Library prosa.results.fifo.rta

From mathcomp Require Import ssreflect ssrbool eqtype ssrnat seq path fintype bigop.

Require Import prosa.model.readiness.basic.
Require Import prosa.model.priority.fifo.
Require Import prosa.analysis.facts.priority.fifo.
Require Import prosa.analysis.abstract.ideal.iw_instantiation.
Require Export prosa.analysis.facts.busy_interval.carry_in.
Require Export prosa.analysis.facts.busy_interval.ideal.inequalities.
Require Import prosa.analysis.facts.model.task_cost.

The formal development and the proofs in this file are described in-depth in the following paper:
  • Bedarkar et al., "From Intuition to Coq: A Case Study in Verified Response-Time Analysis of FIFO Scheduling", RTSS'22.
The interested reader is invited to follow along in parallel both in the paper and here. In particular, the below sections labeled A through H correspond directly to the equivalently labeled subsections in Section IV of the paper.

Response-Time Analysis for FIFO Schedulers

In the following, we derive a response-time analysis for FIFO schedulers, assuming a workload of sporadic real-time tasks characterized by arbitrary arrival curves executing upon an ideal uniprocessor. To this end, we instantiate the abstract Response-Time Analysis (aRTA) as provided in the prosa.analysis.abstract module.

Section AbstractRTAforFIFOwithArrivalCurves.

A. Defining the System Model

Before any formal claims can be stated, an initial setup is needed to define the system model under consideration. To this end, we next introduce and define the following notions using Prosa's standard definitions and behavioral semantics:
  • tasks, jobs, and their parameters,
  • the sequence of job arrivals,
  • worst-case execution time (WCET) and the absence of self-suspensions,
  • the set of tasks under analysis,
  • the task under analysis, and, finally,
  • an arbitrary schedule of the task set.

Tasks and Jobs

Consider any type of tasks, each characterized by a WCET task_cost, an arrival curve max_arrivals, and a run-to-completion threshold task_rtct, ...
  Context {Task : TaskType}.
  Context `{TaskCost Task}.
  Context `{MaxArrivals Task}.
  Context `{TaskRunToCompletionThreshold Task}.

... and any type of jobs associated with these tasks, where each job has an arrival time job_arrival, a cost job_cost, and an arbitrary preemption model indicated by job_preemptable.
  Context {Job : JobType} `{JobTask Job Task}.
  Context `{JobArrival Job}.
  Context `{JobCost Job}.
  Context `{JobPreemptable Job}.

The Job Arrival Sequence

Consider any arrival sequence arr_seq with consistent, non-duplicate arrivals.
  Variable arr_seq : arrival_sequence Job.
  Hypothesis H_valid_arrival_sequence : valid_arrival_sequence arr_seq.

Absence of Self-Suspensions and WCET Compliance

We assume the classic (i.e., Liu & Layland) model of readiness without jitter or self-suspensions, wherein pending jobs are always ready.
  #[local] Existing Instance basic_ready_instance.

We further require that a job's cost cannot exceed its task's stated WCET.
  Hypothesis H_valid_job_cost : arrivals_have_valid_job_costs arr_seq.

The Task Set

We consider an arbitrary task set ts...
  Variable ts : seq Task.

... and assume that all jobs stem from tasks in this task set.
  Hypothesis H_all_jobs_from_taskset : all_jobs_from_taskset arr_seq ts.

Furthermore, we assume that max_arrivals is a family of valid arrival curves that constrains the arrival sequence arr_seq, i.e., for any task tsk in ts, max_arrival tsk is (1) an arrival bound of tsk, and ...
  Hypothesis H_is_arrival_curve : taskset_respects_max_arrivals arr_seq ts.

... (2) a monotonic function that equals 0 for the empty interval delta = 0.
  Hypothesis H_valid_arrival_curve : valid_taskset_arrival_curve ts max_arrivals.

The Task Under Analysis

Let tsk be any task in ts that is to be analyzed.
  Variable tsk : Task.
  Hypothesis H_tsk_in_ts : tsk \in ts.

We assume that tsk is described by a valid task run-to-completion threshold. That is, there exists a task parameter task_rtct such that task_rtct tsk is
  • (1) no larger than tsk's WCET, and
  • (2) for any job of task tsk, the job's run-to-completion threshold job_rtct is bounded by task_rtct tsk.
  Hypothesis H_valid_run_to_completion_threshold :
    valid_task_run_to_completion_threshold arr_seq tsk.

The Schedule

Finally, consider any arbitrary, valid ideal uni-processor schedule of the given arrival sequence arr_seq (and hence the given task set ts).
  Variable sched : schedule (ideal.processor_state Job).
  Hypothesis H_valid_schedule : valid_schedule sched arr_seq.

We assume that the schedule complies with the preemption model ...
  Hypothesis H_valid_preemption_model : valid_preemption_model arr_seq sched.

... and, last but not least, that it respects the FIFO scheduling policy.
  Hypothesis H_respects_policy_at_preemption_point :
    respects_JLFP_policy_at_preemption_point arr_seq sched (FIFO Job).

B. Encoding the Scheduling Policy and Preemption Model

With the system model in place, the next step is to encode the scheduling policy and preemption model such that aRTA becomes applicable. To this end, we encode the semantics of the scheduling policy and preemption model using two functions, by convention called interference and interfering_workload. The paper explains the idea behind these two functions and how they interact in much more detail. At the code level, we fortunately can simply reuse the existing general definitions of interference and interfering workload that apply to any job-level fixed-priority (JLFP) policy (as provided in the module analysis.abstract.ideal.iw_instantiation).
  #[local] Instance ideal_jlfp_interference : Interference Job :=
    ideal_jlfp_interference arr_seq sched.

  #[local] Instance ideal_jlfp_interfering_workload : InterferingWorkload Job :=
    ideal_jlfp_interfering_workload arr_seq sched.

Please refer to the general definitions (by clicking on the links above) to see how they correspond to the definitions provided in Listing 3 of the paper (they are identical).

C. Classic and Abstract Work Conservation

The next step is to connect the classic notion of work conservation with the abstract notion assumed by aRTA. First, let us recall the abstract and classic notations of work conservation as work_conserving_ab and work_conserving_cl, respectively.
  Let work_conserving_ab := abstract.definitions.work_conserving arr_seq sched.
  Let work_conserving_cl := work_conserving.work_conserving arr_seq sched.

In the following, we make the standard assumption that the schedule is work-conserving in the classic sense.
  Hypothesis H_work_conserving : work_conserving_cl.

As explained in much detail in the paper, a general proof obligation of aRTA is to show that its abstract notion of work conservation is also satisfied. That is, the classic, policy-specific notion assumed in H_work_conserving needs to be "translated" into the abstract notion understood by aRTA. Fortunately, in our case the proof is trivial: as a benefit of reusing the general definitions of interference and interfering_workload for JLFP policies, we can reuse the existing general lemma instantiated_i_and_w_are_coherent_with_schedule. This lemma immediately allows us to conclude that the schedule is work-conserving in the abstract sense with respect to interference and interfering_workload.
  Fact abstractly_work_conserving : work_conserving_ab.
  Proof. by apply: instantiated_i_and_w_are_coherent_with_schedule; rt_eauto. Qed.

The preceding fact abstractly_work_conserving corresponds to Lemma 1 in the paper. To see the correspondence, refer to the definition of definitions.work_conserving (by clicking the link in the above definition).

D. Bounding the Maximum Busy-Window Length

The next step is to establish a bound on the maximum busy-window length, which aRTA requires to be given.
To this end, we assume that we are given a positive value L ...
  Variable L : duration.
  Hypothesis H_L_positive : L > 0.

... that is a fixed point of the following equation.
  Hypothesis H_fixed_point : L = total_request_bound_function ts L.

Given this definition of L, it is our proof obligation to show that all busy windows (in the abstract sense) are indeed bounded by L. To this end, let us first recall the notion of a bound on the maximum busy-window length (or, interchangeably, busy-interval length) as understood by aRTA.
  Let busy_windows_are_bounded_by L :=
    busy_intervals_are_bounded_by arr_seq sched tsk L.

We observe that the length of any (abstract) busy window in sched is indeed bounded by L. Again, the proof is trivial because we can reuse a general lemma, namely instantiated_busy_intervals_are_bounded in this case, due to the choice to reuse the existing JLFP definitions of interference and interfering_workload.
  Fact busy_windows_are_bounded : busy_windows_are_bounded_by L.
  Proof. by apply: instantiated_busy_intervals_are_bounded; rt_eauto. Qed.

The preceding fact busy_windows_are_bounded correspond to Lemma 2 in the paper. To clearly see the correspondence, refer to the definition of busy_intervals_are_bounded_by (by clicking on the link in the definition above).

E. Defining the Interference Bound Function (IBF)

Finally, we define the interference bound function (IBF). IBF bounds the cumulative interference incurred by a job in its busy window. In general, aRTA expects to reason about an IBF parametric in two parameters, a relative arrival offset A and an interval length Δ, as described in the paper. In our specific case, for FIFO scheduling, only A is actually relevant. We therefore define IBF as the sum, across all tasks, of the per-task request-bound functions (RBFs) in the interval A + ε minus the WCET of the task under analysis tsk.
  Let IBF tsk (A Δ : duration) :=
        (\sum_(tsko <- ts) task_request_bound_function tsko (A + ε)) - task_cost tsk.

As discussed in the paper, our proof obligation now is to show that the stated IBF is indeed correct. To this end, we first establish two auxiliary lemmas.

Absence of Priority Inversion

Because we reuse the general JLFP notions of interference and interfering_workload, which allowed us to save much proof effort in the preceding sections, we must reason about priority inversion. While priority inversion is conceptually not relevant under FIFO scheduling, it clearly is a factor in the general JLFP case, and hence shows up in the definitions of interference and interfering_workload. We therefore next show it to be actually impossible, too, by proving that, under FIFO scheduling, the cumulative priority inversion experienced by a job j in any interval within its busy window is always 0.
  Section AbsenceOfPriorityInversion.

Consider any job j of the task under analysis tsk.
    Variable j : Job.
    Hypothesis H_j_arrives : arrives_in arr_seq j.
    Hypothesis H_job_of_tsk : job_of_task tsk j.

Assume that the job has a positive cost (as we later do not need to reason about zero-cost jobs).
    Hypothesis H_job_cost_positive: job_cost_positive j.

Assume the busy interval of the job j is given by [t1,t2).
    Variable t1 t2 : duration.
    Hypothesis H_busy_interval :
      definitions.busy_interval sched j t1 t2.

Consider any sub-interval [t1, t1 + Δ) of the busy interval of j.
    Variable Δ : duration.
    Hypothesis H_Δ_in_busy : t1 + Δ < t2.

We prove that the cumulative priority inversion in the interval [t1, t1 + Δ) is indeed 0.
    Lemma no_priority_inversion:
      cumulative_priority_inversion arr_seq sched j t1 (t1 + Δ) = 0.
    Proof.
      apply /eqP; rewrite -leqn0.
      pose zf : nat → nat := (fun⇒ 0).
      have: cumulative_priority_inversion arr_seq sched j t1 (t1 + Δ) ≤ zf (job_arrival j - t1);
        last by apply.
      apply: cumulative_priority_inversion_is_bounded; rt_eauto.
      have → : priority_inversion_is_bounded_by arr_seq sched tsk zf
               = priority_inversion_is_bounded_by_constant arr_seq sched tsk 0;
        by try apply: FIFO_implies_no_pi; rt_eauto.
    Qed.

  End AbsenceOfPriorityInversion.

Higher- and Equal-Priority Interference

Next, we establish a bound on the interference produced by higher- and equal-priority jobs.
  Section BoundOnHEPWorkload.

Consider again a job j of the task under analysis tsk with a positive cost.
    Variable j : Job.
    Hypothesis H_job_of_task : job_of_task tsk j.
    Hypothesis H_j_in_arrivals : arrives_in arr_seq j.
    Hypothesis H_job_cost_positive : job_cost_positive j.

Consider the (abstract) busy window of j and denote it as [t1, t2).
    Variable t1 t2 : instant.
    Hypothesis H_busy_window :
      definitions.busy_interval sched j t1 t2.

Consider any arbitrary sub-interval [t1, Δ) within the busy window of j.
    Variable Δ : instant.
    Hypothesis H_in_busy : t1 + Δ < t2.

The cumulative interference from higher- and equal-priority jobs during [t1, Δ) is bounded as follows.
    Lemma bound_on_hep_workload :
      cumulative_another_hep_job_interference arr_seq sched j t1 (t1 + Δ) ≤
        \sum_(tsko <- ts) task_request_bound_function tsko (job_arrival j - t1 + ε) - task_cost tsk.
    Proof.
      move: H_busy_window ⇒ [ [ /andP [LE GT] [QUIETt1 _ ] ] [QUIETt2 EQNs]].
      rewrite (cumulative_i_ohep_eq_service_of_ohep arr_seq); rt_eauto;
        last by rewrite instantiated_quiet_time_equivalent_quiet_time; rt_eauto.
      eapply leq_trans; first by apply service_of_jobs_le_workload; rt_eauto.
      rewrite (leqRW (workload_equal_subset _ _ _ _ _ _ _)); rt_eauto.
      rewrite (workload_minus_job_cost j); rt_eauto;
        last by rewrite /ε; apply job_in_arrivals_between; rt_auto; lia.
      rewrite /workload_of_jobs /IBF (big_rem tsk) //=
        (addnC (task_request_bound_function tsk (job_arrival j - t1 + ε))).
      rewrite -addnBA; last first.
      - apply leq_trans with (task_request_bound_function tsk ε);
          first by apply : (task_rbf_1_ge_task_cost arr_seq); rt_eauto.
        by apply task_rbf_monotone; [apply H_valid_arrival_curve | lia].
      - eapply leq_trans;
          last by erewrite leq_add2l; eapply task_rbf_excl_tsk_bounds_task_workload_excl_j; eauto 1.
        rewrite addnBA.
        + rewrite leq_sub2r //; eapply leq_trans.
          × apply sum_over_partitions_le ⇒ j' inJOBS.
            by apply H_all_jobs_from_taskset, (in_arrivals_implies_arrived _ _ _ _ inJOBS).
          × rewrite (big_rem tsk) //= addnC leq_add //; last by rewrite subnKC.
            rewrite big_seq_cond [in X in _ ≤ X]big_seq_cond big_mkcond [in X in _ ≤ X]big_mkcond //=.
            apply leq_sum ⇒ tsk' _; rewrite andbC //=.
            destruct (tsk' \in rem (T:=Task) tsk ts) eqn:IN; last by [].
            apply rem_in in IN.
            eapply leq_trans;
              last by apply (task_workload_le_task_rbf _ _ _ IN H_valid_job_cost H_is_arrival_curve t1).
            by rewrite addnBAC //= subnKC //= addn1; apply leqW.
        + move : H_job_of_task ⇒ TSKj.
          rewrite /task_workload_between /task_workload /workload_of_jobs (big_rem j) //=;
            first by rewrite TSKj; apply leq_addr.
          apply job_in_arrivals_between ⇒ //; rt_eauto.
          by apply /andP; split; [| rewrite subnKC; [rewrite addn1 |]].
    Qed.

  End BoundOnHEPWorkload.

Correctness of IBF

Combining the bound on interference due to lower-priority jobs (priority inversion, i.e., no_priority_inversion) and the interference due to higher- or equal-priority jobs (bound_on_hep_workload), we can prove that IBF indeed bounds the total interference.
Notice that, unlike the aRTA framework published at ECRTS'20, generalized aRTA supports custom parameters for IBF. Hence, we have to specify that the first argument of IBF represents the relative arrival time of a job under analysis by passing a proposition relative_arrival_time_of_job_is_A.
  Lemma IBF_correct :
    job_interference_is_bounded_by arr_seq sched tsk IBF (relative_arrival_time_of_job_is_A sched).
  Proof.
    move ⇒ t1 t2 Δ j ARRj TSKj BUSY IN_BUSY NCOMPL A Pred.
    rewrite cumulative_interference_split; rt_eauto.
    have JPOS: job_cost_positive j by rewrite -ltnNge in NCOMPL; unfold job_cost_positive; lia.
    move: (BUSY) ⇒ [ [ /andP [LE GT] [QUIETt1 _ ] ] [QUIETt2 EQNs]].
    rewrite (no_priority_inversion j ARRj _ JPOS _ t2) //= add0n.
    have ->: A = job_arrival j - t1 by erewrite Pred with (t1 := t1); [lia | apply BUSY].
    by rewrite /IBF; eapply bound_on_hep_workload; rt_eauto.
  Qed.

The preceding lemma IBF_correct corresponds to Lemma 3 in the paper. To see the correspondence more clearly, refer to the definition of job_interference_is_bounded_by in the above lemma.

F. Defining the Search Space

In this section, we define the concrete search space for FIFO and relate it to the abstract search space of aRTA. In the case of FIFO, the concrete search space is the set of offsets less than L such that there exists a task tsk' in ts such that r bf tsk' (A) ≠ rbf tsk' (A + ε).
  Definition is_in_concrete_search_space (A : duration) :=
    (A < L) && has (fun tsk' ⇒ task_request_bound_function tsk' (A) !=
                               task_request_bound_function tsk' ( A + ε )) ts.

To enable the use of aRTA, we must now show that any offset A included in the abstract search space is also included in the concrete search space. That is, we must show that the concrete search space is a refinement of the abstract search space assumed by aRTA.
To this end, first recall the notion of the abstract search space in aRTA.
  Let is_in_abstract_search_space A := abstract.search_space.is_in_search_space tsk L IBF A.

  Section SearchSpaceRefinement.

Suppose we are given a job j of the task under analysis tsk with positive cost. We use the existence of such a job in the subsequent proof, even if it does not feature in the claim directly.
    Variable j : Job.
    Hypothesis H_j_arrives : arrives_in arr_seq j.
    Hypothesis H_job_of_tsk : job_of_task tsk j.
    Hypothesis H_positive_cost : 0 < task_cost tsk.

Under this assumption, given any A from the abstract search space, ...
    Variable A : nat.
    Hypothesis H_in_abstract : is_in_abstract_search_space A.

... we prove that A is also in the concrete search space. In other words, we prove that the abstract search space is a subset of the concrete search space.
    Lemma search_space_refinement : is_in_concrete_search_space A.
    Proof.
      move: H_in_abstract ⇒ [INSP | [/andP [POSA LTL] [x [LTx INSP2]]]].
      { subst A.
        apply/andP; split; [by done |].
        apply /hasP. ∃ tsk; first by done.
        rewrite neq_ltn;apply/orP; left.
        erewrite task_rbf_0_zero; eauto 2.
        rewrite add0n ; apply leq_trans with (task_cost tsk).
        - by eapply leq_trans; eauto 2.
        - by eapply task_rbf_1_ge_task_cost; rt_eauto. }
      { apply /andP; split; first by done.
        apply /hasPn.
        move ⇒ EQ2. unfold IBF in INSP2.
        rewrite subnK in INSP2; try by done.
        apply INSP2; clear INSP2.
        have → : \sum_(tsko <- ts) task_request_bound_function tsko A =
                   \sum_(tsko <- ts) task_request_bound_function tsko (A + ε); last by done.
        apply eq_big_seq ⇒ //= task IN.
        by move: (EQ2 task IN) ⇒ /negPn /eqP. }
    Qed.

The preceding lemma search_space_refinement corresponds to Lemma 4 in the paper, which is apparent after consulting the definitions of the abstract and concrete search spaces.

  End SearchSpaceRefinement.

G. Stating the Response-Time Bound R

Having established all necessary preliminaries, it is finally time to state the claimed response-time bound R.
  Variable R : duration.
  Hypothesis H_R_max:
    ∀ (A : duration),
      is_in_concrete_search_space A →
      ∃ (F : nat),
        A + F ≥ \sum_(tsko <- ts) task_request_bound_function tsko (A + ε) ∧
          F ≤ R.

Ultimately, we seek to apply aRTA to prove the correctness of this R. However, in order to connect the concrete definition of R with aRTA, we must first restate the bound in the shape of the abstract response-time bound equation that aRTA expects, which we do next.

  Section ResponseTimeBoundRestated.
Suppose again we are given a job j of the task under analysis tsk with positive cost. We use the existence of such a job in the subsequent proof, even if it does not feature in the claim directly.
    Variable j : Job.
    Hypothesis H_j_arrives : arrives_in arr_seq j.
    Hypothesis H_job_of_tsk : job_of_task tsk j.
    Hypothesis H_job_cost_positive : job_cost_positive j.

We know that:
  • if A is in the abstract search space, then it is also in the concrete search space; and
  • if A is in the concrete search space, then there exists a solution that satisfies the inequalities stated in H_R_is_maximum.
Using these facts, we prove that, if A is in the abstract search space, then there also exists a solution F to the response-time equation as expected by aRTA.
    Lemma soln_abstract_response_time_recurrence :
      ∀ A,
        is_in_abstract_search_space A →
        ∃ (F : nat),
          A + F ≥ task_rtct tsk + IBF tsk A (A + F) ∧
            F + (task_cost tsk - task_rtct tsk) ≤ R.
    Proof.
      move ⇒ A IN.
      eapply search_space_refinement in IN; rt_eauto.
      move: (H_R_max _ IN) ⇒ [F [FIX NEQ]].
      have R_GE_TC: task_cost tsk ≤ R.
      { move : (H_R_max 0) ⇒ SEARCH; feed SEARCH;
          first by eapply search_space_refinement; rt_eauto; left.
        move: SEARCH ⇒ [F' [LE1 LE2]].
        rewrite !add0n in LE1.
        rewrite -(leqRW LE2) -(leqRW LE1).
        by eapply (task_cost_le_sum_rbf arr_seq); rt_eauto. }
      ∃ (R - (task_cost tsk - task_rtct tsk)); split.
      - rewrite /IBF.
        rewrite (leqRW FIX) addnC -subnA; first last.
        + rewrite -(leqRW FIX).
          apply : (task_cost_le_sum_rbf _ _ _ ); rt_eauto.
          by rewrite addn1.
        + by move : H_valid_run_to_completion_threshold ⇒ [TASKvalid JOBvalid].
        + rewrite addnBA; first by rewrite leq_sub2r // leq_add2l.
          by apply leq_trans with (task_cost tsk); [lia|].
      - rewrite subnK; first by done.
        by apply leq_trans with (task_cost tsk); [lia| ].
    Qed.

Lemma soln_abstract_response_time_recurrence is shown in Listing 3 in the paper.
  End ResponseTimeBoundRestated.

H. Soundness of the Response-Time Bound

Finally, we are in a position to establish the soundness of the claimed response-time bound R by applying the main general aRTA theorem uniprocessor_response_time_bound.
  Theorem uniprocessor_response_time_bound_FIFO:
    task_response_time_bound arr_seq sched tsk R.
  Proof.
    move ⇒ js ARRs TSKs.
    rewrite /job_response_time_bound /completed_by.
    case: (posnP (@job_cost _ _ js)) ⇒ [ → |POS]; first by done.
    eapply uniprocessor_response_time_bound_ideal; rt_eauto.
    - exact: abstractly_work_conserving.
    - exact: busy_windows_are_bounded.
    - exact: IBF_correct.
    - by apply: soln_abstract_response_time_recurrence; eauto.
  Qed.

The preceding theorem uniprocessor_response_time_bound_FIFO corresponds to Theorem 2 in the paper. The correspondence becomes clearer when referring to the definition of task_response_time_bound, and then in turn to the definitions of job_of_task and job_response_time_bound.

End AbstractRTAforFIFOwithArrivalCurves.