1 // SPDX-License-Identifier: GPL-2.0
3 * Interface for controlling IO bandwidth on a request queue
5 * Copyright (C) 2010 Vivek Goyal <vgoyal@redhat.com>
8 #include <linux/module.h>
9 #include <linux/slab.h>
10 #include <linux/blkdev.h>
11 #include <linux/bio.h>
12 #include <linux/blktrace_api.h>
13 #include <linux/blk-cgroup.h>
15 #include "blk-cgroup-rwstat.h"
17 /* Max dispatch from a group in 1 round */
18 #define THROTL_GRP_QUANTUM 8
20 /* Total max dispatch from all groups in one round */
21 #define THROTL_QUANTUM 32
23 /* Throttling is performed over a slice and after that slice is renewed */
24 #define DFL_THROTL_SLICE_HD (HZ / 10)
25 #define DFL_THROTL_SLICE_SSD (HZ / 50)
26 #define MAX_THROTL_SLICE (HZ)
27 #define MAX_IDLE_TIME (5L * 1000 * 1000) /* 5 s */
28 #define MIN_THROTL_BPS (320 * 1024)
29 #define MIN_THROTL_IOPS (10)
30 #define DFL_LATENCY_TARGET (-1L)
31 #define DFL_IDLE_THRESHOLD (0)
32 #define DFL_HD_BASELINE_LATENCY (4000L) /* 4ms */
33 #define LATENCY_FILTERED_SSD (0)
35 * For HD, very small latency comes from sequential IO. Such IO is helpless to
36 * help determine if its IO is impacted by others, hence we ignore the IO
38 #define LATENCY_FILTERED_HD (1000L) /* 1ms */
40 static struct blkcg_policy blkcg_policy_throtl
;
42 /* A workqueue to queue throttle related work */
43 static struct workqueue_struct
*kthrotld_workqueue
;
46 * To implement hierarchical throttling, throtl_grps form a tree and bios
47 * are dispatched upwards level by level until they reach the top and get
48 * issued. When dispatching bios from the children and local group at each
49 * level, if the bios are dispatched into a single bio_list, there's a risk
50 * of a local or child group which can queue many bios at once filling up
51 * the list starving others.
53 * To avoid such starvation, dispatched bios are queued separately
54 * according to where they came from. When they are again dispatched to
55 * the parent, they're popped in round-robin order so that no single source
56 * hogs the dispatch window.
58 * throtl_qnode is used to keep the queued bios separated by their sources.
59 * Bios are queued to throtl_qnode which in turn is queued to
60 * throtl_service_queue and then dispatched in round-robin order.
62 * It's also used to track the reference counts on blkg's. A qnode always
63 * belongs to a throtl_grp and gets queued on itself or the parent, so
64 * incrementing the reference of the associated throtl_grp when a qnode is
65 * queued and decrementing when dequeued is enough to keep the whole blkg
66 * tree pinned while bios are in flight.
69 struct list_head node
; /* service_queue->queued[] */
70 struct bio_list bios
; /* queued bios */
71 struct throtl_grp
*tg
; /* tg this qnode belongs to */
74 struct throtl_service_queue
{
75 struct throtl_service_queue
*parent_sq
; /* the parent service_queue */
78 * Bios queued directly to this service_queue or dispatched from
79 * children throtl_grp's.
81 struct list_head queued
[2]; /* throtl_qnode [READ/WRITE] */
82 unsigned int nr_queued
[2]; /* number of queued bios */
85 * RB tree of active children throtl_grp's, which are sorted by
88 struct rb_root_cached pending_tree
; /* RB tree of active tgs */
89 unsigned int nr_pending
; /* # queued in the tree */
90 unsigned long first_pending_disptime
; /* disptime of the first tg */
91 struct timer_list pending_timer
; /* fires on first_pending_disptime */
95 THROTL_TG_PENDING
= 1 << 0, /* on parent's pending tree */
96 THROTL_TG_WAS_EMPTY
= 1 << 1, /* bio_lists[] became non-empty */
99 #define rb_entry_tg(node) rb_entry((node), struct throtl_grp, rb_node)
108 /* must be the first member */
109 struct blkg_policy_data pd
;
111 /* active throtl group service_queue member */
112 struct rb_node rb_node
;
114 /* throtl_data this group belongs to */
115 struct throtl_data
*td
;
117 /* this group's service queue */
118 struct throtl_service_queue service_queue
;
121 * qnode_on_self is used when bios are directly queued to this
122 * throtl_grp so that local bios compete fairly with bios
123 * dispatched from children. qnode_on_parent is used when bios are
124 * dispatched from this throtl_grp into its parent and will compete
125 * with the sibling qnode_on_parents and the parent's
128 struct throtl_qnode qnode_on_self
[2];
129 struct throtl_qnode qnode_on_parent
[2];
132 * Dispatch time in jiffies. This is the estimated time when group
133 * will unthrottle and is ready to dispatch more bio. It is used as
134 * key to sort active groups in service tree.
136 unsigned long disptime
;
140 /* are there any throtl rules between this group and td? */
143 /* internally used bytes per second rate limits */
144 uint64_t bps
[2][LIMIT_CNT
];
145 /* user configured bps limits */
146 uint64_t bps_conf
[2][LIMIT_CNT
];
148 /* internally used IOPS limits */
149 unsigned int iops
[2][LIMIT_CNT
];
150 /* user configured IOPS limits */
151 unsigned int iops_conf
[2][LIMIT_CNT
];
153 /* Number of bytes dispatched in current slice */
154 uint64_t bytes_disp
[2];
155 /* Number of bio's dispatched in current slice */
156 unsigned int io_disp
[2];
158 unsigned long last_low_overflow_time
[2];
160 uint64_t last_bytes_disp
[2];
161 unsigned int last_io_disp
[2];
163 unsigned long last_check_time
;
165 unsigned long latency_target
; /* us */
166 unsigned long latency_target_conf
; /* us */
167 /* When did we start a new slice */
168 unsigned long slice_start
[2];
169 unsigned long slice_end
[2];
171 unsigned long last_finish_time
; /* ns / 1024 */
172 unsigned long checked_last_finish_time
; /* ns / 1024 */
173 unsigned long avg_idletime
; /* ns / 1024 */
174 unsigned long idletime_threshold
; /* us */
175 unsigned long idletime_threshold_conf
; /* us */
177 unsigned int bio_cnt
; /* total bios */
178 unsigned int bad_bio_cnt
; /* bios exceeding latency threshold */
179 unsigned long bio_cnt_reset_time
;
181 struct blkg_rwstat stat_bytes
;
182 struct blkg_rwstat stat_ios
;
185 /* We measure latency for request size from <= 4k to >= 1M */
186 #define LATENCY_BUCKET_SIZE 9
188 struct latency_bucket
{
189 unsigned long total_latency
; /* ns / 1024 */
193 struct avg_latency_bucket
{
194 unsigned long latency
; /* ns / 1024 */
200 /* service tree for active throtl groups */
201 struct throtl_service_queue service_queue
;
203 struct request_queue
*queue
;
205 /* Total Number of queued bios on READ and WRITE lists */
206 unsigned int nr_queued
[2];
208 unsigned int throtl_slice
;
210 /* Work for dispatching throttled bios */
211 struct work_struct dispatch_work
;
212 unsigned int limit_index
;
213 bool limit_valid
[LIMIT_CNT
];
215 unsigned long low_upgrade_time
;
216 unsigned long low_downgrade_time
;
220 struct latency_bucket tmp_buckets
[2][LATENCY_BUCKET_SIZE
];
221 struct avg_latency_bucket avg_buckets
[2][LATENCY_BUCKET_SIZE
];
222 struct latency_bucket __percpu
*latency_buckets
[2];
223 unsigned long last_calculate_time
;
224 unsigned long filtered_latency
;
226 bool track_bio_latency
;
229 static void throtl_pending_timer_fn(struct timer_list
*t
);
231 static inline struct throtl_grp
*pd_to_tg(struct blkg_policy_data
*pd
)
233 return pd
? container_of(pd
, struct throtl_grp
, pd
) : NULL
;
236 static inline struct throtl_grp
*blkg_to_tg(struct blkcg_gq
*blkg
)
238 return pd_to_tg(blkg_to_pd(blkg
, &blkcg_policy_throtl
));
241 static inline struct blkcg_gq
*tg_to_blkg(struct throtl_grp
*tg
)
243 return pd_to_blkg(&tg
->pd
);
247 * sq_to_tg - return the throl_grp the specified service queue belongs to
248 * @sq: the throtl_service_queue of interest
250 * Return the throtl_grp @sq belongs to. If @sq is the top-level one
251 * embedded in throtl_data, %NULL is returned.
253 static struct throtl_grp
*sq_to_tg(struct throtl_service_queue
*sq
)
255 if (sq
&& sq
->parent_sq
)
256 return container_of(sq
, struct throtl_grp
, service_queue
);
262 * sq_to_td - return throtl_data the specified service queue belongs to
263 * @sq: the throtl_service_queue of interest
265 * A service_queue can be embedded in either a throtl_grp or throtl_data.
266 * Determine the associated throtl_data accordingly and return it.
268 static struct throtl_data
*sq_to_td(struct throtl_service_queue
*sq
)
270 struct throtl_grp
*tg
= sq_to_tg(sq
);
275 return container_of(sq
, struct throtl_data
, service_queue
);
279 * cgroup's limit in LIMIT_MAX is scaled if low limit is set. This scale is to
280 * make the IO dispatch more smooth.
281 * Scale up: linearly scale up according to lapsed time since upgrade. For
282 * every throtl_slice, the limit scales up 1/2 .low limit till the
283 * limit hits .max limit
284 * Scale down: exponentially scale down if a cgroup doesn't hit its .low limit
286 static uint64_t throtl_adjusted_limit(uint64_t low
, struct throtl_data
*td
)
288 /* arbitrary value to avoid too big scale */
289 if (td
->scale
< 4096 && time_after_eq(jiffies
,
290 td
->low_upgrade_time
+ td
->scale
* td
->throtl_slice
))
291 td
->scale
= (jiffies
- td
->low_upgrade_time
) / td
->throtl_slice
;
293 return low
+ (low
>> 1) * td
->scale
;
296 static uint64_t tg_bps_limit(struct throtl_grp
*tg
, int rw
)
298 struct blkcg_gq
*blkg
= tg_to_blkg(tg
);
299 struct throtl_data
*td
;
302 if (cgroup_subsys_on_dfl(io_cgrp_subsys
) && !blkg
->parent
)
306 ret
= tg
->bps
[rw
][td
->limit_index
];
307 if (ret
== 0 && td
->limit_index
== LIMIT_LOW
) {
308 /* intermediate node or iops isn't 0 */
309 if (!list_empty(&blkg
->blkcg
->css
.children
) ||
310 tg
->iops
[rw
][td
->limit_index
])
313 return MIN_THROTL_BPS
;
316 if (td
->limit_index
== LIMIT_MAX
&& tg
->bps
[rw
][LIMIT_LOW
] &&
317 tg
->bps
[rw
][LIMIT_LOW
] != tg
->bps
[rw
][LIMIT_MAX
]) {
320 adjusted
= throtl_adjusted_limit(tg
->bps
[rw
][LIMIT_LOW
], td
);
321 ret
= min(tg
->bps
[rw
][LIMIT_MAX
], adjusted
);
326 static unsigned int tg_iops_limit(struct throtl_grp
*tg
, int rw
)
328 struct blkcg_gq
*blkg
= tg_to_blkg(tg
);
329 struct throtl_data
*td
;
332 if (cgroup_subsys_on_dfl(io_cgrp_subsys
) && !blkg
->parent
)
336 ret
= tg
->iops
[rw
][td
->limit_index
];
337 if (ret
== 0 && tg
->td
->limit_index
== LIMIT_LOW
) {
338 /* intermediate node or bps isn't 0 */
339 if (!list_empty(&blkg
->blkcg
->css
.children
) ||
340 tg
->bps
[rw
][td
->limit_index
])
343 return MIN_THROTL_IOPS
;
346 if (td
->limit_index
== LIMIT_MAX
&& tg
->iops
[rw
][LIMIT_LOW
] &&
347 tg
->iops
[rw
][LIMIT_LOW
] != tg
->iops
[rw
][LIMIT_MAX
]) {
350 adjusted
= throtl_adjusted_limit(tg
->iops
[rw
][LIMIT_LOW
], td
);
351 if (adjusted
> UINT_MAX
)
353 ret
= min_t(unsigned int, tg
->iops
[rw
][LIMIT_MAX
], adjusted
);
358 #define request_bucket_index(sectors) \
359 clamp_t(int, order_base_2(sectors) - 3, 0, LATENCY_BUCKET_SIZE - 1)
362 * throtl_log - log debug message via blktrace
363 * @sq: the service_queue being reported
364 * @fmt: printf format string
367 * The messages are prefixed with "throtl BLKG_NAME" if @sq belongs to a
368 * throtl_grp; otherwise, just "throtl".
370 #define throtl_log(sq, fmt, args...) do { \
371 struct throtl_grp *__tg = sq_to_tg((sq)); \
372 struct throtl_data *__td = sq_to_td((sq)); \
375 if (likely(!blk_trace_note_message_enabled(__td->queue))) \
378 blk_add_cgroup_trace_msg(__td->queue, \
379 tg_to_blkg(__tg)->blkcg, "throtl " fmt, ##args);\
381 blk_add_trace_msg(__td->queue, "throtl " fmt, ##args); \
385 static inline unsigned int throtl_bio_data_size(struct bio
*bio
)
387 /* assume it's one sector */
388 if (unlikely(bio_op(bio
) == REQ_OP_DISCARD
))
390 return bio
->bi_iter
.bi_size
;
393 static void throtl_qnode_init(struct throtl_qnode
*qn
, struct throtl_grp
*tg
)
395 INIT_LIST_HEAD(&qn
->node
);
396 bio_list_init(&qn
->bios
);
401 * throtl_qnode_add_bio - add a bio to a throtl_qnode and activate it
402 * @bio: bio being added
403 * @qn: qnode to add bio to
404 * @queued: the service_queue->queued[] list @qn belongs to
406 * Add @bio to @qn and put @qn on @queued if it's not already on.
407 * @qn->tg's reference count is bumped when @qn is activated. See the
408 * comment on top of throtl_qnode definition for details.
410 static void throtl_qnode_add_bio(struct bio
*bio
, struct throtl_qnode
*qn
,
411 struct list_head
*queued
)
413 bio_list_add(&qn
->bios
, bio
);
414 if (list_empty(&qn
->node
)) {
415 list_add_tail(&qn
->node
, queued
);
416 blkg_get(tg_to_blkg(qn
->tg
));
421 * throtl_peek_queued - peek the first bio on a qnode list
422 * @queued: the qnode list to peek
424 static struct bio
*throtl_peek_queued(struct list_head
*queued
)
426 struct throtl_qnode
*qn
;
429 if (list_empty(queued
))
432 qn
= list_first_entry(queued
, struct throtl_qnode
, node
);
433 bio
= bio_list_peek(&qn
->bios
);
439 * throtl_pop_queued - pop the first bio form a qnode list
440 * @queued: the qnode list to pop a bio from
441 * @tg_to_put: optional out argument for throtl_grp to put
443 * Pop the first bio from the qnode list @queued. After popping, the first
444 * qnode is removed from @queued if empty or moved to the end of @queued so
445 * that the popping order is round-robin.
447 * When the first qnode is removed, its associated throtl_grp should be put
448 * too. If @tg_to_put is NULL, this function automatically puts it;
449 * otherwise, *@tg_to_put is set to the throtl_grp to put and the caller is
450 * responsible for putting it.
452 static struct bio
*throtl_pop_queued(struct list_head
*queued
,
453 struct throtl_grp
**tg_to_put
)
455 struct throtl_qnode
*qn
;
458 if (list_empty(queued
))
461 qn
= list_first_entry(queued
, struct throtl_qnode
, node
);
462 bio
= bio_list_pop(&qn
->bios
);
465 if (bio_list_empty(&qn
->bios
)) {
466 list_del_init(&qn
->node
);
470 blkg_put(tg_to_blkg(qn
->tg
));
472 list_move_tail(&qn
->node
, queued
);
478 /* init a service_queue, assumes the caller zeroed it */
479 static void throtl_service_queue_init(struct throtl_service_queue
*sq
)
481 INIT_LIST_HEAD(&sq
->queued
[0]);
482 INIT_LIST_HEAD(&sq
->queued
[1]);
483 sq
->pending_tree
= RB_ROOT_CACHED
;
484 timer_setup(&sq
->pending_timer
, throtl_pending_timer_fn
, 0);
487 static struct blkg_policy_data
*throtl_pd_alloc(gfp_t gfp
,
488 struct request_queue
*q
,
491 struct throtl_grp
*tg
;
494 tg
= kzalloc_node(sizeof(*tg
), gfp
, q
->node
);
498 if (blkg_rwstat_init(&tg
->stat_bytes
, gfp
))
501 if (blkg_rwstat_init(&tg
->stat_ios
, gfp
))
502 goto err_exit_stat_bytes
;
504 throtl_service_queue_init(&tg
->service_queue
);
506 for (rw
= READ
; rw
<= WRITE
; rw
++) {
507 throtl_qnode_init(&tg
->qnode_on_self
[rw
], tg
);
508 throtl_qnode_init(&tg
->qnode_on_parent
[rw
], tg
);
511 RB_CLEAR_NODE(&tg
->rb_node
);
512 tg
->bps
[READ
][LIMIT_MAX
] = U64_MAX
;
513 tg
->bps
[WRITE
][LIMIT_MAX
] = U64_MAX
;
514 tg
->iops
[READ
][LIMIT_MAX
] = UINT_MAX
;
515 tg
->iops
[WRITE
][LIMIT_MAX
] = UINT_MAX
;
516 tg
->bps_conf
[READ
][LIMIT_MAX
] = U64_MAX
;
517 tg
->bps_conf
[WRITE
][LIMIT_MAX
] = U64_MAX
;
518 tg
->iops_conf
[READ
][LIMIT_MAX
] = UINT_MAX
;
519 tg
->iops_conf
[WRITE
][LIMIT_MAX
] = UINT_MAX
;
520 /* LIMIT_LOW will have default value 0 */
522 tg
->latency_target
= DFL_LATENCY_TARGET
;
523 tg
->latency_target_conf
= DFL_LATENCY_TARGET
;
524 tg
->idletime_threshold
= DFL_IDLE_THRESHOLD
;
525 tg
->idletime_threshold_conf
= DFL_IDLE_THRESHOLD
;
530 blkg_rwstat_exit(&tg
->stat_bytes
);
536 static void throtl_pd_init(struct blkg_policy_data
*pd
)
538 struct throtl_grp
*tg
= pd_to_tg(pd
);
539 struct blkcg_gq
*blkg
= tg_to_blkg(tg
);
540 struct throtl_data
*td
= blkg
->q
->td
;
541 struct throtl_service_queue
*sq
= &tg
->service_queue
;
544 * If on the default hierarchy, we switch to properly hierarchical
545 * behavior where limits on a given throtl_grp are applied to the
546 * whole subtree rather than just the group itself. e.g. If 16M
547 * read_bps limit is set on the root group, the whole system can't
548 * exceed 16M for the device.
550 * If not on the default hierarchy, the broken flat hierarchy
551 * behavior is retained where all throtl_grps are treated as if
552 * they're all separate root groups right below throtl_data.
553 * Limits of a group don't interact with limits of other groups
554 * regardless of the position of the group in the hierarchy.
556 sq
->parent_sq
= &td
->service_queue
;
557 if (cgroup_subsys_on_dfl(io_cgrp_subsys
) && blkg
->parent
)
558 sq
->parent_sq
= &blkg_to_tg(blkg
->parent
)->service_queue
;
563 * Set has_rules[] if @tg or any of its parents have limits configured.
564 * This doesn't require walking up to the top of the hierarchy as the
565 * parent's has_rules[] is guaranteed to be correct.
567 static void tg_update_has_rules(struct throtl_grp
*tg
)
569 struct throtl_grp
*parent_tg
= sq_to_tg(tg
->service_queue
.parent_sq
);
570 struct throtl_data
*td
= tg
->td
;
573 for (rw
= READ
; rw
<= WRITE
; rw
++)
574 tg
->has_rules
[rw
] = (parent_tg
&& parent_tg
->has_rules
[rw
]) ||
575 (td
->limit_valid
[td
->limit_index
] &&
576 (tg_bps_limit(tg
, rw
) != U64_MAX
||
577 tg_iops_limit(tg
, rw
) != UINT_MAX
));
580 static void throtl_pd_online(struct blkg_policy_data
*pd
)
582 struct throtl_grp
*tg
= pd_to_tg(pd
);
584 * We don't want new groups to escape the limits of its ancestors.
585 * Update has_rules[] after a new group is brought online.
587 tg_update_has_rules(tg
);
590 static void blk_throtl_update_limit_valid(struct throtl_data
*td
)
592 struct cgroup_subsys_state
*pos_css
;
593 struct blkcg_gq
*blkg
;
594 bool low_valid
= false;
597 blkg_for_each_descendant_post(blkg
, pos_css
, td
->queue
->root_blkg
) {
598 struct throtl_grp
*tg
= blkg_to_tg(blkg
);
600 if (tg
->bps
[READ
][LIMIT_LOW
] || tg
->bps
[WRITE
][LIMIT_LOW
] ||
601 tg
->iops
[READ
][LIMIT_LOW
] || tg
->iops
[WRITE
][LIMIT_LOW
]) {
608 td
->limit_valid
[LIMIT_LOW
] = low_valid
;
611 static void throtl_upgrade_state(struct throtl_data
*td
);
612 static void throtl_pd_offline(struct blkg_policy_data
*pd
)
614 struct throtl_grp
*tg
= pd_to_tg(pd
);
616 tg
->bps
[READ
][LIMIT_LOW
] = 0;
617 tg
->bps
[WRITE
][LIMIT_LOW
] = 0;
618 tg
->iops
[READ
][LIMIT_LOW
] = 0;
619 tg
->iops
[WRITE
][LIMIT_LOW
] = 0;
621 blk_throtl_update_limit_valid(tg
->td
);
623 if (!tg
->td
->limit_valid
[tg
->td
->limit_index
])
624 throtl_upgrade_state(tg
->td
);
627 static void throtl_pd_free(struct blkg_policy_data
*pd
)
629 struct throtl_grp
*tg
= pd_to_tg(pd
);
631 del_timer_sync(&tg
->service_queue
.pending_timer
);
632 blkg_rwstat_exit(&tg
->stat_bytes
);
633 blkg_rwstat_exit(&tg
->stat_ios
);
637 static struct throtl_grp
*
638 throtl_rb_first(struct throtl_service_queue
*parent_sq
)
642 n
= rb_first_cached(&parent_sq
->pending_tree
);
646 return rb_entry_tg(n
);
649 static void throtl_rb_erase(struct rb_node
*n
,
650 struct throtl_service_queue
*parent_sq
)
652 rb_erase_cached(n
, &parent_sq
->pending_tree
);
654 --parent_sq
->nr_pending
;
657 static void update_min_dispatch_time(struct throtl_service_queue
*parent_sq
)
659 struct throtl_grp
*tg
;
661 tg
= throtl_rb_first(parent_sq
);
665 parent_sq
->first_pending_disptime
= tg
->disptime
;
668 static void tg_service_queue_add(struct throtl_grp
*tg
)
670 struct throtl_service_queue
*parent_sq
= tg
->service_queue
.parent_sq
;
671 struct rb_node
**node
= &parent_sq
->pending_tree
.rb_root
.rb_node
;
672 struct rb_node
*parent
= NULL
;
673 struct throtl_grp
*__tg
;
674 unsigned long key
= tg
->disptime
;
675 bool leftmost
= true;
677 while (*node
!= NULL
) {
679 __tg
= rb_entry_tg(parent
);
681 if (time_before(key
, __tg
->disptime
))
682 node
= &parent
->rb_left
;
684 node
= &parent
->rb_right
;
689 rb_link_node(&tg
->rb_node
, parent
, node
);
690 rb_insert_color_cached(&tg
->rb_node
, &parent_sq
->pending_tree
,
694 static void throtl_enqueue_tg(struct throtl_grp
*tg
)
696 if (!(tg
->flags
& THROTL_TG_PENDING
)) {
697 tg_service_queue_add(tg
);
698 tg
->flags
|= THROTL_TG_PENDING
;
699 tg
->service_queue
.parent_sq
->nr_pending
++;
703 static void throtl_dequeue_tg(struct throtl_grp
*tg
)
705 if (tg
->flags
& THROTL_TG_PENDING
) {
706 throtl_rb_erase(&tg
->rb_node
, tg
->service_queue
.parent_sq
);
707 tg
->flags
&= ~THROTL_TG_PENDING
;
711 /* Call with queue lock held */
712 static void throtl_schedule_pending_timer(struct throtl_service_queue
*sq
,
713 unsigned long expires
)
715 unsigned long max_expire
= jiffies
+ 8 * sq_to_td(sq
)->throtl_slice
;
718 * Since we are adjusting the throttle limit dynamically, the sleep
719 * time calculated according to previous limit might be invalid. It's
720 * possible the cgroup sleep time is very long and no other cgroups
721 * have IO running so notify the limit changes. Make sure the cgroup
722 * doesn't sleep too long to avoid the missed notification.
724 if (time_after(expires
, max_expire
))
725 expires
= max_expire
;
726 mod_timer(&sq
->pending_timer
, expires
);
727 throtl_log(sq
, "schedule timer. delay=%lu jiffies=%lu",
728 expires
- jiffies
, jiffies
);
732 * throtl_schedule_next_dispatch - schedule the next dispatch cycle
733 * @sq: the service_queue to schedule dispatch for
734 * @force: force scheduling
736 * Arm @sq->pending_timer so that the next dispatch cycle starts on the
737 * dispatch time of the first pending child. Returns %true if either timer
738 * is armed or there's no pending child left. %false if the current
739 * dispatch window is still open and the caller should continue
742 * If @force is %true, the dispatch timer is always scheduled and this
743 * function is guaranteed to return %true. This is to be used when the
744 * caller can't dispatch itself and needs to invoke pending_timer
745 * unconditionally. Note that forced scheduling is likely to induce short
746 * delay before dispatch starts even if @sq->first_pending_disptime is not
747 * in the future and thus shouldn't be used in hot paths.
749 static bool throtl_schedule_next_dispatch(struct throtl_service_queue
*sq
,
752 /* any pending children left? */
756 update_min_dispatch_time(sq
);
758 /* is the next dispatch time in the future? */
759 if (force
|| time_after(sq
->first_pending_disptime
, jiffies
)) {
760 throtl_schedule_pending_timer(sq
, sq
->first_pending_disptime
);
764 /* tell the caller to continue dispatching */
768 static inline void throtl_start_new_slice_with_credit(struct throtl_grp
*tg
,
769 bool rw
, unsigned long start
)
771 tg
->bytes_disp
[rw
] = 0;
775 * Previous slice has expired. We must have trimmed it after last
776 * bio dispatch. That means since start of last slice, we never used
777 * that bandwidth. Do try to make use of that bandwidth while giving
780 if (time_after_eq(start
, tg
->slice_start
[rw
]))
781 tg
->slice_start
[rw
] = start
;
783 tg
->slice_end
[rw
] = jiffies
+ tg
->td
->throtl_slice
;
784 throtl_log(&tg
->service_queue
,
785 "[%c] new slice with credit start=%lu end=%lu jiffies=%lu",
786 rw
== READ
? 'R' : 'W', tg
->slice_start
[rw
],
787 tg
->slice_end
[rw
], jiffies
);
790 static inline void throtl_start_new_slice(struct throtl_grp
*tg
, bool rw
)
792 tg
->bytes_disp
[rw
] = 0;
794 tg
->slice_start
[rw
] = jiffies
;
795 tg
->slice_end
[rw
] = jiffies
+ tg
->td
->throtl_slice
;
796 throtl_log(&tg
->service_queue
,
797 "[%c] new slice start=%lu end=%lu jiffies=%lu",
798 rw
== READ
? 'R' : 'W', tg
->slice_start
[rw
],
799 tg
->slice_end
[rw
], jiffies
);
802 static inline void throtl_set_slice_end(struct throtl_grp
*tg
, bool rw
,
803 unsigned long jiffy_end
)
805 tg
->slice_end
[rw
] = roundup(jiffy_end
, tg
->td
->throtl_slice
);
808 static inline void throtl_extend_slice(struct throtl_grp
*tg
, bool rw
,
809 unsigned long jiffy_end
)
811 throtl_set_slice_end(tg
, rw
, jiffy_end
);
812 throtl_log(&tg
->service_queue
,
813 "[%c] extend slice start=%lu end=%lu jiffies=%lu",
814 rw
== READ
? 'R' : 'W', tg
->slice_start
[rw
],
815 tg
->slice_end
[rw
], jiffies
);
818 /* Determine if previously allocated or extended slice is complete or not */
819 static bool throtl_slice_used(struct throtl_grp
*tg
, bool rw
)
821 if (time_in_range(jiffies
, tg
->slice_start
[rw
], tg
->slice_end
[rw
]))
827 /* Trim the used slices and adjust slice start accordingly */
828 static inline void throtl_trim_slice(struct throtl_grp
*tg
, bool rw
)
830 unsigned long nr_slices
, time_elapsed
, io_trim
;
833 BUG_ON(time_before(tg
->slice_end
[rw
], tg
->slice_start
[rw
]));
836 * If bps are unlimited (-1), then time slice don't get
837 * renewed. Don't try to trim the slice if slice is used. A new
838 * slice will start when appropriate.
840 if (throtl_slice_used(tg
, rw
))
844 * A bio has been dispatched. Also adjust slice_end. It might happen
845 * that initially cgroup limit was very low resulting in high
846 * slice_end, but later limit was bumped up and bio was dispatched
847 * sooner, then we need to reduce slice_end. A high bogus slice_end
848 * is bad because it does not allow new slice to start.
851 throtl_set_slice_end(tg
, rw
, jiffies
+ tg
->td
->throtl_slice
);
853 time_elapsed
= jiffies
- tg
->slice_start
[rw
];
855 nr_slices
= time_elapsed
/ tg
->td
->throtl_slice
;
859 tmp
= tg_bps_limit(tg
, rw
) * tg
->td
->throtl_slice
* nr_slices
;
863 io_trim
= (tg_iops_limit(tg
, rw
) * tg
->td
->throtl_slice
* nr_slices
) /
866 if (!bytes_trim
&& !io_trim
)
869 if (tg
->bytes_disp
[rw
] >= bytes_trim
)
870 tg
->bytes_disp
[rw
] -= bytes_trim
;
872 tg
->bytes_disp
[rw
] = 0;
874 if (tg
->io_disp
[rw
] >= io_trim
)
875 tg
->io_disp
[rw
] -= io_trim
;
879 tg
->slice_start
[rw
] += nr_slices
* tg
->td
->throtl_slice
;
881 throtl_log(&tg
->service_queue
,
882 "[%c] trim slice nr=%lu bytes=%llu io=%lu start=%lu end=%lu jiffies=%lu",
883 rw
== READ
? 'R' : 'W', nr_slices
, bytes_trim
, io_trim
,
884 tg
->slice_start
[rw
], tg
->slice_end
[rw
], jiffies
);
887 static bool tg_with_in_iops_limit(struct throtl_grp
*tg
, struct bio
*bio
,
888 u32 iops_limit
, unsigned long *wait
)
890 bool rw
= bio_data_dir(bio
);
891 unsigned int io_allowed
;
892 unsigned long jiffy_elapsed
, jiffy_wait
, jiffy_elapsed_rnd
;
895 if (iops_limit
== UINT_MAX
) {
901 jiffy_elapsed
= jiffies
- tg
->slice_start
[rw
];
903 /* Round up to the next throttle slice, wait time must be nonzero */
904 jiffy_elapsed_rnd
= roundup(jiffy_elapsed
+ 1, tg
->td
->throtl_slice
);
907 * jiffy_elapsed_rnd should not be a big value as minimum iops can be
908 * 1 then at max jiffy elapsed should be equivalent of 1 second as we
909 * will allow dispatch after 1 second and after that slice should
913 tmp
= (u64
)iops_limit
* jiffy_elapsed_rnd
;
917 io_allowed
= UINT_MAX
;
921 if (tg
->io_disp
[rw
] + 1 <= io_allowed
) {
927 /* Calc approx time to dispatch */
928 jiffy_wait
= jiffy_elapsed_rnd
- jiffy_elapsed
;
935 static bool tg_with_in_bps_limit(struct throtl_grp
*tg
, struct bio
*bio
,
936 u64 bps_limit
, unsigned long *wait
)
938 bool rw
= bio_data_dir(bio
);
939 u64 bytes_allowed
, extra_bytes
, tmp
;
940 unsigned long jiffy_elapsed
, jiffy_wait
, jiffy_elapsed_rnd
;
941 unsigned int bio_size
= throtl_bio_data_size(bio
);
943 if (bps_limit
== U64_MAX
) {
949 jiffy_elapsed
= jiffy_elapsed_rnd
= jiffies
- tg
->slice_start
[rw
];
951 /* Slice has just started. Consider one slice interval */
953 jiffy_elapsed_rnd
= tg
->td
->throtl_slice
;
955 jiffy_elapsed_rnd
= roundup(jiffy_elapsed_rnd
, tg
->td
->throtl_slice
);
957 tmp
= bps_limit
* jiffy_elapsed_rnd
;
961 if (tg
->bytes_disp
[rw
] + bio_size
<= bytes_allowed
) {
967 /* Calc approx time to dispatch */
968 extra_bytes
= tg
->bytes_disp
[rw
] + bio_size
- bytes_allowed
;
969 jiffy_wait
= div64_u64(extra_bytes
* HZ
, bps_limit
);
975 * This wait time is without taking into consideration the rounding
976 * up we did. Add that time also.
978 jiffy_wait
= jiffy_wait
+ (jiffy_elapsed_rnd
- jiffy_elapsed
);
985 * Returns whether one can dispatch a bio or not. Also returns approx number
986 * of jiffies to wait before this bio is with-in IO rate and can be dispatched
988 static bool tg_may_dispatch(struct throtl_grp
*tg
, struct bio
*bio
,
991 bool rw
= bio_data_dir(bio
);
992 unsigned long bps_wait
= 0, iops_wait
= 0, max_wait
= 0;
993 u64 bps_limit
= tg_bps_limit(tg
, rw
);
994 u32 iops_limit
= tg_iops_limit(tg
, rw
);
997 * Currently whole state machine of group depends on first bio
998 * queued in the group bio list. So one should not be calling
999 * this function with a different bio if there are other bios
1002 BUG_ON(tg
->service_queue
.nr_queued
[rw
] &&
1003 bio
!= throtl_peek_queued(&tg
->service_queue
.queued
[rw
]));
1005 /* If tg->bps = -1, then BW is unlimited */
1006 if (bps_limit
== U64_MAX
&& iops_limit
== UINT_MAX
) {
1013 * If previous slice expired, start a new one otherwise renew/extend
1014 * existing slice to make sure it is at least throtl_slice interval
1015 * long since now. New slice is started only for empty throttle group.
1016 * If there is queued bio, that means there should be an active
1017 * slice and it should be extended instead.
1019 if (throtl_slice_used(tg
, rw
) && !(tg
->service_queue
.nr_queued
[rw
]))
1020 throtl_start_new_slice(tg
, rw
);
1022 if (time_before(tg
->slice_end
[rw
],
1023 jiffies
+ tg
->td
->throtl_slice
))
1024 throtl_extend_slice(tg
, rw
,
1025 jiffies
+ tg
->td
->throtl_slice
);
1028 if (tg_with_in_bps_limit(tg
, bio
, bps_limit
, &bps_wait
) &&
1029 tg_with_in_iops_limit(tg
, bio
, iops_limit
, &iops_wait
)) {
1035 max_wait
= max(bps_wait
, iops_wait
);
1040 if (time_before(tg
->slice_end
[rw
], jiffies
+ max_wait
))
1041 throtl_extend_slice(tg
, rw
, jiffies
+ max_wait
);
1046 static void throtl_charge_bio(struct throtl_grp
*tg
, struct bio
*bio
)
1048 bool rw
= bio_data_dir(bio
);
1049 unsigned int bio_size
= throtl_bio_data_size(bio
);
1051 /* Charge the bio to the group */
1052 tg
->bytes_disp
[rw
] += bio_size
;
1054 tg
->last_bytes_disp
[rw
] += bio_size
;
1055 tg
->last_io_disp
[rw
]++;
1058 * BIO_THROTTLED is used to prevent the same bio to be throttled
1059 * more than once as a throttled bio will go through blk-throtl the
1060 * second time when it eventually gets issued. Set it when a bio
1061 * is being charged to a tg.
1063 if (!bio_flagged(bio
, BIO_THROTTLED
))
1064 bio_set_flag(bio
, BIO_THROTTLED
);
1068 * throtl_add_bio_tg - add a bio to the specified throtl_grp
1071 * @tg: the target throtl_grp
1073 * Add @bio to @tg's service_queue using @qn. If @qn is not specified,
1074 * tg->qnode_on_self[] is used.
1076 static void throtl_add_bio_tg(struct bio
*bio
, struct throtl_qnode
*qn
,
1077 struct throtl_grp
*tg
)
1079 struct throtl_service_queue
*sq
= &tg
->service_queue
;
1080 bool rw
= bio_data_dir(bio
);
1083 qn
= &tg
->qnode_on_self
[rw
];
1086 * If @tg doesn't currently have any bios queued in the same
1087 * direction, queueing @bio can change when @tg should be
1088 * dispatched. Mark that @tg was empty. This is automatically
1089 * cleared on the next tg_update_disptime().
1091 if (!sq
->nr_queued
[rw
])
1092 tg
->flags
|= THROTL_TG_WAS_EMPTY
;
1094 throtl_qnode_add_bio(bio
, qn
, &sq
->queued
[rw
]);
1096 sq
->nr_queued
[rw
]++;
1097 throtl_enqueue_tg(tg
);
1100 static void tg_update_disptime(struct throtl_grp
*tg
)
1102 struct throtl_service_queue
*sq
= &tg
->service_queue
;
1103 unsigned long read_wait
= -1, write_wait
= -1, min_wait
= -1, disptime
;
1106 bio
= throtl_peek_queued(&sq
->queued
[READ
]);
1108 tg_may_dispatch(tg
, bio
, &read_wait
);
1110 bio
= throtl_peek_queued(&sq
->queued
[WRITE
]);
1112 tg_may_dispatch(tg
, bio
, &write_wait
);
1114 min_wait
= min(read_wait
, write_wait
);
1115 disptime
= jiffies
+ min_wait
;
1117 /* Update dispatch time */
1118 throtl_dequeue_tg(tg
);
1119 tg
->disptime
= disptime
;
1120 throtl_enqueue_tg(tg
);
1122 /* see throtl_add_bio_tg() */
1123 tg
->flags
&= ~THROTL_TG_WAS_EMPTY
;
1126 static void start_parent_slice_with_credit(struct throtl_grp
*child_tg
,
1127 struct throtl_grp
*parent_tg
, bool rw
)
1129 if (throtl_slice_used(parent_tg
, rw
)) {
1130 throtl_start_new_slice_with_credit(parent_tg
, rw
,
1131 child_tg
->slice_start
[rw
]);
1136 static void tg_dispatch_one_bio(struct throtl_grp
*tg
, bool rw
)
1138 struct throtl_service_queue
*sq
= &tg
->service_queue
;
1139 struct throtl_service_queue
*parent_sq
= sq
->parent_sq
;
1140 struct throtl_grp
*parent_tg
= sq_to_tg(parent_sq
);
1141 struct throtl_grp
*tg_to_put
= NULL
;
1145 * @bio is being transferred from @tg to @parent_sq. Popping a bio
1146 * from @tg may put its reference and @parent_sq might end up
1147 * getting released prematurely. Remember the tg to put and put it
1148 * after @bio is transferred to @parent_sq.
1150 bio
= throtl_pop_queued(&sq
->queued
[rw
], &tg_to_put
);
1151 sq
->nr_queued
[rw
]--;
1153 throtl_charge_bio(tg
, bio
);
1156 * If our parent is another tg, we just need to transfer @bio to
1157 * the parent using throtl_add_bio_tg(). If our parent is
1158 * @td->service_queue, @bio is ready to be issued. Put it on its
1159 * bio_lists[] and decrease total number queued. The caller is
1160 * responsible for issuing these bios.
1163 throtl_add_bio_tg(bio
, &tg
->qnode_on_parent
[rw
], parent_tg
);
1164 start_parent_slice_with_credit(tg
, parent_tg
, rw
);
1166 throtl_qnode_add_bio(bio
, &tg
->qnode_on_parent
[rw
],
1167 &parent_sq
->queued
[rw
]);
1168 BUG_ON(tg
->td
->nr_queued
[rw
] <= 0);
1169 tg
->td
->nr_queued
[rw
]--;
1172 throtl_trim_slice(tg
, rw
);
1175 blkg_put(tg_to_blkg(tg_to_put
));
1178 static int throtl_dispatch_tg(struct throtl_grp
*tg
)
1180 struct throtl_service_queue
*sq
= &tg
->service_queue
;
1181 unsigned int nr_reads
= 0, nr_writes
= 0;
1182 unsigned int max_nr_reads
= THROTL_GRP_QUANTUM
* 3 / 4;
1183 unsigned int max_nr_writes
= THROTL_GRP_QUANTUM
- max_nr_reads
;
1186 /* Try to dispatch 75% READS and 25% WRITES */
1188 while ((bio
= throtl_peek_queued(&sq
->queued
[READ
])) &&
1189 tg_may_dispatch(tg
, bio
, NULL
)) {
1191 tg_dispatch_one_bio(tg
, bio_data_dir(bio
));
1194 if (nr_reads
>= max_nr_reads
)
1198 while ((bio
= throtl_peek_queued(&sq
->queued
[WRITE
])) &&
1199 tg_may_dispatch(tg
, bio
, NULL
)) {
1201 tg_dispatch_one_bio(tg
, bio_data_dir(bio
));
1204 if (nr_writes
>= max_nr_writes
)
1208 return nr_reads
+ nr_writes
;
1211 static int throtl_select_dispatch(struct throtl_service_queue
*parent_sq
)
1213 unsigned int nr_disp
= 0;
1216 struct throtl_grp
*tg
;
1217 struct throtl_service_queue
*sq
;
1219 if (!parent_sq
->nr_pending
)
1222 tg
= throtl_rb_first(parent_sq
);
1226 if (time_before(jiffies
, tg
->disptime
))
1229 throtl_dequeue_tg(tg
);
1231 nr_disp
+= throtl_dispatch_tg(tg
);
1233 sq
= &tg
->service_queue
;
1234 if (sq
->nr_queued
[0] || sq
->nr_queued
[1])
1235 tg_update_disptime(tg
);
1237 if (nr_disp
>= THROTL_QUANTUM
)
1244 static bool throtl_can_upgrade(struct throtl_data
*td
,
1245 struct throtl_grp
*this_tg
);
1247 * throtl_pending_timer_fn - timer function for service_queue->pending_timer
1248 * @t: the pending_timer member of the throtl_service_queue being serviced
1250 * This timer is armed when a child throtl_grp with active bio's become
1251 * pending and queued on the service_queue's pending_tree and expires when
1252 * the first child throtl_grp should be dispatched. This function
1253 * dispatches bio's from the children throtl_grps to the parent
1256 * If the parent's parent is another throtl_grp, dispatching is propagated
1257 * by either arming its pending_timer or repeating dispatch directly. If
1258 * the top-level service_tree is reached, throtl_data->dispatch_work is
1259 * kicked so that the ready bio's are issued.
1261 static void throtl_pending_timer_fn(struct timer_list
*t
)
1263 struct throtl_service_queue
*sq
= from_timer(sq
, t
, pending_timer
);
1264 struct throtl_grp
*tg
= sq_to_tg(sq
);
1265 struct throtl_data
*td
= sq_to_td(sq
);
1266 struct request_queue
*q
= td
->queue
;
1267 struct throtl_service_queue
*parent_sq
;
1271 spin_lock_irq(&q
->queue_lock
);
1272 if (throtl_can_upgrade(td
, NULL
))
1273 throtl_upgrade_state(td
);
1276 parent_sq
= sq
->parent_sq
;
1280 throtl_log(sq
, "dispatch nr_queued=%u read=%u write=%u",
1281 sq
->nr_queued
[READ
] + sq
->nr_queued
[WRITE
],
1282 sq
->nr_queued
[READ
], sq
->nr_queued
[WRITE
]);
1284 ret
= throtl_select_dispatch(sq
);
1286 throtl_log(sq
, "bios disp=%u", ret
);
1290 if (throtl_schedule_next_dispatch(sq
, false))
1293 /* this dispatch windows is still open, relax and repeat */
1294 spin_unlock_irq(&q
->queue_lock
);
1296 spin_lock_irq(&q
->queue_lock
);
1303 /* @parent_sq is another throl_grp, propagate dispatch */
1304 if (tg
->flags
& THROTL_TG_WAS_EMPTY
) {
1305 tg_update_disptime(tg
);
1306 if (!throtl_schedule_next_dispatch(parent_sq
, false)) {
1307 /* window is already open, repeat dispatching */
1314 /* reached the top-level, queue issuing */
1315 queue_work(kthrotld_workqueue
, &td
->dispatch_work
);
1318 spin_unlock_irq(&q
->queue_lock
);
1322 * blk_throtl_dispatch_work_fn - work function for throtl_data->dispatch_work
1323 * @work: work item being executed
1325 * This function is queued for execution when bios reach the bio_lists[]
1326 * of throtl_data->service_queue. Those bios are ready and issued by this
1329 static void blk_throtl_dispatch_work_fn(struct work_struct
*work
)
1331 struct throtl_data
*td
= container_of(work
, struct throtl_data
,
1333 struct throtl_service_queue
*td_sq
= &td
->service_queue
;
1334 struct request_queue
*q
= td
->queue
;
1335 struct bio_list bio_list_on_stack
;
1337 struct blk_plug plug
;
1340 bio_list_init(&bio_list_on_stack
);
1342 spin_lock_irq(&q
->queue_lock
);
1343 for (rw
= READ
; rw
<= WRITE
; rw
++)
1344 while ((bio
= throtl_pop_queued(&td_sq
->queued
[rw
], NULL
)))
1345 bio_list_add(&bio_list_on_stack
, bio
);
1346 spin_unlock_irq(&q
->queue_lock
);
1348 if (!bio_list_empty(&bio_list_on_stack
)) {
1349 blk_start_plug(&plug
);
1350 while ((bio
= bio_list_pop(&bio_list_on_stack
)))
1351 submit_bio_noacct(bio
);
1352 blk_finish_plug(&plug
);
1356 static u64
tg_prfill_conf_u64(struct seq_file
*sf
, struct blkg_policy_data
*pd
,
1359 struct throtl_grp
*tg
= pd_to_tg(pd
);
1360 u64 v
= *(u64
*)((void *)tg
+ off
);
1364 return __blkg_prfill_u64(sf
, pd
, v
);
1367 static u64
tg_prfill_conf_uint(struct seq_file
*sf
, struct blkg_policy_data
*pd
,
1370 struct throtl_grp
*tg
= pd_to_tg(pd
);
1371 unsigned int v
= *(unsigned int *)((void *)tg
+ off
);
1375 return __blkg_prfill_u64(sf
, pd
, v
);
1378 static int tg_print_conf_u64(struct seq_file
*sf
, void *v
)
1380 blkcg_print_blkgs(sf
, css_to_blkcg(seq_css(sf
)), tg_prfill_conf_u64
,
1381 &blkcg_policy_throtl
, seq_cft(sf
)->private, false);
1385 static int tg_print_conf_uint(struct seq_file
*sf
, void *v
)
1387 blkcg_print_blkgs(sf
, css_to_blkcg(seq_css(sf
)), tg_prfill_conf_uint
,
1388 &blkcg_policy_throtl
, seq_cft(sf
)->private, false);
1392 static void tg_conf_updated(struct throtl_grp
*tg
, bool global
)
1394 struct throtl_service_queue
*sq
= &tg
->service_queue
;
1395 struct cgroup_subsys_state
*pos_css
;
1396 struct blkcg_gq
*blkg
;
1398 throtl_log(&tg
->service_queue
,
1399 "limit change rbps=%llu wbps=%llu riops=%u wiops=%u",
1400 tg_bps_limit(tg
, READ
), tg_bps_limit(tg
, WRITE
),
1401 tg_iops_limit(tg
, READ
), tg_iops_limit(tg
, WRITE
));
1404 * Update has_rules[] flags for the updated tg's subtree. A tg is
1405 * considered to have rules if either the tg itself or any of its
1406 * ancestors has rules. This identifies groups without any
1407 * restrictions in the whole hierarchy and allows them to bypass
1410 blkg_for_each_descendant_pre(blkg
, pos_css
,
1411 global
? tg
->td
->queue
->root_blkg
: tg_to_blkg(tg
)) {
1412 struct throtl_grp
*this_tg
= blkg_to_tg(blkg
);
1413 struct throtl_grp
*parent_tg
;
1415 tg_update_has_rules(this_tg
);
1416 /* ignore root/second level */
1417 if (!cgroup_subsys_on_dfl(io_cgrp_subsys
) || !blkg
->parent
||
1418 !blkg
->parent
->parent
)
1420 parent_tg
= blkg_to_tg(blkg
->parent
);
1422 * make sure all children has lower idle time threshold and
1423 * higher latency target
1425 this_tg
->idletime_threshold
= min(this_tg
->idletime_threshold
,
1426 parent_tg
->idletime_threshold
);
1427 this_tg
->latency_target
= max(this_tg
->latency_target
,
1428 parent_tg
->latency_target
);
1432 * We're already holding queue_lock and know @tg is valid. Let's
1433 * apply the new config directly.
1435 * Restart the slices for both READ and WRITES. It might happen
1436 * that a group's limit are dropped suddenly and we don't want to
1437 * account recently dispatched IO with new low rate.
1439 throtl_start_new_slice(tg
, READ
);
1440 throtl_start_new_slice(tg
, WRITE
);
1442 if (tg
->flags
& THROTL_TG_PENDING
) {
1443 tg_update_disptime(tg
);
1444 throtl_schedule_next_dispatch(sq
->parent_sq
, true);
1448 static ssize_t
tg_set_conf(struct kernfs_open_file
*of
,
1449 char *buf
, size_t nbytes
, loff_t off
, bool is_u64
)
1451 struct blkcg
*blkcg
= css_to_blkcg(of_css(of
));
1452 struct blkg_conf_ctx ctx
;
1453 struct throtl_grp
*tg
;
1457 ret
= blkg_conf_prep(blkcg
, &blkcg_policy_throtl
, buf
, &ctx
);
1462 if (sscanf(ctx
.body
, "%llu", &v
) != 1)
1467 tg
= blkg_to_tg(ctx
.blkg
);
1470 *(u64
*)((void *)tg
+ of_cft(of
)->private) = v
;
1472 *(unsigned int *)((void *)tg
+ of_cft(of
)->private) = v
;
1474 tg_conf_updated(tg
, false);
1477 blkg_conf_finish(&ctx
);
1478 return ret
?: nbytes
;
1481 static ssize_t
tg_set_conf_u64(struct kernfs_open_file
*of
,
1482 char *buf
, size_t nbytes
, loff_t off
)
1484 return tg_set_conf(of
, buf
, nbytes
, off
, true);
1487 static ssize_t
tg_set_conf_uint(struct kernfs_open_file
*of
,
1488 char *buf
, size_t nbytes
, loff_t off
)
1490 return tg_set_conf(of
, buf
, nbytes
, off
, false);
1493 static int tg_print_rwstat(struct seq_file
*sf
, void *v
)
1495 blkcg_print_blkgs(sf
, css_to_blkcg(seq_css(sf
)),
1496 blkg_prfill_rwstat
, &blkcg_policy_throtl
,
1497 seq_cft(sf
)->private, true);
1501 static u64
tg_prfill_rwstat_recursive(struct seq_file
*sf
,
1502 struct blkg_policy_data
*pd
, int off
)
1504 struct blkg_rwstat_sample sum
;
1506 blkg_rwstat_recursive_sum(pd_to_blkg(pd
), &blkcg_policy_throtl
, off
,
1508 return __blkg_prfill_rwstat(sf
, pd
, &sum
);
1511 static int tg_print_rwstat_recursive(struct seq_file
*sf
, void *v
)
1513 blkcg_print_blkgs(sf
, css_to_blkcg(seq_css(sf
)),
1514 tg_prfill_rwstat_recursive
, &blkcg_policy_throtl
,
1515 seq_cft(sf
)->private, true);
1519 static struct cftype throtl_legacy_files
[] = {
1521 .name
= "throttle.read_bps_device",
1522 .private = offsetof(struct throtl_grp
, bps
[READ
][LIMIT_MAX
]),
1523 .seq_show
= tg_print_conf_u64
,
1524 .write
= tg_set_conf_u64
,
1527 .name
= "throttle.write_bps_device",
1528 .private = offsetof(struct throtl_grp
, bps
[WRITE
][LIMIT_MAX
]),
1529 .seq_show
= tg_print_conf_u64
,
1530 .write
= tg_set_conf_u64
,
1533 .name
= "throttle.read_iops_device",
1534 .private = offsetof(struct throtl_grp
, iops
[READ
][LIMIT_MAX
]),
1535 .seq_show
= tg_print_conf_uint
,
1536 .write
= tg_set_conf_uint
,
1539 .name
= "throttle.write_iops_device",
1540 .private = offsetof(struct throtl_grp
, iops
[WRITE
][LIMIT_MAX
]),
1541 .seq_show
= tg_print_conf_uint
,
1542 .write
= tg_set_conf_uint
,
1545 .name
= "throttle.io_service_bytes",
1546 .private = offsetof(struct throtl_grp
, stat_bytes
),
1547 .seq_show
= tg_print_rwstat
,
1550 .name
= "throttle.io_service_bytes_recursive",
1551 .private = offsetof(struct throtl_grp
, stat_bytes
),
1552 .seq_show
= tg_print_rwstat_recursive
,
1555 .name
= "throttle.io_serviced",
1556 .private = offsetof(struct throtl_grp
, stat_ios
),
1557 .seq_show
= tg_print_rwstat
,
1560 .name
= "throttle.io_serviced_recursive",
1561 .private = offsetof(struct throtl_grp
, stat_ios
),
1562 .seq_show
= tg_print_rwstat_recursive
,
1567 static u64
tg_prfill_limit(struct seq_file
*sf
, struct blkg_policy_data
*pd
,
1570 struct throtl_grp
*tg
= pd_to_tg(pd
);
1571 const char *dname
= blkg_dev_name(pd
->blkg
);
1572 char bufs
[4][21] = { "max", "max", "max", "max" };
1574 unsigned int iops_dft
;
1575 char idle_time
[26] = "";
1576 char latency_time
[26] = "";
1581 if (off
== LIMIT_LOW
) {
1586 iops_dft
= UINT_MAX
;
1589 if (tg
->bps_conf
[READ
][off
] == bps_dft
&&
1590 tg
->bps_conf
[WRITE
][off
] == bps_dft
&&
1591 tg
->iops_conf
[READ
][off
] == iops_dft
&&
1592 tg
->iops_conf
[WRITE
][off
] == iops_dft
&&
1593 (off
!= LIMIT_LOW
||
1594 (tg
->idletime_threshold_conf
== DFL_IDLE_THRESHOLD
&&
1595 tg
->latency_target_conf
== DFL_LATENCY_TARGET
)))
1598 if (tg
->bps_conf
[READ
][off
] != U64_MAX
)
1599 snprintf(bufs
[0], sizeof(bufs
[0]), "%llu",
1600 tg
->bps_conf
[READ
][off
]);
1601 if (tg
->bps_conf
[WRITE
][off
] != U64_MAX
)
1602 snprintf(bufs
[1], sizeof(bufs
[1]), "%llu",
1603 tg
->bps_conf
[WRITE
][off
]);
1604 if (tg
->iops_conf
[READ
][off
] != UINT_MAX
)
1605 snprintf(bufs
[2], sizeof(bufs
[2]), "%u",
1606 tg
->iops_conf
[READ
][off
]);
1607 if (tg
->iops_conf
[WRITE
][off
] != UINT_MAX
)
1608 snprintf(bufs
[3], sizeof(bufs
[3]), "%u",
1609 tg
->iops_conf
[WRITE
][off
]);
1610 if (off
== LIMIT_LOW
) {
1611 if (tg
->idletime_threshold_conf
== ULONG_MAX
)
1612 strcpy(idle_time
, " idle=max");
1614 snprintf(idle_time
, sizeof(idle_time
), " idle=%lu",
1615 tg
->idletime_threshold_conf
);
1617 if (tg
->latency_target_conf
== ULONG_MAX
)
1618 strcpy(latency_time
, " latency=max");
1620 snprintf(latency_time
, sizeof(latency_time
),
1621 " latency=%lu", tg
->latency_target_conf
);
1624 seq_printf(sf
, "%s rbps=%s wbps=%s riops=%s wiops=%s%s%s\n",
1625 dname
, bufs
[0], bufs
[1], bufs
[2], bufs
[3], idle_time
,
1630 static int tg_print_limit(struct seq_file
*sf
, void *v
)
1632 blkcg_print_blkgs(sf
, css_to_blkcg(seq_css(sf
)), tg_prfill_limit
,
1633 &blkcg_policy_throtl
, seq_cft(sf
)->private, false);
1637 static ssize_t
tg_set_limit(struct kernfs_open_file
*of
,
1638 char *buf
, size_t nbytes
, loff_t off
)
1640 struct blkcg
*blkcg
= css_to_blkcg(of_css(of
));
1641 struct blkg_conf_ctx ctx
;
1642 struct throtl_grp
*tg
;
1644 unsigned long idle_time
;
1645 unsigned long latency_time
;
1647 int index
= of_cft(of
)->private;
1649 ret
= blkg_conf_prep(blkcg
, &blkcg_policy_throtl
, buf
, &ctx
);
1653 tg
= blkg_to_tg(ctx
.blkg
);
1655 v
[0] = tg
->bps_conf
[READ
][index
];
1656 v
[1] = tg
->bps_conf
[WRITE
][index
];
1657 v
[2] = tg
->iops_conf
[READ
][index
];
1658 v
[3] = tg
->iops_conf
[WRITE
][index
];
1660 idle_time
= tg
->idletime_threshold_conf
;
1661 latency_time
= tg
->latency_target_conf
;
1663 char tok
[27]; /* wiops=18446744073709551616 */
1668 if (sscanf(ctx
.body
, "%26s%n", tok
, &len
) != 1)
1677 if (!p
|| (sscanf(p
, "%llu", &val
) != 1 && strcmp(p
, "max")))
1685 if (!strcmp(tok
, "rbps") && val
> 1)
1687 else if (!strcmp(tok
, "wbps") && val
> 1)
1689 else if (!strcmp(tok
, "riops") && val
> 1)
1690 v
[2] = min_t(u64
, val
, UINT_MAX
);
1691 else if (!strcmp(tok
, "wiops") && val
> 1)
1692 v
[3] = min_t(u64
, val
, UINT_MAX
);
1693 else if (off
== LIMIT_LOW
&& !strcmp(tok
, "idle"))
1695 else if (off
== LIMIT_LOW
&& !strcmp(tok
, "latency"))
1701 tg
->bps_conf
[READ
][index
] = v
[0];
1702 tg
->bps_conf
[WRITE
][index
] = v
[1];
1703 tg
->iops_conf
[READ
][index
] = v
[2];
1704 tg
->iops_conf
[WRITE
][index
] = v
[3];
1706 if (index
== LIMIT_MAX
) {
1707 tg
->bps
[READ
][index
] = v
[0];
1708 tg
->bps
[WRITE
][index
] = v
[1];
1709 tg
->iops
[READ
][index
] = v
[2];
1710 tg
->iops
[WRITE
][index
] = v
[3];
1712 tg
->bps
[READ
][LIMIT_LOW
] = min(tg
->bps_conf
[READ
][LIMIT_LOW
],
1713 tg
->bps_conf
[READ
][LIMIT_MAX
]);
1714 tg
->bps
[WRITE
][LIMIT_LOW
] = min(tg
->bps_conf
[WRITE
][LIMIT_LOW
],
1715 tg
->bps_conf
[WRITE
][LIMIT_MAX
]);
1716 tg
->iops
[READ
][LIMIT_LOW
] = min(tg
->iops_conf
[READ
][LIMIT_LOW
],
1717 tg
->iops_conf
[READ
][LIMIT_MAX
]);
1718 tg
->iops
[WRITE
][LIMIT_LOW
] = min(tg
->iops_conf
[WRITE
][LIMIT_LOW
],
1719 tg
->iops_conf
[WRITE
][LIMIT_MAX
]);
1720 tg
->idletime_threshold_conf
= idle_time
;
1721 tg
->latency_target_conf
= latency_time
;
1723 /* force user to configure all settings for low limit */
1724 if (!(tg
->bps
[READ
][LIMIT_LOW
] || tg
->iops
[READ
][LIMIT_LOW
] ||
1725 tg
->bps
[WRITE
][LIMIT_LOW
] || tg
->iops
[WRITE
][LIMIT_LOW
]) ||
1726 tg
->idletime_threshold_conf
== DFL_IDLE_THRESHOLD
||
1727 tg
->latency_target_conf
== DFL_LATENCY_TARGET
) {
1728 tg
->bps
[READ
][LIMIT_LOW
] = 0;
1729 tg
->bps
[WRITE
][LIMIT_LOW
] = 0;
1730 tg
->iops
[READ
][LIMIT_LOW
] = 0;
1731 tg
->iops
[WRITE
][LIMIT_LOW
] = 0;
1732 tg
->idletime_threshold
= DFL_IDLE_THRESHOLD
;
1733 tg
->latency_target
= DFL_LATENCY_TARGET
;
1734 } else if (index
== LIMIT_LOW
) {
1735 tg
->idletime_threshold
= tg
->idletime_threshold_conf
;
1736 tg
->latency_target
= tg
->latency_target_conf
;
1739 blk_throtl_update_limit_valid(tg
->td
);
1740 if (tg
->td
->limit_valid
[LIMIT_LOW
]) {
1741 if (index
== LIMIT_LOW
)
1742 tg
->td
->limit_index
= LIMIT_LOW
;
1744 tg
->td
->limit_index
= LIMIT_MAX
;
1745 tg_conf_updated(tg
, index
== LIMIT_LOW
&&
1746 tg
->td
->limit_valid
[LIMIT_LOW
]);
1749 blkg_conf_finish(&ctx
);
1750 return ret
?: nbytes
;
1753 static struct cftype throtl_files
[] = {
1754 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
1757 .flags
= CFTYPE_NOT_ON_ROOT
,
1758 .seq_show
= tg_print_limit
,
1759 .write
= tg_set_limit
,
1760 .private = LIMIT_LOW
,
1765 .flags
= CFTYPE_NOT_ON_ROOT
,
1766 .seq_show
= tg_print_limit
,
1767 .write
= tg_set_limit
,
1768 .private = LIMIT_MAX
,
1773 static void throtl_shutdown_wq(struct request_queue
*q
)
1775 struct throtl_data
*td
= q
->td
;
1777 cancel_work_sync(&td
->dispatch_work
);
1780 static struct blkcg_policy blkcg_policy_throtl
= {
1781 .dfl_cftypes
= throtl_files
,
1782 .legacy_cftypes
= throtl_legacy_files
,
1784 .pd_alloc_fn
= throtl_pd_alloc
,
1785 .pd_init_fn
= throtl_pd_init
,
1786 .pd_online_fn
= throtl_pd_online
,
1787 .pd_offline_fn
= throtl_pd_offline
,
1788 .pd_free_fn
= throtl_pd_free
,
1791 static unsigned long __tg_last_low_overflow_time(struct throtl_grp
*tg
)
1793 unsigned long rtime
= jiffies
, wtime
= jiffies
;
1795 if (tg
->bps
[READ
][LIMIT_LOW
] || tg
->iops
[READ
][LIMIT_LOW
])
1796 rtime
= tg
->last_low_overflow_time
[READ
];
1797 if (tg
->bps
[WRITE
][LIMIT_LOW
] || tg
->iops
[WRITE
][LIMIT_LOW
])
1798 wtime
= tg
->last_low_overflow_time
[WRITE
];
1799 return min(rtime
, wtime
);
1802 /* tg should not be an intermediate node */
1803 static unsigned long tg_last_low_overflow_time(struct throtl_grp
*tg
)
1805 struct throtl_service_queue
*parent_sq
;
1806 struct throtl_grp
*parent
= tg
;
1807 unsigned long ret
= __tg_last_low_overflow_time(tg
);
1810 parent_sq
= parent
->service_queue
.parent_sq
;
1811 parent
= sq_to_tg(parent_sq
);
1816 * The parent doesn't have low limit, it always reaches low
1817 * limit. Its overflow time is useless for children
1819 if (!parent
->bps
[READ
][LIMIT_LOW
] &&
1820 !parent
->iops
[READ
][LIMIT_LOW
] &&
1821 !parent
->bps
[WRITE
][LIMIT_LOW
] &&
1822 !parent
->iops
[WRITE
][LIMIT_LOW
])
1824 if (time_after(__tg_last_low_overflow_time(parent
), ret
))
1825 ret
= __tg_last_low_overflow_time(parent
);
1830 static bool throtl_tg_is_idle(struct throtl_grp
*tg
)
1833 * cgroup is idle if:
1834 * - single idle is too long, longer than a fixed value (in case user
1835 * configure a too big threshold) or 4 times of idletime threshold
1836 * - average think time is more than threshold
1837 * - IO latency is largely below threshold
1842 time
= min_t(unsigned long, MAX_IDLE_TIME
, 4 * tg
->idletime_threshold
);
1843 ret
= tg
->latency_target
== DFL_LATENCY_TARGET
||
1844 tg
->idletime_threshold
== DFL_IDLE_THRESHOLD
||
1845 (ktime_get_ns() >> 10) - tg
->last_finish_time
> time
||
1846 tg
->avg_idletime
> tg
->idletime_threshold
||
1847 (tg
->latency_target
&& tg
->bio_cnt
&&
1848 tg
->bad_bio_cnt
* 5 < tg
->bio_cnt
);
1849 throtl_log(&tg
->service_queue
,
1850 "avg_idle=%ld, idle_threshold=%ld, bad_bio=%d, total_bio=%d, is_idle=%d, scale=%d",
1851 tg
->avg_idletime
, tg
->idletime_threshold
, tg
->bad_bio_cnt
,
1852 tg
->bio_cnt
, ret
, tg
->td
->scale
);
1856 static bool throtl_tg_can_upgrade(struct throtl_grp
*tg
)
1858 struct throtl_service_queue
*sq
= &tg
->service_queue
;
1859 bool read_limit
, write_limit
;
1862 * if cgroup reaches low limit (if low limit is 0, the cgroup always
1863 * reaches), it's ok to upgrade to next limit
1865 read_limit
= tg
->bps
[READ
][LIMIT_LOW
] || tg
->iops
[READ
][LIMIT_LOW
];
1866 write_limit
= tg
->bps
[WRITE
][LIMIT_LOW
] || tg
->iops
[WRITE
][LIMIT_LOW
];
1867 if (!read_limit
&& !write_limit
)
1869 if (read_limit
&& sq
->nr_queued
[READ
] &&
1870 (!write_limit
|| sq
->nr_queued
[WRITE
]))
1872 if (write_limit
&& sq
->nr_queued
[WRITE
] &&
1873 (!read_limit
|| sq
->nr_queued
[READ
]))
1876 if (time_after_eq(jiffies
,
1877 tg_last_low_overflow_time(tg
) + tg
->td
->throtl_slice
) &&
1878 throtl_tg_is_idle(tg
))
1883 static bool throtl_hierarchy_can_upgrade(struct throtl_grp
*tg
)
1886 if (throtl_tg_can_upgrade(tg
))
1888 tg
= sq_to_tg(tg
->service_queue
.parent_sq
);
1889 if (!tg
|| !tg_to_blkg(tg
)->parent
)
1895 static bool throtl_can_upgrade(struct throtl_data
*td
,
1896 struct throtl_grp
*this_tg
)
1898 struct cgroup_subsys_state
*pos_css
;
1899 struct blkcg_gq
*blkg
;
1901 if (td
->limit_index
!= LIMIT_LOW
)
1904 if (time_before(jiffies
, td
->low_downgrade_time
+ td
->throtl_slice
))
1908 blkg_for_each_descendant_post(blkg
, pos_css
, td
->queue
->root_blkg
) {
1909 struct throtl_grp
*tg
= blkg_to_tg(blkg
);
1913 if (!list_empty(&tg_to_blkg(tg
)->blkcg
->css
.children
))
1915 if (!throtl_hierarchy_can_upgrade(tg
)) {
1924 static void throtl_upgrade_check(struct throtl_grp
*tg
)
1926 unsigned long now
= jiffies
;
1928 if (tg
->td
->limit_index
!= LIMIT_LOW
)
1931 if (time_after(tg
->last_check_time
+ tg
->td
->throtl_slice
, now
))
1934 tg
->last_check_time
= now
;
1936 if (!time_after_eq(now
,
1937 __tg_last_low_overflow_time(tg
) + tg
->td
->throtl_slice
))
1940 if (throtl_can_upgrade(tg
->td
, NULL
))
1941 throtl_upgrade_state(tg
->td
);
1944 static void throtl_upgrade_state(struct throtl_data
*td
)
1946 struct cgroup_subsys_state
*pos_css
;
1947 struct blkcg_gq
*blkg
;
1949 throtl_log(&td
->service_queue
, "upgrade to max");
1950 td
->limit_index
= LIMIT_MAX
;
1951 td
->low_upgrade_time
= jiffies
;
1954 blkg_for_each_descendant_post(blkg
, pos_css
, td
->queue
->root_blkg
) {
1955 struct throtl_grp
*tg
= blkg_to_tg(blkg
);
1956 struct throtl_service_queue
*sq
= &tg
->service_queue
;
1958 tg
->disptime
= jiffies
- 1;
1959 throtl_select_dispatch(sq
);
1960 throtl_schedule_next_dispatch(sq
, true);
1963 throtl_select_dispatch(&td
->service_queue
);
1964 throtl_schedule_next_dispatch(&td
->service_queue
, true);
1965 queue_work(kthrotld_workqueue
, &td
->dispatch_work
);
1968 static void throtl_downgrade_state(struct throtl_data
*td
)
1972 throtl_log(&td
->service_queue
, "downgrade, scale %d", td
->scale
);
1974 td
->low_upgrade_time
= jiffies
- td
->scale
* td
->throtl_slice
;
1978 td
->limit_index
= LIMIT_LOW
;
1979 td
->low_downgrade_time
= jiffies
;
1982 static bool throtl_tg_can_downgrade(struct throtl_grp
*tg
)
1984 struct throtl_data
*td
= tg
->td
;
1985 unsigned long now
= jiffies
;
1988 * If cgroup is below low limit, consider downgrade and throttle other
1991 if (time_after_eq(now
, td
->low_upgrade_time
+ td
->throtl_slice
) &&
1992 time_after_eq(now
, tg_last_low_overflow_time(tg
) +
1993 td
->throtl_slice
) &&
1994 (!throtl_tg_is_idle(tg
) ||
1995 !list_empty(&tg_to_blkg(tg
)->blkcg
->css
.children
)))
2000 static bool throtl_hierarchy_can_downgrade(struct throtl_grp
*tg
)
2003 if (!throtl_tg_can_downgrade(tg
))
2005 tg
= sq_to_tg(tg
->service_queue
.parent_sq
);
2006 if (!tg
|| !tg_to_blkg(tg
)->parent
)
2012 static void throtl_downgrade_check(struct throtl_grp
*tg
)
2016 unsigned long elapsed_time
;
2017 unsigned long now
= jiffies
;
2019 if (tg
->td
->limit_index
!= LIMIT_MAX
||
2020 !tg
->td
->limit_valid
[LIMIT_LOW
])
2022 if (!list_empty(&tg_to_blkg(tg
)->blkcg
->css
.children
))
2024 if (time_after(tg
->last_check_time
+ tg
->td
->throtl_slice
, now
))
2027 elapsed_time
= now
- tg
->last_check_time
;
2028 tg
->last_check_time
= now
;
2030 if (time_before(now
, tg_last_low_overflow_time(tg
) +
2031 tg
->td
->throtl_slice
))
2034 if (tg
->bps
[READ
][LIMIT_LOW
]) {
2035 bps
= tg
->last_bytes_disp
[READ
] * HZ
;
2036 do_div(bps
, elapsed_time
);
2037 if (bps
>= tg
->bps
[READ
][LIMIT_LOW
])
2038 tg
->last_low_overflow_time
[READ
] = now
;
2041 if (tg
->bps
[WRITE
][LIMIT_LOW
]) {
2042 bps
= tg
->last_bytes_disp
[WRITE
] * HZ
;
2043 do_div(bps
, elapsed_time
);
2044 if (bps
>= tg
->bps
[WRITE
][LIMIT_LOW
])
2045 tg
->last_low_overflow_time
[WRITE
] = now
;
2048 if (tg
->iops
[READ
][LIMIT_LOW
]) {
2049 iops
= tg
->last_io_disp
[READ
] * HZ
/ elapsed_time
;
2050 if (iops
>= tg
->iops
[READ
][LIMIT_LOW
])
2051 tg
->last_low_overflow_time
[READ
] = now
;
2054 if (tg
->iops
[WRITE
][LIMIT_LOW
]) {
2055 iops
= tg
->last_io_disp
[WRITE
] * HZ
/ elapsed_time
;
2056 if (iops
>= tg
->iops
[WRITE
][LIMIT_LOW
])
2057 tg
->last_low_overflow_time
[WRITE
] = now
;
2061 * If cgroup is below low limit, consider downgrade and throttle other
2064 if (throtl_hierarchy_can_downgrade(tg
))
2065 throtl_downgrade_state(tg
->td
);
2067 tg
->last_bytes_disp
[READ
] = 0;
2068 tg
->last_bytes_disp
[WRITE
] = 0;
2069 tg
->last_io_disp
[READ
] = 0;
2070 tg
->last_io_disp
[WRITE
] = 0;
2073 static void blk_throtl_update_idletime(struct throtl_grp
*tg
)
2076 unsigned long last_finish_time
= tg
->last_finish_time
;
2078 if (last_finish_time
== 0)
2081 now
= ktime_get_ns() >> 10;
2082 if (now
<= last_finish_time
||
2083 last_finish_time
== tg
->checked_last_finish_time
)
2086 tg
->avg_idletime
= (tg
->avg_idletime
* 7 + now
- last_finish_time
) >> 3;
2087 tg
->checked_last_finish_time
= last_finish_time
;
2090 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2091 static void throtl_update_latency_buckets(struct throtl_data
*td
)
2093 struct avg_latency_bucket avg_latency
[2][LATENCY_BUCKET_SIZE
];
2095 unsigned long last_latency
[2] = { 0 };
2096 unsigned long latency
[2];
2098 if (!blk_queue_nonrot(td
->queue
) || !td
->limit_valid
[LIMIT_LOW
])
2100 if (time_before(jiffies
, td
->last_calculate_time
+ HZ
))
2102 td
->last_calculate_time
= jiffies
;
2104 memset(avg_latency
, 0, sizeof(avg_latency
));
2105 for (rw
= READ
; rw
<= WRITE
; rw
++) {
2106 for (i
= 0; i
< LATENCY_BUCKET_SIZE
; i
++) {
2107 struct latency_bucket
*tmp
= &td
->tmp_buckets
[rw
][i
];
2109 for_each_possible_cpu(cpu
) {
2110 struct latency_bucket
*bucket
;
2112 /* this isn't race free, but ok in practice */
2113 bucket
= per_cpu_ptr(td
->latency_buckets
[rw
],
2115 tmp
->total_latency
+= bucket
[i
].total_latency
;
2116 tmp
->samples
+= bucket
[i
].samples
;
2117 bucket
[i
].total_latency
= 0;
2118 bucket
[i
].samples
= 0;
2121 if (tmp
->samples
>= 32) {
2122 int samples
= tmp
->samples
;
2124 latency
[rw
] = tmp
->total_latency
;
2126 tmp
->total_latency
= 0;
2128 latency
[rw
] /= samples
;
2129 if (latency
[rw
] == 0)
2131 avg_latency
[rw
][i
].latency
= latency
[rw
];
2136 for (rw
= READ
; rw
<= WRITE
; rw
++) {
2137 for (i
= 0; i
< LATENCY_BUCKET_SIZE
; i
++) {
2138 if (!avg_latency
[rw
][i
].latency
) {
2139 if (td
->avg_buckets
[rw
][i
].latency
< last_latency
[rw
])
2140 td
->avg_buckets
[rw
][i
].latency
=
2145 if (!td
->avg_buckets
[rw
][i
].valid
)
2146 latency
[rw
] = avg_latency
[rw
][i
].latency
;
2148 latency
[rw
] = (td
->avg_buckets
[rw
][i
].latency
* 7 +
2149 avg_latency
[rw
][i
].latency
) >> 3;
2151 td
->avg_buckets
[rw
][i
].latency
= max(latency
[rw
],
2153 td
->avg_buckets
[rw
][i
].valid
= true;
2154 last_latency
[rw
] = td
->avg_buckets
[rw
][i
].latency
;
2158 for (i
= 0; i
< LATENCY_BUCKET_SIZE
; i
++)
2159 throtl_log(&td
->service_queue
,
2160 "Latency bucket %d: read latency=%ld, read valid=%d, "
2161 "write latency=%ld, write valid=%d", i
,
2162 td
->avg_buckets
[READ
][i
].latency
,
2163 td
->avg_buckets
[READ
][i
].valid
,
2164 td
->avg_buckets
[WRITE
][i
].latency
,
2165 td
->avg_buckets
[WRITE
][i
].valid
);
2168 static inline void throtl_update_latency_buckets(struct throtl_data
*td
)
2173 bool blk_throtl_bio(struct bio
*bio
)
2175 struct request_queue
*q
= bio
->bi_disk
->queue
;
2176 struct blkcg_gq
*blkg
= bio
->bi_blkg
;
2177 struct throtl_qnode
*qn
= NULL
;
2178 struct throtl_grp
*tg
= blkg_to_tg(blkg
);
2179 struct throtl_service_queue
*sq
;
2180 bool rw
= bio_data_dir(bio
);
2181 bool throttled
= false;
2182 struct throtl_data
*td
= tg
->td
;
2186 /* see throtl_charge_bio() */
2187 if (bio_flagged(bio
, BIO_THROTTLED
))
2190 if (!cgroup_subsys_on_dfl(io_cgrp_subsys
)) {
2191 blkg_rwstat_add(&tg
->stat_bytes
, bio
->bi_opf
,
2192 bio
->bi_iter
.bi_size
);
2193 blkg_rwstat_add(&tg
->stat_ios
, bio
->bi_opf
, 1);
2196 if (!tg
->has_rules
[rw
])
2199 spin_lock_irq(&q
->queue_lock
);
2201 throtl_update_latency_buckets(td
);
2203 blk_throtl_update_idletime(tg
);
2205 sq
= &tg
->service_queue
;
2209 if (tg
->last_low_overflow_time
[rw
] == 0)
2210 tg
->last_low_overflow_time
[rw
] = jiffies
;
2211 throtl_downgrade_check(tg
);
2212 throtl_upgrade_check(tg
);
2213 /* throtl is FIFO - if bios are already queued, should queue */
2214 if (sq
->nr_queued
[rw
])
2217 /* if above limits, break to queue */
2218 if (!tg_may_dispatch(tg
, bio
, NULL
)) {
2219 tg
->last_low_overflow_time
[rw
] = jiffies
;
2220 if (throtl_can_upgrade(td
, tg
)) {
2221 throtl_upgrade_state(td
);
2227 /* within limits, let's charge and dispatch directly */
2228 throtl_charge_bio(tg
, bio
);
2231 * We need to trim slice even when bios are not being queued
2232 * otherwise it might happen that a bio is not queued for
2233 * a long time and slice keeps on extending and trim is not
2234 * called for a long time. Now if limits are reduced suddenly
2235 * we take into account all the IO dispatched so far at new
2236 * low rate and * newly queued IO gets a really long dispatch
2239 * So keep on trimming slice even if bio is not queued.
2241 throtl_trim_slice(tg
, rw
);
2244 * @bio passed through this layer without being throttled.
2245 * Climb up the ladder. If we're already at the top, it
2246 * can be executed directly.
2248 qn
= &tg
->qnode_on_parent
[rw
];
2255 /* out-of-limit, queue to @tg */
2256 throtl_log(sq
, "[%c] bio. bdisp=%llu sz=%u bps=%llu iodisp=%u iops=%u queued=%d/%d",
2257 rw
== READ
? 'R' : 'W',
2258 tg
->bytes_disp
[rw
], bio
->bi_iter
.bi_size
,
2259 tg_bps_limit(tg
, rw
),
2260 tg
->io_disp
[rw
], tg_iops_limit(tg
, rw
),
2261 sq
->nr_queued
[READ
], sq
->nr_queued
[WRITE
]);
2263 tg
->last_low_overflow_time
[rw
] = jiffies
;
2265 td
->nr_queued
[rw
]++;
2266 throtl_add_bio_tg(bio
, qn
, tg
);
2270 * Update @tg's dispatch time and force schedule dispatch if @tg
2271 * was empty before @bio. The forced scheduling isn't likely to
2272 * cause undue delay as @bio is likely to be dispatched directly if
2273 * its @tg's disptime is not in the future.
2275 if (tg
->flags
& THROTL_TG_WAS_EMPTY
) {
2276 tg_update_disptime(tg
);
2277 throtl_schedule_next_dispatch(tg
->service_queue
.parent_sq
, true);
2281 spin_unlock_irq(&q
->queue_lock
);
2283 bio_set_flag(bio
, BIO_THROTTLED
);
2285 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2286 if (throttled
|| !td
->track_bio_latency
)
2287 bio
->bi_issue
.value
|= BIO_ISSUE_THROTL_SKIP_LATENCY
;
2293 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2294 static void throtl_track_latency(struct throtl_data
*td
, sector_t size
,
2295 int op
, unsigned long time
)
2297 struct latency_bucket
*latency
;
2300 if (!td
|| td
->limit_index
!= LIMIT_LOW
||
2301 !(op
== REQ_OP_READ
|| op
== REQ_OP_WRITE
) ||
2302 !blk_queue_nonrot(td
->queue
))
2305 index
= request_bucket_index(size
);
2307 latency
= get_cpu_ptr(td
->latency_buckets
[op
]);
2308 latency
[index
].total_latency
+= time
;
2309 latency
[index
].samples
++;
2310 put_cpu_ptr(td
->latency_buckets
[op
]);
2313 void blk_throtl_stat_add(struct request
*rq
, u64 time_ns
)
2315 struct request_queue
*q
= rq
->q
;
2316 struct throtl_data
*td
= q
->td
;
2318 throtl_track_latency(td
, blk_rq_stats_sectors(rq
), req_op(rq
),
2322 void blk_throtl_bio_endio(struct bio
*bio
)
2324 struct blkcg_gq
*blkg
;
2325 struct throtl_grp
*tg
;
2327 unsigned long finish_time
;
2328 unsigned long start_time
;
2330 int rw
= bio_data_dir(bio
);
2332 blkg
= bio
->bi_blkg
;
2335 tg
= blkg_to_tg(blkg
);
2336 if (!tg
->td
->limit_valid
[LIMIT_LOW
])
2339 finish_time_ns
= ktime_get_ns();
2340 tg
->last_finish_time
= finish_time_ns
>> 10;
2342 start_time
= bio_issue_time(&bio
->bi_issue
) >> 10;
2343 finish_time
= __bio_issue_time(finish_time_ns
) >> 10;
2344 if (!start_time
|| finish_time
<= start_time
)
2347 lat
= finish_time
- start_time
;
2348 /* this is only for bio based driver */
2349 if (!(bio
->bi_issue
.value
& BIO_ISSUE_THROTL_SKIP_LATENCY
))
2350 throtl_track_latency(tg
->td
, bio_issue_size(&bio
->bi_issue
),
2353 if (tg
->latency_target
&& lat
>= tg
->td
->filtered_latency
) {
2355 unsigned int threshold
;
2357 bucket
= request_bucket_index(bio_issue_size(&bio
->bi_issue
));
2358 threshold
= tg
->td
->avg_buckets
[rw
][bucket
].latency
+
2360 if (lat
> threshold
)
2363 * Not race free, could get wrong count, which means cgroups
2369 if (time_after(jiffies
, tg
->bio_cnt_reset_time
) || tg
->bio_cnt
> 1024) {
2370 tg
->bio_cnt_reset_time
= tg
->td
->throtl_slice
+ jiffies
;
2372 tg
->bad_bio_cnt
/= 2;
2377 int blk_throtl_init(struct request_queue
*q
)
2379 struct throtl_data
*td
;
2382 td
= kzalloc_node(sizeof(*td
), GFP_KERNEL
, q
->node
);
2385 td
->latency_buckets
[READ
] = __alloc_percpu(sizeof(struct latency_bucket
) *
2386 LATENCY_BUCKET_SIZE
, __alignof__(u64
));
2387 if (!td
->latency_buckets
[READ
]) {
2391 td
->latency_buckets
[WRITE
] = __alloc_percpu(sizeof(struct latency_bucket
) *
2392 LATENCY_BUCKET_SIZE
, __alignof__(u64
));
2393 if (!td
->latency_buckets
[WRITE
]) {
2394 free_percpu(td
->latency_buckets
[READ
]);
2399 INIT_WORK(&td
->dispatch_work
, blk_throtl_dispatch_work_fn
);
2400 throtl_service_queue_init(&td
->service_queue
);
2405 td
->limit_valid
[LIMIT_MAX
] = true;
2406 td
->limit_index
= LIMIT_MAX
;
2407 td
->low_upgrade_time
= jiffies
;
2408 td
->low_downgrade_time
= jiffies
;
2410 /* activate policy */
2411 ret
= blkcg_activate_policy(q
, &blkcg_policy_throtl
);
2413 free_percpu(td
->latency_buckets
[READ
]);
2414 free_percpu(td
->latency_buckets
[WRITE
]);
2420 void blk_throtl_exit(struct request_queue
*q
)
2423 throtl_shutdown_wq(q
);
2424 blkcg_deactivate_policy(q
, &blkcg_policy_throtl
);
2425 free_percpu(q
->td
->latency_buckets
[READ
]);
2426 free_percpu(q
->td
->latency_buckets
[WRITE
]);
2430 void blk_throtl_register_queue(struct request_queue
*q
)
2432 struct throtl_data
*td
;
2438 if (blk_queue_nonrot(q
)) {
2439 td
->throtl_slice
= DFL_THROTL_SLICE_SSD
;
2440 td
->filtered_latency
= LATENCY_FILTERED_SSD
;
2442 td
->throtl_slice
= DFL_THROTL_SLICE_HD
;
2443 td
->filtered_latency
= LATENCY_FILTERED_HD
;
2444 for (i
= 0; i
< LATENCY_BUCKET_SIZE
; i
++) {
2445 td
->avg_buckets
[READ
][i
].latency
= DFL_HD_BASELINE_LATENCY
;
2446 td
->avg_buckets
[WRITE
][i
].latency
= DFL_HD_BASELINE_LATENCY
;
2449 #ifndef CONFIG_BLK_DEV_THROTTLING_LOW
2450 /* if no low limit, use previous default */
2451 td
->throtl_slice
= DFL_THROTL_SLICE_HD
;
2454 td
->track_bio_latency
= !queue_is_mq(q
);
2455 if (!td
->track_bio_latency
)
2456 blk_stat_enable_accounting(q
);
2459 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2460 ssize_t
blk_throtl_sample_time_show(struct request_queue
*q
, char *page
)
2464 return sprintf(page
, "%u\n", jiffies_to_msecs(q
->td
->throtl_slice
));
2467 ssize_t
blk_throtl_sample_time_store(struct request_queue
*q
,
2468 const char *page
, size_t count
)
2475 if (kstrtoul(page
, 10, &v
))
2477 t
= msecs_to_jiffies(v
);
2478 if (t
== 0 || t
> MAX_THROTL_SLICE
)
2480 q
->td
->throtl_slice
= t
;
2485 static int __init
throtl_init(void)
2487 kthrotld_workqueue
= alloc_workqueue("kthrotld", WQ_MEM_RECLAIM
, 0);
2488 if (!kthrotld_workqueue
)
2489 panic("Failed to create kthrotld\n");
2491 return blkcg_policy_register(&blkcg_policy_throtl
);
2494 module_init(throtl_init
);