pd_exit_fn
= throtl_pd_exit
,
1106 .pd_reset_stats_fn
= throtl_pd_reset_stats
,
1108 .pd_size
= sizeof(struct throtl_grp
),
1109 .cftypes
= throtl_files
,
1112 bool blk_throtl_bio(struct request_queue
*q
, struct bio
*bio
)
1114 struct throtl_data
*td
= q
->td
;
1115 struct throtl_grp
*tg
;
1116 bool rw
= bio_data_dir(bio
), update_disptime
= true;
1117 struct blkcg
*blkcg
;
1118 bool throttled
= false;
1120 if (bio
->bi_rw
& REQ_THROTTLED
) {
1121 bio
->bi_rw
&= ~REQ_THROTTLED
;
1125 /* bio_associate_current() needs ioc, try creating */
1126 create_io_context(GFP_ATOMIC
, q
->node
);
1129 * A throtl_grp pointer retrieved under rcu can be used to access
1130 * basic fields like stats and io rates. If a group has no rules,
1131 * just update the dispatch stats in lockless manner and return.
1134 blkcg
= bio_blkcg(bio
);
1135 tg
= throtl_lookup_tg(td
, blkcg
);
1137 if (tg_no_rule_group(tg
, rw
)) {
1138 throtl_update_dispatch_stats(tg_to_blkg(tg
),
1139 bio
->bi_size
, bio
->bi_rw
);
1140 goto out_unlock_rcu
;
1145 * Either group has not been allocated yet or it is not an unlimited
1148 spin_lock_irq(q
->queue_lock
);
1149 tg
= throtl_lookup_create_tg(td
, blkcg
);
1153 if (tg
->nr_queued
[rw
]) {
1155 * There is already another bio queued in same dir. No
1156 * need to update dispatch time.
1158 update_disptime
= false;
1163 /* Bio is with-in rate limit of group */
1164 if (tg_may_dispatch(td
, tg
, bio
, NULL
)) {
1165 throtl_charge_bio(tg
, bio
);
1168 * We need to trim slice even when bios are not being queued
1169 * otherwise it might happen that a bio is not queued for
1170 * a long time and slice keeps on extending and trim is not
1171 * called for a long time. Now if limits are reduced suddenly
1172 * we take into account all the IO dispatched so far at new
1173 * low rate and * newly queued IO gets a really long dispatch
1176 * So keep on trimming slice even if bio is not queued.
1178 throtl_trim_slice(td
, tg
, rw
);
1183 throtl_log_tg(td
, tg
, "[%c] bio. bdisp=%llu sz=%u bps=%llu"
1184 " iodisp=%u iops=%u queued=%d/%d",
1185 rw
== READ
? 'R' : 'W',
1186 tg
->bytes_disp
[rw
], bio
->bi_size
, tg
->bps
[rw
],
1187 tg
->io_disp
[rw
], tg
->iops
[rw
],
1188 tg
->nr_queued
[READ
], tg
->nr_queued
[WRITE
]);
1190 bio_associate_current(bio
);
1191 throtl_add_bio_tg(q
->td
, tg
, bio
);
1194 if (update_disptime
) {
1195 tg_update_disptime(td
, tg
);
1196 throtl_schedule_next_dispatch(td
);
1200 spin_unlock_irq(q
->queue_lock
);
1208 * blk_throtl_drain - drain throttled bios
1209 * @q: request_queue to drain throttled bios for
1211 * Dispatch all currently throttled bios on @q through ->make_request_fn().
1213 void blk_throtl_drain(struct request_queue
*q
)
1214 __releases(q
->queue_lock
) __acquires(q
->queue_lock
)
1216 struct throtl_data
*td
= q
->td
;
1217 struct throtl_rb_root
*st
= &td
->tg_service_tree
;
1218 struct throtl_grp
*tg
;
1222 WARN_ON_ONCE(!queue_is_locked(q
));
1226 while ((tg
= throtl_rb_first(st
))) {
1227 throtl_dequeue_tg(td
, tg
);
1229 while ((bio
= bio_list_peek(&tg
->bio_lists
[READ
])))
1230 tg_dispatch_one_bio(td
, tg
, bio_data_dir(bio
), &bl
);
1231 while ((bio
= bio_list_peek(&tg
->bio_lists
[WRITE
])))
1232 tg_dispatch_one_bio(td
, tg
, bio_data_dir(bio
), &bl
);
1234 spin_unlock_irq(q
->queue_lock
);
1236 while ((bio
= bio_list_pop(&bl
)))
1237 generic_make_request(bio
);
1239 spin_lock_irq(q
->queue_lock
);
1242 int blk_throtl_init(struct request_queue
*q
)
1244 struct throtl_data
*td
;
1247 td
= kzalloc_node(sizeof(*td
), GFP_KERNEL
, q
->node
);
1251 td
->tg_service_tree
= THROTL_RB_ROOT
;
1252 td
->limits_changed
= false;
1253 INIT_DELAYED_WORK(&td
->throtl_work
, blk_throtl_work
);
1258 /* activate policy */
1259 ret
= blkcg_activate_policy(q
, &blkcg_policy_throtl
);
1265 void blk_throtl_exit(struct request_queue
*q
)
1268 throtl_shutdown_wq(q
);
1269 blkcg_deactivate_policy(q
, &blkcg_policy_throtl
);
1273 static int __init
throtl_init(void)
1275 kthrotld_workqueue
= alloc_workqueue("kthrotld", WQ_MEM_RECLAIM
, 0);
1276 if (!kthrotld_workqueue
)
1277 panic("Failed to create kthrotld\n");
1279 return blkcg_policy_register(&blkcg_policy_throtl
);
1282 module_init(throtl_init
);