2 * Interface for controlling IO bandwidth on a request queue
4 * Copyright (C) 2010 Vivek Goyal <vgoyal@redhat.com>
7 #include <linux/module.h>
8 #include <linux/slab.h>
9 #include <linux/blkdev.h>
10 #include <linux/bio.h>
11 #include <linux/blktrace_api.h>
12 #include <linux/blk-cgroup.h>
15 /* Max dispatch from a group in 1 round */
16 static int throtl_grp_quantum
= 8;
18 /* Total max dispatch from all groups in one round */
19 static int throtl_quantum
= 32;
21 /* Throttling is performed over a slice and after that slice is renewed */
22 #define DFL_THROTL_SLICE_HD (HZ / 10)
23 #define DFL_THROTL_SLICE_SSD (HZ / 50)
24 #define MAX_THROTL_SLICE (HZ)
25 #define MAX_IDLE_TIME (5L * 1000 * 1000) /* 5 s */
26 #define MIN_THROTL_BPS (320 * 1024)
27 #define MIN_THROTL_IOPS (10)
28 #define DFL_LATENCY_TARGET (-1L)
29 #define DFL_IDLE_THRESHOLD (0)
30 #define DFL_HD_BASELINE_LATENCY (4000L) /* 4ms */
31 #define LATENCY_FILTERED_SSD (0)
33 * For HD, very small latency comes from sequential IO. Such IO is helpless to
34 * help determine if its IO is impacted by others, hence we ignore the IO
36 #define LATENCY_FILTERED_HD (1000L) /* 1ms */
38 #define SKIP_LATENCY (((u64)1) << BLK_STAT_RES_SHIFT)
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 pending_tree
; /* RB tree of active tgs */
89 struct rb_node
*first_pending
; /* first node in the tree */
90 unsigned int nr_pending
; /* # queued in the tree */
91 unsigned long first_pending_disptime
; /* disptime of the first tg */
92 struct timer_list pending_timer
; /* fires on first_pending_disptime */
96 THROTL_TG_PENDING
= 1 << 0, /* on parent's pending tree */
97 THROTL_TG_WAS_EMPTY
= 1 << 1, /* bio_lists[] became non-empty */
100 #define rb_entry_tg(node) rb_entry((node), struct throtl_grp, rb_node)
109 /* must be the first member */
110 struct blkg_policy_data pd
;
112 /* active throtl group service_queue member */
113 struct rb_node rb_node
;
115 /* throtl_data this group belongs to */
116 struct throtl_data
*td
;
118 /* this group's service queue */
119 struct throtl_service_queue service_queue
;
122 * qnode_on_self is used when bios are directly queued to this
123 * throtl_grp so that local bios compete fairly with bios
124 * dispatched from children. qnode_on_parent is used when bios are
125 * dispatched from this throtl_grp into its parent and will compete
126 * with the sibling qnode_on_parents and the parent's
129 struct throtl_qnode qnode_on_self
[2];
130 struct throtl_qnode qnode_on_parent
[2];
133 * Dispatch time in jiffies. This is the estimated time when group
134 * will unthrottle and is ready to dispatch more bio. It is used as
135 * key to sort active groups in service tree.
137 unsigned long disptime
;
141 /* are there any throtl rules between this group and td? */
144 /* internally used bytes per second rate limits */
145 uint64_t bps
[2][LIMIT_CNT
];
146 /* user configured bps limits */
147 uint64_t bps_conf
[2][LIMIT_CNT
];
149 /* internally used IOPS limits */
150 unsigned int iops
[2][LIMIT_CNT
];
151 /* user configured IOPS limits */
152 unsigned int iops_conf
[2][LIMIT_CNT
];
154 /* Number of bytes disptached in current slice */
155 uint64_t bytes_disp
[2];
156 /* Number of bio's dispatched in current slice */
157 unsigned int io_disp
[2];
159 unsigned long last_low_overflow_time
[2];
161 uint64_t last_bytes_disp
[2];
162 unsigned int last_io_disp
[2];
164 unsigned long last_check_time
;
166 unsigned long latency_target
; /* us */
167 unsigned long latency_target_conf
; /* us */
168 /* When did we start a new slice */
169 unsigned long slice_start
[2];
170 unsigned long slice_end
[2];
172 unsigned long last_finish_time
; /* ns / 1024 */
173 unsigned long checked_last_finish_time
; /* ns / 1024 */
174 unsigned long avg_idletime
; /* ns / 1024 */
175 unsigned long idletime_threshold
; /* us */
176 unsigned long idletime_threshold_conf
; /* us */
178 unsigned int bio_cnt
; /* total bios */
179 unsigned int bad_bio_cnt
; /* bios exceeding latency threshold */
180 unsigned long bio_cnt_reset_time
;
183 /* We measure latency for request size from <= 4k to >= 1M */
184 #define LATENCY_BUCKET_SIZE 9
186 struct latency_bucket
{
187 unsigned long total_latency
; /* ns / 1024 */
191 struct avg_latency_bucket
{
192 unsigned long latency
; /* ns / 1024 */
198 /* service tree for active throtl groups */
199 struct throtl_service_queue service_queue
;
201 struct request_queue
*queue
;
203 /* Total Number of queued bios on READ and WRITE lists */
204 unsigned int nr_queued
[2];
206 unsigned int throtl_slice
;
208 /* Work for dispatching throttled bios */
209 struct work_struct dispatch_work
;
210 unsigned int limit_index
;
211 bool limit_valid
[LIMIT_CNT
];
213 unsigned long low_upgrade_time
;
214 unsigned long low_downgrade_time
;
218 struct latency_bucket tmp_buckets
[LATENCY_BUCKET_SIZE
];
219 struct avg_latency_bucket avg_buckets
[LATENCY_BUCKET_SIZE
];
220 struct latency_bucket __percpu
*latency_buckets
;
221 unsigned long last_calculate_time
;
222 unsigned long filtered_latency
;
224 bool track_bio_latency
;
227 static void throtl_pending_timer_fn(unsigned long arg
);
229 static inline struct throtl_grp
*pd_to_tg(struct blkg_policy_data
*pd
)
231 return pd
? container_of(pd
, struct throtl_grp
, pd
) : NULL
;
234 static inline struct throtl_grp
*blkg_to_tg(struct blkcg_gq
*blkg
)
236 return pd_to_tg(blkg_to_pd(blkg
, &blkcg_policy_throtl
));
239 static inline struct blkcg_gq
*tg_to_blkg(struct throtl_grp
*tg
)
241 return pd_to_blkg(&tg
->pd
);
245 * sq_to_tg - return the throl_grp the specified service queue belongs to
246 * @sq: the throtl_service_queue of interest
248 * Return the throtl_grp @sq belongs to. If @sq is the top-level one
249 * embedded in throtl_data, %NULL is returned.
251 static struct throtl_grp
*sq_to_tg(struct throtl_service_queue
*sq
)
253 if (sq
&& sq
->parent_sq
)
254 return container_of(sq
, struct throtl_grp
, service_queue
);
260 * sq_to_td - return throtl_data the specified service queue belongs to
261 * @sq: the throtl_service_queue of interest
263 * A service_queue can be embedded in either a throtl_grp or throtl_data.
264 * Determine the associated throtl_data accordingly and return it.
266 static struct throtl_data
*sq_to_td(struct throtl_service_queue
*sq
)
268 struct throtl_grp
*tg
= sq_to_tg(sq
);
273 return container_of(sq
, struct throtl_data
, service_queue
);
277 * cgroup's limit in LIMIT_MAX is scaled if low limit is set. This scale is to
278 * make the IO dispatch more smooth.
279 * Scale up: linearly scale up according to lapsed time since upgrade. For
280 * every throtl_slice, the limit scales up 1/2 .low limit till the
281 * limit hits .max limit
282 * Scale down: exponentially scale down if a cgroup doesn't hit its .low limit
284 static uint64_t throtl_adjusted_limit(uint64_t low
, struct throtl_data
*td
)
286 /* arbitrary value to avoid too big scale */
287 if (td
->scale
< 4096 && time_after_eq(jiffies
,
288 td
->low_upgrade_time
+ td
->scale
* td
->throtl_slice
))
289 td
->scale
= (jiffies
- td
->low_upgrade_time
) / td
->throtl_slice
;
291 return low
+ (low
>> 1) * td
->scale
;
294 static uint64_t tg_bps_limit(struct throtl_grp
*tg
, int rw
)
296 struct blkcg_gq
*blkg
= tg_to_blkg(tg
);
297 struct throtl_data
*td
;
300 if (cgroup_subsys_on_dfl(io_cgrp_subsys
) && !blkg
->parent
)
304 ret
= tg
->bps
[rw
][td
->limit_index
];
305 if (ret
== 0 && td
->limit_index
== LIMIT_LOW
) {
306 /* intermediate node or iops isn't 0 */
307 if (!list_empty(&blkg
->blkcg
->css
.children
) ||
308 tg
->iops
[rw
][td
->limit_index
])
311 return MIN_THROTL_BPS
;
314 if (td
->limit_index
== LIMIT_MAX
&& tg
->bps
[rw
][LIMIT_LOW
] &&
315 tg
->bps
[rw
][LIMIT_LOW
] != tg
->bps
[rw
][LIMIT_MAX
]) {
318 adjusted
= throtl_adjusted_limit(tg
->bps
[rw
][LIMIT_LOW
], td
);
319 ret
= min(tg
->bps
[rw
][LIMIT_MAX
], adjusted
);
324 static unsigned int tg_iops_limit(struct throtl_grp
*tg
, int rw
)
326 struct blkcg_gq
*blkg
= tg_to_blkg(tg
);
327 struct throtl_data
*td
;
330 if (cgroup_subsys_on_dfl(io_cgrp_subsys
) && !blkg
->parent
)
334 ret
= tg
->iops
[rw
][td
->limit_index
];
335 if (ret
== 0 && tg
->td
->limit_index
== LIMIT_LOW
) {
336 /* intermediate node or bps isn't 0 */
337 if (!list_empty(&blkg
->blkcg
->css
.children
) ||
338 tg
->bps
[rw
][td
->limit_index
])
341 return MIN_THROTL_IOPS
;
344 if (td
->limit_index
== LIMIT_MAX
&& tg
->iops
[rw
][LIMIT_LOW
] &&
345 tg
->iops
[rw
][LIMIT_LOW
] != tg
->iops
[rw
][LIMIT_MAX
]) {
348 adjusted
= throtl_adjusted_limit(tg
->iops
[rw
][LIMIT_LOW
], td
);
349 if (adjusted
> UINT_MAX
)
351 ret
= min_t(unsigned int, tg
->iops
[rw
][LIMIT_MAX
], adjusted
);
356 #define request_bucket_index(sectors) \
357 clamp_t(int, order_base_2(sectors) - 3, 0, LATENCY_BUCKET_SIZE - 1)
360 * throtl_log - log debug message via blktrace
361 * @sq: the service_queue being reported
362 * @fmt: printf format string
365 * The messages are prefixed with "throtl BLKG_NAME" if @sq belongs to a
366 * throtl_grp; otherwise, just "throtl".
368 #define throtl_log(sq, fmt, args...) do { \
369 struct throtl_grp *__tg = sq_to_tg((sq)); \
370 struct throtl_data *__td = sq_to_td((sq)); \
373 if (likely(!blk_trace_note_message_enabled(__td->queue))) \
378 blkg_path(tg_to_blkg(__tg), __pbuf, sizeof(__pbuf)); \
379 blk_add_trace_msg(__td->queue, "throtl %s " fmt, __pbuf, ##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
= list_first_entry(queued
, struct throtl_qnode
, node
);
429 if (list_empty(queued
))
432 bio
= bio_list_peek(&qn
->bios
);
438 * throtl_pop_queued - pop the first bio form a qnode list
439 * @queued: the qnode list to pop a bio from
440 * @tg_to_put: optional out argument for throtl_grp to put
442 * Pop the first bio from the qnode list @queued. After popping, the first
443 * qnode is removed from @queued if empty or moved to the end of @queued so
444 * that the popping order is round-robin.
446 * When the first qnode is removed, its associated throtl_grp should be put
447 * too. If @tg_to_put is NULL, this function automatically puts it;
448 * otherwise, *@tg_to_put is set to the throtl_grp to put and the caller is
449 * responsible for putting it.
451 static struct bio
*throtl_pop_queued(struct list_head
*queued
,
452 struct throtl_grp
**tg_to_put
)
454 struct throtl_qnode
*qn
= list_first_entry(queued
, struct throtl_qnode
, node
);
457 if (list_empty(queued
))
460 bio
= bio_list_pop(&qn
->bios
);
463 if (bio_list_empty(&qn
->bios
)) {
464 list_del_init(&qn
->node
);
468 blkg_put(tg_to_blkg(qn
->tg
));
470 list_move_tail(&qn
->node
, queued
);
476 /* init a service_queue, assumes the caller zeroed it */
477 static void throtl_service_queue_init(struct throtl_service_queue
*sq
)
479 INIT_LIST_HEAD(&sq
->queued
[0]);
480 INIT_LIST_HEAD(&sq
->queued
[1]);
481 sq
->pending_tree
= RB_ROOT
;
482 setup_timer(&sq
->pending_timer
, throtl_pending_timer_fn
,
486 static struct blkg_policy_data
*throtl_pd_alloc(gfp_t gfp
, int node
)
488 struct throtl_grp
*tg
;
491 tg
= kzalloc_node(sizeof(*tg
), gfp
, node
);
495 throtl_service_queue_init(&tg
->service_queue
);
497 for (rw
= READ
; rw
<= WRITE
; rw
++) {
498 throtl_qnode_init(&tg
->qnode_on_self
[rw
], tg
);
499 throtl_qnode_init(&tg
->qnode_on_parent
[rw
], tg
);
502 RB_CLEAR_NODE(&tg
->rb_node
);
503 tg
->bps
[READ
][LIMIT_MAX
] = U64_MAX
;
504 tg
->bps
[WRITE
][LIMIT_MAX
] = U64_MAX
;
505 tg
->iops
[READ
][LIMIT_MAX
] = UINT_MAX
;
506 tg
->iops
[WRITE
][LIMIT_MAX
] = UINT_MAX
;
507 tg
->bps_conf
[READ
][LIMIT_MAX
] = U64_MAX
;
508 tg
->bps_conf
[WRITE
][LIMIT_MAX
] = U64_MAX
;
509 tg
->iops_conf
[READ
][LIMIT_MAX
] = UINT_MAX
;
510 tg
->iops_conf
[WRITE
][LIMIT_MAX
] = UINT_MAX
;
511 /* LIMIT_LOW will have default value 0 */
513 tg
->latency_target
= DFL_LATENCY_TARGET
;
514 tg
->latency_target_conf
= DFL_LATENCY_TARGET
;
515 tg
->idletime_threshold
= DFL_IDLE_THRESHOLD
;
516 tg
->idletime_threshold_conf
= DFL_IDLE_THRESHOLD
;
521 static void throtl_pd_init(struct blkg_policy_data
*pd
)
523 struct throtl_grp
*tg
= pd_to_tg(pd
);
524 struct blkcg_gq
*blkg
= tg_to_blkg(tg
);
525 struct throtl_data
*td
= blkg
->q
->td
;
526 struct throtl_service_queue
*sq
= &tg
->service_queue
;
529 * If on the default hierarchy, we switch to properly hierarchical
530 * behavior where limits on a given throtl_grp are applied to the
531 * whole subtree rather than just the group itself. e.g. If 16M
532 * read_bps limit is set on the root group, the whole system can't
533 * exceed 16M for the device.
535 * If not on the default hierarchy, the broken flat hierarchy
536 * behavior is retained where all throtl_grps are treated as if
537 * they're all separate root groups right below throtl_data.
538 * Limits of a group don't interact with limits of other groups
539 * regardless of the position of the group in the hierarchy.
541 sq
->parent_sq
= &td
->service_queue
;
542 if (cgroup_subsys_on_dfl(io_cgrp_subsys
) && blkg
->parent
)
543 sq
->parent_sq
= &blkg_to_tg(blkg
->parent
)->service_queue
;
548 * Set has_rules[] if @tg or any of its parents have limits configured.
549 * This doesn't require walking up to the top of the hierarchy as the
550 * parent's has_rules[] is guaranteed to be correct.
552 static void tg_update_has_rules(struct throtl_grp
*tg
)
554 struct throtl_grp
*parent_tg
= sq_to_tg(tg
->service_queue
.parent_sq
);
555 struct throtl_data
*td
= tg
->td
;
558 for (rw
= READ
; rw
<= WRITE
; rw
++)
559 tg
->has_rules
[rw
] = (parent_tg
&& parent_tg
->has_rules
[rw
]) ||
560 (td
->limit_valid
[td
->limit_index
] &&
561 (tg_bps_limit(tg
, rw
) != U64_MAX
||
562 tg_iops_limit(tg
, rw
) != UINT_MAX
));
565 static void throtl_pd_online(struct blkg_policy_data
*pd
)
567 struct throtl_grp
*tg
= pd_to_tg(pd
);
569 * We don't want new groups to escape the limits of its ancestors.
570 * Update has_rules[] after a new group is brought online.
572 tg_update_has_rules(tg
);
575 static void blk_throtl_update_limit_valid(struct throtl_data
*td
)
577 struct cgroup_subsys_state
*pos_css
;
578 struct blkcg_gq
*blkg
;
579 bool low_valid
= false;
582 blkg_for_each_descendant_post(blkg
, pos_css
, td
->queue
->root_blkg
) {
583 struct throtl_grp
*tg
= blkg_to_tg(blkg
);
585 if (tg
->bps
[READ
][LIMIT_LOW
] || tg
->bps
[WRITE
][LIMIT_LOW
] ||
586 tg
->iops
[READ
][LIMIT_LOW
] || tg
->iops
[WRITE
][LIMIT_LOW
])
591 td
->limit_valid
[LIMIT_LOW
] = low_valid
;
594 static void throtl_upgrade_state(struct throtl_data
*td
);
595 static void throtl_pd_offline(struct blkg_policy_data
*pd
)
597 struct throtl_grp
*tg
= pd_to_tg(pd
);
599 tg
->bps
[READ
][LIMIT_LOW
] = 0;
600 tg
->bps
[WRITE
][LIMIT_LOW
] = 0;
601 tg
->iops
[READ
][LIMIT_LOW
] = 0;
602 tg
->iops
[WRITE
][LIMIT_LOW
] = 0;
604 blk_throtl_update_limit_valid(tg
->td
);
606 if (!tg
->td
->limit_valid
[tg
->td
->limit_index
])
607 throtl_upgrade_state(tg
->td
);
610 static void throtl_pd_free(struct blkg_policy_data
*pd
)
612 struct throtl_grp
*tg
= pd_to_tg(pd
);
614 del_timer_sync(&tg
->service_queue
.pending_timer
);
618 static struct throtl_grp
*
619 throtl_rb_first(struct throtl_service_queue
*parent_sq
)
621 /* Service tree is empty */
622 if (!parent_sq
->nr_pending
)
625 if (!parent_sq
->first_pending
)
626 parent_sq
->first_pending
= rb_first(&parent_sq
->pending_tree
);
628 if (parent_sq
->first_pending
)
629 return rb_entry_tg(parent_sq
->first_pending
);
634 static void rb_erase_init(struct rb_node
*n
, struct rb_root
*root
)
640 static void throtl_rb_erase(struct rb_node
*n
,
641 struct throtl_service_queue
*parent_sq
)
643 if (parent_sq
->first_pending
== n
)
644 parent_sq
->first_pending
= NULL
;
645 rb_erase_init(n
, &parent_sq
->pending_tree
);
646 --parent_sq
->nr_pending
;
649 static void update_min_dispatch_time(struct throtl_service_queue
*parent_sq
)
651 struct throtl_grp
*tg
;
653 tg
= throtl_rb_first(parent_sq
);
657 parent_sq
->first_pending_disptime
= tg
->disptime
;
660 static void tg_service_queue_add(struct throtl_grp
*tg
)
662 struct throtl_service_queue
*parent_sq
= tg
->service_queue
.parent_sq
;
663 struct rb_node
**node
= &parent_sq
->pending_tree
.rb_node
;
664 struct rb_node
*parent
= NULL
;
665 struct throtl_grp
*__tg
;
666 unsigned long key
= tg
->disptime
;
669 while (*node
!= NULL
) {
671 __tg
= rb_entry_tg(parent
);
673 if (time_before(key
, __tg
->disptime
))
674 node
= &parent
->rb_left
;
676 node
= &parent
->rb_right
;
682 parent_sq
->first_pending
= &tg
->rb_node
;
684 rb_link_node(&tg
->rb_node
, parent
, node
);
685 rb_insert_color(&tg
->rb_node
, &parent_sq
->pending_tree
);
688 static void __throtl_enqueue_tg(struct throtl_grp
*tg
)
690 tg_service_queue_add(tg
);
691 tg
->flags
|= THROTL_TG_PENDING
;
692 tg
->service_queue
.parent_sq
->nr_pending
++;
695 static void throtl_enqueue_tg(struct throtl_grp
*tg
)
697 if (!(tg
->flags
& THROTL_TG_PENDING
))
698 __throtl_enqueue_tg(tg
);
701 static void __throtl_dequeue_tg(struct throtl_grp
*tg
)
703 throtl_rb_erase(&tg
->rb_node
, tg
->service_queue
.parent_sq
);
704 tg
->flags
&= ~THROTL_TG_PENDING
;
707 static void throtl_dequeue_tg(struct throtl_grp
*tg
)
709 if (tg
->flags
& THROTL_TG_PENDING
)
710 __throtl_dequeue_tg(tg
);
713 /* Call with queue lock held */
714 static void throtl_schedule_pending_timer(struct throtl_service_queue
*sq
,
715 unsigned long expires
)
717 unsigned long max_expire
= jiffies
+ 8 * sq_to_td(sq
)->throtl_slice
;
720 * Since we are adjusting the throttle limit dynamically, the sleep
721 * time calculated according to previous limit might be invalid. It's
722 * possible the cgroup sleep time is very long and no other cgroups
723 * have IO running so notify the limit changes. Make sure the cgroup
724 * doesn't sleep too long to avoid the missed notification.
726 if (time_after(expires
, max_expire
))
727 expires
= max_expire
;
728 mod_timer(&sq
->pending_timer
, expires
);
729 throtl_log(sq
, "schedule timer. delay=%lu jiffies=%lu",
730 expires
- jiffies
, jiffies
);
734 * throtl_schedule_next_dispatch - schedule the next dispatch cycle
735 * @sq: the service_queue to schedule dispatch for
736 * @force: force scheduling
738 * Arm @sq->pending_timer so that the next dispatch cycle starts on the
739 * dispatch time of the first pending child. Returns %true if either timer
740 * is armed or there's no pending child left. %false if the current
741 * dispatch window is still open and the caller should continue
744 * If @force is %true, the dispatch timer is always scheduled and this
745 * function is guaranteed to return %true. This is to be used when the
746 * caller can't dispatch itself and needs to invoke pending_timer
747 * unconditionally. Note that forced scheduling is likely to induce short
748 * delay before dispatch starts even if @sq->first_pending_disptime is not
749 * in the future and thus shouldn't be used in hot paths.
751 static bool throtl_schedule_next_dispatch(struct throtl_service_queue
*sq
,
754 /* any pending children left? */
758 update_min_dispatch_time(sq
);
760 /* is the next dispatch time in the future? */
761 if (force
|| time_after(sq
->first_pending_disptime
, jiffies
)) {
762 throtl_schedule_pending_timer(sq
, sq
->first_pending_disptime
);
766 /* tell the caller to continue dispatching */
770 static inline void throtl_start_new_slice_with_credit(struct throtl_grp
*tg
,
771 bool rw
, unsigned long start
)
773 tg
->bytes_disp
[rw
] = 0;
777 * Previous slice has expired. We must have trimmed it after last
778 * bio dispatch. That means since start of last slice, we never used
779 * that bandwidth. Do try to make use of that bandwidth while giving
782 if (time_after_eq(start
, tg
->slice_start
[rw
]))
783 tg
->slice_start
[rw
] = start
;
785 tg
->slice_end
[rw
] = jiffies
+ tg
->td
->throtl_slice
;
786 throtl_log(&tg
->service_queue
,
787 "[%c] new slice with credit start=%lu end=%lu jiffies=%lu",
788 rw
== READ
? 'R' : 'W', tg
->slice_start
[rw
],
789 tg
->slice_end
[rw
], jiffies
);
792 static inline void throtl_start_new_slice(struct throtl_grp
*tg
, bool rw
)
794 tg
->bytes_disp
[rw
] = 0;
796 tg
->slice_start
[rw
] = jiffies
;
797 tg
->slice_end
[rw
] = jiffies
+ tg
->td
->throtl_slice
;
798 throtl_log(&tg
->service_queue
,
799 "[%c] new slice start=%lu end=%lu jiffies=%lu",
800 rw
== READ
? 'R' : 'W', tg
->slice_start
[rw
],
801 tg
->slice_end
[rw
], jiffies
);
804 static inline void throtl_set_slice_end(struct throtl_grp
*tg
, bool rw
,
805 unsigned long jiffy_end
)
807 tg
->slice_end
[rw
] = roundup(jiffy_end
, tg
->td
->throtl_slice
);
810 static inline void throtl_extend_slice(struct throtl_grp
*tg
, bool rw
,
811 unsigned long jiffy_end
)
813 tg
->slice_end
[rw
] = roundup(jiffy_end
, tg
->td
->throtl_slice
);
814 throtl_log(&tg
->service_queue
,
815 "[%c] extend slice start=%lu end=%lu jiffies=%lu",
816 rw
== READ
? 'R' : 'W', tg
->slice_start
[rw
],
817 tg
->slice_end
[rw
], jiffies
);
820 /* Determine if previously allocated or extended slice is complete or not */
821 static bool throtl_slice_used(struct throtl_grp
*tg
, bool rw
)
823 if (time_in_range(jiffies
, tg
->slice_start
[rw
], tg
->slice_end
[rw
]))
829 /* Trim the used slices and adjust slice start accordingly */
830 static inline void throtl_trim_slice(struct throtl_grp
*tg
, bool rw
)
832 unsigned long nr_slices
, time_elapsed
, io_trim
;
835 BUG_ON(time_before(tg
->slice_end
[rw
], tg
->slice_start
[rw
]));
838 * If bps are unlimited (-1), then time slice don't get
839 * renewed. Don't try to trim the slice if slice is used. A new
840 * slice will start when appropriate.
842 if (throtl_slice_used(tg
, rw
))
846 * A bio has been dispatched. Also adjust slice_end. It might happen
847 * that initially cgroup limit was very low resulting in high
848 * slice_end, but later limit was bumped up and bio was dispached
849 * sooner, then we need to reduce slice_end. A high bogus slice_end
850 * is bad because it does not allow new slice to start.
853 throtl_set_slice_end(tg
, rw
, jiffies
+ tg
->td
->throtl_slice
);
855 time_elapsed
= jiffies
- tg
->slice_start
[rw
];
857 nr_slices
= time_elapsed
/ tg
->td
->throtl_slice
;
861 tmp
= tg_bps_limit(tg
, rw
) * tg
->td
->throtl_slice
* nr_slices
;
865 io_trim
= (tg_iops_limit(tg
, rw
) * tg
->td
->throtl_slice
* nr_slices
) /
868 if (!bytes_trim
&& !io_trim
)
871 if (tg
->bytes_disp
[rw
] >= bytes_trim
)
872 tg
->bytes_disp
[rw
] -= bytes_trim
;
874 tg
->bytes_disp
[rw
] = 0;
876 if (tg
->io_disp
[rw
] >= io_trim
)
877 tg
->io_disp
[rw
] -= io_trim
;
881 tg
->slice_start
[rw
] += nr_slices
* tg
->td
->throtl_slice
;
883 throtl_log(&tg
->service_queue
,
884 "[%c] trim slice nr=%lu bytes=%llu io=%lu start=%lu end=%lu jiffies=%lu",
885 rw
== READ
? 'R' : 'W', nr_slices
, bytes_trim
, io_trim
,
886 tg
->slice_start
[rw
], tg
->slice_end
[rw
], jiffies
);
889 static bool tg_with_in_iops_limit(struct throtl_grp
*tg
, struct bio
*bio
,
892 bool rw
= bio_data_dir(bio
);
893 unsigned int io_allowed
;
894 unsigned long jiffy_elapsed
, jiffy_wait
, jiffy_elapsed_rnd
;
897 jiffy_elapsed
= jiffy_elapsed_rnd
= jiffies
- tg
->slice_start
[rw
];
899 /* Slice has just started. Consider one slice interval */
901 jiffy_elapsed_rnd
= tg
->td
->throtl_slice
;
903 jiffy_elapsed_rnd
= roundup(jiffy_elapsed_rnd
, tg
->td
->throtl_slice
);
906 * jiffy_elapsed_rnd should not be a big value as minimum iops can be
907 * 1 then at max jiffy elapsed should be equivalent of 1 second as we
908 * will allow dispatch after 1 second and after that slice should
912 tmp
= (u64
)tg_iops_limit(tg
, rw
) * jiffy_elapsed_rnd
;
916 io_allowed
= UINT_MAX
;
920 if (tg
->io_disp
[rw
] + 1 <= io_allowed
) {
926 /* Calc approx time to dispatch */
927 jiffy_wait
= ((tg
->io_disp
[rw
] + 1) * HZ
) / tg_iops_limit(tg
, rw
) + 1;
929 if (jiffy_wait
> jiffy_elapsed
)
930 jiffy_wait
= jiffy_wait
- jiffy_elapsed
;
939 static bool tg_with_in_bps_limit(struct throtl_grp
*tg
, struct bio
*bio
,
942 bool rw
= bio_data_dir(bio
);
943 u64 bytes_allowed
, extra_bytes
, tmp
;
944 unsigned long jiffy_elapsed
, jiffy_wait
, jiffy_elapsed_rnd
;
945 unsigned int bio_size
= throtl_bio_data_size(bio
);
947 jiffy_elapsed
= jiffy_elapsed_rnd
= jiffies
- tg
->slice_start
[rw
];
949 /* Slice has just started. Consider one slice interval */
951 jiffy_elapsed_rnd
= tg
->td
->throtl_slice
;
953 jiffy_elapsed_rnd
= roundup(jiffy_elapsed_rnd
, tg
->td
->throtl_slice
);
955 tmp
= tg_bps_limit(tg
, rw
) * jiffy_elapsed_rnd
;
959 if (tg
->bytes_disp
[rw
] + bio_size
<= bytes_allowed
) {
965 /* Calc approx time to dispatch */
966 extra_bytes
= tg
->bytes_disp
[rw
] + bio_size
- bytes_allowed
;
967 jiffy_wait
= div64_u64(extra_bytes
* HZ
, tg_bps_limit(tg
, rw
));
973 * This wait time is without taking into consideration the rounding
974 * up we did. Add that time also.
976 jiffy_wait
= jiffy_wait
+ (jiffy_elapsed_rnd
- jiffy_elapsed
);
983 * Returns whether one can dispatch a bio or not. Also returns approx number
984 * of jiffies to wait before this bio is with-in IO rate and can be dispatched
986 static bool tg_may_dispatch(struct throtl_grp
*tg
, struct bio
*bio
,
989 bool rw
= bio_data_dir(bio
);
990 unsigned long bps_wait
= 0, iops_wait
= 0, max_wait
= 0;
993 * Currently whole state machine of group depends on first bio
994 * queued in the group bio list. So one should not be calling
995 * this function with a different bio if there are other bios
998 BUG_ON(tg
->service_queue
.nr_queued
[rw
] &&
999 bio
!= throtl_peek_queued(&tg
->service_queue
.queued
[rw
]));
1001 /* If tg->bps = -1, then BW is unlimited */
1002 if (tg_bps_limit(tg
, rw
) == U64_MAX
&&
1003 tg_iops_limit(tg
, rw
) == UINT_MAX
) {
1010 * If previous slice expired, start a new one otherwise renew/extend
1011 * existing slice to make sure it is at least throtl_slice interval
1012 * long since now. New slice is started only for empty throttle group.
1013 * If there is queued bio, that means there should be an active
1014 * slice and it should be extended instead.
1016 if (throtl_slice_used(tg
, rw
) && !(tg
->service_queue
.nr_queued
[rw
]))
1017 throtl_start_new_slice(tg
, rw
);
1019 if (time_before(tg
->slice_end
[rw
],
1020 jiffies
+ tg
->td
->throtl_slice
))
1021 throtl_extend_slice(tg
, rw
,
1022 jiffies
+ tg
->td
->throtl_slice
);
1025 if (tg_with_in_bps_limit(tg
, bio
, &bps_wait
) &&
1026 tg_with_in_iops_limit(tg
, bio
, &iops_wait
)) {
1032 max_wait
= max(bps_wait
, iops_wait
);
1037 if (time_before(tg
->slice_end
[rw
], jiffies
+ max_wait
))
1038 throtl_extend_slice(tg
, rw
, jiffies
+ max_wait
);
1043 static void throtl_charge_bio(struct throtl_grp
*tg
, struct bio
*bio
)
1045 bool rw
= bio_data_dir(bio
);
1046 unsigned int bio_size
= throtl_bio_data_size(bio
);
1048 /* Charge the bio to the group */
1049 tg
->bytes_disp
[rw
] += bio_size
;
1051 tg
->last_bytes_disp
[rw
] += bio_size
;
1052 tg
->last_io_disp
[rw
]++;
1055 * BIO_THROTTLED is used to prevent the same bio to be throttled
1056 * more than once as a throttled bio will go through blk-throtl the
1057 * second time when it eventually gets issued. Set it when a bio
1058 * is being charged to a tg.
1060 if (!bio_flagged(bio
, BIO_THROTTLED
))
1061 bio_set_flag(bio
, BIO_THROTTLED
);
1065 * throtl_add_bio_tg - add a bio to the specified throtl_grp
1068 * @tg: the target throtl_grp
1070 * Add @bio to @tg's service_queue using @qn. If @qn is not specified,
1071 * tg->qnode_on_self[] is used.
1073 static void throtl_add_bio_tg(struct bio
*bio
, struct throtl_qnode
*qn
,
1074 struct throtl_grp
*tg
)
1076 struct throtl_service_queue
*sq
= &tg
->service_queue
;
1077 bool rw
= bio_data_dir(bio
);
1080 qn
= &tg
->qnode_on_self
[rw
];
1083 * If @tg doesn't currently have any bios queued in the same
1084 * direction, queueing @bio can change when @tg should be
1085 * dispatched. Mark that @tg was empty. This is automatically
1086 * cleaered on the next tg_update_disptime().
1088 if (!sq
->nr_queued
[rw
])
1089 tg
->flags
|= THROTL_TG_WAS_EMPTY
;
1091 throtl_qnode_add_bio(bio
, qn
, &sq
->queued
[rw
]);
1093 sq
->nr_queued
[rw
]++;
1094 throtl_enqueue_tg(tg
);
1097 static void tg_update_disptime(struct throtl_grp
*tg
)
1099 struct throtl_service_queue
*sq
= &tg
->service_queue
;
1100 unsigned long read_wait
= -1, write_wait
= -1, min_wait
= -1, disptime
;
1103 bio
= throtl_peek_queued(&sq
->queued
[READ
]);
1105 tg_may_dispatch(tg
, bio
, &read_wait
);
1107 bio
= throtl_peek_queued(&sq
->queued
[WRITE
]);
1109 tg_may_dispatch(tg
, bio
, &write_wait
);
1111 min_wait
= min(read_wait
, write_wait
);
1112 disptime
= jiffies
+ min_wait
;
1114 /* Update dispatch time */
1115 throtl_dequeue_tg(tg
);
1116 tg
->disptime
= disptime
;
1117 throtl_enqueue_tg(tg
);
1119 /* see throtl_add_bio_tg() */
1120 tg
->flags
&= ~THROTL_TG_WAS_EMPTY
;
1123 static void start_parent_slice_with_credit(struct throtl_grp
*child_tg
,
1124 struct throtl_grp
*parent_tg
, bool rw
)
1126 if (throtl_slice_used(parent_tg
, rw
)) {
1127 throtl_start_new_slice_with_credit(parent_tg
, rw
,
1128 child_tg
->slice_start
[rw
]);
1133 static void tg_dispatch_one_bio(struct throtl_grp
*tg
, bool rw
)
1135 struct throtl_service_queue
*sq
= &tg
->service_queue
;
1136 struct throtl_service_queue
*parent_sq
= sq
->parent_sq
;
1137 struct throtl_grp
*parent_tg
= sq_to_tg(parent_sq
);
1138 struct throtl_grp
*tg_to_put
= NULL
;
1142 * @bio is being transferred from @tg to @parent_sq. Popping a bio
1143 * from @tg may put its reference and @parent_sq might end up
1144 * getting released prematurely. Remember the tg to put and put it
1145 * after @bio is transferred to @parent_sq.
1147 bio
= throtl_pop_queued(&sq
->queued
[rw
], &tg_to_put
);
1148 sq
->nr_queued
[rw
]--;
1150 throtl_charge_bio(tg
, bio
);
1153 * If our parent is another tg, we just need to transfer @bio to
1154 * the parent using throtl_add_bio_tg(). If our parent is
1155 * @td->service_queue, @bio is ready to be issued. Put it on its
1156 * bio_lists[] and decrease total number queued. The caller is
1157 * responsible for issuing these bios.
1160 throtl_add_bio_tg(bio
, &tg
->qnode_on_parent
[rw
], parent_tg
);
1161 start_parent_slice_with_credit(tg
, parent_tg
, rw
);
1163 throtl_qnode_add_bio(bio
, &tg
->qnode_on_parent
[rw
],
1164 &parent_sq
->queued
[rw
]);
1165 BUG_ON(tg
->td
->nr_queued
[rw
] <= 0);
1166 tg
->td
->nr_queued
[rw
]--;
1169 throtl_trim_slice(tg
, rw
);
1172 blkg_put(tg_to_blkg(tg_to_put
));
1175 static int throtl_dispatch_tg(struct throtl_grp
*tg
)
1177 struct throtl_service_queue
*sq
= &tg
->service_queue
;
1178 unsigned int nr_reads
= 0, nr_writes
= 0;
1179 unsigned int max_nr_reads
= throtl_grp_quantum
*3/4;
1180 unsigned int max_nr_writes
= throtl_grp_quantum
- max_nr_reads
;
1183 /* Try to dispatch 75% READS and 25% WRITES */
1185 while ((bio
= throtl_peek_queued(&sq
->queued
[READ
])) &&
1186 tg_may_dispatch(tg
, bio
, NULL
)) {
1188 tg_dispatch_one_bio(tg
, bio_data_dir(bio
));
1191 if (nr_reads
>= max_nr_reads
)
1195 while ((bio
= throtl_peek_queued(&sq
->queued
[WRITE
])) &&
1196 tg_may_dispatch(tg
, bio
, NULL
)) {
1198 tg_dispatch_one_bio(tg
, bio_data_dir(bio
));
1201 if (nr_writes
>= max_nr_writes
)
1205 return nr_reads
+ nr_writes
;
1208 static int throtl_select_dispatch(struct throtl_service_queue
*parent_sq
)
1210 unsigned int nr_disp
= 0;
1213 struct throtl_grp
*tg
= throtl_rb_first(parent_sq
);
1214 struct throtl_service_queue
*sq
= &tg
->service_queue
;
1219 if (time_before(jiffies
, tg
->disptime
))
1222 throtl_dequeue_tg(tg
);
1224 nr_disp
+= throtl_dispatch_tg(tg
);
1226 if (sq
->nr_queued
[0] || sq
->nr_queued
[1])
1227 tg_update_disptime(tg
);
1229 if (nr_disp
>= throtl_quantum
)
1236 static bool throtl_can_upgrade(struct throtl_data
*td
,
1237 struct throtl_grp
*this_tg
);
1239 * throtl_pending_timer_fn - timer function for service_queue->pending_timer
1240 * @arg: the throtl_service_queue being serviced
1242 * This timer is armed when a child throtl_grp with active bio's become
1243 * pending and queued on the service_queue's pending_tree and expires when
1244 * the first child throtl_grp should be dispatched. This function
1245 * dispatches bio's from the children throtl_grps to the parent
1248 * If the parent's parent is another throtl_grp, dispatching is propagated
1249 * by either arming its pending_timer or repeating dispatch directly. If
1250 * the top-level service_tree is reached, throtl_data->dispatch_work is
1251 * kicked so that the ready bio's are issued.
1253 static void throtl_pending_timer_fn(unsigned long arg
)
1255 struct throtl_service_queue
*sq
= (void *)arg
;
1256 struct throtl_grp
*tg
= sq_to_tg(sq
);
1257 struct throtl_data
*td
= sq_to_td(sq
);
1258 struct request_queue
*q
= td
->queue
;
1259 struct throtl_service_queue
*parent_sq
;
1263 spin_lock_irq(q
->queue_lock
);
1264 if (throtl_can_upgrade(td
, NULL
))
1265 throtl_upgrade_state(td
);
1268 parent_sq
= sq
->parent_sq
;
1272 throtl_log(sq
, "dispatch nr_queued=%u read=%u write=%u",
1273 sq
->nr_queued
[READ
] + sq
->nr_queued
[WRITE
],
1274 sq
->nr_queued
[READ
], sq
->nr_queued
[WRITE
]);
1276 ret
= throtl_select_dispatch(sq
);
1278 throtl_log(sq
, "bios disp=%u", ret
);
1282 if (throtl_schedule_next_dispatch(sq
, false))
1285 /* this dispatch windows is still open, relax and repeat */
1286 spin_unlock_irq(q
->queue_lock
);
1288 spin_lock_irq(q
->queue_lock
);
1295 /* @parent_sq is another throl_grp, propagate dispatch */
1296 if (tg
->flags
& THROTL_TG_WAS_EMPTY
) {
1297 tg_update_disptime(tg
);
1298 if (!throtl_schedule_next_dispatch(parent_sq
, false)) {
1299 /* window is already open, repeat dispatching */
1306 /* reached the top-level, queue issueing */
1307 queue_work(kthrotld_workqueue
, &td
->dispatch_work
);
1310 spin_unlock_irq(q
->queue_lock
);
1314 * blk_throtl_dispatch_work_fn - work function for throtl_data->dispatch_work
1315 * @work: work item being executed
1317 * This function is queued for execution when bio's reach the bio_lists[]
1318 * of throtl_data->service_queue. Those bio's are ready and issued by this
1321 static void blk_throtl_dispatch_work_fn(struct work_struct
*work
)
1323 struct throtl_data
*td
= container_of(work
, struct throtl_data
,
1325 struct throtl_service_queue
*td_sq
= &td
->service_queue
;
1326 struct request_queue
*q
= td
->queue
;
1327 struct bio_list bio_list_on_stack
;
1329 struct blk_plug plug
;
1332 bio_list_init(&bio_list_on_stack
);
1334 spin_lock_irq(q
->queue_lock
);
1335 for (rw
= READ
; rw
<= WRITE
; rw
++)
1336 while ((bio
= throtl_pop_queued(&td_sq
->queued
[rw
], NULL
)))
1337 bio_list_add(&bio_list_on_stack
, bio
);
1338 spin_unlock_irq(q
->queue_lock
);
1340 if (!bio_list_empty(&bio_list_on_stack
)) {
1341 blk_start_plug(&plug
);
1342 while((bio
= bio_list_pop(&bio_list_on_stack
)))
1343 generic_make_request(bio
);
1344 blk_finish_plug(&plug
);
1348 static u64
tg_prfill_conf_u64(struct seq_file
*sf
, struct blkg_policy_data
*pd
,
1351 struct throtl_grp
*tg
= pd_to_tg(pd
);
1352 u64 v
= *(u64
*)((void *)tg
+ off
);
1356 return __blkg_prfill_u64(sf
, pd
, v
);
1359 static u64
tg_prfill_conf_uint(struct seq_file
*sf
, struct blkg_policy_data
*pd
,
1362 struct throtl_grp
*tg
= pd_to_tg(pd
);
1363 unsigned int v
= *(unsigned int *)((void *)tg
+ off
);
1367 return __blkg_prfill_u64(sf
, pd
, v
);
1370 static int tg_print_conf_u64(struct seq_file
*sf
, void *v
)
1372 blkcg_print_blkgs(sf
, css_to_blkcg(seq_css(sf
)), tg_prfill_conf_u64
,
1373 &blkcg_policy_throtl
, seq_cft(sf
)->private, false);
1377 static int tg_print_conf_uint(struct seq_file
*sf
, void *v
)
1379 blkcg_print_blkgs(sf
, css_to_blkcg(seq_css(sf
)), tg_prfill_conf_uint
,
1380 &blkcg_policy_throtl
, seq_cft(sf
)->private, false);
1384 static void tg_conf_updated(struct throtl_grp
*tg
, bool global
)
1386 struct throtl_service_queue
*sq
= &tg
->service_queue
;
1387 struct cgroup_subsys_state
*pos_css
;
1388 struct blkcg_gq
*blkg
;
1390 throtl_log(&tg
->service_queue
,
1391 "limit change rbps=%llu wbps=%llu riops=%u wiops=%u",
1392 tg_bps_limit(tg
, READ
), tg_bps_limit(tg
, WRITE
),
1393 tg_iops_limit(tg
, READ
), tg_iops_limit(tg
, WRITE
));
1396 * Update has_rules[] flags for the updated tg's subtree. A tg is
1397 * considered to have rules if either the tg itself or any of its
1398 * ancestors has rules. This identifies groups without any
1399 * restrictions in the whole hierarchy and allows them to bypass
1402 blkg_for_each_descendant_pre(blkg
, pos_css
,
1403 global
? tg
->td
->queue
->root_blkg
: tg_to_blkg(tg
)) {
1404 struct throtl_grp
*this_tg
= blkg_to_tg(blkg
);
1405 struct throtl_grp
*parent_tg
;
1407 tg_update_has_rules(this_tg
);
1408 /* ignore root/second level */
1409 if (!cgroup_subsys_on_dfl(io_cgrp_subsys
) || !blkg
->parent
||
1410 !blkg
->parent
->parent
)
1412 parent_tg
= blkg_to_tg(blkg
->parent
);
1414 * make sure all children has lower idle time threshold and
1415 * higher latency target
1417 this_tg
->idletime_threshold
= min(this_tg
->idletime_threshold
,
1418 parent_tg
->idletime_threshold
);
1419 this_tg
->latency_target
= max(this_tg
->latency_target
,
1420 parent_tg
->latency_target
);
1424 * We're already holding queue_lock and know @tg is valid. Let's
1425 * apply the new config directly.
1427 * Restart the slices for both READ and WRITES. It might happen
1428 * that a group's limit are dropped suddenly and we don't want to
1429 * account recently dispatched IO with new low rate.
1431 throtl_start_new_slice(tg
, 0);
1432 throtl_start_new_slice(tg
, 1);
1434 if (tg
->flags
& THROTL_TG_PENDING
) {
1435 tg_update_disptime(tg
);
1436 throtl_schedule_next_dispatch(sq
->parent_sq
, true);
1440 static ssize_t
tg_set_conf(struct kernfs_open_file
*of
,
1441 char *buf
, size_t nbytes
, loff_t off
, bool is_u64
)
1443 struct blkcg
*blkcg
= css_to_blkcg(of_css(of
));
1444 struct blkg_conf_ctx ctx
;
1445 struct throtl_grp
*tg
;
1449 ret
= blkg_conf_prep(blkcg
, &blkcg_policy_throtl
, buf
, &ctx
);
1454 if (sscanf(ctx
.body
, "%llu", &v
) != 1)
1459 tg
= blkg_to_tg(ctx
.blkg
);
1462 *(u64
*)((void *)tg
+ of_cft(of
)->private) = v
;
1464 *(unsigned int *)((void *)tg
+ of_cft(of
)->private) = v
;
1466 tg_conf_updated(tg
, false);
1469 blkg_conf_finish(&ctx
);
1470 return ret
?: nbytes
;
1473 static ssize_t
tg_set_conf_u64(struct kernfs_open_file
*of
,
1474 char *buf
, size_t nbytes
, loff_t off
)
1476 return tg_set_conf(of
, buf
, nbytes
, off
, true);
1479 static ssize_t
tg_set_conf_uint(struct kernfs_open_file
*of
,
1480 char *buf
, size_t nbytes
, loff_t off
)
1482 return tg_set_conf(of
, buf
, nbytes
, off
, false);
1485 static struct cftype throtl_legacy_files
[] = {
1487 .name
= "throttle.read_bps_device",
1488 .private = offsetof(struct throtl_grp
, bps
[READ
][LIMIT_MAX
]),
1489 .seq_show
= tg_print_conf_u64
,
1490 .write
= tg_set_conf_u64
,
1493 .name
= "throttle.write_bps_device",
1494 .private = offsetof(struct throtl_grp
, bps
[WRITE
][LIMIT_MAX
]),
1495 .seq_show
= tg_print_conf_u64
,
1496 .write
= tg_set_conf_u64
,
1499 .name
= "throttle.read_iops_device",
1500 .private = offsetof(struct throtl_grp
, iops
[READ
][LIMIT_MAX
]),
1501 .seq_show
= tg_print_conf_uint
,
1502 .write
= tg_set_conf_uint
,
1505 .name
= "throttle.write_iops_device",
1506 .private = offsetof(struct throtl_grp
, iops
[WRITE
][LIMIT_MAX
]),
1507 .seq_show
= tg_print_conf_uint
,
1508 .write
= tg_set_conf_uint
,
1511 .name
= "throttle.io_service_bytes",
1512 .private = (unsigned long)&blkcg_policy_throtl
,
1513 .seq_show
= blkg_print_stat_bytes
,
1516 .name
= "throttle.io_serviced",
1517 .private = (unsigned long)&blkcg_policy_throtl
,
1518 .seq_show
= blkg_print_stat_ios
,
1523 static u64
tg_prfill_limit(struct seq_file
*sf
, struct blkg_policy_data
*pd
,
1526 struct throtl_grp
*tg
= pd_to_tg(pd
);
1527 const char *dname
= blkg_dev_name(pd
->blkg
);
1528 char bufs
[4][21] = { "max", "max", "max", "max" };
1530 unsigned int iops_dft
;
1531 char idle_time
[26] = "";
1532 char latency_time
[26] = "";
1537 if (off
== LIMIT_LOW
) {
1542 iops_dft
= UINT_MAX
;
1545 if (tg
->bps_conf
[READ
][off
] == bps_dft
&&
1546 tg
->bps_conf
[WRITE
][off
] == bps_dft
&&
1547 tg
->iops_conf
[READ
][off
] == iops_dft
&&
1548 tg
->iops_conf
[WRITE
][off
] == iops_dft
&&
1549 (off
!= LIMIT_LOW
||
1550 (tg
->idletime_threshold_conf
== DFL_IDLE_THRESHOLD
&&
1551 tg
->latency_target_conf
== DFL_LATENCY_TARGET
)))
1554 if (tg
->bps_conf
[READ
][off
] != U64_MAX
)
1555 snprintf(bufs
[0], sizeof(bufs
[0]), "%llu",
1556 tg
->bps_conf
[READ
][off
]);
1557 if (tg
->bps_conf
[WRITE
][off
] != U64_MAX
)
1558 snprintf(bufs
[1], sizeof(bufs
[1]), "%llu",
1559 tg
->bps_conf
[WRITE
][off
]);
1560 if (tg
->iops_conf
[READ
][off
] != UINT_MAX
)
1561 snprintf(bufs
[2], sizeof(bufs
[2]), "%u",
1562 tg
->iops_conf
[READ
][off
]);
1563 if (tg
->iops_conf
[WRITE
][off
] != UINT_MAX
)
1564 snprintf(bufs
[3], sizeof(bufs
[3]), "%u",
1565 tg
->iops_conf
[WRITE
][off
]);
1566 if (off
== LIMIT_LOW
) {
1567 if (tg
->idletime_threshold_conf
== ULONG_MAX
)
1568 strcpy(idle_time
, " idle=max");
1570 snprintf(idle_time
, sizeof(idle_time
), " idle=%lu",
1571 tg
->idletime_threshold_conf
);
1573 if (tg
->latency_target_conf
== ULONG_MAX
)
1574 strcpy(latency_time
, " latency=max");
1576 snprintf(latency_time
, sizeof(latency_time
),
1577 " latency=%lu", tg
->latency_target_conf
);
1580 seq_printf(sf
, "%s rbps=%s wbps=%s riops=%s wiops=%s%s%s\n",
1581 dname
, bufs
[0], bufs
[1], bufs
[2], bufs
[3], idle_time
,
1586 static int tg_print_limit(struct seq_file
*sf
, void *v
)
1588 blkcg_print_blkgs(sf
, css_to_blkcg(seq_css(sf
)), tg_prfill_limit
,
1589 &blkcg_policy_throtl
, seq_cft(sf
)->private, false);
1593 static ssize_t
tg_set_limit(struct kernfs_open_file
*of
,
1594 char *buf
, size_t nbytes
, loff_t off
)
1596 struct blkcg
*blkcg
= css_to_blkcg(of_css(of
));
1597 struct blkg_conf_ctx ctx
;
1598 struct throtl_grp
*tg
;
1600 unsigned long idle_time
;
1601 unsigned long latency_time
;
1603 int index
= of_cft(of
)->private;
1605 ret
= blkg_conf_prep(blkcg
, &blkcg_policy_throtl
, buf
, &ctx
);
1609 tg
= blkg_to_tg(ctx
.blkg
);
1611 v
[0] = tg
->bps_conf
[READ
][index
];
1612 v
[1] = tg
->bps_conf
[WRITE
][index
];
1613 v
[2] = tg
->iops_conf
[READ
][index
];
1614 v
[3] = tg
->iops_conf
[WRITE
][index
];
1616 idle_time
= tg
->idletime_threshold_conf
;
1617 latency_time
= tg
->latency_target_conf
;
1619 char tok
[27]; /* wiops=18446744073709551616 */
1624 if (sscanf(ctx
.body
, "%26s%n", tok
, &len
) != 1)
1633 if (!p
|| (sscanf(p
, "%llu", &val
) != 1 && strcmp(p
, "max")))
1641 if (!strcmp(tok
, "rbps"))
1643 else if (!strcmp(tok
, "wbps"))
1645 else if (!strcmp(tok
, "riops"))
1646 v
[2] = min_t(u64
, val
, UINT_MAX
);
1647 else if (!strcmp(tok
, "wiops"))
1648 v
[3] = min_t(u64
, val
, UINT_MAX
);
1649 else if (off
== LIMIT_LOW
&& !strcmp(tok
, "idle"))
1651 else if (off
== LIMIT_LOW
&& !strcmp(tok
, "latency"))
1657 tg
->bps_conf
[READ
][index
] = v
[0];
1658 tg
->bps_conf
[WRITE
][index
] = v
[1];
1659 tg
->iops_conf
[READ
][index
] = v
[2];
1660 tg
->iops_conf
[WRITE
][index
] = v
[3];
1662 if (index
== LIMIT_MAX
) {
1663 tg
->bps
[READ
][index
] = v
[0];
1664 tg
->bps
[WRITE
][index
] = v
[1];
1665 tg
->iops
[READ
][index
] = v
[2];
1666 tg
->iops
[WRITE
][index
] = v
[3];
1668 tg
->bps
[READ
][LIMIT_LOW
] = min(tg
->bps_conf
[READ
][LIMIT_LOW
],
1669 tg
->bps_conf
[READ
][LIMIT_MAX
]);
1670 tg
->bps
[WRITE
][LIMIT_LOW
] = min(tg
->bps_conf
[WRITE
][LIMIT_LOW
],
1671 tg
->bps_conf
[WRITE
][LIMIT_MAX
]);
1672 tg
->iops
[READ
][LIMIT_LOW
] = min(tg
->iops_conf
[READ
][LIMIT_LOW
],
1673 tg
->iops_conf
[READ
][LIMIT_MAX
]);
1674 tg
->iops
[WRITE
][LIMIT_LOW
] = min(tg
->iops_conf
[WRITE
][LIMIT_LOW
],
1675 tg
->iops_conf
[WRITE
][LIMIT_MAX
]);
1676 tg
->idletime_threshold_conf
= idle_time
;
1677 tg
->latency_target_conf
= latency_time
;
1679 /* force user to configure all settings for low limit */
1680 if (!(tg
->bps
[READ
][LIMIT_LOW
] || tg
->iops
[READ
][LIMIT_LOW
] ||
1681 tg
->bps
[WRITE
][LIMIT_LOW
] || tg
->iops
[WRITE
][LIMIT_LOW
]) ||
1682 tg
->idletime_threshold_conf
== DFL_IDLE_THRESHOLD
||
1683 tg
->latency_target_conf
== DFL_LATENCY_TARGET
) {
1684 tg
->bps
[READ
][LIMIT_LOW
] = 0;
1685 tg
->bps
[WRITE
][LIMIT_LOW
] = 0;
1686 tg
->iops
[READ
][LIMIT_LOW
] = 0;
1687 tg
->iops
[WRITE
][LIMIT_LOW
] = 0;
1688 tg
->idletime_threshold
= DFL_IDLE_THRESHOLD
;
1689 tg
->latency_target
= DFL_LATENCY_TARGET
;
1690 } else if (index
== LIMIT_LOW
) {
1691 tg
->idletime_threshold
= tg
->idletime_threshold_conf
;
1692 tg
->latency_target
= tg
->latency_target_conf
;
1695 blk_throtl_update_limit_valid(tg
->td
);
1696 if (tg
->td
->limit_valid
[LIMIT_LOW
]) {
1697 if (index
== LIMIT_LOW
)
1698 tg
->td
->limit_index
= LIMIT_LOW
;
1700 tg
->td
->limit_index
= LIMIT_MAX
;
1701 tg_conf_updated(tg
, index
== LIMIT_LOW
&&
1702 tg
->td
->limit_valid
[LIMIT_LOW
]);
1705 blkg_conf_finish(&ctx
);
1706 return ret
?: nbytes
;
1709 static struct cftype throtl_files
[] = {
1710 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
1713 .flags
= CFTYPE_NOT_ON_ROOT
,
1714 .seq_show
= tg_print_limit
,
1715 .write
= tg_set_limit
,
1716 .private = LIMIT_LOW
,
1721 .flags
= CFTYPE_NOT_ON_ROOT
,
1722 .seq_show
= tg_print_limit
,
1723 .write
= tg_set_limit
,
1724 .private = LIMIT_MAX
,
1729 static void throtl_shutdown_wq(struct request_queue
*q
)
1731 struct throtl_data
*td
= q
->td
;
1733 cancel_work_sync(&td
->dispatch_work
);
1736 static struct blkcg_policy blkcg_policy_throtl
= {
1737 .dfl_cftypes
= throtl_files
,
1738 .legacy_cftypes
= throtl_legacy_files
,
1740 .pd_alloc_fn
= throtl_pd_alloc
,
1741 .pd_init_fn
= throtl_pd_init
,
1742 .pd_online_fn
= throtl_pd_online
,
1743 .pd_offline_fn
= throtl_pd_offline
,
1744 .pd_free_fn
= throtl_pd_free
,
1747 static unsigned long __tg_last_low_overflow_time(struct throtl_grp
*tg
)
1749 unsigned long rtime
= jiffies
, wtime
= jiffies
;
1751 if (tg
->bps
[READ
][LIMIT_LOW
] || tg
->iops
[READ
][LIMIT_LOW
])
1752 rtime
= tg
->last_low_overflow_time
[READ
];
1753 if (tg
->bps
[WRITE
][LIMIT_LOW
] || tg
->iops
[WRITE
][LIMIT_LOW
])
1754 wtime
= tg
->last_low_overflow_time
[WRITE
];
1755 return min(rtime
, wtime
);
1758 /* tg should not be an intermediate node */
1759 static unsigned long tg_last_low_overflow_time(struct throtl_grp
*tg
)
1761 struct throtl_service_queue
*parent_sq
;
1762 struct throtl_grp
*parent
= tg
;
1763 unsigned long ret
= __tg_last_low_overflow_time(tg
);
1766 parent_sq
= parent
->service_queue
.parent_sq
;
1767 parent
= sq_to_tg(parent_sq
);
1772 * The parent doesn't have low limit, it always reaches low
1773 * limit. Its overflow time is useless for children
1775 if (!parent
->bps
[READ
][LIMIT_LOW
] &&
1776 !parent
->iops
[READ
][LIMIT_LOW
] &&
1777 !parent
->bps
[WRITE
][LIMIT_LOW
] &&
1778 !parent
->iops
[WRITE
][LIMIT_LOW
])
1780 if (time_after(__tg_last_low_overflow_time(parent
), ret
))
1781 ret
= __tg_last_low_overflow_time(parent
);
1786 static bool throtl_tg_is_idle(struct throtl_grp
*tg
)
1789 * cgroup is idle if:
1790 * - single idle is too long, longer than a fixed value (in case user
1791 * configure a too big threshold) or 4 times of idletime threshold
1792 * - average think time is more than threshold
1793 * - IO latency is largely below threshold
1798 time
= min_t(unsigned long, MAX_IDLE_TIME
, 4 * tg
->idletime_threshold
);
1799 ret
= tg
->latency_target
== DFL_LATENCY_TARGET
||
1800 tg
->idletime_threshold
== DFL_IDLE_THRESHOLD
||
1801 (ktime_get_ns() >> 10) - tg
->last_finish_time
> time
||
1802 tg
->avg_idletime
> tg
->idletime_threshold
||
1803 (tg
->latency_target
&& tg
->bio_cnt
&&
1804 tg
->bad_bio_cnt
* 5 < tg
->bio_cnt
);
1805 throtl_log(&tg
->service_queue
,
1806 "avg_idle=%ld, idle_threshold=%ld, bad_bio=%d, total_bio=%d, is_idle=%d, scale=%d",
1807 tg
->avg_idletime
, tg
->idletime_threshold
, tg
->bad_bio_cnt
,
1808 tg
->bio_cnt
, ret
, tg
->td
->scale
);
1812 static bool throtl_tg_can_upgrade(struct throtl_grp
*tg
)
1814 struct throtl_service_queue
*sq
= &tg
->service_queue
;
1815 bool read_limit
, write_limit
;
1818 * if cgroup reaches low limit (if low limit is 0, the cgroup always
1819 * reaches), it's ok to upgrade to next limit
1821 read_limit
= tg
->bps
[READ
][LIMIT_LOW
] || tg
->iops
[READ
][LIMIT_LOW
];
1822 write_limit
= tg
->bps
[WRITE
][LIMIT_LOW
] || tg
->iops
[WRITE
][LIMIT_LOW
];
1823 if (!read_limit
&& !write_limit
)
1825 if (read_limit
&& sq
->nr_queued
[READ
] &&
1826 (!write_limit
|| sq
->nr_queued
[WRITE
]))
1828 if (write_limit
&& sq
->nr_queued
[WRITE
] &&
1829 (!read_limit
|| sq
->nr_queued
[READ
]))
1832 if (time_after_eq(jiffies
,
1833 tg_last_low_overflow_time(tg
) + tg
->td
->throtl_slice
) &&
1834 throtl_tg_is_idle(tg
))
1839 static bool throtl_hierarchy_can_upgrade(struct throtl_grp
*tg
)
1842 if (throtl_tg_can_upgrade(tg
))
1844 tg
= sq_to_tg(tg
->service_queue
.parent_sq
);
1845 if (!tg
|| !tg_to_blkg(tg
)->parent
)
1851 static bool throtl_can_upgrade(struct throtl_data
*td
,
1852 struct throtl_grp
*this_tg
)
1854 struct cgroup_subsys_state
*pos_css
;
1855 struct blkcg_gq
*blkg
;
1857 if (td
->limit_index
!= LIMIT_LOW
)
1860 if (time_before(jiffies
, td
->low_downgrade_time
+ td
->throtl_slice
))
1864 blkg_for_each_descendant_post(blkg
, pos_css
, td
->queue
->root_blkg
) {
1865 struct throtl_grp
*tg
= blkg_to_tg(blkg
);
1869 if (!list_empty(&tg_to_blkg(tg
)->blkcg
->css
.children
))
1871 if (!throtl_hierarchy_can_upgrade(tg
)) {
1880 static void throtl_upgrade_check(struct throtl_grp
*tg
)
1882 unsigned long now
= jiffies
;
1884 if (tg
->td
->limit_index
!= LIMIT_LOW
)
1887 if (time_after(tg
->last_check_time
+ tg
->td
->throtl_slice
, now
))
1890 tg
->last_check_time
= now
;
1892 if (!time_after_eq(now
,
1893 __tg_last_low_overflow_time(tg
) + tg
->td
->throtl_slice
))
1896 if (throtl_can_upgrade(tg
->td
, NULL
))
1897 throtl_upgrade_state(tg
->td
);
1900 static void throtl_upgrade_state(struct throtl_data
*td
)
1902 struct cgroup_subsys_state
*pos_css
;
1903 struct blkcg_gq
*blkg
;
1905 throtl_log(&td
->service_queue
, "upgrade to max");
1906 td
->limit_index
= LIMIT_MAX
;
1907 td
->low_upgrade_time
= jiffies
;
1910 blkg_for_each_descendant_post(blkg
, pos_css
, td
->queue
->root_blkg
) {
1911 struct throtl_grp
*tg
= blkg_to_tg(blkg
);
1912 struct throtl_service_queue
*sq
= &tg
->service_queue
;
1914 tg
->disptime
= jiffies
- 1;
1915 throtl_select_dispatch(sq
);
1916 throtl_schedule_next_dispatch(sq
, false);
1919 throtl_select_dispatch(&td
->service_queue
);
1920 throtl_schedule_next_dispatch(&td
->service_queue
, false);
1921 queue_work(kthrotld_workqueue
, &td
->dispatch_work
);
1924 static void throtl_downgrade_state(struct throtl_data
*td
, int new)
1928 throtl_log(&td
->service_queue
, "downgrade, scale %d", td
->scale
);
1930 td
->low_upgrade_time
= jiffies
- td
->scale
* td
->throtl_slice
;
1934 td
->limit_index
= new;
1935 td
->low_downgrade_time
= jiffies
;
1938 static bool throtl_tg_can_downgrade(struct throtl_grp
*tg
)
1940 struct throtl_data
*td
= tg
->td
;
1941 unsigned long now
= jiffies
;
1944 * If cgroup is below low limit, consider downgrade and throttle other
1947 if (time_after_eq(now
, td
->low_upgrade_time
+ td
->throtl_slice
) &&
1948 time_after_eq(now
, tg_last_low_overflow_time(tg
) +
1949 td
->throtl_slice
) &&
1950 (!throtl_tg_is_idle(tg
) ||
1951 !list_empty(&tg_to_blkg(tg
)->blkcg
->css
.children
)))
1956 static bool throtl_hierarchy_can_downgrade(struct throtl_grp
*tg
)
1959 if (!throtl_tg_can_downgrade(tg
))
1961 tg
= sq_to_tg(tg
->service_queue
.parent_sq
);
1962 if (!tg
|| !tg_to_blkg(tg
)->parent
)
1968 static void throtl_downgrade_check(struct throtl_grp
*tg
)
1972 unsigned long elapsed_time
;
1973 unsigned long now
= jiffies
;
1975 if (tg
->td
->limit_index
!= LIMIT_MAX
||
1976 !tg
->td
->limit_valid
[LIMIT_LOW
])
1978 if (!list_empty(&tg_to_blkg(tg
)->blkcg
->css
.children
))
1980 if (time_after(tg
->last_check_time
+ tg
->td
->throtl_slice
, now
))
1983 elapsed_time
= now
- tg
->last_check_time
;
1984 tg
->last_check_time
= now
;
1986 if (time_before(now
, tg_last_low_overflow_time(tg
) +
1987 tg
->td
->throtl_slice
))
1990 if (tg
->bps
[READ
][LIMIT_LOW
]) {
1991 bps
= tg
->last_bytes_disp
[READ
] * HZ
;
1992 do_div(bps
, elapsed_time
);
1993 if (bps
>= tg
->bps
[READ
][LIMIT_LOW
])
1994 tg
->last_low_overflow_time
[READ
] = now
;
1997 if (tg
->bps
[WRITE
][LIMIT_LOW
]) {
1998 bps
= tg
->last_bytes_disp
[WRITE
] * HZ
;
1999 do_div(bps
, elapsed_time
);
2000 if (bps
>= tg
->bps
[WRITE
][LIMIT_LOW
])
2001 tg
->last_low_overflow_time
[WRITE
] = now
;
2004 if (tg
->iops
[READ
][LIMIT_LOW
]) {
2005 iops
= tg
->last_io_disp
[READ
] * HZ
/ elapsed_time
;
2006 if (iops
>= tg
->iops
[READ
][LIMIT_LOW
])
2007 tg
->last_low_overflow_time
[READ
] = now
;
2010 if (tg
->iops
[WRITE
][LIMIT_LOW
]) {
2011 iops
= tg
->last_io_disp
[WRITE
] * HZ
/ elapsed_time
;
2012 if (iops
>= tg
->iops
[WRITE
][LIMIT_LOW
])
2013 tg
->last_low_overflow_time
[WRITE
] = now
;
2017 * If cgroup is below low limit, consider downgrade and throttle other
2020 if (throtl_hierarchy_can_downgrade(tg
))
2021 throtl_downgrade_state(tg
->td
, LIMIT_LOW
);
2023 tg
->last_bytes_disp
[READ
] = 0;
2024 tg
->last_bytes_disp
[WRITE
] = 0;
2025 tg
->last_io_disp
[READ
] = 0;
2026 tg
->last_io_disp
[WRITE
] = 0;
2029 static void blk_throtl_update_idletime(struct throtl_grp
*tg
)
2031 unsigned long now
= ktime_get_ns() >> 10;
2032 unsigned long last_finish_time
= tg
->last_finish_time
;
2034 if (now
<= last_finish_time
|| last_finish_time
== 0 ||
2035 last_finish_time
== tg
->checked_last_finish_time
)
2038 tg
->avg_idletime
= (tg
->avg_idletime
* 7 + now
- last_finish_time
) >> 3;
2039 tg
->checked_last_finish_time
= last_finish_time
;
2042 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2043 static void throtl_update_latency_buckets(struct throtl_data
*td
)
2045 struct avg_latency_bucket avg_latency
[LATENCY_BUCKET_SIZE
];
2047 unsigned long last_latency
= 0;
2048 unsigned long latency
;
2050 if (!blk_queue_nonrot(td
->queue
))
2052 if (time_before(jiffies
, td
->last_calculate_time
+ HZ
))
2054 td
->last_calculate_time
= jiffies
;
2056 memset(avg_latency
, 0, sizeof(avg_latency
));
2057 for (i
= 0; i
< LATENCY_BUCKET_SIZE
; i
++) {
2058 struct latency_bucket
*tmp
= &td
->tmp_buckets
[i
];
2060 for_each_possible_cpu(cpu
) {
2061 struct latency_bucket
*bucket
;
2063 /* this isn't race free, but ok in practice */
2064 bucket
= per_cpu_ptr(td
->latency_buckets
, cpu
);
2065 tmp
->total_latency
+= bucket
[i
].total_latency
;
2066 tmp
->samples
+= bucket
[i
].samples
;
2067 bucket
[i
].total_latency
= 0;
2068 bucket
[i
].samples
= 0;
2071 if (tmp
->samples
>= 32) {
2072 int samples
= tmp
->samples
;
2074 latency
= tmp
->total_latency
;
2076 tmp
->total_latency
= 0;
2081 avg_latency
[i
].latency
= latency
;
2085 for (i
= 0; i
< LATENCY_BUCKET_SIZE
; i
++) {
2086 if (!avg_latency
[i
].latency
) {
2087 if (td
->avg_buckets
[i
].latency
< last_latency
)
2088 td
->avg_buckets
[i
].latency
= last_latency
;
2092 if (!td
->avg_buckets
[i
].valid
)
2093 latency
= avg_latency
[i
].latency
;
2095 latency
= (td
->avg_buckets
[i
].latency
* 7 +
2096 avg_latency
[i
].latency
) >> 3;
2098 td
->avg_buckets
[i
].latency
= max(latency
, last_latency
);
2099 td
->avg_buckets
[i
].valid
= true;
2100 last_latency
= td
->avg_buckets
[i
].latency
;
2103 for (i
= 0; i
< LATENCY_BUCKET_SIZE
; i
++)
2104 throtl_log(&td
->service_queue
,
2105 "Latency bucket %d: latency=%ld, valid=%d", i
,
2106 td
->avg_buckets
[i
].latency
, td
->avg_buckets
[i
].valid
);
2109 static inline void throtl_update_latency_buckets(struct throtl_data
*td
)
2114 static void blk_throtl_assoc_bio(struct throtl_grp
*tg
, struct bio
*bio
)
2116 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2119 ret
= bio_associate_current(bio
);
2120 if (ret
== 0 || ret
== -EBUSY
)
2121 bio
->bi_cg_private
= tg
;
2122 blk_stat_set_issue(&bio
->bi_issue_stat
, bio_sectors(bio
));
2124 bio_associate_current(bio
);
2128 bool blk_throtl_bio(struct request_queue
*q
, struct blkcg_gq
*blkg
,
2131 struct throtl_qnode
*qn
= NULL
;
2132 struct throtl_grp
*tg
= blkg_to_tg(blkg
?: q
->root_blkg
);
2133 struct throtl_service_queue
*sq
;
2134 bool rw
= bio_data_dir(bio
);
2135 bool throttled
= false;
2136 struct throtl_data
*td
= tg
->td
;
2138 WARN_ON_ONCE(!rcu_read_lock_held());
2140 /* see throtl_charge_bio() */
2141 if (bio_flagged(bio
, BIO_THROTTLED
) || !tg
->has_rules
[rw
])
2144 spin_lock_irq(q
->queue_lock
);
2146 throtl_update_latency_buckets(td
);
2148 if (unlikely(blk_queue_bypass(q
)))
2151 blk_throtl_assoc_bio(tg
, bio
);
2152 blk_throtl_update_idletime(tg
);
2154 sq
= &tg
->service_queue
;
2158 if (tg
->last_low_overflow_time
[rw
] == 0)
2159 tg
->last_low_overflow_time
[rw
] = jiffies
;
2160 throtl_downgrade_check(tg
);
2161 throtl_upgrade_check(tg
);
2162 /* throtl is FIFO - if bios are already queued, should queue */
2163 if (sq
->nr_queued
[rw
])
2166 /* if above limits, break to queue */
2167 if (!tg_may_dispatch(tg
, bio
, NULL
)) {
2168 tg
->last_low_overflow_time
[rw
] = jiffies
;
2169 if (throtl_can_upgrade(td
, tg
)) {
2170 throtl_upgrade_state(td
);
2176 /* within limits, let's charge and dispatch directly */
2177 throtl_charge_bio(tg
, bio
);
2180 * We need to trim slice even when bios are not being queued
2181 * otherwise it might happen that a bio is not queued for
2182 * a long time and slice keeps on extending and trim is not
2183 * called for a long time. Now if limits are reduced suddenly
2184 * we take into account all the IO dispatched so far at new
2185 * low rate and * newly queued IO gets a really long dispatch
2188 * So keep on trimming slice even if bio is not queued.
2190 throtl_trim_slice(tg
, rw
);
2193 * @bio passed through this layer without being throttled.
2194 * Climb up the ladder. If we''re already at the top, it
2195 * can be executed directly.
2197 qn
= &tg
->qnode_on_parent
[rw
];
2204 /* out-of-limit, queue to @tg */
2205 throtl_log(sq
, "[%c] bio. bdisp=%llu sz=%u bps=%llu iodisp=%u iops=%u queued=%d/%d",
2206 rw
== READ
? 'R' : 'W',
2207 tg
->bytes_disp
[rw
], bio
->bi_iter
.bi_size
,
2208 tg_bps_limit(tg
, rw
),
2209 tg
->io_disp
[rw
], tg_iops_limit(tg
, rw
),
2210 sq
->nr_queued
[READ
], sq
->nr_queued
[WRITE
]);
2212 tg
->last_low_overflow_time
[rw
] = jiffies
;
2214 td
->nr_queued
[rw
]++;
2215 throtl_add_bio_tg(bio
, qn
, tg
);
2219 * Update @tg's dispatch time and force schedule dispatch if @tg
2220 * was empty before @bio. The forced scheduling isn't likely to
2221 * cause undue delay as @bio is likely to be dispatched directly if
2222 * its @tg's disptime is not in the future.
2224 if (tg
->flags
& THROTL_TG_WAS_EMPTY
) {
2225 tg_update_disptime(tg
);
2226 throtl_schedule_next_dispatch(tg
->service_queue
.parent_sq
, true);
2230 spin_unlock_irq(q
->queue_lock
);
2233 * As multiple blk-throtls may stack in the same issue path, we
2234 * don't want bios to leave with the flag set. Clear the flag if
2238 bio_clear_flag(bio
, BIO_THROTTLED
);
2240 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2241 if (throttled
|| !td
->track_bio_latency
)
2242 bio
->bi_issue_stat
.stat
|= SKIP_LATENCY
;
2247 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2248 static void throtl_track_latency(struct throtl_data
*td
, sector_t size
,
2249 int op
, unsigned long time
)
2251 struct latency_bucket
*latency
;
2254 if (!td
|| td
->limit_index
!= LIMIT_LOW
|| op
!= REQ_OP_READ
||
2255 !blk_queue_nonrot(td
->queue
))
2258 index
= request_bucket_index(size
);
2260 latency
= get_cpu_ptr(td
->latency_buckets
);
2261 latency
[index
].total_latency
+= time
;
2262 latency
[index
].samples
++;
2263 put_cpu_ptr(td
->latency_buckets
);
2266 void blk_throtl_stat_add(struct request
*rq
, u64 time_ns
)
2268 struct request_queue
*q
= rq
->q
;
2269 struct throtl_data
*td
= q
->td
;
2271 throtl_track_latency(td
, blk_stat_size(&rq
->issue_stat
),
2272 req_op(rq
), time_ns
>> 10);
2275 void blk_throtl_bio_endio(struct bio
*bio
)
2277 struct throtl_grp
*tg
;
2279 unsigned long finish_time
;
2280 unsigned long start_time
;
2283 tg
= bio
->bi_cg_private
;
2286 bio
->bi_cg_private
= NULL
;
2288 finish_time_ns
= ktime_get_ns();
2289 tg
->last_finish_time
= finish_time_ns
>> 10;
2291 start_time
= blk_stat_time(&bio
->bi_issue_stat
) >> 10;
2292 finish_time
= __blk_stat_time(finish_time_ns
) >> 10;
2293 if (!start_time
|| finish_time
<= start_time
)
2296 lat
= finish_time
- start_time
;
2297 /* this is only for bio based driver */
2298 if (!(bio
->bi_issue_stat
.stat
& SKIP_LATENCY
))
2299 throtl_track_latency(tg
->td
, blk_stat_size(&bio
->bi_issue_stat
),
2302 if (tg
->latency_target
&& lat
>= tg
->td
->filtered_latency
) {
2304 unsigned int threshold
;
2306 bucket
= request_bucket_index(
2307 blk_stat_size(&bio
->bi_issue_stat
));
2308 threshold
= tg
->td
->avg_buckets
[bucket
].latency
+
2310 if (lat
> threshold
)
2313 * Not race free, could get wrong count, which means cgroups
2319 if (time_after(jiffies
, tg
->bio_cnt_reset_time
) || tg
->bio_cnt
> 1024) {
2320 tg
->bio_cnt_reset_time
= tg
->td
->throtl_slice
+ jiffies
;
2322 tg
->bad_bio_cnt
/= 2;
2328 * Dispatch all bios from all children tg's queued on @parent_sq. On
2329 * return, @parent_sq is guaranteed to not have any active children tg's
2330 * and all bios from previously active tg's are on @parent_sq->bio_lists[].
2332 static void tg_drain_bios(struct throtl_service_queue
*parent_sq
)
2334 struct throtl_grp
*tg
;
2336 while ((tg
= throtl_rb_first(parent_sq
))) {
2337 struct throtl_service_queue
*sq
= &tg
->service_queue
;
2340 throtl_dequeue_tg(tg
);
2342 while ((bio
= throtl_peek_queued(&sq
->queued
[READ
])))
2343 tg_dispatch_one_bio(tg
, bio_data_dir(bio
));
2344 while ((bio
= throtl_peek_queued(&sq
->queued
[WRITE
])))
2345 tg_dispatch_one_bio(tg
, bio_data_dir(bio
));
2350 * blk_throtl_drain - drain throttled bios
2351 * @q: request_queue to drain throttled bios for
2353 * Dispatch all currently throttled bios on @q through ->make_request_fn().
2355 void blk_throtl_drain(struct request_queue
*q
)
2356 __releases(q
->queue_lock
) __acquires(q
->queue_lock
)
2358 struct throtl_data
*td
= q
->td
;
2359 struct blkcg_gq
*blkg
;
2360 struct cgroup_subsys_state
*pos_css
;
2364 queue_lockdep_assert_held(q
);
2368 * Drain each tg while doing post-order walk on the blkg tree, so
2369 * that all bios are propagated to td->service_queue. It'd be
2370 * better to walk service_queue tree directly but blkg walk is
2373 blkg_for_each_descendant_post(blkg
, pos_css
, td
->queue
->root_blkg
)
2374 tg_drain_bios(&blkg_to_tg(blkg
)->service_queue
);
2376 /* finally, transfer bios from top-level tg's into the td */
2377 tg_drain_bios(&td
->service_queue
);
2380 spin_unlock_irq(q
->queue_lock
);
2382 /* all bios now should be in td->service_queue, issue them */
2383 for (rw
= READ
; rw
<= WRITE
; rw
++)
2384 while ((bio
= throtl_pop_queued(&td
->service_queue
.queued
[rw
],
2386 generic_make_request(bio
);
2388 spin_lock_irq(q
->queue_lock
);
2391 int blk_throtl_init(struct request_queue
*q
)
2393 struct throtl_data
*td
;
2396 td
= kzalloc_node(sizeof(*td
), GFP_KERNEL
, q
->node
);
2399 td
->latency_buckets
= __alloc_percpu(sizeof(struct latency_bucket
) *
2400 LATENCY_BUCKET_SIZE
, __alignof__(u64
));
2401 if (!td
->latency_buckets
) {
2406 INIT_WORK(&td
->dispatch_work
, blk_throtl_dispatch_work_fn
);
2407 throtl_service_queue_init(&td
->service_queue
);
2412 td
->limit_valid
[LIMIT_MAX
] = true;
2413 td
->limit_index
= LIMIT_MAX
;
2414 td
->low_upgrade_time
= jiffies
;
2415 td
->low_downgrade_time
= jiffies
;
2417 /* activate policy */
2418 ret
= blkcg_activate_policy(q
, &blkcg_policy_throtl
);
2420 free_percpu(td
->latency_buckets
);
2426 void blk_throtl_exit(struct request_queue
*q
)
2429 throtl_shutdown_wq(q
);
2430 blkcg_deactivate_policy(q
, &blkcg_policy_throtl
);
2431 free_percpu(q
->td
->latency_buckets
);
2435 void blk_throtl_register_queue(struct request_queue
*q
)
2437 struct throtl_data
*td
;
2443 if (blk_queue_nonrot(q
)) {
2444 td
->throtl_slice
= DFL_THROTL_SLICE_SSD
;
2445 td
->filtered_latency
= LATENCY_FILTERED_SSD
;
2447 td
->throtl_slice
= DFL_THROTL_SLICE_HD
;
2448 td
->filtered_latency
= LATENCY_FILTERED_HD
;
2449 for (i
= 0; i
< LATENCY_BUCKET_SIZE
; i
++)
2450 td
->avg_buckets
[i
].latency
= DFL_HD_BASELINE_LATENCY
;
2452 #ifndef CONFIG_BLK_DEV_THROTTLING_LOW
2453 /* if no low limit, use previous default */
2454 td
->throtl_slice
= DFL_THROTL_SLICE_HD
;
2457 td
->track_bio_latency
= !q
->mq_ops
&& !q
->request_fn
;
2458 if (!td
->track_bio_latency
)
2459 blk_stat_enable_accounting(q
);
2462 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2463 ssize_t
blk_throtl_sample_time_show(struct request_queue
*q
, char *page
)
2467 return sprintf(page
, "%u\n", jiffies_to_msecs(q
->td
->throtl_slice
));
2470 ssize_t
blk_throtl_sample_time_store(struct request_queue
*q
,
2471 const char *page
, size_t count
)
2478 if (kstrtoul(page
, 10, &v
))
2480 t
= msecs_to_jiffies(v
);
2481 if (t
== 0 || t
> MAX_THROTL_SLICE
)
2483 q
->td
->throtl_slice
= t
;
2488 static int __init
throtl_init(void)
2490 kthrotld_workqueue
= alloc_workqueue("kthrotld", WQ_MEM_RECLAIM
, 0);
2491 if (!kthrotld_workqueue
)
2492 panic("Failed to create kthrotld\n");
2494 return blkcg_policy_register(&blkcg_policy_throtl
);
2497 module_init(throtl_init
);