1 ======================== 2 Deadline Task Scheduling 3 ======================== 4 5 .. CONTENTS 6 7 0. WARNING 8 1. Overview 9 2. Scheduling algorithm 10 2.1 Main algorithm 11 2.2 Bandwidth reclaiming 12 3. Scheduling Real-Time Tasks 13 3.1 Definitions 14 3.2 Schedulability Analysis for Uniproce 15 3.3 Schedulability Analysis for Multipro 16 3.4 Relationship with SCHED_DEADLINE Par 17 4. Bandwidth management 18 4.1 System-wide settings 19 4.2 Task interface 20 4.3 Default behavior 21 4.4 Behavior of sched_yield() 22 5. Tasks CPU affinity 23 5.1 SCHED_DEADLINE and cpusets HOWTO 24 6. Future plans 25 A. Test suite 26 B. Minimal main() 27 28 29 0. WARNING 30 ========== 31 32 Fiddling with these settings can result in an 33 system behavior. As for -rt (group) schedulin 34 know what they're doing. 35 36 37 1. Overview 38 =========== 39 40 The SCHED_DEADLINE policy contained inside th 41 basically an implementation of the Earliest D 42 algorithm, augmented with a mechanism (called 43 that makes it possible to isolate the behavio 44 45 46 2. Scheduling algorithm 47 ======================= 48 49 2.1 Main algorithm 50 ------------------ 51 52 SCHED_DEADLINE [18] uses three parameters, na 53 "deadline", to schedule tasks. A SCHED_DEADLI 54 "runtime" microseconds of execution time ever 55 these "runtime" microseconds are available wi 56 from the beginning of the period. In order t 57 every time the task wakes up, the scheduler c 58 consistent with the guarantee (using the CBS[ 59 scheduled using EDF[1] on these scheduling de 60 earliest scheduling deadline is selected for 61 task actually receives "runtime" time units w 62 "admission control" strategy (see Section "4. 63 (clearly, if the system is overloaded this gu 64 65 Summing up, the CBS[2,3] algorithm assigns sc 66 that each task runs for at most its runtime e 67 interference between different tasks (bandwid 68 algorithm selects the task with the earliest 69 to be executed next. Thanks to this feature, 70 with the "traditional" real-time task model ( 71 use the new policy. 72 73 In more details, the CBS algorithm assigns sc 74 tasks in the following way: 75 76 - Each SCHED_DEADLINE task is characterized 77 "deadline", and "period" parameters; 78 79 - The state of the task is described by a "s 80 a "remaining runtime". These two parameter 81 82 - When a SCHED_DEADLINE task wakes up (becom 83 the scheduler checks if:: 84 85 remaining runtime 86 ---------------------------------- 87 scheduling deadline - current time 88 89 then, if the scheduling deadline is smalle 90 this condition is verified, the scheduling 91 remaining runtime are re-initialized as 92 93 scheduling deadline = current time + 94 remaining runtime = runtime 95 96 otherwise, the scheduling deadline and the 97 left unchanged; 98 99 - When a SCHED_DEADLINE task executes for an 100 remaining runtime is decreased as:: 101 102 remaining runtime = remaining runtime 103 104 (technically, the runtime is decreased at 105 task is descheduled / preempted); 106 107 - When the remaining runtime becomes less or 108 said to be "throttled" (also known as "dep 109 and cannot be scheduled until its scheduli 110 time" for this task (see next item) is set 111 value of the scheduling deadline; 112 113 - When the current time is equal to the repl 114 throttled task, the scheduling deadline an 115 updated as:: 116 117 scheduling deadline = scheduling dead 118 remaining runtime = remaining runtime 119 120 The SCHED_FLAG_DL_OVERRUN flag in sched_attr' 121 to get informed about runtime overruns throug 122 signals. 123 124 125 2.2 Bandwidth reclaiming 126 ------------------------ 127 128 Bandwidth reclaiming for deadline tasks is ba 129 Reclamation of Unused Bandwidth) algorithm [1 130 when flag SCHED_FLAG_RECLAIM is set. 131 132 The following diagram illustrates the state n 133 134 ------------ 135 (d) | Active | 136 ------------->| | 137 | | Contending | 138 | ------------ 139 | A | 140 ---------- | | 141 | | | | 142 | Inactive | |(b) | (a) 143 | | | | 144 ---------- | | 145 A | V 146 | ------------ 147 | | Active | 148 --------------| Non | 149 (c) | Contending | 150 ------------ 151 152 A task can be in one of the following states: 153 154 - ActiveContending: if it is ready for execu 155 156 - ActiveNonContending: if it just blocked an 157 time; 158 159 - Inactive: if it is blocked and has surpass 160 161 State transitions: 162 163 (a) When a task blocks, it does not become i 164 bandwidth cannot be immediately reclaime 165 real-time guarantees. It therefore enter 166 ActiveNonContending. The scheduler arms 167 the 0-lag time, when the task's bandwidt 168 breaking the real-time guarantees. 169 170 The 0-lag time for a task entering the A 171 computed as:: 172 173 (runtime * dl_period) 174 deadline - --------------------- 175 dl_runtime 176 177 where runtime is the remaining runtime, 178 are the reservation parameters. 179 180 (b) If the task wakes up before the inactive 181 the ActiveContending state and the "inac 182 In addition, if the task wakes up on a d 183 the task's utilization must be removed f 184 utilization and must be added to the new 185 In order to avoid races between a task w 186 "inactive timer" is running on a differe 187 flag is used to indicate that a task is 188 (so, the flag is set when the task block 189 "inactive timer" fires or when the task 190 191 (c) When the "inactive timer" fires, the tas 192 its utilization is removed from the runq 193 194 (d) When an inactive task wakes up, it enter 195 its utilization is added to the active u 196 it has been enqueued. 197 198 For each runqueue, the algorithm GRUB keeps t 199 200 - Active bandwidth (running_bw): this is the 201 tasks in active state (i.e., ActiveContend 202 203 - Total bandwidth (this_bw): this is the sum 204 runqueue, including the tasks in Inactive 205 206 - Maximum usable bandwidth (max_bw): This is 207 deadline tasks and is currently set to the 208 209 210 The algorithm reclaims the bandwidth of the t 211 It does so by decrementing the runtime of the 212 to 213 214 dq = -(max{ Ui, (Umax - Uinact - Ue 215 216 where: 217 218 - Ui is the bandwidth of task Ti; 219 - Umax is the maximum reclaimable utilizatio 220 limits); 221 - Uinact is the (per runqueue) inactive util 222 (this_bq - running_bw); 223 - Uextra is the (per runqueue) extra reclaim 224 (subjected to RT throttling limits). 225 226 227 Let's now see a trivial example of two deadli 228 to 4 and period equal to 8 (i.e., bandwidth e 229 230 A Task T1 231 | 232 | | 233 | | 234 |-------- |---- 235 | | V 236 |---|---|---|---|---|---|---|---|---- 237 0 1 2 3 4 5 6 7 8 238 239 240 A Task T2 241 | 242 | | 243 | | 244 | ------------------------| 245 | | V 246 |---|---|---|---|---|---|---|---|---- 247 0 1 2 3 4 5 6 7 8 248 249 250 A running_bw 251 | 252 1 ----------------- ----- 253 | | | 254 0.5- ----------------- 255 | | 256 |---|---|---|---|---|---|---|---|---- 257 0 1 2 3 4 5 6 7 8 258 259 260 - Time t = 0: 261 262 Both tasks are ready for execution and the 263 Suppose Task T1 is the first task to start 264 Since there are no inactive tasks, its run 265 266 - Time t = 2: 267 268 Suppose that task T1 blocks 269 Task T1 therefore enters the ActiveNonCont 270 runtime is equal to 2, its 0-lag time is e 271 Task T2 start execution, with runtime stil 272 there are no inactive tasks. 273 274 - Time t = 4: 275 276 This is the 0-lag time for Task T1. Since 277 meantime, it enters the Inactive state. It 278 running_bw. 279 Task T2 continues its execution. However, 280 dq = - 0.5 dt because Uinact = 0.5. 281 Task T2 therefore reclaims the bandwidth u 282 283 - Time t = 8: 284 285 Task T1 wakes up. It enters the ActiveCont 286 running_bw is incremented. 287 288 289 2.3 Energy-aware scheduling 290 --------------------------- 291 292 When cpufreq's schedutil governor is selected 293 GRUB-PA [19] algorithm, reducing the CPU oper 294 value that still allows to meet the deadlines 295 implemented only for ARM architectures. 296 297 A particular care must be taken in case the t 298 is of the same order of magnitude of the rese 299 setting a fixed CPU frequency results in a lo 300 301 302 3. Scheduling Real-Time Tasks 303 ============================= 304 305 306 307 .. BIG FAT WARNING ************************* 308 309 .. warning:: 310 311 This section contains a (not-thorough) summ 312 scheduling theory, and how it applies to SC 313 The reader can "safely" skip to Section 4 i 314 how the scheduling policy can be used. Anyw 315 to come back here and continue reading (onc 316 satisfied :P) to be sure of fully understan 317 318 .. ****************************************** 319 320 There are no limitations on what kind of task 321 scheduling discipline, even if it must be sai 322 suited for periodic or sporadic real-time tas 323 timing behavior, e.g., multimedia, streaming, 324 325 3.1 Definitions 326 ------------------------ 327 328 A typical real-time task is composed of a rep 329 (task instances, or jobs) which are activated 330 fashion. 331 Each job J_j (where J_j is the j^th job of th 332 arrival time r_j (the time when the job start 333 time c_j needed to finish the job, and a job 334 is the time within which the job should be fi 335 time max{c_j} is called "Worst Case Execution 336 A real-time task can be periodic with period 337 sporadic with minimum inter-arrival time P is 338 d_j = r_j + D, where D is the task's relative 339 Summing up, a real-time task can be described 340 341 Task = (WCET, D, P) 342 343 The utilization of a real-time task is define 344 WCET and its period (or minimum inter-arrival 345 the fraction of CPU time needed to execute th 346 347 If the total utilization U=sum(WCET_i/P_i) is 348 to the number of CPUs), then the scheduler is 349 deadlines. 350 Note that total utilization is defined as the 351 WCET_i/P_i over all the real-time tasks in th 352 multiple real-time tasks, the parameters of t 353 with the "_i" suffix. 354 Moreover, if the total utilization is larger 355 non- real-time tasks by real-time tasks. 356 If, instead, the total utilization is smaller 357 tasks will not be starved and the system migh 358 deadlines. 359 As a matter of fact, in this case it is possi 360 for tardiness (defined as the maximum between 361 between the finishing time of a job and its a 362 More precisely, it can be proven that using a 363 maximum tardiness of each task is smaller or 364 365 ((M − 1) · WCET_max − WCET_min)/( 366 367 where WCET_max = max{WCET_i} is the maximum W 368 is the minimum WCET, and U_max = max{WCET_i/P 369 utilization[12]. 370 371 3.2 Schedulability Analysis for Uniprocessor S 372 ---------------------------------------------- 373 374 If M=1 (uniprocessor system), or in case of p 375 real-time task is statically assigned to one 376 possible to formally check if all the deadlin 377 If D_i = P_i for all tasks, then EDF is able 378 of all the tasks executing on a CPU if and on 379 of the tasks running on such a CPU is smaller 380 If D_i != P_i for some task, then it is possi 381 a task as WCET_i/min{D_i,P_i}, and EDF is abl 382 of all the tasks running on a CPU if the sum 383 running on such a CPU is smaller or equal tha 384 385 sum(WCET_i / min{D_i, P_i}) <= 1 386 387 It is important to notice that this condition 388 necessary: there are task sets that are sched 389 condition. For example, consider the task set 390 Task_1=(50ms,50ms,100ms) and Task_2=(10ms,100 391 EDF is clearly able to schedule the two tasks 392 (Task_1 is scheduled as soon as it is release 393 to respect its deadline; Task_2 is scheduled 394 its response time cannot be larger than 50ms 395 396 50 / min{50,100} + 10 / min{100, 100} 397 398 Of course it is possible to test the exact sc 399 D_i != P_i (checking a condition that is both 400 but this cannot be done by comparing the tota 401 a constant. Instead, the so called "processor 402 computing the total amount of CPU time h(t) n 403 respect all of their deadlines in a time inte 404 such a time with the interval size t. If h(t) 405 the amount of time needed by the tasks in a t 406 smaller than the size of the interval) for al 407 EDF is able to schedule the tasks respecting 408 performing this check for all possible values 409 proven[4,5,6] that it is sufficient to perfor 410 between 0 and a maximum value L. The cited pa 411 mathematical details and explain how to compu 412 In any case, this kind of analysis is too com 413 time-consuming to be performed on-line. Hence 414 4 Linux uses an admission test based on the t 415 416 3.3 Schedulability Analysis for Multiprocessor 417 ---------------------------------------------- 418 419 On multiprocessor systems with global EDF sch 420 systems), a sufficient test for schedulabilit 421 utilizations or densities: it can be shown th 422 sets with utilizations slightly larger than 1 423 of the number of CPUs. 424 425 Consider a set {Task_1,...Task_{M+1}} of M+1 426 CPUs, with the first task Task_1=(P,P,P) havi 427 and WCET equal to P. The remaining M tasks Ta 428 arbitrarily small worst case execution time ( 429 period smaller than the one of the first task 430 activate at the same time t, global EDF sched 431 (because their absolute deadlines are equal t 432 smaller than the absolute deadline of Task_1, 433 result, Task_1 can be scheduled only at time 434 time t + e + P, after its absolute deadline. 435 task set is U = M · e / (P - 1) + P / P = M 436 values of e this can become very close to 1. 437 effect"[7]. Note: the example in the original 438 slightly simplified here (for example, Dhall 439 lim_{e->0}U). 440 441 More complex schedulability tests for global 442 real-time literature[8,9], but they are not b 443 between total utilization (or density) and a 444 have D_i = P_i, a sufficient schedulability c 445 a simple way: 446 447 sum(WCET_i / P_i) <= M - (M - 1) · U_ 448 449 where U_max = max{WCET_i / P_i}[10]. Notice t 450 M - (M - 1) · U_max becomes M - M + 1 = 1 an 451 just confirms the Dhall's effect. A more comp 452 about schedulability tests for multi-processo 453 found in [11]. 454 455 As seen, enforcing that the total utilization 456 guarantee that global EDF schedules the tasks 457 (in other words, global EDF is not an optimal 458 a total utilization smaller than M is enough 459 tasks are not starved and that the tardiness 460 bound[12] (as previously noted). Different bo 461 experienced by real-time tasks have been deve 462 but the theoretical result that is important 463 the total utilization is smaller or equal tha 464 the tasks are limited. 465 466 3.4 Relationship with SCHED_DEADLINE Parameter 467 ---------------------------------------------- 468 469 Finally, it is important to understand the re 470 SCHED_DEADLINE scheduling parameters describe 471 deadline and period) and the real-time task p 472 described in this section. Note that the task 473 represented by its absolute deadlines d_j = r 474 SCHED_DEADLINE schedules the tasks according 475 Section 2). 476 If an admission test is used to guarantee tha 477 are respected, then SCHED_DEADLINE can be use 478 guaranteeing that all the jobs' deadlines of 479 In order to do this, a task must be scheduled 480 481 - runtime >= WCET 482 - deadline = D 483 - period <= P 484 485 IOW, if runtime >= WCET and if period is <= P 486 and the absolute deadlines (d_j) coincide, so 487 allows to respect the jobs' absolute deadline 488 called "hard schedulability property" and is 489 Notice that if runtime > deadline the admissi 490 this task, as it is not possible to respect i 491 492 References: 493 494 1 - C. L. Liu and J. W. Layland. Scheduling 495 ming in a hard-real-time environment. Jo 496 Computing Machinery, 20(1), 1973. 497 2 - L. Abeni , G. Buttazzo. Integrating Mult 498 Real-Time Systems. Proceedings of the 19 499 Symposium, 1998. http://retis.sssup.it/~ 500 3 - L. Abeni. Server Mechanisms for Multimed 501 Technical Report. http://disi.unitn.it/~ 502 4 - J. Y. Leung and M.L. Merril. A Note on P 503 Periodic, Real-Time Tasks. Information P 504 no. 3, pp. 115-118, 1980. 505 5 - S. K. Baruah, A. K. Mok and L. E. Rosier 506 Hard-Real-Time Sporadic Tasks on One Pro 507 11th IEEE Real-time Systems Symposium, 1 508 6 - S. K. Baruah, L. E. Rosier and R. R. How 509 Concerning the Preemptive Scheduling of 510 One Processor. Real-Time Systems Journal 511 1990. 512 7 - S. J. Dhall and C. L. Liu. On a real-tim 513 research, vol. 26, no. 1, pp 127-140, 19 514 8 - T. Baker. Multiprocessor EDF and Deadlin 515 Analysis. Proceedings of the 24th IEEE R 516 9 - T. Baker. An Analysis of EDF Schedulabil 517 IEEE Transactions on Parallel and Distri 518 pp 760-768, 2005. 519 10 - J. Goossens, S. Funk and S. Baruah, Pri 520 Periodic Task Systems on Multiprocessor 521 vol. 25, no. 2–3, pp. 187–205, 2003 522 11 - R. Davis and A. Burns. A Survey of Hard 523 Multiprocessor Systems. ACM Computing S 524 http://www-users.cs.york.ac.uk/~robdavi 525 12 - U. C. Devi and J. H. Anderson. Tardines 526 Scheduling on a Multiprocessor. Real-Ti 527 no. 2, pp 133-189, 2008. 528 13 - P. Valente and G. Lipari. An Upper Boun 529 Real-Time Tasks Scheduled by EDF on Mul 530 the 26th IEEE Real-Time Systems Symposi 531 14 - J. Erickson, U. Devi and S. Baruah. Imp 532 Global EDF. Proceedings of the 22nd Eur 533 Real-Time Systems, 2010. 534 15 - G. Lipari, S. Baruah, Greedy reclamatio 535 constant-bandwidth servers, 12th IEEE E 536 Systems, 2000. 537 16 - L. Abeni, J. Lelli, C. Scordino, L. Pal 538 SCHED DEADLINE. In Proceedings of the R 539 Dusseldorf, Germany, 2014. 540 17 - L. Abeni, G. Lipari, A. Parri, Y. Sun, 541 or sequential?. In Proceedings of the 3 542 Computing, 2016. 543 18 - J. Lelli, C. Scordino, L. Abeni, D. Fag 544 Linux kernel, Software: Practice and Ex 545 2016. 546 19 - C. Scordino, L. Abeni, J. Lelli, Energy 547 the Linux Kernel, 33rd ACM/SIGAPP Sympo 548 2018), Pau, France, April 2018. 549 550 551 4. Bandwidth management 552 ======================= 553 554 As previously mentioned, in order for -deadli 555 effective and useful (that is, to be able to 556 within "deadline"), it is important to have s 557 of the available fractions of CPU time to the 558 This is usually called "admission control" an 559 no guarantee can be given on the actual sched 560 561 As already stated in Section 3, a necessary c 562 correctly schedule a set of real-time tasks i 563 is smaller than M. When talking about -deadli 564 the sum of the ratio between runtime and peri 565 than M. Notice that the ratio runtime/period 566 of a "traditional" real-time task, and is als 567 "bandwidth". 568 The interface used to control the CPU bandwid 569 to -deadline tasks is similar to the one alre 570 tasks with real-time group scheduling (a.k.a. 571 Documentation/scheduler/sched-rt-group.rst), 572 writable control files located in procfs (for 573 Notice that per-group settings (controlled th 574 defined for -deadline tasks, because more dis 575 figure out how we want to manage SCHED_DEADLI 576 level. 577 578 A main difference between deadline bandwidth 579 is that -deadline tasks have bandwidth on the 580 and thus we don't need a higher level throttl 581 desired bandwidth. In other words, this means 582 only used at admission control time (i.e., wh 583 sched_setattr()). Scheduling is then performe 584 parameters, so that CPU bandwidth is allocate 585 respecting their needs in terms of granularit 586 interface we can put a cap on total utilizati 587 \Sum (runtime_i / period_i) < global_dl_utili 588 589 4.1 System wide settings 590 ------------------------ 591 592 The system wide settings are configured under 593 594 For now the -rt knobs are used for -deadline 595 -deadline runtime is accounted against the -r 596 isn't entirely desirable; however, it is bett 597 now, and be able to change it easily later. T 598 run -rt tasks from a -deadline server; in whi 599 direct subset of dl_bw. 600 601 This means that, for a root_domain comprising 602 can be created while the sum of their bandwid 603 604 M * (sched_rt_runtime_us / sched_rt_period_ 605 606 It is also possible to disable this bandwidth 607 be thus free of oversubscribing the system up 608 This is done by writing -1 in /proc/sys/kerne 609 610 611 4.2 Task interface 612 ------------------ 613 614 Specifying a periodic/sporadic task that exec 615 runtime at each instance, and that is schedul 616 its own timing constraints needs, in general, 617 618 - a (maximum/typical) instance execution tim 619 - a minimum interval between consecutive ins 620 - a time constraint by which each instance m 621 622 Therefore: 623 624 * a new struct sched_attr, containing all th 625 provided; 626 * the new scheduling related syscalls that m 627 sched_setattr() and sched_getattr() are im 628 629 For debugging purposes, the leftover runtime 630 SCHED_DEADLINE task can be retrieved through 631 dl.runtime and dl.deadline, both values in ns 632 retrieve these values from production code is 633 634 635 4.3 Default behavior 636 --------------------- 637 638 The default value for SCHED_DEADLINE bandwidt 639 950000. With rt_period equal to 1000000, by d 640 tasks can use at most 95%, multiplied by the 641 root_domain, for each root_domain. 642 This means that non -deadline tasks will rece 643 and that -deadline tasks will receive their r 644 worst-case delay respect to the "deadline" pa 645 and the cpuset mechanism is used to implement 646 Section 5), then this simple setting of the b 647 deterministically guarantee that -deadline ta 648 in a period. 649 650 Finally, notice that in order not to jeopardi 651 -deadline task cannot fork. 652 653 654 4.4 Behavior of sched_yield() 655 ----------------------------- 656 657 When a SCHED_DEADLINE task calls sched_yield( 658 remaining runtime and is immediately throttle 659 period, when its runtime will be replenished 660 dl_yielded is set and used to handle correctl 661 replenishment after a call to sched_yield()). 662 663 This behavior of sched_yield() allows the tas 664 the beginning of the next period. Also, this 665 future with bandwidth reclaiming mechanisms, 666 make the leftoever runtime available for recl 667 SCHED_DEADLINE tasks. 668 669 670 5. Tasks CPU affinity 671 ===================== 672 673 -deadline tasks cannot have an affinity mask 674 root_domain they are created on. However, aff 675 through the cpuset facility (Documentation/ad 676 677 5.1 SCHED_DEADLINE and cpusets HOWTO 678 ------------------------------------ 679 680 An example of a simple configuration (pin a - 681 follows (rt-app is used to create a -deadline 682 683 mkdir /dev/cpuset 684 mount -t cgroup -o cpuset cpuset /dev/cpuse 685 cd /dev/cpuset 686 mkdir cpu0 687 echo 0 > cpu0/cpuset.cpus 688 echo 0 > cpu0/cpuset.mems 689 echo 1 > cpuset.cpu_exclusive 690 echo 0 > cpuset.sched_load_balance 691 echo 1 > cpu0/cpuset.cpu_exclusive 692 echo 1 > cpu0/cpuset.mem_exclusive 693 echo $$ > cpu0/tasks 694 rt-app -t 100000:10000:d:0 -D5 # it is now 695 # task affin 696 697 6. Future plans 698 =============== 699 700 Still missing: 701 702 - programmatic way to retrieve current runti 703 - refinements to deadline inheritance, espec 704 of retaining bandwidth isolation among non 705 being studied from both theoretical and pr 706 hopefully we should be able to produce som 707 - (c)group based bandwidth management, and m 708 - access control for non-root users (and rel 709 address), which is the best way to allow u 710 and how to prevent non-root users "cheat" 711 712 As already discussed, we are planning also to 713 throttling patches [https://lore.kernel.org/r 714 the preliminary phases of the merge and we re 715 help us decide on the direction it should tak 716 717 Appendix A. Test suite 718 ====================== 719 720 The SCHED_DEADLINE policy can be easily teste 721 are part of a wider Linux Scheduler validatio 722 available as a GitHub repository: https://git 723 724 The first testing application is called rt-ap 725 start multiple threads with specific paramete 726 SCHED_{OTHER,FIFO,RR,DEADLINE} scheduling pol 727 parameters (e.g., niceness, priority, runtime 728 is a valuable tool, as it can be used to synt 729 workloads (maybe mimicking real use-cases) an 730 behaves under such workloads. In this way, re 731 rt-app is available at: https://github.com/sc 732 733 Thread parameters can be specified from the c 734 this:: 735 736 # rt-app -t 100000:10000:d -t 150000:20000:f 737 738 The above creates 2 threads. The first one, s 739 executes for 10ms every 100ms. The second one 740 priority 10, executes for 20ms every 150ms. T 741 of 5 seconds. 742 743 More interestingly, configurations can be des 744 can be passed as input to rt-app with somethi 745 746 # rt-app my_config.json 747 748 The parameters that can be specified with the 749 of the command line options. Please refer to 750 details (`<rt-app-sources>/doc/*.json`). 751 752 The second testing application is a modificat 753 schedtool-dl, which can be used to setup SCHE 754 certain pid/application. schedtool-dl is avai 755 https://github.com/scheduler-tools/schedtool- 756 757 The usage is straightforward:: 758 759 # schedtool -E -t 10000000:100000000 -e ./my 760 761 With this, my_cpuhog_app is put to run inside 762 of 10ms every 100ms (note that parameters are 763 You can also use schedtool to create a reserv 764 application, given that you know its pid:: 765 766 # schedtool -E -t 10000000:100000000 my_app_ 767 768 Appendix B. Minimal main() 769 ========================== 770 771 We provide in what follows a simple (ugly) se 772 showing how SCHED_DEADLINE reservations can b 773 application developer:: 774 775 #define _GNU_SOURCE 776 #include <unistd.h> 777 #include <stdio.h> 778 #include <stdlib.h> 779 #include <string.h> 780 #include <time.h> 781 #include <linux/unistd.h> 782 #include <linux/kernel.h> 783 #include <linux/types.h> 784 #include <sys/syscall.h> 785 #include <pthread.h> 786 787 #define gettid() syscall(__NR_gettid) 788 789 #define SCHED_DEADLINE 6 790 791 /* XXX use the proper syscall numbers */ 792 #ifdef __x86_64__ 793 #define __NR_sched_setattr 314 794 #define __NR_sched_getattr 315 795 #endif 796 797 #ifdef __i386__ 798 #define __NR_sched_setattr 351 799 #define __NR_sched_getattr 352 800 #endif 801 802 #ifdef __arm__ 803 #define __NR_sched_setattr 380 804 #define __NR_sched_getattr 381 805 #endif 806 807 static volatile int done; 808 809 struct sched_attr { 810 __u32 size; 811 812 __u32 sched_policy; 813 __u64 sched_flags; 814 815 /* SCHED_NORMAL, SCHED_BATCH */ 816 __s32 sched_nice; 817 818 /* SCHED_FIFO, SCHED_RR */ 819 __u32 sched_priority; 820 821 /* SCHED_DEADLINE (nsec) */ 822 __u64 sched_runtime; 823 __u64 sched_deadline; 824 __u64 sched_period; 825 }; 826 827 int sched_setattr(pid_t pid, 828 const struct sched_attr *att 829 unsigned int flags) 830 { 831 return syscall(__NR_sched_setattr, pid 832 } 833 834 int sched_getattr(pid_t pid, 835 struct sched_attr *attr, 836 unsigned int size, 837 unsigned int flags) 838 { 839 return syscall(__NR_sched_getattr, pid 840 } 841 842 void *run_deadline(void *data) 843 { 844 struct sched_attr attr; 845 int x = 0; 846 int ret; 847 unsigned int flags = 0; 848 849 printf("deadline thread started [%ld]\ 850 851 attr.size = sizeof(attr); 852 attr.sched_flags = 0; 853 attr.sched_nice = 0; 854 attr.sched_priority = 0; 855 856 /* This creates a 10ms/30ms reservatio 857 attr.sched_policy = SCHED_DEADLINE; 858 attr.sched_runtime = 10 * 1000 * 1000; 859 attr.sched_period = attr.sched_deadlin 860 861 ret = sched_setattr(0, &attr, flags); 862 if (ret < 0) { 863 done = 0; 864 perror("sched_setattr"); 865 exit(-1); 866 } 867 868 while (!done) { 869 x++; 870 } 871 872 printf("deadline thread dies [%ld]\n", 873 return NULL; 874 } 875 876 int main (int argc, char **argv) 877 { 878 pthread_t thread; 879 880 printf("main thread [%ld]\n", gettid() 881 882 pthread_create(&thread, NULL, run_dead 883 884 sleep(10); 885 886 done = 1; 887 pthread_join(thread, NULL); 888 889 printf("main dies [%ld]\n", gettid()); 890 return 0; 891 }
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.