1 /* SPDX-License-Identifier: GPL-2.0
3 * IO cost model based controller.
5 * Copyright (C) 2019 Tejun Heo <tj@kernel.org>
6 * Copyright (C) 2019 Andy Newell <newella@fb.com>
7 * Copyright (C) 2019 Facebook
9 * One challenge of controlling IO resources is the lack of trivially
10 * observable cost metric. This is distinguished from CPU and memory where
11 * wallclock time and the number of bytes can serve as accurate enough
14 * Bandwidth and iops are the most commonly used metrics for IO devices but
15 * depending on the type and specifics of the device, different IO patterns
16 * easily lead to multiple orders of magnitude variations rendering them
17 * useless for the purpose of IO capacity distribution. While on-device
18 * time, with a lot of clutches, could serve as a useful approximation for
19 * non-queued rotational devices, this is no longer viable with modern
20 * devices, even the rotational ones.
22 * While there is no cost metric we can trivially observe, it isn't a
23 * complete mystery. For example, on a rotational device, seek cost
24 * dominates while a contiguous transfer contributes a smaller amount
25 * proportional to the size. If we can characterize at least the relative
26 * costs of these different types of IOs, it should be possible to
27 * implement a reasonable work-conserving proportional IO resource
32 * IO cost model estimates the cost of an IO given its basic parameters and
33 * history (e.g. the end sector of the last IO). The cost is measured in
34 * device time. If a given IO is estimated to cost 10ms, the device should
35 * be able to process ~100 of those IOs in a second.
37 * Currently, there's only one builtin cost model - linear. Each IO is
38 * classified as sequential or random and given a base cost accordingly.
39 * On top of that, a size cost proportional to the length of the IO is
40 * added. While simple, this model captures the operational
41 * characteristics of a wide varienty of devices well enough. Default
42 * paramters for several different classes of devices are provided and the
43 * parameters can be configured from userspace via
44 * /sys/fs/cgroup/io.cost.model.
46 * If needed, tools/cgroup/iocost_coef_gen.py can be used to generate
47 * device-specific coefficients.
51 * The device virtual time (vtime) is used as the primary control metric.
52 * The control strategy is composed of the following three parts.
54 * 2-1. Vtime Distribution
56 * When a cgroup becomes active in terms of IOs, its hierarchical share is
57 * calculated. Please consider the following hierarchy where the numbers
58 * inside parentheses denote the configured weights.
64 * A0 (w:100) A1 (w:100)
66 * If B is idle and only A0 and A1 are actively issuing IOs, as the two are
67 * of equal weight, each gets 50% share. If then B starts issuing IOs, B
68 * gets 300/(100+300) or 75% share, and A0 and A1 equally splits the rest,
69 * 12.5% each. The distribution mechanism only cares about these flattened
70 * shares. They're called hweights (hierarchical weights) and always add
71 * upto 1 (HWEIGHT_WHOLE).
73 * A given cgroup's vtime runs slower in inverse proportion to its hweight.
74 * For example, with 12.5% weight, A0's time runs 8 times slower (100/12.5)
75 * against the device vtime - an IO which takes 10ms on the underlying
76 * device is considered to take 80ms on A0.
78 * This constitutes the basis of IO capacity distribution. Each cgroup's
79 * vtime is running at a rate determined by its hweight. A cgroup tracks
80 * the vtime consumed by past IOs and can issue a new IO iff doing so
81 * wouldn't outrun the current device vtime. Otherwise, the IO is
82 * suspended until the vtime has progressed enough to cover it.
84 * 2-2. Vrate Adjustment
86 * It's unrealistic to expect the cost model to be perfect. There are too
87 * many devices and even on the same device the overall performance
88 * fluctuates depending on numerous factors such as IO mixture and device
89 * internal garbage collection. The controller needs to adapt dynamically.
91 * This is achieved by adjusting the overall IO rate according to how busy
92 * the device is. If the device becomes overloaded, we're sending down too
93 * many IOs and should generally slow down. If there are waiting issuers
94 * but the device isn't saturated, we're issuing too few and should
97 * To slow down, we lower the vrate - the rate at which the device vtime
98 * passes compared to the wall clock. For example, if the vtime is running
99 * at the vrate of 75%, all cgroups added up would only be able to issue
100 * 750ms worth of IOs per second, and vice-versa for speeding up.
102 * Device business is determined using two criteria - rq wait and
103 * completion latencies.
105 * When a device gets saturated, the on-device and then the request queues
106 * fill up and a bio which is ready to be issued has to wait for a request
107 * to become available. When this delay becomes noticeable, it's a clear
108 * indication that the device is saturated and we lower the vrate. This
109 * saturation signal is fairly conservative as it only triggers when both
110 * hardware and software queues are filled up, and is used as the default
113 * As devices can have deep queues and be unfair in how the queued commands
114 * are executed, soley depending on rq wait may not result in satisfactory
115 * control quality. For a better control quality, completion latency QoS
116 * parameters can be configured so that the device is considered saturated
117 * if N'th percentile completion latency rises above the set point.
119 * The completion latency requirements are a function of both the
120 * underlying device characteristics and the desired IO latency quality of
121 * service. There is an inherent trade-off - the tighter the latency QoS,
122 * the higher the bandwidth lossage. Latency QoS is disabled by default
123 * and can be set through /sys/fs/cgroup/io.cost.qos.
125 * 2-3. Work Conservation
127 * Imagine two cgroups A and B with equal weights. A is issuing a small IO
128 * periodically while B is sending out enough parallel IOs to saturate the
129 * device on its own. Let's say A's usage amounts to 100ms worth of IO
130 * cost per second, i.e., 10% of the device capacity. The naive
131 * distribution of half and half would lead to 60% utilization of the
132 * device, a significant reduction in the total amount of work done
133 * compared to free-for-all competition. This is too high a cost to pay
136 * To conserve the total amount of work done, we keep track of how much
137 * each active cgroup is actually using and yield part of its weight if
138 * there are other cgroups which can make use of it. In the above case,
139 * A's weight will be lowered so that it hovers above the actual usage and
140 * B would be able to use the rest.
142 * As we don't want to penalize a cgroup for donating its weight, the
143 * surplus weight adjustment factors in a margin and has an immediate
144 * snapback mechanism in case the cgroup needs more IO vtime for itself.
146 * Note that adjusting down surplus weights has the same effects as
147 * accelerating vtime for other cgroups and work conservation can also be
148 * implemented by adjusting vrate dynamically. However, squaring who can
149 * donate and should take back how much requires hweight propagations
150 * anyway making it easier to implement and understand as a separate
155 * Instead of debugfs or other clumsy monitoring mechanisms, this
156 * controller uses a drgn based monitoring script -
157 * tools/cgroup/iocost_monitor.py. For details on drgn, please see
158 * https://github.com/osandov/drgn. The ouput looks like the following.
160 * sdb RUN per=300ms cur_per=234.218:v203.695 busy= +1 vrate= 62.12%
161 * active weight hweight% inflt% dbt delay usages%
162 * test/a * 50/ 50 33.33/ 33.33 27.65 2 0*041 033:033:033
163 * test/b * 100/ 100 66.67/ 66.67 17.56 0 0*000 066:079:077
165 * - per : Timer period
166 * - cur_per : Internal wall and device vtime clock
167 * - vrate : Device virtual time rate against wall clock
168 * - weight : Surplus-adjusted and configured weights
169 * - hweight : Surplus-adjusted and configured hierarchical weights
170 * - inflt : The percentage of in-flight IO cost at the end of last period
171 * - del_ms : Deferred issuer delay induction level and duration
172 * - usages : Usage history
175 #include <linux/kernel.h>
176 #include <linux/module.h>
177 #include <linux/timer.h>
178 #include <linux/time64.h>
179 #include <linux/parser.h>
180 #include <linux/sched/signal.h>
181 #include <linux/blk-cgroup.h>
182 #include <asm/local.h>
183 #include <asm/local64.h>
184 #include "blk-rq-qos.h"
185 #include "blk-stat.h"
188 #ifdef CONFIG_TRACEPOINTS
190 /* copied from TRACE_CGROUP_PATH, see cgroup-internal.h */
191 #define TRACE_IOCG_PATH_LEN 1024
192 static DEFINE_SPINLOCK(trace_iocg_path_lock
);
193 static char trace_iocg_path
[TRACE_IOCG_PATH_LEN
];
195 #define TRACE_IOCG_PATH(type, iocg, ...) \
197 unsigned long flags; \
198 if (trace_iocost_##type##_enabled()) { \
199 spin_lock_irqsave(&trace_iocg_path_lock, flags); \
200 cgroup_path(iocg_to_blkg(iocg)->blkcg->css.cgroup, \
201 trace_iocg_path, TRACE_IOCG_PATH_LEN); \
202 trace_iocost_##type(iocg, trace_iocg_path, \
204 spin_unlock_irqrestore(&trace_iocg_path_lock, flags); \
208 #else /* CONFIG_TRACE_POINTS */
209 #define TRACE_IOCG_PATH(type, iocg, ...) do { } while (0)
210 #endif /* CONFIG_TRACE_POINTS */
215 /* timer period is calculated from latency requirements, bound it */
216 MIN_PERIOD
= USEC_PER_MSEC
,
217 MAX_PERIOD
= USEC_PER_SEC
,
220 * A cgroup's vtime can run 50% behind the device vtime, which
221 * serves as its IO credit buffer. Surplus weight adjustment is
222 * immediately canceled if the vtime margin runs below 10%.
225 INUSE_MARGIN_PCT
= 10,
227 /* Have some play in waitq timer operations */
228 WAITQ_TIMER_MARGIN_PCT
= 5,
231 * vtime can wrap well within a reasonable uptime when vrate is
232 * consistently raised. Don't trust recorded cgroup vtime if the
233 * period counter indicates that it's older than 5mins.
235 VTIME_VALID_DUR
= 300 * USEC_PER_SEC
,
238 * Remember the past three non-zero usages and use the max for
239 * surplus calculation. Three slots guarantee that we remember one
240 * full period usage from the last active stretch even after
241 * partial deactivation and re-activation periods. Don't start
242 * giving away weight before collecting two data points to prevent
243 * hweight adjustments based on one partial activation period.
246 MIN_VALID_USAGES
= 2,
248 /* 1/64k is granular enough and can easily be handled w/ u32 */
249 HWEIGHT_WHOLE
= 1 << 16,
252 * As vtime is used to calculate the cost of each IO, it needs to
253 * be fairly high precision. For example, it should be able to
254 * represent the cost of a single page worth of discard with
255 * suffificient accuracy. At the same time, it should be able to
256 * represent reasonably long enough durations to be useful and
257 * convenient during operation.
259 * 1s worth of vtime is 2^37. This gives us both sub-nanosecond
260 * granularity and days of wrap-around time even at extreme vrates.
262 VTIME_PER_SEC_SHIFT
= 37,
263 VTIME_PER_SEC
= 1LLU << VTIME_PER_SEC_SHIFT
,
264 VTIME_PER_USEC
= VTIME_PER_SEC
/ USEC_PER_SEC
,
265 VTIME_PER_NSEC
= VTIME_PER_SEC
/ NSEC_PER_SEC
,
267 /* bound vrate adjustments within two orders of magnitude */
268 VRATE_MIN_PPM
= 10000, /* 1% */
269 VRATE_MAX_PPM
= 100000000, /* 10000% */
271 VRATE_MIN
= VTIME_PER_USEC
* VRATE_MIN_PPM
/ MILLION
,
272 VRATE_CLAMP_ADJ_PCT
= 4,
274 /* if IOs end up waiting for requests, issue less */
275 RQ_WAIT_BUSY_PCT
= 5,
277 /* unbusy hysterisis */
280 /* don't let cmds which take a very long time pin lagging for too long */
281 MAX_LAGGING_PERIODS
= 10,
284 * If usage% * 1.25 + 2% is lower than hweight% by more than 3%,
285 * donate the surplus.
287 SURPLUS_SCALE_PCT
= 125, /* * 125% */
288 SURPLUS_SCALE_ABS
= HWEIGHT_WHOLE
/ 50, /* + 2% */
289 SURPLUS_MIN_ADJ_DELTA
= HWEIGHT_WHOLE
/ 33, /* 3% */
291 /* switch iff the conditions are met for longer than this */
292 AUTOP_CYCLE_NSEC
= 10LLU * NSEC_PER_SEC
,
295 * Count IO size in 4k pages. The 12bit shift helps keeping
296 * size-proportional components of cost calculation in closer
297 * numbers of digits to per-IO cost components.
300 IOC_PAGE_SIZE
= 1 << IOC_PAGE_SHIFT
,
301 IOC_SECT_TO_PAGE_SHIFT
= IOC_PAGE_SHIFT
- SECTOR_SHIFT
,
303 /* if apart further than 16M, consider randio for linear model */
304 LCOEF_RANDIO_PAGES
= 4096,
313 /* io.cost.qos controls including per-dev enable of the whole controller */
320 /* io.cost.qos params */
331 /* io.cost.model controls */
338 /* builtin linear cost model coefficients */
370 u32 qos
[NR_QOS_PARAMS
];
371 u64 i_lcoefs
[NR_I_LCOEFS
];
372 u64 lcoefs
[NR_LCOEFS
];
373 u32 too_fast_vrate_pct
;
374 u32 too_slow_vrate_pct
;
384 struct ioc_pcpu_stat
{
385 struct ioc_missed missed
[2];
387 local64_t rq_wait_ns
;
397 struct ioc_params params
;
404 struct timer_list timer
;
405 struct list_head active_iocgs
; /* active cgroups */
406 struct ioc_pcpu_stat __percpu
*pcpu_stat
;
408 enum ioc_running running
;
409 atomic64_t vtime_rate
;
411 seqcount_spinlock_t period_seqcount
;
412 u32 period_at
; /* wallclock starttime */
413 u64 period_at_vtime
; /* vtime starttime */
415 atomic64_t cur_period
; /* inc'd each period */
416 int busy_level
; /* saturation history */
418 u64 inuse_margin_vtime
;
419 bool weights_updated
;
420 atomic_t hweight_gen
; /* for lazy hweights */
422 u64 autop_too_fast_at
;
423 u64 autop_too_slow_at
;
425 bool user_qos_params
:1;
426 bool user_cost_model
:1;
429 /* per device-cgroup pair */
431 struct blkg_policy_data pd
;
435 * A iocg can get its weight from two sources - an explicit
436 * per-device-cgroup configuration or the default weight of the
437 * cgroup. `cfg_weight` is the explicit per-device-cgroup
438 * configuration. `weight` is the effective considering both
441 * When an idle cgroup becomes active its `active` goes from 0 to
442 * `weight`. `inuse` is the surplus adjusted active weight.
443 * `active` and `inuse` are used to calculate `hweight_active` and
446 * `last_inuse` remembers `inuse` while an iocg is idle to persist
447 * surplus adjustments.
455 sector_t cursor
; /* to detect randio */
458 * `vtime` is this iocg's vtime cursor which progresses as IOs are
459 * issued. If lagging behind device vtime, the delta represents
460 * the currently available IO budget. If runnning ahead, the
463 * `vtime_done` is the same but progressed on completion rather
464 * than issue. The delta behind `vtime` represents the cost of
465 * currently in-flight IOs.
467 * `last_vtime` is used to remember `vtime` at the end of the last
468 * period to calculate utilization.
471 atomic64_t done_vtime
;
476 * The period this iocg was last active in. Used for deactivation
477 * and invalidating `vtime`.
479 atomic64_t active_period
;
480 struct list_head active_list
;
482 /* see __propagate_weights() and current_hweight() for details */
483 u64 child_active_sum
;
490 struct wait_queue_head waitq
;
491 struct hrtimer waitq_timer
;
492 struct hrtimer delay_timer
;
494 /* usage is recorded as fractions of HWEIGHT_WHOLE */
496 u32 usages
[NR_USAGE_SLOTS
];
498 /* this iocg's depth in the hierarchy and ancestors including self */
500 struct ioc_gq
*ancestors
[];
505 struct blkcg_policy_data cpd
;
506 unsigned int dfl_weight
;
517 struct wait_queue_entry wait
;
523 struct iocg_wake_ctx
{
529 static const struct ioc_params autop
[] = {
532 [QOS_RLAT
] = 250000, /* 250ms */
534 [QOS_MIN
] = VRATE_MIN_PPM
,
535 [QOS_MAX
] = VRATE_MAX_PPM
,
538 [I_LCOEF_RBPS
] = 174019176,
539 [I_LCOEF_RSEQIOPS
] = 41708,
540 [I_LCOEF_RRANDIOPS
] = 370,
541 [I_LCOEF_WBPS
] = 178075866,
542 [I_LCOEF_WSEQIOPS
] = 42705,
543 [I_LCOEF_WRANDIOPS
] = 378,
548 [QOS_RLAT
] = 25000, /* 25ms */
550 [QOS_MIN
] = VRATE_MIN_PPM
,
551 [QOS_MAX
] = VRATE_MAX_PPM
,
554 [I_LCOEF_RBPS
] = 245855193,
555 [I_LCOEF_RSEQIOPS
] = 61575,
556 [I_LCOEF_RRANDIOPS
] = 6946,
557 [I_LCOEF_WBPS
] = 141365009,
558 [I_LCOEF_WSEQIOPS
] = 33716,
559 [I_LCOEF_WRANDIOPS
] = 26796,
564 [QOS_RLAT
] = 25000, /* 25ms */
566 [QOS_MIN
] = VRATE_MIN_PPM
,
567 [QOS_MAX
] = VRATE_MAX_PPM
,
570 [I_LCOEF_RBPS
] = 488636629,
571 [I_LCOEF_RSEQIOPS
] = 8932,
572 [I_LCOEF_RRANDIOPS
] = 8518,
573 [I_LCOEF_WBPS
] = 427891549,
574 [I_LCOEF_WSEQIOPS
] = 28755,
575 [I_LCOEF_WRANDIOPS
] = 21940,
577 .too_fast_vrate_pct
= 500,
581 [QOS_RLAT
] = 5000, /* 5ms */
583 [QOS_MIN
] = VRATE_MIN_PPM
,
584 [QOS_MAX
] = VRATE_MAX_PPM
,
587 [I_LCOEF_RBPS
] = 3102524156LLU,
588 [I_LCOEF_RSEQIOPS
] = 724816,
589 [I_LCOEF_RRANDIOPS
] = 778122,
590 [I_LCOEF_WBPS
] = 1742780862LLU,
591 [I_LCOEF_WSEQIOPS
] = 425702,
592 [I_LCOEF_WRANDIOPS
] = 443193,
594 .too_slow_vrate_pct
= 10,
599 * vrate adjust percentages indexed by ioc->busy_level. We adjust up on
600 * vtime credit shortage and down on device saturation.
602 static u32 vrate_adj_pct
[] =
604 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
605 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2,
606 4, 4, 4, 4, 4, 4, 4, 4, 8, 8, 8, 8, 8, 8, 8, 8, 16 };
608 static struct blkcg_policy blkcg_policy_iocost
;
610 /* accessors and helpers */
611 static struct ioc
*rqos_to_ioc(struct rq_qos
*rqos
)
613 return container_of(rqos
, struct ioc
, rqos
);
616 static struct ioc
*q_to_ioc(struct request_queue
*q
)
618 return rqos_to_ioc(rq_qos_id(q
, RQ_QOS_COST
));
621 static const char *q_name(struct request_queue
*q
)
623 if (test_bit(QUEUE_FLAG_REGISTERED
, &q
->queue_flags
))
624 return kobject_name(q
->kobj
.parent
);
629 static const char __maybe_unused
*ioc_name(struct ioc
*ioc
)
631 return q_name(ioc
->rqos
.q
);
634 static struct ioc_gq
*pd_to_iocg(struct blkg_policy_data
*pd
)
636 return pd
? container_of(pd
, struct ioc_gq
, pd
) : NULL
;
639 static struct ioc_gq
*blkg_to_iocg(struct blkcg_gq
*blkg
)
641 return pd_to_iocg(blkg_to_pd(blkg
, &blkcg_policy_iocost
));
644 static struct blkcg_gq
*iocg_to_blkg(struct ioc_gq
*iocg
)
646 return pd_to_blkg(&iocg
->pd
);
649 static struct ioc_cgrp
*blkcg_to_iocc(struct blkcg
*blkcg
)
651 return container_of(blkcg_to_cpd(blkcg
, &blkcg_policy_iocost
),
652 struct ioc_cgrp
, cpd
);
656 * Scale @abs_cost to the inverse of @hw_inuse. The lower the hierarchical
657 * weight, the more expensive each IO. Must round up.
659 static u64
abs_cost_to_cost(u64 abs_cost
, u32 hw_inuse
)
661 return DIV64_U64_ROUND_UP(abs_cost
* HWEIGHT_WHOLE
, hw_inuse
);
665 * The inverse of abs_cost_to_cost(). Must round up.
667 static u64
cost_to_abs_cost(u64 cost
, u32 hw_inuse
)
669 return DIV64_U64_ROUND_UP(cost
* hw_inuse
, HWEIGHT_WHOLE
);
672 static void iocg_commit_bio(struct ioc_gq
*iocg
, struct bio
*bio
, u64 cost
)
674 bio
->bi_iocost_cost
= cost
;
675 atomic64_add(cost
, &iocg
->vtime
);
678 #define CREATE_TRACE_POINTS
679 #include <trace/events/iocost.h>
681 /* latency Qos params changed, update period_us and all the dependent params */
682 static void ioc_refresh_period_us(struct ioc
*ioc
)
684 u32 ppm
, lat
, multi
, period_us
;
686 lockdep_assert_held(&ioc
->lock
);
688 /* pick the higher latency target */
689 if (ioc
->params
.qos
[QOS_RLAT
] >= ioc
->params
.qos
[QOS_WLAT
]) {
690 ppm
= ioc
->params
.qos
[QOS_RPPM
];
691 lat
= ioc
->params
.qos
[QOS_RLAT
];
693 ppm
= ioc
->params
.qos
[QOS_WPPM
];
694 lat
= ioc
->params
.qos
[QOS_WLAT
];
698 * We want the period to be long enough to contain a healthy number
699 * of IOs while short enough for granular control. Define it as a
700 * multiple of the latency target. Ideally, the multiplier should
701 * be scaled according to the percentile so that it would nominally
702 * contain a certain number of requests. Let's be simpler and
703 * scale it linearly so that it's 2x >= pct(90) and 10x at pct(50).
706 multi
= max_t(u32
, (MILLION
- ppm
) / 50000, 2);
709 period_us
= multi
* lat
;
710 period_us
= clamp_t(u32
, period_us
, MIN_PERIOD
, MAX_PERIOD
);
712 /* calculate dependent params */
713 ioc
->period_us
= period_us
;
714 ioc
->margin_us
= period_us
* MARGIN_PCT
/ 100;
715 ioc
->inuse_margin_vtime
= DIV64_U64_ROUND_UP(
716 period_us
* VTIME_PER_USEC
* INUSE_MARGIN_PCT
, 100);
719 static int ioc_autop_idx(struct ioc
*ioc
)
721 int idx
= ioc
->autop_idx
;
722 const struct ioc_params
*p
= &autop
[idx
];
727 if (!blk_queue_nonrot(ioc
->rqos
.q
))
730 /* handle SATA SSDs w/ broken NCQ */
731 if (blk_queue_depth(ioc
->rqos
.q
) == 1)
732 return AUTOP_SSD_QD1
;
734 /* use one of the normal ssd sets */
735 if (idx
< AUTOP_SSD_DFL
)
736 return AUTOP_SSD_DFL
;
738 /* if user is overriding anything, maintain what was there */
739 if (ioc
->user_qos_params
|| ioc
->user_cost_model
)
742 /* step up/down based on the vrate */
743 vrate_pct
= div64_u64(atomic64_read(&ioc
->vtime_rate
) * 100,
745 now_ns
= ktime_get_ns();
747 if (p
->too_fast_vrate_pct
&& p
->too_fast_vrate_pct
<= vrate_pct
) {
748 if (!ioc
->autop_too_fast_at
)
749 ioc
->autop_too_fast_at
= now_ns
;
750 if (now_ns
- ioc
->autop_too_fast_at
>= AUTOP_CYCLE_NSEC
)
753 ioc
->autop_too_fast_at
= 0;
756 if (p
->too_slow_vrate_pct
&& p
->too_slow_vrate_pct
>= vrate_pct
) {
757 if (!ioc
->autop_too_slow_at
)
758 ioc
->autop_too_slow_at
= now_ns
;
759 if (now_ns
- ioc
->autop_too_slow_at
>= AUTOP_CYCLE_NSEC
)
762 ioc
->autop_too_slow_at
= 0;
769 * Take the followings as input
771 * @bps maximum sequential throughput
772 * @seqiops maximum sequential 4k iops
773 * @randiops maximum random 4k iops
775 * and calculate the linear model cost coefficients.
777 * *@page per-page cost 1s / (@bps / 4096)
778 * *@seqio base cost of a seq IO max((1s / @seqiops) - *@page, 0)
779 * @randiops base cost of a rand IO max((1s / @randiops) - *@page, 0)
781 static void calc_lcoefs(u64 bps
, u64 seqiops
, u64 randiops
,
782 u64
*page
, u64
*seqio
, u64
*randio
)
786 *page
= *seqio
= *randio
= 0;
789 *page
= DIV64_U64_ROUND_UP(VTIME_PER_SEC
,
790 DIV_ROUND_UP_ULL(bps
, IOC_PAGE_SIZE
));
793 v
= DIV64_U64_ROUND_UP(VTIME_PER_SEC
, seqiops
);
799 v
= DIV64_U64_ROUND_UP(VTIME_PER_SEC
, randiops
);
805 static void ioc_refresh_lcoefs(struct ioc
*ioc
)
807 u64
*u
= ioc
->params
.i_lcoefs
;
808 u64
*c
= ioc
->params
.lcoefs
;
810 calc_lcoefs(u
[I_LCOEF_RBPS
], u
[I_LCOEF_RSEQIOPS
], u
[I_LCOEF_RRANDIOPS
],
811 &c
[LCOEF_RPAGE
], &c
[LCOEF_RSEQIO
], &c
[LCOEF_RRANDIO
]);
812 calc_lcoefs(u
[I_LCOEF_WBPS
], u
[I_LCOEF_WSEQIOPS
], u
[I_LCOEF_WRANDIOPS
],
813 &c
[LCOEF_WPAGE
], &c
[LCOEF_WSEQIO
], &c
[LCOEF_WRANDIO
]);
816 static bool ioc_refresh_params(struct ioc
*ioc
, bool force
)
818 const struct ioc_params
*p
;
821 lockdep_assert_held(&ioc
->lock
);
823 idx
= ioc_autop_idx(ioc
);
826 if (idx
== ioc
->autop_idx
&& !force
)
829 if (idx
!= ioc
->autop_idx
)
830 atomic64_set(&ioc
->vtime_rate
, VTIME_PER_USEC
);
832 ioc
->autop_idx
= idx
;
833 ioc
->autop_too_fast_at
= 0;
834 ioc
->autop_too_slow_at
= 0;
836 if (!ioc
->user_qos_params
)
837 memcpy(ioc
->params
.qos
, p
->qos
, sizeof(p
->qos
));
838 if (!ioc
->user_cost_model
)
839 memcpy(ioc
->params
.i_lcoefs
, p
->i_lcoefs
, sizeof(p
->i_lcoefs
));
841 ioc_refresh_period_us(ioc
);
842 ioc_refresh_lcoefs(ioc
);
844 ioc
->vrate_min
= DIV64_U64_ROUND_UP((u64
)ioc
->params
.qos
[QOS_MIN
] *
845 VTIME_PER_USEC
, MILLION
);
846 ioc
->vrate_max
= div64_u64((u64
)ioc
->params
.qos
[QOS_MAX
] *
847 VTIME_PER_USEC
, MILLION
);
852 /* take a snapshot of the current [v]time and vrate */
853 static void ioc_now(struct ioc
*ioc
, struct ioc_now
*now
)
857 now
->now_ns
= ktime_get();
858 now
->now
= ktime_to_us(now
->now_ns
);
859 now
->vrate
= atomic64_read(&ioc
->vtime_rate
);
862 * The current vtime is
864 * vtime at period start + (wallclock time since the start) * vrate
866 * As a consistent snapshot of `period_at_vtime` and `period_at` is
867 * needed, they're seqcount protected.
870 seq
= read_seqcount_begin(&ioc
->period_seqcount
);
871 now
->vnow
= ioc
->period_at_vtime
+
872 (now
->now
- ioc
->period_at
) * now
->vrate
;
873 } while (read_seqcount_retry(&ioc
->period_seqcount
, seq
));
876 static void ioc_start_period(struct ioc
*ioc
, struct ioc_now
*now
)
878 WARN_ON_ONCE(ioc
->running
!= IOC_RUNNING
);
880 write_seqcount_begin(&ioc
->period_seqcount
);
881 ioc
->period_at
= now
->now
;
882 ioc
->period_at_vtime
= now
->vnow
;
883 write_seqcount_end(&ioc
->period_seqcount
);
885 ioc
->timer
.expires
= jiffies
+ usecs_to_jiffies(ioc
->period_us
);
886 add_timer(&ioc
->timer
);
890 * Update @iocg's `active` and `inuse` to @active and @inuse, update level
891 * weight sums and propagate upwards accordingly.
893 static void __propagate_weights(struct ioc_gq
*iocg
, u32 active
, u32 inuse
)
895 struct ioc
*ioc
= iocg
->ioc
;
898 lockdep_assert_held(&ioc
->lock
);
900 inuse
= clamp_t(u32
, inuse
, 1, active
);
902 if (active
== iocg
->active
&& inuse
== iocg
->inuse
)
905 for (lvl
= iocg
->level
- 1; lvl
>= 0; lvl
--) {
906 struct ioc_gq
*parent
= iocg
->ancestors
[lvl
];
907 struct ioc_gq
*child
= iocg
->ancestors
[lvl
+ 1];
908 u32 parent_active
= 0, parent_inuse
= 0;
910 /* update the level sums */
911 parent
->child_active_sum
+= (s32
)(active
- child
->active
);
912 parent
->child_inuse_sum
+= (s32
)(inuse
- child
->inuse
);
913 /* apply the udpates */
914 child
->active
= active
;
915 child
->inuse
= inuse
;
918 * The delta between inuse and active sums indicates that
919 * that much of weight is being given away. Parent's inuse
920 * and active should reflect the ratio.
922 if (parent
->child_active_sum
) {
923 parent_active
= parent
->weight
;
924 parent_inuse
= DIV64_U64_ROUND_UP(
925 parent_active
* parent
->child_inuse_sum
,
926 parent
->child_active_sum
);
929 /* do we need to keep walking up? */
930 if (parent_active
== parent
->active
&&
931 parent_inuse
== parent
->inuse
)
934 active
= parent_active
;
935 inuse
= parent_inuse
;
938 ioc
->weights_updated
= true;
941 static void commit_weights(struct ioc
*ioc
)
943 lockdep_assert_held(&ioc
->lock
);
945 if (ioc
->weights_updated
) {
946 /* paired with rmb in current_hweight(), see there */
948 atomic_inc(&ioc
->hweight_gen
);
949 ioc
->weights_updated
= false;
953 static void propagate_weights(struct ioc_gq
*iocg
, u32 active
, u32 inuse
)
955 __propagate_weights(iocg
, active
, inuse
);
956 commit_weights(iocg
->ioc
);
959 static void current_hweight(struct ioc_gq
*iocg
, u32
*hw_activep
, u32
*hw_inusep
)
961 struct ioc
*ioc
= iocg
->ioc
;
966 /* hot path - if uptodate, use cached */
967 ioc_gen
= atomic_read(&ioc
->hweight_gen
);
968 if (ioc_gen
== iocg
->hweight_gen
)
972 * Paired with wmb in commit_weights(). If we saw the updated
973 * hweight_gen, all the weight updates from __propagate_weights() are
976 * We can race with weight updates during calculation and get it
977 * wrong. However, hweight_gen would have changed and a future
978 * reader will recalculate and we're guaranteed to discard the
983 hwa
= hwi
= HWEIGHT_WHOLE
;
984 for (lvl
= 0; lvl
<= iocg
->level
- 1; lvl
++) {
985 struct ioc_gq
*parent
= iocg
->ancestors
[lvl
];
986 struct ioc_gq
*child
= iocg
->ancestors
[lvl
+ 1];
987 u32 active_sum
= READ_ONCE(parent
->child_active_sum
);
988 u32 inuse_sum
= READ_ONCE(parent
->child_inuse_sum
);
989 u32 active
= READ_ONCE(child
->active
);
990 u32 inuse
= READ_ONCE(child
->inuse
);
992 /* we can race with deactivations and either may read as zero */
993 if (!active_sum
|| !inuse_sum
)
996 active_sum
= max(active
, active_sum
);
997 hwa
= hwa
* active
/ active_sum
; /* max 16bits * 10000 */
999 inuse_sum
= max(inuse
, inuse_sum
);
1000 hwi
= hwi
* inuse
/ inuse_sum
; /* max 16bits * 10000 */
1003 iocg
->hweight_active
= max_t(u32
, hwa
, 1);
1004 iocg
->hweight_inuse
= max_t(u32
, hwi
, 1);
1005 iocg
->hweight_gen
= ioc_gen
;
1008 *hw_activep
= iocg
->hweight_active
;
1010 *hw_inusep
= iocg
->hweight_inuse
;
1013 static void weight_updated(struct ioc_gq
*iocg
)
1015 struct ioc
*ioc
= iocg
->ioc
;
1016 struct blkcg_gq
*blkg
= iocg_to_blkg(iocg
);
1017 struct ioc_cgrp
*iocc
= blkcg_to_iocc(blkg
->blkcg
);
1020 lockdep_assert_held(&ioc
->lock
);
1022 weight
= iocg
->cfg_weight
?: iocc
->dfl_weight
;
1023 if (weight
!= iocg
->weight
&& iocg
->active
)
1024 propagate_weights(iocg
, weight
,
1025 DIV64_U64_ROUND_UP(iocg
->inuse
* weight
, iocg
->weight
));
1026 iocg
->weight
= weight
;
1029 static bool iocg_activate(struct ioc_gq
*iocg
, struct ioc_now
*now
)
1031 struct ioc
*ioc
= iocg
->ioc
;
1032 u64 last_period
, cur_period
, max_period_delta
;
1033 u64 vtime
, vmargin
, vmin
;
1037 * If seem to be already active, just update the stamp to tell the
1038 * timer that we're still active. We don't mind occassional races.
1040 if (!list_empty(&iocg
->active_list
)) {
1042 cur_period
= atomic64_read(&ioc
->cur_period
);
1043 if (atomic64_read(&iocg
->active_period
) != cur_period
)
1044 atomic64_set(&iocg
->active_period
, cur_period
);
1048 /* racy check on internal node IOs, treat as root level IOs */
1049 if (iocg
->child_active_sum
)
1052 spin_lock_irq(&ioc
->lock
);
1057 cur_period
= atomic64_read(&ioc
->cur_period
);
1058 last_period
= atomic64_read(&iocg
->active_period
);
1059 atomic64_set(&iocg
->active_period
, cur_period
);
1061 /* already activated or breaking leaf-only constraint? */
1062 if (!list_empty(&iocg
->active_list
))
1063 goto succeed_unlock
;
1064 for (i
= iocg
->level
- 1; i
> 0; i
--)
1065 if (!list_empty(&iocg
->ancestors
[i
]->active_list
))
1068 if (iocg
->child_active_sum
)
1072 * vtime may wrap when vrate is raised substantially due to
1073 * underestimated IO costs. Look at the period and ignore its
1074 * vtime if the iocg has been idle for too long. Also, cap the
1075 * budget it can start with to the margin.
1077 max_period_delta
= DIV64_U64_ROUND_UP(VTIME_VALID_DUR
, ioc
->period_us
);
1078 vtime
= atomic64_read(&iocg
->vtime
);
1079 vmargin
= ioc
->margin_us
* now
->vrate
;
1080 vmin
= now
->vnow
- vmargin
;
1082 if (last_period
+ max_period_delta
< cur_period
||
1083 time_before64(vtime
, vmin
)) {
1084 atomic64_add(vmin
- vtime
, &iocg
->vtime
);
1085 atomic64_add(vmin
- vtime
, &iocg
->done_vtime
);
1090 * Activate, propagate weight and start period timer if not
1091 * running. Reset hweight_gen to avoid accidental match from
1094 iocg
->hweight_gen
= atomic_read(&ioc
->hweight_gen
) - 1;
1095 list_add(&iocg
->active_list
, &ioc
->active_iocgs
);
1096 propagate_weights(iocg
, iocg
->weight
,
1097 iocg
->last_inuse
?: iocg
->weight
);
1099 TRACE_IOCG_PATH(iocg_activate
, iocg
, now
,
1100 last_period
, cur_period
, vtime
);
1102 iocg
->last_vtime
= vtime
;
1104 if (ioc
->running
== IOC_IDLE
) {
1105 ioc
->running
= IOC_RUNNING
;
1106 ioc_start_period(ioc
, now
);
1110 spin_unlock_irq(&ioc
->lock
);
1114 spin_unlock_irq(&ioc
->lock
);
1118 static int iocg_wake_fn(struct wait_queue_entry
*wq_entry
, unsigned mode
,
1119 int flags
, void *key
)
1121 struct iocg_wait
*wait
= container_of(wq_entry
, struct iocg_wait
, wait
);
1122 struct iocg_wake_ctx
*ctx
= (struct iocg_wake_ctx
*)key
;
1123 u64 cost
= abs_cost_to_cost(wait
->abs_cost
, ctx
->hw_inuse
);
1125 ctx
->vbudget
-= cost
;
1127 if (ctx
->vbudget
< 0)
1130 iocg_commit_bio(ctx
->iocg
, wait
->bio
, cost
);
1133 * autoremove_wake_function() removes the wait entry only when it
1134 * actually changed the task state. We want the wait always
1135 * removed. Remove explicitly and use default_wake_function().
1137 list_del_init(&wq_entry
->entry
);
1138 wait
->committed
= true;
1140 default_wake_function(wq_entry
, mode
, flags
, key
);
1144 static void iocg_kick_waitq(struct ioc_gq
*iocg
, struct ioc_now
*now
)
1146 struct ioc
*ioc
= iocg
->ioc
;
1147 struct iocg_wake_ctx ctx
= { .iocg
= iocg
};
1148 u64 margin_ns
= (u64
)(ioc
->period_us
*
1149 WAITQ_TIMER_MARGIN_PCT
/ 100) * NSEC_PER_USEC
;
1150 u64 vdebt
, vshortage
, expires
, oexpires
;
1154 lockdep_assert_held(&iocg
->waitq
.lock
);
1156 current_hweight(iocg
, NULL
, &hw_inuse
);
1157 vbudget
= now
->vnow
- atomic64_read(&iocg
->vtime
);
1160 vdebt
= abs_cost_to_cost(iocg
->abs_vdebt
, hw_inuse
);
1161 if (vdebt
&& vbudget
> 0) {
1162 u64 delta
= min_t(u64
, vbudget
, vdebt
);
1163 u64 abs_delta
= min(cost_to_abs_cost(delta
, hw_inuse
),
1166 atomic64_add(delta
, &iocg
->vtime
);
1167 atomic64_add(delta
, &iocg
->done_vtime
);
1168 iocg
->abs_vdebt
-= abs_delta
;
1172 * Wake up the ones which are due and see how much vtime we'll need
1175 ctx
.hw_inuse
= hw_inuse
;
1176 ctx
.vbudget
= vbudget
- vdebt
;
1177 __wake_up_locked_key(&iocg
->waitq
, TASK_NORMAL
, &ctx
);
1178 if (!waitqueue_active(&iocg
->waitq
))
1180 if (WARN_ON_ONCE(ctx
.vbudget
>= 0))
1183 /* determine next wakeup, add a quarter margin to guarantee chunking */
1184 vshortage
= -ctx
.vbudget
;
1185 expires
= now
->now_ns
+
1186 DIV64_U64_ROUND_UP(vshortage
, now
->vrate
) * NSEC_PER_USEC
;
1187 expires
+= margin_ns
/ 4;
1189 /* if already active and close enough, don't bother */
1190 oexpires
= ktime_to_ns(hrtimer_get_softexpires(&iocg
->waitq_timer
));
1191 if (hrtimer_is_queued(&iocg
->waitq_timer
) &&
1192 abs(oexpires
- expires
) <= margin_ns
/ 4)
1195 hrtimer_start_range_ns(&iocg
->waitq_timer
, ns_to_ktime(expires
),
1196 margin_ns
/ 4, HRTIMER_MODE_ABS
);
1199 static enum hrtimer_restart
iocg_waitq_timer_fn(struct hrtimer
*timer
)
1201 struct ioc_gq
*iocg
= container_of(timer
, struct ioc_gq
, waitq_timer
);
1203 unsigned long flags
;
1205 ioc_now(iocg
->ioc
, &now
);
1207 spin_lock_irqsave(&iocg
->waitq
.lock
, flags
);
1208 iocg_kick_waitq(iocg
, &now
);
1209 spin_unlock_irqrestore(&iocg
->waitq
.lock
, flags
);
1211 return HRTIMER_NORESTART
;
1214 static bool iocg_kick_delay(struct ioc_gq
*iocg
, struct ioc_now
*now
)
1216 struct ioc
*ioc
= iocg
->ioc
;
1217 struct blkcg_gq
*blkg
= iocg_to_blkg(iocg
);
1218 u64 vtime
= atomic64_read(&iocg
->vtime
);
1219 u64 vmargin
= ioc
->margin_us
* now
->vrate
;
1220 u64 margin_ns
= ioc
->margin_us
* NSEC_PER_USEC
;
1221 u64 delta_ns
, expires
, oexpires
;
1224 lockdep_assert_held(&iocg
->waitq
.lock
);
1226 /* debt-adjust vtime */
1227 current_hweight(iocg
, NULL
, &hw_inuse
);
1228 vtime
+= abs_cost_to_cost(iocg
->abs_vdebt
, hw_inuse
);
1231 * Clear or maintain depending on the overage. Non-zero vdebt is what
1232 * guarantees that @iocg is online and future iocg_kick_delay() will
1233 * clear use_delay. Don't leave it on when there's no vdebt.
1235 if (!iocg
->abs_vdebt
|| time_before_eq64(vtime
, now
->vnow
)) {
1236 blkcg_clear_delay(blkg
);
1239 if (!atomic_read(&blkg
->use_delay
) &&
1240 time_before_eq64(vtime
, now
->vnow
+ vmargin
))
1244 delta_ns
= DIV64_U64_ROUND_UP(vtime
- now
->vnow
,
1245 now
->vrate
) * NSEC_PER_USEC
;
1246 blkcg_set_delay(blkg
, delta_ns
);
1247 expires
= now
->now_ns
+ delta_ns
;
1249 /* if already active and close enough, don't bother */
1250 oexpires
= ktime_to_ns(hrtimer_get_softexpires(&iocg
->delay_timer
));
1251 if (hrtimer_is_queued(&iocg
->delay_timer
) &&
1252 abs(oexpires
- expires
) <= margin_ns
/ 4)
1255 hrtimer_start_range_ns(&iocg
->delay_timer
, ns_to_ktime(expires
),
1256 margin_ns
/ 4, HRTIMER_MODE_ABS
);
1260 static enum hrtimer_restart
iocg_delay_timer_fn(struct hrtimer
*timer
)
1262 struct ioc_gq
*iocg
= container_of(timer
, struct ioc_gq
, delay_timer
);
1264 unsigned long flags
;
1266 spin_lock_irqsave(&iocg
->waitq
.lock
, flags
);
1267 ioc_now(iocg
->ioc
, &now
);
1268 iocg_kick_delay(iocg
, &now
);
1269 spin_unlock_irqrestore(&iocg
->waitq
.lock
, flags
);
1271 return HRTIMER_NORESTART
;
1274 static void ioc_lat_stat(struct ioc
*ioc
, u32
*missed_ppm_ar
, u32
*rq_wait_pct_p
)
1276 u32 nr_met
[2] = { };
1277 u32 nr_missed
[2] = { };
1281 for_each_online_cpu(cpu
) {
1282 struct ioc_pcpu_stat
*stat
= per_cpu_ptr(ioc
->pcpu_stat
, cpu
);
1283 u64 this_rq_wait_ns
;
1285 for (rw
= READ
; rw
<= WRITE
; rw
++) {
1286 u32 this_met
= local_read(&stat
->missed
[rw
].nr_met
);
1287 u32 this_missed
= local_read(&stat
->missed
[rw
].nr_missed
);
1289 nr_met
[rw
] += this_met
- stat
->missed
[rw
].last_met
;
1290 nr_missed
[rw
] += this_missed
- stat
->missed
[rw
].last_missed
;
1291 stat
->missed
[rw
].last_met
= this_met
;
1292 stat
->missed
[rw
].last_missed
= this_missed
;
1295 this_rq_wait_ns
= local64_read(&stat
->rq_wait_ns
);
1296 rq_wait_ns
+= this_rq_wait_ns
- stat
->last_rq_wait_ns
;
1297 stat
->last_rq_wait_ns
= this_rq_wait_ns
;
1300 for (rw
= READ
; rw
<= WRITE
; rw
++) {
1301 if (nr_met
[rw
] + nr_missed
[rw
])
1303 DIV64_U64_ROUND_UP((u64
)nr_missed
[rw
] * MILLION
,
1304 nr_met
[rw
] + nr_missed
[rw
]);
1306 missed_ppm_ar
[rw
] = 0;
1309 *rq_wait_pct_p
= div64_u64(rq_wait_ns
* 100,
1310 ioc
->period_us
* NSEC_PER_USEC
);
1313 /* was iocg idle this period? */
1314 static bool iocg_is_idle(struct ioc_gq
*iocg
)
1316 struct ioc
*ioc
= iocg
->ioc
;
1318 /* did something get issued this period? */
1319 if (atomic64_read(&iocg
->active_period
) ==
1320 atomic64_read(&ioc
->cur_period
))
1323 /* is something in flight? */
1324 if (atomic64_read(&iocg
->done_vtime
) != atomic64_read(&iocg
->vtime
))
1330 /* returns usage with margin added if surplus is large enough */
1331 static u32
surplus_adjusted_hweight_inuse(u32 usage
, u32 hw_inuse
)
1334 usage
= DIV_ROUND_UP(usage
* SURPLUS_SCALE_PCT
, 100);
1335 usage
+= SURPLUS_SCALE_ABS
;
1337 /* don't bother if the surplus is too small */
1338 if (usage
+ SURPLUS_MIN_ADJ_DELTA
> hw_inuse
)
1344 static void ioc_timer_fn(struct timer_list
*timer
)
1346 struct ioc
*ioc
= container_of(timer
, struct ioc
, timer
);
1347 struct ioc_gq
*iocg
, *tiocg
;
1349 int nr_surpluses
= 0, nr_shortages
= 0, nr_lagging
= 0;
1350 u32 ppm_rthr
= MILLION
- ioc
->params
.qos
[QOS_RPPM
];
1351 u32 ppm_wthr
= MILLION
- ioc
->params
.qos
[QOS_WPPM
];
1352 u32 missed_ppm
[2], rq_wait_pct
;
1354 int prev_busy_level
, i
;
1356 /* how were the latencies during the period? */
1357 ioc_lat_stat(ioc
, missed_ppm
, &rq_wait_pct
);
1359 /* take care of active iocgs */
1360 spin_lock_irq(&ioc
->lock
);
1364 period_vtime
= now
.vnow
- ioc
->period_at_vtime
;
1365 if (WARN_ON_ONCE(!period_vtime
)) {
1366 spin_unlock_irq(&ioc
->lock
);
1371 * Waiters determine the sleep durations based on the vrate they
1372 * saw at the time of sleep. If vrate has increased, some waiters
1373 * could be sleeping for too long. Wake up tardy waiters which
1374 * should have woken up in the last period and expire idle iocgs.
1376 list_for_each_entry_safe(iocg
, tiocg
, &ioc
->active_iocgs
, active_list
) {
1377 if (!waitqueue_active(&iocg
->waitq
) && !iocg
->abs_vdebt
&&
1378 !iocg_is_idle(iocg
))
1381 spin_lock(&iocg
->waitq
.lock
);
1383 if (waitqueue_active(&iocg
->waitq
) || iocg
->abs_vdebt
) {
1384 /* might be oversleeping vtime / hweight changes, kick */
1385 iocg_kick_waitq(iocg
, &now
);
1386 iocg_kick_delay(iocg
, &now
);
1387 } else if (iocg_is_idle(iocg
)) {
1388 /* no waiter and idle, deactivate */
1389 iocg
->last_inuse
= iocg
->inuse
;
1390 __propagate_weights(iocg
, 0, 0);
1391 list_del_init(&iocg
->active_list
);
1394 spin_unlock(&iocg
->waitq
.lock
);
1396 commit_weights(ioc
);
1398 /* calc usages and see whether some weights need to be moved around */
1399 list_for_each_entry(iocg
, &ioc
->active_iocgs
, active_list
) {
1400 u64 vdone
, vtime
, vusage
, vmargin
, vmin
;
1401 u32 hw_active
, hw_inuse
, usage
;
1404 * Collect unused and wind vtime closer to vnow to prevent
1405 * iocgs from accumulating a large amount of budget.
1407 vdone
= atomic64_read(&iocg
->done_vtime
);
1408 vtime
= atomic64_read(&iocg
->vtime
);
1409 current_hweight(iocg
, &hw_active
, &hw_inuse
);
1412 * Latency QoS detection doesn't account for IOs which are
1413 * in-flight for longer than a period. Detect them by
1414 * comparing vdone against period start. If lagging behind
1415 * IOs from past periods, don't increase vrate.
1417 if ((ppm_rthr
!= MILLION
|| ppm_wthr
!= MILLION
) &&
1418 !atomic_read(&iocg_to_blkg(iocg
)->use_delay
) &&
1419 time_after64(vtime
, vdone
) &&
1420 time_after64(vtime
, now
.vnow
-
1421 MAX_LAGGING_PERIODS
* period_vtime
) &&
1422 time_before64(vdone
, now
.vnow
- period_vtime
))
1425 if (waitqueue_active(&iocg
->waitq
))
1426 vusage
= now
.vnow
- iocg
->last_vtime
;
1427 else if (time_before64(iocg
->last_vtime
, vtime
))
1428 vusage
= vtime
- iocg
->last_vtime
;
1432 iocg
->last_vtime
+= vusage
;
1434 * Factor in in-flight vtime into vusage to avoid
1435 * high-latency completions appearing as idle. This should
1436 * be done after the above ->last_time adjustment.
1438 vusage
= max(vusage
, vtime
- vdone
);
1440 /* calculate hweight based usage ratio and record */
1442 usage
= DIV64_U64_ROUND_UP(vusage
* hw_inuse
,
1444 iocg
->usage_idx
= (iocg
->usage_idx
+ 1) % NR_USAGE_SLOTS
;
1445 iocg
->usages
[iocg
->usage_idx
] = usage
;
1450 /* see whether there's surplus vtime */
1451 vmargin
= ioc
->margin_us
* now
.vrate
;
1452 vmin
= now
.vnow
- vmargin
;
1454 iocg
->has_surplus
= false;
1456 if (!waitqueue_active(&iocg
->waitq
) &&
1457 time_before64(vtime
, vmin
)) {
1458 u64 delta
= vmin
- vtime
;
1460 /* throw away surplus vtime */
1461 atomic64_add(delta
, &iocg
->vtime
);
1462 atomic64_add(delta
, &iocg
->done_vtime
);
1463 iocg
->last_vtime
+= delta
;
1464 /* if usage is sufficiently low, maybe it can donate */
1465 if (surplus_adjusted_hweight_inuse(usage
, hw_inuse
)) {
1466 iocg
->has_surplus
= true;
1469 } else if (hw_inuse
< hw_active
) {
1470 u32 new_hwi
, new_inuse
;
1472 /* was donating but might need to take back some */
1473 if (waitqueue_active(&iocg
->waitq
)) {
1474 new_hwi
= hw_active
;
1476 new_hwi
= max(hw_inuse
,
1477 usage
* SURPLUS_SCALE_PCT
/ 100 +
1481 new_inuse
= div64_u64((u64
)iocg
->inuse
* new_hwi
,
1483 new_inuse
= clamp_t(u32
, new_inuse
, 1, iocg
->active
);
1485 if (new_inuse
> iocg
->inuse
) {
1486 TRACE_IOCG_PATH(inuse_takeback
, iocg
, &now
,
1487 iocg
->inuse
, new_inuse
,
1489 __propagate_weights(iocg
, iocg
->weight
,
1493 /* genuninely out of vtime */
1498 if (!nr_shortages
|| !nr_surpluses
)
1499 goto skip_surplus_transfers
;
1501 /* there are both shortages and surpluses, transfer surpluses */
1502 list_for_each_entry(iocg
, &ioc
->active_iocgs
, active_list
) {
1503 u32 usage
, hw_active
, hw_inuse
, new_hwi
, new_inuse
;
1506 if (!iocg
->has_surplus
)
1509 /* base the decision on max historical usage */
1510 for (i
= 0, usage
= 0; i
< NR_USAGE_SLOTS
; i
++) {
1511 if (iocg
->usages
[i
]) {
1512 usage
= max(usage
, iocg
->usages
[i
]);
1516 if (nr_valid
< MIN_VALID_USAGES
)
1519 current_hweight(iocg
, &hw_active
, &hw_inuse
);
1520 new_hwi
= surplus_adjusted_hweight_inuse(usage
, hw_inuse
);
1524 new_inuse
= DIV64_U64_ROUND_UP((u64
)iocg
->inuse
* new_hwi
,
1526 if (new_inuse
< iocg
->inuse
) {
1527 TRACE_IOCG_PATH(inuse_giveaway
, iocg
, &now
,
1528 iocg
->inuse
, new_inuse
,
1530 __propagate_weights(iocg
, iocg
->weight
, new_inuse
);
1533 skip_surplus_transfers
:
1534 commit_weights(ioc
);
1537 * If q is getting clogged or we're missing too much, we're issuing
1538 * too much IO and should lower vtime rate. If we're not missing
1539 * and experiencing shortages but not surpluses, we're too stingy
1540 * and should increase vtime rate.
1542 prev_busy_level
= ioc
->busy_level
;
1543 if (rq_wait_pct
> RQ_WAIT_BUSY_PCT
||
1544 missed_ppm
[READ
] > ppm_rthr
||
1545 missed_ppm
[WRITE
] > ppm_wthr
) {
1546 /* clearly missing QoS targets, slow down vrate */
1547 ioc
->busy_level
= max(ioc
->busy_level
, 0);
1549 } else if (rq_wait_pct
<= RQ_WAIT_BUSY_PCT
* UNBUSY_THR_PCT
/ 100 &&
1550 missed_ppm
[READ
] <= ppm_rthr
* UNBUSY_THR_PCT
/ 100 &&
1551 missed_ppm
[WRITE
] <= ppm_wthr
* UNBUSY_THR_PCT
/ 100) {
1552 /* QoS targets are being met with >25% margin */
1555 * We're throttling while the device has spare
1556 * capacity. If vrate was being slowed down, stop.
1558 ioc
->busy_level
= min(ioc
->busy_level
, 0);
1561 * If there are IOs spanning multiple periods, wait
1562 * them out before pushing the device harder. If
1563 * there are surpluses, let redistribution work it
1566 if (!nr_lagging
&& !nr_surpluses
)
1570 * Nobody is being throttled and the users aren't
1571 * issuing enough IOs to saturate the device. We
1572 * simply don't know how close the device is to
1573 * saturation. Coast.
1575 ioc
->busy_level
= 0;
1578 /* inside the hysterisis margin, we're good */
1579 ioc
->busy_level
= 0;
1582 ioc
->busy_level
= clamp(ioc
->busy_level
, -1000, 1000);
1584 if (ioc
->busy_level
> 0 || (ioc
->busy_level
< 0 && !nr_lagging
)) {
1585 u64 vrate
= atomic64_read(&ioc
->vtime_rate
);
1586 u64 vrate_min
= ioc
->vrate_min
, vrate_max
= ioc
->vrate_max
;
1588 /* rq_wait signal is always reliable, ignore user vrate_min */
1589 if (rq_wait_pct
> RQ_WAIT_BUSY_PCT
)
1590 vrate_min
= VRATE_MIN
;
1593 * If vrate is out of bounds, apply clamp gradually as the
1594 * bounds can change abruptly. Otherwise, apply busy_level
1597 if (vrate
< vrate_min
) {
1598 vrate
= div64_u64(vrate
* (100 + VRATE_CLAMP_ADJ_PCT
),
1600 vrate
= min(vrate
, vrate_min
);
1601 } else if (vrate
> vrate_max
) {
1602 vrate
= div64_u64(vrate
* (100 - VRATE_CLAMP_ADJ_PCT
),
1604 vrate
= max(vrate
, vrate_max
);
1606 int idx
= min_t(int, abs(ioc
->busy_level
),
1607 ARRAY_SIZE(vrate_adj_pct
) - 1);
1608 u32 adj_pct
= vrate_adj_pct
[idx
];
1610 if (ioc
->busy_level
> 0)
1611 adj_pct
= 100 - adj_pct
;
1613 adj_pct
= 100 + adj_pct
;
1615 vrate
= clamp(DIV64_U64_ROUND_UP(vrate
* adj_pct
, 100),
1616 vrate_min
, vrate_max
);
1619 trace_iocost_ioc_vrate_adj(ioc
, vrate
, missed_ppm
, rq_wait_pct
,
1620 nr_lagging
, nr_shortages
,
1623 atomic64_set(&ioc
->vtime_rate
, vrate
);
1624 ioc
->inuse_margin_vtime
= DIV64_U64_ROUND_UP(
1625 ioc
->period_us
* vrate
* INUSE_MARGIN_PCT
, 100);
1626 } else if (ioc
->busy_level
!= prev_busy_level
|| nr_lagging
) {
1627 trace_iocost_ioc_vrate_adj(ioc
, atomic64_read(&ioc
->vtime_rate
),
1628 missed_ppm
, rq_wait_pct
, nr_lagging
,
1629 nr_shortages
, nr_surpluses
);
1632 ioc_refresh_params(ioc
, false);
1635 * This period is done. Move onto the next one. If nothing's
1636 * going on with the device, stop the timer.
1638 atomic64_inc(&ioc
->cur_period
);
1640 if (ioc
->running
!= IOC_STOP
) {
1641 if (!list_empty(&ioc
->active_iocgs
)) {
1642 ioc_start_period(ioc
, &now
);
1644 ioc
->busy_level
= 0;
1645 ioc
->running
= IOC_IDLE
;
1649 spin_unlock_irq(&ioc
->lock
);
1652 static void calc_vtime_cost_builtin(struct bio
*bio
, struct ioc_gq
*iocg
,
1653 bool is_merge
, u64
*costp
)
1655 struct ioc
*ioc
= iocg
->ioc
;
1656 u64 coef_seqio
, coef_randio
, coef_page
;
1657 u64 pages
= max_t(u64
, bio_sectors(bio
) >> IOC_SECT_TO_PAGE_SHIFT
, 1);
1661 switch (bio_op(bio
)) {
1663 coef_seqio
= ioc
->params
.lcoefs
[LCOEF_RSEQIO
];
1664 coef_randio
= ioc
->params
.lcoefs
[LCOEF_RRANDIO
];
1665 coef_page
= ioc
->params
.lcoefs
[LCOEF_RPAGE
];
1668 coef_seqio
= ioc
->params
.lcoefs
[LCOEF_WSEQIO
];
1669 coef_randio
= ioc
->params
.lcoefs
[LCOEF_WRANDIO
];
1670 coef_page
= ioc
->params
.lcoefs
[LCOEF_WPAGE
];
1677 seek_pages
= abs(bio
->bi_iter
.bi_sector
- iocg
->cursor
);
1678 seek_pages
>>= IOC_SECT_TO_PAGE_SHIFT
;
1682 if (seek_pages
> LCOEF_RANDIO_PAGES
) {
1683 cost
+= coef_randio
;
1688 cost
+= pages
* coef_page
;
1693 static u64
calc_vtime_cost(struct bio
*bio
, struct ioc_gq
*iocg
, bool is_merge
)
1697 calc_vtime_cost_builtin(bio
, iocg
, is_merge
, &cost
);
1701 static void calc_size_vtime_cost_builtin(struct request
*rq
, struct ioc
*ioc
,
1704 unsigned int pages
= blk_rq_stats_sectors(rq
) >> IOC_SECT_TO_PAGE_SHIFT
;
1706 switch (req_op(rq
)) {
1708 *costp
= pages
* ioc
->params
.lcoefs
[LCOEF_RPAGE
];
1711 *costp
= pages
* ioc
->params
.lcoefs
[LCOEF_WPAGE
];
1718 static u64
calc_size_vtime_cost(struct request
*rq
, struct ioc
*ioc
)
1722 calc_size_vtime_cost_builtin(rq
, ioc
, &cost
);
1726 static void ioc_rqos_throttle(struct rq_qos
*rqos
, struct bio
*bio
)
1728 struct blkcg_gq
*blkg
= bio
->bi_blkg
;
1729 struct ioc
*ioc
= rqos_to_ioc(rqos
);
1730 struct ioc_gq
*iocg
= blkg_to_iocg(blkg
);
1732 struct iocg_wait wait
;
1733 u32 hw_active
, hw_inuse
;
1734 u64 abs_cost
, cost
, vtime
;
1736 /* bypass IOs if disabled or for root cgroup */
1737 if (!ioc
->enabled
|| !iocg
->level
)
1740 /* always activate so that even 0 cost IOs get protected to some level */
1741 if (!iocg_activate(iocg
, &now
))
1744 /* calculate the absolute vtime cost */
1745 abs_cost
= calc_vtime_cost(bio
, iocg
, false);
1749 iocg
->cursor
= bio_end_sector(bio
);
1751 vtime
= atomic64_read(&iocg
->vtime
);
1752 current_hweight(iocg
, &hw_active
, &hw_inuse
);
1754 if (hw_inuse
< hw_active
&&
1755 time_after_eq64(vtime
+ ioc
->inuse_margin_vtime
, now
.vnow
)) {
1756 TRACE_IOCG_PATH(inuse_reset
, iocg
, &now
,
1757 iocg
->inuse
, iocg
->weight
, hw_inuse
, hw_active
);
1758 spin_lock_irq(&ioc
->lock
);
1759 propagate_weights(iocg
, iocg
->weight
, iocg
->weight
);
1760 spin_unlock_irq(&ioc
->lock
);
1761 current_hweight(iocg
, &hw_active
, &hw_inuse
);
1764 cost
= abs_cost_to_cost(abs_cost
, hw_inuse
);
1767 * If no one's waiting and within budget, issue right away. The
1768 * tests are racy but the races aren't systemic - we only miss once
1769 * in a while which is fine.
1771 if (!waitqueue_active(&iocg
->waitq
) && !iocg
->abs_vdebt
&&
1772 time_before_eq64(vtime
+ cost
, now
.vnow
)) {
1773 iocg_commit_bio(iocg
, bio
, cost
);
1778 * We activated above but w/o any synchronization. Deactivation is
1779 * synchronized with waitq.lock and we won't get deactivated as long
1780 * as we're waiting or has debt, so we're good if we're activated
1781 * here. In the unlikely case that we aren't, just issue the IO.
1783 spin_lock_irq(&iocg
->waitq
.lock
);
1785 if (unlikely(list_empty(&iocg
->active_list
))) {
1786 spin_unlock_irq(&iocg
->waitq
.lock
);
1787 iocg_commit_bio(iocg
, bio
, cost
);
1792 * We're over budget. If @bio has to be issued regardless, remember
1793 * the abs_cost instead of advancing vtime. iocg_kick_waitq() will pay
1794 * off the debt before waking more IOs.
1796 * This way, the debt is continuously paid off each period with the
1797 * actual budget available to the cgroup. If we just wound vtime, we
1798 * would incorrectly use the current hw_inuse for the entire amount
1799 * which, for example, can lead to the cgroup staying blocked for a
1800 * long time even with substantially raised hw_inuse.
1802 * An iocg with vdebt should stay online so that the timer can keep
1803 * deducting its vdebt and [de]activate use_delay mechanism
1804 * accordingly. We don't want to race against the timer trying to
1805 * clear them and leave @iocg inactive w/ dangling use_delay heavily
1806 * penalizing the cgroup and its descendants.
1808 if (bio_issue_as_root_blkg(bio
) || fatal_signal_pending(current
)) {
1809 iocg
->abs_vdebt
+= abs_cost
;
1810 if (iocg_kick_delay(iocg
, &now
))
1811 blkcg_schedule_throttle(rqos
->q
,
1812 (bio
->bi_opf
& REQ_SWAP
) == REQ_SWAP
);
1813 spin_unlock_irq(&iocg
->waitq
.lock
);
1818 * Append self to the waitq and schedule the wakeup timer if we're
1819 * the first waiter. The timer duration is calculated based on the
1820 * current vrate. vtime and hweight changes can make it too short
1821 * or too long. Each wait entry records the absolute cost it's
1822 * waiting for to allow re-evaluation using a custom wait entry.
1824 * If too short, the timer simply reschedules itself. If too long,
1825 * the period timer will notice and trigger wakeups.
1827 * All waiters are on iocg->waitq and the wait states are
1828 * synchronized using waitq.lock.
1830 init_waitqueue_func_entry(&wait
.wait
, iocg_wake_fn
);
1831 wait
.wait
.private = current
;
1833 wait
.abs_cost
= abs_cost
;
1834 wait
.committed
= false; /* will be set true by waker */
1836 __add_wait_queue_entry_tail(&iocg
->waitq
, &wait
.wait
);
1837 iocg_kick_waitq(iocg
, &now
);
1839 spin_unlock_irq(&iocg
->waitq
.lock
);
1842 set_current_state(TASK_UNINTERRUPTIBLE
);
1848 /* waker already committed us, proceed */
1849 finish_wait(&iocg
->waitq
, &wait
.wait
);
1852 static void ioc_rqos_merge(struct rq_qos
*rqos
, struct request
*rq
,
1855 struct ioc_gq
*iocg
= blkg_to_iocg(bio
->bi_blkg
);
1856 struct ioc
*ioc
= iocg
->ioc
;
1857 sector_t bio_end
= bio_end_sector(bio
);
1861 unsigned long flags
;
1863 /* bypass if disabled or for root cgroup */
1864 if (!ioc
->enabled
|| !iocg
->level
)
1867 abs_cost
= calc_vtime_cost(bio
, iocg
, true);
1872 current_hweight(iocg
, NULL
, &hw_inuse
);
1873 cost
= abs_cost_to_cost(abs_cost
, hw_inuse
);
1875 /* update cursor if backmerging into the request at the cursor */
1876 if (blk_rq_pos(rq
) < bio_end
&&
1877 blk_rq_pos(rq
) + blk_rq_sectors(rq
) == iocg
->cursor
)
1878 iocg
->cursor
= bio_end
;
1881 * Charge if there's enough vtime budget and the existing request has
1884 if (rq
->bio
&& rq
->bio
->bi_iocost_cost
&&
1885 time_before_eq64(atomic64_read(&iocg
->vtime
) + cost
, now
.vnow
)) {
1886 iocg_commit_bio(iocg
, bio
, cost
);
1891 * Otherwise, account it as debt if @iocg is online, which it should
1892 * be for the vast majority of cases. See debt handling in
1893 * ioc_rqos_throttle() for details.
1895 spin_lock_irqsave(&iocg
->waitq
.lock
, flags
);
1896 if (likely(!list_empty(&iocg
->active_list
))) {
1897 iocg
->abs_vdebt
+= abs_cost
;
1898 iocg_kick_delay(iocg
, &now
);
1900 iocg_commit_bio(iocg
, bio
, cost
);
1902 spin_unlock_irqrestore(&iocg
->waitq
.lock
, flags
);
1905 static void ioc_rqos_done_bio(struct rq_qos
*rqos
, struct bio
*bio
)
1907 struct ioc_gq
*iocg
= blkg_to_iocg(bio
->bi_blkg
);
1909 if (iocg
&& bio
->bi_iocost_cost
)
1910 atomic64_add(bio
->bi_iocost_cost
, &iocg
->done_vtime
);
1913 static void ioc_rqos_done(struct rq_qos
*rqos
, struct request
*rq
)
1915 struct ioc
*ioc
= rqos_to_ioc(rqos
);
1916 struct ioc_pcpu_stat
*ccs
;
1917 u64 on_q_ns
, rq_wait_ns
, size_nsec
;
1920 if (!ioc
->enabled
|| !rq
->alloc_time_ns
|| !rq
->start_time_ns
)
1923 switch (req_op(rq
) & REQ_OP_MASK
) {
1936 on_q_ns
= ktime_get_ns() - rq
->alloc_time_ns
;
1937 rq_wait_ns
= rq
->start_time_ns
- rq
->alloc_time_ns
;
1938 size_nsec
= div64_u64(calc_size_vtime_cost(rq
, ioc
), VTIME_PER_NSEC
);
1940 ccs
= get_cpu_ptr(ioc
->pcpu_stat
);
1942 if (on_q_ns
<= size_nsec
||
1943 on_q_ns
- size_nsec
<= ioc
->params
.qos
[pidx
] * NSEC_PER_USEC
)
1944 local_inc(&ccs
->missed
[rw
].nr_met
);
1946 local_inc(&ccs
->missed
[rw
].nr_missed
);
1948 local64_add(rq_wait_ns
, &ccs
->rq_wait_ns
);
1953 static void ioc_rqos_queue_depth_changed(struct rq_qos
*rqos
)
1955 struct ioc
*ioc
= rqos_to_ioc(rqos
);
1957 spin_lock_irq(&ioc
->lock
);
1958 ioc_refresh_params(ioc
, false);
1959 spin_unlock_irq(&ioc
->lock
);
1962 static void ioc_rqos_exit(struct rq_qos
*rqos
)
1964 struct ioc
*ioc
= rqos_to_ioc(rqos
);
1966 blkcg_deactivate_policy(rqos
->q
, &blkcg_policy_iocost
);
1968 spin_lock_irq(&ioc
->lock
);
1969 ioc
->running
= IOC_STOP
;
1970 spin_unlock_irq(&ioc
->lock
);
1972 del_timer_sync(&ioc
->timer
);
1973 free_percpu(ioc
->pcpu_stat
);
1977 static struct rq_qos_ops ioc_rqos_ops
= {
1978 .throttle
= ioc_rqos_throttle
,
1979 .merge
= ioc_rqos_merge
,
1980 .done_bio
= ioc_rqos_done_bio
,
1981 .done
= ioc_rqos_done
,
1982 .queue_depth_changed
= ioc_rqos_queue_depth_changed
,
1983 .exit
= ioc_rqos_exit
,
1986 static int blk_iocost_init(struct request_queue
*q
)
1989 struct rq_qos
*rqos
;
1992 ioc
= kzalloc(sizeof(*ioc
), GFP_KERNEL
);
1996 ioc
->pcpu_stat
= alloc_percpu(struct ioc_pcpu_stat
);
1997 if (!ioc
->pcpu_stat
) {
2002 for_each_possible_cpu(cpu
) {
2003 struct ioc_pcpu_stat
*ccs
= per_cpu_ptr(ioc
->pcpu_stat
, cpu
);
2005 for (i
= 0; i
< ARRAY_SIZE(ccs
->missed
); i
++) {
2006 local_set(&ccs
->missed
[i
].nr_met
, 0);
2007 local_set(&ccs
->missed
[i
].nr_missed
, 0);
2009 local64_set(&ccs
->rq_wait_ns
, 0);
2013 rqos
->id
= RQ_QOS_COST
;
2014 rqos
->ops
= &ioc_rqos_ops
;
2017 spin_lock_init(&ioc
->lock
);
2018 timer_setup(&ioc
->timer
, ioc_timer_fn
, 0);
2019 INIT_LIST_HEAD(&ioc
->active_iocgs
);
2021 ioc
->running
= IOC_IDLE
;
2022 atomic64_set(&ioc
->vtime_rate
, VTIME_PER_USEC
);
2023 seqcount_spinlock_init(&ioc
->period_seqcount
, &ioc
->lock
);
2024 ioc
->period_at
= ktime_to_us(ktime_get());
2025 atomic64_set(&ioc
->cur_period
, 0);
2026 atomic_set(&ioc
->hweight_gen
, 0);
2028 spin_lock_irq(&ioc
->lock
);
2029 ioc
->autop_idx
= AUTOP_INVALID
;
2030 ioc_refresh_params(ioc
, true);
2031 spin_unlock_irq(&ioc
->lock
);
2033 rq_qos_add(q
, rqos
);
2034 ret
= blkcg_activate_policy(q
, &blkcg_policy_iocost
);
2036 rq_qos_del(q
, rqos
);
2037 free_percpu(ioc
->pcpu_stat
);
2044 static struct blkcg_policy_data
*ioc_cpd_alloc(gfp_t gfp
)
2046 struct ioc_cgrp
*iocc
;
2048 iocc
= kzalloc(sizeof(struct ioc_cgrp
), gfp
);
2052 iocc
->dfl_weight
= CGROUP_WEIGHT_DFL
;
2056 static void ioc_cpd_free(struct blkcg_policy_data
*cpd
)
2058 kfree(container_of(cpd
, struct ioc_cgrp
, cpd
));
2061 static struct blkg_policy_data
*ioc_pd_alloc(gfp_t gfp
, struct request_queue
*q
,
2062 struct blkcg
*blkcg
)
2064 int levels
= blkcg
->css
.cgroup
->level
+ 1;
2065 struct ioc_gq
*iocg
;
2067 iocg
= kzalloc_node(struct_size(iocg
, ancestors
, levels
), gfp
, q
->node
);
2074 static void ioc_pd_init(struct blkg_policy_data
*pd
)
2076 struct ioc_gq
*iocg
= pd_to_iocg(pd
);
2077 struct blkcg_gq
*blkg
= pd_to_blkg(&iocg
->pd
);
2078 struct ioc
*ioc
= q_to_ioc(blkg
->q
);
2080 struct blkcg_gq
*tblkg
;
2081 unsigned long flags
;
2086 atomic64_set(&iocg
->vtime
, now
.vnow
);
2087 atomic64_set(&iocg
->done_vtime
, now
.vnow
);
2088 atomic64_set(&iocg
->active_period
, atomic64_read(&ioc
->cur_period
));
2089 INIT_LIST_HEAD(&iocg
->active_list
);
2090 iocg
->hweight_active
= HWEIGHT_WHOLE
;
2091 iocg
->hweight_inuse
= HWEIGHT_WHOLE
;
2093 init_waitqueue_head(&iocg
->waitq
);
2094 hrtimer_init(&iocg
->waitq_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_ABS
);
2095 iocg
->waitq_timer
.function
= iocg_waitq_timer_fn
;
2096 hrtimer_init(&iocg
->delay_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_ABS
);
2097 iocg
->delay_timer
.function
= iocg_delay_timer_fn
;
2099 iocg
->level
= blkg
->blkcg
->css
.cgroup
->level
;
2101 for (tblkg
= blkg
; tblkg
; tblkg
= tblkg
->parent
) {
2102 struct ioc_gq
*tiocg
= blkg_to_iocg(tblkg
);
2103 iocg
->ancestors
[tiocg
->level
] = tiocg
;
2106 spin_lock_irqsave(&ioc
->lock
, flags
);
2107 weight_updated(iocg
);
2108 spin_unlock_irqrestore(&ioc
->lock
, flags
);
2111 static void ioc_pd_free(struct blkg_policy_data
*pd
)
2113 struct ioc_gq
*iocg
= pd_to_iocg(pd
);
2114 struct ioc
*ioc
= iocg
->ioc
;
2115 unsigned long flags
;
2118 spin_lock_irqsave(&ioc
->lock
, flags
);
2119 if (!list_empty(&iocg
->active_list
)) {
2120 propagate_weights(iocg
, 0, 0);
2121 list_del_init(&iocg
->active_list
);
2123 spin_unlock_irqrestore(&ioc
->lock
, flags
);
2125 hrtimer_cancel(&iocg
->waitq_timer
);
2126 hrtimer_cancel(&iocg
->delay_timer
);
2131 static u64
ioc_weight_prfill(struct seq_file
*sf
, struct blkg_policy_data
*pd
,
2134 const char *dname
= blkg_dev_name(pd
->blkg
);
2135 struct ioc_gq
*iocg
= pd_to_iocg(pd
);
2137 if (dname
&& iocg
->cfg_weight
)
2138 seq_printf(sf
, "%s %u\n", dname
, iocg
->cfg_weight
);
2143 static int ioc_weight_show(struct seq_file
*sf
, void *v
)
2145 struct blkcg
*blkcg
= css_to_blkcg(seq_css(sf
));
2146 struct ioc_cgrp
*iocc
= blkcg_to_iocc(blkcg
);
2148 seq_printf(sf
, "default %u\n", iocc
->dfl_weight
);
2149 blkcg_print_blkgs(sf
, blkcg
, ioc_weight_prfill
,
2150 &blkcg_policy_iocost
, seq_cft(sf
)->private, false);
2154 static ssize_t
ioc_weight_write(struct kernfs_open_file
*of
, char *buf
,
2155 size_t nbytes
, loff_t off
)
2157 struct blkcg
*blkcg
= css_to_blkcg(of_css(of
));
2158 struct ioc_cgrp
*iocc
= blkcg_to_iocc(blkcg
);
2159 struct blkg_conf_ctx ctx
;
2160 struct ioc_gq
*iocg
;
2164 if (!strchr(buf
, ':')) {
2165 struct blkcg_gq
*blkg
;
2167 if (!sscanf(buf
, "default %u", &v
) && !sscanf(buf
, "%u", &v
))
2170 if (v
< CGROUP_WEIGHT_MIN
|| v
> CGROUP_WEIGHT_MAX
)
2173 spin_lock(&blkcg
->lock
);
2174 iocc
->dfl_weight
= v
;
2175 hlist_for_each_entry(blkg
, &blkcg
->blkg_list
, blkcg_node
) {
2176 struct ioc_gq
*iocg
= blkg_to_iocg(blkg
);
2179 spin_lock_irq(&iocg
->ioc
->lock
);
2180 weight_updated(iocg
);
2181 spin_unlock_irq(&iocg
->ioc
->lock
);
2184 spin_unlock(&blkcg
->lock
);
2189 ret
= blkg_conf_prep(blkcg
, &blkcg_policy_iocost
, buf
, &ctx
);
2193 iocg
= blkg_to_iocg(ctx
.blkg
);
2195 if (!strncmp(ctx
.body
, "default", 7)) {
2198 if (!sscanf(ctx
.body
, "%u", &v
))
2200 if (v
< CGROUP_WEIGHT_MIN
|| v
> CGROUP_WEIGHT_MAX
)
2204 spin_lock(&iocg
->ioc
->lock
);
2205 iocg
->cfg_weight
= v
;
2206 weight_updated(iocg
);
2207 spin_unlock(&iocg
->ioc
->lock
);
2209 blkg_conf_finish(&ctx
);
2213 blkg_conf_finish(&ctx
);
2217 static u64
ioc_qos_prfill(struct seq_file
*sf
, struct blkg_policy_data
*pd
,
2220 const char *dname
= blkg_dev_name(pd
->blkg
);
2221 struct ioc
*ioc
= pd_to_iocg(pd
)->ioc
;
2226 seq_printf(sf
, "%s enable=%d ctrl=%s rpct=%u.%02u rlat=%u wpct=%u.%02u wlat=%u min=%u.%02u max=%u.%02u\n",
2227 dname
, ioc
->enabled
, ioc
->user_qos_params
? "user" : "auto",
2228 ioc
->params
.qos
[QOS_RPPM
] / 10000,
2229 ioc
->params
.qos
[QOS_RPPM
] % 10000 / 100,
2230 ioc
->params
.qos
[QOS_RLAT
],
2231 ioc
->params
.qos
[QOS_WPPM
] / 10000,
2232 ioc
->params
.qos
[QOS_WPPM
] % 10000 / 100,
2233 ioc
->params
.qos
[QOS_WLAT
],
2234 ioc
->params
.qos
[QOS_MIN
] / 10000,
2235 ioc
->params
.qos
[QOS_MIN
] % 10000 / 100,
2236 ioc
->params
.qos
[QOS_MAX
] / 10000,
2237 ioc
->params
.qos
[QOS_MAX
] % 10000 / 100);
2241 static int ioc_qos_show(struct seq_file
*sf
, void *v
)
2243 struct blkcg
*blkcg
= css_to_blkcg(seq_css(sf
));
2245 blkcg_print_blkgs(sf
, blkcg
, ioc_qos_prfill
,
2246 &blkcg_policy_iocost
, seq_cft(sf
)->private, false);
2250 static const match_table_t qos_ctrl_tokens
= {
2251 { QOS_ENABLE
, "enable=%u" },
2252 { QOS_CTRL
, "ctrl=%s" },
2253 { NR_QOS_CTRL_PARAMS
, NULL
},
2256 static const match_table_t qos_tokens
= {
2257 { QOS_RPPM
, "rpct=%s" },
2258 { QOS_RLAT
, "rlat=%u" },
2259 { QOS_WPPM
, "wpct=%s" },
2260 { QOS_WLAT
, "wlat=%u" },
2261 { QOS_MIN
, "min=%s" },
2262 { QOS_MAX
, "max=%s" },
2263 { NR_QOS_PARAMS
, NULL
},
2266 static ssize_t
ioc_qos_write(struct kernfs_open_file
*of
, char *input
,
2267 size_t nbytes
, loff_t off
)
2269 struct gendisk
*disk
;
2271 u32 qos
[NR_QOS_PARAMS
];
2276 disk
= blkcg_conf_get_disk(&input
);
2278 return PTR_ERR(disk
);
2280 ioc
= q_to_ioc(disk
->queue
);
2282 ret
= blk_iocost_init(disk
->queue
);
2285 ioc
= q_to_ioc(disk
->queue
);
2288 spin_lock_irq(&ioc
->lock
);
2289 memcpy(qos
, ioc
->params
.qos
, sizeof(qos
));
2290 enable
= ioc
->enabled
;
2291 user
= ioc
->user_qos_params
;
2292 spin_unlock_irq(&ioc
->lock
);
2294 while ((p
= strsep(&input
, " \t\n"))) {
2295 substring_t args
[MAX_OPT_ARGS
];
2303 switch (match_token(p
, qos_ctrl_tokens
, args
)) {
2305 match_u64(&args
[0], &v
);
2309 match_strlcpy(buf
, &args
[0], sizeof(buf
));
2310 if (!strcmp(buf
, "auto"))
2312 else if (!strcmp(buf
, "user"))
2319 tok
= match_token(p
, qos_tokens
, args
);
2323 if (match_strlcpy(buf
, &args
[0], sizeof(buf
)) >=
2326 if (cgroup_parse_float(buf
, 2, &v
))
2328 if (v
< 0 || v
> 10000)
2334 if (match_u64(&args
[0], &v
))
2340 if (match_strlcpy(buf
, &args
[0], sizeof(buf
)) >=
2343 if (cgroup_parse_float(buf
, 2, &v
))
2347 qos
[tok
] = clamp_t(s64
, v
* 100,
2348 VRATE_MIN_PPM
, VRATE_MAX_PPM
);
2356 if (qos
[QOS_MIN
] > qos
[QOS_MAX
])
2359 spin_lock_irq(&ioc
->lock
);
2362 blk_stat_enable_accounting(ioc
->rqos
.q
);
2363 blk_queue_flag_set(QUEUE_FLAG_RQ_ALLOC_TIME
, ioc
->rqos
.q
);
2364 ioc
->enabled
= true;
2366 blk_queue_flag_clear(QUEUE_FLAG_RQ_ALLOC_TIME
, ioc
->rqos
.q
);
2367 ioc
->enabled
= false;
2371 memcpy(ioc
->params
.qos
, qos
, sizeof(qos
));
2372 ioc
->user_qos_params
= true;
2374 ioc
->user_qos_params
= false;
2377 ioc_refresh_params(ioc
, true);
2378 spin_unlock_irq(&ioc
->lock
);
2380 put_disk_and_module(disk
);
2385 put_disk_and_module(disk
);
2389 static u64
ioc_cost_model_prfill(struct seq_file
*sf
,
2390 struct blkg_policy_data
*pd
, int off
)
2392 const char *dname
= blkg_dev_name(pd
->blkg
);
2393 struct ioc
*ioc
= pd_to_iocg(pd
)->ioc
;
2394 u64
*u
= ioc
->params
.i_lcoefs
;
2399 seq_printf(sf
, "%s ctrl=%s model=linear "
2400 "rbps=%llu rseqiops=%llu rrandiops=%llu "
2401 "wbps=%llu wseqiops=%llu wrandiops=%llu\n",
2402 dname
, ioc
->user_cost_model
? "user" : "auto",
2403 u
[I_LCOEF_RBPS
], u
[I_LCOEF_RSEQIOPS
], u
[I_LCOEF_RRANDIOPS
],
2404 u
[I_LCOEF_WBPS
], u
[I_LCOEF_WSEQIOPS
], u
[I_LCOEF_WRANDIOPS
]);
2408 static int ioc_cost_model_show(struct seq_file
*sf
, void *v
)
2410 struct blkcg
*blkcg
= css_to_blkcg(seq_css(sf
));
2412 blkcg_print_blkgs(sf
, blkcg
, ioc_cost_model_prfill
,
2413 &blkcg_policy_iocost
, seq_cft(sf
)->private, false);
2417 static const match_table_t cost_ctrl_tokens
= {
2418 { COST_CTRL
, "ctrl=%s" },
2419 { COST_MODEL
, "model=%s" },
2420 { NR_COST_CTRL_PARAMS
, NULL
},
2423 static const match_table_t i_lcoef_tokens
= {
2424 { I_LCOEF_RBPS
, "rbps=%u" },
2425 { I_LCOEF_RSEQIOPS
, "rseqiops=%u" },
2426 { I_LCOEF_RRANDIOPS
, "rrandiops=%u" },
2427 { I_LCOEF_WBPS
, "wbps=%u" },
2428 { I_LCOEF_WSEQIOPS
, "wseqiops=%u" },
2429 { I_LCOEF_WRANDIOPS
, "wrandiops=%u" },
2430 { NR_I_LCOEFS
, NULL
},
2433 static ssize_t
ioc_cost_model_write(struct kernfs_open_file
*of
, char *input
,
2434 size_t nbytes
, loff_t off
)
2436 struct gendisk
*disk
;
2443 disk
= blkcg_conf_get_disk(&input
);
2445 return PTR_ERR(disk
);
2447 ioc
= q_to_ioc(disk
->queue
);
2449 ret
= blk_iocost_init(disk
->queue
);
2452 ioc
= q_to_ioc(disk
->queue
);
2455 spin_lock_irq(&ioc
->lock
);
2456 memcpy(u
, ioc
->params
.i_lcoefs
, sizeof(u
));
2457 user
= ioc
->user_cost_model
;
2458 spin_unlock_irq(&ioc
->lock
);
2460 while ((p
= strsep(&input
, " \t\n"))) {
2461 substring_t args
[MAX_OPT_ARGS
];
2469 switch (match_token(p
, cost_ctrl_tokens
, args
)) {
2471 match_strlcpy(buf
, &args
[0], sizeof(buf
));
2472 if (!strcmp(buf
, "auto"))
2474 else if (!strcmp(buf
, "user"))
2480 match_strlcpy(buf
, &args
[0], sizeof(buf
));
2481 if (strcmp(buf
, "linear"))
2486 tok
= match_token(p
, i_lcoef_tokens
, args
);
2487 if (tok
== NR_I_LCOEFS
)
2489 if (match_u64(&args
[0], &v
))
2495 spin_lock_irq(&ioc
->lock
);
2497 memcpy(ioc
->params
.i_lcoefs
, u
, sizeof(u
));
2498 ioc
->user_cost_model
= true;
2500 ioc
->user_cost_model
= false;
2502 ioc_refresh_params(ioc
, true);
2503 spin_unlock_irq(&ioc
->lock
);
2505 put_disk_and_module(disk
);
2511 put_disk_and_module(disk
);
2515 static struct cftype ioc_files
[] = {
2518 .flags
= CFTYPE_NOT_ON_ROOT
,
2519 .seq_show
= ioc_weight_show
,
2520 .write
= ioc_weight_write
,
2524 .flags
= CFTYPE_ONLY_ON_ROOT
,
2525 .seq_show
= ioc_qos_show
,
2526 .write
= ioc_qos_write
,
2529 .name
= "cost.model",
2530 .flags
= CFTYPE_ONLY_ON_ROOT
,
2531 .seq_show
= ioc_cost_model_show
,
2532 .write
= ioc_cost_model_write
,
2537 static struct blkcg_policy blkcg_policy_iocost
= {
2538 .dfl_cftypes
= ioc_files
,
2539 .cpd_alloc_fn
= ioc_cpd_alloc
,
2540 .cpd_free_fn
= ioc_cpd_free
,
2541 .pd_alloc_fn
= ioc_pd_alloc
,
2542 .pd_init_fn
= ioc_pd_init
,
2543 .pd_free_fn
= ioc_pd_free
,
2546 static int __init
ioc_init(void)
2548 return blkcg_policy_register(&blkcg_policy_iocost
);
2551 static void __exit
ioc_exit(void)
2553 return blkcg_policy_unregister(&blkcg_policy_iocost
);
2556 module_init(ioc_init
);
2557 module_exit(ioc_exit
);