xref: /openbmc/linux/block/blk-iocost.c (revision ecd25094)
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]		=         50000, /* 50ms */
533 			[QOS_WLAT]		=         50000,
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 	for (i = iocg->level; i > 0; i--)
1061 		if (!list_empty(&iocg->active_list))
1062 			goto fail_unlock;
1063 	if (iocg->child_active_sum)
1064 		goto fail_unlock;
1065 
1066 	/*
1067 	 * vtime may wrap when vrate is raised substantially due to
1068 	 * underestimated IO costs.  Look at the period and ignore its
1069 	 * vtime if the iocg has been idle for too long.  Also, cap the
1070 	 * budget it can start with to the margin.
1071 	 */
1072 	max_period_delta = DIV64_U64_ROUND_UP(VTIME_VALID_DUR, ioc->period_us);
1073 	vtime = atomic64_read(&iocg->vtime);
1074 	vmargin = ioc->margin_us * now->vrate;
1075 	vmin = now->vnow - vmargin;
1076 
1077 	if (last_period + max_period_delta < cur_period ||
1078 	    time_before64(vtime, vmin)) {
1079 		atomic64_add(vmin - vtime, &iocg->vtime);
1080 		atomic64_add(vmin - vtime, &iocg->done_vtime);
1081 		vtime = vmin;
1082 	}
1083 
1084 	/*
1085 	 * Activate, propagate weight and start period timer if not
1086 	 * running.  Reset hweight_gen to avoid accidental match from
1087 	 * wrapping.
1088 	 */
1089 	iocg->hweight_gen = atomic_read(&ioc->hweight_gen) - 1;
1090 	list_add(&iocg->active_list, &ioc->active_iocgs);
1091 	propagate_active_weight(iocg, iocg->weight,
1092 				iocg->last_inuse ?: iocg->weight);
1093 
1094 	TRACE_IOCG_PATH(iocg_activate, iocg, now,
1095 			last_period, cur_period, vtime);
1096 
1097 	iocg->last_vtime = vtime;
1098 
1099 	if (ioc->running == IOC_IDLE) {
1100 		ioc->running = IOC_RUNNING;
1101 		ioc_start_period(ioc, now);
1102 	}
1103 
1104 	spin_unlock_irq(&ioc->lock);
1105 	return true;
1106 
1107 fail_unlock:
1108 	spin_unlock_irq(&ioc->lock);
1109 	return false;
1110 }
1111 
1112 static int iocg_wake_fn(struct wait_queue_entry *wq_entry, unsigned mode,
1113 			int flags, void *key)
1114 {
1115 	struct iocg_wait *wait = container_of(wq_entry, struct iocg_wait, wait);
1116 	struct iocg_wake_ctx *ctx = (struct iocg_wake_ctx *)key;
1117 	u64 cost = abs_cost_to_cost(wait->abs_cost, ctx->hw_inuse);
1118 
1119 	ctx->vbudget -= cost;
1120 
1121 	if (ctx->vbudget < 0)
1122 		return -1;
1123 
1124 	iocg_commit_bio(ctx->iocg, wait->bio, cost);
1125 
1126 	/*
1127 	 * autoremove_wake_function() removes the wait entry only when it
1128 	 * actually changed the task state.  We want the wait always
1129 	 * removed.  Remove explicitly and use default_wake_function().
1130 	 */
1131 	list_del_init(&wq_entry->entry);
1132 	wait->committed = true;
1133 
1134 	default_wake_function(wq_entry, mode, flags, key);
1135 	return 0;
1136 }
1137 
1138 static void iocg_kick_waitq(struct ioc_gq *iocg, struct ioc_now *now)
1139 {
1140 	struct ioc *ioc = iocg->ioc;
1141 	struct iocg_wake_ctx ctx = { .iocg = iocg };
1142 	u64 margin_ns = (u64)(ioc->period_us *
1143 			      WAITQ_TIMER_MARGIN_PCT / 100) * NSEC_PER_USEC;
1144 	u64 abs_vdebt, vdebt, vshortage, expires, oexpires;
1145 	s64 vbudget;
1146 	u32 hw_inuse;
1147 
1148 	lockdep_assert_held(&iocg->waitq.lock);
1149 
1150 	current_hweight(iocg, NULL, &hw_inuse);
1151 	vbudget = now->vnow - atomic64_read(&iocg->vtime);
1152 
1153 	/* pay off debt */
1154 	abs_vdebt = atomic64_read(&iocg->abs_vdebt);
1155 	vdebt = abs_cost_to_cost(abs_vdebt, hw_inuse);
1156 	if (vdebt && vbudget > 0) {
1157 		u64 delta = min_t(u64, vbudget, vdebt);
1158 		u64 abs_delta = min(cost_to_abs_cost(delta, hw_inuse),
1159 				    abs_vdebt);
1160 
1161 		atomic64_add(delta, &iocg->vtime);
1162 		atomic64_add(delta, &iocg->done_vtime);
1163 		atomic64_sub(abs_delta, &iocg->abs_vdebt);
1164 		if (WARN_ON_ONCE(atomic64_read(&iocg->abs_vdebt) < 0))
1165 			atomic64_set(&iocg->abs_vdebt, 0);
1166 	}
1167 
1168 	/*
1169 	 * Wake up the ones which are due and see how much vtime we'll need
1170 	 * for the next one.
1171 	 */
1172 	ctx.hw_inuse = hw_inuse;
1173 	ctx.vbudget = vbudget - vdebt;
1174 	__wake_up_locked_key(&iocg->waitq, TASK_NORMAL, &ctx);
1175 	if (!waitqueue_active(&iocg->waitq))
1176 		return;
1177 	if (WARN_ON_ONCE(ctx.vbudget >= 0))
1178 		return;
1179 
1180 	/* determine next wakeup, add a quarter margin to guarantee chunking */
1181 	vshortage = -ctx.vbudget;
1182 	expires = now->now_ns +
1183 		DIV64_U64_ROUND_UP(vshortage, now->vrate) * NSEC_PER_USEC;
1184 	expires += margin_ns / 4;
1185 
1186 	/* if already active and close enough, don't bother */
1187 	oexpires = ktime_to_ns(hrtimer_get_softexpires(&iocg->waitq_timer));
1188 	if (hrtimer_is_queued(&iocg->waitq_timer) &&
1189 	    abs(oexpires - expires) <= margin_ns / 4)
1190 		return;
1191 
1192 	hrtimer_start_range_ns(&iocg->waitq_timer, ns_to_ktime(expires),
1193 			       margin_ns / 4, HRTIMER_MODE_ABS);
1194 }
1195 
1196 static enum hrtimer_restart iocg_waitq_timer_fn(struct hrtimer *timer)
1197 {
1198 	struct ioc_gq *iocg = container_of(timer, struct ioc_gq, waitq_timer);
1199 	struct ioc_now now;
1200 	unsigned long flags;
1201 
1202 	ioc_now(iocg->ioc, &now);
1203 
1204 	spin_lock_irqsave(&iocg->waitq.lock, flags);
1205 	iocg_kick_waitq(iocg, &now);
1206 	spin_unlock_irqrestore(&iocg->waitq.lock, flags);
1207 
1208 	return HRTIMER_NORESTART;
1209 }
1210 
1211 static void iocg_kick_delay(struct ioc_gq *iocg, struct ioc_now *now, u64 cost)
1212 {
1213 	struct ioc *ioc = iocg->ioc;
1214 	struct blkcg_gq *blkg = iocg_to_blkg(iocg);
1215 	u64 vtime = atomic64_read(&iocg->vtime);
1216 	u64 vmargin = ioc->margin_us * now->vrate;
1217 	u64 margin_ns = ioc->margin_us * NSEC_PER_USEC;
1218 	u64 expires, oexpires;
1219 	u32 hw_inuse;
1220 
1221 	/* debt-adjust vtime */
1222 	current_hweight(iocg, NULL, &hw_inuse);
1223 	vtime += abs_cost_to_cost(atomic64_read(&iocg->abs_vdebt), hw_inuse);
1224 
1225 	/* clear or maintain depending on the overage */
1226 	if (time_before_eq64(vtime, now->vnow)) {
1227 		blkcg_clear_delay(blkg);
1228 		return;
1229 	}
1230 	if (!atomic_read(&blkg->use_delay) &&
1231 	    time_before_eq64(vtime, now->vnow + vmargin))
1232 		return;
1233 
1234 	/* use delay */
1235 	if (cost) {
1236 		u64 cost_ns = DIV64_U64_ROUND_UP(cost * NSEC_PER_USEC,
1237 						 now->vrate);
1238 		blkcg_add_delay(blkg, now->now_ns, cost_ns);
1239 	}
1240 	blkcg_use_delay(blkg);
1241 
1242 	expires = now->now_ns + DIV64_U64_ROUND_UP(vtime - now->vnow,
1243 						   now->vrate) * NSEC_PER_USEC;
1244 
1245 	/* if already active and close enough, don't bother */
1246 	oexpires = ktime_to_ns(hrtimer_get_softexpires(&iocg->delay_timer));
1247 	if (hrtimer_is_queued(&iocg->delay_timer) &&
1248 	    abs(oexpires - expires) <= margin_ns / 4)
1249 		return;
1250 
1251 	hrtimer_start_range_ns(&iocg->delay_timer, ns_to_ktime(expires),
1252 			       margin_ns / 4, HRTIMER_MODE_ABS);
1253 }
1254 
1255 static enum hrtimer_restart iocg_delay_timer_fn(struct hrtimer *timer)
1256 {
1257 	struct ioc_gq *iocg = container_of(timer, struct ioc_gq, delay_timer);
1258 	struct ioc_now now;
1259 
1260 	ioc_now(iocg->ioc, &now);
1261 	iocg_kick_delay(iocg, &now, 0);
1262 
1263 	return HRTIMER_NORESTART;
1264 }
1265 
1266 static void ioc_lat_stat(struct ioc *ioc, u32 *missed_ppm_ar, u32 *rq_wait_pct_p)
1267 {
1268 	u32 nr_met[2] = { };
1269 	u32 nr_missed[2] = { };
1270 	u64 rq_wait_ns = 0;
1271 	int cpu, rw;
1272 
1273 	for_each_online_cpu(cpu) {
1274 		struct ioc_pcpu_stat *stat = per_cpu_ptr(ioc->pcpu_stat, cpu);
1275 		u64 this_rq_wait_ns;
1276 
1277 		for (rw = READ; rw <= WRITE; rw++) {
1278 			u32 this_met = READ_ONCE(stat->missed[rw].nr_met);
1279 			u32 this_missed = READ_ONCE(stat->missed[rw].nr_missed);
1280 
1281 			nr_met[rw] += this_met - stat->missed[rw].last_met;
1282 			nr_missed[rw] += this_missed - stat->missed[rw].last_missed;
1283 			stat->missed[rw].last_met = this_met;
1284 			stat->missed[rw].last_missed = this_missed;
1285 		}
1286 
1287 		this_rq_wait_ns = READ_ONCE(stat->rq_wait_ns);
1288 		rq_wait_ns += this_rq_wait_ns - stat->last_rq_wait_ns;
1289 		stat->last_rq_wait_ns = this_rq_wait_ns;
1290 	}
1291 
1292 	for (rw = READ; rw <= WRITE; rw++) {
1293 		if (nr_met[rw] + nr_missed[rw])
1294 			missed_ppm_ar[rw] =
1295 				DIV64_U64_ROUND_UP((u64)nr_missed[rw] * MILLION,
1296 						   nr_met[rw] + nr_missed[rw]);
1297 		else
1298 			missed_ppm_ar[rw] = 0;
1299 	}
1300 
1301 	*rq_wait_pct_p = div64_u64(rq_wait_ns * 100,
1302 				   ioc->period_us * NSEC_PER_USEC);
1303 }
1304 
1305 /* was iocg idle this period? */
1306 static bool iocg_is_idle(struct ioc_gq *iocg)
1307 {
1308 	struct ioc *ioc = iocg->ioc;
1309 
1310 	/* did something get issued this period? */
1311 	if (atomic64_read(&iocg->active_period) ==
1312 	    atomic64_read(&ioc->cur_period))
1313 		return false;
1314 
1315 	/* is something in flight? */
1316 	if (atomic64_read(&iocg->done_vtime) < atomic64_read(&iocg->vtime))
1317 		return false;
1318 
1319 	return true;
1320 }
1321 
1322 /* returns usage with margin added if surplus is large enough */
1323 static u32 surplus_adjusted_hweight_inuse(u32 usage, u32 hw_inuse)
1324 {
1325 	/* add margin */
1326 	usage = DIV_ROUND_UP(usage * SURPLUS_SCALE_PCT, 100);
1327 	usage += SURPLUS_SCALE_ABS;
1328 
1329 	/* don't bother if the surplus is too small */
1330 	if (usage + SURPLUS_MIN_ADJ_DELTA > hw_inuse)
1331 		return 0;
1332 
1333 	return usage;
1334 }
1335 
1336 static void ioc_timer_fn(struct timer_list *timer)
1337 {
1338 	struct ioc *ioc = container_of(timer, struct ioc, timer);
1339 	struct ioc_gq *iocg, *tiocg;
1340 	struct ioc_now now;
1341 	int nr_surpluses = 0, nr_shortages = 0, nr_lagging = 0;
1342 	u32 ppm_rthr = MILLION - ioc->params.qos[QOS_RPPM];
1343 	u32 ppm_wthr = MILLION - ioc->params.qos[QOS_WPPM];
1344 	u32 missed_ppm[2], rq_wait_pct;
1345 	u64 period_vtime;
1346 	int i;
1347 
1348 	/* how were the latencies during the period? */
1349 	ioc_lat_stat(ioc, missed_ppm, &rq_wait_pct);
1350 
1351 	/* take care of active iocgs */
1352 	spin_lock_irq(&ioc->lock);
1353 
1354 	ioc_now(ioc, &now);
1355 
1356 	period_vtime = now.vnow - ioc->period_at_vtime;
1357 	if (WARN_ON_ONCE(!period_vtime)) {
1358 		spin_unlock_irq(&ioc->lock);
1359 		return;
1360 	}
1361 
1362 	/*
1363 	 * Waiters determine the sleep durations based on the vrate they
1364 	 * saw at the time of sleep.  If vrate has increased, some waiters
1365 	 * could be sleeping for too long.  Wake up tardy waiters which
1366 	 * should have woken up in the last period and expire idle iocgs.
1367 	 */
1368 	list_for_each_entry_safe(iocg, tiocg, &ioc->active_iocgs, active_list) {
1369 		if (!waitqueue_active(&iocg->waitq) &&
1370 		    !atomic64_read(&iocg->abs_vdebt) && !iocg_is_idle(iocg))
1371 			continue;
1372 
1373 		spin_lock(&iocg->waitq.lock);
1374 
1375 		if (waitqueue_active(&iocg->waitq) ||
1376 		    atomic64_read(&iocg->abs_vdebt)) {
1377 			/* might be oversleeping vtime / hweight changes, kick */
1378 			iocg_kick_waitq(iocg, &now);
1379 			iocg_kick_delay(iocg, &now, 0);
1380 		} else if (iocg_is_idle(iocg)) {
1381 			/* no waiter and idle, deactivate */
1382 			iocg->last_inuse = iocg->inuse;
1383 			__propagate_active_weight(iocg, 0, 0);
1384 			list_del_init(&iocg->active_list);
1385 		}
1386 
1387 		spin_unlock(&iocg->waitq.lock);
1388 	}
1389 	commit_active_weights(ioc);
1390 
1391 	/* calc usages and see whether some weights need to be moved around */
1392 	list_for_each_entry(iocg, &ioc->active_iocgs, active_list) {
1393 		u64 vdone, vtime, vusage, vmargin, vmin;
1394 		u32 hw_active, hw_inuse, usage;
1395 
1396 		/*
1397 		 * Collect unused and wind vtime closer to vnow to prevent
1398 		 * iocgs from accumulating a large amount of budget.
1399 		 */
1400 		vdone = atomic64_read(&iocg->done_vtime);
1401 		vtime = atomic64_read(&iocg->vtime);
1402 		current_hweight(iocg, &hw_active, &hw_inuse);
1403 
1404 		/*
1405 		 * Latency QoS detection doesn't account for IOs which are
1406 		 * in-flight for longer than a period.  Detect them by
1407 		 * comparing vdone against period start.  If lagging behind
1408 		 * IOs from past periods, don't increase vrate.
1409 		 */
1410 		if (!atomic_read(&iocg_to_blkg(iocg)->use_delay) &&
1411 		    time_after64(vtime, vdone) &&
1412 		    time_after64(vtime, now.vnow -
1413 				 MAX_LAGGING_PERIODS * period_vtime) &&
1414 		    time_before64(vdone, now.vnow - period_vtime))
1415 			nr_lagging++;
1416 
1417 		if (waitqueue_active(&iocg->waitq))
1418 			vusage = now.vnow - iocg->last_vtime;
1419 		else if (time_before64(iocg->last_vtime, vtime))
1420 			vusage = vtime - iocg->last_vtime;
1421 		else
1422 			vusage = 0;
1423 
1424 		iocg->last_vtime += vusage;
1425 		/*
1426 		 * Factor in in-flight vtime into vusage to avoid
1427 		 * high-latency completions appearing as idle.  This should
1428 		 * be done after the above ->last_time adjustment.
1429 		 */
1430 		vusage = max(vusage, vtime - vdone);
1431 
1432 		/* calculate hweight based usage ratio and record */
1433 		if (vusage) {
1434 			usage = DIV64_U64_ROUND_UP(vusage * hw_inuse,
1435 						   period_vtime);
1436 			iocg->usage_idx = (iocg->usage_idx + 1) % NR_USAGE_SLOTS;
1437 			iocg->usages[iocg->usage_idx] = usage;
1438 		} else {
1439 			usage = 0;
1440 		}
1441 
1442 		/* see whether there's surplus vtime */
1443 		vmargin = ioc->margin_us * now.vrate;
1444 		vmin = now.vnow - vmargin;
1445 
1446 		iocg->has_surplus = false;
1447 
1448 		if (!waitqueue_active(&iocg->waitq) &&
1449 		    time_before64(vtime, vmin)) {
1450 			u64 delta = vmin - vtime;
1451 
1452 			/* throw away surplus vtime */
1453 			atomic64_add(delta, &iocg->vtime);
1454 			atomic64_add(delta, &iocg->done_vtime);
1455 			iocg->last_vtime += delta;
1456 			/* if usage is sufficiently low, maybe it can donate */
1457 			if (surplus_adjusted_hweight_inuse(usage, hw_inuse)) {
1458 				iocg->has_surplus = true;
1459 				nr_surpluses++;
1460 			}
1461 		} else if (hw_inuse < hw_active) {
1462 			u32 new_hwi, new_inuse;
1463 
1464 			/* was donating but might need to take back some */
1465 			if (waitqueue_active(&iocg->waitq)) {
1466 				new_hwi = hw_active;
1467 			} else {
1468 				new_hwi = max(hw_inuse,
1469 					      usage * SURPLUS_SCALE_PCT / 100 +
1470 					      SURPLUS_SCALE_ABS);
1471 			}
1472 
1473 			new_inuse = div64_u64((u64)iocg->inuse * new_hwi,
1474 					      hw_inuse);
1475 			new_inuse = clamp_t(u32, new_inuse, 1, iocg->active);
1476 
1477 			if (new_inuse > iocg->inuse) {
1478 				TRACE_IOCG_PATH(inuse_takeback, iocg, &now,
1479 						iocg->inuse, new_inuse,
1480 						hw_inuse, new_hwi);
1481 				__propagate_active_weight(iocg, iocg->weight,
1482 							  new_inuse);
1483 			}
1484 		} else {
1485 			/* genuninely out of vtime */
1486 			nr_shortages++;
1487 		}
1488 	}
1489 
1490 	if (!nr_shortages || !nr_surpluses)
1491 		goto skip_surplus_transfers;
1492 
1493 	/* there are both shortages and surpluses, transfer surpluses */
1494 	list_for_each_entry(iocg, &ioc->active_iocgs, active_list) {
1495 		u32 usage, hw_active, hw_inuse, new_hwi, new_inuse;
1496 		int nr_valid = 0;
1497 
1498 		if (!iocg->has_surplus)
1499 			continue;
1500 
1501 		/* base the decision on max historical usage */
1502 		for (i = 0, usage = 0; i < NR_USAGE_SLOTS; i++) {
1503 			if (iocg->usages[i]) {
1504 				usage = max(usage, iocg->usages[i]);
1505 				nr_valid++;
1506 			}
1507 		}
1508 		if (nr_valid < MIN_VALID_USAGES)
1509 			continue;
1510 
1511 		current_hweight(iocg, &hw_active, &hw_inuse);
1512 		new_hwi = surplus_adjusted_hweight_inuse(usage, hw_inuse);
1513 		if (!new_hwi)
1514 			continue;
1515 
1516 		new_inuse = DIV64_U64_ROUND_UP((u64)iocg->inuse * new_hwi,
1517 					       hw_inuse);
1518 		if (new_inuse < iocg->inuse) {
1519 			TRACE_IOCG_PATH(inuse_giveaway, iocg, &now,
1520 					iocg->inuse, new_inuse,
1521 					hw_inuse, new_hwi);
1522 			__propagate_active_weight(iocg, iocg->weight, new_inuse);
1523 		}
1524 	}
1525 skip_surplus_transfers:
1526 	commit_active_weights(ioc);
1527 
1528 	/*
1529 	 * If q is getting clogged or we're missing too much, we're issuing
1530 	 * too much IO and should lower vtime rate.  If we're not missing
1531 	 * and experiencing shortages but not surpluses, we're too stingy
1532 	 * and should increase vtime rate.
1533 	 */
1534 	if (rq_wait_pct > RQ_WAIT_BUSY_PCT ||
1535 	    missed_ppm[READ] > ppm_rthr ||
1536 	    missed_ppm[WRITE] > ppm_wthr) {
1537 		ioc->busy_level = max(ioc->busy_level, 0);
1538 		ioc->busy_level++;
1539 	} else if (nr_lagging) {
1540 		ioc->busy_level = max(ioc->busy_level, 0);
1541 	} else if (nr_shortages && !nr_surpluses &&
1542 		   rq_wait_pct <= RQ_WAIT_BUSY_PCT * UNBUSY_THR_PCT / 100 &&
1543 		   missed_ppm[READ] <= ppm_rthr * UNBUSY_THR_PCT / 100 &&
1544 		   missed_ppm[WRITE] <= ppm_wthr * UNBUSY_THR_PCT / 100) {
1545 		ioc->busy_level = min(ioc->busy_level, 0);
1546 		ioc->busy_level--;
1547 	} else {
1548 		ioc->busy_level = 0;
1549 	}
1550 
1551 	ioc->busy_level = clamp(ioc->busy_level, -1000, 1000);
1552 
1553 	if (ioc->busy_level) {
1554 		u64 vrate = atomic64_read(&ioc->vtime_rate);
1555 		u64 vrate_min = ioc->vrate_min, vrate_max = ioc->vrate_max;
1556 
1557 		/* rq_wait signal is always reliable, ignore user vrate_min */
1558 		if (rq_wait_pct > RQ_WAIT_BUSY_PCT)
1559 			vrate_min = VRATE_MIN;
1560 
1561 		/*
1562 		 * If vrate is out of bounds, apply clamp gradually as the
1563 		 * bounds can change abruptly.  Otherwise, apply busy_level
1564 		 * based adjustment.
1565 		 */
1566 		if (vrate < vrate_min) {
1567 			vrate = div64_u64(vrate * (100 + VRATE_CLAMP_ADJ_PCT),
1568 					  100);
1569 			vrate = min(vrate, vrate_min);
1570 		} else if (vrate > vrate_max) {
1571 			vrate = div64_u64(vrate * (100 - VRATE_CLAMP_ADJ_PCT),
1572 					  100);
1573 			vrate = max(vrate, vrate_max);
1574 		} else {
1575 			int idx = min_t(int, abs(ioc->busy_level),
1576 					ARRAY_SIZE(vrate_adj_pct) - 1);
1577 			u32 adj_pct = vrate_adj_pct[idx];
1578 
1579 			if (ioc->busy_level > 0)
1580 				adj_pct = 100 - adj_pct;
1581 			else
1582 				adj_pct = 100 + adj_pct;
1583 
1584 			vrate = clamp(DIV64_U64_ROUND_UP(vrate * adj_pct, 100),
1585 				      vrate_min, vrate_max);
1586 		}
1587 
1588 		trace_iocost_ioc_vrate_adj(ioc, vrate, &missed_ppm, rq_wait_pct,
1589 					   nr_lagging, nr_shortages,
1590 					   nr_surpluses);
1591 
1592 		atomic64_set(&ioc->vtime_rate, vrate);
1593 		ioc->inuse_margin_vtime = DIV64_U64_ROUND_UP(
1594 			ioc->period_us * vrate * INUSE_MARGIN_PCT, 100);
1595 	}
1596 
1597 	ioc_refresh_params(ioc, false);
1598 
1599 	/*
1600 	 * This period is done.  Move onto the next one.  If nothing's
1601 	 * going on with the device, stop the timer.
1602 	 */
1603 	atomic64_inc(&ioc->cur_period);
1604 
1605 	if (ioc->running != IOC_STOP) {
1606 		if (!list_empty(&ioc->active_iocgs)) {
1607 			ioc_start_period(ioc, &now);
1608 		} else {
1609 			ioc->busy_level = 0;
1610 			ioc->running = IOC_IDLE;
1611 		}
1612 	}
1613 
1614 	spin_unlock_irq(&ioc->lock);
1615 }
1616 
1617 static void calc_vtime_cost_builtin(struct bio *bio, struct ioc_gq *iocg,
1618 				    bool is_merge, u64 *costp)
1619 {
1620 	struct ioc *ioc = iocg->ioc;
1621 	u64 coef_seqio, coef_randio, coef_page;
1622 	u64 pages = max_t(u64, bio_sectors(bio) >> IOC_SECT_TO_PAGE_SHIFT, 1);
1623 	u64 seek_pages = 0;
1624 	u64 cost = 0;
1625 
1626 	switch (bio_op(bio)) {
1627 	case REQ_OP_READ:
1628 		coef_seqio	= ioc->params.lcoefs[LCOEF_RSEQIO];
1629 		coef_randio	= ioc->params.lcoefs[LCOEF_RRANDIO];
1630 		coef_page	= ioc->params.lcoefs[LCOEF_RPAGE];
1631 		break;
1632 	case REQ_OP_WRITE:
1633 		coef_seqio	= ioc->params.lcoefs[LCOEF_WSEQIO];
1634 		coef_randio	= ioc->params.lcoefs[LCOEF_WRANDIO];
1635 		coef_page	= ioc->params.lcoefs[LCOEF_WPAGE];
1636 		break;
1637 	default:
1638 		goto out;
1639 	}
1640 
1641 	if (iocg->cursor) {
1642 		seek_pages = abs(bio->bi_iter.bi_sector - iocg->cursor);
1643 		seek_pages >>= IOC_SECT_TO_PAGE_SHIFT;
1644 	}
1645 
1646 	if (!is_merge) {
1647 		if (seek_pages > LCOEF_RANDIO_PAGES) {
1648 			cost += coef_randio;
1649 		} else {
1650 			cost += coef_seqio;
1651 		}
1652 	}
1653 	cost += pages * coef_page;
1654 out:
1655 	*costp = cost;
1656 }
1657 
1658 static u64 calc_vtime_cost(struct bio *bio, struct ioc_gq *iocg, bool is_merge)
1659 {
1660 	u64 cost;
1661 
1662 	calc_vtime_cost_builtin(bio, iocg, is_merge, &cost);
1663 	return cost;
1664 }
1665 
1666 static void ioc_rqos_throttle(struct rq_qos *rqos, struct bio *bio)
1667 {
1668 	struct blkcg_gq *blkg = bio->bi_blkg;
1669 	struct ioc *ioc = rqos_to_ioc(rqos);
1670 	struct ioc_gq *iocg = blkg_to_iocg(blkg);
1671 	struct ioc_now now;
1672 	struct iocg_wait wait;
1673 	u32 hw_active, hw_inuse;
1674 	u64 abs_cost, cost, vtime;
1675 
1676 	/* bypass IOs if disabled or for root cgroup */
1677 	if (!ioc->enabled || !iocg->level)
1678 		return;
1679 
1680 	/* always activate so that even 0 cost IOs get protected to some level */
1681 	if (!iocg_activate(iocg, &now))
1682 		return;
1683 
1684 	/* calculate the absolute vtime cost */
1685 	abs_cost = calc_vtime_cost(bio, iocg, false);
1686 	if (!abs_cost)
1687 		return;
1688 
1689 	iocg->cursor = bio_end_sector(bio);
1690 
1691 	vtime = atomic64_read(&iocg->vtime);
1692 	current_hweight(iocg, &hw_active, &hw_inuse);
1693 
1694 	if (hw_inuse < hw_active &&
1695 	    time_after_eq64(vtime + ioc->inuse_margin_vtime, now.vnow)) {
1696 		TRACE_IOCG_PATH(inuse_reset, iocg, &now,
1697 				iocg->inuse, iocg->weight, hw_inuse, hw_active);
1698 		spin_lock_irq(&ioc->lock);
1699 		propagate_active_weight(iocg, iocg->weight, iocg->weight);
1700 		spin_unlock_irq(&ioc->lock);
1701 		current_hweight(iocg, &hw_active, &hw_inuse);
1702 	}
1703 
1704 	cost = abs_cost_to_cost(abs_cost, hw_inuse);
1705 
1706 	/*
1707 	 * If no one's waiting and within budget, issue right away.  The
1708 	 * tests are racy but the races aren't systemic - we only miss once
1709 	 * in a while which is fine.
1710 	 */
1711 	if (!waitqueue_active(&iocg->waitq) &&
1712 	    !atomic64_read(&iocg->abs_vdebt) &&
1713 	    time_before_eq64(vtime + cost, now.vnow)) {
1714 		iocg_commit_bio(iocg, bio, cost);
1715 		return;
1716 	}
1717 
1718 	/*
1719 	 * We're over budget.  If @bio has to be issued regardless,
1720 	 * remember the abs_cost instead of advancing vtime.
1721 	 * iocg_kick_waitq() will pay off the debt before waking more IOs.
1722 	 * This way, the debt is continuously paid off each period with the
1723 	 * actual budget available to the cgroup.  If we just wound vtime,
1724 	 * we would incorrectly use the current hw_inuse for the entire
1725 	 * amount which, for example, can lead to the cgroup staying
1726 	 * blocked for a long time even with substantially raised hw_inuse.
1727 	 */
1728 	if (bio_issue_as_root_blkg(bio) || fatal_signal_pending(current)) {
1729 		atomic64_add(abs_cost, &iocg->abs_vdebt);
1730 		iocg_kick_delay(iocg, &now, cost);
1731 		return;
1732 	}
1733 
1734 	/*
1735 	 * Append self to the waitq and schedule the wakeup timer if we're
1736 	 * the first waiter.  The timer duration is calculated based on the
1737 	 * current vrate.  vtime and hweight changes can make it too short
1738 	 * or too long.  Each wait entry records the absolute cost it's
1739 	 * waiting for to allow re-evaluation using a custom wait entry.
1740 	 *
1741 	 * If too short, the timer simply reschedules itself.  If too long,
1742 	 * the period timer will notice and trigger wakeups.
1743 	 *
1744 	 * All waiters are on iocg->waitq and the wait states are
1745 	 * synchronized using waitq.lock.
1746 	 */
1747 	spin_lock_irq(&iocg->waitq.lock);
1748 
1749 	/*
1750 	 * We activated above but w/o any synchronization.  Deactivation is
1751 	 * synchronized with waitq.lock and we won't get deactivated as
1752 	 * long as we're waiting, so we're good if we're activated here.
1753 	 * In the unlikely case that we are deactivated, just issue the IO.
1754 	 */
1755 	if (unlikely(list_empty(&iocg->active_list))) {
1756 		spin_unlock_irq(&iocg->waitq.lock);
1757 		iocg_commit_bio(iocg, bio, cost);
1758 		return;
1759 	}
1760 
1761 	init_waitqueue_func_entry(&wait.wait, iocg_wake_fn);
1762 	wait.wait.private = current;
1763 	wait.bio = bio;
1764 	wait.abs_cost = abs_cost;
1765 	wait.committed = false;	/* will be set true by waker */
1766 
1767 	__add_wait_queue_entry_tail(&iocg->waitq, &wait.wait);
1768 	iocg_kick_waitq(iocg, &now);
1769 
1770 	spin_unlock_irq(&iocg->waitq.lock);
1771 
1772 	while (true) {
1773 		set_current_state(TASK_UNINTERRUPTIBLE);
1774 		if (wait.committed)
1775 			break;
1776 		io_schedule();
1777 	}
1778 
1779 	/* waker already committed us, proceed */
1780 	finish_wait(&iocg->waitq, &wait.wait);
1781 }
1782 
1783 static void ioc_rqos_merge(struct rq_qos *rqos, struct request *rq,
1784 			   struct bio *bio)
1785 {
1786 	struct ioc_gq *iocg = blkg_to_iocg(bio->bi_blkg);
1787 	struct ioc *ioc = iocg->ioc;
1788 	sector_t bio_end = bio_end_sector(bio);
1789 	struct ioc_now now;
1790 	u32 hw_inuse;
1791 	u64 abs_cost, cost;
1792 
1793 	/* bypass if disabled or for root cgroup */
1794 	if (!ioc->enabled || !iocg->level)
1795 		return;
1796 
1797 	abs_cost = calc_vtime_cost(bio, iocg, true);
1798 	if (!abs_cost)
1799 		return;
1800 
1801 	ioc_now(ioc, &now);
1802 	current_hweight(iocg, NULL, &hw_inuse);
1803 	cost = abs_cost_to_cost(abs_cost, hw_inuse);
1804 
1805 	/* update cursor if backmerging into the request at the cursor */
1806 	if (blk_rq_pos(rq) < bio_end &&
1807 	    blk_rq_pos(rq) + blk_rq_sectors(rq) == iocg->cursor)
1808 		iocg->cursor = bio_end;
1809 
1810 	/*
1811 	 * Charge if there's enough vtime budget and the existing request
1812 	 * has cost assigned.  Otherwise, account it as debt.  See debt
1813 	 * handling in ioc_rqos_throttle() for details.
1814 	 */
1815 	if (rq->bio && rq->bio->bi_iocost_cost &&
1816 	    time_before_eq64(atomic64_read(&iocg->vtime) + cost, now.vnow))
1817 		iocg_commit_bio(iocg, bio, cost);
1818 	else
1819 		atomic64_add(abs_cost, &iocg->abs_vdebt);
1820 }
1821 
1822 static void ioc_rqos_done_bio(struct rq_qos *rqos, struct bio *bio)
1823 {
1824 	struct ioc_gq *iocg = blkg_to_iocg(bio->bi_blkg);
1825 
1826 	if (iocg && bio->bi_iocost_cost)
1827 		atomic64_add(bio->bi_iocost_cost, &iocg->done_vtime);
1828 }
1829 
1830 static void ioc_rqos_done(struct rq_qos *rqos, struct request *rq)
1831 {
1832 	struct ioc *ioc = rqos_to_ioc(rqos);
1833 	u64 on_q_ns, rq_wait_ns;
1834 	int pidx, rw;
1835 
1836 	if (!ioc->enabled || !rq->alloc_time_ns || !rq->start_time_ns)
1837 		return;
1838 
1839 	switch (req_op(rq) & REQ_OP_MASK) {
1840 	case REQ_OP_READ:
1841 		pidx = QOS_RLAT;
1842 		rw = READ;
1843 		break;
1844 	case REQ_OP_WRITE:
1845 		pidx = QOS_WLAT;
1846 		rw = WRITE;
1847 		break;
1848 	default:
1849 		return;
1850 	}
1851 
1852 	on_q_ns = ktime_get_ns() - rq->alloc_time_ns;
1853 	rq_wait_ns = rq->start_time_ns - rq->alloc_time_ns;
1854 
1855 	if (on_q_ns <= ioc->params.qos[pidx] * NSEC_PER_USEC)
1856 		this_cpu_inc(ioc->pcpu_stat->missed[rw].nr_met);
1857 	else
1858 		this_cpu_inc(ioc->pcpu_stat->missed[rw].nr_missed);
1859 
1860 	this_cpu_add(ioc->pcpu_stat->rq_wait_ns, rq_wait_ns);
1861 }
1862 
1863 static void ioc_rqos_queue_depth_changed(struct rq_qos *rqos)
1864 {
1865 	struct ioc *ioc = rqos_to_ioc(rqos);
1866 
1867 	spin_lock_irq(&ioc->lock);
1868 	ioc_refresh_params(ioc, false);
1869 	spin_unlock_irq(&ioc->lock);
1870 }
1871 
1872 static void ioc_rqos_exit(struct rq_qos *rqos)
1873 {
1874 	struct ioc *ioc = rqos_to_ioc(rqos);
1875 
1876 	blkcg_deactivate_policy(rqos->q, &blkcg_policy_iocost);
1877 
1878 	spin_lock_irq(&ioc->lock);
1879 	ioc->running = IOC_STOP;
1880 	spin_unlock_irq(&ioc->lock);
1881 
1882 	del_timer_sync(&ioc->timer);
1883 	free_percpu(ioc->pcpu_stat);
1884 	kfree(ioc);
1885 }
1886 
1887 static struct rq_qos_ops ioc_rqos_ops = {
1888 	.throttle = ioc_rqos_throttle,
1889 	.merge = ioc_rqos_merge,
1890 	.done_bio = ioc_rqos_done_bio,
1891 	.done = ioc_rqos_done,
1892 	.queue_depth_changed = ioc_rqos_queue_depth_changed,
1893 	.exit = ioc_rqos_exit,
1894 };
1895 
1896 static int blk_iocost_init(struct request_queue *q)
1897 {
1898 	struct ioc *ioc;
1899 	struct rq_qos *rqos;
1900 	int ret;
1901 
1902 	ioc = kzalloc(sizeof(*ioc), GFP_KERNEL);
1903 	if (!ioc)
1904 		return -ENOMEM;
1905 
1906 	ioc->pcpu_stat = alloc_percpu(struct ioc_pcpu_stat);
1907 	if (!ioc->pcpu_stat) {
1908 		kfree(ioc);
1909 		return -ENOMEM;
1910 	}
1911 
1912 	rqos = &ioc->rqos;
1913 	rqos->id = RQ_QOS_COST;
1914 	rqos->ops = &ioc_rqos_ops;
1915 	rqos->q = q;
1916 
1917 	spin_lock_init(&ioc->lock);
1918 	timer_setup(&ioc->timer, ioc_timer_fn, 0);
1919 	INIT_LIST_HEAD(&ioc->active_iocgs);
1920 
1921 	ioc->running = IOC_IDLE;
1922 	atomic64_set(&ioc->vtime_rate, VTIME_PER_USEC);
1923 	seqcount_init(&ioc->period_seqcount);
1924 	ioc->period_at = ktime_to_us(ktime_get());
1925 	atomic64_set(&ioc->cur_period, 0);
1926 	atomic_set(&ioc->hweight_gen, 0);
1927 
1928 	spin_lock_irq(&ioc->lock);
1929 	ioc->autop_idx = AUTOP_INVALID;
1930 	ioc_refresh_params(ioc, true);
1931 	spin_unlock_irq(&ioc->lock);
1932 
1933 	rq_qos_add(q, rqos);
1934 	ret = blkcg_activate_policy(q, &blkcg_policy_iocost);
1935 	if (ret) {
1936 		rq_qos_del(q, rqos);
1937 		free_percpu(ioc->pcpu_stat);
1938 		kfree(ioc);
1939 		return ret;
1940 	}
1941 	return 0;
1942 }
1943 
1944 static struct blkcg_policy_data *ioc_cpd_alloc(gfp_t gfp)
1945 {
1946 	struct ioc_cgrp *iocc;
1947 
1948 	iocc = kzalloc(sizeof(struct ioc_cgrp), gfp);
1949 	if (!iocc)
1950 		return NULL;
1951 
1952 	iocc->dfl_weight = CGROUP_WEIGHT_DFL;
1953 	return &iocc->cpd;
1954 }
1955 
1956 static void ioc_cpd_free(struct blkcg_policy_data *cpd)
1957 {
1958 	kfree(container_of(cpd, struct ioc_cgrp, cpd));
1959 }
1960 
1961 static struct blkg_policy_data *ioc_pd_alloc(gfp_t gfp, struct request_queue *q,
1962 					     struct blkcg *blkcg)
1963 {
1964 	int levels = blkcg->css.cgroup->level + 1;
1965 	struct ioc_gq *iocg;
1966 
1967 	iocg = kzalloc_node(sizeof(*iocg) + levels * sizeof(iocg->ancestors[0]),
1968 			    gfp, q->node);
1969 	if (!iocg)
1970 		return NULL;
1971 
1972 	return &iocg->pd;
1973 }
1974 
1975 static void ioc_pd_init(struct blkg_policy_data *pd)
1976 {
1977 	struct ioc_gq *iocg = pd_to_iocg(pd);
1978 	struct blkcg_gq *blkg = pd_to_blkg(&iocg->pd);
1979 	struct ioc *ioc = q_to_ioc(blkg->q);
1980 	struct ioc_now now;
1981 	struct blkcg_gq *tblkg;
1982 	unsigned long flags;
1983 
1984 	ioc_now(ioc, &now);
1985 
1986 	iocg->ioc = ioc;
1987 	atomic64_set(&iocg->vtime, now.vnow);
1988 	atomic64_set(&iocg->done_vtime, now.vnow);
1989 	atomic64_set(&iocg->abs_vdebt, 0);
1990 	atomic64_set(&iocg->active_period, atomic64_read(&ioc->cur_period));
1991 	INIT_LIST_HEAD(&iocg->active_list);
1992 	iocg->hweight_active = HWEIGHT_WHOLE;
1993 	iocg->hweight_inuse = HWEIGHT_WHOLE;
1994 
1995 	init_waitqueue_head(&iocg->waitq);
1996 	hrtimer_init(&iocg->waitq_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
1997 	iocg->waitq_timer.function = iocg_waitq_timer_fn;
1998 	hrtimer_init(&iocg->delay_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
1999 	iocg->delay_timer.function = iocg_delay_timer_fn;
2000 
2001 	iocg->level = blkg->blkcg->css.cgroup->level;
2002 
2003 	for (tblkg = blkg; tblkg; tblkg = tblkg->parent) {
2004 		struct ioc_gq *tiocg = blkg_to_iocg(tblkg);
2005 		iocg->ancestors[tiocg->level] = tiocg;
2006 	}
2007 
2008 	spin_lock_irqsave(&ioc->lock, flags);
2009 	weight_updated(iocg);
2010 	spin_unlock_irqrestore(&ioc->lock, flags);
2011 }
2012 
2013 static void ioc_pd_free(struct blkg_policy_data *pd)
2014 {
2015 	struct ioc_gq *iocg = pd_to_iocg(pd);
2016 	struct ioc *ioc = iocg->ioc;
2017 
2018 	if (ioc) {
2019 		spin_lock(&ioc->lock);
2020 		if (!list_empty(&iocg->active_list)) {
2021 			propagate_active_weight(iocg, 0, 0);
2022 			list_del_init(&iocg->active_list);
2023 		}
2024 		spin_unlock(&ioc->lock);
2025 
2026 		hrtimer_cancel(&iocg->waitq_timer);
2027 		hrtimer_cancel(&iocg->delay_timer);
2028 	}
2029 	kfree(iocg);
2030 }
2031 
2032 static u64 ioc_weight_prfill(struct seq_file *sf, struct blkg_policy_data *pd,
2033 			     int off)
2034 {
2035 	const char *dname = blkg_dev_name(pd->blkg);
2036 	struct ioc_gq *iocg = pd_to_iocg(pd);
2037 
2038 	if (dname && iocg->cfg_weight)
2039 		seq_printf(sf, "%s %u\n", dname, iocg->cfg_weight);
2040 	return 0;
2041 }
2042 
2043 
2044 static int ioc_weight_show(struct seq_file *sf, void *v)
2045 {
2046 	struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
2047 	struct ioc_cgrp *iocc = blkcg_to_iocc(blkcg);
2048 
2049 	seq_printf(sf, "default %u\n", iocc->dfl_weight);
2050 	blkcg_print_blkgs(sf, blkcg, ioc_weight_prfill,
2051 			  &blkcg_policy_iocost, seq_cft(sf)->private, false);
2052 	return 0;
2053 }
2054 
2055 static ssize_t ioc_weight_write(struct kernfs_open_file *of, char *buf,
2056 				size_t nbytes, loff_t off)
2057 {
2058 	struct blkcg *blkcg = css_to_blkcg(of_css(of));
2059 	struct ioc_cgrp *iocc = blkcg_to_iocc(blkcg);
2060 	struct blkg_conf_ctx ctx;
2061 	struct ioc_gq *iocg;
2062 	u32 v;
2063 	int ret;
2064 
2065 	if (!strchr(buf, ':')) {
2066 		struct blkcg_gq *blkg;
2067 
2068 		if (!sscanf(buf, "default %u", &v) && !sscanf(buf, "%u", &v))
2069 			return -EINVAL;
2070 
2071 		if (v < CGROUP_WEIGHT_MIN || v > CGROUP_WEIGHT_MAX)
2072 			return -EINVAL;
2073 
2074 		spin_lock(&blkcg->lock);
2075 		iocc->dfl_weight = v;
2076 		hlist_for_each_entry(blkg, &blkcg->blkg_list, blkcg_node) {
2077 			struct ioc_gq *iocg = blkg_to_iocg(blkg);
2078 
2079 			if (iocg) {
2080 				spin_lock_irq(&iocg->ioc->lock);
2081 				weight_updated(iocg);
2082 				spin_unlock_irq(&iocg->ioc->lock);
2083 			}
2084 		}
2085 		spin_unlock(&blkcg->lock);
2086 
2087 		return nbytes;
2088 	}
2089 
2090 	ret = blkg_conf_prep(blkcg, &blkcg_policy_iocost, buf, &ctx);
2091 	if (ret)
2092 		return ret;
2093 
2094 	iocg = blkg_to_iocg(ctx.blkg);
2095 
2096 	if (!strncmp(ctx.body, "default", 7)) {
2097 		v = 0;
2098 	} else {
2099 		if (!sscanf(ctx.body, "%u", &v))
2100 			goto einval;
2101 		if (v < CGROUP_WEIGHT_MIN || v > CGROUP_WEIGHT_MAX)
2102 			goto einval;
2103 	}
2104 
2105 	spin_lock_irq(&iocg->ioc->lock);
2106 	iocg->cfg_weight = v;
2107 	weight_updated(iocg);
2108 	spin_unlock_irq(&iocg->ioc->lock);
2109 
2110 	blkg_conf_finish(&ctx);
2111 	return nbytes;
2112 
2113 einval:
2114 	blkg_conf_finish(&ctx);
2115 	return -EINVAL;
2116 }
2117 
2118 static u64 ioc_qos_prfill(struct seq_file *sf, struct blkg_policy_data *pd,
2119 			  int off)
2120 {
2121 	const char *dname = blkg_dev_name(pd->blkg);
2122 	struct ioc *ioc = pd_to_iocg(pd)->ioc;
2123 
2124 	if (!dname)
2125 		return 0;
2126 
2127 	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",
2128 		   dname, ioc->enabled, ioc->user_qos_params ? "user" : "auto",
2129 		   ioc->params.qos[QOS_RPPM] / 10000,
2130 		   ioc->params.qos[QOS_RPPM] % 10000 / 100,
2131 		   ioc->params.qos[QOS_RLAT],
2132 		   ioc->params.qos[QOS_WPPM] / 10000,
2133 		   ioc->params.qos[QOS_WPPM] % 10000 / 100,
2134 		   ioc->params.qos[QOS_WLAT],
2135 		   ioc->params.qos[QOS_MIN] / 10000,
2136 		   ioc->params.qos[QOS_MIN] % 10000 / 100,
2137 		   ioc->params.qos[QOS_MAX] / 10000,
2138 		   ioc->params.qos[QOS_MAX] % 10000 / 100);
2139 	return 0;
2140 }
2141 
2142 static int ioc_qos_show(struct seq_file *sf, void *v)
2143 {
2144 	struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
2145 
2146 	blkcg_print_blkgs(sf, blkcg, ioc_qos_prfill,
2147 			  &blkcg_policy_iocost, seq_cft(sf)->private, false);
2148 	return 0;
2149 }
2150 
2151 static const match_table_t qos_ctrl_tokens = {
2152 	{ QOS_ENABLE,		"enable=%u"	},
2153 	{ QOS_CTRL,		"ctrl=%s"	},
2154 	{ NR_QOS_CTRL_PARAMS,	NULL		},
2155 };
2156 
2157 static const match_table_t qos_tokens = {
2158 	{ QOS_RPPM,		"rpct=%s"	},
2159 	{ QOS_RLAT,		"rlat=%u"	},
2160 	{ QOS_WPPM,		"wpct=%s"	},
2161 	{ QOS_WLAT,		"wlat=%u"	},
2162 	{ QOS_MIN,		"min=%s"	},
2163 	{ QOS_MAX,		"max=%s"	},
2164 	{ NR_QOS_PARAMS,	NULL		},
2165 };
2166 
2167 static ssize_t ioc_qos_write(struct kernfs_open_file *of, char *input,
2168 			     size_t nbytes, loff_t off)
2169 {
2170 	struct gendisk *disk;
2171 	struct ioc *ioc;
2172 	u32 qos[NR_QOS_PARAMS];
2173 	bool enable, user;
2174 	char *p;
2175 	int ret;
2176 
2177 	disk = blkcg_conf_get_disk(&input);
2178 	if (IS_ERR(disk))
2179 		return PTR_ERR(disk);
2180 
2181 	ioc = q_to_ioc(disk->queue);
2182 	if (!ioc) {
2183 		ret = blk_iocost_init(disk->queue);
2184 		if (ret)
2185 			goto err;
2186 		ioc = q_to_ioc(disk->queue);
2187 	}
2188 
2189 	spin_lock_irq(&ioc->lock);
2190 	memcpy(qos, ioc->params.qos, sizeof(qos));
2191 	enable = ioc->enabled;
2192 	user = ioc->user_qos_params;
2193 	spin_unlock_irq(&ioc->lock);
2194 
2195 	while ((p = strsep(&input, " \t\n"))) {
2196 		substring_t args[MAX_OPT_ARGS];
2197 		char buf[32];
2198 		int tok;
2199 		s64 v;
2200 
2201 		if (!*p)
2202 			continue;
2203 
2204 		switch (match_token(p, qos_ctrl_tokens, args)) {
2205 		case QOS_ENABLE:
2206 			match_u64(&args[0], &v);
2207 			enable = v;
2208 			continue;
2209 		case QOS_CTRL:
2210 			match_strlcpy(buf, &args[0], sizeof(buf));
2211 			if (!strcmp(buf, "auto"))
2212 				user = false;
2213 			else if (!strcmp(buf, "user"))
2214 				user = true;
2215 			else
2216 				goto einval;
2217 			continue;
2218 		}
2219 
2220 		tok = match_token(p, qos_tokens, args);
2221 		switch (tok) {
2222 		case QOS_RPPM:
2223 		case QOS_WPPM:
2224 			if (match_strlcpy(buf, &args[0], sizeof(buf)) >=
2225 			    sizeof(buf))
2226 				goto einval;
2227 			if (cgroup_parse_float(buf, 2, &v))
2228 				goto einval;
2229 			if (v < 0 || v > 10000)
2230 				goto einval;
2231 			qos[tok] = v * 100;
2232 			break;
2233 		case QOS_RLAT:
2234 		case QOS_WLAT:
2235 			if (match_u64(&args[0], &v))
2236 				goto einval;
2237 			qos[tok] = v;
2238 			break;
2239 		case QOS_MIN:
2240 		case QOS_MAX:
2241 			if (match_strlcpy(buf, &args[0], sizeof(buf)) >=
2242 			    sizeof(buf))
2243 				goto einval;
2244 			if (cgroup_parse_float(buf, 2, &v))
2245 				goto einval;
2246 			if (v < 0)
2247 				goto einval;
2248 			qos[tok] = clamp_t(s64, v * 100,
2249 					   VRATE_MIN_PPM, VRATE_MAX_PPM);
2250 			break;
2251 		default:
2252 			goto einval;
2253 		}
2254 		user = true;
2255 	}
2256 
2257 	if (qos[QOS_MIN] > qos[QOS_MAX])
2258 		goto einval;
2259 
2260 	spin_lock_irq(&ioc->lock);
2261 
2262 	if (enable) {
2263 		blk_queue_flag_set(QUEUE_FLAG_RQ_ALLOC_TIME, ioc->rqos.q);
2264 		ioc->enabled = true;
2265 	} else {
2266 		blk_queue_flag_clear(QUEUE_FLAG_RQ_ALLOC_TIME, ioc->rqos.q);
2267 		ioc->enabled = false;
2268 	}
2269 
2270 	if (user) {
2271 		memcpy(ioc->params.qos, qos, sizeof(qos));
2272 		ioc->user_qos_params = true;
2273 	} else {
2274 		ioc->user_qos_params = false;
2275 	}
2276 
2277 	ioc_refresh_params(ioc, true);
2278 	spin_unlock_irq(&ioc->lock);
2279 
2280 	put_disk_and_module(disk);
2281 	return nbytes;
2282 einval:
2283 	ret = -EINVAL;
2284 err:
2285 	put_disk_and_module(disk);
2286 	return ret;
2287 }
2288 
2289 static u64 ioc_cost_model_prfill(struct seq_file *sf,
2290 				 struct blkg_policy_data *pd, int off)
2291 {
2292 	const char *dname = blkg_dev_name(pd->blkg);
2293 	struct ioc *ioc = pd_to_iocg(pd)->ioc;
2294 	u64 *u = ioc->params.i_lcoefs;
2295 
2296 	if (!dname)
2297 		return 0;
2298 
2299 	seq_printf(sf, "%s ctrl=%s model=linear "
2300 		   "rbps=%llu rseqiops=%llu rrandiops=%llu "
2301 		   "wbps=%llu wseqiops=%llu wrandiops=%llu\n",
2302 		   dname, ioc->user_cost_model ? "user" : "auto",
2303 		   u[I_LCOEF_RBPS], u[I_LCOEF_RSEQIOPS], u[I_LCOEF_RRANDIOPS],
2304 		   u[I_LCOEF_WBPS], u[I_LCOEF_WSEQIOPS], u[I_LCOEF_WRANDIOPS]);
2305 	return 0;
2306 }
2307 
2308 static int ioc_cost_model_show(struct seq_file *sf, void *v)
2309 {
2310 	struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
2311 
2312 	blkcg_print_blkgs(sf, blkcg, ioc_cost_model_prfill,
2313 			  &blkcg_policy_iocost, seq_cft(sf)->private, false);
2314 	return 0;
2315 }
2316 
2317 static const match_table_t cost_ctrl_tokens = {
2318 	{ COST_CTRL,		"ctrl=%s"	},
2319 	{ COST_MODEL,		"model=%s"	},
2320 	{ NR_COST_CTRL_PARAMS,	NULL		},
2321 };
2322 
2323 static const match_table_t i_lcoef_tokens = {
2324 	{ I_LCOEF_RBPS,		"rbps=%u"	},
2325 	{ I_LCOEF_RSEQIOPS,	"rseqiops=%u"	},
2326 	{ I_LCOEF_RRANDIOPS,	"rrandiops=%u"	},
2327 	{ I_LCOEF_WBPS,		"wbps=%u"	},
2328 	{ I_LCOEF_WSEQIOPS,	"wseqiops=%u"	},
2329 	{ I_LCOEF_WRANDIOPS,	"wrandiops=%u"	},
2330 	{ NR_I_LCOEFS,		NULL		},
2331 };
2332 
2333 static ssize_t ioc_cost_model_write(struct kernfs_open_file *of, char *input,
2334 				    size_t nbytes, loff_t off)
2335 {
2336 	struct gendisk *disk;
2337 	struct ioc *ioc;
2338 	u64 u[NR_I_LCOEFS];
2339 	bool user;
2340 	char *p;
2341 	int ret;
2342 
2343 	disk = blkcg_conf_get_disk(&input);
2344 	if (IS_ERR(disk))
2345 		return PTR_ERR(disk);
2346 
2347 	ioc = q_to_ioc(disk->queue);
2348 	if (!ioc) {
2349 		ret = blk_iocost_init(disk->queue);
2350 		if (ret)
2351 			goto err;
2352 		ioc = q_to_ioc(disk->queue);
2353 	}
2354 
2355 	spin_lock_irq(&ioc->lock);
2356 	memcpy(u, ioc->params.i_lcoefs, sizeof(u));
2357 	user = ioc->user_cost_model;
2358 	spin_unlock_irq(&ioc->lock);
2359 
2360 	while ((p = strsep(&input, " \t\n"))) {
2361 		substring_t args[MAX_OPT_ARGS];
2362 		char buf[32];
2363 		int tok;
2364 		u64 v;
2365 
2366 		if (!*p)
2367 			continue;
2368 
2369 		switch (match_token(p, cost_ctrl_tokens, args)) {
2370 		case COST_CTRL:
2371 			match_strlcpy(buf, &args[0], sizeof(buf));
2372 			if (!strcmp(buf, "auto"))
2373 				user = false;
2374 			else if (!strcmp(buf, "user"))
2375 				user = true;
2376 			else
2377 				goto einval;
2378 			continue;
2379 		case COST_MODEL:
2380 			match_strlcpy(buf, &args[0], sizeof(buf));
2381 			if (strcmp(buf, "linear"))
2382 				goto einval;
2383 			continue;
2384 		}
2385 
2386 		tok = match_token(p, i_lcoef_tokens, args);
2387 		if (tok == NR_I_LCOEFS)
2388 			goto einval;
2389 		if (match_u64(&args[0], &v))
2390 			goto einval;
2391 		u[tok] = v;
2392 		user = true;
2393 	}
2394 
2395 	spin_lock_irq(&ioc->lock);
2396 	if (user) {
2397 		memcpy(ioc->params.i_lcoefs, u, sizeof(u));
2398 		ioc->user_cost_model = true;
2399 	} else {
2400 		ioc->user_cost_model = false;
2401 	}
2402 	ioc_refresh_params(ioc, true);
2403 	spin_unlock_irq(&ioc->lock);
2404 
2405 	put_disk_and_module(disk);
2406 	return nbytes;
2407 
2408 einval:
2409 	ret = -EINVAL;
2410 err:
2411 	put_disk_and_module(disk);
2412 	return ret;
2413 }
2414 
2415 static struct cftype ioc_files[] = {
2416 	{
2417 		.name = "weight",
2418 		.flags = CFTYPE_NOT_ON_ROOT,
2419 		.seq_show = ioc_weight_show,
2420 		.write = ioc_weight_write,
2421 	},
2422 	{
2423 		.name = "cost.qos",
2424 		.flags = CFTYPE_ONLY_ON_ROOT,
2425 		.seq_show = ioc_qos_show,
2426 		.write = ioc_qos_write,
2427 	},
2428 	{
2429 		.name = "cost.model",
2430 		.flags = CFTYPE_ONLY_ON_ROOT,
2431 		.seq_show = ioc_cost_model_show,
2432 		.write = ioc_cost_model_write,
2433 	},
2434 	{}
2435 };
2436 
2437 static struct blkcg_policy blkcg_policy_iocost = {
2438 	.dfl_cftypes	= ioc_files,
2439 	.cpd_alloc_fn	= ioc_cpd_alloc,
2440 	.cpd_free_fn	= ioc_cpd_free,
2441 	.pd_alloc_fn	= ioc_pd_alloc,
2442 	.pd_init_fn	= ioc_pd_init,
2443 	.pd_free_fn	= ioc_pd_free,
2444 };
2445 
2446 static int __init ioc_init(void)
2447 {
2448 	return blkcg_policy_register(&blkcg_policy_iocost);
2449 }
2450 
2451 static void __exit ioc_exit(void)
2452 {
2453 	return blkcg_policy_unregister(&blkcg_policy_iocost);
2454 }
2455 
2456 module_init(ioc_init);
2457 module_exit(ioc_exit);
2458