Library prosa.classic.analysis.global.parallel.workload_bound

Require Import prosa.classic.util.all.
Require Import prosa.classic.model.arrival.basic.task prosa.classic.model.arrival.basic.job prosa.classic.model.arrival.basic.task_arrival.
Require Import prosa.classic.model.schedule.global.response_time prosa.classic.model.schedule.global.workload
               prosa.classic.model.schedule.global.schedulability.
Require Import prosa.classic.model.schedule.global.basic.schedule.
From mathcomp Require Import ssreflect ssrbool eqtype ssrnat seq div fintype bigop path.

Module WorkloadBound.

  Import Job SporadicTaskset Schedule ScheduleOfSporadicTask TaskArrival ResponseTime Schedulability Workload.

  Section WorkloadBoundDef.

    Context {sporadic_task: eqType}.
    Variable task_cost: sporadic_task time.
    Variable task_period: sporadic_task time.

    Variable tsk: sporadic_task.
    Variable R_tsk: time. (* Known response-time bound for the task *)
    Variable delta: time. (* Length of the interval *)

    (* Bound on the number of jobs that execute on the interval *)
    Definition max_jobs :=
      div_ceil (delta + R_tsk) (task_period tsk).

    (* Simplified workload bound for parallel jobs.*)
    Definition W := max_jobs × task_cost tsk.

  End WorkloadBoundDef.

  Section BasicLemmas.

    Context {sporadic_task: eqType}.
    Variable task_cost: sporadic_task time.
    Variable task_period: sporadic_task time.

    (* Let tsk be any task...*)
    Variable tsk: sporadic_task.

    (* ...with period > 0. *)
    Hypothesis H_period_positive: task_period tsk > 0.

    (* Let R1 <= R2 be two response-time bounds that
       are larger than the cost of the tsk. *)

    Variable R1 R2: time.
    Hypothesis H_R_lower_bound: R1 task_cost tsk.
    Hypothesis H_R1_le_R2: R1 R2.

    Let workload_bound := W task_cost task_period tsk.

    (* The workload bound W is monotonically increasing. *)
    Lemma W_monotonic :
       t1 t2,
        t1 t2
        workload_bound R1 t1 workload_bound R2 t2.
    Proof.
      intros t1 t2 LEt.
      unfold workload_bound, W, max_jobs.
      rewrite leq_mul2r; apply/orP; right.
      by apply leq_divceil2r; last by apply leq_add.
    Qed.

  End BasicLemmas.

  Section ProofWorkloadBound.

    Context {sporadic_task: eqType}.
    Variable task_cost: sporadic_task time.
    Variable task_period: sporadic_task time.
    Variable task_deadline: sporadic_task time.

    Context {Job: eqType}.
    Variable job_arrival: Job time.
    Variable job_cost: Job time.
    Variable job_task: Job sporadic_task.
    Variable job_deadline: Job time.

    Variable arr_seq: arrival_sequence Job.

    (* Assume that all jobs have valid parameters *)
    Hypothesis H_jobs_have_valid_parameters :
       j,
        arrives_in arr_seq j
        valid_sporadic_job task_cost task_deadline job_cost job_deadline job_task j.

    (* Consider any schedule. *)
    Context {num_cpus: nat}.
    Variable sched: schedule Job num_cpus.
    Hypothesis H_jobs_come_from_arrival_sequence:
      jobs_come_from_arrival_sequence sched arr_seq.

    (* Assumption: jobs only execute if they arrived.
       This is used to eliminate jobs that arrive after end of the interval t1 + delta. *)

    Hypothesis H_jobs_must_arrive_to_execute:
      jobs_must_arrive_to_execute job_arrival sched.

    (* Assumption: jobs do not execute after they completed.
       This is used to eliminate jobs that complete before the start of the interval t1. *)

    Hypothesis H_completed_jobs_dont_execute:
      completed_jobs_dont_execute job_cost sched.

    (* Assumption: sporadic task model.
       This is necessary to conclude that consecutive jobs ordered by arrival times
       are separated by at least 'period' times units. *)

    Hypothesis H_sporadic_tasks:
      sporadic_task_model task_period job_arrival job_task arr_seq.

    (* Before starting the proof, let's give simpler names to the definitions. *)
    Let job_has_completed_by := completed job_cost sched.

    Let workload_of (tsk: sporadic_task) (t1 t2: time) :=
      workload job_task sched tsk t1 t2.

    (* Now we define the theorem. Let tsk be any task in the taskset. *)
    Variable tsk: sporadic_task.

    (* Assumption: the task must have valid parameters:
         a) period > 0 (used in divisions)
         b) deadline of the jobs = deadline of the task
         c) cost <= period
            (used to prove that the distance between the first and last
             jobs is at least (cost + n*period), where n is the number
             of middle jobs. If cost >> period, the claim does not hold
             for every task set. *)

    Hypothesis H_valid_task_parameters:
      is_valid_sporadic_task task_cost task_period task_deadline tsk.

    (* Consider an interval t1, t1 + delta). *)
    Variable t1 delta: time.

    (* Assume that any job with response-time bound R_tsk before
       t1 + delta indeed completes within the response-time bound. *)

    Variable R_tsk: time.

    Hypothesis H_response_time_bound :
       j,
        arrives_in arr_seq j
        job_task j = tsk
        job_arrival j + R_tsk < t1 + delta
        job_has_completed_by j (job_arrival j + R_tsk).

    Section MainProof.

      (* In this section, we prove that the workload of a task in the
         interval t1, t1 + delta) is bounded by W. *)


      (* Let's simplify the names a bit. *)
      Let t2 := t1 + delta.
      Let n_k := max_jobs task_period tsk R_tsk delta.
      Let workload_bound := W task_cost task_period tsk R_tsk delta.

      (* Since we only care about the workload of tsk, we restrict
         our view to the set of jobs of tsk scheduled in t1, t2). *)

      Let scheduled_jobs :=
        jobs_of_task_scheduled_between job_task sched tsk t1 t2.

      (* Now, let's consider the list of interfering jobs sorted by arrival time. *)
      Let earlier_arrival := fun x yjob_arrival x job_arrival y.
      Let sorted_jobs := (sort earlier_arrival scheduled_jobs).

      (* The first step consists in simplifying the sum corresponding
         to the workload. *)

      Section SimplifyJobSequence.

        (* After switching to the definition of workload based on a list
           of jobs, we show that sorting the list preserves the sum. *)

        Lemma workload_bound_simpl_by_sorting_scheduled_jobs :
          workload_joblist job_task sched tsk t1 t2 =
           \sum_(i <- sorted_jobs) service_during sched i t1 t2.
        Proof.
          unfold workload_joblist; fold scheduled_jobs.
          rewrite (perm_big sorted_jobs) /= //.
          by rewrite -(perm_sort earlier_arrival).
        Qed.

        (* Remember that both sequences have the same set of elements *)
        Lemma workload_bound_job_in_same_sequence :
           j,
            (j \in scheduled_jobs) = (j \in sorted_jobs).
        Proof.
          by apply perm_mem; rewrite -(perm_sort earlier_arrival).
        Qed.

        (* Remember that all jobs in the sorted sequence is an
           interfering job of task tsk. *)

        Lemma workload_bound_all_jobs_from_tsk :
           j_i,
            j_i \in sorted_jobs
            arrives_in arr_seq j_i
            job_task j_i = tsk
            service_during sched j_i t1 t2 != 0
            j_i \in jobs_scheduled_between sched t1 t2.
        Proof.
          rename H_jobs_come_from_arrival_sequence into FROMarr.
          intros j_i LTi.
          rewrite -workload_bound_job_in_same_sequence mem_filter in LTi; des.
          unfold jobs_scheduled_between in *; rewrite mem_undup in LTi0.
          apply mem_bigcat_nat_exists in LTi0; des.
          rewrite mem_scheduled_jobs_eq_scheduled in LTi0.
          repeat split; try (by done); first by apply (FROMarr j_i i).
          {
            apply service_implies_cumulative_service with (t := i);
              first by apply/andP; split.
            by rewrite -not_scheduled_no_service negbK.
          }
          {
            rewrite mem_undup.
            apply mem_bigcat_nat with (j := i); first by auto.
            by rewrite mem_scheduled_jobs_eq_scheduled.
          }
        Qed.

        (* Remember that consecutive jobs are ordered by arrival. *)
        Lemma workload_bound_jobs_ordered_by_arrival :
           i elem,
            i < (size sorted_jobs).-1
            earlier_arrival (nth elem sorted_jobs i) (nth elem sorted_jobs i.+1).
        Proof.
          intros i elem LT.
          assert (SORT: sorted earlier_arrival sorted_jobs).
            by apply sort_sorted; unfold total, earlier_arrival; ins; apply leq_total.
          by destruct sorted_jobs; simpl in *; [by rewrite ltn0 in LT | by apply/pathP].
        Qed.

      End SimplifyJobSequence.

      (* Next, we show that if the number of jobs is no larger than n_k,
         the workload bound trivially holds. *)

      Section WorkloadNotManyJobs.

        Lemma workload_bound_holds_for_at_most_n_k_jobs :
          size sorted_jobs n_k
          \sum_(i <- sorted_jobs) service_during sched i t1 t2
            workload_bound.
        Proof.
          unfold workload_bound, W; intros LEnk.
          apply leq_trans with (n := \sum_(x <- sorted_jobs) task_cost tsk);
            last by rewrite big_const_seq iter_addn addn0 mulnC leq_mul2r; apply/orP; right.
          rewrite [\sum_(_ <- _) service_during _ _ _ _]big_seq_cond.
          rewrite [\sum_(_ <- _) task_cost _]big_seq_cond.
          apply leq_sum; intros j_i; move/andP ⇒ [INi _].
          apply workload_bound_all_jobs_from_tsk in INi; des.
          eapply cumulative_service_le_task_cost;
            [by apply H_completed_jobs_dont_execute | by apply INi0 |].
          by apply H_jobs_have_valid_parameters.
        Qed.

      End WorkloadNotManyJobs.

      (* Otherwise, assume that the number of jobs is larger than n_k >= 0.
         First, consider the simple case with only one job. *)

      Section WorkloadSingleJob.

        (* Assume that there's at least one job in the sorted list. *)
        Hypothesis H_at_least_one_job: size sorted_jobs > 0.

        Variable elem: Job.
        Let j_fst := nth elem sorted_jobs 0.

        (* The first job is an interfering job of task tsk. *)
        Lemma workload_bound_j_fst_is_job_of_tsk :
          arrives_in arr_seq j_fst
          job_task j_fst = tsk
          service_during sched j_fst t1 t2 != 0
          j_fst \in jobs_scheduled_between sched t1 t2.
        Proof.
          by apply workload_bound_all_jobs_from_tsk, mem_nth.
        Qed.

        (* The workload bound holds for the single job. *)
        Lemma workload_bound_holds_for_a_single_job :
          \sum_(0 i < 1) service_during sched (nth elem sorted_jobs i) t1 t2
          workload_bound.
        Proof.
          rename H_valid_task_parameters into PARAMS.
          unfold is_valid_sporadic_task in ×.
          unfold workload_bound, W; fold n_k.
          have INfst := workload_bound_j_fst_is_job_of_tsk; des.
          rewrite big_nat_recr // big_geq // [nth]lock /= -lock add0n.
          apply leq_trans with (n := task_cost tsk).
          {
            eapply cumulative_service_le_task_cost; [by eauto 1 | by apply INfst0 |].
            by apply H_jobs_have_valid_parameters.
          }
          rewrite -{1}[task_cost tsk]mul1n leq_mul2r; apply/orP; right.
          apply ceil_neq0; last by apply PARAMS0.
          {
            apply leq_trans with (n := delta); last by apply leq_addr.
            rewrite lt0n; apply/eqP; intro EQ0.
            move: INfst1 ⇒ /eqP BUG; apply BUG.
            unfold t2; rewrite EQ0 addn0.
            by unfold service_during; rewrite big_geq.
          }
        Qed.

      End WorkloadSingleJob.

      (* Next, consider the last case where there are at least two jobs:
         the first job j_fst, and the last job j_lst. *)

      Section WorkloadTwoOrMoreJobs.

        (* There are at least two jobs. *)
        Variable num_mid_jobs: nat.
        Hypothesis H_at_least_two_jobs : size sorted_jobs = num_mid_jobs.+2.

        Variable elem: Job.
        Let j_fst := nth elem sorted_jobs 0.
        Let j_lst := nth elem sorted_jobs num_mid_jobs.+1.

        (* The last job is an interfering job of task tsk. *)
        Lemma workload_bound_j_lst_is_job_of_tsk :
          arrives_in arr_seq j_lst
          job_task j_lst = tsk
          service_during sched j_lst t1 t2 != 0
          j_lst \in jobs_scheduled_between sched t1 t2.
        Proof.
          apply workload_bound_all_jobs_from_tsk, mem_nth.
          by rewrite H_at_least_two_jobs.
        Qed.

        (* The response time of the first job must fall inside the interval. *)
        Lemma workload_bound_response_time_of_first_job_inside_interval :
          t1 job_arrival j_fst + R_tsk.
        Proof.
          rewrite leqNgt; apply /negP; unfold not; intro LTt1.
          exploit workload_bound_all_jobs_from_tsk.
          {
            apply mem_nth; instantiate (1 := 0).
            apply ltn_trans with (n := 1); [by done | by rewrite H_at_least_two_jobs].
          }
          instantiate (1 := elem); move ⇒ [FSTarr [FSTtsk [/eqP FSTserv FSTin]]].
          apply FSTserv.
          apply (cumulative_service_after_job_rt_zero job_arrival job_cost) with (R := R_tsk);
            try (by done); last by apply ltnW.
          apply H_response_time_bound; try (by done).
          by apply leq_trans with (n := t1); last by apply leq_addr.
        Qed.

        (* The arrival of the last job must also fall inside the interval. *)
        Lemma workload_bound_last_job_arrives_before_end_of_interval :
          job_arrival j_lst < t2.
        Proof.
          rewrite leqNgt; apply/negP; unfold not; intro LT2.
          exploit workload_bound_all_jobs_from_tsk.
          {
            apply mem_nth; instantiate (1 := num_mid_jobs.+1).
            by rewrite -(ltn_add2r 1) addn1 H_at_least_two_jobs addn1.
          }
          instantiate (1 := elem); move ⇒ [LSTarr [LSTtsk [/eqP LSTserv LSTin]]].
          unfold service_during; apply LSTserv.
          by eapply cumulative_service_before_job_arrival_zero; eauto 1.
        Qed.

        (* Bound the service of the middle jobs. *)
        Lemma workload_bound_service_of_middle_jobs :
          \sum_(0 i < num_mid_jobs)
            service_during sched (nth elem sorted_jobs i.+1) t1 t2
            num_mid_jobs × task_cost tsk.
        Proof.
          apply leq_trans with (n := num_mid_jobs × task_cost tsk);
            last by rewrite leq_mul2l; apply/orP; right.
          apply leq_trans with (n := \sum_(0 i < num_mid_jobs) task_cost tsk);
            last by rewrite big_const_nat iter_addn addn0 mulnC subn0.
          rewrite big_nat_cond [\sum_(0 i < num_mid_jobs) task_cost _]big_nat_cond.
          apply leq_sum; intros i; rewrite andbT; move ⇒ /andP LT; des.
          exploit workload_bound_all_jobs_from_tsk.
          {
            instantiate (1 := nth elem sorted_jobs i.+1).
            apply mem_nth; rewrite H_at_least_two_jobs.
            by rewrite ltnS; apply leq_trans with (n := num_mid_jobs).
          }
          move ⇒ [ARR [TSK _]].
          eapply cumulative_service_le_task_cost; eauto 1.
          by apply H_jobs_have_valid_parameters.
        Qed.

        (* Conclude that the distance between first and last is at least num_mid_jobs + 1 periods. *)
        Lemma workload_bound_many_periods_in_between :
          job_arrival j_lst - job_arrival j_fst num_mid_jobs.+1 × (task_period tsk).
        Proof.
          assert (EQnk: num_mid_jobs.+1=(size sorted_jobs).-1).
            by rewrite H_at_least_two_jobs.
          unfold j_fst, j_lst; rewrite EQnk telescoping_sum;
            last by ins; apply workload_bound_jobs_ordered_by_arrival.
          rewrite -[_ × _ tsk]addn0 mulnC -iter_addn -{1}[_.-1]subn0 -big_const_nat.
          rewrite big_nat_cond [\sum_(0 i < _)(_-_)]big_nat_cond.
          apply leq_sum; intros i; rewrite andbT; move ⇒ /andP LT; des.

          (* To simplify, call the jobs 'cur' and 'next' *)
          set cur := nth elem sorted_jobs i.
          set next := nth elem sorted_jobs i.+1.

          (* Show that cur arrives earlier than next *)
          assert (ARRle: job_arrival cur job_arrival next).
            by unfold cur, next; apply workload_bound_jobs_ordered_by_arrival.

          feed (workload_bound_all_jobs_from_tsk cur).
            by apply mem_nth, (ltn_trans LT0); destruct sorted_jobs.
          intros [CURarr [CURtsk [_ CURin]]].

          feed (workload_bound_all_jobs_from_tsk next).
            by apply mem_nth; destruct sorted_jobs.
          intros [NEXTarr [NEXTtsk [_ NEXTin]]].

          (* Use the sporadic task model to conclude that cur and next are separated
             by at least (task_period tsk) units. Of course this only holds if cur != next.
             Since we don't know much about the list (except that it's sorted), we must
             also prove that it doesn't contain duplicates. *)

          assert (CUR_LE_NEXT: job_arrival cur + task_period (job_task cur) job_arrival next).
          {
            apply H_sporadic_tasks; try (by done).
            unfold cur, next, not; intro EQ; move: EQ ⇒ /eqP EQ.
            rewrite nth_uniq in EQ; first by move: EQ ⇒ /eqP EQ; lia.
              by apply ltn_trans with (n := (size sorted_jobs).-1); destruct sorted_jobs; ins.
              by destruct sorted_jobs; ins.
              by rewrite sort_uniq -/scheduled_jobs filter_uniq // undup_uniq.
              by rewrite CURtsk.
          }
          by rewrite leq_subRL_impl // -CURtsk.
        Qed.

        (* Next, we prove that n_k covers every scheduled job, ... *)
        Lemma workload_bound_n_k_covers_all_jobs :
          n_k num_mid_jobs.+2.
        Proof.
          rename H_valid_task_parameters into PARAMS.
          unfold is_valid_sporadic_task in *; des.
          rewrite leqNgt; apply/negP; unfold not; intro LTnk.
          assert (DISTmax: job_arrival j_lst - job_arrival j_fst delta + R_tsk).
          {
            apply leq_trans with (n := num_mid_jobs.+1 × task_period tsk);
              last by apply workload_bound_many_periods_in_between.
            apply leq_trans with (n := n_k × task_period tsk);
              last by rewrite leq_mul2r; apply/orP; right.
            unfold n_k, max_jobs, div_ceil.
            destruct (task_period tsk %| delta + R_tsk) eqn:DIV.
            {
              rewrite dvdn_eq in DIV; move: DIV ⇒ /eqP DIV.
              by rewrite DIV.
            }
            by apply ltnW, ltn_ceil, PARAMS0.
          }
          rewrite <- leq_add2r with (p := job_arrival j_fst) in DISTmax.
          rewrite addnC addnBAC in DISTmax; last first.
          {
            unfold j_fst, j_lst; rewrite -[_.+1]add0n.
            apply prev_le_next; last by rewrite H_at_least_two_jobs add0n leqnn.
            by ins; apply workload_bound_jobs_ordered_by_arrival.
          }
          rewrite -[job_arrival j_lst + _ - _]subnBA // subnn subn0 in DISTmax.
          rewrite [delta + _]addnC addnA in DISTmax.
          apply leq_trans with (m := t1 + delta) in DISTmax; last first.
          {
            rewrite leq_add2r.
            by apply workload_bound_response_time_of_first_job_inside_interval.
          }
          have LST := workload_bound_last_job_arrives_before_end_of_interval.
          by apply leq_ltn_trans with (m := t2) in LST; first by rewrite ltnn in LST.
        Qed.

        (* ... so the workload bound trivially holds. *)
        Lemma workload_bound_holds :
            \sum_(0 i < num_mid_jobs.+2)
             service_during sched (nth elem sorted_jobs i) t1 t2
           workload_bound.
        Proof.
          unfold workload_bound, W; fold n_k.
          have ALL := workload_bound_n_k_covers_all_jobs.
          apply leq_trans with (n := num_mid_jobs.+2 × task_cost tsk);
            last by rewrite leq_mul2r; apply/orP; right.
          apply leq_trans with (n := \sum_(0 i < num_mid_jobs.+2) task_cost tsk);
            last by rewrite big_const_nat iter_addn addn0 subn0 mulnC.
          rewrite big_nat_cond [\sum_(_ _ < _ | true)_]big_nat_cond.
          apply leq_sum; intro i; move ⇒ /andP [/andP [_ LEi] _].
          exploit workload_bound_all_jobs_from_tsk.
          {
            instantiate (1 := nth elem sorted_jobs i).
            by apply mem_nth; rewrite H_at_least_two_jobs.
          }
          move ⇒ [ARR [TSK _]].
          eapply cumulative_service_le_task_cost; eauto 1.
          by apply H_jobs_have_valid_parameters.
        Qed.

      End WorkloadTwoOrMoreJobs.

      (* Using the lemmas above, we prove the main theorem about the workload bound. *)
      Theorem workload_bounded_by_W :
        workload_of tsk t1 (t1 + delta) workload_bound.
      Proof.
        unfold workload_of, workload_bound, W in *; ins; des.
        fold n_k.

        (* Use the definition of workload based on list of jobs. *)
        rewrite workload_eq_workload_joblist.

        (* Now we order the list by job arrival time. *)
        rewrite workload_bound_simpl_by_sorting_scheduled_jobs.

        (* Next, we show that the workload bound holds if n_k
           is no larger than the number of interferings jobs. *)

        destruct (size sorted_jobs n_k) eqn:NUM;
          first by apply workload_bound_holds_for_at_most_n_k_jobs.
        apply negbT in NUM; rewrite -ltnNge in NUM.

        (* Find some dummy element to use in the nth function *)
        assert (EX: elem: Job, True).
          destruct sorted_jobs; [ by rewrite ltn0 in NUM | by s].
        destruct EX as [elem _].

        (* Now we index the sum to access the first and last elements. *)
        rewrite (big_nth elem).

        (* First, we show that the bound holds for an empty list of jobs. *)
        destruct (size sorted_jobs) as [| n] eqn:SIZE;
          first by rewrite big_geq.

        (* Then, we show the same for a singleton set of jobs. *)
        destruct n as [| num_mid_jobs];
          first by apply workload_bound_holds_for_a_single_job; rewrite SIZE.
        (* Since num_mid_jobs + 2 <= n_k >= num_mid_jobs + 2, the proof follows easily. *)
        by apply workload_bound_holds.
      Qed.

    End MainProof.

  End ProofWorkloadBound.

End WorkloadBound.