1 /* SPDX-License-Identifier: GPL-2.0 2 * 3 * IO cost model based controller. 4 * 5 * Copyright (C) 2019 Tejun Heo <tj@kernel.org> 6 * Copyright (C) 2019 Andy Newell <newella@fb.com> 7 * Copyright (C) 2019 Facebook 8 * 9 * One challenge of controlling IO resources is the lack of trivially 10 * observable cost metric. This is distinguished from CPU and memory where 11 * wallclock time and the number of bytes can serve as accurate enough 12 * approximations. 13 * 14 * Bandwidth and iops are the most commonly used metrics for IO devices but 15 * depending on the type and specifics of the device, different IO patterns 16 * easily lead to multiple orders of magnitude variations rendering them 17 * useless for the purpose of IO capacity distribution. While on-device 18 * time, with a lot of clutches, could serve as a useful approximation for 19 * non-queued rotational devices, this is no longer viable with modern 20 * devices, even the rotational ones. 21 * 22 * While there is no cost metric we can trivially observe, it isn't a 23 * complete mystery. For example, on a rotational device, seek cost 24 * dominates while a contiguous transfer contributes a smaller amount 25 * proportional to the size. If we can characterize at least the relative 26 * costs of these different types of IOs, it should be possible to 27 * implement a reasonable work-conserving proportional IO resource 28 * distribution. 29 * 30 * 1. IO Cost Model 31 * 32 * IO cost model estimates the cost of an IO given its basic parameters and 33 * history (e.g. the end sector of the last IO). The cost is measured in 34 * device time. If a given IO is estimated to cost 10ms, the device should 35 * be able to process ~100 of those IOs in a second. 36 * 37 * Currently, there's only one builtin cost model - linear. Each IO is 38 * classified as sequential or random and given a base cost accordingly. 39 * On top of that, a size cost proportional to the length of the IO is 40 * added. While simple, this model captures the operational 41 * characteristics of a wide varienty of devices well enough. Default 42 * paramters for several different classes of devices are provided and the 43 * parameters can be configured from userspace via 44 * /sys/fs/cgroup/io.cost.model. 45 * 46 * If needed, tools/cgroup/iocost_coef_gen.py can be used to generate 47 * device-specific coefficients. 48 * 49 * If needed, tools/cgroup/iocost_coef_gen.py can be used to generate 50 * device-specific coefficients. 51 * 52 * 2. Control Strategy 53 * 54 * The device virtual time (vtime) is used as the primary control metric. 55 * The control strategy is composed of the following three parts. 56 * 57 * 2-1. Vtime Distribution 58 * 59 * When a cgroup becomes active in terms of IOs, its hierarchical share is 60 * calculated. Please consider the following hierarchy where the numbers 61 * inside parentheses denote the configured weights. 62 * 63 * root 64 * / \ 65 * A (w:100) B (w:300) 66 * / \ 67 * A0 (w:100) A1 (w:100) 68 * 69 * If B is idle and only A0 and A1 are actively issuing IOs, as the two are 70 * of equal weight, each gets 50% share. If then B starts issuing IOs, B 71 * gets 300/(100+300) or 75% share, and A0 and A1 equally splits the rest, 72 * 12.5% each. The distribution mechanism only cares about these flattened 73 * shares. They're called hweights (hierarchical weights) and always add 74 * upto 1 (HWEIGHT_WHOLE). 75 * 76 * A given cgroup's vtime runs slower in inverse proportion to its hweight. 77 * For example, with 12.5% weight, A0's time runs 8 times slower (100/12.5) 78 * against the device vtime - an IO which takes 10ms on the underlying 79 * device is considered to take 80ms on A0. 80 * 81 * This constitutes the basis of IO capacity distribution. Each cgroup's 82 * vtime is running at a rate determined by its hweight. A cgroup tracks 83 * the vtime consumed by past IOs and can issue a new IO iff doing so 84 * wouldn't outrun the current device vtime. Otherwise, the IO is 85 * suspended until the vtime has progressed enough to cover it. 86 * 87 * 2-2. Vrate Adjustment 88 * 89 * It's unrealistic to expect the cost model to be perfect. There are too 90 * many devices and even on the same device the overall performance 91 * fluctuates depending on numerous factors such as IO mixture and device 92 * internal garbage collection. The controller needs to adapt dynamically. 93 * 94 * This is achieved by adjusting the overall IO rate according to how busy 95 * the device is. If the device becomes overloaded, we're sending down too 96 * many IOs and should generally slow down. If there are waiting issuers 97 * but the device isn't saturated, we're issuing too few and should 98 * generally speed up. 99 * 100 * To slow down, we lower the vrate - the rate at which the device vtime 101 * passes compared to the wall clock. For example, if the vtime is running 102 * at the vrate of 75%, all cgroups added up would only be able to issue 103 * 750ms worth of IOs per second, and vice-versa for speeding up. 104 * 105 * Device business is determined using two criteria - rq wait and 106 * completion latencies. 107 * 108 * When a device gets saturated, the on-device and then the request queues 109 * fill up and a bio which is ready to be issued has to wait for a request 110 * to become available. When this delay becomes noticeable, it's a clear 111 * indication that the device is saturated and we lower the vrate. This 112 * saturation signal is fairly conservative as it only triggers when both 113 * hardware and software queues are filled up, and is used as the default 114 * busy signal. 115 * 116 * As devices can have deep queues and be unfair in how the queued commands 117 * are executed, soley depending on rq wait may not result in satisfactory 118 * control quality. For a better control quality, completion latency QoS 119 * parameters can be configured so that the device is considered saturated 120 * if N'th percentile completion latency rises above the set point. 121 * 122 * The completion latency requirements are a function of both the 123 * underlying device characteristics and the desired IO latency quality of 124 * service. There is an inherent trade-off - the tighter the latency QoS, 125 * the higher the bandwidth lossage. Latency QoS is disabled by default 126 * and can be set through /sys/fs/cgroup/io.cost.qos. 127 * 128 * 2-3. Work Conservation 129 * 130 * Imagine two cgroups A and B with equal weights. A is issuing a small IO 131 * periodically while B is sending out enough parallel IOs to saturate the 132 * device on its own. Let's say A's usage amounts to 100ms worth of IO 133 * cost per second, i.e., 10% of the device capacity. The naive 134 * distribution of half and half would lead to 60% utilization of the 135 * device, a significant reduction in the total amount of work done 136 * compared to free-for-all competition. This is too high a cost to pay 137 * for IO control. 138 * 139 * To conserve the total amount of work done, we keep track of how much 140 * each active cgroup is actually using and yield part of its weight if 141 * there are other cgroups which can make use of it. In the above case, 142 * A's weight will be lowered so that it hovers above the actual usage and 143 * B would be able to use the rest. 144 * 145 * As we don't want to penalize a cgroup for donating its weight, the 146 * surplus weight adjustment factors in a margin and has an immediate 147 * snapback mechanism in case the cgroup needs more IO vtime for itself. 148 * 149 * Note that adjusting down surplus weights has the same effects as 150 * accelerating vtime for other cgroups and work conservation can also be 151 * implemented by adjusting vrate dynamically. However, squaring who can 152 * donate and should take back how much requires hweight propagations 153 * anyway making it easier to implement and understand as a separate 154 * mechanism. 155 * 156 * 3. Monitoring 157 * 158 * Instead of debugfs or other clumsy monitoring mechanisms, this 159 * controller uses a drgn based monitoring script - 160 * tools/cgroup/iocost_monitor.py. For details on drgn, please see 161 * https://github.com/osandov/drgn. The ouput looks like the following. 162 * 163 * sdb RUN per=300ms cur_per=234.218:v203.695 busy= +1 vrate= 62.12% 164 * active weight hweight% inflt% dbt delay usages% 165 * test/a * 50/ 50 33.33/ 33.33 27.65 2 0*041 033:033:033 166 * test/b * 100/ 100 66.67/ 66.67 17.56 0 0*000 066:079:077 167 * 168 * - per : Timer period 169 * - cur_per : Internal wall and device vtime clock 170 * - vrate : Device virtual time rate against wall clock 171 * - weight : Surplus-adjusted and configured weights 172 * - hweight : Surplus-adjusted and configured hierarchical weights 173 * - inflt : The percentage of in-flight IO cost at the end of last period 174 * - del_ms : Deferred issuer delay induction level and duration 175 * - usages : Usage history 176 */ 177 178 #include <linux/kernel.h> 179 #include <linux/module.h> 180 #include <linux/timer.h> 181 #include <linux/time64.h> 182 #include <linux/parser.h> 183 #include <linux/sched/signal.h> 184 #include <linux/blk-cgroup.h> 185 #include "blk-rq-qos.h" 186 #include "blk-stat.h" 187 #include "blk-wbt.h" 188 189 #ifdef CONFIG_TRACEPOINTS 190 191 /* copied from TRACE_CGROUP_PATH, see cgroup-internal.h */ 192 #define TRACE_IOCG_PATH_LEN 1024 193 static DEFINE_SPINLOCK(trace_iocg_path_lock); 194 static char trace_iocg_path[TRACE_IOCG_PATH_LEN]; 195 196 #define TRACE_IOCG_PATH(type, iocg, ...) \ 197 do { \ 198 unsigned long flags; \ 199 if (trace_iocost_##type##_enabled()) { \ 200 spin_lock_irqsave(&trace_iocg_path_lock, flags); \ 201 cgroup_path(iocg_to_blkg(iocg)->blkcg->css.cgroup, \ 202 trace_iocg_path, TRACE_IOCG_PATH_LEN); \ 203 trace_iocost_##type(iocg, trace_iocg_path, \ 204 ##__VA_ARGS__); \ 205 spin_unlock_irqrestore(&trace_iocg_path_lock, flags); \ 206 } \ 207 } while (0) 208 209 #else /* CONFIG_TRACE_POINTS */ 210 #define TRACE_IOCG_PATH(type, iocg, ...) do { } while (0) 211 #endif /* CONFIG_TRACE_POINTS */ 212 213 enum { 214 MILLION = 1000000, 215 216 /* timer period is calculated from latency requirements, bound it */ 217 MIN_PERIOD = USEC_PER_MSEC, 218 MAX_PERIOD = USEC_PER_SEC, 219 220 /* 221 * A cgroup's vtime can run 50% behind the device vtime, which 222 * serves as its IO credit buffer. Surplus weight adjustment is 223 * immediately canceled if the vtime margin runs below 10%. 224 */ 225 MARGIN_PCT = 50, 226 INUSE_MARGIN_PCT = 10, 227 228 /* Have some play in waitq timer operations */ 229 WAITQ_TIMER_MARGIN_PCT = 5, 230 231 /* 232 * vtime can wrap well within a reasonable uptime when vrate is 233 * consistently raised. Don't trust recorded cgroup vtime if the 234 * period counter indicates that it's older than 5mins. 235 */ 236 VTIME_VALID_DUR = 300 * USEC_PER_SEC, 237 238 /* 239 * Remember the past three non-zero usages and use the max for 240 * surplus calculation. Three slots guarantee that we remember one 241 * full period usage from the last active stretch even after 242 * partial deactivation and re-activation periods. Don't start 243 * giving away weight before collecting two data points to prevent 244 * hweight adjustments based on one partial activation period. 245 */ 246 NR_USAGE_SLOTS = 3, 247 MIN_VALID_USAGES = 2, 248 249 /* 1/64k is granular enough and can easily be handled w/ u32 */ 250 HWEIGHT_WHOLE = 1 << 16, 251 252 /* 253 * As vtime is used to calculate the cost of each IO, it needs to 254 * be fairly high precision. For example, it should be able to 255 * represent the cost of a single page worth of discard with 256 * suffificient accuracy. At the same time, it should be able to 257 * represent reasonably long enough durations to be useful and 258 * convenient during operation. 259 * 260 * 1s worth of vtime is 2^37. This gives us both sub-nanosecond 261 * granularity and days of wrap-around time even at extreme vrates. 262 */ 263 VTIME_PER_SEC_SHIFT = 37, 264 VTIME_PER_SEC = 1LLU << VTIME_PER_SEC_SHIFT, 265 VTIME_PER_USEC = VTIME_PER_SEC / USEC_PER_SEC, 266 267 /* bound vrate adjustments within two orders of magnitude */ 268 VRATE_MIN_PPM = 10000, /* 1% */ 269 VRATE_MAX_PPM = 100000000, /* 10000% */ 270 271 VRATE_MIN = VTIME_PER_USEC * VRATE_MIN_PPM / MILLION, 272 VRATE_CLAMP_ADJ_PCT = 4, 273 274 /* if IOs end up waiting for requests, issue less */ 275 RQ_WAIT_BUSY_PCT = 5, 276 277 /* unbusy hysterisis */ 278 UNBUSY_THR_PCT = 75, 279 280 /* don't let cmds which take a very long time pin lagging for too long */ 281 MAX_LAGGING_PERIODS = 10, 282 283 /* 284 * If usage% * 1.25 + 2% is lower than hweight% by more than 3%, 285 * donate the surplus. 286 */ 287 SURPLUS_SCALE_PCT = 125, /* * 125% */ 288 SURPLUS_SCALE_ABS = HWEIGHT_WHOLE / 50, /* + 2% */ 289 SURPLUS_MIN_ADJ_DELTA = HWEIGHT_WHOLE / 33, /* 3% */ 290 291 /* switch iff the conditions are met for longer than this */ 292 AUTOP_CYCLE_NSEC = 10LLU * NSEC_PER_SEC, 293 294 /* 295 * Count IO size in 4k pages. The 12bit shift helps keeping 296 * size-proportional components of cost calculation in closer 297 * numbers of digits to per-IO cost components. 298 */ 299 IOC_PAGE_SHIFT = 12, 300 IOC_PAGE_SIZE = 1 << IOC_PAGE_SHIFT, 301 IOC_SECT_TO_PAGE_SHIFT = IOC_PAGE_SHIFT - SECTOR_SHIFT, 302 303 /* if apart further than 16M, consider randio for linear model */ 304 LCOEF_RANDIO_PAGES = 4096, 305 }; 306 307 enum ioc_running { 308 IOC_IDLE, 309 IOC_RUNNING, 310 IOC_STOP, 311 }; 312 313 /* io.cost.qos controls including per-dev enable of the whole controller */ 314 enum { 315 QOS_ENABLE, 316 QOS_CTRL, 317 NR_QOS_CTRL_PARAMS, 318 }; 319 320 /* io.cost.qos params */ 321 enum { 322 QOS_RPPM, 323 QOS_RLAT, 324 QOS_WPPM, 325 QOS_WLAT, 326 QOS_MIN, 327 QOS_MAX, 328 NR_QOS_PARAMS, 329 }; 330 331 /* io.cost.model controls */ 332 enum { 333 COST_CTRL, 334 COST_MODEL, 335 NR_COST_CTRL_PARAMS, 336 }; 337 338 /* builtin linear cost model coefficients */ 339 enum { 340 I_LCOEF_RBPS, 341 I_LCOEF_RSEQIOPS, 342 I_LCOEF_RRANDIOPS, 343 I_LCOEF_WBPS, 344 I_LCOEF_WSEQIOPS, 345 I_LCOEF_WRANDIOPS, 346 NR_I_LCOEFS, 347 }; 348 349 enum { 350 LCOEF_RPAGE, 351 LCOEF_RSEQIO, 352 LCOEF_RRANDIO, 353 LCOEF_WPAGE, 354 LCOEF_WSEQIO, 355 LCOEF_WRANDIO, 356 NR_LCOEFS, 357 }; 358 359 enum { 360 AUTOP_INVALID, 361 AUTOP_HDD, 362 AUTOP_SSD_QD1, 363 AUTOP_SSD_DFL, 364 AUTOP_SSD_FAST, 365 }; 366 367 struct ioc_gq; 368 369 struct ioc_params { 370 u32 qos[NR_QOS_PARAMS]; 371 u64 i_lcoefs[NR_I_LCOEFS]; 372 u64 lcoefs[NR_LCOEFS]; 373 u32 too_fast_vrate_pct; 374 u32 too_slow_vrate_pct; 375 }; 376 377 struct ioc_missed { 378 u32 nr_met; 379 u32 nr_missed; 380 u32 last_met; 381 u32 last_missed; 382 }; 383 384 struct ioc_pcpu_stat { 385 struct ioc_missed missed[2]; 386 387 u64 rq_wait_ns; 388 u64 last_rq_wait_ns; 389 }; 390 391 /* per device */ 392 struct ioc { 393 struct rq_qos rqos; 394 395 bool enabled; 396 397 struct ioc_params params; 398 u32 period_us; 399 u32 margin_us; 400 u64 vrate_min; 401 u64 vrate_max; 402 403 spinlock_t lock; 404 struct timer_list timer; 405 struct list_head active_iocgs; /* active cgroups */ 406 struct ioc_pcpu_stat __percpu *pcpu_stat; 407 408 enum ioc_running running; 409 atomic64_t vtime_rate; 410 411 seqcount_t period_seqcount; 412 u32 period_at; /* wallclock starttime */ 413 u64 period_at_vtime; /* vtime starttime */ 414 415 atomic64_t cur_period; /* inc'd each period */ 416 int busy_level; /* saturation history */ 417 418 u64 inuse_margin_vtime; 419 bool weights_updated; 420 atomic_t hweight_gen; /* for lazy hweights */ 421 422 u64 autop_too_fast_at; 423 u64 autop_too_slow_at; 424 int autop_idx; 425 bool user_qos_params:1; 426 bool user_cost_model:1; 427 }; 428 429 /* per device-cgroup pair */ 430 struct ioc_gq { 431 struct blkg_policy_data pd; 432 struct ioc *ioc; 433 434 /* 435 * A iocg can get its weight from two sources - an explicit 436 * per-device-cgroup configuration or the default weight of the 437 * cgroup. `cfg_weight` is the explicit per-device-cgroup 438 * configuration. `weight` is the effective considering both 439 * sources. 440 * 441 * When an idle cgroup becomes active its `active` goes from 0 to 442 * `weight`. `inuse` is the surplus adjusted active weight. 443 * `active` and `inuse` are used to calculate `hweight_active` and 444 * `hweight_inuse`. 445 * 446 * `last_inuse` remembers `inuse` while an iocg is idle to persist 447 * surplus adjustments. 448 */ 449 u32 cfg_weight; 450 u32 weight; 451 u32 active; 452 u32 inuse; 453 u32 last_inuse; 454 455 sector_t cursor; /* to detect randio */ 456 457 /* 458 * `vtime` is this iocg's vtime cursor which progresses as IOs are 459 * issued. If lagging behind device vtime, the delta represents 460 * the currently available IO budget. If runnning ahead, the 461 * overage. 462 * 463 * `vtime_done` is the same but progressed on completion rather 464 * than issue. The delta behind `vtime` represents the cost of 465 * currently in-flight IOs. 466 * 467 * `last_vtime` is used to remember `vtime` at the end of the last 468 * period to calculate utilization. 469 */ 470 atomic64_t vtime; 471 atomic64_t done_vtime; 472 atomic64_t abs_vdebt; 473 u64 last_vtime; 474 475 /* 476 * The period this iocg was last active in. Used for deactivation 477 * and invalidating `vtime`. 478 */ 479 atomic64_t active_period; 480 struct list_head active_list; 481 482 /* see __propagate_active_weight() and current_hweight() for details */ 483 u64 child_active_sum; 484 u64 child_inuse_sum; 485 int hweight_gen; 486 u32 hweight_active; 487 u32 hweight_inuse; 488 bool has_surplus; 489 490 struct wait_queue_head waitq; 491 struct hrtimer waitq_timer; 492 struct hrtimer delay_timer; 493 494 /* usage is recorded as fractions of HWEIGHT_WHOLE */ 495 int usage_idx; 496 u32 usages[NR_USAGE_SLOTS]; 497 498 /* this iocg's depth in the hierarchy and ancestors including self */ 499 int level; 500 struct ioc_gq *ancestors[]; 501 }; 502 503 /* per cgroup */ 504 struct ioc_cgrp { 505 struct blkcg_policy_data cpd; 506 unsigned int dfl_weight; 507 }; 508 509 struct ioc_now { 510 u64 now_ns; 511 u32 now; 512 u64 vnow; 513 u64 vrate; 514 }; 515 516 struct iocg_wait { 517 struct wait_queue_entry wait; 518 struct bio *bio; 519 u64 abs_cost; 520 bool committed; 521 }; 522 523 struct iocg_wake_ctx { 524 struct ioc_gq *iocg; 525 u32 hw_inuse; 526 s64 vbudget; 527 }; 528 529 static const struct ioc_params autop[] = { 530 [AUTOP_HDD] = { 531 .qos = { 532 [QOS_RLAT] = 250000, /* 250ms */ 533 [QOS_WLAT] = 250000, 534 [QOS_MIN] = VRATE_MIN_PPM, 535 [QOS_MAX] = VRATE_MAX_PPM, 536 }, 537 .i_lcoefs = { 538 [I_LCOEF_RBPS] = 174019176, 539 [I_LCOEF_RSEQIOPS] = 41708, 540 [I_LCOEF_RRANDIOPS] = 370, 541 [I_LCOEF_WBPS] = 178075866, 542 [I_LCOEF_WSEQIOPS] = 42705, 543 [I_LCOEF_WRANDIOPS] = 378, 544 }, 545 }, 546 [AUTOP_SSD_QD1] = { 547 .qos = { 548 [QOS_RLAT] = 25000, /* 25ms */ 549 [QOS_WLAT] = 25000, 550 [QOS_MIN] = VRATE_MIN_PPM, 551 [QOS_MAX] = VRATE_MAX_PPM, 552 }, 553 .i_lcoefs = { 554 [I_LCOEF_RBPS] = 245855193, 555 [I_LCOEF_RSEQIOPS] = 61575, 556 [I_LCOEF_RRANDIOPS] = 6946, 557 [I_LCOEF_WBPS] = 141365009, 558 [I_LCOEF_WSEQIOPS] = 33716, 559 [I_LCOEF_WRANDIOPS] = 26796, 560 }, 561 }, 562 [AUTOP_SSD_DFL] = { 563 .qos = { 564 [QOS_RLAT] = 25000, /* 25ms */ 565 [QOS_WLAT] = 25000, 566 [QOS_MIN] = VRATE_MIN_PPM, 567 [QOS_MAX] = VRATE_MAX_PPM, 568 }, 569 .i_lcoefs = { 570 [I_LCOEF_RBPS] = 488636629, 571 [I_LCOEF_RSEQIOPS] = 8932, 572 [I_LCOEF_RRANDIOPS] = 8518, 573 [I_LCOEF_WBPS] = 427891549, 574 [I_LCOEF_WSEQIOPS] = 28755, 575 [I_LCOEF_WRANDIOPS] = 21940, 576 }, 577 .too_fast_vrate_pct = 500, 578 }, 579 [AUTOP_SSD_FAST] = { 580 .qos = { 581 [QOS_RLAT] = 5000, /* 5ms */ 582 [QOS_WLAT] = 5000, 583 [QOS_MIN] = VRATE_MIN_PPM, 584 [QOS_MAX] = VRATE_MAX_PPM, 585 }, 586 .i_lcoefs = { 587 [I_LCOEF_RBPS] = 3102524156LLU, 588 [I_LCOEF_RSEQIOPS] = 724816, 589 [I_LCOEF_RRANDIOPS] = 778122, 590 [I_LCOEF_WBPS] = 1742780862LLU, 591 [I_LCOEF_WSEQIOPS] = 425702, 592 [I_LCOEF_WRANDIOPS] = 443193, 593 }, 594 .too_slow_vrate_pct = 10, 595 }, 596 }; 597 598 /* 599 * vrate adjust percentages indexed by ioc->busy_level. We adjust up on 600 * vtime credit shortage and down on device saturation. 601 */ 602 static u32 vrate_adj_pct[] = 603 { 0, 0, 0, 0, 604 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 605 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 606 4, 4, 4, 4, 4, 4, 4, 4, 8, 8, 8, 8, 8, 8, 8, 8, 16 }; 607 608 static struct blkcg_policy blkcg_policy_iocost; 609 610 /* accessors and helpers */ 611 static struct ioc *rqos_to_ioc(struct rq_qos *rqos) 612 { 613 return container_of(rqos, struct ioc, rqos); 614 } 615 616 static struct ioc *q_to_ioc(struct request_queue *q) 617 { 618 return rqos_to_ioc(rq_qos_id(q, RQ_QOS_COST)); 619 } 620 621 static const char *q_name(struct request_queue *q) 622 { 623 if (test_bit(QUEUE_FLAG_REGISTERED, &q->queue_flags)) 624 return kobject_name(q->kobj.parent); 625 else 626 return "<unknown>"; 627 } 628 629 static const char __maybe_unused *ioc_name(struct ioc *ioc) 630 { 631 return q_name(ioc->rqos.q); 632 } 633 634 static struct ioc_gq *pd_to_iocg(struct blkg_policy_data *pd) 635 { 636 return pd ? container_of(pd, struct ioc_gq, pd) : NULL; 637 } 638 639 static struct ioc_gq *blkg_to_iocg(struct blkcg_gq *blkg) 640 { 641 return pd_to_iocg(blkg_to_pd(blkg, &blkcg_policy_iocost)); 642 } 643 644 static struct blkcg_gq *iocg_to_blkg(struct ioc_gq *iocg) 645 { 646 return pd_to_blkg(&iocg->pd); 647 } 648 649 static struct ioc_cgrp *blkcg_to_iocc(struct blkcg *blkcg) 650 { 651 return container_of(blkcg_to_cpd(blkcg, &blkcg_policy_iocost), 652 struct ioc_cgrp, cpd); 653 } 654 655 /* 656 * Scale @abs_cost to the inverse of @hw_inuse. The lower the hierarchical 657 * weight, the more expensive each IO. Must round up. 658 */ 659 static u64 abs_cost_to_cost(u64 abs_cost, u32 hw_inuse) 660 { 661 return DIV64_U64_ROUND_UP(abs_cost * HWEIGHT_WHOLE, hw_inuse); 662 } 663 664 /* 665 * The inverse of abs_cost_to_cost(). Must round up. 666 */ 667 static u64 cost_to_abs_cost(u64 cost, u32 hw_inuse) 668 { 669 return DIV64_U64_ROUND_UP(cost * hw_inuse, HWEIGHT_WHOLE); 670 } 671 672 static void iocg_commit_bio(struct ioc_gq *iocg, struct bio *bio, u64 cost) 673 { 674 bio->bi_iocost_cost = cost; 675 atomic64_add(cost, &iocg->vtime); 676 } 677 678 #define CREATE_TRACE_POINTS 679 #include <trace/events/iocost.h> 680 681 /* latency Qos params changed, update period_us and all the dependent params */ 682 static void ioc_refresh_period_us(struct ioc *ioc) 683 { 684 u32 ppm, lat, multi, period_us; 685 686 lockdep_assert_held(&ioc->lock); 687 688 /* pick the higher latency target */ 689 if (ioc->params.qos[QOS_RLAT] >= ioc->params.qos[QOS_WLAT]) { 690 ppm = ioc->params.qos[QOS_RPPM]; 691 lat = ioc->params.qos[QOS_RLAT]; 692 } else { 693 ppm = ioc->params.qos[QOS_WPPM]; 694 lat = ioc->params.qos[QOS_WLAT]; 695 } 696 697 /* 698 * We want the period to be long enough to contain a healthy number 699 * of IOs while short enough for granular control. Define it as a 700 * multiple of the latency target. Ideally, the multiplier should 701 * be scaled according to the percentile so that it would nominally 702 * contain a certain number of requests. Let's be simpler and 703 * scale it linearly so that it's 2x >= pct(90) and 10x at pct(50). 704 */ 705 if (ppm) 706 multi = max_t(u32, (MILLION - ppm) / 50000, 2); 707 else 708 multi = 2; 709 period_us = multi * lat; 710 period_us = clamp_t(u32, period_us, MIN_PERIOD, MAX_PERIOD); 711 712 /* calculate dependent params */ 713 ioc->period_us = period_us; 714 ioc->margin_us = period_us * MARGIN_PCT / 100; 715 ioc->inuse_margin_vtime = DIV64_U64_ROUND_UP( 716 period_us * VTIME_PER_USEC * INUSE_MARGIN_PCT, 100); 717 } 718 719 static int ioc_autop_idx(struct ioc *ioc) 720 { 721 int idx = ioc->autop_idx; 722 const struct ioc_params *p = &autop[idx]; 723 u32 vrate_pct; 724 u64 now_ns; 725 726 /* rotational? */ 727 if (!blk_queue_nonrot(ioc->rqos.q)) 728 return AUTOP_HDD; 729 730 /* handle SATA SSDs w/ broken NCQ */ 731 if (blk_queue_depth(ioc->rqos.q) == 1) 732 return AUTOP_SSD_QD1; 733 734 /* use one of the normal ssd sets */ 735 if (idx < AUTOP_SSD_DFL) 736 return AUTOP_SSD_DFL; 737 738 /* if user is overriding anything, maintain what was there */ 739 if (ioc->user_qos_params || ioc->user_cost_model) 740 return idx; 741 742 /* step up/down based on the vrate */ 743 vrate_pct = div64_u64(atomic64_read(&ioc->vtime_rate) * 100, 744 VTIME_PER_USEC); 745 now_ns = ktime_get_ns(); 746 747 if (p->too_fast_vrate_pct && p->too_fast_vrate_pct <= vrate_pct) { 748 if (!ioc->autop_too_fast_at) 749 ioc->autop_too_fast_at = now_ns; 750 if (now_ns - ioc->autop_too_fast_at >= AUTOP_CYCLE_NSEC) 751 return idx + 1; 752 } else { 753 ioc->autop_too_fast_at = 0; 754 } 755 756 if (p->too_slow_vrate_pct && p->too_slow_vrate_pct >= vrate_pct) { 757 if (!ioc->autop_too_slow_at) 758 ioc->autop_too_slow_at = now_ns; 759 if (now_ns - ioc->autop_too_slow_at >= AUTOP_CYCLE_NSEC) 760 return idx - 1; 761 } else { 762 ioc->autop_too_slow_at = 0; 763 } 764 765 return idx; 766 } 767 768 /* 769 * Take the followings as input 770 * 771 * @bps maximum sequential throughput 772 * @seqiops maximum sequential 4k iops 773 * @randiops maximum random 4k iops 774 * 775 * and calculate the linear model cost coefficients. 776 * 777 * *@page per-page cost 1s / (@bps / 4096) 778 * *@seqio base cost of a seq IO max((1s / @seqiops) - *@page, 0) 779 * @randiops base cost of a rand IO max((1s / @randiops) - *@page, 0) 780 */ 781 static void calc_lcoefs(u64 bps, u64 seqiops, u64 randiops, 782 u64 *page, u64 *seqio, u64 *randio) 783 { 784 u64 v; 785 786 *page = *seqio = *randio = 0; 787 788 if (bps) 789 *page = DIV64_U64_ROUND_UP(VTIME_PER_SEC, 790 DIV_ROUND_UP_ULL(bps, IOC_PAGE_SIZE)); 791 792 if (seqiops) { 793 v = DIV64_U64_ROUND_UP(VTIME_PER_SEC, seqiops); 794 if (v > *page) 795 *seqio = v - *page; 796 } 797 798 if (randiops) { 799 v = DIV64_U64_ROUND_UP(VTIME_PER_SEC, randiops); 800 if (v > *page) 801 *randio = v - *page; 802 } 803 } 804 805 static void ioc_refresh_lcoefs(struct ioc *ioc) 806 { 807 u64 *u = ioc->params.i_lcoefs; 808 u64 *c = ioc->params.lcoefs; 809 810 calc_lcoefs(u[I_LCOEF_RBPS], u[I_LCOEF_RSEQIOPS], u[I_LCOEF_RRANDIOPS], 811 &c[LCOEF_RPAGE], &c[LCOEF_RSEQIO], &c[LCOEF_RRANDIO]); 812 calc_lcoefs(u[I_LCOEF_WBPS], u[I_LCOEF_WSEQIOPS], u[I_LCOEF_WRANDIOPS], 813 &c[LCOEF_WPAGE], &c[LCOEF_WSEQIO], &c[LCOEF_WRANDIO]); 814 } 815 816 static bool ioc_refresh_params(struct ioc *ioc, bool force) 817 { 818 const struct ioc_params *p; 819 int idx; 820 821 lockdep_assert_held(&ioc->lock); 822 823 idx = ioc_autop_idx(ioc); 824 p = &autop[idx]; 825 826 if (idx == ioc->autop_idx && !force) 827 return false; 828 829 if (idx != ioc->autop_idx) 830 atomic64_set(&ioc->vtime_rate, VTIME_PER_USEC); 831 832 ioc->autop_idx = idx; 833 ioc->autop_too_fast_at = 0; 834 ioc->autop_too_slow_at = 0; 835 836 if (!ioc->user_qos_params) 837 memcpy(ioc->params.qos, p->qos, sizeof(p->qos)); 838 if (!ioc->user_cost_model) 839 memcpy(ioc->params.i_lcoefs, p->i_lcoefs, sizeof(p->i_lcoefs)); 840 841 ioc_refresh_period_us(ioc); 842 ioc_refresh_lcoefs(ioc); 843 844 ioc->vrate_min = DIV64_U64_ROUND_UP((u64)ioc->params.qos[QOS_MIN] * 845 VTIME_PER_USEC, MILLION); 846 ioc->vrate_max = div64_u64((u64)ioc->params.qos[QOS_MAX] * 847 VTIME_PER_USEC, MILLION); 848 849 return true; 850 } 851 852 /* take a snapshot of the current [v]time and vrate */ 853 static void ioc_now(struct ioc *ioc, struct ioc_now *now) 854 { 855 unsigned seq; 856 857 now->now_ns = ktime_get(); 858 now->now = ktime_to_us(now->now_ns); 859 now->vrate = atomic64_read(&ioc->vtime_rate); 860 861 /* 862 * The current vtime is 863 * 864 * vtime at period start + (wallclock time since the start) * vrate 865 * 866 * As a consistent snapshot of `period_at_vtime` and `period_at` is 867 * needed, they're seqcount protected. 868 */ 869 do { 870 seq = read_seqcount_begin(&ioc->period_seqcount); 871 now->vnow = ioc->period_at_vtime + 872 (now->now - ioc->period_at) * now->vrate; 873 } while (read_seqcount_retry(&ioc->period_seqcount, seq)); 874 } 875 876 static void ioc_start_period(struct ioc *ioc, struct ioc_now *now) 877 { 878 lockdep_assert_held(&ioc->lock); 879 WARN_ON_ONCE(ioc->running != IOC_RUNNING); 880 881 write_seqcount_begin(&ioc->period_seqcount); 882 ioc->period_at = now->now; 883 ioc->period_at_vtime = now->vnow; 884 write_seqcount_end(&ioc->period_seqcount); 885 886 ioc->timer.expires = jiffies + usecs_to_jiffies(ioc->period_us); 887 add_timer(&ioc->timer); 888 } 889 890 /* 891 * Update @iocg's `active` and `inuse` to @active and @inuse, update level 892 * weight sums and propagate upwards accordingly. 893 */ 894 static void __propagate_active_weight(struct ioc_gq *iocg, u32 active, u32 inuse) 895 { 896 struct ioc *ioc = iocg->ioc; 897 int lvl; 898 899 lockdep_assert_held(&ioc->lock); 900 901 inuse = min(active, inuse); 902 903 for (lvl = iocg->level - 1; lvl >= 0; lvl--) { 904 struct ioc_gq *parent = iocg->ancestors[lvl]; 905 struct ioc_gq *child = iocg->ancestors[lvl + 1]; 906 u32 parent_active = 0, parent_inuse = 0; 907 908 /* update the level sums */ 909 parent->child_active_sum += (s32)(active - child->active); 910 parent->child_inuse_sum += (s32)(inuse - child->inuse); 911 /* apply the udpates */ 912 child->active = active; 913 child->inuse = inuse; 914 915 /* 916 * The delta between inuse and active sums indicates that 917 * that much of weight is being given away. Parent's inuse 918 * and active should reflect the ratio. 919 */ 920 if (parent->child_active_sum) { 921 parent_active = parent->weight; 922 parent_inuse = DIV64_U64_ROUND_UP( 923 parent_active * parent->child_inuse_sum, 924 parent->child_active_sum); 925 } 926 927 /* do we need to keep walking up? */ 928 if (parent_active == parent->active && 929 parent_inuse == parent->inuse) 930 break; 931 932 active = parent_active; 933 inuse = parent_inuse; 934 } 935 936 ioc->weights_updated = true; 937 } 938 939 static void commit_active_weights(struct ioc *ioc) 940 { 941 lockdep_assert_held(&ioc->lock); 942 943 if (ioc->weights_updated) { 944 /* paired with rmb in current_hweight(), see there */ 945 smp_wmb(); 946 atomic_inc(&ioc->hweight_gen); 947 ioc->weights_updated = false; 948 } 949 } 950 951 static void propagate_active_weight(struct ioc_gq *iocg, u32 active, u32 inuse) 952 { 953 __propagate_active_weight(iocg, active, inuse); 954 commit_active_weights(iocg->ioc); 955 } 956 957 static void current_hweight(struct ioc_gq *iocg, u32 *hw_activep, u32 *hw_inusep) 958 { 959 struct ioc *ioc = iocg->ioc; 960 int lvl; 961 u32 hwa, hwi; 962 int ioc_gen; 963 964 /* hot path - if uptodate, use cached */ 965 ioc_gen = atomic_read(&ioc->hweight_gen); 966 if (ioc_gen == iocg->hweight_gen) 967 goto out; 968 969 /* 970 * Paired with wmb in commit_active_weights(). If we saw the 971 * updated hweight_gen, all the weight updates from 972 * __propagate_active_weight() are visible too. 973 * 974 * We can race with weight updates during calculation and get it 975 * wrong. However, hweight_gen would have changed and a future 976 * reader will recalculate and we're guaranteed to discard the 977 * wrong result soon. 978 */ 979 smp_rmb(); 980 981 hwa = hwi = HWEIGHT_WHOLE; 982 for (lvl = 0; lvl <= iocg->level - 1; lvl++) { 983 struct ioc_gq *parent = iocg->ancestors[lvl]; 984 struct ioc_gq *child = iocg->ancestors[lvl + 1]; 985 u32 active_sum = READ_ONCE(parent->child_active_sum); 986 u32 inuse_sum = READ_ONCE(parent->child_inuse_sum); 987 u32 active = READ_ONCE(child->active); 988 u32 inuse = READ_ONCE(child->inuse); 989 990 /* we can race with deactivations and either may read as zero */ 991 if (!active_sum || !inuse_sum) 992 continue; 993 994 active_sum = max(active, active_sum); 995 hwa = hwa * active / active_sum; /* max 16bits * 10000 */ 996 997 inuse_sum = max(inuse, inuse_sum); 998 hwi = hwi * inuse / inuse_sum; /* max 16bits * 10000 */ 999 } 1000 1001 iocg->hweight_active = max_t(u32, hwa, 1); 1002 iocg->hweight_inuse = max_t(u32, hwi, 1); 1003 iocg->hweight_gen = ioc_gen; 1004 out: 1005 if (hw_activep) 1006 *hw_activep = iocg->hweight_active; 1007 if (hw_inusep) 1008 *hw_inusep = iocg->hweight_inuse; 1009 } 1010 1011 static void weight_updated(struct ioc_gq *iocg) 1012 { 1013 struct ioc *ioc = iocg->ioc; 1014 struct blkcg_gq *blkg = iocg_to_blkg(iocg); 1015 struct ioc_cgrp *iocc = blkcg_to_iocc(blkg->blkcg); 1016 u32 weight; 1017 1018 lockdep_assert_held(&ioc->lock); 1019 1020 weight = iocg->cfg_weight ?: iocc->dfl_weight; 1021 if (weight != iocg->weight && iocg->active) 1022 propagate_active_weight(iocg, weight, 1023 DIV64_U64_ROUND_UP(iocg->inuse * weight, iocg->weight)); 1024 iocg->weight = weight; 1025 } 1026 1027 static bool iocg_activate(struct ioc_gq *iocg, struct ioc_now *now) 1028 { 1029 struct ioc *ioc = iocg->ioc; 1030 u64 last_period, cur_period, max_period_delta; 1031 u64 vtime, vmargin, vmin; 1032 int i; 1033 1034 /* 1035 * If seem to be already active, just update the stamp to tell the 1036 * timer that we're still active. We don't mind occassional races. 1037 */ 1038 if (!list_empty(&iocg->active_list)) { 1039 ioc_now(ioc, now); 1040 cur_period = atomic64_read(&ioc->cur_period); 1041 if (atomic64_read(&iocg->active_period) != cur_period) 1042 atomic64_set(&iocg->active_period, cur_period); 1043 return true; 1044 } 1045 1046 /* racy check on internal node IOs, treat as root level IOs */ 1047 if (iocg->child_active_sum) 1048 return false; 1049 1050 spin_lock_irq(&ioc->lock); 1051 1052 ioc_now(ioc, now); 1053 1054 /* update period */ 1055 cur_period = atomic64_read(&ioc->cur_period); 1056 last_period = atomic64_read(&iocg->active_period); 1057 atomic64_set(&iocg->active_period, cur_period); 1058 1059 /* already activated or breaking leaf-only constraint? */ 1060 for (i = iocg->level; i > 0; i--) 1061 if (!list_empty(&iocg->active_list)) 1062 goto fail_unlock; 1063 if (iocg->child_active_sum) 1064 goto fail_unlock; 1065 1066 /* 1067 * vtime may wrap when vrate is raised substantially due to 1068 * underestimated IO costs. Look at the period and ignore its 1069 * vtime if the iocg has been idle for too long. Also, cap the 1070 * budget it can start with to the margin. 1071 */ 1072 max_period_delta = DIV64_U64_ROUND_UP(VTIME_VALID_DUR, ioc->period_us); 1073 vtime = atomic64_read(&iocg->vtime); 1074 vmargin = ioc->margin_us * now->vrate; 1075 vmin = now->vnow - vmargin; 1076 1077 if (last_period + max_period_delta < cur_period || 1078 time_before64(vtime, vmin)) { 1079 atomic64_add(vmin - vtime, &iocg->vtime); 1080 atomic64_add(vmin - vtime, &iocg->done_vtime); 1081 vtime = vmin; 1082 } 1083 1084 /* 1085 * Activate, propagate weight and start period timer if not 1086 * running. Reset hweight_gen to avoid accidental match from 1087 * wrapping. 1088 */ 1089 iocg->hweight_gen = atomic_read(&ioc->hweight_gen) - 1; 1090 list_add(&iocg->active_list, &ioc->active_iocgs); 1091 propagate_active_weight(iocg, iocg->weight, 1092 iocg->last_inuse ?: iocg->weight); 1093 1094 TRACE_IOCG_PATH(iocg_activate, iocg, now, 1095 last_period, cur_period, vtime); 1096 1097 iocg->last_vtime = vtime; 1098 1099 if (ioc->running == IOC_IDLE) { 1100 ioc->running = IOC_RUNNING; 1101 ioc_start_period(ioc, now); 1102 } 1103 1104 spin_unlock_irq(&ioc->lock); 1105 return true; 1106 1107 fail_unlock: 1108 spin_unlock_irq(&ioc->lock); 1109 return false; 1110 } 1111 1112 static int iocg_wake_fn(struct wait_queue_entry *wq_entry, unsigned mode, 1113 int flags, void *key) 1114 { 1115 struct iocg_wait *wait = container_of(wq_entry, struct iocg_wait, wait); 1116 struct iocg_wake_ctx *ctx = (struct iocg_wake_ctx *)key; 1117 u64 cost = abs_cost_to_cost(wait->abs_cost, ctx->hw_inuse); 1118 1119 ctx->vbudget -= cost; 1120 1121 if (ctx->vbudget < 0) 1122 return -1; 1123 1124 iocg_commit_bio(ctx->iocg, wait->bio, cost); 1125 1126 /* 1127 * autoremove_wake_function() removes the wait entry only when it 1128 * actually changed the task state. We want the wait always 1129 * removed. Remove explicitly and use default_wake_function(). 1130 */ 1131 list_del_init(&wq_entry->entry); 1132 wait->committed = true; 1133 1134 default_wake_function(wq_entry, mode, flags, key); 1135 return 0; 1136 } 1137 1138 static void iocg_kick_waitq(struct ioc_gq *iocg, struct ioc_now *now) 1139 { 1140 struct ioc *ioc = iocg->ioc; 1141 struct iocg_wake_ctx ctx = { .iocg = iocg }; 1142 u64 margin_ns = (u64)(ioc->period_us * 1143 WAITQ_TIMER_MARGIN_PCT / 100) * NSEC_PER_USEC; 1144 u64 abs_vdebt, vdebt, vshortage, expires, oexpires; 1145 s64 vbudget; 1146 u32 hw_inuse; 1147 1148 lockdep_assert_held(&iocg->waitq.lock); 1149 1150 current_hweight(iocg, NULL, &hw_inuse); 1151 vbudget = now->vnow - atomic64_read(&iocg->vtime); 1152 1153 /* pay off debt */ 1154 abs_vdebt = atomic64_read(&iocg->abs_vdebt); 1155 vdebt = abs_cost_to_cost(abs_vdebt, hw_inuse); 1156 if (vdebt && vbudget > 0) { 1157 u64 delta = min_t(u64, vbudget, vdebt); 1158 u64 abs_delta = min(cost_to_abs_cost(delta, hw_inuse), 1159 abs_vdebt); 1160 1161 atomic64_add(delta, &iocg->vtime); 1162 atomic64_add(delta, &iocg->done_vtime); 1163 atomic64_sub(abs_delta, &iocg->abs_vdebt); 1164 if (WARN_ON_ONCE(atomic64_read(&iocg->abs_vdebt) < 0)) 1165 atomic64_set(&iocg->abs_vdebt, 0); 1166 } 1167 1168 /* 1169 * Wake up the ones which are due and see how much vtime we'll need 1170 * for the next one. 1171 */ 1172 ctx.hw_inuse = hw_inuse; 1173 ctx.vbudget = vbudget - vdebt; 1174 __wake_up_locked_key(&iocg->waitq, TASK_NORMAL, &ctx); 1175 if (!waitqueue_active(&iocg->waitq)) 1176 return; 1177 if (WARN_ON_ONCE(ctx.vbudget >= 0)) 1178 return; 1179 1180 /* determine next wakeup, add a quarter margin to guarantee chunking */ 1181 vshortage = -ctx.vbudget; 1182 expires = now->now_ns + 1183 DIV64_U64_ROUND_UP(vshortage, now->vrate) * NSEC_PER_USEC; 1184 expires += margin_ns / 4; 1185 1186 /* if already active and close enough, don't bother */ 1187 oexpires = ktime_to_ns(hrtimer_get_softexpires(&iocg->waitq_timer)); 1188 if (hrtimer_is_queued(&iocg->waitq_timer) && 1189 abs(oexpires - expires) <= margin_ns / 4) 1190 return; 1191 1192 hrtimer_start_range_ns(&iocg->waitq_timer, ns_to_ktime(expires), 1193 margin_ns / 4, HRTIMER_MODE_ABS); 1194 } 1195 1196 static enum hrtimer_restart iocg_waitq_timer_fn(struct hrtimer *timer) 1197 { 1198 struct ioc_gq *iocg = container_of(timer, struct ioc_gq, waitq_timer); 1199 struct ioc_now now; 1200 unsigned long flags; 1201 1202 ioc_now(iocg->ioc, &now); 1203 1204 spin_lock_irqsave(&iocg->waitq.lock, flags); 1205 iocg_kick_waitq(iocg, &now); 1206 spin_unlock_irqrestore(&iocg->waitq.lock, flags); 1207 1208 return HRTIMER_NORESTART; 1209 } 1210 1211 static void iocg_kick_delay(struct ioc_gq *iocg, struct ioc_now *now, u64 cost) 1212 { 1213 struct ioc *ioc = iocg->ioc; 1214 struct blkcg_gq *blkg = iocg_to_blkg(iocg); 1215 u64 vtime = atomic64_read(&iocg->vtime); 1216 u64 vmargin = ioc->margin_us * now->vrate; 1217 u64 margin_ns = ioc->margin_us * NSEC_PER_USEC; 1218 u64 expires, oexpires; 1219 u32 hw_inuse; 1220 1221 /* debt-adjust vtime */ 1222 current_hweight(iocg, NULL, &hw_inuse); 1223 vtime += abs_cost_to_cost(atomic64_read(&iocg->abs_vdebt), hw_inuse); 1224 1225 /* clear or maintain depending on the overage */ 1226 if (time_before_eq64(vtime, now->vnow)) { 1227 blkcg_clear_delay(blkg); 1228 return; 1229 } 1230 if (!atomic_read(&blkg->use_delay) && 1231 time_before_eq64(vtime, now->vnow + vmargin)) 1232 return; 1233 1234 /* use delay */ 1235 if (cost) { 1236 u64 cost_ns = DIV64_U64_ROUND_UP(cost * NSEC_PER_USEC, 1237 now->vrate); 1238 blkcg_add_delay(blkg, now->now_ns, cost_ns); 1239 } 1240 blkcg_use_delay(blkg); 1241 1242 expires = now->now_ns + DIV64_U64_ROUND_UP(vtime - now->vnow, 1243 now->vrate) * NSEC_PER_USEC; 1244 1245 /* if already active and close enough, don't bother */ 1246 oexpires = ktime_to_ns(hrtimer_get_softexpires(&iocg->delay_timer)); 1247 if (hrtimer_is_queued(&iocg->delay_timer) && 1248 abs(oexpires - expires) <= margin_ns / 4) 1249 return; 1250 1251 hrtimer_start_range_ns(&iocg->delay_timer, ns_to_ktime(expires), 1252 margin_ns / 4, HRTIMER_MODE_ABS); 1253 } 1254 1255 static enum hrtimer_restart iocg_delay_timer_fn(struct hrtimer *timer) 1256 { 1257 struct ioc_gq *iocg = container_of(timer, struct ioc_gq, delay_timer); 1258 struct ioc_now now; 1259 1260 ioc_now(iocg->ioc, &now); 1261 iocg_kick_delay(iocg, &now, 0); 1262 1263 return HRTIMER_NORESTART; 1264 } 1265 1266 static void ioc_lat_stat(struct ioc *ioc, u32 *missed_ppm_ar, u32 *rq_wait_pct_p) 1267 { 1268 u32 nr_met[2] = { }; 1269 u32 nr_missed[2] = { }; 1270 u64 rq_wait_ns = 0; 1271 int cpu, rw; 1272 1273 for_each_online_cpu(cpu) { 1274 struct ioc_pcpu_stat *stat = per_cpu_ptr(ioc->pcpu_stat, cpu); 1275 u64 this_rq_wait_ns; 1276 1277 for (rw = READ; rw <= WRITE; rw++) { 1278 u32 this_met = READ_ONCE(stat->missed[rw].nr_met); 1279 u32 this_missed = READ_ONCE(stat->missed[rw].nr_missed); 1280 1281 nr_met[rw] += this_met - stat->missed[rw].last_met; 1282 nr_missed[rw] += this_missed - stat->missed[rw].last_missed; 1283 stat->missed[rw].last_met = this_met; 1284 stat->missed[rw].last_missed = this_missed; 1285 } 1286 1287 this_rq_wait_ns = READ_ONCE(stat->rq_wait_ns); 1288 rq_wait_ns += this_rq_wait_ns - stat->last_rq_wait_ns; 1289 stat->last_rq_wait_ns = this_rq_wait_ns; 1290 } 1291 1292 for (rw = READ; rw <= WRITE; rw++) { 1293 if (nr_met[rw] + nr_missed[rw]) 1294 missed_ppm_ar[rw] = 1295 DIV64_U64_ROUND_UP((u64)nr_missed[rw] * MILLION, 1296 nr_met[rw] + nr_missed[rw]); 1297 else 1298 missed_ppm_ar[rw] = 0; 1299 } 1300 1301 *rq_wait_pct_p = div64_u64(rq_wait_ns * 100, 1302 ioc->period_us * NSEC_PER_USEC); 1303 } 1304 1305 /* was iocg idle this period? */ 1306 static bool iocg_is_idle(struct ioc_gq *iocg) 1307 { 1308 struct ioc *ioc = iocg->ioc; 1309 1310 /* did something get issued this period? */ 1311 if (atomic64_read(&iocg->active_period) == 1312 atomic64_read(&ioc->cur_period)) 1313 return false; 1314 1315 /* is something in flight? */ 1316 if (atomic64_read(&iocg->done_vtime) < atomic64_read(&iocg->vtime)) 1317 return false; 1318 1319 return true; 1320 } 1321 1322 /* returns usage with margin added if surplus is large enough */ 1323 static u32 surplus_adjusted_hweight_inuse(u32 usage, u32 hw_inuse) 1324 { 1325 /* add margin */ 1326 usage = DIV_ROUND_UP(usage * SURPLUS_SCALE_PCT, 100); 1327 usage += SURPLUS_SCALE_ABS; 1328 1329 /* don't bother if the surplus is too small */ 1330 if (usage + SURPLUS_MIN_ADJ_DELTA > hw_inuse) 1331 return 0; 1332 1333 return usage; 1334 } 1335 1336 static void ioc_timer_fn(struct timer_list *timer) 1337 { 1338 struct ioc *ioc = container_of(timer, struct ioc, timer); 1339 struct ioc_gq *iocg, *tiocg; 1340 struct ioc_now now; 1341 int nr_surpluses = 0, nr_shortages = 0, nr_lagging = 0; 1342 u32 ppm_rthr = MILLION - ioc->params.qos[QOS_RPPM]; 1343 u32 ppm_wthr = MILLION - ioc->params.qos[QOS_WPPM]; 1344 u32 missed_ppm[2], rq_wait_pct; 1345 u64 period_vtime; 1346 int prev_busy_level, i; 1347 1348 /* how were the latencies during the period? */ 1349 ioc_lat_stat(ioc, missed_ppm, &rq_wait_pct); 1350 1351 /* take care of active iocgs */ 1352 spin_lock_irq(&ioc->lock); 1353 1354 ioc_now(ioc, &now); 1355 1356 period_vtime = now.vnow - ioc->period_at_vtime; 1357 if (WARN_ON_ONCE(!period_vtime)) { 1358 spin_unlock_irq(&ioc->lock); 1359 return; 1360 } 1361 1362 /* 1363 * Waiters determine the sleep durations based on the vrate they 1364 * saw at the time of sleep. If vrate has increased, some waiters 1365 * could be sleeping for too long. Wake up tardy waiters which 1366 * should have woken up in the last period and expire idle iocgs. 1367 */ 1368 list_for_each_entry_safe(iocg, tiocg, &ioc->active_iocgs, active_list) { 1369 if (!waitqueue_active(&iocg->waitq) && 1370 !atomic64_read(&iocg->abs_vdebt) && !iocg_is_idle(iocg)) 1371 continue; 1372 1373 spin_lock(&iocg->waitq.lock); 1374 1375 if (waitqueue_active(&iocg->waitq) || 1376 atomic64_read(&iocg->abs_vdebt)) { 1377 /* might be oversleeping vtime / hweight changes, kick */ 1378 iocg_kick_waitq(iocg, &now); 1379 iocg_kick_delay(iocg, &now, 0); 1380 } else if (iocg_is_idle(iocg)) { 1381 /* no waiter and idle, deactivate */ 1382 iocg->last_inuse = iocg->inuse; 1383 __propagate_active_weight(iocg, 0, 0); 1384 list_del_init(&iocg->active_list); 1385 } 1386 1387 spin_unlock(&iocg->waitq.lock); 1388 } 1389 commit_active_weights(ioc); 1390 1391 /* calc usages and see whether some weights need to be moved around */ 1392 list_for_each_entry(iocg, &ioc->active_iocgs, active_list) { 1393 u64 vdone, vtime, vusage, vmargin, vmin; 1394 u32 hw_active, hw_inuse, usage; 1395 1396 /* 1397 * Collect unused and wind vtime closer to vnow to prevent 1398 * iocgs from accumulating a large amount of budget. 1399 */ 1400 vdone = atomic64_read(&iocg->done_vtime); 1401 vtime = atomic64_read(&iocg->vtime); 1402 current_hweight(iocg, &hw_active, &hw_inuse); 1403 1404 /* 1405 * Latency QoS detection doesn't account for IOs which are 1406 * in-flight for longer than a period. Detect them by 1407 * comparing vdone against period start. If lagging behind 1408 * IOs from past periods, don't increase vrate. 1409 */ 1410 if ((ppm_rthr != MILLION || ppm_wthr != MILLION) && 1411 !atomic_read(&iocg_to_blkg(iocg)->use_delay) && 1412 time_after64(vtime, vdone) && 1413 time_after64(vtime, now.vnow - 1414 MAX_LAGGING_PERIODS * period_vtime) && 1415 time_before64(vdone, now.vnow - period_vtime)) 1416 nr_lagging++; 1417 1418 if (waitqueue_active(&iocg->waitq)) 1419 vusage = now.vnow - iocg->last_vtime; 1420 else if (time_before64(iocg->last_vtime, vtime)) 1421 vusage = vtime - iocg->last_vtime; 1422 else 1423 vusage = 0; 1424 1425 iocg->last_vtime += vusage; 1426 /* 1427 * Factor in in-flight vtime into vusage to avoid 1428 * high-latency completions appearing as idle. This should 1429 * be done after the above ->last_time adjustment. 1430 */ 1431 vusage = max(vusage, vtime - vdone); 1432 1433 /* calculate hweight based usage ratio and record */ 1434 if (vusage) { 1435 usage = DIV64_U64_ROUND_UP(vusage * hw_inuse, 1436 period_vtime); 1437 iocg->usage_idx = (iocg->usage_idx + 1) % NR_USAGE_SLOTS; 1438 iocg->usages[iocg->usage_idx] = usage; 1439 } else { 1440 usage = 0; 1441 } 1442 1443 /* see whether there's surplus vtime */ 1444 vmargin = ioc->margin_us * now.vrate; 1445 vmin = now.vnow - vmargin; 1446 1447 iocg->has_surplus = false; 1448 1449 if (!waitqueue_active(&iocg->waitq) && 1450 time_before64(vtime, vmin)) { 1451 u64 delta = vmin - vtime; 1452 1453 /* throw away surplus vtime */ 1454 atomic64_add(delta, &iocg->vtime); 1455 atomic64_add(delta, &iocg->done_vtime); 1456 iocg->last_vtime += delta; 1457 /* if usage is sufficiently low, maybe it can donate */ 1458 if (surplus_adjusted_hweight_inuse(usage, hw_inuse)) { 1459 iocg->has_surplus = true; 1460 nr_surpluses++; 1461 } 1462 } else if (hw_inuse < hw_active) { 1463 u32 new_hwi, new_inuse; 1464 1465 /* was donating but might need to take back some */ 1466 if (waitqueue_active(&iocg->waitq)) { 1467 new_hwi = hw_active; 1468 } else { 1469 new_hwi = max(hw_inuse, 1470 usage * SURPLUS_SCALE_PCT / 100 + 1471 SURPLUS_SCALE_ABS); 1472 } 1473 1474 new_inuse = div64_u64((u64)iocg->inuse * new_hwi, 1475 hw_inuse); 1476 new_inuse = clamp_t(u32, new_inuse, 1, iocg->active); 1477 1478 if (new_inuse > iocg->inuse) { 1479 TRACE_IOCG_PATH(inuse_takeback, iocg, &now, 1480 iocg->inuse, new_inuse, 1481 hw_inuse, new_hwi); 1482 __propagate_active_weight(iocg, iocg->weight, 1483 new_inuse); 1484 } 1485 } else { 1486 /* genuninely out of vtime */ 1487 nr_shortages++; 1488 } 1489 } 1490 1491 if (!nr_shortages || !nr_surpluses) 1492 goto skip_surplus_transfers; 1493 1494 /* there are both shortages and surpluses, transfer surpluses */ 1495 list_for_each_entry(iocg, &ioc->active_iocgs, active_list) { 1496 u32 usage, hw_active, hw_inuse, new_hwi, new_inuse; 1497 int nr_valid = 0; 1498 1499 if (!iocg->has_surplus) 1500 continue; 1501 1502 /* base the decision on max historical usage */ 1503 for (i = 0, usage = 0; i < NR_USAGE_SLOTS; i++) { 1504 if (iocg->usages[i]) { 1505 usage = max(usage, iocg->usages[i]); 1506 nr_valid++; 1507 } 1508 } 1509 if (nr_valid < MIN_VALID_USAGES) 1510 continue; 1511 1512 current_hweight(iocg, &hw_active, &hw_inuse); 1513 new_hwi = surplus_adjusted_hweight_inuse(usage, hw_inuse); 1514 if (!new_hwi) 1515 continue; 1516 1517 new_inuse = DIV64_U64_ROUND_UP((u64)iocg->inuse * new_hwi, 1518 hw_inuse); 1519 if (new_inuse < iocg->inuse) { 1520 TRACE_IOCG_PATH(inuse_giveaway, iocg, &now, 1521 iocg->inuse, new_inuse, 1522 hw_inuse, new_hwi); 1523 __propagate_active_weight(iocg, iocg->weight, new_inuse); 1524 } 1525 } 1526 skip_surplus_transfers: 1527 commit_active_weights(ioc); 1528 1529 /* 1530 * If q is getting clogged or we're missing too much, we're issuing 1531 * too much IO and should lower vtime rate. If we're not missing 1532 * and experiencing shortages but not surpluses, we're too stingy 1533 * and should increase vtime rate. 1534 */ 1535 prev_busy_level = ioc->busy_level; 1536 if (rq_wait_pct > RQ_WAIT_BUSY_PCT || 1537 missed_ppm[READ] > ppm_rthr || 1538 missed_ppm[WRITE] > ppm_wthr) { 1539 ioc->busy_level = max(ioc->busy_level, 0); 1540 ioc->busy_level++; 1541 } else if (rq_wait_pct <= RQ_WAIT_BUSY_PCT * UNBUSY_THR_PCT / 100 && 1542 missed_ppm[READ] <= ppm_rthr * UNBUSY_THR_PCT / 100 && 1543 missed_ppm[WRITE] <= ppm_wthr * UNBUSY_THR_PCT / 100) { 1544 /* take action iff there is contention */ 1545 if (nr_shortages && !nr_lagging) { 1546 ioc->busy_level = min(ioc->busy_level, 0); 1547 /* redistribute surpluses first */ 1548 if (!nr_surpluses) 1549 ioc->busy_level--; 1550 } 1551 } else { 1552 ioc->busy_level = 0; 1553 } 1554 1555 ioc->busy_level = clamp(ioc->busy_level, -1000, 1000); 1556 1557 if (ioc->busy_level > 0 || (ioc->busy_level < 0 && !nr_lagging)) { 1558 u64 vrate = atomic64_read(&ioc->vtime_rate); 1559 u64 vrate_min = ioc->vrate_min, vrate_max = ioc->vrate_max; 1560 1561 /* rq_wait signal is always reliable, ignore user vrate_min */ 1562 if (rq_wait_pct > RQ_WAIT_BUSY_PCT) 1563 vrate_min = VRATE_MIN; 1564 1565 /* 1566 * If vrate is out of bounds, apply clamp gradually as the 1567 * bounds can change abruptly. Otherwise, apply busy_level 1568 * based adjustment. 1569 */ 1570 if (vrate < vrate_min) { 1571 vrate = div64_u64(vrate * (100 + VRATE_CLAMP_ADJ_PCT), 1572 100); 1573 vrate = min(vrate, vrate_min); 1574 } else if (vrate > vrate_max) { 1575 vrate = div64_u64(vrate * (100 - VRATE_CLAMP_ADJ_PCT), 1576 100); 1577 vrate = max(vrate, vrate_max); 1578 } else { 1579 int idx = min_t(int, abs(ioc->busy_level), 1580 ARRAY_SIZE(vrate_adj_pct) - 1); 1581 u32 adj_pct = vrate_adj_pct[idx]; 1582 1583 if (ioc->busy_level > 0) 1584 adj_pct = 100 - adj_pct; 1585 else 1586 adj_pct = 100 + adj_pct; 1587 1588 vrate = clamp(DIV64_U64_ROUND_UP(vrate * adj_pct, 100), 1589 vrate_min, vrate_max); 1590 } 1591 1592 trace_iocost_ioc_vrate_adj(ioc, vrate, &missed_ppm, rq_wait_pct, 1593 nr_lagging, nr_shortages, 1594 nr_surpluses); 1595 1596 atomic64_set(&ioc->vtime_rate, vrate); 1597 ioc->inuse_margin_vtime = DIV64_U64_ROUND_UP( 1598 ioc->period_us * vrate * INUSE_MARGIN_PCT, 100); 1599 } else if (ioc->busy_level != prev_busy_level || nr_lagging) { 1600 trace_iocost_ioc_vrate_adj(ioc, atomic64_read(&ioc->vtime_rate), 1601 &missed_ppm, rq_wait_pct, nr_lagging, 1602 nr_shortages, nr_surpluses); 1603 } 1604 1605 ioc_refresh_params(ioc, false); 1606 1607 /* 1608 * This period is done. Move onto the next one. If nothing's 1609 * going on with the device, stop the timer. 1610 */ 1611 atomic64_inc(&ioc->cur_period); 1612 1613 if (ioc->running != IOC_STOP) { 1614 if (!list_empty(&ioc->active_iocgs)) { 1615 ioc_start_period(ioc, &now); 1616 } else { 1617 ioc->busy_level = 0; 1618 ioc->running = IOC_IDLE; 1619 } 1620 } 1621 1622 spin_unlock_irq(&ioc->lock); 1623 } 1624 1625 static void calc_vtime_cost_builtin(struct bio *bio, struct ioc_gq *iocg, 1626 bool is_merge, u64 *costp) 1627 { 1628 struct ioc *ioc = iocg->ioc; 1629 u64 coef_seqio, coef_randio, coef_page; 1630 u64 pages = max_t(u64, bio_sectors(bio) >> IOC_SECT_TO_PAGE_SHIFT, 1); 1631 u64 seek_pages = 0; 1632 u64 cost = 0; 1633 1634 switch (bio_op(bio)) { 1635 case REQ_OP_READ: 1636 coef_seqio = ioc->params.lcoefs[LCOEF_RSEQIO]; 1637 coef_randio = ioc->params.lcoefs[LCOEF_RRANDIO]; 1638 coef_page = ioc->params.lcoefs[LCOEF_RPAGE]; 1639 break; 1640 case REQ_OP_WRITE: 1641 coef_seqio = ioc->params.lcoefs[LCOEF_WSEQIO]; 1642 coef_randio = ioc->params.lcoefs[LCOEF_WRANDIO]; 1643 coef_page = ioc->params.lcoefs[LCOEF_WPAGE]; 1644 break; 1645 default: 1646 goto out; 1647 } 1648 1649 if (iocg->cursor) { 1650 seek_pages = abs(bio->bi_iter.bi_sector - iocg->cursor); 1651 seek_pages >>= IOC_SECT_TO_PAGE_SHIFT; 1652 } 1653 1654 if (!is_merge) { 1655 if (seek_pages > LCOEF_RANDIO_PAGES) { 1656 cost += coef_randio; 1657 } else { 1658 cost += coef_seqio; 1659 } 1660 } 1661 cost += pages * coef_page; 1662 out: 1663 *costp = cost; 1664 } 1665 1666 static u64 calc_vtime_cost(struct bio *bio, struct ioc_gq *iocg, bool is_merge) 1667 { 1668 u64 cost; 1669 1670 calc_vtime_cost_builtin(bio, iocg, is_merge, &cost); 1671 return cost; 1672 } 1673 1674 static void ioc_rqos_throttle(struct rq_qos *rqos, struct bio *bio) 1675 { 1676 struct blkcg_gq *blkg = bio->bi_blkg; 1677 struct ioc *ioc = rqos_to_ioc(rqos); 1678 struct ioc_gq *iocg = blkg_to_iocg(blkg); 1679 struct ioc_now now; 1680 struct iocg_wait wait; 1681 u32 hw_active, hw_inuse; 1682 u64 abs_cost, cost, vtime; 1683 1684 /* bypass IOs if disabled or for root cgroup */ 1685 if (!ioc->enabled || !iocg->level) 1686 return; 1687 1688 /* always activate so that even 0 cost IOs get protected to some level */ 1689 if (!iocg_activate(iocg, &now)) 1690 return; 1691 1692 /* calculate the absolute vtime cost */ 1693 abs_cost = calc_vtime_cost(bio, iocg, false); 1694 if (!abs_cost) 1695 return; 1696 1697 iocg->cursor = bio_end_sector(bio); 1698 1699 vtime = atomic64_read(&iocg->vtime); 1700 current_hweight(iocg, &hw_active, &hw_inuse); 1701 1702 if (hw_inuse < hw_active && 1703 time_after_eq64(vtime + ioc->inuse_margin_vtime, now.vnow)) { 1704 TRACE_IOCG_PATH(inuse_reset, iocg, &now, 1705 iocg->inuse, iocg->weight, hw_inuse, hw_active); 1706 spin_lock_irq(&ioc->lock); 1707 propagate_active_weight(iocg, iocg->weight, iocg->weight); 1708 spin_unlock_irq(&ioc->lock); 1709 current_hweight(iocg, &hw_active, &hw_inuse); 1710 } 1711 1712 cost = abs_cost_to_cost(abs_cost, hw_inuse); 1713 1714 /* 1715 * If no one's waiting and within budget, issue right away. The 1716 * tests are racy but the races aren't systemic - we only miss once 1717 * in a while which is fine. 1718 */ 1719 if (!waitqueue_active(&iocg->waitq) && 1720 !atomic64_read(&iocg->abs_vdebt) && 1721 time_before_eq64(vtime + cost, now.vnow)) { 1722 iocg_commit_bio(iocg, bio, cost); 1723 return; 1724 } 1725 1726 /* 1727 * We're over budget. If @bio has to be issued regardless, 1728 * remember the abs_cost instead of advancing vtime. 1729 * iocg_kick_waitq() will pay off the debt before waking more IOs. 1730 * This way, the debt is continuously paid off each period with the 1731 * actual budget available to the cgroup. If we just wound vtime, 1732 * we would incorrectly use the current hw_inuse for the entire 1733 * amount which, for example, can lead to the cgroup staying 1734 * blocked for a long time even with substantially raised hw_inuse. 1735 */ 1736 if (bio_issue_as_root_blkg(bio) || fatal_signal_pending(current)) { 1737 atomic64_add(abs_cost, &iocg->abs_vdebt); 1738 iocg_kick_delay(iocg, &now, cost); 1739 return; 1740 } 1741 1742 /* 1743 * Append self to the waitq and schedule the wakeup timer if we're 1744 * the first waiter. The timer duration is calculated based on the 1745 * current vrate. vtime and hweight changes can make it too short 1746 * or too long. Each wait entry records the absolute cost it's 1747 * waiting for to allow re-evaluation using a custom wait entry. 1748 * 1749 * If too short, the timer simply reschedules itself. If too long, 1750 * the period timer will notice and trigger wakeups. 1751 * 1752 * All waiters are on iocg->waitq and the wait states are 1753 * synchronized using waitq.lock. 1754 */ 1755 spin_lock_irq(&iocg->waitq.lock); 1756 1757 /* 1758 * We activated above but w/o any synchronization. Deactivation is 1759 * synchronized with waitq.lock and we won't get deactivated as 1760 * long as we're waiting, so we're good if we're activated here. 1761 * In the unlikely case that we are deactivated, just issue the IO. 1762 */ 1763 if (unlikely(list_empty(&iocg->active_list))) { 1764 spin_unlock_irq(&iocg->waitq.lock); 1765 iocg_commit_bio(iocg, bio, cost); 1766 return; 1767 } 1768 1769 init_waitqueue_func_entry(&wait.wait, iocg_wake_fn); 1770 wait.wait.private = current; 1771 wait.bio = bio; 1772 wait.abs_cost = abs_cost; 1773 wait.committed = false; /* will be set true by waker */ 1774 1775 __add_wait_queue_entry_tail(&iocg->waitq, &wait.wait); 1776 iocg_kick_waitq(iocg, &now); 1777 1778 spin_unlock_irq(&iocg->waitq.lock); 1779 1780 while (true) { 1781 set_current_state(TASK_UNINTERRUPTIBLE); 1782 if (wait.committed) 1783 break; 1784 io_schedule(); 1785 } 1786 1787 /* waker already committed us, proceed */ 1788 finish_wait(&iocg->waitq, &wait.wait); 1789 } 1790 1791 static void ioc_rqos_merge(struct rq_qos *rqos, struct request *rq, 1792 struct bio *bio) 1793 { 1794 struct ioc_gq *iocg = blkg_to_iocg(bio->bi_blkg); 1795 struct ioc *ioc = iocg->ioc; 1796 sector_t bio_end = bio_end_sector(bio); 1797 struct ioc_now now; 1798 u32 hw_inuse; 1799 u64 abs_cost, cost; 1800 1801 /* bypass if disabled or for root cgroup */ 1802 if (!ioc->enabled || !iocg->level) 1803 return; 1804 1805 abs_cost = calc_vtime_cost(bio, iocg, true); 1806 if (!abs_cost) 1807 return; 1808 1809 ioc_now(ioc, &now); 1810 current_hweight(iocg, NULL, &hw_inuse); 1811 cost = abs_cost_to_cost(abs_cost, hw_inuse); 1812 1813 /* update cursor if backmerging into the request at the cursor */ 1814 if (blk_rq_pos(rq) < bio_end && 1815 blk_rq_pos(rq) + blk_rq_sectors(rq) == iocg->cursor) 1816 iocg->cursor = bio_end; 1817 1818 /* 1819 * Charge if there's enough vtime budget and the existing request 1820 * has cost assigned. Otherwise, account it as debt. See debt 1821 * handling in ioc_rqos_throttle() for details. 1822 */ 1823 if (rq->bio && rq->bio->bi_iocost_cost && 1824 time_before_eq64(atomic64_read(&iocg->vtime) + cost, now.vnow)) 1825 iocg_commit_bio(iocg, bio, cost); 1826 else 1827 atomic64_add(abs_cost, &iocg->abs_vdebt); 1828 } 1829 1830 static void ioc_rqos_done_bio(struct rq_qos *rqos, struct bio *bio) 1831 { 1832 struct ioc_gq *iocg = blkg_to_iocg(bio->bi_blkg); 1833 1834 if (iocg && bio->bi_iocost_cost) 1835 atomic64_add(bio->bi_iocost_cost, &iocg->done_vtime); 1836 } 1837 1838 static void ioc_rqos_done(struct rq_qos *rqos, struct request *rq) 1839 { 1840 struct ioc *ioc = rqos_to_ioc(rqos); 1841 u64 on_q_ns, rq_wait_ns; 1842 int pidx, rw; 1843 1844 if (!ioc->enabled || !rq->alloc_time_ns || !rq->start_time_ns) 1845 return; 1846 1847 switch (req_op(rq) & REQ_OP_MASK) { 1848 case REQ_OP_READ: 1849 pidx = QOS_RLAT; 1850 rw = READ; 1851 break; 1852 case REQ_OP_WRITE: 1853 pidx = QOS_WLAT; 1854 rw = WRITE; 1855 break; 1856 default: 1857 return; 1858 } 1859 1860 on_q_ns = ktime_get_ns() - rq->alloc_time_ns; 1861 rq_wait_ns = rq->start_time_ns - rq->alloc_time_ns; 1862 1863 if (on_q_ns <= ioc->params.qos[pidx] * NSEC_PER_USEC) 1864 this_cpu_inc(ioc->pcpu_stat->missed[rw].nr_met); 1865 else 1866 this_cpu_inc(ioc->pcpu_stat->missed[rw].nr_missed); 1867 1868 this_cpu_add(ioc->pcpu_stat->rq_wait_ns, rq_wait_ns); 1869 } 1870 1871 static void ioc_rqos_queue_depth_changed(struct rq_qos *rqos) 1872 { 1873 struct ioc *ioc = rqos_to_ioc(rqos); 1874 1875 spin_lock_irq(&ioc->lock); 1876 ioc_refresh_params(ioc, false); 1877 spin_unlock_irq(&ioc->lock); 1878 } 1879 1880 static void ioc_rqos_exit(struct rq_qos *rqos) 1881 { 1882 struct ioc *ioc = rqos_to_ioc(rqos); 1883 1884 blkcg_deactivate_policy(rqos->q, &blkcg_policy_iocost); 1885 1886 spin_lock_irq(&ioc->lock); 1887 ioc->running = IOC_STOP; 1888 spin_unlock_irq(&ioc->lock); 1889 1890 del_timer_sync(&ioc->timer); 1891 free_percpu(ioc->pcpu_stat); 1892 kfree(ioc); 1893 } 1894 1895 static struct rq_qos_ops ioc_rqos_ops = { 1896 .throttle = ioc_rqos_throttle, 1897 .merge = ioc_rqos_merge, 1898 .done_bio = ioc_rqos_done_bio, 1899 .done = ioc_rqos_done, 1900 .queue_depth_changed = ioc_rqos_queue_depth_changed, 1901 .exit = ioc_rqos_exit, 1902 }; 1903 1904 static int blk_iocost_init(struct request_queue *q) 1905 { 1906 struct ioc *ioc; 1907 struct rq_qos *rqos; 1908 int ret; 1909 1910 ioc = kzalloc(sizeof(*ioc), GFP_KERNEL); 1911 if (!ioc) 1912 return -ENOMEM; 1913 1914 ioc->pcpu_stat = alloc_percpu(struct ioc_pcpu_stat); 1915 if (!ioc->pcpu_stat) { 1916 kfree(ioc); 1917 return -ENOMEM; 1918 } 1919 1920 rqos = &ioc->rqos; 1921 rqos->id = RQ_QOS_COST; 1922 rqos->ops = &ioc_rqos_ops; 1923 rqos->q = q; 1924 1925 spin_lock_init(&ioc->lock); 1926 timer_setup(&ioc->timer, ioc_timer_fn, 0); 1927 INIT_LIST_HEAD(&ioc->active_iocgs); 1928 1929 ioc->running = IOC_IDLE; 1930 atomic64_set(&ioc->vtime_rate, VTIME_PER_USEC); 1931 seqcount_init(&ioc->period_seqcount); 1932 ioc->period_at = ktime_to_us(ktime_get()); 1933 atomic64_set(&ioc->cur_period, 0); 1934 atomic_set(&ioc->hweight_gen, 0); 1935 1936 spin_lock_irq(&ioc->lock); 1937 ioc->autop_idx = AUTOP_INVALID; 1938 ioc_refresh_params(ioc, true); 1939 spin_unlock_irq(&ioc->lock); 1940 1941 rq_qos_add(q, rqos); 1942 ret = blkcg_activate_policy(q, &blkcg_policy_iocost); 1943 if (ret) { 1944 rq_qos_del(q, rqos); 1945 free_percpu(ioc->pcpu_stat); 1946 kfree(ioc); 1947 return ret; 1948 } 1949 return 0; 1950 } 1951 1952 static struct blkcg_policy_data *ioc_cpd_alloc(gfp_t gfp) 1953 { 1954 struct ioc_cgrp *iocc; 1955 1956 iocc = kzalloc(sizeof(struct ioc_cgrp), gfp); 1957 if (!iocc) 1958 return NULL; 1959 1960 iocc->dfl_weight = CGROUP_WEIGHT_DFL; 1961 return &iocc->cpd; 1962 } 1963 1964 static void ioc_cpd_free(struct blkcg_policy_data *cpd) 1965 { 1966 kfree(container_of(cpd, struct ioc_cgrp, cpd)); 1967 } 1968 1969 static struct blkg_policy_data *ioc_pd_alloc(gfp_t gfp, struct request_queue *q, 1970 struct blkcg *blkcg) 1971 { 1972 int levels = blkcg->css.cgroup->level + 1; 1973 struct ioc_gq *iocg; 1974 1975 iocg = kzalloc_node(sizeof(*iocg) + levels * sizeof(iocg->ancestors[0]), 1976 gfp, q->node); 1977 if (!iocg) 1978 return NULL; 1979 1980 return &iocg->pd; 1981 } 1982 1983 static void ioc_pd_init(struct blkg_policy_data *pd) 1984 { 1985 struct ioc_gq *iocg = pd_to_iocg(pd); 1986 struct blkcg_gq *blkg = pd_to_blkg(&iocg->pd); 1987 struct ioc *ioc = q_to_ioc(blkg->q); 1988 struct ioc_now now; 1989 struct blkcg_gq *tblkg; 1990 unsigned long flags; 1991 1992 ioc_now(ioc, &now); 1993 1994 iocg->ioc = ioc; 1995 atomic64_set(&iocg->vtime, now.vnow); 1996 atomic64_set(&iocg->done_vtime, now.vnow); 1997 atomic64_set(&iocg->abs_vdebt, 0); 1998 atomic64_set(&iocg->active_period, atomic64_read(&ioc->cur_period)); 1999 INIT_LIST_HEAD(&iocg->active_list); 2000 iocg->hweight_active = HWEIGHT_WHOLE; 2001 iocg->hweight_inuse = HWEIGHT_WHOLE; 2002 2003 init_waitqueue_head(&iocg->waitq); 2004 hrtimer_init(&iocg->waitq_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS); 2005 iocg->waitq_timer.function = iocg_waitq_timer_fn; 2006 hrtimer_init(&iocg->delay_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS); 2007 iocg->delay_timer.function = iocg_delay_timer_fn; 2008 2009 iocg->level = blkg->blkcg->css.cgroup->level; 2010 2011 for (tblkg = blkg; tblkg; tblkg = tblkg->parent) { 2012 struct ioc_gq *tiocg = blkg_to_iocg(tblkg); 2013 iocg->ancestors[tiocg->level] = tiocg; 2014 } 2015 2016 spin_lock_irqsave(&ioc->lock, flags); 2017 weight_updated(iocg); 2018 spin_unlock_irqrestore(&ioc->lock, flags); 2019 } 2020 2021 static void ioc_pd_free(struct blkg_policy_data *pd) 2022 { 2023 struct ioc_gq *iocg = pd_to_iocg(pd); 2024 struct ioc *ioc = iocg->ioc; 2025 2026 if (ioc) { 2027 spin_lock(&ioc->lock); 2028 if (!list_empty(&iocg->active_list)) { 2029 propagate_active_weight(iocg, 0, 0); 2030 list_del_init(&iocg->active_list); 2031 } 2032 spin_unlock(&ioc->lock); 2033 2034 hrtimer_cancel(&iocg->waitq_timer); 2035 hrtimer_cancel(&iocg->delay_timer); 2036 } 2037 kfree(iocg); 2038 } 2039 2040 static u64 ioc_weight_prfill(struct seq_file *sf, struct blkg_policy_data *pd, 2041 int off) 2042 { 2043 const char *dname = blkg_dev_name(pd->blkg); 2044 struct ioc_gq *iocg = pd_to_iocg(pd); 2045 2046 if (dname && iocg->cfg_weight) 2047 seq_printf(sf, "%s %u\n", dname, iocg->cfg_weight); 2048 return 0; 2049 } 2050 2051 2052 static int ioc_weight_show(struct seq_file *sf, void *v) 2053 { 2054 struct blkcg *blkcg = css_to_blkcg(seq_css(sf)); 2055 struct ioc_cgrp *iocc = blkcg_to_iocc(blkcg); 2056 2057 seq_printf(sf, "default %u\n", iocc->dfl_weight); 2058 blkcg_print_blkgs(sf, blkcg, ioc_weight_prfill, 2059 &blkcg_policy_iocost, seq_cft(sf)->private, false); 2060 return 0; 2061 } 2062 2063 static ssize_t ioc_weight_write(struct kernfs_open_file *of, char *buf, 2064 size_t nbytes, loff_t off) 2065 { 2066 struct blkcg *blkcg = css_to_blkcg(of_css(of)); 2067 struct ioc_cgrp *iocc = blkcg_to_iocc(blkcg); 2068 struct blkg_conf_ctx ctx; 2069 struct ioc_gq *iocg; 2070 u32 v; 2071 int ret; 2072 2073 if (!strchr(buf, ':')) { 2074 struct blkcg_gq *blkg; 2075 2076 if (!sscanf(buf, "default %u", &v) && !sscanf(buf, "%u", &v)) 2077 return -EINVAL; 2078 2079 if (v < CGROUP_WEIGHT_MIN || v > CGROUP_WEIGHT_MAX) 2080 return -EINVAL; 2081 2082 spin_lock(&blkcg->lock); 2083 iocc->dfl_weight = v; 2084 hlist_for_each_entry(blkg, &blkcg->blkg_list, blkcg_node) { 2085 struct ioc_gq *iocg = blkg_to_iocg(blkg); 2086 2087 if (iocg) { 2088 spin_lock_irq(&iocg->ioc->lock); 2089 weight_updated(iocg); 2090 spin_unlock_irq(&iocg->ioc->lock); 2091 } 2092 } 2093 spin_unlock(&blkcg->lock); 2094 2095 return nbytes; 2096 } 2097 2098 ret = blkg_conf_prep(blkcg, &blkcg_policy_iocost, buf, &ctx); 2099 if (ret) 2100 return ret; 2101 2102 iocg = blkg_to_iocg(ctx.blkg); 2103 2104 if (!strncmp(ctx.body, "default", 7)) { 2105 v = 0; 2106 } else { 2107 if (!sscanf(ctx.body, "%u", &v)) 2108 goto einval; 2109 if (v < CGROUP_WEIGHT_MIN || v > CGROUP_WEIGHT_MAX) 2110 goto einval; 2111 } 2112 2113 spin_lock(&iocg->ioc->lock); 2114 iocg->cfg_weight = v; 2115 weight_updated(iocg); 2116 spin_unlock(&iocg->ioc->lock); 2117 2118 blkg_conf_finish(&ctx); 2119 return nbytes; 2120 2121 einval: 2122 blkg_conf_finish(&ctx); 2123 return -EINVAL; 2124 } 2125 2126 static u64 ioc_qos_prfill(struct seq_file *sf, struct blkg_policy_data *pd, 2127 int off) 2128 { 2129 const char *dname = blkg_dev_name(pd->blkg); 2130 struct ioc *ioc = pd_to_iocg(pd)->ioc; 2131 2132 if (!dname) 2133 return 0; 2134 2135 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", 2136 dname, ioc->enabled, ioc->user_qos_params ? "user" : "auto", 2137 ioc->params.qos[QOS_RPPM] / 10000, 2138 ioc->params.qos[QOS_RPPM] % 10000 / 100, 2139 ioc->params.qos[QOS_RLAT], 2140 ioc->params.qos[QOS_WPPM] / 10000, 2141 ioc->params.qos[QOS_WPPM] % 10000 / 100, 2142 ioc->params.qos[QOS_WLAT], 2143 ioc->params.qos[QOS_MIN] / 10000, 2144 ioc->params.qos[QOS_MIN] % 10000 / 100, 2145 ioc->params.qos[QOS_MAX] / 10000, 2146 ioc->params.qos[QOS_MAX] % 10000 / 100); 2147 return 0; 2148 } 2149 2150 static int ioc_qos_show(struct seq_file *sf, void *v) 2151 { 2152 struct blkcg *blkcg = css_to_blkcg(seq_css(sf)); 2153 2154 blkcg_print_blkgs(sf, blkcg, ioc_qos_prfill, 2155 &blkcg_policy_iocost, seq_cft(sf)->private, false); 2156 return 0; 2157 } 2158 2159 static const match_table_t qos_ctrl_tokens = { 2160 { QOS_ENABLE, "enable=%u" }, 2161 { QOS_CTRL, "ctrl=%s" }, 2162 { NR_QOS_CTRL_PARAMS, NULL }, 2163 }; 2164 2165 static const match_table_t qos_tokens = { 2166 { QOS_RPPM, "rpct=%s" }, 2167 { QOS_RLAT, "rlat=%u" }, 2168 { QOS_WPPM, "wpct=%s" }, 2169 { QOS_WLAT, "wlat=%u" }, 2170 { QOS_MIN, "min=%s" }, 2171 { QOS_MAX, "max=%s" }, 2172 { NR_QOS_PARAMS, NULL }, 2173 }; 2174 2175 static ssize_t ioc_qos_write(struct kernfs_open_file *of, char *input, 2176 size_t nbytes, loff_t off) 2177 { 2178 struct gendisk *disk; 2179 struct ioc *ioc; 2180 u32 qos[NR_QOS_PARAMS]; 2181 bool enable, user; 2182 char *p; 2183 int ret; 2184 2185 disk = blkcg_conf_get_disk(&input); 2186 if (IS_ERR(disk)) 2187 return PTR_ERR(disk); 2188 2189 ioc = q_to_ioc(disk->queue); 2190 if (!ioc) { 2191 ret = blk_iocost_init(disk->queue); 2192 if (ret) 2193 goto err; 2194 ioc = q_to_ioc(disk->queue); 2195 } 2196 2197 spin_lock_irq(&ioc->lock); 2198 memcpy(qos, ioc->params.qos, sizeof(qos)); 2199 enable = ioc->enabled; 2200 user = ioc->user_qos_params; 2201 spin_unlock_irq(&ioc->lock); 2202 2203 while ((p = strsep(&input, " \t\n"))) { 2204 substring_t args[MAX_OPT_ARGS]; 2205 char buf[32]; 2206 int tok; 2207 s64 v; 2208 2209 if (!*p) 2210 continue; 2211 2212 switch (match_token(p, qos_ctrl_tokens, args)) { 2213 case QOS_ENABLE: 2214 match_u64(&args[0], &v); 2215 enable = v; 2216 continue; 2217 case QOS_CTRL: 2218 match_strlcpy(buf, &args[0], sizeof(buf)); 2219 if (!strcmp(buf, "auto")) 2220 user = false; 2221 else if (!strcmp(buf, "user")) 2222 user = true; 2223 else 2224 goto einval; 2225 continue; 2226 } 2227 2228 tok = match_token(p, qos_tokens, args); 2229 switch (tok) { 2230 case QOS_RPPM: 2231 case QOS_WPPM: 2232 if (match_strlcpy(buf, &args[0], sizeof(buf)) >= 2233 sizeof(buf)) 2234 goto einval; 2235 if (cgroup_parse_float(buf, 2, &v)) 2236 goto einval; 2237 if (v < 0 || v > 10000) 2238 goto einval; 2239 qos[tok] = v * 100; 2240 break; 2241 case QOS_RLAT: 2242 case QOS_WLAT: 2243 if (match_u64(&args[0], &v)) 2244 goto einval; 2245 qos[tok] = v; 2246 break; 2247 case QOS_MIN: 2248 case QOS_MAX: 2249 if (match_strlcpy(buf, &args[0], sizeof(buf)) >= 2250 sizeof(buf)) 2251 goto einval; 2252 if (cgroup_parse_float(buf, 2, &v)) 2253 goto einval; 2254 if (v < 0) 2255 goto einval; 2256 qos[tok] = clamp_t(s64, v * 100, 2257 VRATE_MIN_PPM, VRATE_MAX_PPM); 2258 break; 2259 default: 2260 goto einval; 2261 } 2262 user = true; 2263 } 2264 2265 if (qos[QOS_MIN] > qos[QOS_MAX]) 2266 goto einval; 2267 2268 spin_lock_irq(&ioc->lock); 2269 2270 if (enable) { 2271 blk_queue_flag_set(QUEUE_FLAG_RQ_ALLOC_TIME, ioc->rqos.q); 2272 ioc->enabled = true; 2273 } else { 2274 blk_queue_flag_clear(QUEUE_FLAG_RQ_ALLOC_TIME, ioc->rqos.q); 2275 ioc->enabled = false; 2276 } 2277 2278 if (user) { 2279 memcpy(ioc->params.qos, qos, sizeof(qos)); 2280 ioc->user_qos_params = true; 2281 } else { 2282 ioc->user_qos_params = false; 2283 } 2284 2285 ioc_refresh_params(ioc, true); 2286 spin_unlock_irq(&ioc->lock); 2287 2288 put_disk_and_module(disk); 2289 return nbytes; 2290 einval: 2291 ret = -EINVAL; 2292 err: 2293 put_disk_and_module(disk); 2294 return ret; 2295 } 2296 2297 static u64 ioc_cost_model_prfill(struct seq_file *sf, 2298 struct blkg_policy_data *pd, int off) 2299 { 2300 const char *dname = blkg_dev_name(pd->blkg); 2301 struct ioc *ioc = pd_to_iocg(pd)->ioc; 2302 u64 *u = ioc->params.i_lcoefs; 2303 2304 if (!dname) 2305 return 0; 2306 2307 seq_printf(sf, "%s ctrl=%s model=linear " 2308 "rbps=%llu rseqiops=%llu rrandiops=%llu " 2309 "wbps=%llu wseqiops=%llu wrandiops=%llu\n", 2310 dname, ioc->user_cost_model ? "user" : "auto", 2311 u[I_LCOEF_RBPS], u[I_LCOEF_RSEQIOPS], u[I_LCOEF_RRANDIOPS], 2312 u[I_LCOEF_WBPS], u[I_LCOEF_WSEQIOPS], u[I_LCOEF_WRANDIOPS]); 2313 return 0; 2314 } 2315 2316 static int ioc_cost_model_show(struct seq_file *sf, void *v) 2317 { 2318 struct blkcg *blkcg = css_to_blkcg(seq_css(sf)); 2319 2320 blkcg_print_blkgs(sf, blkcg, ioc_cost_model_prfill, 2321 &blkcg_policy_iocost, seq_cft(sf)->private, false); 2322 return 0; 2323 } 2324 2325 static const match_table_t cost_ctrl_tokens = { 2326 { COST_CTRL, "ctrl=%s" }, 2327 { COST_MODEL, "model=%s" }, 2328 { NR_COST_CTRL_PARAMS, NULL }, 2329 }; 2330 2331 static const match_table_t i_lcoef_tokens = { 2332 { I_LCOEF_RBPS, "rbps=%u" }, 2333 { I_LCOEF_RSEQIOPS, "rseqiops=%u" }, 2334 { I_LCOEF_RRANDIOPS, "rrandiops=%u" }, 2335 { I_LCOEF_WBPS, "wbps=%u" }, 2336 { I_LCOEF_WSEQIOPS, "wseqiops=%u" }, 2337 { I_LCOEF_WRANDIOPS, "wrandiops=%u" }, 2338 { NR_I_LCOEFS, NULL }, 2339 }; 2340 2341 static ssize_t ioc_cost_model_write(struct kernfs_open_file *of, char *input, 2342 size_t nbytes, loff_t off) 2343 { 2344 struct gendisk *disk; 2345 struct ioc *ioc; 2346 u64 u[NR_I_LCOEFS]; 2347 bool user; 2348 char *p; 2349 int ret; 2350 2351 disk = blkcg_conf_get_disk(&input); 2352 if (IS_ERR(disk)) 2353 return PTR_ERR(disk); 2354 2355 ioc = q_to_ioc(disk->queue); 2356 if (!ioc) { 2357 ret = blk_iocost_init(disk->queue); 2358 if (ret) 2359 goto err; 2360 ioc = q_to_ioc(disk->queue); 2361 } 2362 2363 spin_lock_irq(&ioc->lock); 2364 memcpy(u, ioc->params.i_lcoefs, sizeof(u)); 2365 user = ioc->user_cost_model; 2366 spin_unlock_irq(&ioc->lock); 2367 2368 while ((p = strsep(&input, " \t\n"))) { 2369 substring_t args[MAX_OPT_ARGS]; 2370 char buf[32]; 2371 int tok; 2372 u64 v; 2373 2374 if (!*p) 2375 continue; 2376 2377 switch (match_token(p, cost_ctrl_tokens, args)) { 2378 case COST_CTRL: 2379 match_strlcpy(buf, &args[0], sizeof(buf)); 2380 if (!strcmp(buf, "auto")) 2381 user = false; 2382 else if (!strcmp(buf, "user")) 2383 user = true; 2384 else 2385 goto einval; 2386 continue; 2387 case COST_MODEL: 2388 match_strlcpy(buf, &args[0], sizeof(buf)); 2389 if (strcmp(buf, "linear")) 2390 goto einval; 2391 continue; 2392 } 2393 2394 tok = match_token(p, i_lcoef_tokens, args); 2395 if (tok == NR_I_LCOEFS) 2396 goto einval; 2397 if (match_u64(&args[0], &v)) 2398 goto einval; 2399 u[tok] = v; 2400 user = true; 2401 } 2402 2403 spin_lock_irq(&ioc->lock); 2404 if (user) { 2405 memcpy(ioc->params.i_lcoefs, u, sizeof(u)); 2406 ioc->user_cost_model = true; 2407 } else { 2408 ioc->user_cost_model = false; 2409 } 2410 ioc_refresh_params(ioc, true); 2411 spin_unlock_irq(&ioc->lock); 2412 2413 put_disk_and_module(disk); 2414 return nbytes; 2415 2416 einval: 2417 ret = -EINVAL; 2418 err: 2419 put_disk_and_module(disk); 2420 return ret; 2421 } 2422 2423 static struct cftype ioc_files[] = { 2424 { 2425 .name = "weight", 2426 .flags = CFTYPE_NOT_ON_ROOT, 2427 .seq_show = ioc_weight_show, 2428 .write = ioc_weight_write, 2429 }, 2430 { 2431 .name = "cost.qos", 2432 .flags = CFTYPE_ONLY_ON_ROOT, 2433 .seq_show = ioc_qos_show, 2434 .write = ioc_qos_write, 2435 }, 2436 { 2437 .name = "cost.model", 2438 .flags = CFTYPE_ONLY_ON_ROOT, 2439 .seq_show = ioc_cost_model_show, 2440 .write = ioc_cost_model_write, 2441 }, 2442 {} 2443 }; 2444 2445 static struct blkcg_policy blkcg_policy_iocost = { 2446 .dfl_cftypes = ioc_files, 2447 .cpd_alloc_fn = ioc_cpd_alloc, 2448 .cpd_free_fn = ioc_cpd_free, 2449 .pd_alloc_fn = ioc_pd_alloc, 2450 .pd_init_fn = ioc_pd_init, 2451 .pd_free_fn = ioc_pd_free, 2452 }; 2453 2454 static int __init ioc_init(void) 2455 { 2456 return blkcg_policy_register(&blkcg_policy_iocost); 2457 } 2458 2459 static void __exit ioc_exit(void) 2460 { 2461 return blkcg_policy_unregister(&blkcg_policy_iocost); 2462 } 2463 2464 module_init(ioc_init); 2465 module_exit(ioc_exit); 2466