xref: /openbmc/linux/block/blk-iocost.c (revision eb431351)
1 /* SPDX-License-Identifier: GPL-2.0
2  *
3  * IO cost model based controller.
4  *
5  * Copyright (C) 2019 Tejun Heo <tj@kernel.org>
6  * Copyright (C) 2019 Andy Newell <newella@fb.com>
7  * Copyright (C) 2019 Facebook
8  *
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
12  * approximations.
13  *
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.
21  *
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
28  * distribution.
29  *
30  * 1. IO Cost Model
31  *
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.
36  *
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.
45  *
46  * If needed, tools/cgroup/iocost_coef_gen.py can be used to generate
47  * device-specific coefficients.
48  *
49  * If needed, tools/cgroup/iocost_coef_gen.py can be used to generate
50  * device-specific coefficients.
51  *
52  * 2. Control Strategy
53  *
54  * The device virtual time (vtime) is used as the primary control metric.
55  * The control strategy is composed of the following three parts.
56  *
57  * 2-1. Vtime Distribution
58  *
59  * When a cgroup becomes active in terms of IOs, its hierarchical share is
60  * calculated.  Please consider the following hierarchy where the numbers
61  * inside parentheses denote the configured weights.
62  *
63  *           root
64  *         /       \
65  *      A (w:100)  B (w:300)
66  *      /       \
67  *  A0 (w:100)  A1 (w:100)
68  *
69  * If B is idle and only A0 and A1 are actively issuing IOs, as the two are
70  * of equal weight, each gets 50% share.  If then B starts issuing IOs, B
71  * gets 300/(100+300) or 75% share, and A0 and A1 equally splits the rest,
72  * 12.5% each.  The distribution mechanism only cares about these flattened
73  * shares.  They're called hweights (hierarchical weights) and always add
74  * upto 1 (HWEIGHT_WHOLE).
75  *
76  * A given cgroup's vtime runs slower in inverse proportion to its hweight.
77  * For example, with 12.5% weight, A0's time runs 8 times slower (100/12.5)
78  * against the device vtime - an IO which takes 10ms on the underlying
79  * device is considered to take 80ms on A0.
80  *
81  * This constitutes the basis of IO capacity distribution.  Each cgroup's
82  * vtime is running at a rate determined by its hweight.  A cgroup tracks
83  * the vtime consumed by past IOs and can issue a new IO iff doing so
84  * wouldn't outrun the current device vtime.  Otherwise, the IO is
85  * suspended until the vtime has progressed enough to cover it.
86  *
87  * 2-2. Vrate Adjustment
88  *
89  * It's unrealistic to expect the cost model to be perfect.  There are too
90  * many devices and even on the same device the overall performance
91  * fluctuates depending on numerous factors such as IO mixture and device
92  * internal garbage collection.  The controller needs to adapt dynamically.
93  *
94  * This is achieved by adjusting the overall IO rate according to how busy
95  * the device is.  If the device becomes overloaded, we're sending down too
96  * many IOs and should generally slow down.  If there are waiting issuers
97  * but the device isn't saturated, we're issuing too few and should
98  * generally speed up.
99  *
100  * To slow down, we lower the vrate - the rate at which the device vtime
101  * passes compared to the wall clock.  For example, if the vtime is running
102  * at the vrate of 75%, all cgroups added up would only be able to issue
103  * 750ms worth of IOs per second, and vice-versa for speeding up.
104  *
105  * Device business is determined using two criteria - rq wait and
106  * completion latencies.
107  *
108  * When a device gets saturated, the on-device and then the request queues
109  * fill up and a bio which is ready to be issued has to wait for a request
110  * to become available.  When this delay becomes noticeable, it's a clear
111  * indication that the device is saturated and we lower the vrate.  This
112  * saturation signal is fairly conservative as it only triggers when both
113  * hardware and software queues are filled up, and is used as the default
114  * busy signal.
115  *
116  * As devices can have deep queues and be unfair in how the queued commands
117  * are executed, soley depending on rq wait may not result in satisfactory
118  * control quality.  For a better control quality, completion latency QoS
119  * parameters can be configured so that the device is considered saturated
120  * if N'th percentile completion latency rises above the set point.
121  *
122  * The completion latency requirements are a function of both the
123  * underlying device characteristics and the desired IO latency quality of
124  * service.  There is an inherent trade-off - the tighter the latency QoS,
125  * the higher the bandwidth lossage.  Latency QoS is disabled by default
126  * and can be set through /sys/fs/cgroup/io.cost.qos.
127  *
128  * 2-3. Work Conservation
129  *
130  * Imagine two cgroups A and B with equal weights.  A is issuing a small IO
131  * periodically while B is sending out enough parallel IOs to saturate the
132  * device on its own.  Let's say A's usage amounts to 100ms worth of IO
133  * cost per second, i.e., 10% of the device capacity.  The naive
134  * distribution of half and half would lead to 60% utilization of the
135  * device, a significant reduction in the total amount of work done
136  * compared to free-for-all competition.  This is too high a cost to pay
137  * for IO control.
138  *
139  * To conserve the total amount of work done, we keep track of how much
140  * each active cgroup is actually using and yield part of its weight if
141  * there are other cgroups which can make use of it.  In the above case,
142  * A's weight will be lowered so that it hovers above the actual usage and
143  * B would be able to use the rest.
144  *
145  * As we don't want to penalize a cgroup for donating its weight, the
146  * surplus weight adjustment factors in a margin and has an immediate
147  * snapback mechanism in case the cgroup needs more IO vtime for itself.
148  *
149  * Note that adjusting down surplus weights has the same effects as
150  * accelerating vtime for other cgroups and work conservation can also be
151  * implemented by adjusting vrate dynamically.  However, squaring who can
152  * donate and should take back how much requires hweight propagations
153  * anyway making it easier to implement and understand as a separate
154  * mechanism.
155  *
156  * 3. Monitoring
157  *
158  * Instead of debugfs or other clumsy monitoring mechanisms, this
159  * controller uses a drgn based monitoring script -
160  * tools/cgroup/iocost_monitor.py.  For details on drgn, please see
161  * https://github.com/osandov/drgn.  The ouput looks like the following.
162  *
163  *  sdb RUN   per=300ms cur_per=234.218:v203.695 busy= +1 vrate= 62.12%
164  *                 active      weight      hweight% inflt% dbt  delay usages%
165  *  test/a              *    50/   50  33.33/ 33.33  27.65   2  0*041 033:033:033
166  *  test/b              *   100/  100  66.67/ 66.67  17.56   0  0*000 066:079:077
167  *
168  * - per	: Timer period
169  * - cur_per	: Internal wall and device vtime clock
170  * - vrate	: Device virtual time rate against wall clock
171  * - weight	: Surplus-adjusted and configured weights
172  * - hweight	: Surplus-adjusted and configured hierarchical weights
173  * - inflt	: The percentage of in-flight IO cost at the end of last period
174  * - del_ms	: Deferred issuer delay induction level and duration
175  * - usages	: Usage history
176  */
177 
178 #include <linux/kernel.h>
179 #include <linux/module.h>
180 #include <linux/timer.h>
181 #include <linux/time64.h>
182 #include <linux/parser.h>
183 #include <linux/sched/signal.h>
184 #include <linux/blk-cgroup.h>
185 #include "blk-rq-qos.h"
186 #include "blk-stat.h"
187 #include "blk-wbt.h"
188 
189 #ifdef CONFIG_TRACEPOINTS
190 
191 /* copied from TRACE_CGROUP_PATH, see cgroup-internal.h */
192 #define TRACE_IOCG_PATH_LEN 1024
193 static DEFINE_SPINLOCK(trace_iocg_path_lock);
194 static char trace_iocg_path[TRACE_IOCG_PATH_LEN];
195 
196 #define TRACE_IOCG_PATH(type, iocg, ...)					\
197 	do {									\
198 		unsigned long flags;						\
199 		if (trace_iocost_##type##_enabled()) {				\
200 			spin_lock_irqsave(&trace_iocg_path_lock, flags);	\
201 			cgroup_path(iocg_to_blkg(iocg)->blkcg->css.cgroup,	\
202 				    trace_iocg_path, TRACE_IOCG_PATH_LEN);	\
203 			trace_iocost_##type(iocg, trace_iocg_path,		\
204 					      ##__VA_ARGS__);			\
205 			spin_unlock_irqrestore(&trace_iocg_path_lock, flags);	\
206 		}								\
207 	} while (0)
208 
209 #else	/* CONFIG_TRACE_POINTS */
210 #define TRACE_IOCG_PATH(type, iocg, ...)	do { } while (0)
211 #endif	/* CONFIG_TRACE_POINTS */
212 
213 enum {
214 	MILLION			= 1000000,
215 
216 	/* timer period is calculated from latency requirements, bound it */
217 	MIN_PERIOD		= USEC_PER_MSEC,
218 	MAX_PERIOD		= USEC_PER_SEC,
219 
220 	/*
221 	 * A cgroup's vtime can run 50% behind the device vtime, which
222 	 * serves as its IO credit buffer.  Surplus weight adjustment is
223 	 * immediately canceled if the vtime margin runs below 10%.
224 	 */
225 	MARGIN_PCT		= 50,
226 	INUSE_MARGIN_PCT	= 10,
227 
228 	/* Have some play in waitq timer operations */
229 	WAITQ_TIMER_MARGIN_PCT	= 5,
230 
231 	/*
232 	 * vtime can wrap well within a reasonable uptime when vrate is
233 	 * consistently raised.  Don't trust recorded cgroup vtime if the
234 	 * period counter indicates that it's older than 5mins.
235 	 */
236 	VTIME_VALID_DUR		= 300 * USEC_PER_SEC,
237 
238 	/*
239 	 * Remember the past three non-zero usages and use the max for
240 	 * surplus calculation.  Three slots guarantee that we remember one
241 	 * full period usage from the last active stretch even after
242 	 * partial deactivation and re-activation periods.  Don't start
243 	 * giving away weight before collecting two data points to prevent
244 	 * hweight adjustments based on one partial activation period.
245 	 */
246 	NR_USAGE_SLOTS		= 3,
247 	MIN_VALID_USAGES	= 2,
248 
249 	/* 1/64k is granular enough and can easily be handled w/ u32 */
250 	HWEIGHT_WHOLE		= 1 << 16,
251 
252 	/*
253 	 * As vtime is used to calculate the cost of each IO, it needs to
254 	 * be fairly high precision.  For example, it should be able to
255 	 * represent the cost of a single page worth of discard with
256 	 * suffificient accuracy.  At the same time, it should be able to
257 	 * represent reasonably long enough durations to be useful and
258 	 * convenient during operation.
259 	 *
260 	 * 1s worth of vtime is 2^37.  This gives us both sub-nanosecond
261 	 * granularity and days of wrap-around time even at extreme vrates.
262 	 */
263 	VTIME_PER_SEC_SHIFT	= 37,
264 	VTIME_PER_SEC		= 1LLU << VTIME_PER_SEC_SHIFT,
265 	VTIME_PER_USEC		= VTIME_PER_SEC / USEC_PER_SEC,
266 
267 	/* bound vrate adjustments within two orders of magnitude */
268 	VRATE_MIN_PPM		= 10000,	/* 1% */
269 	VRATE_MAX_PPM		= 100000000,	/* 10000% */
270 
271 	VRATE_MIN		= VTIME_PER_USEC * VRATE_MIN_PPM / MILLION,
272 	VRATE_CLAMP_ADJ_PCT	= 4,
273 
274 	/* if IOs end up waiting for requests, issue less */
275 	RQ_WAIT_BUSY_PCT	= 5,
276 
277 	/* unbusy hysterisis */
278 	UNBUSY_THR_PCT		= 75,
279 
280 	/* don't let cmds which take a very long time pin lagging for too long */
281 	MAX_LAGGING_PERIODS	= 10,
282 
283 	/*
284 	 * If usage% * 1.25 + 2% is lower than hweight% by more than 3%,
285 	 * donate the surplus.
286 	 */
287 	SURPLUS_SCALE_PCT	= 125,			/* * 125% */
288 	SURPLUS_SCALE_ABS	= HWEIGHT_WHOLE / 50,	/* + 2% */
289 	SURPLUS_MIN_ADJ_DELTA	= HWEIGHT_WHOLE / 33,	/* 3% */
290 
291 	/* switch iff the conditions are met for longer than this */
292 	AUTOP_CYCLE_NSEC	= 10LLU * NSEC_PER_SEC,
293 
294 	/*
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.
298 	 */
299 	IOC_PAGE_SHIFT		= 12,
300 	IOC_PAGE_SIZE		= 1 << IOC_PAGE_SHIFT,
301 	IOC_SECT_TO_PAGE_SHIFT	= IOC_PAGE_SHIFT - SECTOR_SHIFT,
302 
303 	/* if apart further than 16M, consider randio for linear model */
304 	LCOEF_RANDIO_PAGES	= 4096,
305 };
306 
307 enum ioc_running {
308 	IOC_IDLE,
309 	IOC_RUNNING,
310 	IOC_STOP,
311 };
312 
313 /* io.cost.qos controls including per-dev enable of the whole controller */
314 enum {
315 	QOS_ENABLE,
316 	QOS_CTRL,
317 	NR_QOS_CTRL_PARAMS,
318 };
319 
320 /* io.cost.qos params */
321 enum {
322 	QOS_RPPM,
323 	QOS_RLAT,
324 	QOS_WPPM,
325 	QOS_WLAT,
326 	QOS_MIN,
327 	QOS_MAX,
328 	NR_QOS_PARAMS,
329 };
330 
331 /* io.cost.model controls */
332 enum {
333 	COST_CTRL,
334 	COST_MODEL,
335 	NR_COST_CTRL_PARAMS,
336 };
337 
338 /* builtin linear cost model coefficients */
339 enum {
340 	I_LCOEF_RBPS,
341 	I_LCOEF_RSEQIOPS,
342 	I_LCOEF_RRANDIOPS,
343 	I_LCOEF_WBPS,
344 	I_LCOEF_WSEQIOPS,
345 	I_LCOEF_WRANDIOPS,
346 	NR_I_LCOEFS,
347 };
348 
349 enum {
350 	LCOEF_RPAGE,
351 	LCOEF_RSEQIO,
352 	LCOEF_RRANDIO,
353 	LCOEF_WPAGE,
354 	LCOEF_WSEQIO,
355 	LCOEF_WRANDIO,
356 	NR_LCOEFS,
357 };
358 
359 enum {
360 	AUTOP_INVALID,
361 	AUTOP_HDD,
362 	AUTOP_SSD_QD1,
363 	AUTOP_SSD_DFL,
364 	AUTOP_SSD_FAST,
365 };
366 
367 struct ioc_gq;
368 
369 struct ioc_params {
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;
375 };
376 
377 struct ioc_missed {
378 	u32				nr_met;
379 	u32				nr_missed;
380 	u32				last_met;
381 	u32				last_missed;
382 };
383 
384 struct ioc_pcpu_stat {
385 	struct ioc_missed		missed[2];
386 
387 	u64				rq_wait_ns;
388 	u64				last_rq_wait_ns;
389 };
390 
391 /* per device */
392 struct ioc {
393 	struct rq_qos			rqos;
394 
395 	bool				enabled;
396 
397 	struct ioc_params		params;
398 	u32				period_us;
399 	u32				margin_us;
400 	u64				vrate_min;
401 	u64				vrate_max;
402 
403 	spinlock_t			lock;
404 	struct timer_list		timer;
405 	struct list_head		active_iocgs;	/* active cgroups */
406 	struct ioc_pcpu_stat __percpu	*pcpu_stat;
407 
408 	enum ioc_running		running;
409 	atomic64_t			vtime_rate;
410 
411 	seqcount_t			period_seqcount;
412 	u32				period_at;	/* wallclock starttime */
413 	u64				period_at_vtime; /* vtime starttime */
414 
415 	atomic64_t			cur_period;	/* inc'd each period */
416 	int				busy_level;	/* saturation history */
417 
418 	u64				inuse_margin_vtime;
419 	bool				weights_updated;
420 	atomic_t			hweight_gen;	/* for lazy hweights */
421 
422 	u64				autop_too_fast_at;
423 	u64				autop_too_slow_at;
424 	int				autop_idx;
425 	bool				user_qos_params:1;
426 	bool				user_cost_model:1;
427 };
428 
429 /* per device-cgroup pair */
430 struct ioc_gq {
431 	struct blkg_policy_data		pd;
432 	struct ioc			*ioc;
433 
434 	/*
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
439 	 * sources.
440 	 *
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
444 	 * `hweight_inuse`.
445 	 *
446 	 * `last_inuse` remembers `inuse` while an iocg is idle to persist
447 	 * surplus adjustments.
448 	 */
449 	u32				cfg_weight;
450 	u32				weight;
451 	u32				active;
452 	u32				inuse;
453 	u32				last_inuse;
454 
455 	sector_t			cursor;		/* to detect randio */
456 
457 	/*
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
461 	 * overage.
462 	 *
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.
466 	 *
467 	 * `last_vtime` is used to remember `vtime` at the end of the last
468 	 * period to calculate utilization.
469 	 */
470 	atomic64_t			vtime;
471 	atomic64_t			done_vtime;
472 	atomic64_t			abs_vdebt;
473 	u64				last_vtime;
474 
475 	/*
476 	 * The period this iocg was last active in.  Used for deactivation
477 	 * and invalidating `vtime`.
478 	 */
479 	atomic64_t			active_period;
480 	struct list_head		active_list;
481 
482 	/* see __propagate_active_weight() and current_hweight() for details */
483 	u64				child_active_sum;
484 	u64				child_inuse_sum;
485 	int				hweight_gen;
486 	u32				hweight_active;
487 	u32				hweight_inuse;
488 	bool				has_surplus;
489 
490 	struct wait_queue_head		waitq;
491 	struct hrtimer			waitq_timer;
492 	struct hrtimer			delay_timer;
493 
494 	/* usage is recorded as fractions of HWEIGHT_WHOLE */
495 	int				usage_idx;
496 	u32				usages[NR_USAGE_SLOTS];
497 
498 	/* this iocg's depth in the hierarchy and ancestors including self */
499 	int				level;
500 	struct ioc_gq			*ancestors[];
501 };
502 
503 /* per cgroup */
504 struct ioc_cgrp {
505 	struct blkcg_policy_data	cpd;
506 	unsigned int			dfl_weight;
507 };
508 
509 struct ioc_now {
510 	u64				now_ns;
511 	u32				now;
512 	u64				vnow;
513 	u64				vrate;
514 };
515 
516 struct iocg_wait {
517 	struct wait_queue_entry		wait;
518 	struct bio			*bio;
519 	u64				abs_cost;
520 	bool				committed;
521 };
522 
523 struct iocg_wake_ctx {
524 	struct ioc_gq			*iocg;
525 	u32				hw_inuse;
526 	s64				vbudget;
527 };
528 
529 static const struct ioc_params autop[] = {
530 	[AUTOP_HDD] = {
531 		.qos				= {
532 			[QOS_RLAT]		=        250000, /* 250ms */
533 			[QOS_WLAT]		=        250000,
534 			[QOS_MIN]		= VRATE_MIN_PPM,
535 			[QOS_MAX]		= VRATE_MAX_PPM,
536 		},
537 		.i_lcoefs			= {
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,
544 		},
545 	},
546 	[AUTOP_SSD_QD1] = {
547 		.qos				= {
548 			[QOS_RLAT]		=         25000, /* 25ms */
549 			[QOS_WLAT]		=         25000,
550 			[QOS_MIN]		= VRATE_MIN_PPM,
551 			[QOS_MAX]		= VRATE_MAX_PPM,
552 		},
553 		.i_lcoefs			= {
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,
560 		},
561 	},
562 	[AUTOP_SSD_DFL] = {
563 		.qos				= {
564 			[QOS_RLAT]		=         25000, /* 25ms */
565 			[QOS_WLAT]		=         25000,
566 			[QOS_MIN]		= VRATE_MIN_PPM,
567 			[QOS_MAX]		= VRATE_MAX_PPM,
568 		},
569 		.i_lcoefs			= {
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,
576 		},
577 		.too_fast_vrate_pct		=           500,
578 	},
579 	[AUTOP_SSD_FAST] = {
580 		.qos				= {
581 			[QOS_RLAT]		=          5000, /* 5ms */
582 			[QOS_WLAT]		=          5000,
583 			[QOS_MIN]		= VRATE_MIN_PPM,
584 			[QOS_MAX]		= VRATE_MAX_PPM,
585 		},
586 		.i_lcoefs			= {
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,
593 		},
594 		.too_slow_vrate_pct		=            10,
595 	},
596 };
597 
598 /*
599  * vrate adjust percentages indexed by ioc->busy_level.  We adjust up on
600  * vtime credit shortage and down on device saturation.
601  */
602 static u32 vrate_adj_pct[] =
603 	{ 0, 0, 0, 0,
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 };
607 
608 static struct blkcg_policy blkcg_policy_iocost;
609 
610 /* accessors and helpers */
611 static struct ioc *rqos_to_ioc(struct rq_qos *rqos)
612 {
613 	return container_of(rqos, struct ioc, rqos);
614 }
615 
616 static struct ioc *q_to_ioc(struct request_queue *q)
617 {
618 	return rqos_to_ioc(rq_qos_id(q, RQ_QOS_COST));
619 }
620 
621 static const char *q_name(struct request_queue *q)
622 {
623 	if (test_bit(QUEUE_FLAG_REGISTERED, &q->queue_flags))
624 		return kobject_name(q->kobj.parent);
625 	else
626 		return "<unknown>";
627 }
628 
629 static const char __maybe_unused *ioc_name(struct ioc *ioc)
630 {
631 	return q_name(ioc->rqos.q);
632 }
633 
634 static struct ioc_gq *pd_to_iocg(struct blkg_policy_data *pd)
635 {
636 	return pd ? container_of(pd, struct ioc_gq, pd) : NULL;
637 }
638 
639 static struct ioc_gq *blkg_to_iocg(struct blkcg_gq *blkg)
640 {
641 	return pd_to_iocg(blkg_to_pd(blkg, &blkcg_policy_iocost));
642 }
643 
644 static struct blkcg_gq *iocg_to_blkg(struct ioc_gq *iocg)
645 {
646 	return pd_to_blkg(&iocg->pd);
647 }
648 
649 static struct ioc_cgrp *blkcg_to_iocc(struct blkcg *blkcg)
650 {
651 	return container_of(blkcg_to_cpd(blkcg, &blkcg_policy_iocost),
652 			    struct ioc_cgrp, cpd);
653 }
654 
655 /*
656  * Scale @abs_cost to the inverse of @hw_inuse.  The lower the hierarchical
657  * weight, the more expensive each IO.  Must round up.
658  */
659 static u64 abs_cost_to_cost(u64 abs_cost, u32 hw_inuse)
660 {
661 	return DIV64_U64_ROUND_UP(abs_cost * HWEIGHT_WHOLE, hw_inuse);
662 }
663 
664 /*
665  * The inverse of abs_cost_to_cost().  Must round up.
666  */
667 static u64 cost_to_abs_cost(u64 cost, u32 hw_inuse)
668 {
669 	return DIV64_U64_ROUND_UP(cost * hw_inuse, HWEIGHT_WHOLE);
670 }
671 
672 static void iocg_commit_bio(struct ioc_gq *iocg, struct bio *bio, u64 cost)
673 {
674 	bio->bi_iocost_cost = cost;
675 	atomic64_add(cost, &iocg->vtime);
676 }
677 
678 #define CREATE_TRACE_POINTS
679 #include <trace/events/iocost.h>
680 
681 /* latency Qos params changed, update period_us and all the dependent params */
682 static void ioc_refresh_period_us(struct ioc *ioc)
683 {
684 	u32 ppm, lat, multi, period_us;
685 
686 	lockdep_assert_held(&ioc->lock);
687 
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];
692 	} else {
693 		ppm = ioc->params.qos[QOS_WPPM];
694 		lat = ioc->params.qos[QOS_WLAT];
695 	}
696 
697 	/*
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).
704 	 */
705 	if (ppm)
706 		multi = max_t(u32, (MILLION - ppm) / 50000, 2);
707 	else
708 		multi = 2;
709 	period_us = multi * lat;
710 	period_us = clamp_t(u32, period_us, MIN_PERIOD, MAX_PERIOD);
711 
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);
717 }
718 
719 static int ioc_autop_idx(struct ioc *ioc)
720 {
721 	int idx = ioc->autop_idx;
722 	const struct ioc_params *p = &autop[idx];
723 	u32 vrate_pct;
724 	u64 now_ns;
725 
726 	/* rotational? */
727 	if (!blk_queue_nonrot(ioc->rqos.q))
728 		return AUTOP_HDD;
729 
730 	/* handle SATA SSDs w/ broken NCQ */
731 	if (blk_queue_depth(ioc->rqos.q) == 1)
732 		return AUTOP_SSD_QD1;
733 
734 	/* use one of the normal ssd sets */
735 	if (idx < AUTOP_SSD_DFL)
736 		return AUTOP_SSD_DFL;
737 
738 	/* if user is overriding anything, maintain what was there */
739 	if (ioc->user_qos_params || ioc->user_cost_model)
740 		return idx;
741 
742 	/* step up/down based on the vrate */
743 	vrate_pct = div64_u64(atomic64_read(&ioc->vtime_rate) * 100,
744 			      VTIME_PER_USEC);
745 	now_ns = ktime_get_ns();
746 
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)
751 			return idx + 1;
752 	} else {
753 		ioc->autop_too_fast_at = 0;
754 	}
755 
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)
760 			return idx - 1;
761 	} else {
762 		ioc->autop_too_slow_at = 0;
763 	}
764 
765 	return idx;
766 }
767 
768 /*
769  * Take the followings as input
770  *
771  *  @bps	maximum sequential throughput
772  *  @seqiops	maximum sequential 4k iops
773  *  @randiops	maximum random 4k iops
774  *
775  * and calculate the linear model cost coefficients.
776  *
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)
780  */
781 static void calc_lcoefs(u64 bps, u64 seqiops, u64 randiops,
782 			u64 *page, u64 *seqio, u64 *randio)
783 {
784 	u64 v;
785 
786 	*page = *seqio = *randio = 0;
787 
788 	if (bps)
789 		*page = DIV64_U64_ROUND_UP(VTIME_PER_SEC,
790 					   DIV_ROUND_UP_ULL(bps, IOC_PAGE_SIZE));
791 
792 	if (seqiops) {
793 		v = DIV64_U64_ROUND_UP(VTIME_PER_SEC, seqiops);
794 		if (v > *page)
795 			*seqio = v - *page;
796 	}
797 
798 	if (randiops) {
799 		v = DIV64_U64_ROUND_UP(VTIME_PER_SEC, randiops);
800 		if (v > *page)
801 			*randio = v - *page;
802 	}
803 }
804 
805 static void ioc_refresh_lcoefs(struct ioc *ioc)
806 {
807 	u64 *u = ioc->params.i_lcoefs;
808 	u64 *c = ioc->params.lcoefs;
809 
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]);
814 }
815 
816 static bool ioc_refresh_params(struct ioc *ioc, bool force)
817 {
818 	const struct ioc_params *p;
819 	int idx;
820 
821 	lockdep_assert_held(&ioc->lock);
822 
823 	idx = ioc_autop_idx(ioc);
824 	p = &autop[idx];
825 
826 	if (idx == ioc->autop_idx && !force)
827 		return false;
828 
829 	if (idx != ioc->autop_idx)
830 		atomic64_set(&ioc->vtime_rate, VTIME_PER_USEC);
831 
832 	ioc->autop_idx = idx;
833 	ioc->autop_too_fast_at = 0;
834 	ioc->autop_too_slow_at = 0;
835 
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));
840 
841 	ioc_refresh_period_us(ioc);
842 	ioc_refresh_lcoefs(ioc);
843 
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);
848 
849 	return true;
850 }
851 
852 /* take a snapshot of the current [v]time and vrate */
853 static void ioc_now(struct ioc *ioc, struct ioc_now *now)
854 {
855 	unsigned seq;
856 
857 	now->now_ns = ktime_get();
858 	now->now = ktime_to_us(now->now_ns);
859 	now->vrate = atomic64_read(&ioc->vtime_rate);
860 
861 	/*
862 	 * The current vtime is
863 	 *
864 	 *   vtime at period start + (wallclock time since the start) * vrate
865 	 *
866 	 * As a consistent snapshot of `period_at_vtime` and `period_at` is
867 	 * needed, they're seqcount protected.
868 	 */
869 	do {
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));
874 }
875 
876 static void ioc_start_period(struct ioc *ioc, struct ioc_now *now)
877 {
878 	lockdep_assert_held(&ioc->lock);
879 	WARN_ON_ONCE(ioc->running != IOC_RUNNING);
880 
881 	write_seqcount_begin(&ioc->period_seqcount);
882 	ioc->period_at = now->now;
883 	ioc->period_at_vtime = now->vnow;
884 	write_seqcount_end(&ioc->period_seqcount);
885 
886 	ioc->timer.expires = jiffies + usecs_to_jiffies(ioc->period_us);
887 	add_timer(&ioc->timer);
888 }
889 
890 /*
891  * Update @iocg's `active` and `inuse` to @active and @inuse, update level
892  * weight sums and propagate upwards accordingly.
893  */
894 static void __propagate_active_weight(struct ioc_gq *iocg, u32 active, u32 inuse)
895 {
896 	struct ioc *ioc = iocg->ioc;
897 	int lvl;
898 
899 	lockdep_assert_held(&ioc->lock);
900 
901 	inuse = min(active, inuse);
902 
903 	for (lvl = iocg->level - 1; lvl >= 0; lvl--) {
904 		struct ioc_gq *parent = iocg->ancestors[lvl];
905 		struct ioc_gq *child = iocg->ancestors[lvl + 1];
906 		u32 parent_active = 0, parent_inuse = 0;
907 
908 		/* update the level sums */
909 		parent->child_active_sum += (s32)(active - child->active);
910 		parent->child_inuse_sum += (s32)(inuse - child->inuse);
911 		/* apply the udpates */
912 		child->active = active;
913 		child->inuse = inuse;
914 
915 		/*
916 		 * The delta between inuse and active sums indicates that
917 		 * that much of weight is being given away.  Parent's inuse
918 		 * and active should reflect the ratio.
919 		 */
920 		if (parent->child_active_sum) {
921 			parent_active = parent->weight;
922 			parent_inuse = DIV64_U64_ROUND_UP(
923 				parent_active * parent->child_inuse_sum,
924 				parent->child_active_sum);
925 		}
926 
927 		/* do we need to keep walking up? */
928 		if (parent_active == parent->active &&
929 		    parent_inuse == parent->inuse)
930 			break;
931 
932 		active = parent_active;
933 		inuse = parent_inuse;
934 	}
935 
936 	ioc->weights_updated = true;
937 }
938 
939 static void commit_active_weights(struct ioc *ioc)
940 {
941 	lockdep_assert_held(&ioc->lock);
942 
943 	if (ioc->weights_updated) {
944 		/* paired with rmb in current_hweight(), see there */
945 		smp_wmb();
946 		atomic_inc(&ioc->hweight_gen);
947 		ioc->weights_updated = false;
948 	}
949 }
950 
951 static void propagate_active_weight(struct ioc_gq *iocg, u32 active, u32 inuse)
952 {
953 	__propagate_active_weight(iocg, active, inuse);
954 	commit_active_weights(iocg->ioc);
955 }
956 
957 static void current_hweight(struct ioc_gq *iocg, u32 *hw_activep, u32 *hw_inusep)
958 {
959 	struct ioc *ioc = iocg->ioc;
960 	int lvl;
961 	u32 hwa, hwi;
962 	int ioc_gen;
963 
964 	/* hot path - if uptodate, use cached */
965 	ioc_gen = atomic_read(&ioc->hweight_gen);
966 	if (ioc_gen == iocg->hweight_gen)
967 		goto out;
968 
969 	/*
970 	 * Paired with wmb in commit_active_weights().  If we saw the
971 	 * updated hweight_gen, all the weight updates from
972 	 * __propagate_active_weight() are visible too.
973 	 *
974 	 * We can race with weight updates during calculation and get it
975 	 * wrong.  However, hweight_gen would have changed and a future
976 	 * reader will recalculate and we're guaranteed to discard the
977 	 * wrong result soon.
978 	 */
979 	smp_rmb();
980 
981 	hwa = hwi = HWEIGHT_WHOLE;
982 	for (lvl = 0; lvl <= iocg->level - 1; lvl++) {
983 		struct ioc_gq *parent = iocg->ancestors[lvl];
984 		struct ioc_gq *child = iocg->ancestors[lvl + 1];
985 		u32 active_sum = READ_ONCE(parent->child_active_sum);
986 		u32 inuse_sum = READ_ONCE(parent->child_inuse_sum);
987 		u32 active = READ_ONCE(child->active);
988 		u32 inuse = READ_ONCE(child->inuse);
989 
990 		/* we can race with deactivations and either may read as zero */
991 		if (!active_sum || !inuse_sum)
992 			continue;
993 
994 		active_sum = max(active, active_sum);
995 		hwa = hwa * active / active_sum;	/* max 16bits * 10000 */
996 
997 		inuse_sum = max(inuse, inuse_sum);
998 		hwi = hwi * inuse / inuse_sum;		/* max 16bits * 10000 */
999 	}
1000 
1001 	iocg->hweight_active = max_t(u32, hwa, 1);
1002 	iocg->hweight_inuse = max_t(u32, hwi, 1);
1003 	iocg->hweight_gen = ioc_gen;
1004 out:
1005 	if (hw_activep)
1006 		*hw_activep = iocg->hweight_active;
1007 	if (hw_inusep)
1008 		*hw_inusep = iocg->hweight_inuse;
1009 }
1010 
1011 static void weight_updated(struct ioc_gq *iocg)
1012 {
1013 	struct ioc *ioc = iocg->ioc;
1014 	struct blkcg_gq *blkg = iocg_to_blkg(iocg);
1015 	struct ioc_cgrp *iocc = blkcg_to_iocc(blkg->blkcg);
1016 	u32 weight;
1017 
1018 	lockdep_assert_held(&ioc->lock);
1019 
1020 	weight = iocg->cfg_weight ?: iocc->dfl_weight;
1021 	if (weight != iocg->weight && iocg->active)
1022 		propagate_active_weight(iocg, weight,
1023 			DIV64_U64_ROUND_UP(iocg->inuse * weight, iocg->weight));
1024 	iocg->weight = weight;
1025 }
1026 
1027 static bool iocg_activate(struct ioc_gq *iocg, struct ioc_now *now)
1028 {
1029 	struct ioc *ioc = iocg->ioc;
1030 	u64 last_period, cur_period, max_period_delta;
1031 	u64 vtime, vmargin, vmin;
1032 	int i;
1033 
1034 	/*
1035 	 * If seem to be already active, just update the stamp to tell the
1036 	 * timer that we're still active.  We don't mind occassional races.
1037 	 */
1038 	if (!list_empty(&iocg->active_list)) {
1039 		ioc_now(ioc, now);
1040 		cur_period = atomic64_read(&ioc->cur_period);
1041 		if (atomic64_read(&iocg->active_period) != cur_period)
1042 			atomic64_set(&iocg->active_period, cur_period);
1043 		return true;
1044 	}
1045 
1046 	/* racy check on internal node IOs, treat as root level IOs */
1047 	if (iocg->child_active_sum)
1048 		return false;
1049 
1050 	spin_lock_irq(&ioc->lock);
1051 
1052 	ioc_now(ioc, now);
1053 
1054 	/* update period */
1055 	cur_period = atomic64_read(&ioc->cur_period);
1056 	last_period = atomic64_read(&iocg->active_period);
1057 	atomic64_set(&iocg->active_period, cur_period);
1058 
1059 	/* already activated or breaking leaf-only constraint? */
1060 	if (!list_empty(&iocg->active_list))
1061 		goto succeed_unlock;
1062 	for (i = iocg->level - 1; i > 0; i--)
1063 		if (!list_empty(&iocg->ancestors[i]->active_list))
1064 			goto fail_unlock;
1065 
1066 	if (iocg->child_active_sum)
1067 		goto fail_unlock;
1068 
1069 	/*
1070 	 * vtime may wrap when vrate is raised substantially due to
1071 	 * underestimated IO costs.  Look at the period and ignore its
1072 	 * vtime if the iocg has been idle for too long.  Also, cap the
1073 	 * budget it can start with to the margin.
1074 	 */
1075 	max_period_delta = DIV64_U64_ROUND_UP(VTIME_VALID_DUR, ioc->period_us);
1076 	vtime = atomic64_read(&iocg->vtime);
1077 	vmargin = ioc->margin_us * now->vrate;
1078 	vmin = now->vnow - vmargin;
1079 
1080 	if (last_period + max_period_delta < cur_period ||
1081 	    time_before64(vtime, vmin)) {
1082 		atomic64_add(vmin - vtime, &iocg->vtime);
1083 		atomic64_add(vmin - vtime, &iocg->done_vtime);
1084 		vtime = vmin;
1085 	}
1086 
1087 	/*
1088 	 * Activate, propagate weight and start period timer if not
1089 	 * running.  Reset hweight_gen to avoid accidental match from
1090 	 * wrapping.
1091 	 */
1092 	iocg->hweight_gen = atomic_read(&ioc->hweight_gen) - 1;
1093 	list_add(&iocg->active_list, &ioc->active_iocgs);
1094 	propagate_active_weight(iocg, iocg->weight,
1095 				iocg->last_inuse ?: iocg->weight);
1096 
1097 	TRACE_IOCG_PATH(iocg_activate, iocg, now,
1098 			last_period, cur_period, vtime);
1099 
1100 	iocg->last_vtime = vtime;
1101 
1102 	if (ioc->running == IOC_IDLE) {
1103 		ioc->running = IOC_RUNNING;
1104 		ioc_start_period(ioc, now);
1105 	}
1106 
1107 succeed_unlock:
1108 	spin_unlock_irq(&ioc->lock);
1109 	return true;
1110 
1111 fail_unlock:
1112 	spin_unlock_irq(&ioc->lock);
1113 	return false;
1114 }
1115 
1116 static int iocg_wake_fn(struct wait_queue_entry *wq_entry, unsigned mode,
1117 			int flags, void *key)
1118 {
1119 	struct iocg_wait *wait = container_of(wq_entry, struct iocg_wait, wait);
1120 	struct iocg_wake_ctx *ctx = (struct iocg_wake_ctx *)key;
1121 	u64 cost = abs_cost_to_cost(wait->abs_cost, ctx->hw_inuse);
1122 
1123 	ctx->vbudget -= cost;
1124 
1125 	if (ctx->vbudget < 0)
1126 		return -1;
1127 
1128 	iocg_commit_bio(ctx->iocg, wait->bio, cost);
1129 
1130 	/*
1131 	 * autoremove_wake_function() removes the wait entry only when it
1132 	 * actually changed the task state.  We want the wait always
1133 	 * removed.  Remove explicitly and use default_wake_function().
1134 	 */
1135 	list_del_init(&wq_entry->entry);
1136 	wait->committed = true;
1137 
1138 	default_wake_function(wq_entry, mode, flags, key);
1139 	return 0;
1140 }
1141 
1142 static void iocg_kick_waitq(struct ioc_gq *iocg, struct ioc_now *now)
1143 {
1144 	struct ioc *ioc = iocg->ioc;
1145 	struct iocg_wake_ctx ctx = { .iocg = iocg };
1146 	u64 margin_ns = (u64)(ioc->period_us *
1147 			      WAITQ_TIMER_MARGIN_PCT / 100) * NSEC_PER_USEC;
1148 	u64 abs_vdebt, vdebt, vshortage, expires, oexpires;
1149 	s64 vbudget;
1150 	u32 hw_inuse;
1151 
1152 	lockdep_assert_held(&iocg->waitq.lock);
1153 
1154 	current_hweight(iocg, NULL, &hw_inuse);
1155 	vbudget = now->vnow - atomic64_read(&iocg->vtime);
1156 
1157 	/* pay off debt */
1158 	abs_vdebt = atomic64_read(&iocg->abs_vdebt);
1159 	vdebt = abs_cost_to_cost(abs_vdebt, hw_inuse);
1160 	if (vdebt && vbudget > 0) {
1161 		u64 delta = min_t(u64, vbudget, vdebt);
1162 		u64 abs_delta = min(cost_to_abs_cost(delta, hw_inuse),
1163 				    abs_vdebt);
1164 
1165 		atomic64_add(delta, &iocg->vtime);
1166 		atomic64_add(delta, &iocg->done_vtime);
1167 		atomic64_sub(abs_delta, &iocg->abs_vdebt);
1168 		if (WARN_ON_ONCE(atomic64_read(&iocg->abs_vdebt) < 0))
1169 			atomic64_set(&iocg->abs_vdebt, 0);
1170 	}
1171 
1172 	/*
1173 	 * Wake up the ones which are due and see how much vtime we'll need
1174 	 * for the next one.
1175 	 */
1176 	ctx.hw_inuse = hw_inuse;
1177 	ctx.vbudget = vbudget - vdebt;
1178 	__wake_up_locked_key(&iocg->waitq, TASK_NORMAL, &ctx);
1179 	if (!waitqueue_active(&iocg->waitq))
1180 		return;
1181 	if (WARN_ON_ONCE(ctx.vbudget >= 0))
1182 		return;
1183 
1184 	/* determine next wakeup, add a quarter margin to guarantee chunking */
1185 	vshortage = -ctx.vbudget;
1186 	expires = now->now_ns +
1187 		DIV64_U64_ROUND_UP(vshortage, now->vrate) * NSEC_PER_USEC;
1188 	expires += margin_ns / 4;
1189 
1190 	/* if already active and close enough, don't bother */
1191 	oexpires = ktime_to_ns(hrtimer_get_softexpires(&iocg->waitq_timer));
1192 	if (hrtimer_is_queued(&iocg->waitq_timer) &&
1193 	    abs(oexpires - expires) <= margin_ns / 4)
1194 		return;
1195 
1196 	hrtimer_start_range_ns(&iocg->waitq_timer, ns_to_ktime(expires),
1197 			       margin_ns / 4, HRTIMER_MODE_ABS);
1198 }
1199 
1200 static enum hrtimer_restart iocg_waitq_timer_fn(struct hrtimer *timer)
1201 {
1202 	struct ioc_gq *iocg = container_of(timer, struct ioc_gq, waitq_timer);
1203 	struct ioc_now now;
1204 	unsigned long flags;
1205 
1206 	ioc_now(iocg->ioc, &now);
1207 
1208 	spin_lock_irqsave(&iocg->waitq.lock, flags);
1209 	iocg_kick_waitq(iocg, &now);
1210 	spin_unlock_irqrestore(&iocg->waitq.lock, flags);
1211 
1212 	return HRTIMER_NORESTART;
1213 }
1214 
1215 static void iocg_kick_delay(struct ioc_gq *iocg, struct ioc_now *now, u64 cost)
1216 {
1217 	struct ioc *ioc = iocg->ioc;
1218 	struct blkcg_gq *blkg = iocg_to_blkg(iocg);
1219 	u64 vtime = atomic64_read(&iocg->vtime);
1220 	u64 vmargin = ioc->margin_us * now->vrate;
1221 	u64 margin_ns = ioc->margin_us * NSEC_PER_USEC;
1222 	u64 expires, oexpires;
1223 	u32 hw_inuse;
1224 
1225 	/* debt-adjust vtime */
1226 	current_hweight(iocg, NULL, &hw_inuse);
1227 	vtime += abs_cost_to_cost(atomic64_read(&iocg->abs_vdebt), hw_inuse);
1228 
1229 	/* clear or maintain depending on the overage */
1230 	if (time_before_eq64(vtime, now->vnow)) {
1231 		blkcg_clear_delay(blkg);
1232 		return;
1233 	}
1234 	if (!atomic_read(&blkg->use_delay) &&
1235 	    time_before_eq64(vtime, now->vnow + vmargin))
1236 		return;
1237 
1238 	/* use delay */
1239 	if (cost) {
1240 		u64 cost_ns = DIV64_U64_ROUND_UP(cost * NSEC_PER_USEC,
1241 						 now->vrate);
1242 		blkcg_add_delay(blkg, now->now_ns, cost_ns);
1243 	}
1244 	blkcg_use_delay(blkg);
1245 
1246 	expires = now->now_ns + DIV64_U64_ROUND_UP(vtime - now->vnow,
1247 						   now->vrate) * NSEC_PER_USEC;
1248 
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)
1253 		return;
1254 
1255 	hrtimer_start_range_ns(&iocg->delay_timer, ns_to_ktime(expires),
1256 			       margin_ns / 4, HRTIMER_MODE_ABS);
1257 }
1258 
1259 static enum hrtimer_restart iocg_delay_timer_fn(struct hrtimer *timer)
1260 {
1261 	struct ioc_gq *iocg = container_of(timer, struct ioc_gq, delay_timer);
1262 	struct ioc_now now;
1263 
1264 	ioc_now(iocg->ioc, &now);
1265 	iocg_kick_delay(iocg, &now, 0);
1266 
1267 	return HRTIMER_NORESTART;
1268 }
1269 
1270 static void ioc_lat_stat(struct ioc *ioc, u32 *missed_ppm_ar, u32 *rq_wait_pct_p)
1271 {
1272 	u32 nr_met[2] = { };
1273 	u32 nr_missed[2] = { };
1274 	u64 rq_wait_ns = 0;
1275 	int cpu, rw;
1276 
1277 	for_each_online_cpu(cpu) {
1278 		struct ioc_pcpu_stat *stat = per_cpu_ptr(ioc->pcpu_stat, cpu);
1279 		u64 this_rq_wait_ns;
1280 
1281 		for (rw = READ; rw <= WRITE; rw++) {
1282 			u32 this_met = READ_ONCE(stat->missed[rw].nr_met);
1283 			u32 this_missed = READ_ONCE(stat->missed[rw].nr_missed);
1284 
1285 			nr_met[rw] += this_met - stat->missed[rw].last_met;
1286 			nr_missed[rw] += this_missed - stat->missed[rw].last_missed;
1287 			stat->missed[rw].last_met = this_met;
1288 			stat->missed[rw].last_missed = this_missed;
1289 		}
1290 
1291 		this_rq_wait_ns = READ_ONCE(stat->rq_wait_ns);
1292 		rq_wait_ns += this_rq_wait_ns - stat->last_rq_wait_ns;
1293 		stat->last_rq_wait_ns = this_rq_wait_ns;
1294 	}
1295 
1296 	for (rw = READ; rw <= WRITE; rw++) {
1297 		if (nr_met[rw] + nr_missed[rw])
1298 			missed_ppm_ar[rw] =
1299 				DIV64_U64_ROUND_UP((u64)nr_missed[rw] * MILLION,
1300 						   nr_met[rw] + nr_missed[rw]);
1301 		else
1302 			missed_ppm_ar[rw] = 0;
1303 	}
1304 
1305 	*rq_wait_pct_p = div64_u64(rq_wait_ns * 100,
1306 				   ioc->period_us * NSEC_PER_USEC);
1307 }
1308 
1309 /* was iocg idle this period? */
1310 static bool iocg_is_idle(struct ioc_gq *iocg)
1311 {
1312 	struct ioc *ioc = iocg->ioc;
1313 
1314 	/* did something get issued this period? */
1315 	if (atomic64_read(&iocg->active_period) ==
1316 	    atomic64_read(&ioc->cur_period))
1317 		return false;
1318 
1319 	/* is something in flight? */
1320 	if (atomic64_read(&iocg->done_vtime) < atomic64_read(&iocg->vtime))
1321 		return false;
1322 
1323 	return true;
1324 }
1325 
1326 /* returns usage with margin added if surplus is large enough */
1327 static u32 surplus_adjusted_hweight_inuse(u32 usage, u32 hw_inuse)
1328 {
1329 	/* add margin */
1330 	usage = DIV_ROUND_UP(usage * SURPLUS_SCALE_PCT, 100);
1331 	usage += SURPLUS_SCALE_ABS;
1332 
1333 	/* don't bother if the surplus is too small */
1334 	if (usage + SURPLUS_MIN_ADJ_DELTA > hw_inuse)
1335 		return 0;
1336 
1337 	return usage;
1338 }
1339 
1340 static void ioc_timer_fn(struct timer_list *timer)
1341 {
1342 	struct ioc *ioc = container_of(timer, struct ioc, timer);
1343 	struct ioc_gq *iocg, *tiocg;
1344 	struct ioc_now now;
1345 	int nr_surpluses = 0, nr_shortages = 0, nr_lagging = 0;
1346 	u32 ppm_rthr = MILLION - ioc->params.qos[QOS_RPPM];
1347 	u32 ppm_wthr = MILLION - ioc->params.qos[QOS_WPPM];
1348 	u32 missed_ppm[2], rq_wait_pct;
1349 	u64 period_vtime;
1350 	int prev_busy_level, i;
1351 
1352 	/* how were the latencies during the period? */
1353 	ioc_lat_stat(ioc, missed_ppm, &rq_wait_pct);
1354 
1355 	/* take care of active iocgs */
1356 	spin_lock_irq(&ioc->lock);
1357 
1358 	ioc_now(ioc, &now);
1359 
1360 	period_vtime = now.vnow - ioc->period_at_vtime;
1361 	if (WARN_ON_ONCE(!period_vtime)) {
1362 		spin_unlock_irq(&ioc->lock);
1363 		return;
1364 	}
1365 
1366 	/*
1367 	 * Waiters determine the sleep durations based on the vrate they
1368 	 * saw at the time of sleep.  If vrate has increased, some waiters
1369 	 * could be sleeping for too long.  Wake up tardy waiters which
1370 	 * should have woken up in the last period and expire idle iocgs.
1371 	 */
1372 	list_for_each_entry_safe(iocg, tiocg, &ioc->active_iocgs, active_list) {
1373 		if (!waitqueue_active(&iocg->waitq) &&
1374 		    !atomic64_read(&iocg->abs_vdebt) && !iocg_is_idle(iocg))
1375 			continue;
1376 
1377 		spin_lock(&iocg->waitq.lock);
1378 
1379 		if (waitqueue_active(&iocg->waitq) ||
1380 		    atomic64_read(&iocg->abs_vdebt)) {
1381 			/* might be oversleeping vtime / hweight changes, kick */
1382 			iocg_kick_waitq(iocg, &now);
1383 			iocg_kick_delay(iocg, &now, 0);
1384 		} else if (iocg_is_idle(iocg)) {
1385 			/* no waiter and idle, deactivate */
1386 			iocg->last_inuse = iocg->inuse;
1387 			__propagate_active_weight(iocg, 0, 0);
1388 			list_del_init(&iocg->active_list);
1389 		}
1390 
1391 		spin_unlock(&iocg->waitq.lock);
1392 	}
1393 	commit_active_weights(ioc);
1394 
1395 	/* calc usages and see whether some weights need to be moved around */
1396 	list_for_each_entry(iocg, &ioc->active_iocgs, active_list) {
1397 		u64 vdone, vtime, vusage, vmargin, vmin;
1398 		u32 hw_active, hw_inuse, usage;
1399 
1400 		/*
1401 		 * Collect unused and wind vtime closer to vnow to prevent
1402 		 * iocgs from accumulating a large amount of budget.
1403 		 */
1404 		vdone = atomic64_read(&iocg->done_vtime);
1405 		vtime = atomic64_read(&iocg->vtime);
1406 		current_hweight(iocg, &hw_active, &hw_inuse);
1407 
1408 		/*
1409 		 * Latency QoS detection doesn't account for IOs which are
1410 		 * in-flight for longer than a period.  Detect them by
1411 		 * comparing vdone against period start.  If lagging behind
1412 		 * IOs from past periods, don't increase vrate.
1413 		 */
1414 		if ((ppm_rthr != MILLION || ppm_wthr != MILLION) &&
1415 		    !atomic_read(&iocg_to_blkg(iocg)->use_delay) &&
1416 		    time_after64(vtime, vdone) &&
1417 		    time_after64(vtime, now.vnow -
1418 				 MAX_LAGGING_PERIODS * period_vtime) &&
1419 		    time_before64(vdone, now.vnow - period_vtime))
1420 			nr_lagging++;
1421 
1422 		if (waitqueue_active(&iocg->waitq))
1423 			vusage = now.vnow - iocg->last_vtime;
1424 		else if (time_before64(iocg->last_vtime, vtime))
1425 			vusage = vtime - iocg->last_vtime;
1426 		else
1427 			vusage = 0;
1428 
1429 		iocg->last_vtime += vusage;
1430 		/*
1431 		 * Factor in in-flight vtime into vusage to avoid
1432 		 * high-latency completions appearing as idle.  This should
1433 		 * be done after the above ->last_time adjustment.
1434 		 */
1435 		vusage = max(vusage, vtime - vdone);
1436 
1437 		/* calculate hweight based usage ratio and record */
1438 		if (vusage) {
1439 			usage = DIV64_U64_ROUND_UP(vusage * hw_inuse,
1440 						   period_vtime);
1441 			iocg->usage_idx = (iocg->usage_idx + 1) % NR_USAGE_SLOTS;
1442 			iocg->usages[iocg->usage_idx] = usage;
1443 		} else {
1444 			usage = 0;
1445 		}
1446 
1447 		/* see whether there's surplus vtime */
1448 		vmargin = ioc->margin_us * now.vrate;
1449 		vmin = now.vnow - vmargin;
1450 
1451 		iocg->has_surplus = false;
1452 
1453 		if (!waitqueue_active(&iocg->waitq) &&
1454 		    time_before64(vtime, vmin)) {
1455 			u64 delta = vmin - vtime;
1456 
1457 			/* throw away surplus vtime */
1458 			atomic64_add(delta, &iocg->vtime);
1459 			atomic64_add(delta, &iocg->done_vtime);
1460 			iocg->last_vtime += delta;
1461 			/* if usage is sufficiently low, maybe it can donate */
1462 			if (surplus_adjusted_hweight_inuse(usage, hw_inuse)) {
1463 				iocg->has_surplus = true;
1464 				nr_surpluses++;
1465 			}
1466 		} else if (hw_inuse < hw_active) {
1467 			u32 new_hwi, new_inuse;
1468 
1469 			/* was donating but might need to take back some */
1470 			if (waitqueue_active(&iocg->waitq)) {
1471 				new_hwi = hw_active;
1472 			} else {
1473 				new_hwi = max(hw_inuse,
1474 					      usage * SURPLUS_SCALE_PCT / 100 +
1475 					      SURPLUS_SCALE_ABS);
1476 			}
1477 
1478 			new_inuse = div64_u64((u64)iocg->inuse * new_hwi,
1479 					      hw_inuse);
1480 			new_inuse = clamp_t(u32, new_inuse, 1, iocg->active);
1481 
1482 			if (new_inuse > iocg->inuse) {
1483 				TRACE_IOCG_PATH(inuse_takeback, iocg, &now,
1484 						iocg->inuse, new_inuse,
1485 						hw_inuse, new_hwi);
1486 				__propagate_active_weight(iocg, iocg->weight,
1487 							  new_inuse);
1488 			}
1489 		} else {
1490 			/* genuninely out of vtime */
1491 			nr_shortages++;
1492 		}
1493 	}
1494 
1495 	if (!nr_shortages || !nr_surpluses)
1496 		goto skip_surplus_transfers;
1497 
1498 	/* there are both shortages and surpluses, transfer surpluses */
1499 	list_for_each_entry(iocg, &ioc->active_iocgs, active_list) {
1500 		u32 usage, hw_active, hw_inuse, new_hwi, new_inuse;
1501 		int nr_valid = 0;
1502 
1503 		if (!iocg->has_surplus)
1504 			continue;
1505 
1506 		/* base the decision on max historical usage */
1507 		for (i = 0, usage = 0; i < NR_USAGE_SLOTS; i++) {
1508 			if (iocg->usages[i]) {
1509 				usage = max(usage, iocg->usages[i]);
1510 				nr_valid++;
1511 			}
1512 		}
1513 		if (nr_valid < MIN_VALID_USAGES)
1514 			continue;
1515 
1516 		current_hweight(iocg, &hw_active, &hw_inuse);
1517 		new_hwi = surplus_adjusted_hweight_inuse(usage, hw_inuse);
1518 		if (!new_hwi)
1519 			continue;
1520 
1521 		new_inuse = DIV64_U64_ROUND_UP((u64)iocg->inuse * new_hwi,
1522 					       hw_inuse);
1523 		if (new_inuse < iocg->inuse) {
1524 			TRACE_IOCG_PATH(inuse_giveaway, iocg, &now,
1525 					iocg->inuse, new_inuse,
1526 					hw_inuse, new_hwi);
1527 			__propagate_active_weight(iocg, iocg->weight, new_inuse);
1528 		}
1529 	}
1530 skip_surplus_transfers:
1531 	commit_active_weights(ioc);
1532 
1533 	/*
1534 	 * If q is getting clogged or we're missing too much, we're issuing
1535 	 * too much IO and should lower vtime rate.  If we're not missing
1536 	 * and experiencing shortages but not surpluses, we're too stingy
1537 	 * and should increase vtime rate.
1538 	 */
1539 	prev_busy_level = ioc->busy_level;
1540 	if (rq_wait_pct > RQ_WAIT_BUSY_PCT ||
1541 	    missed_ppm[READ] > ppm_rthr ||
1542 	    missed_ppm[WRITE] > ppm_wthr) {
1543 		ioc->busy_level = max(ioc->busy_level, 0);
1544 		ioc->busy_level++;
1545 	} else if (rq_wait_pct <= RQ_WAIT_BUSY_PCT * UNBUSY_THR_PCT / 100 &&
1546 		   missed_ppm[READ] <= ppm_rthr * UNBUSY_THR_PCT / 100 &&
1547 		   missed_ppm[WRITE] <= ppm_wthr * UNBUSY_THR_PCT / 100) {
1548 		/* take action iff there is contention */
1549 		if (nr_shortages && !nr_lagging) {
1550 			ioc->busy_level = min(ioc->busy_level, 0);
1551 			/* redistribute surpluses first */
1552 			if (!nr_surpluses)
1553 				ioc->busy_level--;
1554 		}
1555 	} else {
1556 		ioc->busy_level = 0;
1557 	}
1558 
1559 	ioc->busy_level = clamp(ioc->busy_level, -1000, 1000);
1560 
1561 	if (ioc->busy_level > 0 || (ioc->busy_level < 0 && !nr_lagging)) {
1562 		u64 vrate = atomic64_read(&ioc->vtime_rate);
1563 		u64 vrate_min = ioc->vrate_min, vrate_max = ioc->vrate_max;
1564 
1565 		/* rq_wait signal is always reliable, ignore user vrate_min */
1566 		if (rq_wait_pct > RQ_WAIT_BUSY_PCT)
1567 			vrate_min = VRATE_MIN;
1568 
1569 		/*
1570 		 * If vrate is out of bounds, apply clamp gradually as the
1571 		 * bounds can change abruptly.  Otherwise, apply busy_level
1572 		 * based adjustment.
1573 		 */
1574 		if (vrate < vrate_min) {
1575 			vrate = div64_u64(vrate * (100 + VRATE_CLAMP_ADJ_PCT),
1576 					  100);
1577 			vrate = min(vrate, vrate_min);
1578 		} else if (vrate > vrate_max) {
1579 			vrate = div64_u64(vrate * (100 - VRATE_CLAMP_ADJ_PCT),
1580 					  100);
1581 			vrate = max(vrate, vrate_max);
1582 		} else {
1583 			int idx = min_t(int, abs(ioc->busy_level),
1584 					ARRAY_SIZE(vrate_adj_pct) - 1);
1585 			u32 adj_pct = vrate_adj_pct[idx];
1586 
1587 			if (ioc->busy_level > 0)
1588 				adj_pct = 100 - adj_pct;
1589 			else
1590 				adj_pct = 100 + adj_pct;
1591 
1592 			vrate = clamp(DIV64_U64_ROUND_UP(vrate * adj_pct, 100),
1593 				      vrate_min, vrate_max);
1594 		}
1595 
1596 		trace_iocost_ioc_vrate_adj(ioc, vrate, &missed_ppm, rq_wait_pct,
1597 					   nr_lagging, nr_shortages,
1598 					   nr_surpluses);
1599 
1600 		atomic64_set(&ioc->vtime_rate, vrate);
1601 		ioc->inuse_margin_vtime = DIV64_U64_ROUND_UP(
1602 			ioc->period_us * vrate * INUSE_MARGIN_PCT, 100);
1603 	} else if (ioc->busy_level != prev_busy_level || nr_lagging) {
1604 		trace_iocost_ioc_vrate_adj(ioc, atomic64_read(&ioc->vtime_rate),
1605 					   &missed_ppm, rq_wait_pct, nr_lagging,
1606 					   nr_shortages, nr_surpluses);
1607 	}
1608 
1609 	ioc_refresh_params(ioc, false);
1610 
1611 	/*
1612 	 * This period is done.  Move onto the next one.  If nothing's
1613 	 * going on with the device, stop the timer.
1614 	 */
1615 	atomic64_inc(&ioc->cur_period);
1616 
1617 	if (ioc->running != IOC_STOP) {
1618 		if (!list_empty(&ioc->active_iocgs)) {
1619 			ioc_start_period(ioc, &now);
1620 		} else {
1621 			ioc->busy_level = 0;
1622 			ioc->running = IOC_IDLE;
1623 		}
1624 	}
1625 
1626 	spin_unlock_irq(&ioc->lock);
1627 }
1628 
1629 static void calc_vtime_cost_builtin(struct bio *bio, struct ioc_gq *iocg,
1630 				    bool is_merge, u64 *costp)
1631 {
1632 	struct ioc *ioc = iocg->ioc;
1633 	u64 coef_seqio, coef_randio, coef_page;
1634 	u64 pages = max_t(u64, bio_sectors(bio) >> IOC_SECT_TO_PAGE_SHIFT, 1);
1635 	u64 seek_pages = 0;
1636 	u64 cost = 0;
1637 
1638 	switch (bio_op(bio)) {
1639 	case REQ_OP_READ:
1640 		coef_seqio	= ioc->params.lcoefs[LCOEF_RSEQIO];
1641 		coef_randio	= ioc->params.lcoefs[LCOEF_RRANDIO];
1642 		coef_page	= ioc->params.lcoefs[LCOEF_RPAGE];
1643 		break;
1644 	case REQ_OP_WRITE:
1645 		coef_seqio	= ioc->params.lcoefs[LCOEF_WSEQIO];
1646 		coef_randio	= ioc->params.lcoefs[LCOEF_WRANDIO];
1647 		coef_page	= ioc->params.lcoefs[LCOEF_WPAGE];
1648 		break;
1649 	default:
1650 		goto out;
1651 	}
1652 
1653 	if (iocg->cursor) {
1654 		seek_pages = abs(bio->bi_iter.bi_sector - iocg->cursor);
1655 		seek_pages >>= IOC_SECT_TO_PAGE_SHIFT;
1656 	}
1657 
1658 	if (!is_merge) {
1659 		if (seek_pages > LCOEF_RANDIO_PAGES) {
1660 			cost += coef_randio;
1661 		} else {
1662 			cost += coef_seqio;
1663 		}
1664 	}
1665 	cost += pages * coef_page;
1666 out:
1667 	*costp = cost;
1668 }
1669 
1670 static u64 calc_vtime_cost(struct bio *bio, struct ioc_gq *iocg, bool is_merge)
1671 {
1672 	u64 cost;
1673 
1674 	calc_vtime_cost_builtin(bio, iocg, is_merge, &cost);
1675 	return cost;
1676 }
1677 
1678 static void ioc_rqos_throttle(struct rq_qos *rqos, struct bio *bio)
1679 {
1680 	struct blkcg_gq *blkg = bio->bi_blkg;
1681 	struct ioc *ioc = rqos_to_ioc(rqos);
1682 	struct ioc_gq *iocg = blkg_to_iocg(blkg);
1683 	struct ioc_now now;
1684 	struct iocg_wait wait;
1685 	u32 hw_active, hw_inuse;
1686 	u64 abs_cost, cost, vtime;
1687 
1688 	/* bypass IOs if disabled or for root cgroup */
1689 	if (!ioc->enabled || !iocg->level)
1690 		return;
1691 
1692 	/* always activate so that even 0 cost IOs get protected to some level */
1693 	if (!iocg_activate(iocg, &now))
1694 		return;
1695 
1696 	/* calculate the absolute vtime cost */
1697 	abs_cost = calc_vtime_cost(bio, iocg, false);
1698 	if (!abs_cost)
1699 		return;
1700 
1701 	iocg->cursor = bio_end_sector(bio);
1702 
1703 	vtime = atomic64_read(&iocg->vtime);
1704 	current_hweight(iocg, &hw_active, &hw_inuse);
1705 
1706 	if (hw_inuse < hw_active &&
1707 	    time_after_eq64(vtime + ioc->inuse_margin_vtime, now.vnow)) {
1708 		TRACE_IOCG_PATH(inuse_reset, iocg, &now,
1709 				iocg->inuse, iocg->weight, hw_inuse, hw_active);
1710 		spin_lock_irq(&ioc->lock);
1711 		propagate_active_weight(iocg, iocg->weight, iocg->weight);
1712 		spin_unlock_irq(&ioc->lock);
1713 		current_hweight(iocg, &hw_active, &hw_inuse);
1714 	}
1715 
1716 	cost = abs_cost_to_cost(abs_cost, hw_inuse);
1717 
1718 	/*
1719 	 * If no one's waiting and within budget, issue right away.  The
1720 	 * tests are racy but the races aren't systemic - we only miss once
1721 	 * in a while which is fine.
1722 	 */
1723 	if (!waitqueue_active(&iocg->waitq) &&
1724 	    !atomic64_read(&iocg->abs_vdebt) &&
1725 	    time_before_eq64(vtime + cost, now.vnow)) {
1726 		iocg_commit_bio(iocg, bio, cost);
1727 		return;
1728 	}
1729 
1730 	/*
1731 	 * We're over budget.  If @bio has to be issued regardless,
1732 	 * remember the abs_cost instead of advancing vtime.
1733 	 * iocg_kick_waitq() will pay off the debt before waking more IOs.
1734 	 * This way, the debt is continuously paid off each period with the
1735 	 * actual budget available to the cgroup.  If we just wound vtime,
1736 	 * we would incorrectly use the current hw_inuse for the entire
1737 	 * amount which, for example, can lead to the cgroup staying
1738 	 * blocked for a long time even with substantially raised hw_inuse.
1739 	 */
1740 	if (bio_issue_as_root_blkg(bio) || fatal_signal_pending(current)) {
1741 		atomic64_add(abs_cost, &iocg->abs_vdebt);
1742 		iocg_kick_delay(iocg, &now, cost);
1743 		return;
1744 	}
1745 
1746 	/*
1747 	 * Append self to the waitq and schedule the wakeup timer if we're
1748 	 * the first waiter.  The timer duration is calculated based on the
1749 	 * current vrate.  vtime and hweight changes can make it too short
1750 	 * or too long.  Each wait entry records the absolute cost it's
1751 	 * waiting for to allow re-evaluation using a custom wait entry.
1752 	 *
1753 	 * If too short, the timer simply reschedules itself.  If too long,
1754 	 * the period timer will notice and trigger wakeups.
1755 	 *
1756 	 * All waiters are on iocg->waitq and the wait states are
1757 	 * synchronized using waitq.lock.
1758 	 */
1759 	spin_lock_irq(&iocg->waitq.lock);
1760 
1761 	/*
1762 	 * We activated above but w/o any synchronization.  Deactivation is
1763 	 * synchronized with waitq.lock and we won't get deactivated as
1764 	 * long as we're waiting, so we're good if we're activated here.
1765 	 * In the unlikely case that we are deactivated, just issue the IO.
1766 	 */
1767 	if (unlikely(list_empty(&iocg->active_list))) {
1768 		spin_unlock_irq(&iocg->waitq.lock);
1769 		iocg_commit_bio(iocg, bio, cost);
1770 		return;
1771 	}
1772 
1773 	init_waitqueue_func_entry(&wait.wait, iocg_wake_fn);
1774 	wait.wait.private = current;
1775 	wait.bio = bio;
1776 	wait.abs_cost = abs_cost;
1777 	wait.committed = false;	/* will be set true by waker */
1778 
1779 	__add_wait_queue_entry_tail(&iocg->waitq, &wait.wait);
1780 	iocg_kick_waitq(iocg, &now);
1781 
1782 	spin_unlock_irq(&iocg->waitq.lock);
1783 
1784 	while (true) {
1785 		set_current_state(TASK_UNINTERRUPTIBLE);
1786 		if (wait.committed)
1787 			break;
1788 		io_schedule();
1789 	}
1790 
1791 	/* waker already committed us, proceed */
1792 	finish_wait(&iocg->waitq, &wait.wait);
1793 }
1794 
1795 static void ioc_rqos_merge(struct rq_qos *rqos, struct request *rq,
1796 			   struct bio *bio)
1797 {
1798 	struct ioc_gq *iocg = blkg_to_iocg(bio->bi_blkg);
1799 	struct ioc *ioc = iocg->ioc;
1800 	sector_t bio_end = bio_end_sector(bio);
1801 	struct ioc_now now;
1802 	u32 hw_inuse;
1803 	u64 abs_cost, cost;
1804 
1805 	/* bypass if disabled or for root cgroup */
1806 	if (!ioc->enabled || !iocg->level)
1807 		return;
1808 
1809 	abs_cost = calc_vtime_cost(bio, iocg, true);
1810 	if (!abs_cost)
1811 		return;
1812 
1813 	ioc_now(ioc, &now);
1814 	current_hweight(iocg, NULL, &hw_inuse);
1815 	cost = abs_cost_to_cost(abs_cost, hw_inuse);
1816 
1817 	/* update cursor if backmerging into the request at the cursor */
1818 	if (blk_rq_pos(rq) < bio_end &&
1819 	    blk_rq_pos(rq) + blk_rq_sectors(rq) == iocg->cursor)
1820 		iocg->cursor = bio_end;
1821 
1822 	/*
1823 	 * Charge if there's enough vtime budget and the existing request
1824 	 * has cost assigned.  Otherwise, account it as debt.  See debt
1825 	 * handling in ioc_rqos_throttle() for details.
1826 	 */
1827 	if (rq->bio && rq->bio->bi_iocost_cost &&
1828 	    time_before_eq64(atomic64_read(&iocg->vtime) + cost, now.vnow))
1829 		iocg_commit_bio(iocg, bio, cost);
1830 	else
1831 		atomic64_add(abs_cost, &iocg->abs_vdebt);
1832 }
1833 
1834 static void ioc_rqos_done_bio(struct rq_qos *rqos, struct bio *bio)
1835 {
1836 	struct ioc_gq *iocg = blkg_to_iocg(bio->bi_blkg);
1837 
1838 	if (iocg && bio->bi_iocost_cost)
1839 		atomic64_add(bio->bi_iocost_cost, &iocg->done_vtime);
1840 }
1841 
1842 static void ioc_rqos_done(struct rq_qos *rqos, struct request *rq)
1843 {
1844 	struct ioc *ioc = rqos_to_ioc(rqos);
1845 	u64 on_q_ns, rq_wait_ns;
1846 	int pidx, rw;
1847 
1848 	if (!ioc->enabled || !rq->alloc_time_ns || !rq->start_time_ns)
1849 		return;
1850 
1851 	switch (req_op(rq) & REQ_OP_MASK) {
1852 	case REQ_OP_READ:
1853 		pidx = QOS_RLAT;
1854 		rw = READ;
1855 		break;
1856 	case REQ_OP_WRITE:
1857 		pidx = QOS_WLAT;
1858 		rw = WRITE;
1859 		break;
1860 	default:
1861 		return;
1862 	}
1863 
1864 	on_q_ns = ktime_get_ns() - rq->alloc_time_ns;
1865 	rq_wait_ns = rq->start_time_ns - rq->alloc_time_ns;
1866 
1867 	if (on_q_ns <= ioc->params.qos[pidx] * NSEC_PER_USEC)
1868 		this_cpu_inc(ioc->pcpu_stat->missed[rw].nr_met);
1869 	else
1870 		this_cpu_inc(ioc->pcpu_stat->missed[rw].nr_missed);
1871 
1872 	this_cpu_add(ioc->pcpu_stat->rq_wait_ns, rq_wait_ns);
1873 }
1874 
1875 static void ioc_rqos_queue_depth_changed(struct rq_qos *rqos)
1876 {
1877 	struct ioc *ioc = rqos_to_ioc(rqos);
1878 
1879 	spin_lock_irq(&ioc->lock);
1880 	ioc_refresh_params(ioc, false);
1881 	spin_unlock_irq(&ioc->lock);
1882 }
1883 
1884 static void ioc_rqos_exit(struct rq_qos *rqos)
1885 {
1886 	struct ioc *ioc = rqos_to_ioc(rqos);
1887 
1888 	blkcg_deactivate_policy(rqos->q, &blkcg_policy_iocost);
1889 
1890 	spin_lock_irq(&ioc->lock);
1891 	ioc->running = IOC_STOP;
1892 	spin_unlock_irq(&ioc->lock);
1893 
1894 	del_timer_sync(&ioc->timer);
1895 	free_percpu(ioc->pcpu_stat);
1896 	kfree(ioc);
1897 }
1898 
1899 static struct rq_qos_ops ioc_rqos_ops = {
1900 	.throttle = ioc_rqos_throttle,
1901 	.merge = ioc_rqos_merge,
1902 	.done_bio = ioc_rqos_done_bio,
1903 	.done = ioc_rqos_done,
1904 	.queue_depth_changed = ioc_rqos_queue_depth_changed,
1905 	.exit = ioc_rqos_exit,
1906 };
1907 
1908 static int blk_iocost_init(struct request_queue *q)
1909 {
1910 	struct ioc *ioc;
1911 	struct rq_qos *rqos;
1912 	int ret;
1913 
1914 	ioc = kzalloc(sizeof(*ioc), GFP_KERNEL);
1915 	if (!ioc)
1916 		return -ENOMEM;
1917 
1918 	ioc->pcpu_stat = alloc_percpu(struct ioc_pcpu_stat);
1919 	if (!ioc->pcpu_stat) {
1920 		kfree(ioc);
1921 		return -ENOMEM;
1922 	}
1923 
1924 	rqos = &ioc->rqos;
1925 	rqos->id = RQ_QOS_COST;
1926 	rqos->ops = &ioc_rqos_ops;
1927 	rqos->q = q;
1928 
1929 	spin_lock_init(&ioc->lock);
1930 	timer_setup(&ioc->timer, ioc_timer_fn, 0);
1931 	INIT_LIST_HEAD(&ioc->active_iocgs);
1932 
1933 	ioc->running = IOC_IDLE;
1934 	atomic64_set(&ioc->vtime_rate, VTIME_PER_USEC);
1935 	seqcount_init(&ioc->period_seqcount);
1936 	ioc->period_at = ktime_to_us(ktime_get());
1937 	atomic64_set(&ioc->cur_period, 0);
1938 	atomic_set(&ioc->hweight_gen, 0);
1939 
1940 	spin_lock_irq(&ioc->lock);
1941 	ioc->autop_idx = AUTOP_INVALID;
1942 	ioc_refresh_params(ioc, true);
1943 	spin_unlock_irq(&ioc->lock);
1944 
1945 	rq_qos_add(q, rqos);
1946 	ret = blkcg_activate_policy(q, &blkcg_policy_iocost);
1947 	if (ret) {
1948 		rq_qos_del(q, rqos);
1949 		free_percpu(ioc->pcpu_stat);
1950 		kfree(ioc);
1951 		return ret;
1952 	}
1953 	return 0;
1954 }
1955 
1956 static struct blkcg_policy_data *ioc_cpd_alloc(gfp_t gfp)
1957 {
1958 	struct ioc_cgrp *iocc;
1959 
1960 	iocc = kzalloc(sizeof(struct ioc_cgrp), gfp);
1961 	if (!iocc)
1962 		return NULL;
1963 
1964 	iocc->dfl_weight = CGROUP_WEIGHT_DFL;
1965 	return &iocc->cpd;
1966 }
1967 
1968 static void ioc_cpd_free(struct blkcg_policy_data *cpd)
1969 {
1970 	kfree(container_of(cpd, struct ioc_cgrp, cpd));
1971 }
1972 
1973 static struct blkg_policy_data *ioc_pd_alloc(gfp_t gfp, struct request_queue *q,
1974 					     struct blkcg *blkcg)
1975 {
1976 	int levels = blkcg->css.cgroup->level + 1;
1977 	struct ioc_gq *iocg;
1978 
1979 	iocg = kzalloc_node(sizeof(*iocg) + levels * sizeof(iocg->ancestors[0]),
1980 			    gfp, q->node);
1981 	if (!iocg)
1982 		return NULL;
1983 
1984 	return &iocg->pd;
1985 }
1986 
1987 static void ioc_pd_init(struct blkg_policy_data *pd)
1988 {
1989 	struct ioc_gq *iocg = pd_to_iocg(pd);
1990 	struct blkcg_gq *blkg = pd_to_blkg(&iocg->pd);
1991 	struct ioc *ioc = q_to_ioc(blkg->q);
1992 	struct ioc_now now;
1993 	struct blkcg_gq *tblkg;
1994 	unsigned long flags;
1995 
1996 	ioc_now(ioc, &now);
1997 
1998 	iocg->ioc = ioc;
1999 	atomic64_set(&iocg->vtime, now.vnow);
2000 	atomic64_set(&iocg->done_vtime, now.vnow);
2001 	atomic64_set(&iocg->abs_vdebt, 0);
2002 	atomic64_set(&iocg->active_period, atomic64_read(&ioc->cur_period));
2003 	INIT_LIST_HEAD(&iocg->active_list);
2004 	iocg->hweight_active = HWEIGHT_WHOLE;
2005 	iocg->hweight_inuse = HWEIGHT_WHOLE;
2006 
2007 	init_waitqueue_head(&iocg->waitq);
2008 	hrtimer_init(&iocg->waitq_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
2009 	iocg->waitq_timer.function = iocg_waitq_timer_fn;
2010 	hrtimer_init(&iocg->delay_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
2011 	iocg->delay_timer.function = iocg_delay_timer_fn;
2012 
2013 	iocg->level = blkg->blkcg->css.cgroup->level;
2014 
2015 	for (tblkg = blkg; tblkg; tblkg = tblkg->parent) {
2016 		struct ioc_gq *tiocg = blkg_to_iocg(tblkg);
2017 		iocg->ancestors[tiocg->level] = tiocg;
2018 	}
2019 
2020 	spin_lock_irqsave(&ioc->lock, flags);
2021 	weight_updated(iocg);
2022 	spin_unlock_irqrestore(&ioc->lock, flags);
2023 }
2024 
2025 static void ioc_pd_free(struct blkg_policy_data *pd)
2026 {
2027 	struct ioc_gq *iocg = pd_to_iocg(pd);
2028 	struct ioc *ioc = iocg->ioc;
2029 
2030 	if (ioc) {
2031 		spin_lock(&ioc->lock);
2032 		if (!list_empty(&iocg->active_list)) {
2033 			propagate_active_weight(iocg, 0, 0);
2034 			list_del_init(&iocg->active_list);
2035 		}
2036 		spin_unlock(&ioc->lock);
2037 
2038 		hrtimer_cancel(&iocg->waitq_timer);
2039 		hrtimer_cancel(&iocg->delay_timer);
2040 	}
2041 	kfree(iocg);
2042 }
2043 
2044 static u64 ioc_weight_prfill(struct seq_file *sf, struct blkg_policy_data *pd,
2045 			     int off)
2046 {
2047 	const char *dname = blkg_dev_name(pd->blkg);
2048 	struct ioc_gq *iocg = pd_to_iocg(pd);
2049 
2050 	if (dname && iocg->cfg_weight)
2051 		seq_printf(sf, "%s %u\n", dname, iocg->cfg_weight);
2052 	return 0;
2053 }
2054 
2055 
2056 static int ioc_weight_show(struct seq_file *sf, void *v)
2057 {
2058 	struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
2059 	struct ioc_cgrp *iocc = blkcg_to_iocc(blkcg);
2060 
2061 	seq_printf(sf, "default %u\n", iocc->dfl_weight);
2062 	blkcg_print_blkgs(sf, blkcg, ioc_weight_prfill,
2063 			  &blkcg_policy_iocost, seq_cft(sf)->private, false);
2064 	return 0;
2065 }
2066 
2067 static ssize_t ioc_weight_write(struct kernfs_open_file *of, char *buf,
2068 				size_t nbytes, loff_t off)
2069 {
2070 	struct blkcg *blkcg = css_to_blkcg(of_css(of));
2071 	struct ioc_cgrp *iocc = blkcg_to_iocc(blkcg);
2072 	struct blkg_conf_ctx ctx;
2073 	struct ioc_gq *iocg;
2074 	u32 v;
2075 	int ret;
2076 
2077 	if (!strchr(buf, ':')) {
2078 		struct blkcg_gq *blkg;
2079 
2080 		if (!sscanf(buf, "default %u", &v) && !sscanf(buf, "%u", &v))
2081 			return -EINVAL;
2082 
2083 		if (v < CGROUP_WEIGHT_MIN || v > CGROUP_WEIGHT_MAX)
2084 			return -EINVAL;
2085 
2086 		spin_lock(&blkcg->lock);
2087 		iocc->dfl_weight = v;
2088 		hlist_for_each_entry(blkg, &blkcg->blkg_list, blkcg_node) {
2089 			struct ioc_gq *iocg = blkg_to_iocg(blkg);
2090 
2091 			if (iocg) {
2092 				spin_lock_irq(&iocg->ioc->lock);
2093 				weight_updated(iocg);
2094 				spin_unlock_irq(&iocg->ioc->lock);
2095 			}
2096 		}
2097 		spin_unlock(&blkcg->lock);
2098 
2099 		return nbytes;
2100 	}
2101 
2102 	ret = blkg_conf_prep(blkcg, &blkcg_policy_iocost, buf, &ctx);
2103 	if (ret)
2104 		return ret;
2105 
2106 	iocg = blkg_to_iocg(ctx.blkg);
2107 
2108 	if (!strncmp(ctx.body, "default", 7)) {
2109 		v = 0;
2110 	} else {
2111 		if (!sscanf(ctx.body, "%u", &v))
2112 			goto einval;
2113 		if (v < CGROUP_WEIGHT_MIN || v > CGROUP_WEIGHT_MAX)
2114 			goto einval;
2115 	}
2116 
2117 	spin_lock(&iocg->ioc->lock);
2118 	iocg->cfg_weight = v;
2119 	weight_updated(iocg);
2120 	spin_unlock(&iocg->ioc->lock);
2121 
2122 	blkg_conf_finish(&ctx);
2123 	return nbytes;
2124 
2125 einval:
2126 	blkg_conf_finish(&ctx);
2127 	return -EINVAL;
2128 }
2129 
2130 static u64 ioc_qos_prfill(struct seq_file *sf, struct blkg_policy_data *pd,
2131 			  int off)
2132 {
2133 	const char *dname = blkg_dev_name(pd->blkg);
2134 	struct ioc *ioc = pd_to_iocg(pd)->ioc;
2135 
2136 	if (!dname)
2137 		return 0;
2138 
2139 	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",
2140 		   dname, ioc->enabled, ioc->user_qos_params ? "user" : "auto",
2141 		   ioc->params.qos[QOS_RPPM] / 10000,
2142 		   ioc->params.qos[QOS_RPPM] % 10000 / 100,
2143 		   ioc->params.qos[QOS_RLAT],
2144 		   ioc->params.qos[QOS_WPPM] / 10000,
2145 		   ioc->params.qos[QOS_WPPM] % 10000 / 100,
2146 		   ioc->params.qos[QOS_WLAT],
2147 		   ioc->params.qos[QOS_MIN] / 10000,
2148 		   ioc->params.qos[QOS_MIN] % 10000 / 100,
2149 		   ioc->params.qos[QOS_MAX] / 10000,
2150 		   ioc->params.qos[QOS_MAX] % 10000 / 100);
2151 	return 0;
2152 }
2153 
2154 static int ioc_qos_show(struct seq_file *sf, void *v)
2155 {
2156 	struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
2157 
2158 	blkcg_print_blkgs(sf, blkcg, ioc_qos_prfill,
2159 			  &blkcg_policy_iocost, seq_cft(sf)->private, false);
2160 	return 0;
2161 }
2162 
2163 static const match_table_t qos_ctrl_tokens = {
2164 	{ QOS_ENABLE,		"enable=%u"	},
2165 	{ QOS_CTRL,		"ctrl=%s"	},
2166 	{ NR_QOS_CTRL_PARAMS,	NULL		},
2167 };
2168 
2169 static const match_table_t qos_tokens = {
2170 	{ QOS_RPPM,		"rpct=%s"	},
2171 	{ QOS_RLAT,		"rlat=%u"	},
2172 	{ QOS_WPPM,		"wpct=%s"	},
2173 	{ QOS_WLAT,		"wlat=%u"	},
2174 	{ QOS_MIN,		"min=%s"	},
2175 	{ QOS_MAX,		"max=%s"	},
2176 	{ NR_QOS_PARAMS,	NULL		},
2177 };
2178 
2179 static ssize_t ioc_qos_write(struct kernfs_open_file *of, char *input,
2180 			     size_t nbytes, loff_t off)
2181 {
2182 	struct gendisk *disk;
2183 	struct ioc *ioc;
2184 	u32 qos[NR_QOS_PARAMS];
2185 	bool enable, user;
2186 	char *p;
2187 	int ret;
2188 
2189 	disk = blkcg_conf_get_disk(&input);
2190 	if (IS_ERR(disk))
2191 		return PTR_ERR(disk);
2192 
2193 	ioc = q_to_ioc(disk->queue);
2194 	if (!ioc) {
2195 		ret = blk_iocost_init(disk->queue);
2196 		if (ret)
2197 			goto err;
2198 		ioc = q_to_ioc(disk->queue);
2199 	}
2200 
2201 	spin_lock_irq(&ioc->lock);
2202 	memcpy(qos, ioc->params.qos, sizeof(qos));
2203 	enable = ioc->enabled;
2204 	user = ioc->user_qos_params;
2205 	spin_unlock_irq(&ioc->lock);
2206 
2207 	while ((p = strsep(&input, " \t\n"))) {
2208 		substring_t args[MAX_OPT_ARGS];
2209 		char buf[32];
2210 		int tok;
2211 		s64 v;
2212 
2213 		if (!*p)
2214 			continue;
2215 
2216 		switch (match_token(p, qos_ctrl_tokens, args)) {
2217 		case QOS_ENABLE:
2218 			match_u64(&args[0], &v);
2219 			enable = v;
2220 			continue;
2221 		case QOS_CTRL:
2222 			match_strlcpy(buf, &args[0], sizeof(buf));
2223 			if (!strcmp(buf, "auto"))
2224 				user = false;
2225 			else if (!strcmp(buf, "user"))
2226 				user = true;
2227 			else
2228 				goto einval;
2229 			continue;
2230 		}
2231 
2232 		tok = match_token(p, qos_tokens, args);
2233 		switch (tok) {
2234 		case QOS_RPPM:
2235 		case QOS_WPPM:
2236 			if (match_strlcpy(buf, &args[0], sizeof(buf)) >=
2237 			    sizeof(buf))
2238 				goto einval;
2239 			if (cgroup_parse_float(buf, 2, &v))
2240 				goto einval;
2241 			if (v < 0 || v > 10000)
2242 				goto einval;
2243 			qos[tok] = v * 100;
2244 			break;
2245 		case QOS_RLAT:
2246 		case QOS_WLAT:
2247 			if (match_u64(&args[0], &v))
2248 				goto einval;
2249 			qos[tok] = v;
2250 			break;
2251 		case QOS_MIN:
2252 		case QOS_MAX:
2253 			if (match_strlcpy(buf, &args[0], sizeof(buf)) >=
2254 			    sizeof(buf))
2255 				goto einval;
2256 			if (cgroup_parse_float(buf, 2, &v))
2257 				goto einval;
2258 			if (v < 0)
2259 				goto einval;
2260 			qos[tok] = clamp_t(s64, v * 100,
2261 					   VRATE_MIN_PPM, VRATE_MAX_PPM);
2262 			break;
2263 		default:
2264 			goto einval;
2265 		}
2266 		user = true;
2267 	}
2268 
2269 	if (qos[QOS_MIN] > qos[QOS_MAX])
2270 		goto einval;
2271 
2272 	spin_lock_irq(&ioc->lock);
2273 
2274 	if (enable) {
2275 		blk_queue_flag_set(QUEUE_FLAG_RQ_ALLOC_TIME, ioc->rqos.q);
2276 		ioc->enabled = true;
2277 	} else {
2278 		blk_queue_flag_clear(QUEUE_FLAG_RQ_ALLOC_TIME, ioc->rqos.q);
2279 		ioc->enabled = false;
2280 	}
2281 
2282 	if (user) {
2283 		memcpy(ioc->params.qos, qos, sizeof(qos));
2284 		ioc->user_qos_params = true;
2285 	} else {
2286 		ioc->user_qos_params = false;
2287 	}
2288 
2289 	ioc_refresh_params(ioc, true);
2290 	spin_unlock_irq(&ioc->lock);
2291 
2292 	put_disk_and_module(disk);
2293 	return nbytes;
2294 einval:
2295 	ret = -EINVAL;
2296 err:
2297 	put_disk_and_module(disk);
2298 	return ret;
2299 }
2300 
2301 static u64 ioc_cost_model_prfill(struct seq_file *sf,
2302 				 struct blkg_policy_data *pd, int off)
2303 {
2304 	const char *dname = blkg_dev_name(pd->blkg);
2305 	struct ioc *ioc = pd_to_iocg(pd)->ioc;
2306 	u64 *u = ioc->params.i_lcoefs;
2307 
2308 	if (!dname)
2309 		return 0;
2310 
2311 	seq_printf(sf, "%s ctrl=%s model=linear "
2312 		   "rbps=%llu rseqiops=%llu rrandiops=%llu "
2313 		   "wbps=%llu wseqiops=%llu wrandiops=%llu\n",
2314 		   dname, ioc->user_cost_model ? "user" : "auto",
2315 		   u[I_LCOEF_RBPS], u[I_LCOEF_RSEQIOPS], u[I_LCOEF_RRANDIOPS],
2316 		   u[I_LCOEF_WBPS], u[I_LCOEF_WSEQIOPS], u[I_LCOEF_WRANDIOPS]);
2317 	return 0;
2318 }
2319 
2320 static int ioc_cost_model_show(struct seq_file *sf, void *v)
2321 {
2322 	struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
2323 
2324 	blkcg_print_blkgs(sf, blkcg, ioc_cost_model_prfill,
2325 			  &blkcg_policy_iocost, seq_cft(sf)->private, false);
2326 	return 0;
2327 }
2328 
2329 static const match_table_t cost_ctrl_tokens = {
2330 	{ COST_CTRL,		"ctrl=%s"	},
2331 	{ COST_MODEL,		"model=%s"	},
2332 	{ NR_COST_CTRL_PARAMS,	NULL		},
2333 };
2334 
2335 static const match_table_t i_lcoef_tokens = {
2336 	{ I_LCOEF_RBPS,		"rbps=%u"	},
2337 	{ I_LCOEF_RSEQIOPS,	"rseqiops=%u"	},
2338 	{ I_LCOEF_RRANDIOPS,	"rrandiops=%u"	},
2339 	{ I_LCOEF_WBPS,		"wbps=%u"	},
2340 	{ I_LCOEF_WSEQIOPS,	"wseqiops=%u"	},
2341 	{ I_LCOEF_WRANDIOPS,	"wrandiops=%u"	},
2342 	{ NR_I_LCOEFS,		NULL		},
2343 };
2344 
2345 static ssize_t ioc_cost_model_write(struct kernfs_open_file *of, char *input,
2346 				    size_t nbytes, loff_t off)
2347 {
2348 	struct gendisk *disk;
2349 	struct ioc *ioc;
2350 	u64 u[NR_I_LCOEFS];
2351 	bool user;
2352 	char *p;
2353 	int ret;
2354 
2355 	disk = blkcg_conf_get_disk(&input);
2356 	if (IS_ERR(disk))
2357 		return PTR_ERR(disk);
2358 
2359 	ioc = q_to_ioc(disk->queue);
2360 	if (!ioc) {
2361 		ret = blk_iocost_init(disk->queue);
2362 		if (ret)
2363 			goto err;
2364 		ioc = q_to_ioc(disk->queue);
2365 	}
2366 
2367 	spin_lock_irq(&ioc->lock);
2368 	memcpy(u, ioc->params.i_lcoefs, sizeof(u));
2369 	user = ioc->user_cost_model;
2370 	spin_unlock_irq(&ioc->lock);
2371 
2372 	while ((p = strsep(&input, " \t\n"))) {
2373 		substring_t args[MAX_OPT_ARGS];
2374 		char buf[32];
2375 		int tok;
2376 		u64 v;
2377 
2378 		if (!*p)
2379 			continue;
2380 
2381 		switch (match_token(p, cost_ctrl_tokens, args)) {
2382 		case COST_CTRL:
2383 			match_strlcpy(buf, &args[0], sizeof(buf));
2384 			if (!strcmp(buf, "auto"))
2385 				user = false;
2386 			else if (!strcmp(buf, "user"))
2387 				user = true;
2388 			else
2389 				goto einval;
2390 			continue;
2391 		case COST_MODEL:
2392 			match_strlcpy(buf, &args[0], sizeof(buf));
2393 			if (strcmp(buf, "linear"))
2394 				goto einval;
2395 			continue;
2396 		}
2397 
2398 		tok = match_token(p, i_lcoef_tokens, args);
2399 		if (tok == NR_I_LCOEFS)
2400 			goto einval;
2401 		if (match_u64(&args[0], &v))
2402 			goto einval;
2403 		u[tok] = v;
2404 		user = true;
2405 	}
2406 
2407 	spin_lock_irq(&ioc->lock);
2408 	if (user) {
2409 		memcpy(ioc->params.i_lcoefs, u, sizeof(u));
2410 		ioc->user_cost_model = true;
2411 	} else {
2412 		ioc->user_cost_model = false;
2413 	}
2414 	ioc_refresh_params(ioc, true);
2415 	spin_unlock_irq(&ioc->lock);
2416 
2417 	put_disk_and_module(disk);
2418 	return nbytes;
2419 
2420 einval:
2421 	ret = -EINVAL;
2422 err:
2423 	put_disk_and_module(disk);
2424 	return ret;
2425 }
2426 
2427 static struct cftype ioc_files[] = {
2428 	{
2429 		.name = "weight",
2430 		.flags = CFTYPE_NOT_ON_ROOT,
2431 		.seq_show = ioc_weight_show,
2432 		.write = ioc_weight_write,
2433 	},
2434 	{
2435 		.name = "cost.qos",
2436 		.flags = CFTYPE_ONLY_ON_ROOT,
2437 		.seq_show = ioc_qos_show,
2438 		.write = ioc_qos_write,
2439 	},
2440 	{
2441 		.name = "cost.model",
2442 		.flags = CFTYPE_ONLY_ON_ROOT,
2443 		.seq_show = ioc_cost_model_show,
2444 		.write = ioc_cost_model_write,
2445 	},
2446 	{}
2447 };
2448 
2449 static struct blkcg_policy blkcg_policy_iocost = {
2450 	.dfl_cftypes	= ioc_files,
2451 	.cpd_alloc_fn	= ioc_cpd_alloc,
2452 	.cpd_free_fn	= ioc_cpd_free,
2453 	.pd_alloc_fn	= ioc_pd_alloc,
2454 	.pd_init_fn	= ioc_pd_init,
2455 	.pd_free_fn	= ioc_pd_free,
2456 };
2457 
2458 static int __init ioc_init(void)
2459 {
2460 	return blkcg_policy_register(&blkcg_policy_iocost);
2461 }
2462 
2463 static void __exit ioc_exit(void)
2464 {
2465 	return blkcg_policy_unregister(&blkcg_policy_iocost);
2466 }
2467 
2468 module_init(ioc_init);
2469 module_exit(ioc_exit);
2470