blk-iocost.c

   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 * parameters 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 * 2. Control Strategy
  50 *
  51 * The device virtual time (vtime) is used as the primary control metric.
  52 * The control strategy is composed of the following three parts.
  53 *
  54 * 2-1. Vtime Distribution
  55 *
  56 * When a cgroup becomes active in terms of IOs, its hierarchical share is
  57 * calculated.  Please consider the following hierarchy where the numbers
  58 * inside parentheses denote the configured weights.
  59 *
  60 *           root
  61 *         /       \
  62 *      A (w:100)  B (w:300)
  63 *      /       \
  64 *  A0 (w:100)  A1 (w:100)
  65 *
  66 * If B is idle and only A0 and A1 are actively issuing IOs, as the two are
  67 * of equal weight, each gets 50% share.  If then B starts issuing IOs, B
  68 * gets 300/(100+300) or 75% share, and A0 and A1 equally splits the rest,
  69 * 12.5% each.  The distribution mechanism only cares about these flattened
  70 * shares.  They're called hweights (hierarchical weights) and always add
  71 * upto 1 (WEIGHT_ONE).
  72 *
  73 * A given cgroup's vtime runs slower in inverse proportion to its hweight.
  74 * For example, with 12.5% weight, A0's time runs 8 times slower (100/12.5)
  75 * against the device vtime - an IO which takes 10ms on the underlying
  76 * device is considered to take 80ms on A0.
  77 *
  78 * This constitutes the basis of IO capacity distribution.  Each cgroup's
  79 * vtime is running at a rate determined by its hweight.  A cgroup tracks
  80 * the vtime consumed by past IOs and can issue a new IO if doing so
  81 * wouldn't outrun the current device vtime.  Otherwise, the IO is
  82 * suspended until the vtime has progressed enough to cover it.
  83 *
  84 * 2-2. Vrate Adjustment
  85 *
  86 * It's unrealistic to expect the cost model to be perfect.  There are too
  87 * many devices and even on the same device the overall performance
  88 * fluctuates depending on numerous factors such as IO mixture and device
  89 * internal garbage collection.  The controller needs to adapt dynamically.
  90 *
  91 * This is achieved by adjusting the overall IO rate according to how busy
  92 * the device is.  If the device becomes overloaded, we're sending down too
  93 * many IOs and should generally slow down.  If there are waiting issuers
  94 * but the device isn't saturated, we're issuing too few and should
  95 * generally speed up.
  96 *
  97 * To slow down, we lower the vrate - the rate at which the device vtime
  98 * passes compared to the wall clock.  For example, if the vtime is running
  99 * at the vrate of 75%, all cgroups added up would only be able to issue
 100 * 750ms worth of IOs per second, and vice-versa for speeding up.
 101 *
 102 * Device business is determined using two criteria - rq wait and
 103 * completion latencies.
 104 *
 105 * When a device gets saturated, the on-device and then the request queues
 106 * fill up and a bio which is ready to be issued has to wait for a request
 107 * to become available.  When this delay becomes noticeable, it's a clear
 108 * indication that the device is saturated and we lower the vrate.  This
 109 * saturation signal is fairly conservative as it only triggers when both
 110 * hardware and software queues are filled up, and is used as the default
 111 * busy signal.
 112 *
 113 * As devices can have deep queues and be unfair in how the queued commands
 114 * are executed, solely depending on rq wait may not result in satisfactory
 115 * control quality.  For a better control quality, completion latency QoS
 116 * parameters can be configured so that the device is considered saturated
 117 * if N'th percentile completion latency rises above the set point.
 118 *
 119 * The completion latency requirements are a function of both the
 120 * underlying device characteristics and the desired IO latency quality of
 121 * service.  There is an inherent trade-off - the tighter the latency QoS,
 122 * the higher the bandwidth lossage.  Latency QoS is disabled by default
 123 * and can be set through /sys/fs/cgroup/io.cost.qos.
 124 *
 125 * 2-3. Work Conservation
 126 *
 127 * Imagine two cgroups A and B with equal weights.  A is issuing a small IO
 128 * periodically while B is sending out enough parallel IOs to saturate the
 129 * device on its own.  Let's say A's usage amounts to 100ms worth of IO
 130 * cost per second, i.e., 10% of the device capacity.  The naive
 131 * distribution of half and half would lead to 60% utilization of the
 132 * device, a significant reduction in the total amount of work done
 133 * compared to free-for-all competition.  This is too high a cost to pay
 134 * for IO control.
 135 *
 136 * To conserve the total amount of work done, we keep track of how much
 137 * each active cgroup is actually using and yield part of its weight if
 138 * there are other cgroups which can make use of it.  In the above case,
 139 * A's weight will be lowered so that it hovers above the actual usage and
 140 * B would be able to use the rest.
 141 *
 142 * As we don't want to penalize a cgroup for donating its weight, the
 143 * surplus weight adjustment factors in a margin and has an immediate
 144 * snapback mechanism in case the cgroup needs more IO vtime for itself.
 145 *
 146 * Note that adjusting down surplus weights has the same effects as
 147 * accelerating vtime for other cgroups and work conservation can also be
 148 * implemented by adjusting vrate dynamically.  However, squaring who can
 149 * donate and should take back how much requires hweight propagations
 150 * anyway making it easier to implement and understand as a separate
 151 * mechanism.
 152 *
 153 * 3. Monitoring
 154 *
 155 * Instead of debugfs or other clumsy monitoring mechanisms, this
 156 * controller uses a drgn based monitoring script -
 157 * tools/cgroup/iocost_monitor.py.  For details on drgn, please see
 158 * https://github.com/osandov/drgn.  The output looks like the following.
 159 *
 160 *  sdb RUN   per=300ms cur_per=234.218:v203.695 busy= +1 vrate= 62.12%
 161 *                 active      weight      hweight% inflt% dbt  delay usages%
 162 *  test/a              *    50/   50  33.33/ 33.33  27.65   2  0*041 033:033:033
 163 *  test/b              *   100/  100  66.67/ 66.67  17.56   0  0*000 066:079:077
 164 *
 165 * - per	: Timer period
 166 * - cur_per	: Internal wall and device vtime clock
 167 * - vrate	: Device virtual time rate against wall clock
 168 * - weight	: Surplus-adjusted and configured weights
 169 * - hweight	: Surplus-adjusted and configured hierarchical weights
 170 * - inflt	: The percentage of in-flight IO cost at the end of last period
 171 * - del_ms	: Deferred issuer delay induction level and duration
 172 * - usages	: Usage history
 173 */
 174
 175#include <linux/kernel.h>
 176#include <linux/module.h>
 177#include <linux/timer.h>
 178#include <linux/time64.h>
 179#include <linux/parser.h>
 180#include <linux/sched/signal.h>
 181#include <asm/local.h>
 182#include <asm/local64.h>
 183#include "blk-rq-qos.h"
 184#include "blk-stat.h"
 185#include "blk-wbt.h"
 186#include "blk-cgroup.h"
 187
 188#ifdef CONFIG_TRACEPOINTS
 189
 190/* copied from TRACE_CGROUP_PATH, see cgroup-internal.h */
 191#define TRACE_IOCG_PATH_LEN 1024
 192static DEFINE_SPINLOCK(trace_iocg_path_lock);
 193static char trace_iocg_path[TRACE_IOCG_PATH_LEN];
 194
 195#define TRACE_IOCG_PATH(type, iocg, ...)					\
 196	do {									\
 197		unsigned long flags;						\
 198		if (trace_iocost_##type##_enabled()) {				\
 199			spin_lock_irqsave(&trace_iocg_path_lock, flags);	\
 200			cgroup_path(iocg_to_blkg(iocg)->blkcg->css.cgroup,	\
 201				    trace_iocg_path, TRACE_IOCG_PATH_LEN);	\
 202			trace_iocost_##type(iocg, trace_iocg_path,		\
 203					      ##__VA_ARGS__);			\
 204			spin_unlock_irqrestore(&trace_iocg_path_lock, flags);	\
 205		}								\
 206	} while (0)
 207
 208#else	/* CONFIG_TRACE_POINTS */
 209#define TRACE_IOCG_PATH(type, iocg, ...)	do { } while (0)
 210#endif	/* CONFIG_TRACE_POINTS */
 211
 212enum {
 213	MILLION			= 1000000,
 214
 215	/* timer period is calculated from latency requirements, bound it */
 216	MIN_PERIOD		= USEC_PER_MSEC,
 217	MAX_PERIOD		= USEC_PER_SEC,
 218
 219	/*
 220	 * iocg->vtime is targeted at 50% behind the device vtime, which
 221	 * serves as its IO credit buffer.  Surplus weight adjustment is
 222	 * immediately canceled if the vtime margin runs below 10%.
 223	 */
 224	MARGIN_MIN_PCT		= 10,
 225	MARGIN_LOW_PCT		= 20,
 226	MARGIN_TARGET_PCT	= 50,
 227
 228	INUSE_ADJ_STEP_PCT	= 25,
 229
 230	/* Have some play in timer operations */
 231	TIMER_SLACK_PCT		= 1,
 232
 233	/* 1/64k is granular enough and can easily be handled w/ u32 */
 234	WEIGHT_ONE		= 1 << 16,
 235};
 236
 237enum {
 238	/*
 239	 * As vtime is used to calculate the cost of each IO, it needs to
 240	 * be fairly high precision.  For example, it should be able to
 241	 * represent the cost of a single page worth of discard with
 242	 * suffificient accuracy.  At the same time, it should be able to
 243	 * represent reasonably long enough durations to be useful and
 244	 * convenient during operation.
 245	 *
 246	 * 1s worth of vtime is 2^37.  This gives us both sub-nanosecond
 247	 * granularity and days of wrap-around time even at extreme vrates.
 248	 */
 249	VTIME_PER_SEC_SHIFT	= 37,
 250	VTIME_PER_SEC		= 1LLU << VTIME_PER_SEC_SHIFT,
 251	VTIME_PER_USEC		= VTIME_PER_SEC / USEC_PER_SEC,
 252	VTIME_PER_NSEC		= VTIME_PER_SEC / NSEC_PER_SEC,
 253
 254	/* bound vrate adjustments within two orders of magnitude */
 255	VRATE_MIN_PPM		= 10000,	/* 1% */
 256	VRATE_MAX_PPM		= 100000000,	/* 10000% */
 257
 258	VRATE_MIN		= VTIME_PER_USEC * VRATE_MIN_PPM / MILLION,
 259	VRATE_CLAMP_ADJ_PCT	= 4,
 260
 261	/* switch iff the conditions are met for longer than this */
 262	AUTOP_CYCLE_NSEC	= 10LLU * NSEC_PER_SEC,
 263};
 264
 265enum {
 266	/* if IOs end up waiting for requests, issue less */
 267	RQ_WAIT_BUSY_PCT	= 5,
 268
 269	/* unbusy hysterisis */
 270	UNBUSY_THR_PCT		= 75,
 271
 272	/*
 273	 * The effect of delay is indirect and non-linear and a huge amount of
 274	 * future debt can accumulate abruptly while unthrottled. Linearly scale
 275	 * up delay as debt is going up and then let it decay exponentially.
 276	 * This gives us quick ramp ups while delay is accumulating and long
 277	 * tails which can help reducing the frequency of debt explosions on
 278	 * unthrottle. The parameters are experimentally determined.
 279	 *
 280	 * The delay mechanism provides adequate protection and behavior in many
 281	 * cases. However, this is far from ideal and falls shorts on both
 282	 * fronts. The debtors are often throttled too harshly costing a
 283	 * significant level of fairness and possibly total work while the
 284	 * protection against their impacts on the system can be choppy and
 285	 * unreliable.
 286	 *
 287	 * The shortcoming primarily stems from the fact that, unlike for page
 288	 * cache, the kernel doesn't have well-defined back-pressure propagation
 289	 * mechanism and policies for anonymous memory. Fully addressing this
 290	 * issue will likely require substantial improvements in the area.
 291	 */
 292	MIN_DELAY_THR_PCT	= 500,
 293	MAX_DELAY_THR_PCT	= 25000,
 294	MIN_DELAY		= 250,
 295	MAX_DELAY		= 250 * USEC_PER_MSEC,
 296
 297	/* halve debts if avg usage over 100ms is under 50% */
 298	DFGV_USAGE_PCT		= 50,
 299	DFGV_PERIOD		= 100 * USEC_PER_MSEC,
 300
 301	/* don't let cmds which take a very long time pin lagging for too long */
 302	MAX_LAGGING_PERIODS	= 10,
 303
 304	/*
 305	 * Count IO size in 4k pages.  The 12bit shift helps keeping
 306	 * size-proportional components of cost calculation in closer
 307	 * numbers of digits to per-IO cost components.
 308	 */
 309	IOC_PAGE_SHIFT		= 12,
 310	IOC_PAGE_SIZE		= 1 << IOC_PAGE_SHIFT,
 311	IOC_SECT_TO_PAGE_SHIFT	= IOC_PAGE_SHIFT - SECTOR_SHIFT,
 312
 313	/* if apart further than 16M, consider randio for linear model */
 314	LCOEF_RANDIO_PAGES	= 4096,
 315};
 316
 317enum ioc_running {
 318	IOC_IDLE,
 319	IOC_RUNNING,
 320	IOC_STOP,
 321};
 322
 323/* io.cost.qos controls including per-dev enable of the whole controller */
 324enum {
 325	QOS_ENABLE,
 326	QOS_CTRL,
 327	NR_QOS_CTRL_PARAMS,
 328};
 329
 330/* io.cost.qos params */
 331enum {
 332	QOS_RPPM,
 333	QOS_RLAT,
 334	QOS_WPPM,
 335	QOS_WLAT,
 336	QOS_MIN,
 337	QOS_MAX,
 338	NR_QOS_PARAMS,
 339};
 340
 341/* io.cost.model controls */
 342enum {
 343	COST_CTRL,
 344	COST_MODEL,
 345	NR_COST_CTRL_PARAMS,
 346};
 347
 348/* builtin linear cost model coefficients */
 349enum {
 350	I_LCOEF_RBPS,
 351	I_LCOEF_RSEQIOPS,
 352	I_LCOEF_RRANDIOPS,
 353	I_LCOEF_WBPS,
 354	I_LCOEF_WSEQIOPS,
 355	I_LCOEF_WRANDIOPS,
 356	NR_I_LCOEFS,
 357};
 358
 359enum {
 360	LCOEF_RPAGE,
 361	LCOEF_RSEQIO,
 362	LCOEF_RRANDIO,
 363	LCOEF_WPAGE,
 364	LCOEF_WSEQIO,
 365	LCOEF_WRANDIO,
 366	NR_LCOEFS,
 367};
 368
 369enum {
 370	AUTOP_INVALID,
 371	AUTOP_HDD,
 372	AUTOP_SSD_QD1,
 373	AUTOP_SSD_DFL,
 374	AUTOP_SSD_FAST,
 375};
 376
 377struct ioc_params {
 378	u32				qos[NR_QOS_PARAMS];
 379	u64				i_lcoefs[NR_I_LCOEFS];
 380	u64				lcoefs[NR_LCOEFS];
 381	u32				too_fast_vrate_pct;
 382	u32				too_slow_vrate_pct;
 383};
 384
 385struct ioc_margins {
 386	s64				min;
 387	s64				low;
 388	s64				target;
 389};
 390
 391struct ioc_missed {
 392	local_t				nr_met;
 393	local_t				nr_missed;
 394	u32				last_met;
 395	u32				last_missed;
 396};
 397
 398struct ioc_pcpu_stat {
 399	struct ioc_missed		missed[2];
 400
 401	local64_t			rq_wait_ns;
 402	u64				last_rq_wait_ns;
 403};
 404
 405/* per device */
 406struct ioc {
 407	struct rq_qos			rqos;
 408
 409	bool				enabled;
 410
 411	struct ioc_params		params;
 412	struct ioc_margins		margins;
 413	u32				period_us;
 414	u32				timer_slack_ns;
 415	u64				vrate_min;
 416	u64				vrate_max;
 417
 418	spinlock_t			lock;
 419	struct timer_list		timer;
 420	struct list_head		active_iocgs;	/* active cgroups */
 421	struct ioc_pcpu_stat __percpu	*pcpu_stat;
 422
 423	enum ioc_running		running;
 424	atomic64_t			vtime_rate;
 425	u64				vtime_base_rate;
 426	s64				vtime_err;
 427
 428	seqcount_spinlock_t		period_seqcount;
 429	u64				period_at;	/* wallclock starttime */
 430	u64				period_at_vtime; /* vtime starttime */
 431
 432	atomic64_t			cur_period;	/* inc'd each period */
 433	int				busy_level;	/* saturation history */
 434
 435	bool				weights_updated;
 436	atomic_t			hweight_gen;	/* for lazy hweights */
 437
 438	/* debt forgivness */
 439	u64				dfgv_period_at;
 440	u64				dfgv_period_rem;
 441	u64				dfgv_usage_us_sum;
 442
 443	u64				autop_too_fast_at;
 444	u64				autop_too_slow_at;
 445	int				autop_idx;
 446	bool				user_qos_params:1;
 447	bool				user_cost_model:1;
 448};
 449
 450struct iocg_pcpu_stat {
 451	local64_t			abs_vusage;
 452};
 453
 454struct iocg_stat {
 455	u64				usage_us;
 456	u64				wait_us;
 457	u64				indebt_us;
 458	u64				indelay_us;
 459};
 460
 461/* per device-cgroup pair */
 462struct ioc_gq {
 463	struct blkg_policy_data		pd;
 464	struct ioc			*ioc;
 465
 466	/*
 467	 * A iocg can get its weight from two sources - an explicit
 468	 * per-device-cgroup configuration or the default weight of the
 469	 * cgroup.  `cfg_weight` is the explicit per-device-cgroup
 470	 * configuration.  `weight` is the effective considering both
 471	 * sources.
 472	 *
 473	 * When an idle cgroup becomes active its `active` goes from 0 to
 474	 * `weight`.  `inuse` is the surplus adjusted active weight.
 475	 * `active` and `inuse` are used to calculate `hweight_active` and
 476	 * `hweight_inuse`.
 477	 *
 478	 * `last_inuse` remembers `inuse` while an iocg is idle to persist
 479	 * surplus adjustments.
 480	 *
 481	 * `inuse` may be adjusted dynamically during period. `saved_*` are used
 482	 * to determine and track adjustments.
 483	 */
 484	u32				cfg_weight;
 485	u32				weight;
 486	u32				active;
 487	u32				inuse;
 488
 489	u32				last_inuse;
 490	s64				saved_margin;
 491
 492	sector_t			cursor;		/* to detect randio */
 493
 494	/*
 495	 * `vtime` is this iocg's vtime cursor which progresses as IOs are
 496	 * issued.  If lagging behind device vtime, the delta represents
 497	 * the currently available IO budget.  If running ahead, the
 498	 * overage.
 499	 *
 500	 * `vtime_done` is the same but progressed on completion rather
 501	 * than issue.  The delta behind `vtime` represents the cost of
 502	 * currently in-flight IOs.
 503	 */
 504	atomic64_t			vtime;
 505	atomic64_t			done_vtime;
 506	u64				abs_vdebt;
 507
 508	/* current delay in effect and when it started */
 509	u64				delay;
 510	u64				delay_at;
 511
 512	/*
 513	 * The period this iocg was last active in.  Used for deactivation
 514	 * and invalidating `vtime`.
 515	 */
 516	atomic64_t			active_period;
 517	struct list_head		active_list;
 518
 519	/* see __propagate_weights() and current_hweight() for details */
 520	u64				child_active_sum;
 521	u64				child_inuse_sum;
 522	u64				child_adjusted_sum;
 523	int				hweight_gen;
 524	u32				hweight_active;
 525	u32				hweight_inuse;
 526	u32				hweight_donating;
 527	u32				hweight_after_donation;
 528
 529	struct list_head		walk_list;
 530	struct list_head		surplus_list;
 531
 532	struct wait_queue_head		waitq;
 533	struct hrtimer			waitq_timer;
 534
 535	/* timestamp at the latest activation */
 536	u64				activated_at;
 537
 538	/* statistics */
 539	struct iocg_pcpu_stat __percpu	*pcpu_stat;
 540	struct iocg_stat		stat;
 541	struct iocg_stat		last_stat;
 542	u64				last_stat_abs_vusage;
 543	u64				usage_delta_us;
 544	u64				wait_since;
 545	u64				indebt_since;
 546	u64				indelay_since;
 547
 548	/* this iocg's depth in the hierarchy and ancestors including self */
 549	int				level;
 550	struct ioc_gq			*ancestors[];
 551};
 552
 553/* per cgroup */
 554struct ioc_cgrp {
 555	struct blkcg_policy_data	cpd;
 556	unsigned int			dfl_weight;
 557};
 558
 559struct ioc_now {
 560	u64				now_ns;
 561	u64				now;
 562	u64				vnow;
 563};
 564
 565struct iocg_wait {
 566	struct wait_queue_entry		wait;
 567	struct bio			*bio;
 568	u64				abs_cost;
 569	bool				committed;
 570};
 571
 572struct iocg_wake_ctx {
 573	struct ioc_gq			*iocg;
 574	u32				hw_inuse;
 575	s64				vbudget;
 576};
 577
 578static const struct ioc_params autop[] = {
 579	[AUTOP_HDD] = {
 580		.qos				= {
 581			[QOS_RLAT]		=        250000, /* 250ms */
 582			[QOS_WLAT]		=        250000,
 583			[QOS_MIN]		= VRATE_MIN_PPM,
 584			[QOS_MAX]		= VRATE_MAX_PPM,
 585		},
 586		.i_lcoefs			= {
 587			[I_LCOEF_RBPS]		=     174019176,
 588			[I_LCOEF_RSEQIOPS]	=         41708,
 589			[I_LCOEF_RRANDIOPS]	=           370,
 590			[I_LCOEF_WBPS]		=     178075866,
 591			[I_LCOEF_WSEQIOPS]	=         42705,
 592			[I_LCOEF_WRANDIOPS]	=           378,
 593		},
 594	},
 595	[AUTOP_SSD_QD1] = {
 596		.qos				= {
 597			[QOS_RLAT]		=         25000, /* 25ms */
 598			[QOS_WLAT]		=         25000,
 599			[QOS_MIN]		= VRATE_MIN_PPM,
 600			[QOS_MAX]		= VRATE_MAX_PPM,
 601		},
 602		.i_lcoefs			= {
 603			[I_LCOEF_RBPS]		=     245855193,
 604			[I_LCOEF_RSEQIOPS]	=         61575,
 605			[I_LCOEF_RRANDIOPS]	=          6946,
 606			[I_LCOEF_WBPS]		=     141365009,
 607			[I_LCOEF_WSEQIOPS]	=         33716,
 608			[I_LCOEF_WRANDIOPS]	=         26796,
 609		},
 610	},
 611	[AUTOP_SSD_DFL] = {
 612		.qos				= {
 613			[QOS_RLAT]		=         25000, /* 25ms */
 614			[QOS_WLAT]		=         25000,
 615			[QOS_MIN]		= VRATE_MIN_PPM,
 616			[QOS_MAX]		= VRATE_MAX_PPM,
 617		},
 618		.i_lcoefs			= {
 619			[I_LCOEF_RBPS]		=     488636629,
 620			[I_LCOEF_RSEQIOPS]	=          8932,
 621			[I_LCOEF_RRANDIOPS]	=          8518,
 622			[I_LCOEF_WBPS]		=     427891549,
 623			[I_LCOEF_WSEQIOPS]	=         28755,
 624			[I_LCOEF_WRANDIOPS]	=         21940,
 625		},
 626		.too_fast_vrate_pct		=           500,
 627	},
 628	[AUTOP_SSD_FAST] = {
 629		.qos				= {
 630			[QOS_RLAT]		=          5000, /* 5ms */
 631			[QOS_WLAT]		=          5000,
 632			[QOS_MIN]		= VRATE_MIN_PPM,
 633			[QOS_MAX]		= VRATE_MAX_PPM,
 634		},
 635		.i_lcoefs			= {
 636			[I_LCOEF_RBPS]		=    3102524156LLU,
 637			[I_LCOEF_RSEQIOPS]	=        724816,
 638			[I_LCOEF_RRANDIOPS]	=        778122,
 639			[I_LCOEF_WBPS]		=    1742780862LLU,
 640			[I_LCOEF_WSEQIOPS]	=        425702,
 641			[I_LCOEF_WRANDIOPS]	=	 443193,
 642		},
 643		.too_slow_vrate_pct		=            10,
 644	},
 645};
 646
 647/*
 648 * vrate adjust percentages indexed by ioc->busy_level.  We adjust up on
 649 * vtime credit shortage and down on device saturation.
 650 */
 651static const u32 vrate_adj_pct[] =
 652	{ 0, 0, 0, 0,
 653	  1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
 654	  2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2,
 655	  4, 4, 4, 4, 4, 4, 4, 4, 8, 8, 8, 8, 8, 8, 8, 8, 16 };
 656
 657static struct blkcg_policy blkcg_policy_iocost;
 658
 659/* accessors and helpers */
 660static struct ioc *rqos_to_ioc(struct rq_qos *rqos)
 661{
 662	return container_of(rqos, struct ioc, rqos);
 663}
 664
 665static struct ioc *q_to_ioc(struct request_queue *q)
 666{
 667	return rqos_to_ioc(rq_qos_id(q, RQ_QOS_COST));
 668}
 669
 670static const char __maybe_unused *ioc_name(struct ioc *ioc)
 671{
 672	struct gendisk *disk = ioc->rqos.disk;
 673
 674	if (!disk)
 675		return "<unknown>";
 676	return disk->disk_name;
 677}
 678
 679static struct ioc_gq *pd_to_iocg(struct blkg_policy_data *pd)
 680{
 681	return pd ? container_of(pd, struct ioc_gq, pd) : NULL;
 682}
 683
 684static struct ioc_gq *blkg_to_iocg(struct blkcg_gq *blkg)
 685{
 686	return pd_to_iocg(blkg_to_pd(blkg, &blkcg_policy_iocost));
 687}
 688
 689static struct blkcg_gq *iocg_to_blkg(struct ioc_gq *iocg)
 690{
 691	return pd_to_blkg(&iocg->pd);
 692}
 693
 694static struct ioc_cgrp *blkcg_to_iocc(struct blkcg *blkcg)
 695{
 696	return container_of(blkcg_to_cpd(blkcg, &blkcg_policy_iocost),
 697			    struct ioc_cgrp, cpd);
 698}
 699
 700/*
 701 * Scale @abs_cost to the inverse of @hw_inuse.  The lower the hierarchical
 702 * weight, the more expensive each IO.  Must round up.
 703 */
 704static u64 abs_cost_to_cost(u64 abs_cost, u32 hw_inuse)
 705{
 706	return DIV64_U64_ROUND_UP(abs_cost * WEIGHT_ONE, hw_inuse);
 707}
 708
 709/*
 710 * The inverse of abs_cost_to_cost().  Must round up.
 711 */
 712static u64 cost_to_abs_cost(u64 cost, u32 hw_inuse)
 713{
 714	return DIV64_U64_ROUND_UP(cost * hw_inuse, WEIGHT_ONE);
 715}
 716
 717static void iocg_commit_bio(struct ioc_gq *iocg, struct bio *bio,
 718			    u64 abs_cost, u64 cost)
 719{
 720	struct iocg_pcpu_stat *gcs;
 721
 722	bio->bi_iocost_cost = cost;
 723	atomic64_add(cost, &iocg->vtime);
 724
 725	gcs = get_cpu_ptr(iocg->pcpu_stat);
 726	local64_add(abs_cost, &gcs->abs_vusage);
 727	put_cpu_ptr(gcs);
 728}
 729
 730static void iocg_lock(struct ioc_gq *iocg, bool lock_ioc, unsigned long *flags)
 731{
 732	if (lock_ioc) {
 733		spin_lock_irqsave(&iocg->ioc->lock, *flags);
 734		spin_lock(&iocg->waitq.lock);
 735	} else {
 736		spin_lock_irqsave(&iocg->waitq.lock, *flags);
 737	}
 738}
 739
 740static void iocg_unlock(struct ioc_gq *iocg, bool unlock_ioc, unsigned long *flags)
 741{
 742	if (unlock_ioc) {
 743		spin_unlock(&iocg->waitq.lock);
 744		spin_unlock_irqrestore(&iocg->ioc->lock, *flags);
 745	} else {
 746		spin_unlock_irqrestore(&iocg->waitq.lock, *flags);
 747	}
 748}
 749
 750#define CREATE_TRACE_POINTS
 751#include <trace/events/iocost.h>
 752
 753static void ioc_refresh_margins(struct ioc *ioc)
 754{
 755	struct ioc_margins *margins = &ioc->margins;
 756	u32 period_us = ioc->period_us;
 757	u64 vrate = ioc->vtime_base_rate;
 758
 759	margins->min = (period_us * MARGIN_MIN_PCT / 100) * vrate;
 760	margins->low = (period_us * MARGIN_LOW_PCT / 100) * vrate;
 761	margins->target = (period_us * MARGIN_TARGET_PCT / 100) * vrate;
 762}
 763
 764/* latency Qos params changed, update period_us and all the dependent params */
 765static void ioc_refresh_period_us(struct ioc *ioc)
 766{
 767	u32 ppm, lat, multi, period_us;
 768
 769	lockdep_assert_held(&ioc->lock);
 770
 771	/* pick the higher latency target */
 772	if (ioc->params.qos[QOS_RLAT] >= ioc->params.qos[QOS_WLAT]) {
 773		ppm = ioc->params.qos[QOS_RPPM];
 774		lat = ioc->params.qos[QOS_RLAT];
 775	} else {
 776		ppm = ioc->params.qos[QOS_WPPM];
 777		lat = ioc->params.qos[QOS_WLAT];
 778	}
 779
 780	/*
 781	 * We want the period to be long enough to contain a healthy number
 782	 * of IOs while short enough for granular control.  Define it as a
 783	 * multiple of the latency target.  Ideally, the multiplier should
 784	 * be scaled according to the percentile so that it would nominally
 785	 * contain a certain number of requests.  Let's be simpler and
 786	 * scale it linearly so that it's 2x >= pct(90) and 10x at pct(50).
 787	 */
 788	if (ppm)
 789		multi = max_t(u32, (MILLION - ppm) / 50000, 2);
 790	else
 791		multi = 2;
 792	period_us = multi * lat;
 793	period_us = clamp_t(u32, period_us, MIN_PERIOD, MAX_PERIOD);
 794
 795	/* calculate dependent params */
 796	ioc->period_us = period_us;
 797	ioc->timer_slack_ns = div64_u64(
 798		(u64)period_us * NSEC_PER_USEC * TIMER_SLACK_PCT,
 799		100);
 800	ioc_refresh_margins(ioc);
 801}
 802
 803/*
 804 *  ioc->rqos.disk isn't initialized when this function is called from
 805 *  the init path.
 806 */
 807static int ioc_autop_idx(struct ioc *ioc, struct gendisk *disk)
 808{
 809	int idx = ioc->autop_idx;
 810	const struct ioc_params *p = &autop[idx];
 811	u32 vrate_pct;
 812	u64 now_ns;
 813
 814	/* rotational? */
 815	if (!blk_queue_nonrot(disk->queue))
 816		return AUTOP_HDD;
 817
 818	/* handle SATA SSDs w/ broken NCQ */
 819	if (blk_queue_depth(disk->queue) == 1)
 820		return AUTOP_SSD_QD1;
 821
 822	/* use one of the normal ssd sets */
 823	if (idx < AUTOP_SSD_DFL)
 824		return AUTOP_SSD_DFL;
 825
 826	/* if user is overriding anything, maintain what was there */
 827	if (ioc->user_qos_params || ioc->user_cost_model)
 828		return idx;
 829
 830	/* step up/down based on the vrate */
 831	vrate_pct = div64_u64(ioc->vtime_base_rate * 100, VTIME_PER_USEC);
 832	now_ns = blk_time_get_ns();
 833
 834	if (p->too_fast_vrate_pct && p->too_fast_vrate_pct <= vrate_pct) {
 835		if (!ioc->autop_too_fast_at)
 836			ioc->autop_too_fast_at = now_ns;
 837		if (now_ns - ioc->autop_too_fast_at >= AUTOP_CYCLE_NSEC)
 838			return idx + 1;
 839	} else {
 840		ioc->autop_too_fast_at = 0;
 841	}
 842
 843	if (p->too_slow_vrate_pct && p->too_slow_vrate_pct >= vrate_pct) {
 844		if (!ioc->autop_too_slow_at)
 845			ioc->autop_too_slow_at = now_ns;
 846		if (now_ns - ioc->autop_too_slow_at >= AUTOP_CYCLE_NSEC)
 847			return idx - 1;
 848	} else {
 849		ioc->autop_too_slow_at = 0;
 850	}
 851
 852	return idx;
 853}
 854
 855/*
 856 * Take the followings as input
 857 *
 858 *  @bps	maximum sequential throughput
 859 *  @seqiops	maximum sequential 4k iops
 860 *  @randiops	maximum random 4k iops
 861 *
 862 * and calculate the linear model cost coefficients.
 863 *
 864 *  *@page	per-page cost		1s / (@bps / 4096)
 865 *  *@seqio	base cost of a seq IO	max((1s / @seqiops) - *@page, 0)
 866 *  @randiops	base cost of a rand IO	max((1s / @randiops) - *@page, 0)
 867 */
 868static void calc_lcoefs(u64 bps, u64 seqiops, u64 randiops,
 869			u64 *page, u64 *seqio, u64 *randio)
 870{
 871	u64 v;
 872
 873	*page = *seqio = *randio = 0;
 874
 875	if (bps) {
 876		u64 bps_pages = DIV_ROUND_UP_ULL(bps, IOC_PAGE_SIZE);
 877
 878		if (bps_pages)
 879			*page = DIV64_U64_ROUND_UP(VTIME_PER_SEC, bps_pages);
 880		else
 881			*page = 1;
 882	}
 883
 884	if (seqiops) {
 885		v = DIV64_U64_ROUND_UP(VTIME_PER_SEC, seqiops);
 886		if (v > *page)
 887			*seqio = v - *page;
 888	}
 889
 890	if (randiops) {
 891		v = DIV64_U64_ROUND_UP(VTIME_PER_SEC, randiops);
 892		if (v > *page)
 893			*randio = v - *page;
 894	}
 895}
 896
 897static void ioc_refresh_lcoefs(struct ioc *ioc)
 898{
 899	u64 *u = ioc->params.i_lcoefs;
 900	u64 *c = ioc->params.lcoefs;
 901
 902	calc_lcoefs(u[I_LCOEF_RBPS], u[I_LCOEF_RSEQIOPS], u[I_LCOEF_RRANDIOPS],
 903		    &c[LCOEF_RPAGE], &c[LCOEF_RSEQIO], &c[LCOEF_RRANDIO]);
 904	calc_lcoefs(u[I_LCOEF_WBPS], u[I_LCOEF_WSEQIOPS], u[I_LCOEF_WRANDIOPS],
 905		    &c[LCOEF_WPAGE], &c[LCOEF_WSEQIO], &c[LCOEF_WRANDIO]);
 906}
 907
 908/*
 909 * struct gendisk is required as an argument because ioc->rqos.disk
 910 * is not properly initialized when called from the init path.
 911 */
 912static bool ioc_refresh_params_disk(struct ioc *ioc, bool force,
 913				    struct gendisk *disk)
 914{
 915	const struct ioc_params *p;
 916	int idx;
 917
 918	lockdep_assert_held(&ioc->lock);
 919
 920	idx = ioc_autop_idx(ioc, disk);
 921	p = &autop[idx];
 922
 923	if (idx == ioc->autop_idx && !force)
 924		return false;
 925
 926	if (idx != ioc->autop_idx) {
 927		atomic64_set(&ioc->vtime_rate, VTIME_PER_USEC);
 928		ioc->vtime_base_rate = VTIME_PER_USEC;
 929	}
 930
 931	ioc->autop_idx = idx;
 932	ioc->autop_too_fast_at = 0;
 933	ioc->autop_too_slow_at = 0;
 934
 935	if (!ioc->user_qos_params)
 936		memcpy(ioc->params.qos, p->qos, sizeof(p->qos));
 937	if (!ioc->user_cost_model)
 938		memcpy(ioc->params.i_lcoefs, p->i_lcoefs, sizeof(p->i_lcoefs));
 939
 940	ioc_refresh_period_us(ioc);
 941	ioc_refresh_lcoefs(ioc);
 942
 943	ioc->vrate_min = DIV64_U64_ROUND_UP((u64)ioc->params.qos[QOS_MIN] *
 944					    VTIME_PER_USEC, MILLION);
 945	ioc->vrate_max = DIV64_U64_ROUND_UP((u64)ioc->params.qos[QOS_MAX] *
 946					    VTIME_PER_USEC, MILLION);
 947
 948	return true;
 949}
 950
 951static bool ioc_refresh_params(struct ioc *ioc, bool force)
 952{
 953	return ioc_refresh_params_disk(ioc, force, ioc->rqos.disk);
 954}
 955
 956/*
 957 * When an iocg accumulates too much vtime or gets deactivated, we throw away
 958 * some vtime, which lowers the overall device utilization. As the exact amount
 959 * which is being thrown away is known, we can compensate by accelerating the
 960 * vrate accordingly so that the extra vtime generated in the current period
 961 * matches what got lost.
 962 */
 963static void ioc_refresh_vrate(struct ioc *ioc, struct ioc_now *now)
 964{
 965	s64 pleft = ioc->period_at + ioc->period_us - now->now;
 966	s64 vperiod = ioc->period_us * ioc->vtime_base_rate;
 967	s64 vcomp, vcomp_min, vcomp_max;
 968
 969	lockdep_assert_held(&ioc->lock);
 970
 971	/* we need some time left in this period */
 972	if (pleft <= 0)
 973		goto done;
 974
 975	/*
 976	 * Calculate how much vrate should be adjusted to offset the error.
 977	 * Limit the amount of adjustment and deduct the adjusted amount from
 978	 * the error.
 979	 */
 980	vcomp = -div64_s64(ioc->vtime_err, pleft);
 981	vcomp_min = -(ioc->vtime_base_rate >> 1);
 982	vcomp_max = ioc->vtime_base_rate;
 983	vcomp = clamp(vcomp, vcomp_min, vcomp_max);
 984
 985	ioc->vtime_err += vcomp * pleft;
 986
 987	atomic64_set(&ioc->vtime_rate, ioc->vtime_base_rate + vcomp);
 988done:
 989	/* bound how much error can accumulate */
 990	ioc->vtime_err = clamp(ioc->vtime_err, -vperiod, vperiod);
 991}
 992
 993static void ioc_adjust_base_vrate(struct ioc *ioc, u32 rq_wait_pct,
 994				  int nr_lagging, int nr_shortages,
 995				  int prev_busy_level, u32 *missed_ppm)
 996{
 997	u64 vrate = ioc->vtime_base_rate;
 998	u64 vrate_min = ioc->vrate_min, vrate_max = ioc->vrate_max;
 999
1000	if (!ioc->busy_level || (ioc->busy_level < 0 && nr_lagging)) {
1001		if (ioc->busy_level != prev_busy_level || nr_lagging)
1002			trace_iocost_ioc_vrate_adj(ioc, vrate,
1003						   missed_ppm, rq_wait_pct,
1004						   nr_lagging, nr_shortages);
1005
1006		return;
1007	}
1008
1009	/*
1010	 * If vrate is out of bounds, apply clamp gradually as the
1011	 * bounds can change abruptly.  Otherwise, apply busy_level
1012	 * based adjustment.
1013	 */
1014	if (vrate < vrate_min) {
1015		vrate = div64_u64(vrate * (100 + VRATE_CLAMP_ADJ_PCT), 100);
1016		vrate = min(vrate, vrate_min);
1017	} else if (vrate > vrate_max) {
1018		vrate = div64_u64(vrate * (100 - VRATE_CLAMP_ADJ_PCT), 100);
1019		vrate = max(vrate, vrate_max);
1020	} else {
1021		int idx = min_t(int, abs(ioc->busy_level),
1022				ARRAY_SIZE(vrate_adj_pct) - 1);
1023		u32 adj_pct = vrate_adj_pct[idx];
1024
1025		if (ioc->busy_level > 0)
1026			adj_pct = 100 - adj_pct;
1027		else
1028			adj_pct = 100 + adj_pct;
1029
1030		vrate = clamp(DIV64_U64_ROUND_UP(vrate * adj_pct, 100),
1031			      vrate_min, vrate_max);
1032	}
1033
1034	trace_iocost_ioc_vrate_adj(ioc, vrate, missed_ppm, rq_wait_pct,
1035				   nr_lagging, nr_shortages);
1036
1037	ioc->vtime_base_rate = vrate;
1038	ioc_refresh_margins(ioc);
1039}
1040
1041/* take a snapshot of the current [v]time and vrate */
1042static void ioc_now(struct ioc *ioc, struct ioc_now *now)
1043{
1044	unsigned seq;
1045	u64 vrate;
1046
1047	now->now_ns = blk_time_get_ns();
1048	now->now = ktime_to_us(now->now_ns);
1049	vrate = atomic64_read(&ioc->vtime_rate);
1050
1051	/*
1052	 * The current vtime is
1053	 *
1054	 *   vtime at period start + (wallclock time since the start) * vrate
1055	 *
1056	 * As a consistent snapshot of `period_at_vtime` and `period_at` is
1057	 * needed, they're seqcount protected.
1058	 */
1059	do {
1060		seq = read_seqcount_begin(&ioc->period_seqcount);
1061		now->vnow = ioc->period_at_vtime +
1062			(now->now - ioc->period_at) * vrate;
1063	} while (read_seqcount_retry(&ioc->period_seqcount, seq));
1064}
1065
1066static void ioc_start_period(struct ioc *ioc, struct ioc_now *now)
1067{
1068	WARN_ON_ONCE(ioc->running != IOC_RUNNING);
1069
1070	write_seqcount_begin(&ioc->period_seqcount);
1071	ioc->period_at = now->now;
1072	ioc->period_at_vtime = now->vnow;
1073	write_seqcount_end(&ioc->period_seqcount);
1074
1075	ioc->timer.expires = jiffies + usecs_to_jiffies(ioc->period_us);
1076	add_timer(&ioc->timer);
1077}
1078
1079/*
1080 * Update @iocg's `active` and `inuse` to @active and @inuse, update level
1081 * weight sums and propagate upwards accordingly. If @save, the current margin
1082 * is saved to be used as reference for later inuse in-period adjustments.
1083 */
1084static void __propagate_weights(struct ioc_gq *iocg, u32 active, u32 inuse,
1085				bool save, struct ioc_now *now)
1086{
1087	struct ioc *ioc = iocg->ioc;
1088	int lvl;
1089
1090	lockdep_assert_held(&ioc->lock);
1091
1092	/*
1093	 * For an active leaf node, its inuse shouldn't be zero or exceed
1094	 * @active. An active internal node's inuse is solely determined by the
1095	 * inuse to active ratio of its children regardless of @inuse.
1096	 */
1097	if (list_empty(&iocg->active_list) && iocg->child_active_sum) {
1098		inuse = DIV64_U64_ROUND_UP(active * iocg->child_inuse_sum,
1099					   iocg->child_active_sum);
1100	} else {
1101		/*
1102		 * It may be tempting to turn this into a clamp expression with
1103		 * a lower limit of 1 but active may be 0, which cannot be used
1104		 * as an upper limit in that situation. This expression allows
1105		 * active to clamp inuse unless it is 0, in which case inuse
1106		 * becomes 1.
1107		 */
1108		inuse = min(inuse, active) ?: 1;
1109	}
1110
1111	iocg->last_inuse = iocg->inuse;
1112	if (save)
1113		iocg->saved_margin = now->vnow - atomic64_read(&iocg->vtime);
1114
1115	if (active == iocg->active && inuse == iocg->inuse)
1116		return;
1117
1118	for (lvl = iocg->level - 1; lvl >= 0; lvl--) {
1119		struct ioc_gq *parent = iocg->ancestors[lvl];
1120		struct ioc_gq *child = iocg->ancestors[lvl + 1];
1121		u32 parent_active = 0, parent_inuse = 0;
1122
1123		/* update the level sums */
1124		parent->child_active_sum += (s32)(active - child->active);
1125		parent->child_inuse_sum += (s32)(inuse - child->inuse);
1126		/* apply the updates */
1127		child->active = active;
1128		child->inuse = inuse;
1129
1130		/*
1131		 * The delta between inuse and active sums indicates that
1132		 * much of weight is being given away.  Parent's inuse
1133		 * and active should reflect the ratio.
1134		 */
1135		if (parent->child_active_sum) {
1136			parent_active = parent->weight;
1137			parent_inuse = DIV64_U64_ROUND_UP(
1138				parent_active * parent->child_inuse_sum,
1139				parent->child_active_sum);
1140		}
1141
1142		/* do we need to keep walking up? */
1143		if (parent_active == parent->active &&
1144		    parent_inuse == parent->inuse)
1145			break;
1146
1147		active = parent_active;
1148		inuse = parent_inuse;
1149	}
1150
1151	ioc->weights_updated = true;
1152}
1153
1154static void commit_weights(struct ioc *ioc)
1155{
1156	lockdep_assert_held(&ioc->lock);
1157
1158	if (ioc->weights_updated) {
1159		/* paired with rmb in current_hweight(), see there */
1160		smp_wmb();
1161		atomic_inc(&ioc->hweight_gen);
1162		ioc->weights_updated = false;
1163	}
1164}
1165
1166static void propagate_weights(struct ioc_gq *iocg, u32 active, u32 inuse,
1167			      bool save, struct ioc_now *now)
1168{
1169	__propagate_weights(iocg, active, inuse, save, now);
1170	commit_weights(iocg->ioc);
1171}
1172
1173static void current_hweight(struct ioc_gq *iocg, u32 *hw_activep, u32 *hw_inusep)
1174{
1175	struct ioc *ioc = iocg->ioc;
1176	int lvl;
1177	u32 hwa, hwi;
1178	int ioc_gen;
1179
1180	/* hot path - if uptodate, use cached */
1181	ioc_gen = atomic_read(&ioc->hweight_gen);
1182	if (ioc_gen == iocg->hweight_gen)
1183		goto out;
1184
1185	/*
1186	 * Paired with wmb in commit_weights(). If we saw the updated
1187	 * hweight_gen, all the weight updates from __propagate_weights() are
1188	 * visible too.
1189	 *
1190	 * We can race with weight updates during calculation and get it
1191	 * wrong.  However, hweight_gen would have changed and a future
1192	 * reader will recalculate and we're guaranteed to discard the
1193	 * wrong result soon.
1194	 */
1195	smp_rmb();
1196
1197	hwa = hwi = WEIGHT_ONE;
1198	for (lvl = 0; lvl <= iocg->level - 1; lvl++) {
1199		struct ioc_gq *parent = iocg->ancestors[lvl];
1200		struct ioc_gq *child = iocg->ancestors[lvl + 1];
1201		u64 active_sum = READ_ONCE(parent->child_active_sum);
1202		u64 inuse_sum = READ_ONCE(parent->child_inuse_sum);
1203		u32 active = READ_ONCE(child->active);
1204		u32 inuse = READ_ONCE(child->inuse);
1205
1206		/* we can race with deactivations and either may read as zero */
1207		if (!active_sum || !inuse_sum)
1208			continue;
1209
1210		active_sum = max_t(u64, active, active_sum);
1211		hwa = div64_u64((u64)hwa * active, active_sum);
1212
1213		inuse_sum = max_t(u64, inuse, inuse_sum);
1214		hwi = div64_u64((u64)hwi * inuse, inuse_sum);
1215	}
1216
1217	iocg->hweight_active = max_t(u32, hwa, 1);
1218	iocg->hweight_inuse = max_t(u32, hwi, 1);
1219	iocg->hweight_gen = ioc_gen;
1220out:
1221	if (hw_activep)
1222		*hw_activep = iocg->hweight_active;
1223	if (hw_inusep)
1224		*hw_inusep = iocg->hweight_inuse;
1225}
1226
1227/*
1228 * Calculate the hweight_inuse @iocg would get with max @inuse assuming all the
1229 * other weights stay unchanged.
1230 */
1231static u32 current_hweight_max(struct ioc_gq *iocg)
1232{
1233	u32 hwm = WEIGHT_ONE;
1234	u32 inuse = iocg->active;
1235	u64 child_inuse_sum;
1236	int lvl;
1237
1238	lockdep_assert_held(&iocg->ioc->lock);
1239
1240	for (lvl = iocg->level - 1; lvl >= 0; lvl--) {
1241		struct ioc_gq *parent = iocg->ancestors[lvl];
1242		struct ioc_gq *child = iocg->ancestors[lvl + 1];
1243
1244		child_inuse_sum = parent->child_inuse_sum + inuse - child->inuse;
1245		hwm = div64_u64((u64)hwm * inuse, child_inuse_sum);
1246		inuse = DIV64_U64_ROUND_UP(parent->active * child_inuse_sum,
1247					   parent->child_active_sum);
1248	}
1249
1250	return max_t(u32, hwm, 1);
1251}
1252
1253static void weight_updated(struct ioc_gq *iocg, struct ioc_now *now)
1254{
1255	struct ioc *ioc = iocg->ioc;
1256	struct blkcg_gq *blkg = iocg_to_blkg(iocg);
1257	struct ioc_cgrp *iocc = blkcg_to_iocc(blkg->blkcg);
1258	u32 weight;
1259
1260	lockdep_assert_held(&ioc->lock);
1261
1262	weight = iocg->cfg_weight ?: iocc->dfl_weight;
1263	if (weight != iocg->weight && iocg->active)
1264		propagate_weights(iocg, weight, iocg->inuse, true, now);
1265	iocg->weight = weight;
1266}
1267
1268static bool iocg_activate(struct ioc_gq *iocg, struct ioc_now *now)
1269{
1270	struct ioc *ioc = iocg->ioc;
1271	u64 __maybe_unused last_period, cur_period;
1272	u64 vtime, vtarget;
1273	int i;
1274
1275	/*
1276	 * If seem to be already active, just update the stamp to tell the
1277	 * timer that we're still active.  We don't mind occassional races.
1278	 */
1279	if (!list_empty(&iocg->active_list)) {
1280		ioc_now(ioc, now);
1281		cur_period = atomic64_read(&ioc->cur_period);
1282		if (atomic64_read(&iocg->active_period) != cur_period)
1283			atomic64_set(&iocg->active_period, cur_period);
1284		return true;
1285	}
1286
1287	/* racy check on internal node IOs, treat as root level IOs */
1288	if (iocg->child_active_sum)
1289		return false;
1290
1291	spin_lock_irq(&ioc->lock);
1292
1293	ioc_now(ioc, now);
1294
1295	/* update period */
1296	cur_period = atomic64_read(&ioc->cur_period);
1297	last_period = atomic64_read(&iocg->active_period);
1298	atomic64_set(&iocg->active_period, cur_period);
1299
1300	/* already activated or breaking leaf-only constraint? */
1301	if (!list_empty(&iocg->active_list))
1302		goto succeed_unlock;
1303	for (i = iocg->level - 1; i > 0; i--)
1304		if (!list_empty(&iocg->ancestors[i]->active_list))
1305			goto fail_unlock;
1306
1307	if (iocg->child_active_sum)
1308		goto fail_unlock;
1309
1310	/*
1311	 * Always start with the target budget. On deactivation, we throw away
1312	 * anything above it.
1313	 */
1314	vtarget = now->vnow - ioc->margins.target;
1315	vtime = atomic64_read(&iocg->vtime);
1316
1317	atomic64_add(vtarget - vtime, &iocg->vtime);
1318	atomic64_add(vtarget - vtime, &iocg->done_vtime);
1319	vtime = vtarget;
1320
1321	/*
1322	 * Activate, propagate weight and start period timer if not
1323	 * running.  Reset hweight_gen to avoid accidental match from
1324	 * wrapping.
1325	 */
1326	iocg->hweight_gen = atomic_read(&ioc->hweight_gen) - 1;
1327	list_add(&iocg->active_list, &ioc->active_iocgs);
1328
1329	propagate_weights(iocg, iocg->weight,
1330			  iocg->last_inuse ?: iocg->weight, true, now);
1331
1332	TRACE_IOCG_PATH(iocg_activate, iocg, now,
1333			last_period, cur_period, vtime);
1334
1335	iocg->activated_at = now->now;
1336
1337	if (ioc->running == IOC_IDLE) {
1338		ioc->running = IOC_RUNNING;
1339		ioc->dfgv_period_at = now->now;
1340		ioc->dfgv_period_rem = 0;
1341		ioc_start_period(ioc, now);
1342	}
1343
1344succeed_unlock:
1345	spin_unlock_irq(&ioc->lock);
1346	return true;
1347
1348fail_unlock:
1349	spin_unlock_irq(&ioc->lock);
1350	return false;
1351}
1352
1353static bool iocg_kick_delay(struct ioc_gq *iocg, struct ioc_now *now)
1354{
1355	struct ioc *ioc = iocg->ioc;
1356	struct blkcg_gq *blkg = iocg_to_blkg(iocg);
1357	u64 tdelta, delay, new_delay, shift;
1358	s64 vover, vover_pct;
1359	u32 hwa;
1360
1361	lockdep_assert_held(&iocg->waitq.lock);
1362
1363	/*
1364	 * If the delay is set by another CPU, we may be in the past. No need to
1365	 * change anything if so. This avoids decay calculation underflow.
1366	 */
1367	if (time_before64(now->now, iocg->delay_at))
1368		return false;
1369
1370	/* calculate the current delay in effect - 1/2 every second */
1371	tdelta = now->now - iocg->delay_at;
1372	shift = div64_u64(tdelta, USEC_PER_SEC);
1373	if (iocg->delay && shift < BITS_PER_LONG)
1374		delay = iocg->delay >> shift;
1375	else
1376		delay = 0;
1377
1378	/* calculate the new delay from the debt amount */
1379	current_hweight(iocg, &hwa, NULL);
1380	vover = atomic64_read(&iocg->vtime) +
1381		abs_cost_to_cost(iocg->abs_vdebt, hwa) - now->vnow;
1382	vover_pct = div64_s64(100 * vover,
1383			      ioc->period_us * ioc->vtime_base_rate);
1384
1385	if (vover_pct <= MIN_DELAY_THR_PCT)
1386		new_delay = 0;
1387	else if (vover_pct >= MAX_DELAY_THR_PCT)
1388		new_delay = MAX_DELAY;
1389	else
1390		new_delay = MIN_DELAY +
1391			div_u64((MAX_DELAY - MIN_DELAY) *
1392				(vover_pct - MIN_DELAY_THR_PCT),
1393				MAX_DELAY_THR_PCT - MIN_DELAY_THR_PCT);
1394
1395	/* pick the higher one and apply */
1396	if (new_delay > delay) {
1397		iocg->delay = new_delay;
1398		iocg->delay_at = now->now;
1399		delay = new_delay;
1400	}
1401
1402	if (delay >= MIN_DELAY) {
1403		if (!iocg->indelay_since)
1404			iocg->indelay_since = now->now;
1405		blkcg_set_delay(blkg, delay * NSEC_PER_USEC);
1406		return true;
1407	} else {
1408		if (iocg->indelay_since) {
1409			iocg->stat.indelay_us += now->now - iocg->indelay_since;
1410			iocg->indelay_since = 0;
1411		}
1412		iocg->delay = 0;
1413		blkcg_clear_delay(blkg);
1414		return false;
1415	}
1416}
1417
1418static void iocg_incur_debt(struct ioc_gq *iocg, u64 abs_cost,
1419			    struct ioc_now *now)
1420{
1421	struct iocg_pcpu_stat *gcs;
1422
1423	lockdep_assert_held(&iocg->ioc->lock);
1424	lockdep_assert_held(&iocg->waitq.lock);
1425	WARN_ON_ONCE(list_empty(&iocg->active_list));
1426
1427	/*
1428	 * Once in debt, debt handling owns inuse. @iocg stays at the minimum
1429	 * inuse donating all of it share to others until its debt is paid off.
1430	 */
1431	if (!iocg->abs_vdebt && abs_cost) {
1432		iocg->indebt_since = now->now;
1433		propagate_weights(iocg, iocg->active, 0, false, now);
1434	}
1435
1436	iocg->abs_vdebt += abs_cost;
1437
1438	gcs = get_cpu_ptr(iocg->pcpu_stat);
1439	local64_add(abs_cost, &gcs->abs_vusage);
1440	put_cpu_ptr(gcs);
1441}
1442
1443static void iocg_pay_debt(struct ioc_gq *iocg, u64 abs_vpay,
1444			  struct ioc_now *now)
1445{
1446	lockdep_assert_held(&iocg->ioc->lock);
1447	lockdep_assert_held(&iocg->waitq.lock);
1448
1449	/*
1450	 * make sure that nobody messed with @iocg. Check iocg->pd.online
1451	 * to avoid warn when removing blkcg or disk.
1452	 */
1453	WARN_ON_ONCE(list_empty(&iocg->active_list) && iocg->pd.online);
1454	WARN_ON_ONCE(iocg->inuse > 1);
1455
1456	iocg->abs_vdebt -= min(abs_vpay, iocg->abs_vdebt);
1457
1458	/* if debt is paid in full, restore inuse */
1459	if (!iocg->abs_vdebt) {
1460		iocg->stat.indebt_us += now->now - iocg->indebt_since;
1461		iocg->indebt_since = 0;
1462
1463		propagate_weights(iocg, iocg->active, iocg->last_inuse,
1464				  false, now);
1465	}
1466}
1467
1468static int iocg_wake_fn(struct wait_queue_entry *wq_entry, unsigned mode,
1469			int flags, void *key)
1470{
1471	struct iocg_wait *wait = container_of(wq_entry, struct iocg_wait, wait);
1472	struct iocg_wake_ctx *ctx = key;
1473	u64 cost = abs_cost_to_cost(wait->abs_cost, ctx->hw_inuse);
1474
1475	ctx->vbudget -= cost;
1476
1477	if (ctx->vbudget < 0)
1478		return -1;
1479
1480	iocg_commit_bio(ctx->iocg, wait->bio, wait->abs_cost, cost);
1481	wait->committed = true;
1482
1483	/*
1484	 * autoremove_wake_function() removes the wait entry only when it
1485	 * actually changed the task state. We want the wait always removed.
1486	 * Remove explicitly and use default_wake_function(). Note that the
1487	 * order of operations is important as finish_wait() tests whether
1488	 * @wq_entry is removed without grabbing the lock.
1489	 */
1490	default_wake_function(wq_entry, mode, flags, key);
1491	list_del_init_careful(&wq_entry->entry);
1492	return 0;
1493}
1494
1495/*
1496 * Calculate the accumulated budget, pay debt if @pay_debt and wake up waiters
1497 * accordingly. When @pay_debt is %true, the caller must be holding ioc->lock in
1498 * addition to iocg->waitq.lock.
1499 */
1500static void iocg_kick_waitq(struct ioc_gq *iocg, bool pay_debt,
1501			    struct ioc_now *now)
1502{
1503	struct ioc *ioc = iocg->ioc;
1504	struct iocg_wake_ctx ctx = { .iocg = iocg };
1505	u64 vshortage, expires, oexpires;
1506	s64 vbudget;
1507	u32 hwa;
1508
1509	lockdep_assert_held(&iocg->waitq.lock);
1510
1511	current_hweight(iocg, &hwa, NULL);
1512	vbudget = now->vnow - atomic64_read(&iocg->vtime);
1513
1514	/* pay off debt */
1515	if (pay_debt && iocg->abs_vdebt && vbudget > 0) {
1516		u64 abs_vbudget = cost_to_abs_cost(vbudget, hwa);
1517		u64 abs_vpay = min_t(u64, abs_vbudget, iocg->abs_vdebt);
1518		u64 vpay = abs_cost_to_cost(abs_vpay, hwa);
1519
1520		lockdep_assert_held(&ioc->lock);
1521
1522		atomic64_add(vpay, &iocg->vtime);
1523		atomic64_add(vpay, &iocg->done_vtime);
1524		iocg_pay_debt(iocg, abs_vpay, now);
1525		vbudget -= vpay;
1526	}
1527
1528	if (iocg->abs_vdebt || iocg->delay)
1529		iocg_kick_delay(iocg, now);
1530
1531	/*
1532	 * Debt can still be outstanding if we haven't paid all yet or the
1533	 * caller raced and called without @pay_debt. Shouldn't wake up waiters
1534	 * under debt. Make sure @vbudget reflects the outstanding amount and is
1535	 * not positive.
1536	 */
1537	if (iocg->abs_vdebt) {
1538		s64 vdebt = abs_cost_to_cost(iocg->abs_vdebt, hwa);
1539		vbudget = min_t(s64, 0, vbudget - vdebt);
1540	}
1541
1542	/*
1543	 * Wake up the ones which are due and see how much vtime we'll need for
1544	 * the next one. As paying off debt restores hw_inuse, it must be read
1545	 * after the above debt payment.
1546	 */
1547	ctx.vbudget = vbudget;
1548	current_hweight(iocg, NULL, &ctx.hw_inuse);
1549
1550	__wake_up_locked_key(&iocg->waitq, TASK_NORMAL, &ctx);
1551
1552	if (!waitqueue_active(&iocg->waitq)) {
1553		if (iocg->wait_since) {
1554			iocg->stat.wait_us += now->now - iocg->wait_since;
1555			iocg->wait_since = 0;
1556		}
1557		return;
1558	}
1559
1560	if (!iocg->wait_since)
1561		iocg->wait_since = now->now;
1562
1563	if (WARN_ON_ONCE(ctx.vbudget >= 0))
1564		return;
1565
1566	/* determine next wakeup, add a timer margin to guarantee chunking */
1567	vshortage = -ctx.vbudget;
1568	expires = now->now_ns +
1569		DIV64_U64_ROUND_UP(vshortage, ioc->vtime_base_rate) *
1570		NSEC_PER_USEC;
1571	expires += ioc->timer_slack_ns;
1572
1573	/* if already active and close enough, don't bother */
1574	oexpires = ktime_to_ns(hrtimer_get_softexpires(&iocg->waitq_timer));
1575	if (hrtimer_is_queued(&iocg->waitq_timer) &&
1576	    abs(oexpires - expires) <= ioc->timer_slack_ns)
1577		return;
1578
1579	hrtimer_start_range_ns(&iocg->waitq_timer, ns_to_ktime(expires),
1580			       ioc->timer_slack_ns, HRTIMER_MODE_ABS);
1581}
1582
1583static enum hrtimer_restart iocg_waitq_timer_fn(struct hrtimer *timer)
1584{
1585	struct ioc_gq *iocg = container_of(timer, struct ioc_gq, waitq_timer);
1586	bool pay_debt = READ_ONCE(iocg->abs_vdebt);
1587	struct ioc_now now;
1588	unsigned long flags;
1589
1590	ioc_now(iocg->ioc, &now);
1591
1592	iocg_lock(iocg, pay_debt, &flags);
1593	iocg_kick_waitq(iocg, pay_debt, &now);
1594	iocg_unlock(iocg, pay_debt, &flags);
1595
1596	return HRTIMER_NORESTART;
1597}
1598
1599static void ioc_lat_stat(struct ioc *ioc, u32 *missed_ppm_ar, u32 *rq_wait_pct_p)
1600{
1601	u32 nr_met[2] = { };
1602	u32 nr_missed[2] = { };
1603	u64 rq_wait_ns = 0;
1604	int cpu, rw;
1605
1606	for_each_online_cpu(cpu) {
1607		struct ioc_pcpu_stat *stat = per_cpu_ptr(ioc->pcpu_stat, cpu);
1608		u64 this_rq_wait_ns;
1609
1610		for (rw = READ; rw <= WRITE; rw++) {
1611			u32 this_met = local_read(&stat->missed[rw].nr_met);
1612			u32 this_missed = local_read(&stat->missed[rw].nr_missed);
1613
1614			nr_met[rw] += this_met - stat->missed[rw].last_met;
1615			nr_missed[rw] += this_missed - stat->missed[rw].last_missed;
1616			stat->missed[rw].last_met = this_met;
1617			stat->missed[rw].last_missed = this_missed;
1618		}
1619
1620		this_rq_wait_ns = local64_read(&stat->rq_wait_ns);
1621		rq_wait_ns += this_rq_wait_ns - stat->last_rq_wait_ns;
1622		stat->last_rq_wait_ns = this_rq_wait_ns;
1623	}
1624
1625	for (rw = READ; rw <= WRITE; rw++) {
1626		if (nr_met[rw] + nr_missed[rw])
1627			missed_ppm_ar[rw] =
1628				DIV64_U64_ROUND_UP((u64)nr_missed[rw] * MILLION,
1629						   nr_met[rw] + nr_missed[rw]);
1630		else
1631			missed_ppm_ar[rw] = 0;
1632	}
1633
1634	*rq_wait_pct_p = div64_u64(rq_wait_ns * 100,
1635				   ioc->period_us * NSEC_PER_USEC);
1636}
1637
1638/* was iocg idle this period? */
1639static bool iocg_is_idle(struct ioc_gq *iocg)
1640{
1641	struct ioc *ioc = iocg->ioc;
1642
1643	/* did something get issued this period? */
1644	if (atomic64_read(&iocg->active_period) ==
1645	    atomic64_read(&ioc->cur_period))
1646		return false;
1647
1648	/* is something in flight? */
1649	if (atomic64_read(&iocg->done_vtime) != atomic64_read(&iocg->vtime))
1650		return false;
1651
1652	return true;
1653}
1654
1655/*
1656 * Call this function on the target leaf @iocg's to build pre-order traversal
1657 * list of all the ancestors in @inner_walk. The inner nodes are linked through
1658 * ->walk_list and the caller is responsible for dissolving the list after use.
1659 */
1660static void iocg_build_inner_walk(struct ioc_gq *iocg,
1661				  struct list_head *inner_walk)
1662{
1663	int lvl;
1664
1665	WARN_ON_ONCE(!list_empty(&iocg->walk_list));
1666
1667	/* find the first ancestor which hasn't been visited yet */
1668	for (lvl = iocg->level - 1; lvl >= 0; lvl--) {
1669		if (!list_empty(&iocg->ancestors[lvl]->walk_list))
1670			break;
1671	}
1672
1673	/* walk down and visit the inner nodes to get pre-order traversal */
1674	while (++lvl <= iocg->level - 1) {
1675		struct ioc_gq *inner = iocg->ancestors[lvl];
1676
1677		/* record traversal order */
1678		list_add_tail(&inner->walk_list, inner_walk);
1679	}
1680}
1681
1682/* propagate the deltas to the parent */
1683static void iocg_flush_stat_upward(struct ioc_gq *iocg)
1684{
1685	if (iocg->level > 0) {
1686		struct iocg_stat *parent_stat =
1687			&iocg->ancestors[iocg->level - 1]->stat;
1688
1689		parent_stat->usage_us +=
1690			iocg->stat.usage_us - iocg->last_stat.usage_us;
1691		parent_stat->wait_us +=
1692			iocg->stat.wait_us - iocg->last_stat.wait_us;
1693		parent_stat->indebt_us +=
1694			iocg->stat.indebt_us - iocg->last_stat.indebt_us;
1695		parent_stat->indelay_us +=
1696			iocg->stat.indelay_us - iocg->last_stat.indelay_us;
1697	}
1698
1699	iocg->last_stat = iocg->stat;
1700}
1701
1702/* collect per-cpu counters and propagate the deltas to the parent */
1703static void iocg_flush_stat_leaf(struct ioc_gq *iocg, struct ioc_now *now)
1704{
1705	struct ioc *ioc = iocg->ioc;
1706	u64 abs_vusage = 0;
1707	u64 vusage_delta;
1708	int cpu;
1709
1710	lockdep_assert_held(&iocg->ioc->lock);
1711
1712	/* collect per-cpu counters */
1713	for_each_possible_cpu(cpu) {
1714		abs_vusage += local64_read(
1715				per_cpu_ptr(&iocg->pcpu_stat->abs_vusage, cpu));
1716	}
1717	vusage_delta = abs_vusage - iocg->last_stat_abs_vusage;
1718	iocg->last_stat_abs_vusage = abs_vusage;
1719
1720	iocg->usage_delta_us = div64_u64(vusage_delta, ioc->vtime_base_rate);
1721	iocg->stat.usage_us += iocg->usage_delta_us;
1722
1723	iocg_flush_stat_upward(iocg);
1724}
1725
1726/* get stat counters ready for reading on all active iocgs */
1727static void iocg_flush_stat(struct list_head *target_iocgs, struct ioc_now *now)
1728{
1729	LIST_HEAD(inner_walk);
1730	struct ioc_gq *iocg, *tiocg;
1731
1732	/* flush leaves and build inner node walk list */
1733	list_for_each_entry(iocg, target_iocgs, active_list) {
1734		iocg_flush_stat_leaf(iocg, now);
1735		iocg_build_inner_walk(iocg, &inner_walk);
1736	}
1737
1738	/* keep flushing upwards by walking the inner list backwards */
1739	list_for_each_entry_safe_reverse(iocg, tiocg, &inner_walk, walk_list) {
1740		iocg_flush_stat_upward(iocg);
1741		list_del_init(&iocg->walk_list);
1742	}
1743}
1744
1745/*
1746 * Determine what @iocg's hweight_inuse should be after donating unused
1747 * capacity. @hwm is the upper bound and used to signal no donation. This
1748 * function also throws away @iocg's excess budget.
1749 */
1750static u32 hweight_after_donation(struct ioc_gq *iocg, u32 old_hwi, u32 hwm,
1751				  u32 usage, struct ioc_now *now)
1752{
1753	struct ioc *ioc = iocg->ioc;
1754	u64 vtime = atomic64_read(&iocg->vtime);
1755	s64 excess, delta, target, new_hwi;
1756
1757	/* debt handling owns inuse for debtors */
1758	if (iocg->abs_vdebt)
1759		return 1;
1760
1761	/* see whether minimum margin requirement is met */
1762	if (waitqueue_active(&iocg->waitq) ||
1763	    time_after64(vtime, now->vnow - ioc->margins.min))
1764		return hwm;
1765
1766	/* throw away excess above target */
1767	excess = now->vnow - vtime - ioc->margins.target;
1768	if (excess > 0) {
1769		atomic64_add(excess, &iocg->vtime);
1770		atomic64_add(excess, &iocg->done_vtime);
1771		vtime += excess;
1772		ioc->vtime_err -= div64_u64(excess * old_hwi, WEIGHT_ONE);
1773	}
1774
1775	/*
1776	 * Let's say the distance between iocg's and device's vtimes as a
1777	 * fraction of period duration is delta. Assuming that the iocg will
1778	 * consume the usage determined above, we want to determine new_hwi so
1779	 * that delta equals MARGIN_TARGET at the end of the next period.
1780	 *
1781	 * We need to execute usage worth of IOs while spending the sum of the
1782	 * new budget (1 - MARGIN_TARGET) and the leftover from the last period
1783	 * (delta):
1784	 *
1785	 *   usage = (1 - MARGIN_TARGET + delta) * new_hwi
1786	 *
1787	 * Therefore, the new_hwi is:
1788	 *
1789	 *   new_hwi = usage / (1 - MARGIN_TARGET + delta)
1790	 */
1791	delta = div64_s64(WEIGHT_ONE * (now->vnow - vtime),
1792			  now->vnow - ioc->period_at_vtime);
1793	target = WEIGHT_ONE * MARGIN_TARGET_PCT / 100;
1794	new_hwi = div64_s64(WEIGHT_ONE * usage, WEIGHT_ONE - target + delta);
1795
1796	return clamp_t(s64, new_hwi, 1, hwm);
1797}
1798
1799/*
1800 * For work-conservation, an iocg which isn't using all of its share should
1801 * donate the leftover to other iocgs. There are two ways to achieve this - 1.
1802 * bumping up vrate accordingly 2. lowering the donating iocg's inuse weight.
1803 *
1804 * #1 is mathematically simpler but has the drawback of requiring synchronous
1805 * global hweight_inuse updates when idle iocg's get activated or inuse weights
1806 * change due to donation snapbacks as it has the possibility of grossly
1807 * overshooting what's allowed by the model and vrate.
1808 *
1809 * #2 is inherently safe with local operations. The donating iocg can easily
1810 * snap back to higher weights when needed without worrying about impacts on
1811 * other nodes as the impacts will be inherently correct. This also makes idle
1812 * iocg activations safe. The only effect activations have is decreasing
1813 * hweight_inuse of others, the right solution to which is for those iocgs to
1814 * snap back to higher weights.
1815 *
1816 * So, we go with #2. The challenge is calculating how each donating iocg's
1817 * inuse should be adjusted to achieve the target donation amounts. This is done
1818 * using Andy's method described in the following pdf.
1819 *
1820 *   https://drive.google.com/file/d/1PsJwxPFtjUnwOY1QJ5AeICCcsL7BM3bo
1821 *
1822 * Given the weights and target after-donation hweight_inuse values, Andy's
1823 * method determines how the proportional distribution should look like at each
1824 * sibling level to maintain the relative relationship between all non-donating
1825 * pairs. To roughly summarize, it divides the tree into donating and
1826 * non-donating parts, calculates global donation rate which is used to
1827 * determine the target hweight_inuse for each node, and then derives per-level
1828 * proportions.
1829 *
1830 * The following pdf shows that global distribution calculated this way can be
1831 * achieved by scaling inuse weights of donating leaves and propagating the
1832 * adjustments upwards proportionally.
1833 *
1834 *   https://drive.google.com/file/d/1vONz1-fzVO7oY5DXXsLjSxEtYYQbOvsE
1835 *
1836 * Combining the above two, we can determine how each leaf iocg's inuse should
1837 * be adjusted to achieve the target donation.
1838 *
1839 *   https://drive.google.com/file/d/1WcrltBOSPN0qXVdBgnKm4mdp9FhuEFQN
1840 *
1841 * The inline comments use symbols from the last pdf.
1842 *
1843 *   b is the sum of the absolute budgets in the subtree. 1 for the root node.
1844 *   f is the sum of the absolute budgets of non-donating nodes in the subtree.
1845 *   t is the sum of the absolute budgets of donating nodes in the subtree.
1846 *   w is the weight of the node. w = w_f + w_t
1847 *   w_f is the non-donating portion of w. w_f = w * f / b
1848 *   w_b is the donating portion of w. w_t = w * t / b
1849 *   s is the sum of all sibling weights. s = Sum(w) for siblings
1850 *   s_f and s_t are the non-donating and donating portions of s.
1851 *
1852 * Subscript p denotes the parent's counterpart and ' the adjusted value - e.g.
1853 * w_pt is the donating portion of the parent's weight and w'_pt the same value
1854 * after adjustments. Subscript r denotes the root node's values.
1855 */
1856static void transfer_surpluses(struct list_head *surpluses, struct ioc_now *now)
1857{
1858	LIST_HEAD(over_hwa);
1859	LIST_HEAD(inner_walk);
1860	struct ioc_gq *iocg, *tiocg, *root_iocg;
1861	u32 after_sum, over_sum, over_target, gamma;
1862
1863	/*
1864	 * It's pretty unlikely but possible for the total sum of
1865	 * hweight_after_donation's to be higher than WEIGHT_ONE, which will
1866	 * confuse the following calculations. If such condition is detected,
1867	 * scale down everyone over its full share equally to keep the sum below
1868	 * WEIGHT_ONE.
1869	 */
1870	after_sum = 0;
1871	over_sum = 0;
1872	list_for_each_entry(iocg, surpluses, surplus_list) {
1873		u32 hwa;
1874
1875		current_hweight(iocg, &hwa, NULL);
1876		after_sum += iocg->hweight_after_donation;
1877
1878		if (iocg->hweight_after_donation > hwa) {
1879			over_sum += iocg->hweight_after_donation;
1880			list_add(&iocg->walk_list, &over_hwa);
1881		}
1882	}
1883
1884	if (after_sum >= WEIGHT_ONE) {
1885		/*
1886		 * The delta should be deducted from the over_sum, calculate
1887		 * target over_sum value.
1888		 */
1889		u32 over_delta = after_sum - (WEIGHT_ONE - 1);
1890		WARN_ON_ONCE(over_sum <= over_delta);
1891		over_target = over_sum - over_delta;
1892	} else {
1893		over_target = 0;
1894	}
1895
1896	list_for_each_entry_safe(iocg, tiocg, &over_hwa, walk_list) {
1897		if (over_target)
1898			iocg->hweight_after_donation =
1899				div_u64((u64)iocg->hweight_after_donation *
1900					over_target, over_sum);
1901		list_del_init(&iocg->walk_list);
1902	}
1903
1904	/*
1905	 * Build pre-order inner node walk list and prepare for donation
1906	 * adjustment calculations.
1907	 */
1908	list_for_each_entry(iocg, surpluses, surplus_list) {
1909		iocg_build_inner_walk(iocg, &inner_walk);
1910	}
1911
1912	root_iocg = list_first_entry(&inner_walk, struct ioc_gq, walk_list);
1913	WARN_ON_ONCE(root_iocg->level > 0);
1914
1915	list_for_each_entry(iocg, &inner_walk, walk_list) {
1916		iocg->child_adjusted_sum = 0;
1917		iocg->hweight_donating = 0;
1918		iocg->hweight_after_donation = 0;
1919	}
1920
1921	/*
1922	 * Propagate the donating budget (b_t) and after donation budget (b'_t)
1923	 * up the hierarchy.
1924	 */
1925	list_for_each_entry(iocg, surpluses, surplus_list) {
1926		struct ioc_gq *parent = iocg->ancestors[iocg->level - 1];
1927
1928		parent->hweight_donating += iocg->hweight_donating;
1929		parent->hweight_after_donation += iocg->hweight_after_donation;
1930	}
1931
1932	list_for_each_entry_reverse(iocg, &inner_walk, walk_list) {
1933		if (iocg->level > 0) {
1934			struct ioc_gq *parent = iocg->ancestors[iocg->level - 1];
1935
1936			parent->hweight_donating += iocg->hweight_donating;
1937			parent->hweight_after_donation += iocg->hweight_after_donation;
1938		}
1939	}
1940
1941	/*
1942	 * Calculate inner hwa's (b) and make sure the donation values are
1943	 * within the accepted ranges as we're doing low res calculations with
1944	 * roundups.
1945	 */
1946	list_for_each_entry(iocg, &inner_walk, walk_list) {
1947		if (iocg->level) {
1948			struct ioc_gq *parent = iocg->ancestors[iocg->level - 1];
1949
1950			iocg->hweight_active = DIV64_U64_ROUND_UP(
1951				(u64)parent->hweight_active * iocg->active,
1952				parent->child_active_sum);
1953
1954		}
1955
1956		iocg->hweight_donating = min(iocg->hweight_donating,
1957					     iocg->hweight_active);
1958		iocg->hweight_after_donation = min(iocg->hweight_after_donation,
1959						   iocg->hweight_donating - 1);
1960		if (WARN_ON_ONCE(iocg->hweight_active <= 1 ||
1961				 iocg->hweight_donating <= 1 ||
1962				 iocg->hweight_after_donation == 0)) {
1963			pr_warn("iocg: invalid donation weights in ");
1964			pr_cont_cgroup_path(iocg_to_blkg(iocg)->blkcg->css.cgroup);
1965			pr_cont(": active=%u donating=%u after=%u\n",
1966				iocg->hweight_active, iocg->hweight_donating,
1967				iocg->hweight_after_donation);
1968		}
1969	}
1970
1971	/*
1972	 * Calculate the global donation rate (gamma) - the rate to adjust
1973	 * non-donating budgets by.
1974	 *
1975	 * No need to use 64bit multiplication here as the first operand is
1976	 * guaranteed to be smaller than WEIGHT_ONE (1<<16).
1977	 *
1978	 * We know that there are beneficiary nodes and the sum of the donating
1979	 * hweights can't be whole; however, due to the round-ups during hweight
1980	 * calculations, root_iocg->hweight_donating might still end up equal to
1981	 * or greater than whole. Limit the range when calculating the divider.
1982	 *
1983	 * gamma = (1 - t_r') / (1 - t_r)
1984	 */
1985	gamma = DIV_ROUND_UP(
1986		(WEIGHT_ONE - root_iocg->hweight_after_donation) * WEIGHT_ONE,
1987		WEIGHT_ONE - min_t(u32, root_iocg->hweight_donating, WEIGHT_ONE - 1));
1988
1989	/*
1990	 * Calculate adjusted hwi, child_adjusted_sum and inuse for the inner
1991	 * nodes.
1992	 */
1993	list_for_each_entry(iocg, &inner_walk, walk_list) {
1994		struct ioc_gq *parent;
1995		u32 inuse, wpt, wptp;
1996		u64 st, sf;
1997
1998		if (iocg->level == 0) {
1999			/* adjusted weight sum for 1st level: s' = s * b_pf / b'_pf */
2000			iocg->child_adjusted_sum = DIV64_U64_ROUND_UP(
2001				iocg->child_active_sum * (WEIGHT_ONE - iocg->hweight_donating),
2002				WEIGHT_ONE - iocg->hweight_after_donation);
2003			continue;
2004		}
2005
2006		parent = iocg->ancestors[iocg->level - 1];
2007
2008		/* b' = gamma * b_f + b_t' */
2009		iocg->hweight_inuse = DIV64_U64_ROUND_UP(
2010			(u64)gamma * (iocg->hweight_active - iocg->hweight_donating),
2011			WEIGHT_ONE) + iocg->hweight_after_donation;
2012
2013		/* w' = s' * b' / b'_p */
2014		inuse = DIV64_U64_ROUND_UP(
2015			(u64)parent->child_adjusted_sum * iocg->hweight_inuse,
2016			parent->hweight_inuse);
2017
2018		/* adjusted weight sum for children: s' = s_f + s_t * w'_pt / w_pt */
2019		st = DIV64_U64_ROUND_UP(
2020			iocg->child_active_sum * iocg->hweight_donating,
2021			iocg->hweight_active);
2022		sf = iocg->child_active_sum - st;
2023		wpt = DIV64_U64_ROUND_UP(
2024			(u64)iocg->active * iocg->hweight_donating,
2025			iocg->hweight_active);
2026		wptp = DIV64_U64_ROUND_UP(
2027			(u64)inuse * iocg->hweight_after_donation,
2028			iocg->hweight_inuse);
2029
2030		iocg->child_adjusted_sum = sf + DIV64_U64_ROUND_UP(st * wptp, wpt);
2031	}
2032
2033	/*
2034	 * All inner nodes now have ->hweight_inuse and ->child_adjusted_sum and
2035	 * we can finally determine leaf adjustments.
2036	 */
2037	list_for_each_entry(iocg, surpluses, surplus_list) {
2038		struct ioc_gq *parent = iocg->ancestors[iocg->level - 1];
2039		u32 inuse;
2040
2041		/*
2042		 * In-debt iocgs participated in the donation calculation with
2043		 * the minimum target hweight_inuse. Configuring inuse
2044		 * accordingly would work fine but debt handling expects
2045		 * @iocg->inuse stay at the minimum and we don't wanna
2046		 * interfere.
2047		 */
2048		if (iocg->abs_vdebt) {
2049			WARN_ON_ONCE(iocg->inuse > 1);
2050			continue;
2051		}
2052
2053		/* w' = s' * b' / b'_p, note that b' == b'_t for donating leaves */
2054		inuse = DIV64_U64_ROUND_UP(
2055			parent->child_adjusted_sum * iocg->hweight_after_donation,
2056			parent->hweight_inuse);
2057
2058		TRACE_IOCG_PATH(inuse_transfer, iocg, now,
2059				iocg->inuse, inuse,
2060				iocg->hweight_inuse,
2061				iocg->hweight_after_donation);
2062
2063		__propagate_weights(iocg, iocg->active, inuse, true, now);
2064	}
2065
2066	/* walk list should be dissolved after use */
2067	list_for_each_entry_safe(iocg, tiocg, &inner_walk, walk_list)
2068		list_del_init(&iocg->walk_list);
2069}
2070
2071/*
2072 * A low weight iocg can amass a large amount of debt, for example, when
2073 * anonymous memory gets reclaimed aggressively. If the system has a lot of
2074 * memory paired with a slow IO device, the debt can span multiple seconds or
2075 * more. If there are no other subsequent IO issuers, the in-debt iocg may end
2076 * up blocked paying its debt while the IO device is idle.
2077 *
2078 * The following protects against such cases. If the device has been
2079 * sufficiently idle for a while, the debts are halved and delays are
2080 * recalculated.
2081 */
2082static void ioc_forgive_debts(struct ioc *ioc, u64 usage_us_sum, int nr_debtors,
2083			      struct ioc_now *now)
2084{
2085	struct ioc_gq *iocg;
2086	u64 dur, usage_pct, nr_cycles, nr_cycles_shift;
2087
2088	/* if no debtor, reset the cycle */
2089	if (!nr_debtors) {
2090		ioc->dfgv_period_at = now->now;
2091		ioc->dfgv_period_rem = 0;
2092		ioc->dfgv_usage_us_sum = 0;
2093		return;
2094	}
2095
2096	/*
2097	 * Debtors can pass through a lot of writes choking the device and we
2098	 * don't want to be forgiving debts while the device is struggling from
2099	 * write bursts. If we're missing latency targets, consider the device
2100	 * fully utilized.
2101	 */
2102	if (ioc->busy_level > 0)
2103		usage_us_sum = max_t(u64, usage_us_sum, ioc->period_us);
2104
2105	ioc->dfgv_usage_us_sum += usage_us_sum;
2106	if (time_before64(now->now, ioc->dfgv_period_at + DFGV_PERIOD))
2107		return;
2108
2109	/*
2110	 * At least DFGV_PERIOD has passed since the last period. Calculate the
2111	 * average usage and reset the period counters.
2112	 */
2113	dur = now->now - ioc->dfgv_period_at;
2114	usage_pct = div64_u64(100 * ioc->dfgv_usage_us_sum, dur);
2115
2116	ioc->dfgv_period_at = now->now;
2117	ioc->dfgv_usage_us_sum = 0;
2118
2119	/* if was too busy, reset everything */
2120	if (usage_pct > DFGV_USAGE_PCT) {
2121		ioc->dfgv_period_rem = 0;
2122		return;
2123	}
2124
2125	/*
2126	 * Usage is lower than threshold. Let's forgive some debts. Debt
2127	 * forgiveness runs off of the usual ioc timer but its period usually
2128	 * doesn't match ioc's. Compensate the difference by performing the
2129	 * reduction as many times as would fit in the duration since the last
2130	 * run and carrying over the left-over duration in @ioc->dfgv_period_rem
2131	 * - if ioc period is 75% of DFGV_PERIOD, one out of three consecutive
2132	 * reductions is doubled.
2133	 */
2134	nr_cycles = dur + ioc->dfgv_period_rem;
2135	ioc->dfgv_period_rem = do_div(nr_cycles, DFGV_PERIOD);
2136
2137	list_for_each_entry(iocg, &ioc->active_iocgs, active_list) {
2138		u64 __maybe_unused old_debt, __maybe_unused old_delay;
2139
2140		if (!iocg->abs_vdebt && !iocg->delay)
2141			continue;
2142
2143		spin_lock(&iocg->waitq.lock);
2144
2145		old_debt = iocg->abs_vdebt;
2146		old_delay = iocg->delay;
2147
2148		nr_cycles_shift = min_t(u64, nr_cycles, BITS_PER_LONG - 1);
2149		if (iocg->abs_vdebt)
2150			iocg->abs_vdebt = iocg->abs_vdebt >> nr_cycles_shift ?: 1;
2151
2152		if (iocg->delay)
2153			iocg->delay = iocg->delay >> nr_cycles_shift ?: 1;
2154
2155		iocg_kick_waitq(iocg, true, now);
2156
2157		TRACE_IOCG_PATH(iocg_forgive_debt, iocg, now, usage_pct,
2158				old_debt, iocg->abs_vdebt,
2159				old_delay, iocg->delay);
2160
2161		spin_unlock(&iocg->waitq.lock);
2162	}
2163}
2164
2165/*
2166 * Check the active iocgs' state to avoid oversleeping and deactive
2167 * idle iocgs.
2168 *
2169 * Since waiters determine the sleep durations based on the vrate
2170 * they saw at the time of sleep, if vrate has increased, some
2171 * waiters could be sleeping for too long. Wake up tardy waiters
2172 * which should have woken up in the last period and expire idle
2173 * iocgs.
2174 */
2175static int ioc_check_iocgs(struct ioc *ioc, struct ioc_now *now)
2176{
2177	int nr_debtors = 0;
2178	struct ioc_gq *iocg, *tiocg;
2179
2180	list_for_each_entry_safe(iocg, tiocg, &ioc->active_iocgs, active_list) {
2181		if (!waitqueue_active(&iocg->waitq) && !iocg->abs_vdebt &&
2182		    !iocg->delay && !iocg_is_idle(iocg))
2183			continue;
2184
2185		spin_lock(&iocg->waitq.lock);
2186
2187		/* flush wait and indebt stat deltas */
2188		if (iocg->wait_since) {
2189			iocg->stat.wait_us += now->now - iocg->wait_since;
2190			iocg->wait_since = now->now;
2191		}
2192		if (iocg->indebt_since) {
2193			iocg->stat.indebt_us +=
2194				now->now - iocg->indebt_since;
2195			iocg->indebt_since = now->now;
2196		}
2197		if (iocg->indelay_since) {
2198			iocg->stat.indelay_us +=
2199				now->now - iocg->indelay_since;
2200			iocg->indelay_since = now->now;
2201		}
2202
2203		if (waitqueue_active(&iocg->waitq) || iocg->abs_vdebt ||
2204		    iocg->delay) {
2205			/* might be oversleeping vtime / hweight changes, kick */
2206			iocg_kick_waitq(iocg, true, now);
2207			if (iocg->abs_vdebt || iocg->delay)
2208				nr_debtors++;
2209		} else if (iocg_is_idle(iocg)) {
2210			/* no waiter and idle, deactivate */
2211			u64 vtime = atomic64_read(&iocg->vtime);
2212			s64 excess;
2213
2214			/*
2215			 * @iocg has been inactive for a full duration and will
2216			 * have a high budget. Account anything above target as
2217			 * error and throw away. On reactivation, it'll start
2218			 * with the target budget.
2219			 */
2220			excess = now->vnow - vtime - ioc->margins.target;
2221			if (excess > 0) {
2222				u32 old_hwi;
2223
2224				current_hweight(iocg, NULL, &old_hwi);
2225				ioc->vtime_err -= div64_u64(excess * old_hwi,
2226							    WEIGHT_ONE);
2227			}
2228
2229			TRACE_IOCG_PATH(iocg_idle, iocg, now,
2230					atomic64_read(&iocg->active_period),
2231					atomic64_read(&ioc->cur_period), vtime);
2232			__propagate_weights(iocg, 0, 0, false, now);
2233			list_del_init(&iocg->active_list);
2234		}
2235
2236		spin_unlock(&iocg->waitq.lock);
2237	}
2238
2239	commit_weights(ioc);
2240	return nr_debtors;
2241}
2242
2243static void ioc_timer_fn(struct timer_list *timer)
2244{
2245	struct ioc *ioc = container_of(timer, struct ioc, timer);
2246	struct ioc_gq *iocg, *tiocg;
2247	struct ioc_now now;
2248	LIST_HEAD(surpluses);
2249	int nr_debtors, nr_shortages = 0, nr_lagging = 0;
2250	u64 usage_us_sum = 0;
2251	u32 ppm_rthr;
2252	u32 ppm_wthr;
2253	u32 missed_ppm[2], rq_wait_pct;
2254	u64 period_vtime;
2255	int prev_busy_level;
2256
2257	/* how were the latencies during the period? */
2258	ioc_lat_stat(ioc, missed_ppm, &rq_wait_pct);
2259
2260	/* take care of active iocgs */
2261	spin_lock_irq(&ioc->lock);
2262
2263	ppm_rthr = MILLION - ioc->params.qos[QOS_RPPM];
2264	ppm_wthr = MILLION - ioc->params.qos[QOS_WPPM];
2265	ioc_now(ioc, &now);
2266
2267	period_vtime = now.vnow - ioc->period_at_vtime;
2268	if (WARN_ON_ONCE(!period_vtime)) {
2269		spin_unlock_irq(&ioc->lock);
2270		return;
2271	}
2272
2273	nr_debtors = ioc_check_iocgs(ioc, &now);
2274
2275	/*
2276	 * Wait and indebt stat are flushed above and the donation calculation
2277	 * below needs updated usage stat. Let's bring stat up-to-date.
2278	 */
2279	iocg_flush_stat(&ioc->active_iocgs, &now);
2280
2281	/* calc usage and see whether some weights need to be moved around */
2282	list_for_each_entry(iocg, &ioc->active_iocgs, active_list) {
2283		u64 vdone, vtime, usage_us;
2284		u32 hw_active, hw_inuse;
2285
2286		/*
2287		 * Collect unused and wind vtime closer to vnow to prevent
2288		 * iocgs from accumulating a large amount of budget.
2289		 */
2290		vdone = atomic64_read(&iocg->done_vtime);
2291		vtime = atomic64_read(&iocg->vtime);
2292		current_hweight(iocg, &hw_active, &hw_inuse);
2293
2294		/*
2295		 * Latency QoS detection doesn't account for IOs which are
2296		 * in-flight for longer than a period.  Detect them by
2297		 * comparing vdone against period start.  If lagging behind
2298		 * IOs from past periods, don't increase vrate.
2299		 */
2300		if ((ppm_rthr != MILLION || ppm_wthr != MILLION) &&
2301		    !atomic_read(&iocg_to_blkg(iocg)->use_delay) &&
2302		    time_after64(vtime, vdone) &&
2303		    time_after64(vtime, now.vnow -
2304				 MAX_LAGGING_PERIODS * period_vtime) &&
2305		    time_before64(vdone, now.vnow - period_vtime))
2306			nr_lagging++;
2307
2308		/*
2309		 * Determine absolute usage factoring in in-flight IOs to avoid
2310		 * high-latency completions appearing as idle.
2311		 */
2312		usage_us = iocg->usage_delta_us;
2313		usage_us_sum += usage_us;
2314
2315		/* see whether there's surplus vtime */
2316		WARN_ON_ONCE(!list_empty(&iocg->surplus_list));
2317		if (hw_inuse < hw_active ||
2318		    (!waitqueue_active(&iocg->waitq) &&
2319		     time_before64(vtime, now.vnow - ioc->margins.low))) {
2320			u32 hwa, old_hwi, hwm, new_hwi, usage;
2321			u64 usage_dur;
2322
2323			if (vdone != vtime) {
2324				u64 inflight_us = DIV64_U64_ROUND_UP(
2325					cost_to_abs_cost(vtime - vdone, hw_inuse),
2326					ioc->vtime_base_rate);
2327
2328				usage_us = max(usage_us, inflight_us);
2329			}
2330
2331			/* convert to hweight based usage ratio */
2332			if (time_after64(iocg->activated_at, ioc->period_at))
2333				usage_dur = max_t(u64, now.now - iocg->activated_at, 1);
2334			else
2335				usage_dur = max_t(u64, now.now - ioc->period_at, 1);
2336
2337			usage = clamp(DIV64_U64_ROUND_UP(usage_us * WEIGHT_ONE, usage_dur),
2338				      1, WEIGHT_ONE);
2339
2340			/*
2341			 * Already donating or accumulated enough to start.
2342			 * Determine the donation amount.
2343			 */
2344			current_hweight(iocg, &hwa, &old_hwi);
2345			hwm = current_hweight_max(iocg);
2346			new_hwi = hweight_after_donation(iocg, old_hwi, hwm,
2347							 usage, &now);
2348			/*
2349			 * Donation calculation assumes hweight_after_donation
2350			 * to be positive, a condition that a donor w/ hwa < 2
2351			 * can't meet. Don't bother with donation if hwa is
2352			 * below 2. It's not gonna make a meaningful difference
2353			 * anyway.
2354			 */
2355			if (new_hwi < hwm && hwa >= 2) {
2356				iocg->hweight_donating = hwa;
2357				iocg->hweight_after_donation = new_hwi;
2358				list_add(&iocg->surplus_list, &surpluses);
2359			} else if (!iocg->abs_vdebt) {
2360				/*
2361				 * @iocg doesn't have enough to donate. Reset
2362				 * its inuse to active.
2363				 *
2364				 * Don't reset debtors as their inuse's are
2365				 * owned by debt handling. This shouldn't affect
2366				 * donation calculuation in any meaningful way
2367				 * as @iocg doesn't have a meaningful amount of
2368				 * share anyway.
2369				 */
2370				TRACE_IOCG_PATH(inuse_shortage, iocg, &now,
2371						iocg->inuse, iocg->active,
2372						iocg->hweight_inuse, new_hwi);
2373
2374				__propagate_weights(iocg, iocg->active,
2375						    iocg->active, true, &now);
2376				nr_shortages++;
2377			}
2378		} else {
2379			/* genuinely short on vtime */
2380			nr_shortages++;
2381		}
2382	}
2383
2384	if (!list_empty(&surpluses) && nr_shortages)
2385		transfer_surpluses(&surpluses, &now);
2386
2387	commit_weights(ioc);
2388
2389	/* surplus list should be dissolved after use */
2390	list_for_each_entry_safe(iocg, tiocg, &surpluses, surplus_list)
2391		list_del_init(&iocg->surplus_list);
2392
2393	/*
2394	 * If q is getting clogged or we're missing too much, we're issuing
2395	 * too much IO and should lower vtime rate.  If we're not missing
2396	 * and experiencing shortages but not surpluses, we're too stingy
2397	 * and should increase vtime rate.
2398	 */
2399	prev_busy_level = ioc->busy_level;
2400	if (rq_wait_pct > RQ_WAIT_BUSY_PCT ||
2401	    missed_ppm[READ] > ppm_rthr ||
2402	    missed_ppm[WRITE] > ppm_wthr) {
2403		/* clearly missing QoS targets, slow down vrate */
2404		ioc->busy_level = max(ioc->busy_level, 0);
2405		ioc->busy_level++;
2406	} else if (rq_wait_pct <= RQ_WAIT_BUSY_PCT * UNBUSY_THR_PCT / 100 &&
2407		   missed_ppm[READ] <= ppm_rthr * UNBUSY_THR_PCT / 100 &&
2408		   missed_ppm[WRITE] <= ppm_wthr * UNBUSY_THR_PCT / 100) {
2409		/* QoS targets are being met with >25% margin */
2410		if (nr_shortages) {
2411			/*
2412			 * We're throttling while the device has spare
2413			 * capacity.  If vrate was being slowed down, stop.
2414			 */
2415			ioc->busy_level = min(ioc->busy_level, 0);
2416
2417			/*
2418			 * If there are IOs spanning multiple periods, wait
2419			 * them out before pushing the device harder.
2420			 */
2421			if (!nr_lagging)
2422				ioc->busy_level--;
2423		} else {
2424			/*
2425			 * Nobody is being throttled and the users aren't
2426			 * issuing enough IOs to saturate the device.  We
2427			 * simply don't know how close the device is to
2428			 * saturation.  Coast.
2429			 */
2430			ioc->busy_level = 0;
2431		}
2432	} else {
2433		/* inside the hysterisis margin, we're good */
2434		ioc->busy_level = 0;
2435	}
2436
2437	ioc->busy_level = clamp(ioc->busy_level, -1000, 1000);
2438
2439	ioc_adjust_base_vrate(ioc, rq_wait_pct, nr_lagging, nr_shortages,
2440			      prev_busy_level, missed_ppm);
2441
2442	ioc_refresh_params(ioc, false);
2443
2444	ioc_forgive_debts(ioc, usage_us_sum, nr_debtors, &now);
2445
2446	/*
2447	 * This period is done.  Move onto the next one.  If nothing's
2448	 * going on with the device, stop the timer.
2449	 */
2450	atomic64_inc(&ioc->cur_period);
2451
2452	if (ioc->running != IOC_STOP) {
2453		if (!list_empty(&ioc->active_iocgs)) {
2454			ioc_start_period(ioc, &now);
2455		} else {
2456			ioc->busy_level = 0;
2457			ioc->vtime_err = 0;
2458			ioc->running = IOC_IDLE;
2459		}
2460
2461		ioc_refresh_vrate(ioc, &now);
2462	}
2463
2464	spin_unlock_irq(&ioc->lock);
2465}
2466
2467static u64 adjust_inuse_and_calc_cost(struct ioc_gq *iocg, u64 vtime,
2468				      u64 abs_cost, struct ioc_now *now)
2469{
2470	struct ioc *ioc = iocg->ioc;
2471	struct ioc_margins *margins = &ioc->margins;
2472	u32 __maybe_unused old_inuse = iocg->inuse, __maybe_unused old_hwi;
2473	u32 hwi, adj_step;
2474	s64 margin;
2475	u64 cost, new_inuse;
2476	unsigned long flags;
2477
2478	current_hweight(iocg, NULL, &hwi);
2479	old_hwi = hwi;
2480	cost = abs_cost_to_cost(abs_cost, hwi);
2481	margin = now->vnow - vtime - cost;
2482
2483	/* debt handling owns inuse for debtors */
2484	if (iocg->abs_vdebt)
2485		return cost;
2486
2487	/*
2488	 * We only increase inuse during period and do so if the margin has
2489	 * deteriorated since the previous adjustment.
2490	 */
2491	if (margin >= iocg->saved_margin || margin >= margins->low ||
2492	    iocg->inuse == iocg->active)
2493		return cost;
2494
2495	spin_lock_irqsave(&ioc->lock, flags);
2496
2497	/* we own inuse only when @iocg is in the normal active state */
2498	if (iocg->abs_vdebt || list_empty(&iocg->active_list)) {
2499		spin_unlock_irqrestore(&ioc->lock, flags);
2500		return cost;
2501	}
2502
2503	/*
2504	 * Bump up inuse till @abs_cost fits in the existing budget.
2505	 * adj_step must be determined after acquiring ioc->lock - we might
2506	 * have raced and lost to another thread for activation and could
2507	 * be reading 0 iocg->active before ioc->lock which will lead to
2508	 * infinite loop.
2509	 */
2510	new_inuse = iocg->inuse;
2511	adj_step = DIV_ROUND_UP(iocg->active * INUSE_ADJ_STEP_PCT, 100);
2512	do {
2513		new_inuse = new_inuse + adj_step;
2514		propagate_weights(iocg, iocg->active, new_inuse, true, now);
2515		current_hweight(iocg, NULL, &hwi);
2516		cost = abs_cost_to_cost(abs_cost, hwi);
2517	} while (time_after64(vtime + cost, now->vnow) &&
2518		 iocg->inuse != iocg->active);
2519
2520	spin_unlock_irqrestore(&ioc->lock, flags);
2521
2522	TRACE_IOCG_PATH(inuse_adjust, iocg, now,
2523			old_inuse, iocg->inuse, old_hwi, hwi);
2524
2525	return cost;
2526}
2527
2528static void calc_vtime_cost_builtin(struct bio *bio, struct ioc_gq *iocg,
2529				    bool is_merge, u64 *costp)
2530{
2531	struct ioc *ioc = iocg->ioc;
2532	u64 coef_seqio, coef_randio, coef_page;
2533	u64 pages = max_t(u64, bio_sectors(bio) >> IOC_SECT_TO_PAGE_SHIFT, 1);
2534	u64 seek_pages = 0;
2535	u64 cost = 0;
2536
2537	/* Can't calculate cost for empty bio */
2538	if (!bio->bi_iter.bi_size)
2539		goto out;
2540
2541	switch (bio_op(bio)) {
2542	case REQ_OP_READ:
2543		coef_seqio	= ioc->params.lcoefs[LCOEF_RSEQIO];
2544		coef_randio	= ioc->params.lcoefs[LCOEF_RRANDIO];
2545		coef_page	= ioc->params.lcoefs[LCOEF_RPAGE];
2546		break;
2547	case REQ_OP_WRITE:
2548		coef_seqio	= ioc->params.lcoefs[LCOEF_WSEQIO];
2549		coef_randio	= ioc->params.lcoefs[LCOEF_WRANDIO];
2550		coef_page	= ioc->params.lcoefs[LCOEF_WPAGE];
2551		break;
2552	default:
2553		goto out;
2554	}
2555
2556	if (iocg->cursor) {
2557		seek_pages = abs(bio->bi_iter.bi_sector - iocg->cursor);
2558		seek_pages >>= IOC_SECT_TO_PAGE_SHIFT;
2559	}
2560
2561	if (!is_merge) {
2562		if (seek_pages > LCOEF_RANDIO_PAGES) {
2563			cost += coef_randio;
2564		} else {
2565			cost += coef_seqio;
2566		}
2567	}
2568	cost += pages * coef_page;
2569out:
2570	*costp = cost;
2571}
2572
2573static u64 calc_vtime_cost(struct bio *bio, struct ioc_gq *iocg, bool is_merge)
2574{
2575	u64 cost;
2576
2577	calc_vtime_cost_builtin(bio, iocg, is_merge, &cost);
2578	return cost;
2579}
2580
2581static void calc_size_vtime_cost_builtin(struct request *rq, struct ioc *ioc,
2582					 u64 *costp)
2583{
2584	unsigned int pages = blk_rq_stats_sectors(rq) >> IOC_SECT_TO_PAGE_SHIFT;
2585
2586	switch (req_op(rq)) {
2587	case REQ_OP_READ:
2588		*costp = pages * ioc->params.lcoefs[LCOEF_RPAGE];
2589		break;
2590	case REQ_OP_WRITE:
2591		*costp = pages * ioc->params.lcoefs[LCOEF_WPAGE];
2592		break;
2593	default:
2594		*costp = 0;
2595	}
2596}
2597
2598static u64 calc_size_vtime_cost(struct request *rq, struct ioc *ioc)
2599{
2600	u64 cost;
2601
2602	calc_size_vtime_cost_builtin(rq, ioc, &cost);
2603	return cost;
2604}
2605
2606static void ioc_rqos_throttle(struct rq_qos *rqos, struct bio *bio)
2607{
2608	struct blkcg_gq *blkg = bio->bi_blkg;
2609	struct ioc *ioc = rqos_to_ioc(rqos);
2610	struct ioc_gq *iocg = blkg_to_iocg(blkg);
2611	struct ioc_now now;
2612	struct iocg_wait wait;
2613	u64 abs_cost, cost, vtime;
2614	bool use_debt, ioc_locked;
2615	unsigned long flags;
2616
2617	/* bypass IOs if disabled, still initializing, or for root cgroup */
2618	if (!ioc->enabled || !iocg || !iocg->level)
2619		return;
2620
2621	/* calculate the absolute vtime cost */
2622	abs_cost = calc_vtime_cost(bio, iocg, false);
2623	if (!abs_cost)
2624		return;
2625
2626	if (!iocg_activate(iocg, &now))
2627		return;
2628
2629	iocg->cursor = bio_end_sector(bio);
2630	vtime = atomic64_read(&iocg->vtime);
2631	cost = adjust_inuse_and_calc_cost(iocg, vtime, abs_cost, &now);
2632
2633	/*
2634	 * If no one's waiting and within budget, issue right away.  The
2635	 * tests are racy but the races aren't systemic - we only miss once
2636	 * in a while which is fine.
2637	 */
2638	if (!waitqueue_active(&iocg->waitq) && !iocg->abs_vdebt &&
2639	    time_before_eq64(vtime + cost, now.vnow)) {
2640		iocg_commit_bio(iocg, bio, abs_cost, cost);
2641		return;
2642	}
2643
2644	/*
2645	 * We're over budget. This can be handled in two ways. IOs which may
2646	 * cause priority inversions are punted to @ioc->aux_iocg and charged as
2647	 * debt. Otherwise, the issuer is blocked on @iocg->waitq. Debt handling
2648	 * requires @ioc->lock, waitq handling @iocg->waitq.lock. Determine
2649	 * whether debt handling is needed and acquire locks accordingly.
2650	 */
2651	use_debt = bio_issue_as_root_blkg(bio) || fatal_signal_pending(current);
2652	ioc_locked = use_debt || READ_ONCE(iocg->abs_vdebt);
2653retry_lock:
2654	iocg_lock(iocg, ioc_locked, &flags);
2655
2656	/*
2657	 * @iocg must stay activated for debt and waitq handling. Deactivation
2658	 * is synchronized against both ioc->lock and waitq.lock and we won't
2659	 * get deactivated as long as we're waiting or has debt, so we're good
2660	 * if we're activated here. In the unlikely cases that we aren't, just
2661	 * issue the IO.
2662	 */
2663	if (unlikely(list_empty(&iocg->active_list))) {
2664		iocg_unlock(iocg, ioc_locked, &flags);
2665		iocg_commit_bio(iocg, bio, abs_cost, cost);
2666		return;
2667	}
2668
2669	/*
2670	 * We're over budget. If @bio has to be issued regardless, remember
2671	 * the abs_cost instead of advancing vtime. iocg_kick_waitq() will pay
2672	 * off the debt before waking more IOs.
2673	 *
2674	 * This way, the debt is continuously paid off each period with the
2675	 * actual budget available to the cgroup. If we just wound vtime, we
2676	 * would incorrectly use the current hw_inuse for the entire amount
2677	 * which, for example, can lead to the cgroup staying blocked for a
2678	 * long time even with substantially raised hw_inuse.
2679	 *
2680	 * An iocg with vdebt should stay online so that the timer can keep
2681	 * deducting its vdebt and [de]activate use_delay mechanism
2682	 * accordingly. We don't want to race against the timer trying to
2683	 * clear them and leave @iocg inactive w/ dangling use_delay heavily
2684	 * penalizing the cgroup and its descendants.
2685	 */
2686	if (use_debt) {
2687		iocg_incur_debt(iocg, abs_cost, &now);
2688		if (iocg_kick_delay(iocg, &now))
2689			blkcg_schedule_throttle(rqos->disk,
2690					(bio->bi_opf & REQ_SWAP) == REQ_SWAP);
2691		iocg_unlock(iocg, ioc_locked, &flags);
2692		return;
2693	}
2694
2695	/* guarantee that iocgs w/ waiters have maximum inuse */
2696	if (!iocg->abs_vdebt && iocg->inuse != iocg->active) {
2697		if (!ioc_locked) {
2698			iocg_unlock(iocg, false, &flags);
2699			ioc_locked = true;
2700			goto retry_lock;
2701		}
2702		propagate_weights(iocg, iocg->active, iocg->active, true,
2703				  &now);
2704	}
2705
2706	/*
2707	 * Append self to the waitq and schedule the wakeup timer if we're
2708	 * the first waiter.  The timer duration is calculated based on the
2709	 * current vrate.  vtime and hweight changes can make it too short
2710	 * or too long.  Each wait entry records the absolute cost it's
2711	 * waiting for to allow re-evaluation using a custom wait entry.
2712	 *
2713	 * If too short, the timer simply reschedules itself.  If too long,
2714	 * the period timer will notice and trigger wakeups.
2715	 *
2716	 * All waiters are on iocg->waitq and the wait states are
2717	 * synchronized using waitq.lock.
2718	 */
2719	init_wait_func(&wait.wait, iocg_wake_fn);
2720	wait.bio = bio;
2721	wait.abs_cost = abs_cost;
2722	wait.committed = false;	/* will be set true by waker */
2723
2724	__add_wait_queue_entry_tail(&iocg->waitq, &wait.wait);
2725	iocg_kick_waitq(iocg, ioc_locked, &now);
2726
2727	iocg_unlock(iocg, ioc_locked, &flags);
2728
2729	while (true) {
2730		set_current_state(TASK_UNINTERRUPTIBLE);
2731		if (wait.committed)
2732			break;
2733		io_schedule();
2734	}
2735
2736	/* waker already committed us, proceed */
2737	finish_wait(&iocg->waitq, &wait.wait);
2738}
2739
2740static void ioc_rqos_merge(struct rq_qos *rqos, struct request *rq,
2741			   struct bio *bio)
2742{
2743	struct ioc_gq *iocg = blkg_to_iocg(bio->bi_blkg);
2744	struct ioc *ioc = rqos_to_ioc(rqos);
2745	sector_t bio_end = bio_end_sector(bio);
2746	struct ioc_now now;
2747	u64 vtime, abs_cost, cost;
2748	unsigned long flags;
2749
2750	/* bypass if disabled, still initializing, or for root cgroup */
2751	if (!ioc->enabled || !iocg || !iocg->level)
2752		return;
2753
2754	abs_cost = calc_vtime_cost(bio, iocg, true);
2755	if (!abs_cost)
2756		return;
2757
2758	ioc_now(ioc, &now);
2759
2760	vtime = atomic64_read(&iocg->vtime);
2761	cost = adjust_inuse_and_calc_cost(iocg, vtime, abs_cost, &now);
2762
2763	/* update cursor if backmerging into the request at the cursor */
2764	if (blk_rq_pos(rq) < bio_end &&
2765	    blk_rq_pos(rq) + blk_rq_sectors(rq) == iocg->cursor)
2766		iocg->cursor = bio_end;
2767
2768	/*
2769	 * Charge if there's enough vtime budget and the existing request has
2770	 * cost assigned.
2771	 */
2772	if (rq->bio && rq->bio->bi_iocost_cost &&
2773	    time_before_eq64(atomic64_read(&iocg->vtime) + cost, now.vnow)) {
2774		iocg_commit_bio(iocg, bio, abs_cost, cost);
2775		return;
2776	}
2777
2778	/*
2779	 * Otherwise, account it as debt if @iocg is online, which it should
2780	 * be for the vast majority of cases. See debt handling in
2781	 * ioc_rqos_throttle() for details.
2782	 */
2783	spin_lock_irqsave(&ioc->lock, flags);
2784	spin_lock(&iocg->waitq.lock);
2785
2786	if (likely(!list_empty(&iocg->active_list))) {
2787		iocg_incur_debt(iocg, abs_cost, &now);
2788		if (iocg_kick_delay(iocg, &now))
2789			blkcg_schedule_throttle(rqos->disk,
2790					(bio->bi_opf & REQ_SWAP) == REQ_SWAP);
2791	} else {
2792		iocg_commit_bio(iocg, bio, abs_cost, cost);
2793	}
2794
2795	spin_unlock(&iocg->waitq.lock);
2796	spin_unlock_irqrestore(&ioc->lock, flags);
2797}
2798
2799static void ioc_rqos_done_bio(struct rq_qos *rqos, struct bio *bio)
2800{
2801	struct ioc_gq *iocg = blkg_to_iocg(bio->bi_blkg);
2802
2803	if (iocg && bio->bi_iocost_cost)
2804		atomic64_add(bio->bi_iocost_cost, &iocg->done_vtime);
2805}
2806
2807static void ioc_rqos_done(struct rq_qos *rqos, struct request *rq)
2808{
2809	struct ioc *ioc = rqos_to_ioc(rqos);
2810	struct ioc_pcpu_stat *ccs;
2811	u64 on_q_ns, rq_wait_ns, size_nsec;
2812	int pidx, rw;
2813
2814	if (!ioc->enabled || !rq->alloc_time_ns || !rq->start_time_ns)
2815		return;
2816
2817	switch (req_op(rq)) {
2818	case REQ_OP_READ:
2819		pidx = QOS_RLAT;
2820		rw = READ;
2821		break;
2822	case REQ_OP_WRITE:
2823		pidx = QOS_WLAT;
2824		rw = WRITE;
2825		break;
2826	default:
2827		return;
2828	}
2829
2830	on_q_ns = blk_time_get_ns() - rq->alloc_time_ns;
2831	rq_wait_ns = rq->start_time_ns - rq->alloc_time_ns;
2832	size_nsec = div64_u64(calc_size_vtime_cost(rq, ioc), VTIME_PER_NSEC);
2833
2834	ccs = get_cpu_ptr(ioc->pcpu_stat);
2835
2836	if (on_q_ns <= size_nsec ||
2837	    on_q_ns - size_nsec <= ioc->params.qos[pidx] * NSEC_PER_USEC)
2838		local_inc(&ccs->missed[rw].nr_met);
2839	else
2840		local_inc(&ccs->missed[rw].nr_missed);
2841
2842	local64_add(rq_wait_ns, &ccs->rq_wait_ns);
2843
2844	put_cpu_ptr(ccs);
2845}
2846
2847static void ioc_rqos_queue_depth_changed(struct rq_qos *rqos)
2848{
2849	struct ioc *ioc = rqos_to_ioc(rqos);
2850
2851	spin_lock_irq(&ioc->lock);
2852	ioc_refresh_params(ioc, false);
2853	spin_unlock_irq(&ioc->lock);
2854}
2855
2856static void ioc_rqos_exit(struct rq_qos *rqos)
2857{
2858	struct ioc *ioc = rqos_to_ioc(rqos);
2859
2860	blkcg_deactivate_policy(rqos->disk, &blkcg_policy_iocost);
2861
2862	spin_lock_irq(&ioc->lock);
2863	ioc->running = IOC_STOP;
2864	spin_unlock_irq(&ioc->lock);
2865
2866	timer_shutdown_sync(&ioc->timer);
2867	free_percpu(ioc->pcpu_stat);
2868	kfree(ioc);
2869}
2870
2871static const struct rq_qos_ops ioc_rqos_ops = {
2872	.throttle = ioc_rqos_throttle,
2873	.merge = ioc_rqos_merge,
2874	.done_bio = ioc_rqos_done_bio,
2875	.done = ioc_rqos_done,
2876	.queue_depth_changed = ioc_rqos_queue_depth_changed,
2877	.exit = ioc_rqos_exit,
2878};
2879
2880static int blk_iocost_init(struct gendisk *disk)
2881{
2882	struct ioc *ioc;
2883	int i, cpu, ret;
2884
2885	ioc = kzalloc(sizeof(*ioc), GFP_KERNEL);
2886	if (!ioc)
2887		return -ENOMEM;
2888
2889	ioc->pcpu_stat = alloc_percpu(struct ioc_pcpu_stat);
2890	if (!ioc->pcpu_stat) {
2891		kfree(ioc);
2892		return -ENOMEM;
2893	}
2894
2895	for_each_possible_cpu(cpu) {
2896		struct ioc_pcpu_stat *ccs = per_cpu_ptr(ioc->pcpu_stat, cpu);
2897
2898		for (i = 0; i < ARRAY_SIZE(ccs->missed); i++) {
2899			local_set(&ccs->missed[i].nr_met, 0);
2900			local_set(&ccs->missed[i].nr_missed, 0);
2901		}
2902		local64_set(&ccs->rq_wait_ns, 0);
2903	}
2904
2905	spin_lock_init(&ioc->lock);
2906	timer_setup(&ioc->timer, ioc_timer_fn, 0);
2907	INIT_LIST_HEAD(&ioc->active_iocgs);
2908
2909	ioc->running = IOC_IDLE;
2910	ioc->vtime_base_rate = VTIME_PER_USEC;
2911	atomic64_set(&ioc->vtime_rate, VTIME_PER_USEC);
2912	seqcount_spinlock_init(&ioc->period_seqcount, &ioc->lock);
2913	ioc->period_at = ktime_to_us(blk_time_get());
2914	atomic64_set(&ioc->cur_period, 0);
2915	atomic_set(&ioc->hweight_gen, 0);
2916
2917	spin_lock_irq(&ioc->lock);
2918	ioc->autop_idx = AUTOP_INVALID;
2919	ioc_refresh_params_disk(ioc, true, disk);
2920	spin_unlock_irq(&ioc->lock);
2921
2922	/*
2923	 * rqos must be added before activation to allow ioc_pd_init() to
2924	 * lookup the ioc from q. This means that the rqos methods may get
2925	 * called before policy activation completion, can't assume that the
2926	 * target bio has an iocg associated and need to test for NULL iocg.
2927	 */
2928	ret = rq_qos_add(&ioc->rqos, disk, RQ_QOS_COST, &ioc_rqos_ops);
2929	if (ret)
2930		goto err_free_ioc;
2931
2932	ret = blkcg_activate_policy(disk, &blkcg_policy_iocost);
2933	if (ret)
2934		goto err_del_qos;
2935	return 0;
2936
2937err_del_qos:
2938	rq_qos_del(&ioc->rqos);
2939err_free_ioc:
2940	free_percpu(ioc->pcpu_stat);
2941	kfree(ioc);
2942	return ret;
2943}
2944
2945static struct blkcg_policy_data *ioc_cpd_alloc(gfp_t gfp)
2946{
2947	struct ioc_cgrp *iocc;
2948
2949	iocc = kzalloc(sizeof(struct ioc_cgrp), gfp);
2950	if (!iocc)
2951		return NULL;
2952
2953	iocc->dfl_weight = CGROUP_WEIGHT_DFL * WEIGHT_ONE;
2954	return &iocc->cpd;
2955}
2956
2957static void ioc_cpd_free(struct blkcg_policy_data *cpd)
2958{
2959	kfree(container_of(cpd, struct ioc_cgrp, cpd));
2960}
2961
2962static struct blkg_policy_data *ioc_pd_alloc(struct gendisk *disk,
2963		struct blkcg *blkcg, gfp_t gfp)
2964{
2965	int levels = blkcg->css.cgroup->level + 1;
2966	struct ioc_gq *iocg;
2967
2968	iocg = kzalloc_node(struct_size(iocg, ancestors, levels), gfp,
2969			    disk->node_id);
2970	if (!iocg)
2971		return NULL;
2972
2973	iocg->pcpu_stat = alloc_percpu_gfp(struct iocg_pcpu_stat, gfp);
2974	if (!iocg->pcpu_stat) {
2975		kfree(iocg);
2976		return NULL;
2977	}
2978
2979	return &iocg->pd;
2980}
2981
2982static void ioc_pd_init(struct blkg_policy_data *pd)
2983{
2984	struct ioc_gq *iocg = pd_to_iocg(pd);
2985	struct blkcg_gq *blkg = pd_to_blkg(&iocg->pd);
2986	struct ioc *ioc = q_to_ioc(blkg->q);
2987	struct ioc_now now;
2988	struct blkcg_gq *tblkg;
2989	unsigned long flags;
2990
2991	ioc_now(ioc, &now);
2992
2993	iocg->ioc = ioc;
2994	atomic64_set(&iocg->vtime, now.vnow);
2995	atomic64_set(&iocg->done_vtime, now.vnow);
2996	atomic64_set(&iocg->active_period, atomic64_read(&ioc->cur_period));
2997	INIT_LIST_HEAD(&iocg->active_list);
2998	INIT_LIST_HEAD(&iocg->walk_list);
2999	INIT_LIST_HEAD(&iocg->surplus_list);
3000	iocg->hweight_active = WEIGHT_ONE;
3001	iocg->hweight_inuse = WEIGHT_ONE;
3002
3003	init_waitqueue_head(&iocg->waitq);
3004	hrtimer_setup(&iocg->waitq_timer, iocg_waitq_timer_fn, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
3005
3006	iocg->level = blkg->blkcg->css.cgroup->level;
3007
3008	for (tblkg = blkg; tblkg; tblkg = tblkg->parent) {
3009		struct ioc_gq *tiocg = blkg_to_iocg(tblkg);
3010		iocg->ancestors[tiocg->level] = tiocg;
3011	}
3012
3013	spin_lock_irqsave(&ioc->lock, flags);
3014	weight_updated(iocg, &now);
3015	spin_unlock_irqrestore(&ioc->lock, flags);
3016}
3017
3018static void ioc_pd_free(struct blkg_policy_data *pd)
3019{
3020	struct ioc_gq *iocg = pd_to_iocg(pd);
3021	struct ioc *ioc = iocg->ioc;
3022	unsigned long flags;
3023
3024	if (ioc) {
3025		spin_lock_irqsave(&ioc->lock, flags);
3026
3027		if (!list_empty(&iocg->active_list)) {
3028			struct ioc_now now;
3029
3030			ioc_now(ioc, &now);
3031			propagate_weights(iocg, 0, 0, false, &now);
3032			list_del_init(&iocg->active_list);
3033		}
3034
3035		WARN_ON_ONCE(!list_empty(&iocg->walk_list));
3036		WARN_ON_ONCE(!list_empty(&iocg->surplus_list));
3037
3038		spin_unlock_irqrestore(&ioc->lock, flags);
3039
3040		hrtimer_cancel(&iocg->waitq_timer);
3041	}
3042	free_percpu(iocg->pcpu_stat);
3043	kfree(iocg);
3044}
3045
3046static void ioc_pd_stat(struct blkg_policy_data *pd, struct seq_file *s)
3047{
3048	struct ioc_gq *iocg = pd_to_iocg(pd);
3049	struct ioc *ioc = iocg->ioc;
3050
3051	if (!ioc->enabled)
3052		return;
3053
3054	if (iocg->level == 0) {
3055		unsigned vp10k = DIV64_U64_ROUND_CLOSEST(
3056			ioc->vtime_base_rate * 10000,
3057			VTIME_PER_USEC);
3058		seq_printf(s, " cost.vrate=%u.%02u", vp10k / 100, vp10k % 100);
3059	}
3060
3061	seq_printf(s, " cost.usage=%llu", iocg->last_stat.usage_us);
3062
3063	if (blkcg_debug_stats)
3064		seq_printf(s, " cost.wait=%llu cost.indebt=%llu cost.indelay=%llu",
3065			iocg->last_stat.wait_us,
3066			iocg->last_stat.indebt_us,
3067			iocg->last_stat.indelay_us);
3068}
3069
3070static u64 ioc_weight_prfill(struct seq_file *sf, struct blkg_policy_data *pd,
3071			     int off)
3072{
3073	const char *dname = blkg_dev_name(pd->blkg);
3074	struct ioc_gq *iocg = pd_to_iocg(pd);
3075
3076	if (dname && iocg->cfg_weight)
3077		seq_printf(sf, "%s %u\n", dname, iocg->cfg_weight / WEIGHT_ONE);
3078	return 0;
3079}
3080
3081
3082static int ioc_weight_show(struct seq_file *sf, void *v)
3083{
3084	struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
3085	struct ioc_cgrp *iocc = blkcg_to_iocc(blkcg);
3086
3087	seq_printf(sf, "default %u\n", iocc->dfl_weight / WEIGHT_ONE);
3088	blkcg_print_blkgs(sf, blkcg, ioc_weight_prfill,
3089			  &blkcg_policy_iocost, seq_cft(sf)->private, false);
3090	return 0;
3091}
3092
3093static ssize_t ioc_weight_write(struct kernfs_open_file *of, char *buf,
3094				size_t nbytes, loff_t off)
3095{
3096	struct blkcg *blkcg = css_to_blkcg(of_css(of));
3097	struct ioc_cgrp *iocc = blkcg_to_iocc(blkcg);
3098	struct blkg_conf_ctx ctx;
3099	struct ioc_now now;
3100	struct ioc_gq *iocg;
3101	u32 v;
3102	int ret;
3103
3104	if (!strchr(buf, ':')) {
3105		struct blkcg_gq *blkg;
3106
3107		if (!sscanf(buf, "default %u", &v) && !sscanf(buf, "%u", &v))
3108			return -EINVAL;
3109
3110		if (v < CGROUP_WEIGHT_MIN || v > CGROUP_WEIGHT_MAX)
3111			return -EINVAL;
3112
3113		spin_lock_irq(&blkcg->lock);
3114		iocc->dfl_weight = v * WEIGHT_ONE;
3115		hlist_for_each_entry(blkg, &blkcg->blkg_list, blkcg_node) {
3116			struct ioc_gq *iocg = blkg_to_iocg(blkg);
3117
3118			if (iocg) {
3119				spin_lock(&iocg->ioc->lock);
3120				ioc_now(iocg->ioc, &now);
3121				weight_updated(iocg, &now);
3122				spin_unlock(&iocg->ioc->lock);
3123			}
3124		}
3125		spin_unlock_irq(&blkcg->lock);
3126
3127		return nbytes;
3128	}
3129
3130	blkg_conf_init(&ctx, buf);
3131
3132	ret = blkg_conf_prep(blkcg, &blkcg_policy_iocost, &ctx);
3133	if (ret)
3134		goto err;
3135
3136	iocg = blkg_to_iocg(ctx.blkg);
3137
3138	if (!strncmp(ctx.body, "default", 7)) {
3139		v = 0;
3140	} else {
3141		if (!sscanf(ctx.body, "%u", &v))
3142			goto einval;
3143		if (v < CGROUP_WEIGHT_MIN || v > CGROUP_WEIGHT_MAX)
3144			goto einval;
3145	}
3146
3147	spin_lock(&iocg->ioc->lock);
3148	iocg->cfg_weight = v * WEIGHT_ONE;
3149	ioc_now(iocg->ioc, &now);
3150	weight_updated(iocg, &now);
3151	spin_unlock(&iocg->ioc->lock);
3152
3153	blkg_conf_exit(&ctx);
3154	return nbytes;
3155
3156einval:
3157	ret = -EINVAL;
3158err:
3159	blkg_conf_exit(&ctx);
3160	return ret;
3161}
3162
3163static u64 ioc_qos_prfill(struct seq_file *sf, struct blkg_policy_data *pd,
3164			  int off)
3165{
3166	const char *dname = blkg_dev_name(pd->blkg);
3167	struct ioc *ioc = pd_to_iocg(pd)->ioc;
3168
3169	if (!dname)
3170		return 0;
3171
3172	spin_lock(&ioc->lock);
3173	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",
3174		   dname, ioc->enabled, ioc->user_qos_params ? "user" : "auto",
3175		   ioc->params.qos[QOS_RPPM] / 10000,
3176		   ioc->params.qos[QOS_RPPM] % 10000 / 100,
3177		   ioc->params.qos[QOS_RLAT],
3178		   ioc->params.qos[QOS_WPPM] / 10000,
3179		   ioc->params.qos[QOS_WPPM] % 10000 / 100,
3180		   ioc->params.qos[QOS_WLAT],
3181		   ioc->params.qos[QOS_MIN] / 10000,
3182		   ioc->params.qos[QOS_MIN] % 10000 / 100,
3183		   ioc->params.qos[QOS_MAX] / 10000,
3184		   ioc->params.qos[QOS_MAX] % 10000 / 100);
3185	spin_unlock(&ioc->lock);
3186	return 0;
3187}
3188
3189static int ioc_qos_show(struct seq_file *sf, void *v)
3190{
3191	struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
3192
3193	blkcg_print_blkgs(sf, blkcg, ioc_qos_prfill,
3194			  &blkcg_policy_iocost, seq_cft(sf)->private, false);
3195	return 0;
3196}
3197
3198static const match_table_t qos_ctrl_tokens = {
3199	{ QOS_ENABLE,		"enable=%u"	},
3200	{ QOS_CTRL,		"ctrl=%s"	},
3201	{ NR_QOS_CTRL_PARAMS,	NULL		},
3202};
3203
3204static const match_table_t qos_tokens = {
3205	{ QOS_RPPM,		"rpct=%s"	},
3206	{ QOS_RLAT,		"rlat=%u"	},
3207	{ QOS_WPPM,		"wpct=%s"	},
3208	{ QOS_WLAT,		"wlat=%u"	},
3209	{ QOS_MIN,		"min=%s"	},
3210	{ QOS_MAX,		"max=%s"	},
3211	{ NR_QOS_PARAMS,	NULL		},
3212};
3213
3214static ssize_t ioc_qos_write(struct kernfs_open_file *of, char *input,
3215			     size_t nbytes, loff_t off)
3216{
3217	struct blkg_conf_ctx ctx;
3218	struct gendisk *disk;
3219	struct ioc *ioc;
3220	u32 qos[NR_QOS_PARAMS];
3221	bool enable, user;
3222	char *body, *p;
3223	unsigned long memflags;
3224	int ret;
3225
3226	blkg_conf_init(&ctx, input);
3227
3228	memflags = blkg_conf_open_bdev_frozen(&ctx);
3229	if (IS_ERR_VALUE(memflags)) {
3230		ret = memflags;
3231		goto err;
3232	}
3233
3234	body = ctx.body;
3235	disk = ctx.bdev->bd_disk;
3236	if (!queue_is_mq(disk->queue)) {
3237		ret = -EOPNOTSUPP;
3238		goto err;
3239	}
3240
3241	ioc = q_to_ioc(disk->queue);
3242	if (!ioc) {
3243		ret = blk_iocost_init(disk);
3244		if (ret)
3245			goto err;
3246		ioc = q_to_ioc(disk->queue);
3247	}
3248
3249	blk_mq_quiesce_queue(disk->queue);
3250
3251	spin_lock_irq(&ioc->lock);
3252	memcpy(qos, ioc->params.qos, sizeof(qos));
3253	enable = ioc->enabled;
3254	user = ioc->user_qos_params;
3255
3256	while ((p = strsep(&body, " \t\n"))) {
3257		substring_t args[MAX_OPT_ARGS];
3258		char buf[32];
3259		int tok;
3260		s64 v;
3261
3262		if (!*p)
3263			continue;
3264
3265		switch (match_token(p, qos_ctrl_tokens, args)) {
3266		case QOS_ENABLE:
3267			if (match_u64(&args[0], &v))
3268				goto einval;
3269			enable = v;
3270			continue;
3271		case QOS_CTRL:
3272			match_strlcpy(buf, &args[0], sizeof(buf));
3273			if (!strcmp(buf, "auto"))
3274				user = false;
3275			else if (!strcmp(buf, "user"))
3276				user = true;
3277			else
3278				goto einval;
3279			continue;
3280		}
3281
3282		tok = match_token(p, qos_tokens, args);
3283		switch (tok) {
3284		case QOS_RPPM:
3285		case QOS_WPPM:
3286			if (match_strlcpy(buf, &args[0], sizeof(buf)) >=
3287			    sizeof(buf))
3288				goto einval;
3289			if (cgroup_parse_float(buf, 2, &v))
3290				goto einval;
3291			if (v < 0 || v > 10000)
3292				goto einval;
3293			qos[tok] = v * 100;
3294			break;
3295		case QOS_RLAT:
3296		case QOS_WLAT:
3297			if (match_u64(&args[0], &v))
3298				goto einval;
3299			qos[tok] = v;
3300			break;
3301		case QOS_MIN:
3302		case QOS_MAX:
3303			if (match_strlcpy(buf, &args[0], sizeof(buf)) >=
3304			    sizeof(buf))
3305				goto einval;
3306			if (cgroup_parse_float(buf, 2, &v))
3307				goto einval;
3308			if (v < 0)
3309				goto einval;
3310			qos[tok] = clamp_t(s64, v * 100,
3311					   VRATE_MIN_PPM, VRATE_MAX_PPM);
3312			break;
3313		default:
3314			goto einval;
3315		}
3316		user = true;
3317	}
3318
3319	if (qos[QOS_MIN] > qos[QOS_MAX])
3320		goto einval;
3321
3322	if (enable && !ioc->enabled) {
3323		blk_stat_enable_accounting(disk->queue);
3324		blk_queue_flag_set(QUEUE_FLAG_RQ_ALLOC_TIME, disk->queue);
3325		ioc->enabled = true;
3326	} else if (!enable && ioc->enabled) {
3327		blk_stat_disable_accounting(disk->queue);
3328		blk_queue_flag_clear(QUEUE_FLAG_RQ_ALLOC_TIME, disk->queue);
3329		ioc->enabled = false;
3330	}
3331
3332	if (user) {
3333		memcpy(ioc->params.qos, qos, sizeof(qos));
3334		ioc->user_qos_params = true;
3335	} else {
3336		ioc->user_qos_params = false;
3337	}
3338
3339	ioc_refresh_params(ioc, true);
3340	spin_unlock_irq(&ioc->lock);
3341
3342	if (enable)
3343		wbt_disable_default(disk);
3344	else
3345		wbt_enable_default(disk);
3346
3347	blk_mq_unquiesce_queue(disk->queue);
3348
3349	blkg_conf_exit_frozen(&ctx, memflags);
3350	return nbytes;
3351einval:
3352	spin_unlock_irq(&ioc->lock);
3353	blk_mq_unquiesce_queue(disk->queue);
3354	ret = -EINVAL;
3355err:
3356	blkg_conf_exit_frozen(&ctx, memflags);
3357	return ret;
3358}
3359
3360static u64 ioc_cost_model_prfill(struct seq_file *sf,
3361				 struct blkg_policy_data *pd, int off)
3362{
3363	const char *dname = blkg_dev_name(pd->blkg);
3364	struct ioc *ioc = pd_to_iocg(pd)->ioc;
3365	u64 *u = ioc->params.i_lcoefs;
3366
3367	if (!dname)
3368		return 0;
3369
3370	spin_lock(&ioc->lock);
3371	seq_printf(sf, "%s ctrl=%s model=linear "
3372		   "rbps=%llu rseqiops=%llu rrandiops=%llu "
3373		   "wbps=%llu wseqiops=%llu wrandiops=%llu\n",
3374		   dname, ioc->user_cost_model ? "user" : "auto",
3375		   u[I_LCOEF_RBPS], u[I_LCOEF_RSEQIOPS], u[I_LCOEF_RRANDIOPS],
3376		   u[I_LCOEF_WBPS], u[I_LCOEF_WSEQIOPS], u[I_LCOEF_WRANDIOPS]);
3377	spin_unlock(&ioc->lock);
3378	return 0;
3379}
3380
3381static int ioc_cost_model_show(struct seq_file *sf, void *v)
3382{
3383	struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
3384
3385	blkcg_print_blkgs(sf, blkcg, ioc_cost_model_prfill,
3386			  &blkcg_policy_iocost, seq_cft(sf)->private, false);
3387	return 0;
3388}
3389
3390static const match_table_t cost_ctrl_tokens = {
3391	{ COST_CTRL,		"ctrl=%s"	},
3392	{ COST_MODEL,		"model=%s"	},
3393	{ NR_COST_CTRL_PARAMS,	NULL		},
3394};
3395
3396static const match_table_t i_lcoef_tokens = {
3397	{ I_LCOEF_RBPS,		"rbps=%u"	},
3398	{ I_LCOEF_RSEQIOPS,	"rseqiops=%u"	},
3399	{ I_LCOEF_RRANDIOPS,	"rrandiops=%u"	},
3400	{ I_LCOEF_WBPS,		"wbps=%u"	},
3401	{ I_LCOEF_WSEQIOPS,	"wseqiops=%u"	},
3402	{ I_LCOEF_WRANDIOPS,	"wrandiops=%u"	},
3403	{ NR_I_LCOEFS,		NULL		},
3404};
3405
3406static ssize_t ioc_cost_model_write(struct kernfs_open_file *of, char *input,
3407				    size_t nbytes, loff_t off)
3408{
3409	struct blkg_conf_ctx ctx;
3410	struct request_queue *q;
3411	unsigned int memflags;
3412	struct ioc *ioc;
3413	u64 u[NR_I_LCOEFS];
3414	bool user;
3415	char *body, *p;
3416	int ret;
3417
3418	blkg_conf_init(&ctx, input);
3419
3420	ret = blkg_conf_open_bdev(&ctx);
3421	if (ret)
3422		goto err;
3423
3424	body = ctx.body;
3425	q = bdev_get_queue(ctx.bdev);
3426	if (!queue_is_mq(q)) {
3427		ret = -EOPNOTSUPP;
3428		goto err;
3429	}
3430
3431	ioc = q_to_ioc(q);
3432	if (!ioc) {
3433		ret = blk_iocost_init(ctx.bdev->bd_disk);
3434		if (ret)
3435			goto err;
3436		ioc = q_to_ioc(q);
3437	}
3438
3439	memflags = blk_mq_freeze_queue(q);
3440	blk_mq_quiesce_queue(q);
3441
3442	spin_lock_irq(&ioc->lock);
3443	memcpy(u, ioc->params.i_lcoefs, sizeof(u));
3444	user = ioc->user_cost_model;
3445
3446	while ((p = strsep(&body, " \t\n"))) {
3447		substring_t args[MAX_OPT_ARGS];
3448		char buf[32];
3449		int tok;
3450		u64 v;
3451
3452		if (!*p)
3453			continue;
3454
3455		switch (match_token(p, cost_ctrl_tokens, args)) {
3456		case COST_CTRL:
3457			match_strlcpy(buf, &args[0], sizeof(buf));
3458			if (!strcmp(buf, "auto"))
3459				user = false;
3460			else if (!strcmp(buf, "user"))
3461				user = true;
3462			else
3463				goto einval;
3464			continue;
3465		case COST_MODEL:
3466			match_strlcpy(buf, &args[0], sizeof(buf));
3467			if (strcmp(buf, "linear"))
3468				goto einval;
3469			continue;
3470		}
3471
3472		tok = match_token(p, i_lcoef_tokens, args);
3473		if (tok == NR_I_LCOEFS)
3474			goto einval;
3475		if (match_u64(&args[0], &v))
3476			goto einval;
3477		u[tok] = v;
3478		user = true;
3479	}
3480
3481	if (user) {
3482		memcpy(ioc->params.i_lcoefs, u, sizeof(u));
3483		ioc->user_cost_model = true;
3484	} else {
3485		ioc->user_cost_model = false;
3486	}
3487	ioc_refresh_params(ioc, true);
3488	spin_unlock_irq(&ioc->lock);
3489
3490	blk_mq_unquiesce_queue(q);
3491	blk_mq_unfreeze_queue(q, memflags);
3492
3493	blkg_conf_exit(&ctx);
3494	return nbytes;
3495
3496einval:
3497	spin_unlock_irq(&ioc->lock);
3498
3499	blk_mq_unquiesce_queue(q);
3500	blk_mq_unfreeze_queue(q, memflags);
3501
3502	ret = -EINVAL;
3503err:
3504	blkg_conf_exit(&ctx);
3505	return ret;
3506}
3507
3508static struct cftype ioc_files[] = {
3509	{
3510		.name = "weight",
3511		.flags = CFTYPE_NOT_ON_ROOT,
3512		.seq_show = ioc_weight_show,
3513		.write = ioc_weight_write,
3514	},
3515	{
3516		.name = "cost.qos",
3517		.flags = CFTYPE_ONLY_ON_ROOT,
3518		.seq_show = ioc_qos_show,
3519		.write = ioc_qos_write,
3520	},
3521	{
3522		.name = "cost.model",
3523		.flags = CFTYPE_ONLY_ON_ROOT,
3524		.seq_show = ioc_cost_model_show,
3525		.write = ioc_cost_model_write,
3526	},
3527	{}
3528};
3529
3530static struct blkcg_policy blkcg_policy_iocost = {
3531	.dfl_cftypes	= ioc_files,
3532	.cpd_alloc_fn	= ioc_cpd_alloc,
3533	.cpd_free_fn	= ioc_cpd_free,
3534	.pd_alloc_fn	= ioc_pd_alloc,
3535	.pd_init_fn	= ioc_pd_init,
3536	.pd_free_fn	= ioc_pd_free,
3537	.pd_stat_fn	= ioc_pd_stat,
3538};
3539
3540static int __init ioc_init(void)
3541{
3542	return blkcg_policy_register(&blkcg_policy_iocost);
3543}
3544
3545static void __exit ioc_exit(void)
3546{
3547	blkcg_policy_unregister(&blkcg_policy_iocost);
3548}
3549
3550module_init(ioc_init);
3551module_exit(ioc_exit);