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>
16 /* Max dispatch from a group in 1 round */
17 static int throtl_grp_quantum
= 8;
19 /* Total max dispatch from all groups in one round */
20 static int throtl_quantum
= 32;
22 /* Throttling is performed over a slice and after that slice is renewed */
23 #define DFL_THROTL_SLICE_HD (HZ / 10)
24 #define DFL_THROTL_SLICE_SSD (HZ / 50)
25 #define MAX_THROTL_SLICE (HZ)
26 #define MAX_IDLE_TIME (5L * 1000 * 1000) /* 5 s */
27 #define MIN_THROTL_BPS (320 * 1024)
28 #define MIN_THROTL_IOPS (10)
29 #define DFL_LATENCY_TARGET (-1L)
30 #define DFL_IDLE_THRESHOLD (0)
31 #define DFL_HD_BASELINE_LATENCY (4000L) /* 4ms */
32 #define LATENCY_FILTERED_SSD (0)
34 * For HD, very small latency comes from sequential IO. Such IO is helpless to
35 * help determine if its IO is impacted by others, hence we ignore the IO
37 #define LATENCY_FILTERED_HD (1000L) /* 1ms */
39 #define SKIP_LATENCY (((u64)1) << BLK_STAT_RES_SHIFT)
41 static struct blkcg_policy blkcg_policy_throtl
;
43 /* A workqueue to queue throttle related work */
44 static struct workqueue_struct
*kthrotld_workqueue
;
47 * To implement hierarchical throttling, throtl_grps form a tree and bios
48 * are dispatched upwards level by level until they reach the top and get
49 * issued. When dispatching bios from the children and local group at each
50 * level, if the bios are dispatched into a single bio_list, there's a risk
51 * of a local or child group which can queue many bios at once filling up
52 * the list starving others.
54 * To avoid such starvation, dispatched bios are queued separately
55 * according to where they came from. When they are again dispatched to
56 * the parent, they're popped in round-robin order so that no single source
57 * hogs the dispatch window.
59 * throtl_qnode is used to keep the queued bios separated by their sources.
60 * Bios are queued to throtl_qnode which in turn is queued to
61 * throtl_service_queue and then dispatched in round-robin order.
63 * It's also used to track the reference counts on blkg's. A qnode always
64 * belongs to a throtl_grp and gets queued on itself or the parent, so
65 * incrementing the reference of the associated throtl_grp when a qnode is
66 * queued and decrementing when dequeued is enough to keep the whole blkg
67 * tree pinned while bios are in flight.
70 struct list_head node
; /* service_queue->queued[] */
71 struct bio_list bios
; /* queued bios */
72 struct throtl_grp
*tg
; /* tg this qnode belongs to */
75 struct throtl_service_queue
{
76 struct throtl_service_queue
*parent_sq
; /* the parent service_queue */
79 * Bios queued directly to this service_queue or dispatched from
80 * children throtl_grp's.
82 struct list_head queued
[2]; /* throtl_qnode [READ/WRITE] */
83 unsigned int nr_queued
[2]; /* number of queued bios */
86 * RB tree of active children throtl_grp's, which are sorted by
89 struct rb_root pending_tree
; /* RB tree of active tgs */
90 struct rb_node
*first_pending
; /* first node in the tree */
91 unsigned int nr_pending
; /* # queued in the tree */
92 unsigned long first_pending_disptime
; /* disptime of the first tg */
93 struct timer_list pending_timer
; /* fires on first_pending_disptime */
97 THROTL_TG_PENDING
= 1 << 0, /* on parent's pending tree */
98 THROTL_TG_WAS_EMPTY
= 1 << 1, /* bio_lists[] became non-empty */
101 #define rb_entry_tg(node) rb_entry((node), struct throtl_grp, rb_node)
110 /* must be the first member */
111 struct blkg_policy_data pd
;
113 /* active throtl group service_queue member */
114 struct rb_node rb_node
;
116 /* throtl_data this group belongs to */
117 struct throtl_data
*td
;
119 /* this group's service queue */
120 struct throtl_service_queue service_queue
;
123 * qnode_on_self is used when bios are directly queued to this
124 * throtl_grp so that local bios compete fairly with bios
125 * dispatched from children. qnode_on_parent is used when bios are
126 * dispatched from this throtl_grp into its parent and will compete
127 * with the sibling qnode_on_parents and the parent's
130 struct throtl_qnode qnode_on_self
[2];
131 struct throtl_qnode qnode_on_parent
[2];
134 * Dispatch time in jiffies. This is the estimated time when group
135 * will unthrottle and is ready to dispatch more bio. It is used as
136 * key to sort active groups in service tree.
138 unsigned long disptime
;
142 /* are there any throtl rules between this group and td? */
145 /* internally used bytes per second rate limits */
146 uint64_t bps
[2][LIMIT_CNT
];
147 /* user configured bps limits */
148 uint64_t bps_conf
[2][LIMIT_CNT
];
150 /* internally used IOPS limits */
151 unsigned int iops
[2][LIMIT_CNT
];
152 /* user configured IOPS limits */
153 unsigned int iops_conf
[2][LIMIT_CNT
];
155 /* Number of bytes disptached in current slice */
156 uint64_t bytes_disp
[2];
157 /* Number of bio's dispatched in current slice */
158 unsigned int io_disp
[2];
160 unsigned long last_low_overflow_time
[2];
162 uint64_t last_bytes_disp
[2];
163 unsigned int last_io_disp
[2];
165 unsigned long last_check_time
;
167 unsigned long latency_target
; /* us */
168 unsigned long latency_target_conf
; /* us */
169 /* When did we start a new slice */
170 unsigned long slice_start
[2];
171 unsigned long slice_end
[2];
173 unsigned long last_finish_time
; /* ns / 1024 */
174 unsigned long checked_last_finish_time
; /* ns / 1024 */
175 unsigned long avg_idletime
; /* ns / 1024 */
176 unsigned long idletime_threshold
; /* us */
177 unsigned long idletime_threshold_conf
; /* us */
179 unsigned int bio_cnt
; /* total bios */
180 unsigned int bad_bio_cnt
; /* bios exceeding latency threshold */
181 unsigned long bio_cnt_reset_time
;
184 /* We measure latency for request size from <= 4k to >= 1M */
185 #define LATENCY_BUCKET_SIZE 9
187 struct latency_bucket
{
188 unsigned long total_latency
; /* ns / 1024 */
192 struct avg_latency_bucket
{
193 unsigned long latency
; /* ns / 1024 */
199 /* service tree for active throtl groups */
200 struct throtl_service_queue service_queue
;
202 struct request_queue
*queue
;
204 /* Total Number of queued bios on READ and WRITE lists */
205 unsigned int nr_queued
[2];
207 unsigned int throtl_slice
;
209 /* Work for dispatching throttled bios */
210 struct work_struct dispatch_work
;
211 unsigned int limit_index
;
212 bool limit_valid
[LIMIT_CNT
];
214 unsigned long low_upgrade_time
;
215 unsigned long low_downgrade_time
;
219 struct latency_bucket tmp_buckets
[LATENCY_BUCKET_SIZE
];
220 struct avg_latency_bucket avg_buckets
[LATENCY_BUCKET_SIZE
];
221 struct latency_bucket __percpu
*latency_buckets
;
222 unsigned long last_calculate_time
;
223 unsigned long filtered_latency
;
225 bool track_bio_latency
;
228 static void throtl_pending_timer_fn(struct timer_list
*t
);
230 static inline struct throtl_grp
*pd_to_tg(struct blkg_policy_data
*pd
)
232 return pd
? container_of(pd
, struct throtl_grp
, pd
) : NULL
;
235 static inline struct throtl_grp
*blkg_to_tg(struct blkcg_gq
*blkg
)
237 return pd_to_tg(blkg_to_pd(blkg
, &blkcg_policy_throtl
));
240 static inline struct blkcg_gq
*tg_to_blkg(struct throtl_grp
*tg
)
242 return pd_to_blkg(&tg
->pd
);
246 * sq_to_tg - return the throl_grp the specified service queue belongs to
247 * @sq: the throtl_service_queue of interest
249 * Return the throtl_grp @sq belongs to. If @sq is the top-level one
250 * embedded in throtl_data, %NULL is returned.
252 static struct throtl_grp
*sq_to_tg(struct throtl_service_queue
*sq
)
254 if (sq
&& sq
->parent_sq
)
255 return container_of(sq
, struct throtl_grp
, service_queue
);
261 * sq_to_td - return throtl_data the specified service queue belongs to
262 * @sq: the throtl_service_queue of interest
264 * A service_queue can be embedded in either a throtl_grp or throtl_data.
265 * Determine the associated throtl_data accordingly and return it.
267 static struct throtl_data
*sq_to_td(struct throtl_service_queue
*sq
)
269 struct throtl_grp
*tg
= sq_to_tg(sq
);
274 return container_of(sq
, struct throtl_data
, service_queue
);
278 * cgroup's limit in LIMIT_MAX is scaled if low limit is set. This scale is to
279 * make the IO dispatch more smooth.
280 * Scale up: linearly scale up according to lapsed time since upgrade. For
281 * every throtl_slice, the limit scales up 1/2 .low limit till the
282 * limit hits .max limit
283 * Scale down: exponentially scale down if a cgroup doesn't hit its .low limit
285 static uint64_t throtl_adjusted_limit(uint64_t low
, struct throtl_data
*td
)
287 /* arbitrary value to avoid too big scale */
288 if (td
->scale
< 4096 && time_after_eq(jiffies
,
289 td
->low_upgrade_time
+ td
->scale
* td
->throtl_slice
))
290 td
->scale
= (jiffies
- td
->low_upgrade_time
) / td
->throtl_slice
;
292 return low
+ (low
>> 1) * td
->scale
;
295 static uint64_t tg_bps_limit(struct throtl_grp
*tg
, int rw
)
297 struct blkcg_gq
*blkg
= tg_to_blkg(tg
);
298 struct throtl_data
*td
;
301 if (cgroup_subsys_on_dfl(io_cgrp_subsys
) && !blkg
->parent
)
305 ret
= tg
->bps
[rw
][td
->limit_index
];
306 if (ret
== 0 && td
->limit_index
== LIMIT_LOW
) {
307 /* intermediate node or iops isn't 0 */
308 if (!list_empty(&blkg
->blkcg
->css
.children
) ||
309 tg
->iops
[rw
][td
->limit_index
])
312 return MIN_THROTL_BPS
;
315 if (td
->limit_index
== LIMIT_MAX
&& tg
->bps
[rw
][LIMIT_LOW
] &&
316 tg
->bps
[rw
][LIMIT_LOW
] != tg
->bps
[rw
][LIMIT_MAX
]) {
319 adjusted
= throtl_adjusted_limit(tg
->bps
[rw
][LIMIT_LOW
], td
);
320 ret
= min(tg
->bps
[rw
][LIMIT_MAX
], adjusted
);
325 static unsigned int tg_iops_limit(struct throtl_grp
*tg
, int rw
)
327 struct blkcg_gq
*blkg
= tg_to_blkg(tg
);
328 struct throtl_data
*td
;
331 if (cgroup_subsys_on_dfl(io_cgrp_subsys
) && !blkg
->parent
)
335 ret
= tg
->iops
[rw
][td
->limit_index
];
336 if (ret
== 0 && tg
->td
->limit_index
== LIMIT_LOW
) {
337 /* intermediate node or bps isn't 0 */
338 if (!list_empty(&blkg
->blkcg
->css
.children
) ||
339 tg
->bps
[rw
][td
->limit_index
])
342 return MIN_THROTL_IOPS
;
345 if (td
->limit_index
== LIMIT_MAX
&& tg
->iops
[rw
][LIMIT_LOW
] &&
346 tg
->iops
[rw
][LIMIT_LOW
] != tg
->iops
[rw
][LIMIT_MAX
]) {
349 adjusted
= throtl_adjusted_limit(tg
->iops
[rw
][LIMIT_LOW
], td
);
350 if (adjusted
> UINT_MAX
)
352 ret
= min_t(unsigned int, tg
->iops
[rw
][LIMIT_MAX
], adjusted
);
357 #define request_bucket_index(sectors) \
358 clamp_t(int, order_base_2(sectors) - 3, 0, LATENCY_BUCKET_SIZE - 1)
361 * throtl_log - log debug message via blktrace
362 * @sq: the service_queue being reported
363 * @fmt: printf format string
366 * The messages are prefixed with "throtl BLKG_NAME" if @sq belongs to a
367 * throtl_grp; otherwise, just "throtl".
369 #define throtl_log(sq, fmt, args...) do { \
370 struct throtl_grp *__tg = sq_to_tg((sq)); \
371 struct throtl_data *__td = sq_to_td((sq)); \
374 if (likely(!blk_trace_note_message_enabled(__td->queue))) \
377 blk_add_cgroup_trace_msg(__td->queue, \
378 tg_to_blkg(__tg)->blkcg, "throtl " fmt, ##args);\
380 blk_add_trace_msg(__td->queue, "throtl " fmt, ##args); \
384 static inline unsigned int throtl_bio_data_size(struct bio
*bio
)
386 /* assume it's one sector */
387 if (unlikely(bio_op(bio
) == REQ_OP_DISCARD
))
389 return bio
->bi_iter
.bi_size
;
392 static void throtl_qnode_init(struct throtl_qnode
*qn
, struct throtl_grp
*tg
)
394 INIT_LIST_HEAD(&qn
->node
);
395 bio_list_init(&qn
->bios
);
400 * throtl_qnode_add_bio - add a bio to a throtl_qnode and activate it
401 * @bio: bio being added
402 * @qn: qnode to add bio to
403 * @queued: the service_queue->queued[] list @qn belongs to
405 * Add @bio to @qn and put @qn on @queued if it's not already on.
406 * @qn->tg's reference count is bumped when @qn is activated. See the
407 * comment on top of throtl_qnode definition for details.
409 static void throtl_qnode_add_bio(struct bio
*bio
, struct throtl_qnode
*qn
,
410 struct list_head
*queued
)
412 bio_list_add(&qn
->bios
, bio
);
413 if (list_empty(&qn
->node
)) {
414 list_add_tail(&qn
->node
, queued
);
415 blkg_get(tg_to_blkg(qn
->tg
));
420 * throtl_peek_queued - peek the first bio on a qnode list
421 * @queued: the qnode list to peek
423 static struct bio
*throtl_peek_queued(struct list_head
*queued
)
425 struct throtl_qnode
*qn
= list_first_entry(queued
, struct throtl_qnode
, node
);
428 if (list_empty(queued
))
431 bio
= bio_list_peek(&qn
->bios
);
437 * throtl_pop_queued - pop the first bio form a qnode list
438 * @queued: the qnode list to pop a bio from
439 * @tg_to_put: optional out argument for throtl_grp to put
441 * Pop the first bio from the qnode list @queued. After popping, the first
442 * qnode is removed from @queued if empty or moved to the end of @queued so
443 * that the popping order is round-robin.
445 * When the first qnode is removed, its associated throtl_grp should be put
446 * too. If @tg_to_put is NULL, this function automatically puts it;
447 * otherwise, *@tg_to_put is set to the throtl_grp to put and the caller is
448 * responsible for putting it.
450 static struct bio
*throtl_pop_queued(struct list_head
*queued
,
451 struct throtl_grp
**tg_to_put
)
453 struct throtl_qnode
*qn
= list_first_entry(queued
, struct throtl_qnode
, node
);
456 if (list_empty(queued
))
459 bio
= bio_list_pop(&qn
->bios
);
462 if (bio_list_empty(&qn
->bios
)) {
463 list_del_init(&qn
->node
);
467 blkg_put(tg_to_blkg(qn
->tg
));
469 list_move_tail(&qn
->node
, queued
);
475 /* init a service_queue, assumes the caller zeroed it */
476 static void throtl_service_queue_init(struct throtl_service_queue
*sq
)
478 INIT_LIST_HEAD(&sq
->queued
[0]);
479 INIT_LIST_HEAD(&sq
->queued
[1]);
480 sq
->pending_tree
= RB_ROOT
;
481 timer_setup(&sq
->pending_timer
, throtl_pending_timer_fn
, 0);
484 static struct blkg_policy_data
*throtl_pd_alloc(gfp_t gfp
, int node
)
486 struct throtl_grp
*tg
;
489 tg
= kzalloc_node(sizeof(*tg
), gfp
, node
);
493 throtl_service_queue_init(&tg
->service_queue
);
495 for (rw
= READ
; rw
<= WRITE
; rw
++) {
496 throtl_qnode_init(&tg
->qnode_on_self
[rw
], tg
);
497 throtl_qnode_init(&tg
->qnode_on_parent
[rw
], tg
);
500 RB_CLEAR_NODE(&tg
->rb_node
);
501 tg
->bps
[READ
][LIMIT_MAX
] = U64_MAX
;
502 tg
->bps
[WRITE
][LIMIT_MAX
] = U64_MAX
;
503 tg
->iops
[READ
][LIMIT_MAX
] = UINT_MAX
;
504 tg
->iops
[WRITE
][LIMIT_MAX
] = UINT_MAX
;
505 tg
->bps_conf
[READ
][LIMIT_MAX
] = U64_MAX
;
506 tg
->bps_conf
[WRITE
][LIMIT_MAX
] = U64_MAX
;
507 tg
->iops_conf
[READ
][LIMIT_MAX
] = UINT_MAX
;
508 tg
->iops_conf
[WRITE
][LIMIT_MAX
] = UINT_MAX
;
509 /* LIMIT_LOW will have default value 0 */
511 tg
->latency_target
= DFL_LATENCY_TARGET
;
512 tg
->latency_target_conf
= DFL_LATENCY_TARGET
;
513 tg
->idletime_threshold
= DFL_IDLE_THRESHOLD
;
514 tg
->idletime_threshold_conf
= DFL_IDLE_THRESHOLD
;
519 static void throtl_pd_init(struct blkg_policy_data
*pd
)
521 struct throtl_grp
*tg
= pd_to_tg(pd
);
522 struct blkcg_gq
*blkg
= tg_to_blkg(tg
);
523 struct throtl_data
*td
= blkg
->q
->td
;
524 struct throtl_service_queue
*sq
= &tg
->service_queue
;
527 * If on the default hierarchy, we switch to properly hierarchical
528 * behavior where limits on a given throtl_grp are applied to the
529 * whole subtree rather than just the group itself. e.g. If 16M
530 * read_bps limit is set on the root group, the whole system can't
531 * exceed 16M for the device.
533 * If not on the default hierarchy, the broken flat hierarchy
534 * behavior is retained where all throtl_grps are treated as if
535 * they're all separate root groups right below throtl_data.
536 * Limits of a group don't interact with limits of other groups
537 * regardless of the position of the group in the hierarchy.
539 sq
->parent_sq
= &td
->service_queue
;
540 if (cgroup_subsys_on_dfl(io_cgrp_subsys
) && blkg
->parent
)
541 sq
->parent_sq
= &blkg_to_tg(blkg
->parent
)->service_queue
;
546 * Set has_rules[] if @tg or any of its parents have limits configured.
547 * This doesn't require walking up to the top of the hierarchy as the
548 * parent's has_rules[] is guaranteed to be correct.
550 static void tg_update_has_rules(struct throtl_grp
*tg
)
552 struct throtl_grp
*parent_tg
= sq_to_tg(tg
->service_queue
.parent_sq
);
553 struct throtl_data
*td
= tg
->td
;
556 for (rw
= READ
; rw
<= WRITE
; rw
++)
557 tg
->has_rules
[rw
] = (parent_tg
&& parent_tg
->has_rules
[rw
]) ||
558 (td
->limit_valid
[td
->limit_index
] &&
559 (tg_bps_limit(tg
, rw
) != U64_MAX
||
560 tg_iops_limit(tg
, rw
) != UINT_MAX
));
563 static void throtl_pd_online(struct blkg_policy_data
*pd
)
565 struct throtl_grp
*tg
= pd_to_tg(pd
);
567 * We don't want new groups to escape the limits of its ancestors.
568 * Update has_rules[] after a new group is brought online.
570 tg_update_has_rules(tg
);
573 static void blk_throtl_update_limit_valid(struct throtl_data
*td
)
575 struct cgroup_subsys_state
*pos_css
;
576 struct blkcg_gq
*blkg
;
577 bool low_valid
= false;
580 blkg_for_each_descendant_post(blkg
, pos_css
, td
->queue
->root_blkg
) {
581 struct throtl_grp
*tg
= blkg_to_tg(blkg
);
583 if (tg
->bps
[READ
][LIMIT_LOW
] || tg
->bps
[WRITE
][LIMIT_LOW
] ||
584 tg
->iops
[READ
][LIMIT_LOW
] || tg
->iops
[WRITE
][LIMIT_LOW
])
589 td
->limit_valid
[LIMIT_LOW
] = low_valid
;
592 static void throtl_upgrade_state(struct throtl_data
*td
);
593 static void throtl_pd_offline(struct blkg_policy_data
*pd
)
595 struct throtl_grp
*tg
= pd_to_tg(pd
);
597 tg
->bps
[READ
][LIMIT_LOW
] = 0;
598 tg
->bps
[WRITE
][LIMIT_LOW
] = 0;
599 tg
->iops
[READ
][LIMIT_LOW
] = 0;
600 tg
->iops
[WRITE
][LIMIT_LOW
] = 0;
602 blk_throtl_update_limit_valid(tg
->td
);
604 if (!tg
->td
->limit_valid
[tg
->td
->limit_index
])
605 throtl_upgrade_state(tg
->td
);
608 static void throtl_pd_free(struct blkg_policy_data
*pd
)
610 struct throtl_grp
*tg
= pd_to_tg(pd
);
612 del_timer_sync(&tg
->service_queue
.pending_timer
);
616 static struct throtl_grp
*
617 throtl_rb_first(struct throtl_service_queue
*parent_sq
)
619 /* Service tree is empty */
620 if (!parent_sq
->nr_pending
)
623 if (!parent_sq
->first_pending
)
624 parent_sq
->first_pending
= rb_first(&parent_sq
->pending_tree
);
626 if (parent_sq
->first_pending
)
627 return rb_entry_tg(parent_sq
->first_pending
);
632 static void rb_erase_init(struct rb_node
*n
, struct rb_root
*root
)
638 static void throtl_rb_erase(struct rb_node
*n
,
639 struct throtl_service_queue
*parent_sq
)
641 if (parent_sq
->first_pending
== n
)
642 parent_sq
->first_pending
= NULL
;
643 rb_erase_init(n
, &parent_sq
->pending_tree
);
644 --parent_sq
->nr_pending
;
647 static void update_min_dispatch_time(struct throtl_service_queue
*parent_sq
)
649 struct throtl_grp
*tg
;
651 tg
= throtl_rb_first(parent_sq
);
655 parent_sq
->first_pending_disptime
= tg
->disptime
;
658 static void tg_service_queue_add(struct throtl_grp
*tg
)
660 struct throtl_service_queue
*parent_sq
= tg
->service_queue
.parent_sq
;
661 struct rb_node
**node
= &parent_sq
->pending_tree
.rb_node
;
662 struct rb_node
*parent
= NULL
;
663 struct throtl_grp
*__tg
;
664 unsigned long key
= tg
->disptime
;
667 while (*node
!= NULL
) {
669 __tg
= rb_entry_tg(parent
);
671 if (time_before(key
, __tg
->disptime
))
672 node
= &parent
->rb_left
;
674 node
= &parent
->rb_right
;
680 parent_sq
->first_pending
= &tg
->rb_node
;
682 rb_link_node(&tg
->rb_node
, parent
, node
);
683 rb_insert_color(&tg
->rb_node
, &parent_sq
->pending_tree
);
686 static void __throtl_enqueue_tg(struct throtl_grp
*tg
)
688 tg_service_queue_add(tg
);
689 tg
->flags
|= THROTL_TG_PENDING
;
690 tg
->service_queue
.parent_sq
->nr_pending
++;
693 static void throtl_enqueue_tg(struct throtl_grp
*tg
)
695 if (!(tg
->flags
& THROTL_TG_PENDING
))
696 __throtl_enqueue_tg(tg
);
699 static void __throtl_dequeue_tg(struct throtl_grp
*tg
)
701 throtl_rb_erase(&tg
->rb_node
, tg
->service_queue
.parent_sq
);
702 tg
->flags
&= ~THROTL_TG_PENDING
;
705 static void throtl_dequeue_tg(struct throtl_grp
*tg
)
707 if (tg
->flags
& THROTL_TG_PENDING
)
708 __throtl_dequeue_tg(tg
);
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 tg
->slice_end
[rw
] = roundup(jiffy_end
, tg
->td
->throtl_slice
);
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 dispached
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
,
890 bool rw
= bio_data_dir(bio
);
891 unsigned int io_allowed
;
892 unsigned long jiffy_elapsed
, jiffy_wait
, jiffy_elapsed_rnd
;
895 jiffy_elapsed
= jiffies
- tg
->slice_start
[rw
];
897 /* Round up to the next throttle slice, wait time must be nonzero */
898 jiffy_elapsed_rnd
= roundup(jiffy_elapsed
+ 1, tg
->td
->throtl_slice
);
901 * jiffy_elapsed_rnd should not be a big value as minimum iops can be
902 * 1 then at max jiffy elapsed should be equivalent of 1 second as we
903 * will allow dispatch after 1 second and after that slice should
907 tmp
= (u64
)tg_iops_limit(tg
, rw
) * jiffy_elapsed_rnd
;
911 io_allowed
= UINT_MAX
;
915 if (tg
->io_disp
[rw
] + 1 <= io_allowed
) {
921 /* Calc approx time to dispatch */
922 jiffy_wait
= ((tg
->io_disp
[rw
] + 1) * HZ
) / tg_iops_limit(tg
, rw
) + 1;
924 if (jiffy_wait
> jiffy_elapsed
)
925 jiffy_wait
= jiffy_wait
- jiffy_elapsed
;
934 static bool tg_with_in_bps_limit(struct throtl_grp
*tg
, struct bio
*bio
,
937 bool rw
= bio_data_dir(bio
);
938 u64 bytes_allowed
, extra_bytes
, tmp
;
939 unsigned long jiffy_elapsed
, jiffy_wait
, jiffy_elapsed_rnd
;
940 unsigned int bio_size
= throtl_bio_data_size(bio
);
942 jiffy_elapsed
= jiffy_elapsed_rnd
= jiffies
- tg
->slice_start
[rw
];
944 /* Slice has just started. Consider one slice interval */
946 jiffy_elapsed_rnd
= tg
->td
->throtl_slice
;
948 jiffy_elapsed_rnd
= roundup(jiffy_elapsed_rnd
, tg
->td
->throtl_slice
);
950 tmp
= tg_bps_limit(tg
, rw
) * jiffy_elapsed_rnd
;
954 if (tg
->bytes_disp
[rw
] + bio_size
<= bytes_allowed
) {
960 /* Calc approx time to dispatch */
961 extra_bytes
= tg
->bytes_disp
[rw
] + bio_size
- bytes_allowed
;
962 jiffy_wait
= div64_u64(extra_bytes
* HZ
, tg_bps_limit(tg
, rw
));
968 * This wait time is without taking into consideration the rounding
969 * up we did. Add that time also.
971 jiffy_wait
= jiffy_wait
+ (jiffy_elapsed_rnd
- jiffy_elapsed
);
978 * Returns whether one can dispatch a bio or not. Also returns approx number
979 * of jiffies to wait before this bio is with-in IO rate and can be dispatched
981 static bool tg_may_dispatch(struct throtl_grp
*tg
, struct bio
*bio
,
984 bool rw
= bio_data_dir(bio
);
985 unsigned long bps_wait
= 0, iops_wait
= 0, max_wait
= 0;
988 * Currently whole state machine of group depends on first bio
989 * queued in the group bio list. So one should not be calling
990 * this function with a different bio if there are other bios
993 BUG_ON(tg
->service_queue
.nr_queued
[rw
] &&
994 bio
!= throtl_peek_queued(&tg
->service_queue
.queued
[rw
]));
996 /* If tg->bps = -1, then BW is unlimited */
997 if (tg_bps_limit(tg
, rw
) == U64_MAX
&&
998 tg_iops_limit(tg
, rw
) == UINT_MAX
) {
1005 * If previous slice expired, start a new one otherwise renew/extend
1006 * existing slice to make sure it is at least throtl_slice interval
1007 * long since now. New slice is started only for empty throttle group.
1008 * If there is queued bio, that means there should be an active
1009 * slice and it should be extended instead.
1011 if (throtl_slice_used(tg
, rw
) && !(tg
->service_queue
.nr_queued
[rw
]))
1012 throtl_start_new_slice(tg
, rw
);
1014 if (time_before(tg
->slice_end
[rw
],
1015 jiffies
+ tg
->td
->throtl_slice
))
1016 throtl_extend_slice(tg
, rw
,
1017 jiffies
+ tg
->td
->throtl_slice
);
1020 if (tg_with_in_bps_limit(tg
, bio
, &bps_wait
) &&
1021 tg_with_in_iops_limit(tg
, bio
, &iops_wait
)) {
1027 max_wait
= max(bps_wait
, iops_wait
);
1032 if (time_before(tg
->slice_end
[rw
], jiffies
+ max_wait
))
1033 throtl_extend_slice(tg
, rw
, jiffies
+ max_wait
);
1038 static void throtl_charge_bio(struct throtl_grp
*tg
, struct bio
*bio
)
1040 bool rw
= bio_data_dir(bio
);
1041 unsigned int bio_size
= throtl_bio_data_size(bio
);
1043 /* Charge the bio to the group */
1044 tg
->bytes_disp
[rw
] += bio_size
;
1046 tg
->last_bytes_disp
[rw
] += bio_size
;
1047 tg
->last_io_disp
[rw
]++;
1050 * BIO_THROTTLED is used to prevent the same bio to be throttled
1051 * more than once as a throttled bio will go through blk-throtl the
1052 * second time when it eventually gets issued. Set it when a bio
1053 * is being charged to a tg.
1055 if (!bio_flagged(bio
, BIO_THROTTLED
))
1056 bio_set_flag(bio
, BIO_THROTTLED
);
1060 * throtl_add_bio_tg - add a bio to the specified throtl_grp
1063 * @tg: the target throtl_grp
1065 * Add @bio to @tg's service_queue using @qn. If @qn is not specified,
1066 * tg->qnode_on_self[] is used.
1068 static void throtl_add_bio_tg(struct bio
*bio
, struct throtl_qnode
*qn
,
1069 struct throtl_grp
*tg
)
1071 struct throtl_service_queue
*sq
= &tg
->service_queue
;
1072 bool rw
= bio_data_dir(bio
);
1075 qn
= &tg
->qnode_on_self
[rw
];
1078 * If @tg doesn't currently have any bios queued in the same
1079 * direction, queueing @bio can change when @tg should be
1080 * dispatched. Mark that @tg was empty. This is automatically
1081 * cleaered on the next tg_update_disptime().
1083 if (!sq
->nr_queued
[rw
])
1084 tg
->flags
|= THROTL_TG_WAS_EMPTY
;
1086 throtl_qnode_add_bio(bio
, qn
, &sq
->queued
[rw
]);
1088 sq
->nr_queued
[rw
]++;
1089 throtl_enqueue_tg(tg
);
1092 static void tg_update_disptime(struct throtl_grp
*tg
)
1094 struct throtl_service_queue
*sq
= &tg
->service_queue
;
1095 unsigned long read_wait
= -1, write_wait
= -1, min_wait
= -1, disptime
;
1098 bio
= throtl_peek_queued(&sq
->queued
[READ
]);
1100 tg_may_dispatch(tg
, bio
, &read_wait
);
1102 bio
= throtl_peek_queued(&sq
->queued
[WRITE
]);
1104 tg_may_dispatch(tg
, bio
, &write_wait
);
1106 min_wait
= min(read_wait
, write_wait
);
1107 disptime
= jiffies
+ min_wait
;
1109 /* Update dispatch time */
1110 throtl_dequeue_tg(tg
);
1111 tg
->disptime
= disptime
;
1112 throtl_enqueue_tg(tg
);
1114 /* see throtl_add_bio_tg() */
1115 tg
->flags
&= ~THROTL_TG_WAS_EMPTY
;
1118 static void start_parent_slice_with_credit(struct throtl_grp
*child_tg
,
1119 struct throtl_grp
*parent_tg
, bool rw
)
1121 if (throtl_slice_used(parent_tg
, rw
)) {
1122 throtl_start_new_slice_with_credit(parent_tg
, rw
,
1123 child_tg
->slice_start
[rw
]);
1128 static void tg_dispatch_one_bio(struct throtl_grp
*tg
, bool rw
)
1130 struct throtl_service_queue
*sq
= &tg
->service_queue
;
1131 struct throtl_service_queue
*parent_sq
= sq
->parent_sq
;
1132 struct throtl_grp
*parent_tg
= sq_to_tg(parent_sq
);
1133 struct throtl_grp
*tg_to_put
= NULL
;
1137 * @bio is being transferred from @tg to @parent_sq. Popping a bio
1138 * from @tg may put its reference and @parent_sq might end up
1139 * getting released prematurely. Remember the tg to put and put it
1140 * after @bio is transferred to @parent_sq.
1142 bio
= throtl_pop_queued(&sq
->queued
[rw
], &tg_to_put
);
1143 sq
->nr_queued
[rw
]--;
1145 throtl_charge_bio(tg
, bio
);
1148 * If our parent is another tg, we just need to transfer @bio to
1149 * the parent using throtl_add_bio_tg(). If our parent is
1150 * @td->service_queue, @bio is ready to be issued. Put it on its
1151 * bio_lists[] and decrease total number queued. The caller is
1152 * responsible for issuing these bios.
1155 throtl_add_bio_tg(bio
, &tg
->qnode_on_parent
[rw
], parent_tg
);
1156 start_parent_slice_with_credit(tg
, parent_tg
, rw
);
1158 throtl_qnode_add_bio(bio
, &tg
->qnode_on_parent
[rw
],
1159 &parent_sq
->queued
[rw
]);
1160 BUG_ON(tg
->td
->nr_queued
[rw
] <= 0);
1161 tg
->td
->nr_queued
[rw
]--;
1164 throtl_trim_slice(tg
, rw
);
1167 blkg_put(tg_to_blkg(tg_to_put
));
1170 static int throtl_dispatch_tg(struct throtl_grp
*tg
)
1172 struct throtl_service_queue
*sq
= &tg
->service_queue
;
1173 unsigned int nr_reads
= 0, nr_writes
= 0;
1174 unsigned int max_nr_reads
= throtl_grp_quantum
*3/4;
1175 unsigned int max_nr_writes
= throtl_grp_quantum
- max_nr_reads
;
1178 /* Try to dispatch 75% READS and 25% WRITES */
1180 while ((bio
= throtl_peek_queued(&sq
->queued
[READ
])) &&
1181 tg_may_dispatch(tg
, bio
, NULL
)) {
1183 tg_dispatch_one_bio(tg
, bio_data_dir(bio
));
1186 if (nr_reads
>= max_nr_reads
)
1190 while ((bio
= throtl_peek_queued(&sq
->queued
[WRITE
])) &&
1191 tg_may_dispatch(tg
, bio
, NULL
)) {
1193 tg_dispatch_one_bio(tg
, bio_data_dir(bio
));
1196 if (nr_writes
>= max_nr_writes
)
1200 return nr_reads
+ nr_writes
;
1203 static int throtl_select_dispatch(struct throtl_service_queue
*parent_sq
)
1205 unsigned int nr_disp
= 0;
1208 struct throtl_grp
*tg
= throtl_rb_first(parent_sq
);
1209 struct throtl_service_queue
*sq
= &tg
->service_queue
;
1214 if (time_before(jiffies
, tg
->disptime
))
1217 throtl_dequeue_tg(tg
);
1219 nr_disp
+= throtl_dispatch_tg(tg
);
1221 if (sq
->nr_queued
[0] || sq
->nr_queued
[1])
1222 tg_update_disptime(tg
);
1224 if (nr_disp
>= throtl_quantum
)
1231 static bool throtl_can_upgrade(struct throtl_data
*td
,
1232 struct throtl_grp
*this_tg
);
1234 * throtl_pending_timer_fn - timer function for service_queue->pending_timer
1235 * @arg: the throtl_service_queue being serviced
1237 * This timer is armed when a child throtl_grp with active bio's become
1238 * pending and queued on the service_queue's pending_tree and expires when
1239 * the first child throtl_grp should be dispatched. This function
1240 * dispatches bio's from the children throtl_grps to the parent
1243 * If the parent's parent is another throtl_grp, dispatching is propagated
1244 * by either arming its pending_timer or repeating dispatch directly. If
1245 * the top-level service_tree is reached, throtl_data->dispatch_work is
1246 * kicked so that the ready bio's are issued.
1248 static void throtl_pending_timer_fn(struct timer_list
*t
)
1250 struct throtl_service_queue
*sq
= from_timer(sq
, t
, pending_timer
);
1251 struct throtl_grp
*tg
= sq_to_tg(sq
);
1252 struct throtl_data
*td
= sq_to_td(sq
);
1253 struct request_queue
*q
= td
->queue
;
1254 struct throtl_service_queue
*parent_sq
;
1258 spin_lock_irq(q
->queue_lock
);
1259 if (throtl_can_upgrade(td
, NULL
))
1260 throtl_upgrade_state(td
);
1263 parent_sq
= sq
->parent_sq
;
1267 throtl_log(sq
, "dispatch nr_queued=%u read=%u write=%u",
1268 sq
->nr_queued
[READ
] + sq
->nr_queued
[WRITE
],
1269 sq
->nr_queued
[READ
], sq
->nr_queued
[WRITE
]);
1271 ret
= throtl_select_dispatch(sq
);
1273 throtl_log(sq
, "bios disp=%u", ret
);
1277 if (throtl_schedule_next_dispatch(sq
, false))
1280 /* this dispatch windows is still open, relax and repeat */
1281 spin_unlock_irq(q
->queue_lock
);
1283 spin_lock_irq(q
->queue_lock
);
1290 /* @parent_sq is another throl_grp, propagate dispatch */
1291 if (tg
->flags
& THROTL_TG_WAS_EMPTY
) {
1292 tg_update_disptime(tg
);
1293 if (!throtl_schedule_next_dispatch(parent_sq
, false)) {
1294 /* window is already open, repeat dispatching */
1301 /* reached the top-level, queue issueing */
1302 queue_work(kthrotld_workqueue
, &td
->dispatch_work
);
1305 spin_unlock_irq(q
->queue_lock
);
1309 * blk_throtl_dispatch_work_fn - work function for throtl_data->dispatch_work
1310 * @work: work item being executed
1312 * This function is queued for execution when bio's reach the bio_lists[]
1313 * of throtl_data->service_queue. Those bio's are ready and issued by this
1316 static void blk_throtl_dispatch_work_fn(struct work_struct
*work
)
1318 struct throtl_data
*td
= container_of(work
, struct throtl_data
,
1320 struct throtl_service_queue
*td_sq
= &td
->service_queue
;
1321 struct request_queue
*q
= td
->queue
;
1322 struct bio_list bio_list_on_stack
;
1324 struct blk_plug plug
;
1327 bio_list_init(&bio_list_on_stack
);
1329 spin_lock_irq(q
->queue_lock
);
1330 for (rw
= READ
; rw
<= WRITE
; rw
++)
1331 while ((bio
= throtl_pop_queued(&td_sq
->queued
[rw
], NULL
)))
1332 bio_list_add(&bio_list_on_stack
, bio
);
1333 spin_unlock_irq(q
->queue_lock
);
1335 if (!bio_list_empty(&bio_list_on_stack
)) {
1336 blk_start_plug(&plug
);
1337 while((bio
= bio_list_pop(&bio_list_on_stack
)))
1338 generic_make_request(bio
);
1339 blk_finish_plug(&plug
);
1343 static u64
tg_prfill_conf_u64(struct seq_file
*sf
, struct blkg_policy_data
*pd
,
1346 struct throtl_grp
*tg
= pd_to_tg(pd
);
1347 u64 v
= *(u64
*)((void *)tg
+ off
);
1351 return __blkg_prfill_u64(sf
, pd
, v
);
1354 static u64
tg_prfill_conf_uint(struct seq_file
*sf
, struct blkg_policy_data
*pd
,
1357 struct throtl_grp
*tg
= pd_to_tg(pd
);
1358 unsigned int v
= *(unsigned int *)((void *)tg
+ off
);
1362 return __blkg_prfill_u64(sf
, pd
, v
);
1365 static int tg_print_conf_u64(struct seq_file
*sf
, void *v
)
1367 blkcg_print_blkgs(sf
, css_to_blkcg(seq_css(sf
)), tg_prfill_conf_u64
,
1368 &blkcg_policy_throtl
, seq_cft(sf
)->private, false);
1372 static int tg_print_conf_uint(struct seq_file
*sf
, void *v
)
1374 blkcg_print_blkgs(sf
, css_to_blkcg(seq_css(sf
)), tg_prfill_conf_uint
,
1375 &blkcg_policy_throtl
, seq_cft(sf
)->private, false);
1379 static void tg_conf_updated(struct throtl_grp
*tg
, bool global
)
1381 struct throtl_service_queue
*sq
= &tg
->service_queue
;
1382 struct cgroup_subsys_state
*pos_css
;
1383 struct blkcg_gq
*blkg
;
1385 throtl_log(&tg
->service_queue
,
1386 "limit change rbps=%llu wbps=%llu riops=%u wiops=%u",
1387 tg_bps_limit(tg
, READ
), tg_bps_limit(tg
, WRITE
),
1388 tg_iops_limit(tg
, READ
), tg_iops_limit(tg
, WRITE
));
1391 * Update has_rules[] flags for the updated tg's subtree. A tg is
1392 * considered to have rules if either the tg itself or any of its
1393 * ancestors has rules. This identifies groups without any
1394 * restrictions in the whole hierarchy and allows them to bypass
1397 blkg_for_each_descendant_pre(blkg
, pos_css
,
1398 global
? tg
->td
->queue
->root_blkg
: tg_to_blkg(tg
)) {
1399 struct throtl_grp
*this_tg
= blkg_to_tg(blkg
);
1400 struct throtl_grp
*parent_tg
;
1402 tg_update_has_rules(this_tg
);
1403 /* ignore root/second level */
1404 if (!cgroup_subsys_on_dfl(io_cgrp_subsys
) || !blkg
->parent
||
1405 !blkg
->parent
->parent
)
1407 parent_tg
= blkg_to_tg(blkg
->parent
);
1409 * make sure all children has lower idle time threshold and
1410 * higher latency target
1412 this_tg
->idletime_threshold
= min(this_tg
->idletime_threshold
,
1413 parent_tg
->idletime_threshold
);
1414 this_tg
->latency_target
= max(this_tg
->latency_target
,
1415 parent_tg
->latency_target
);
1419 * We're already holding queue_lock and know @tg is valid. Let's
1420 * apply the new config directly.
1422 * Restart the slices for both READ and WRITES. It might happen
1423 * that a group's limit are dropped suddenly and we don't want to
1424 * account recently dispatched IO with new low rate.
1426 throtl_start_new_slice(tg
, 0);
1427 throtl_start_new_slice(tg
, 1);
1429 if (tg
->flags
& THROTL_TG_PENDING
) {
1430 tg_update_disptime(tg
);
1431 throtl_schedule_next_dispatch(sq
->parent_sq
, true);
1435 static ssize_t
tg_set_conf(struct kernfs_open_file
*of
,
1436 char *buf
, size_t nbytes
, loff_t off
, bool is_u64
)
1438 struct blkcg
*blkcg
= css_to_blkcg(of_css(of
));
1439 struct blkg_conf_ctx ctx
;
1440 struct throtl_grp
*tg
;
1444 ret
= blkg_conf_prep(blkcg
, &blkcg_policy_throtl
, buf
, &ctx
);
1449 if (sscanf(ctx
.body
, "%llu", &v
) != 1)
1454 tg
= blkg_to_tg(ctx
.blkg
);
1457 *(u64
*)((void *)tg
+ of_cft(of
)->private) = v
;
1459 *(unsigned int *)((void *)tg
+ of_cft(of
)->private) = v
;
1461 tg_conf_updated(tg
, false);
1464 blkg_conf_finish(&ctx
);
1465 return ret
?: nbytes
;
1468 static ssize_t
tg_set_conf_u64(struct kernfs_open_file
*of
,
1469 char *buf
, size_t nbytes
, loff_t off
)
1471 return tg_set_conf(of
, buf
, nbytes
, off
, true);
1474 static ssize_t
tg_set_conf_uint(struct kernfs_open_file
*of
,
1475 char *buf
, size_t nbytes
, loff_t off
)
1477 return tg_set_conf(of
, buf
, nbytes
, off
, false);
1480 static struct cftype throtl_legacy_files
[] = {
1482 .name
= "throttle.read_bps_device",
1483 .private = offsetof(struct throtl_grp
, bps
[READ
][LIMIT_MAX
]),
1484 .seq_show
= tg_print_conf_u64
,
1485 .write
= tg_set_conf_u64
,
1488 .name
= "throttle.write_bps_device",
1489 .private = offsetof(struct throtl_grp
, bps
[WRITE
][LIMIT_MAX
]),
1490 .seq_show
= tg_print_conf_u64
,
1491 .write
= tg_set_conf_u64
,
1494 .name
= "throttle.read_iops_device",
1495 .private = offsetof(struct throtl_grp
, iops
[READ
][LIMIT_MAX
]),
1496 .seq_show
= tg_print_conf_uint
,
1497 .write
= tg_set_conf_uint
,
1500 .name
= "throttle.write_iops_device",
1501 .private = offsetof(struct throtl_grp
, iops
[WRITE
][LIMIT_MAX
]),
1502 .seq_show
= tg_print_conf_uint
,
1503 .write
= tg_set_conf_uint
,
1506 .name
= "throttle.io_service_bytes",
1507 .private = (unsigned long)&blkcg_policy_throtl
,
1508 .seq_show
= blkg_print_stat_bytes
,
1511 .name
= "throttle.io_serviced",
1512 .private = (unsigned long)&blkcg_policy_throtl
,
1513 .seq_show
= blkg_print_stat_ios
,
1518 static u64
tg_prfill_limit(struct seq_file
*sf
, struct blkg_policy_data
*pd
,
1521 struct throtl_grp
*tg
= pd_to_tg(pd
);
1522 const char *dname
= blkg_dev_name(pd
->blkg
);
1523 char bufs
[4][21] = { "max", "max", "max", "max" };
1525 unsigned int iops_dft
;
1526 char idle_time
[26] = "";
1527 char latency_time
[26] = "";
1532 if (off
== LIMIT_LOW
) {
1537 iops_dft
= UINT_MAX
;
1540 if (tg
->bps_conf
[READ
][off
] == bps_dft
&&
1541 tg
->bps_conf
[WRITE
][off
] == bps_dft
&&
1542 tg
->iops_conf
[READ
][off
] == iops_dft
&&
1543 tg
->iops_conf
[WRITE
][off
] == iops_dft
&&
1544 (off
!= LIMIT_LOW
||
1545 (tg
->idletime_threshold_conf
== DFL_IDLE_THRESHOLD
&&
1546 tg
->latency_target_conf
== DFL_LATENCY_TARGET
)))
1549 if (tg
->bps_conf
[READ
][off
] != U64_MAX
)
1550 snprintf(bufs
[0], sizeof(bufs
[0]), "%llu",
1551 tg
->bps_conf
[READ
][off
]);
1552 if (tg
->bps_conf
[WRITE
][off
] != U64_MAX
)
1553 snprintf(bufs
[1], sizeof(bufs
[1]), "%llu",
1554 tg
->bps_conf
[WRITE
][off
]);
1555 if (tg
->iops_conf
[READ
][off
] != UINT_MAX
)
1556 snprintf(bufs
[2], sizeof(bufs
[2]), "%u",
1557 tg
->iops_conf
[READ
][off
]);
1558 if (tg
->iops_conf
[WRITE
][off
] != UINT_MAX
)
1559 snprintf(bufs
[3], sizeof(bufs
[3]), "%u",
1560 tg
->iops_conf
[WRITE
][off
]);
1561 if (off
== LIMIT_LOW
) {
1562 if (tg
->idletime_threshold_conf
== ULONG_MAX
)
1563 strcpy(idle_time
, " idle=max");
1565 snprintf(idle_time
, sizeof(idle_time
), " idle=%lu",
1566 tg
->idletime_threshold_conf
);
1568 if (tg
->latency_target_conf
== ULONG_MAX
)
1569 strcpy(latency_time
, " latency=max");
1571 snprintf(latency_time
, sizeof(latency_time
),
1572 " latency=%lu", tg
->latency_target_conf
);
1575 seq_printf(sf
, "%s rbps=%s wbps=%s riops=%s wiops=%s%s%s\n",
1576 dname
, bufs
[0], bufs
[1], bufs
[2], bufs
[3], idle_time
,
1581 static int tg_print_limit(struct seq_file
*sf
, void *v
)
1583 blkcg_print_blkgs(sf
, css_to_blkcg(seq_css(sf
)), tg_prfill_limit
,
1584 &blkcg_policy_throtl
, seq_cft(sf
)->private, false);
1588 static ssize_t
tg_set_limit(struct kernfs_open_file
*of
,
1589 char *buf
, size_t nbytes
, loff_t off
)
1591 struct blkcg
*blkcg
= css_to_blkcg(of_css(of
));
1592 struct blkg_conf_ctx ctx
;
1593 struct throtl_grp
*tg
;
1595 unsigned long idle_time
;
1596 unsigned long latency_time
;
1598 int index
= of_cft(of
)->private;
1600 ret
= blkg_conf_prep(blkcg
, &blkcg_policy_throtl
, buf
, &ctx
);
1604 tg
= blkg_to_tg(ctx
.blkg
);
1606 v
[0] = tg
->bps_conf
[READ
][index
];
1607 v
[1] = tg
->bps_conf
[WRITE
][index
];
1608 v
[2] = tg
->iops_conf
[READ
][index
];
1609 v
[3] = tg
->iops_conf
[WRITE
][index
];
1611 idle_time
= tg
->idletime_threshold_conf
;
1612 latency_time
= tg
->latency_target_conf
;
1614 char tok
[27]; /* wiops=18446744073709551616 */
1619 if (sscanf(ctx
.body
, "%26s%n", tok
, &len
) != 1)
1628 if (!p
|| (sscanf(p
, "%llu", &val
) != 1 && strcmp(p
, "max")))
1636 if (!strcmp(tok
, "rbps"))
1638 else if (!strcmp(tok
, "wbps"))
1640 else if (!strcmp(tok
, "riops"))
1641 v
[2] = min_t(u64
, val
, UINT_MAX
);
1642 else if (!strcmp(tok
, "wiops"))
1643 v
[3] = min_t(u64
, val
, UINT_MAX
);
1644 else if (off
== LIMIT_LOW
&& !strcmp(tok
, "idle"))
1646 else if (off
== LIMIT_LOW
&& !strcmp(tok
, "latency"))
1652 tg
->bps_conf
[READ
][index
] = v
[0];
1653 tg
->bps_conf
[WRITE
][index
] = v
[1];
1654 tg
->iops_conf
[READ
][index
] = v
[2];
1655 tg
->iops_conf
[WRITE
][index
] = v
[3];
1657 if (index
== LIMIT_MAX
) {
1658 tg
->bps
[READ
][index
] = v
[0];
1659 tg
->bps
[WRITE
][index
] = v
[1];
1660 tg
->iops
[READ
][index
] = v
[2];
1661 tg
->iops
[WRITE
][index
] = v
[3];
1663 tg
->bps
[READ
][LIMIT_LOW
] = min(tg
->bps_conf
[READ
][LIMIT_LOW
],
1664 tg
->bps_conf
[READ
][LIMIT_MAX
]);
1665 tg
->bps
[WRITE
][LIMIT_LOW
] = min(tg
->bps_conf
[WRITE
][LIMIT_LOW
],
1666 tg
->bps_conf
[WRITE
][LIMIT_MAX
]);
1667 tg
->iops
[READ
][LIMIT_LOW
] = min(tg
->iops_conf
[READ
][LIMIT_LOW
],
1668 tg
->iops_conf
[READ
][LIMIT_MAX
]);
1669 tg
->iops
[WRITE
][LIMIT_LOW
] = min(tg
->iops_conf
[WRITE
][LIMIT_LOW
],
1670 tg
->iops_conf
[WRITE
][LIMIT_MAX
]);
1671 tg
->idletime_threshold_conf
= idle_time
;
1672 tg
->latency_target_conf
= latency_time
;
1674 /* force user to configure all settings for low limit */
1675 if (!(tg
->bps
[READ
][LIMIT_LOW
] || tg
->iops
[READ
][LIMIT_LOW
] ||
1676 tg
->bps
[WRITE
][LIMIT_LOW
] || tg
->iops
[WRITE
][LIMIT_LOW
]) ||
1677 tg
->idletime_threshold_conf
== DFL_IDLE_THRESHOLD
||
1678 tg
->latency_target_conf
== DFL_LATENCY_TARGET
) {
1679 tg
->bps
[READ
][LIMIT_LOW
] = 0;
1680 tg
->bps
[WRITE
][LIMIT_LOW
] = 0;
1681 tg
->iops
[READ
][LIMIT_LOW
] = 0;
1682 tg
->iops
[WRITE
][LIMIT_LOW
] = 0;
1683 tg
->idletime_threshold
= DFL_IDLE_THRESHOLD
;
1684 tg
->latency_target
= DFL_LATENCY_TARGET
;
1685 } else if (index
== LIMIT_LOW
) {
1686 tg
->idletime_threshold
= tg
->idletime_threshold_conf
;
1687 tg
->latency_target
= tg
->latency_target_conf
;
1690 blk_throtl_update_limit_valid(tg
->td
);
1691 if (tg
->td
->limit_valid
[LIMIT_LOW
]) {
1692 if (index
== LIMIT_LOW
)
1693 tg
->td
->limit_index
= LIMIT_LOW
;
1695 tg
->td
->limit_index
= LIMIT_MAX
;
1696 tg_conf_updated(tg
, index
== LIMIT_LOW
&&
1697 tg
->td
->limit_valid
[LIMIT_LOW
]);
1700 blkg_conf_finish(&ctx
);
1701 return ret
?: nbytes
;
1704 static struct cftype throtl_files
[] = {
1705 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
1708 .flags
= CFTYPE_NOT_ON_ROOT
,
1709 .seq_show
= tg_print_limit
,
1710 .write
= tg_set_limit
,
1711 .private = LIMIT_LOW
,
1716 .flags
= CFTYPE_NOT_ON_ROOT
,
1717 .seq_show
= tg_print_limit
,
1718 .write
= tg_set_limit
,
1719 .private = LIMIT_MAX
,
1724 static void throtl_shutdown_wq(struct request_queue
*q
)
1726 struct throtl_data
*td
= q
->td
;
1728 cancel_work_sync(&td
->dispatch_work
);
1731 static struct blkcg_policy blkcg_policy_throtl
= {
1732 .dfl_cftypes
= throtl_files
,
1733 .legacy_cftypes
= throtl_legacy_files
,
1735 .pd_alloc_fn
= throtl_pd_alloc
,
1736 .pd_init_fn
= throtl_pd_init
,
1737 .pd_online_fn
= throtl_pd_online
,
1738 .pd_offline_fn
= throtl_pd_offline
,
1739 .pd_free_fn
= throtl_pd_free
,
1742 static unsigned long __tg_last_low_overflow_time(struct throtl_grp
*tg
)
1744 unsigned long rtime
= jiffies
, wtime
= jiffies
;
1746 if (tg
->bps
[READ
][LIMIT_LOW
] || tg
->iops
[READ
][LIMIT_LOW
])
1747 rtime
= tg
->last_low_overflow_time
[READ
];
1748 if (tg
->bps
[WRITE
][LIMIT_LOW
] || tg
->iops
[WRITE
][LIMIT_LOW
])
1749 wtime
= tg
->last_low_overflow_time
[WRITE
];
1750 return min(rtime
, wtime
);
1753 /* tg should not be an intermediate node */
1754 static unsigned long tg_last_low_overflow_time(struct throtl_grp
*tg
)
1756 struct throtl_service_queue
*parent_sq
;
1757 struct throtl_grp
*parent
= tg
;
1758 unsigned long ret
= __tg_last_low_overflow_time(tg
);
1761 parent_sq
= parent
->service_queue
.parent_sq
;
1762 parent
= sq_to_tg(parent_sq
);
1767 * The parent doesn't have low limit, it always reaches low
1768 * limit. Its overflow time is useless for children
1770 if (!parent
->bps
[READ
][LIMIT_LOW
] &&
1771 !parent
->iops
[READ
][LIMIT_LOW
] &&
1772 !parent
->bps
[WRITE
][LIMIT_LOW
] &&
1773 !parent
->iops
[WRITE
][LIMIT_LOW
])
1775 if (time_after(__tg_last_low_overflow_time(parent
), ret
))
1776 ret
= __tg_last_low_overflow_time(parent
);
1781 static bool throtl_tg_is_idle(struct throtl_grp
*tg
)
1784 * cgroup is idle if:
1785 * - single idle is too long, longer than a fixed value (in case user
1786 * configure a too big threshold) or 4 times of idletime threshold
1787 * - average think time is more than threshold
1788 * - IO latency is largely below threshold
1793 time
= min_t(unsigned long, MAX_IDLE_TIME
, 4 * tg
->idletime_threshold
);
1794 ret
= tg
->latency_target
== DFL_LATENCY_TARGET
||
1795 tg
->idletime_threshold
== DFL_IDLE_THRESHOLD
||
1796 (ktime_get_ns() >> 10) - tg
->last_finish_time
> time
||
1797 tg
->avg_idletime
> tg
->idletime_threshold
||
1798 (tg
->latency_target
&& tg
->bio_cnt
&&
1799 tg
->bad_bio_cnt
* 5 < tg
->bio_cnt
);
1800 throtl_log(&tg
->service_queue
,
1801 "avg_idle=%ld, idle_threshold=%ld, bad_bio=%d, total_bio=%d, is_idle=%d, scale=%d",
1802 tg
->avg_idletime
, tg
->idletime_threshold
, tg
->bad_bio_cnt
,
1803 tg
->bio_cnt
, ret
, tg
->td
->scale
);
1807 static bool throtl_tg_can_upgrade(struct throtl_grp
*tg
)
1809 struct throtl_service_queue
*sq
= &tg
->service_queue
;
1810 bool read_limit
, write_limit
;
1813 * if cgroup reaches low limit (if low limit is 0, the cgroup always
1814 * reaches), it's ok to upgrade to next limit
1816 read_limit
= tg
->bps
[READ
][LIMIT_LOW
] || tg
->iops
[READ
][LIMIT_LOW
];
1817 write_limit
= tg
->bps
[WRITE
][LIMIT_LOW
] || tg
->iops
[WRITE
][LIMIT_LOW
];
1818 if (!read_limit
&& !write_limit
)
1820 if (read_limit
&& sq
->nr_queued
[READ
] &&
1821 (!write_limit
|| sq
->nr_queued
[WRITE
]))
1823 if (write_limit
&& sq
->nr_queued
[WRITE
] &&
1824 (!read_limit
|| sq
->nr_queued
[READ
]))
1827 if (time_after_eq(jiffies
,
1828 tg_last_low_overflow_time(tg
) + tg
->td
->throtl_slice
) &&
1829 throtl_tg_is_idle(tg
))
1834 static bool throtl_hierarchy_can_upgrade(struct throtl_grp
*tg
)
1837 if (throtl_tg_can_upgrade(tg
))
1839 tg
= sq_to_tg(tg
->service_queue
.parent_sq
);
1840 if (!tg
|| !tg_to_blkg(tg
)->parent
)
1846 static bool throtl_can_upgrade(struct throtl_data
*td
,
1847 struct throtl_grp
*this_tg
)
1849 struct cgroup_subsys_state
*pos_css
;
1850 struct blkcg_gq
*blkg
;
1852 if (td
->limit_index
!= LIMIT_LOW
)
1855 if (time_before(jiffies
, td
->low_downgrade_time
+ td
->throtl_slice
))
1859 blkg_for_each_descendant_post(blkg
, pos_css
, td
->queue
->root_blkg
) {
1860 struct throtl_grp
*tg
= blkg_to_tg(blkg
);
1864 if (!list_empty(&tg_to_blkg(tg
)->blkcg
->css
.children
))
1866 if (!throtl_hierarchy_can_upgrade(tg
)) {
1875 static void throtl_upgrade_check(struct throtl_grp
*tg
)
1877 unsigned long now
= jiffies
;
1879 if (tg
->td
->limit_index
!= LIMIT_LOW
)
1882 if (time_after(tg
->last_check_time
+ tg
->td
->throtl_slice
, now
))
1885 tg
->last_check_time
= now
;
1887 if (!time_after_eq(now
,
1888 __tg_last_low_overflow_time(tg
) + tg
->td
->throtl_slice
))
1891 if (throtl_can_upgrade(tg
->td
, NULL
))
1892 throtl_upgrade_state(tg
->td
);
1895 static void throtl_upgrade_state(struct throtl_data
*td
)
1897 struct cgroup_subsys_state
*pos_css
;
1898 struct blkcg_gq
*blkg
;
1900 throtl_log(&td
->service_queue
, "upgrade to max");
1901 td
->limit_index
= LIMIT_MAX
;
1902 td
->low_upgrade_time
= jiffies
;
1905 blkg_for_each_descendant_post(blkg
, pos_css
, td
->queue
->root_blkg
) {
1906 struct throtl_grp
*tg
= blkg_to_tg(blkg
);
1907 struct throtl_service_queue
*sq
= &tg
->service_queue
;
1909 tg
->disptime
= jiffies
- 1;
1910 throtl_select_dispatch(sq
);
1911 throtl_schedule_next_dispatch(sq
, true);
1914 throtl_select_dispatch(&td
->service_queue
);
1915 throtl_schedule_next_dispatch(&td
->service_queue
, true);
1916 queue_work(kthrotld_workqueue
, &td
->dispatch_work
);
1919 static void throtl_downgrade_state(struct throtl_data
*td
, int new)
1923 throtl_log(&td
->service_queue
, "downgrade, scale %d", td
->scale
);
1925 td
->low_upgrade_time
= jiffies
- td
->scale
* td
->throtl_slice
;
1929 td
->limit_index
= new;
1930 td
->low_downgrade_time
= jiffies
;
1933 static bool throtl_tg_can_downgrade(struct throtl_grp
*tg
)
1935 struct throtl_data
*td
= tg
->td
;
1936 unsigned long now
= jiffies
;
1939 * If cgroup is below low limit, consider downgrade and throttle other
1942 if (time_after_eq(now
, td
->low_upgrade_time
+ td
->throtl_slice
) &&
1943 time_after_eq(now
, tg_last_low_overflow_time(tg
) +
1944 td
->throtl_slice
) &&
1945 (!throtl_tg_is_idle(tg
) ||
1946 !list_empty(&tg_to_blkg(tg
)->blkcg
->css
.children
)))
1951 static bool throtl_hierarchy_can_downgrade(struct throtl_grp
*tg
)
1954 if (!throtl_tg_can_downgrade(tg
))
1956 tg
= sq_to_tg(tg
->service_queue
.parent_sq
);
1957 if (!tg
|| !tg_to_blkg(tg
)->parent
)
1963 static void throtl_downgrade_check(struct throtl_grp
*tg
)
1967 unsigned long elapsed_time
;
1968 unsigned long now
= jiffies
;
1970 if (tg
->td
->limit_index
!= LIMIT_MAX
||
1971 !tg
->td
->limit_valid
[LIMIT_LOW
])
1973 if (!list_empty(&tg_to_blkg(tg
)->blkcg
->css
.children
))
1975 if (time_after(tg
->last_check_time
+ tg
->td
->throtl_slice
, now
))
1978 elapsed_time
= now
- tg
->last_check_time
;
1979 tg
->last_check_time
= now
;
1981 if (time_before(now
, tg_last_low_overflow_time(tg
) +
1982 tg
->td
->throtl_slice
))
1985 if (tg
->bps
[READ
][LIMIT_LOW
]) {
1986 bps
= tg
->last_bytes_disp
[READ
] * HZ
;
1987 do_div(bps
, elapsed_time
);
1988 if (bps
>= tg
->bps
[READ
][LIMIT_LOW
])
1989 tg
->last_low_overflow_time
[READ
] = now
;
1992 if (tg
->bps
[WRITE
][LIMIT_LOW
]) {
1993 bps
= tg
->last_bytes_disp
[WRITE
] * HZ
;
1994 do_div(bps
, elapsed_time
);
1995 if (bps
>= tg
->bps
[WRITE
][LIMIT_LOW
])
1996 tg
->last_low_overflow_time
[WRITE
] = now
;
1999 if (tg
->iops
[READ
][LIMIT_LOW
]) {
2000 iops
= tg
->last_io_disp
[READ
] * HZ
/ elapsed_time
;
2001 if (iops
>= tg
->iops
[READ
][LIMIT_LOW
])
2002 tg
->last_low_overflow_time
[READ
] = now
;
2005 if (tg
->iops
[WRITE
][LIMIT_LOW
]) {
2006 iops
= tg
->last_io_disp
[WRITE
] * HZ
/ elapsed_time
;
2007 if (iops
>= tg
->iops
[WRITE
][LIMIT_LOW
])
2008 tg
->last_low_overflow_time
[WRITE
] = now
;
2012 * If cgroup is below low limit, consider downgrade and throttle other
2015 if (throtl_hierarchy_can_downgrade(tg
))
2016 throtl_downgrade_state(tg
->td
, LIMIT_LOW
);
2018 tg
->last_bytes_disp
[READ
] = 0;
2019 tg
->last_bytes_disp
[WRITE
] = 0;
2020 tg
->last_io_disp
[READ
] = 0;
2021 tg
->last_io_disp
[WRITE
] = 0;
2024 static void blk_throtl_update_idletime(struct throtl_grp
*tg
)
2026 unsigned long now
= ktime_get_ns() >> 10;
2027 unsigned long last_finish_time
= tg
->last_finish_time
;
2029 if (now
<= last_finish_time
|| last_finish_time
== 0 ||
2030 last_finish_time
== tg
->checked_last_finish_time
)
2033 tg
->avg_idletime
= (tg
->avg_idletime
* 7 + now
- last_finish_time
) >> 3;
2034 tg
->checked_last_finish_time
= last_finish_time
;
2037 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2038 static void throtl_update_latency_buckets(struct throtl_data
*td
)
2040 struct avg_latency_bucket avg_latency
[LATENCY_BUCKET_SIZE
];
2042 unsigned long last_latency
= 0;
2043 unsigned long latency
;
2045 if (!blk_queue_nonrot(td
->queue
))
2047 if (time_before(jiffies
, td
->last_calculate_time
+ HZ
))
2049 td
->last_calculate_time
= jiffies
;
2051 memset(avg_latency
, 0, sizeof(avg_latency
));
2052 for (i
= 0; i
< LATENCY_BUCKET_SIZE
; i
++) {
2053 struct latency_bucket
*tmp
= &td
->tmp_buckets
[i
];
2055 for_each_possible_cpu(cpu
) {
2056 struct latency_bucket
*bucket
;
2058 /* this isn't race free, but ok in practice */
2059 bucket
= per_cpu_ptr(td
->latency_buckets
, cpu
);
2060 tmp
->total_latency
+= bucket
[i
].total_latency
;
2061 tmp
->samples
+= bucket
[i
].samples
;
2062 bucket
[i
].total_latency
= 0;
2063 bucket
[i
].samples
= 0;
2066 if (tmp
->samples
>= 32) {
2067 int samples
= tmp
->samples
;
2069 latency
= tmp
->total_latency
;
2071 tmp
->total_latency
= 0;
2076 avg_latency
[i
].latency
= latency
;
2080 for (i
= 0; i
< LATENCY_BUCKET_SIZE
; i
++) {
2081 if (!avg_latency
[i
].latency
) {
2082 if (td
->avg_buckets
[i
].latency
< last_latency
)
2083 td
->avg_buckets
[i
].latency
= last_latency
;
2087 if (!td
->avg_buckets
[i
].valid
)
2088 latency
= avg_latency
[i
].latency
;
2090 latency
= (td
->avg_buckets
[i
].latency
* 7 +
2091 avg_latency
[i
].latency
) >> 3;
2093 td
->avg_buckets
[i
].latency
= max(latency
, last_latency
);
2094 td
->avg_buckets
[i
].valid
= true;
2095 last_latency
= td
->avg_buckets
[i
].latency
;
2098 for (i
= 0; i
< LATENCY_BUCKET_SIZE
; i
++)
2099 throtl_log(&td
->service_queue
,
2100 "Latency bucket %d: latency=%ld, valid=%d", i
,
2101 td
->avg_buckets
[i
].latency
, td
->avg_buckets
[i
].valid
);
2104 static inline void throtl_update_latency_buckets(struct throtl_data
*td
)
2109 static void blk_throtl_assoc_bio(struct throtl_grp
*tg
, struct bio
*bio
)
2111 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2113 if (bio
->bi_cg_private
)
2114 blkg_put(tg_to_blkg(bio
->bi_cg_private
));
2115 bio
->bi_cg_private
= tg
;
2116 blkg_get(tg_to_blkg(tg
));
2118 blk_stat_set_issue(&bio
->bi_issue_stat
, bio_sectors(bio
));
2122 bool blk_throtl_bio(struct request_queue
*q
, struct blkcg_gq
*blkg
,
2125 struct throtl_qnode
*qn
= NULL
;
2126 struct throtl_grp
*tg
= blkg_to_tg(blkg
?: q
->root_blkg
);
2127 struct throtl_service_queue
*sq
;
2128 bool rw
= bio_data_dir(bio
);
2129 bool throttled
= false;
2130 struct throtl_data
*td
= tg
->td
;
2132 WARN_ON_ONCE(!rcu_read_lock_held());
2134 /* see throtl_charge_bio() */
2135 if (bio_flagged(bio
, BIO_THROTTLED
) || !tg
->has_rules
[rw
])
2138 spin_lock_irq(q
->queue_lock
);
2140 throtl_update_latency_buckets(td
);
2142 if (unlikely(blk_queue_bypass(q
)))
2145 blk_throtl_assoc_bio(tg
, bio
);
2146 blk_throtl_update_idletime(tg
);
2148 sq
= &tg
->service_queue
;
2152 if (tg
->last_low_overflow_time
[rw
] == 0)
2153 tg
->last_low_overflow_time
[rw
] = jiffies
;
2154 throtl_downgrade_check(tg
);
2155 throtl_upgrade_check(tg
);
2156 /* throtl is FIFO - if bios are already queued, should queue */
2157 if (sq
->nr_queued
[rw
])
2160 /* if above limits, break to queue */
2161 if (!tg_may_dispatch(tg
, bio
, NULL
)) {
2162 tg
->last_low_overflow_time
[rw
] = jiffies
;
2163 if (throtl_can_upgrade(td
, tg
)) {
2164 throtl_upgrade_state(td
);
2170 /* within limits, let's charge and dispatch directly */
2171 throtl_charge_bio(tg
, bio
);
2174 * We need to trim slice even when bios are not being queued
2175 * otherwise it might happen that a bio is not queued for
2176 * a long time and slice keeps on extending and trim is not
2177 * called for a long time. Now if limits are reduced suddenly
2178 * we take into account all the IO dispatched so far at new
2179 * low rate and * newly queued IO gets a really long dispatch
2182 * So keep on trimming slice even if bio is not queued.
2184 throtl_trim_slice(tg
, rw
);
2187 * @bio passed through this layer without being throttled.
2188 * Climb up the ladder. If we''re already at the top, it
2189 * can be executed directly.
2191 qn
= &tg
->qnode_on_parent
[rw
];
2198 /* out-of-limit, queue to @tg */
2199 throtl_log(sq
, "[%c] bio. bdisp=%llu sz=%u bps=%llu iodisp=%u iops=%u queued=%d/%d",
2200 rw
== READ
? 'R' : 'W',
2201 tg
->bytes_disp
[rw
], bio
->bi_iter
.bi_size
,
2202 tg_bps_limit(tg
, rw
),
2203 tg
->io_disp
[rw
], tg_iops_limit(tg
, rw
),
2204 sq
->nr_queued
[READ
], sq
->nr_queued
[WRITE
]);
2206 tg
->last_low_overflow_time
[rw
] = jiffies
;
2208 td
->nr_queued
[rw
]++;
2209 throtl_add_bio_tg(bio
, qn
, tg
);
2213 * Update @tg's dispatch time and force schedule dispatch if @tg
2214 * was empty before @bio. The forced scheduling isn't likely to
2215 * cause undue delay as @bio is likely to be dispatched directly if
2216 * its @tg's disptime is not in the future.
2218 if (tg
->flags
& THROTL_TG_WAS_EMPTY
) {
2219 tg_update_disptime(tg
);
2220 throtl_schedule_next_dispatch(tg
->service_queue
.parent_sq
, true);
2224 spin_unlock_irq(q
->queue_lock
);
2226 bio_set_flag(bio
, BIO_THROTTLED
);
2228 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2229 if (throttled
|| !td
->track_bio_latency
)
2230 bio
->bi_issue_stat
.stat
|= SKIP_LATENCY
;
2235 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2236 static void throtl_track_latency(struct throtl_data
*td
, sector_t size
,
2237 int op
, unsigned long time
)
2239 struct latency_bucket
*latency
;
2242 if (!td
|| td
->limit_index
!= LIMIT_LOW
|| op
!= REQ_OP_READ
||
2243 !blk_queue_nonrot(td
->queue
))
2246 index
= request_bucket_index(size
);
2248 latency
= get_cpu_ptr(td
->latency_buckets
);
2249 latency
[index
].total_latency
+= time
;
2250 latency
[index
].samples
++;
2251 put_cpu_ptr(td
->latency_buckets
);
2254 void blk_throtl_stat_add(struct request
*rq
, u64 time_ns
)
2256 struct request_queue
*q
= rq
->q
;
2257 struct throtl_data
*td
= q
->td
;
2259 throtl_track_latency(td
, blk_stat_size(&rq
->issue_stat
),
2260 req_op(rq
), time_ns
>> 10);
2263 void blk_throtl_bio_endio(struct bio
*bio
)
2265 struct throtl_grp
*tg
;
2267 unsigned long finish_time
;
2268 unsigned long start_time
;
2271 tg
= bio
->bi_cg_private
;
2274 bio
->bi_cg_private
= NULL
;
2276 finish_time_ns
= ktime_get_ns();
2277 tg
->last_finish_time
= finish_time_ns
>> 10;
2279 start_time
= blk_stat_time(&bio
->bi_issue_stat
) >> 10;
2280 finish_time
= __blk_stat_time(finish_time_ns
) >> 10;
2281 if (!start_time
|| finish_time
<= start_time
) {
2282 blkg_put(tg_to_blkg(tg
));
2286 lat
= finish_time
- start_time
;
2287 /* this is only for bio based driver */
2288 if (!(bio
->bi_issue_stat
.stat
& SKIP_LATENCY
))
2289 throtl_track_latency(tg
->td
, blk_stat_size(&bio
->bi_issue_stat
),
2292 if (tg
->latency_target
&& lat
>= tg
->td
->filtered_latency
) {
2294 unsigned int threshold
;
2296 bucket
= request_bucket_index(
2297 blk_stat_size(&bio
->bi_issue_stat
));
2298 threshold
= tg
->td
->avg_buckets
[bucket
].latency
+
2300 if (lat
> threshold
)
2303 * Not race free, could get wrong count, which means cgroups
2309 if (time_after(jiffies
, tg
->bio_cnt_reset_time
) || tg
->bio_cnt
> 1024) {
2310 tg
->bio_cnt_reset_time
= tg
->td
->throtl_slice
+ jiffies
;
2312 tg
->bad_bio_cnt
/= 2;
2315 blkg_put(tg_to_blkg(tg
));
2320 * Dispatch all bios from all children tg's queued on @parent_sq. On
2321 * return, @parent_sq is guaranteed to not have any active children tg's
2322 * and all bios from previously active tg's are on @parent_sq->bio_lists[].
2324 static void tg_drain_bios(struct throtl_service_queue
*parent_sq
)
2326 struct throtl_grp
*tg
;
2328 while ((tg
= throtl_rb_first(parent_sq
))) {
2329 struct throtl_service_queue
*sq
= &tg
->service_queue
;
2332 throtl_dequeue_tg(tg
);
2334 while ((bio
= throtl_peek_queued(&sq
->queued
[READ
])))
2335 tg_dispatch_one_bio(tg
, bio_data_dir(bio
));
2336 while ((bio
= throtl_peek_queued(&sq
->queued
[WRITE
])))
2337 tg_dispatch_one_bio(tg
, bio_data_dir(bio
));
2342 * blk_throtl_drain - drain throttled bios
2343 * @q: request_queue to drain throttled bios for
2345 * Dispatch all currently throttled bios on @q through ->make_request_fn().
2347 void blk_throtl_drain(struct request_queue
*q
)
2348 __releases(q
->queue_lock
) __acquires(q
->queue_lock
)
2350 struct throtl_data
*td
= q
->td
;
2351 struct blkcg_gq
*blkg
;
2352 struct cgroup_subsys_state
*pos_css
;
2356 queue_lockdep_assert_held(q
);
2360 * Drain each tg while doing post-order walk on the blkg tree, so
2361 * that all bios are propagated to td->service_queue. It'd be
2362 * better to walk service_queue tree directly but blkg walk is
2365 blkg_for_each_descendant_post(blkg
, pos_css
, td
->queue
->root_blkg
)
2366 tg_drain_bios(&blkg_to_tg(blkg
)->service_queue
);
2368 /* finally, transfer bios from top-level tg's into the td */
2369 tg_drain_bios(&td
->service_queue
);
2372 spin_unlock_irq(q
->queue_lock
);
2374 /* all bios now should be in td->service_queue, issue them */
2375 for (rw
= READ
; rw
<= WRITE
; rw
++)
2376 while ((bio
= throtl_pop_queued(&td
->service_queue
.queued
[rw
],
2378 generic_make_request(bio
);
2380 spin_lock_irq(q
->queue_lock
);
2383 int blk_throtl_init(struct request_queue
*q
)
2385 struct throtl_data
*td
;
2388 td
= kzalloc_node(sizeof(*td
), GFP_KERNEL
, q
->node
);
2391 td
->latency_buckets
= __alloc_percpu(sizeof(struct latency_bucket
) *
2392 LATENCY_BUCKET_SIZE
, __alignof__(u64
));
2393 if (!td
->latency_buckets
) {
2398 INIT_WORK(&td
->dispatch_work
, blk_throtl_dispatch_work_fn
);
2399 throtl_service_queue_init(&td
->service_queue
);
2404 td
->limit_valid
[LIMIT_MAX
] = true;
2405 td
->limit_index
= LIMIT_MAX
;
2406 td
->low_upgrade_time
= jiffies
;
2407 td
->low_downgrade_time
= jiffies
;
2409 /* activate policy */
2410 ret
= blkcg_activate_policy(q
, &blkcg_policy_throtl
);
2412 free_percpu(td
->latency_buckets
);
2418 void blk_throtl_exit(struct request_queue
*q
)
2421 throtl_shutdown_wq(q
);
2422 blkcg_deactivate_policy(q
, &blkcg_policy_throtl
);
2423 free_percpu(q
->td
->latency_buckets
);
2427 void blk_throtl_register_queue(struct request_queue
*q
)
2429 struct throtl_data
*td
;
2435 if (blk_queue_nonrot(q
)) {
2436 td
->throtl_slice
= DFL_THROTL_SLICE_SSD
;
2437 td
->filtered_latency
= LATENCY_FILTERED_SSD
;
2439 td
->throtl_slice
= DFL_THROTL_SLICE_HD
;
2440 td
->filtered_latency
= LATENCY_FILTERED_HD
;
2441 for (i
= 0; i
< LATENCY_BUCKET_SIZE
; i
++)
2442 td
->avg_buckets
[i
].latency
= DFL_HD_BASELINE_LATENCY
;
2444 #ifndef CONFIG_BLK_DEV_THROTTLING_LOW
2445 /* if no low limit, use previous default */
2446 td
->throtl_slice
= DFL_THROTL_SLICE_HD
;
2449 td
->track_bio_latency
= !q
->mq_ops
&& !q
->request_fn
;
2450 if (!td
->track_bio_latency
)
2451 blk_stat_enable_accounting(q
);
2454 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2455 ssize_t
blk_throtl_sample_time_show(struct request_queue
*q
, char *page
)
2459 return sprintf(page
, "%u\n", jiffies_to_msecs(q
->td
->throtl_slice
));
2462 ssize_t
blk_throtl_sample_time_store(struct request_queue
*q
,
2463 const char *page
, size_t count
)
2470 if (kstrtoul(page
, 10, &v
))
2472 t
= msecs_to_jiffies(v
);
2473 if (t
== 0 || t
> MAX_THROTL_SLICE
)
2475 q
->td
->throtl_slice
= t
;
2480 static int __init
throtl_init(void)
2482 kthrotld_workqueue
= alloc_workqueue("kthrotld", WQ_MEM_RECLAIM
, 0);
2483 if (!kthrotld_workqueue
)
2484 panic("Failed to create kthrotld\n");
2486 return blkcg_policy_register(&blkcg_policy_throtl
);
2489 module_init(throtl_init
);