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.

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, ...

... 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.

The Job Arrival Sequence

Consider any arrival sequence arr_seq with consistent, non-duplicate arrivals.

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.

We further require that a job's cost cannot exceed its task's stated WCET.

The Task Set

We consider an arbitrary task set ts...

... and assume that all jobs stem from tasks in this task set.

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 ...

... (2) a monotonic function that equals 0 for the empty interval delta = 0.

The Task Under Analysis

Let tsk be any task in ts that is to be analyzed.

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.

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).

We assume that the schedule complies with the preemption model ...

... and, last but not least, that it respects the FIFO scheduling policy.

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 ideal_jlfp_rta).

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.

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

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.

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 ...

... that is a fixed point of the following equation.

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.

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.

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.

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.

Consider any job j of the task under analysis tsk.

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

Assume the busy interval of the job j is given by [t1,t2).

Consider any sub-interval [t1, t1 + Δ) of the busy interval of j.

We prove that the cumulative priority inversion in the interval [t1, t1 + Δ) is indeed 0.

Higher- and Equal-Priority Interference

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

Consider again a job j of the task under analysis tsk with a positive cost.

Consider the (abstract) busy window of j and denote it as [t1, t2).

Consider any arbitrary sub-interval [t1, Δ) within the busy window of j.

The cumulative interference from higher- and equal-priority jobs during [t1, Δ) is bounded as follows.

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.

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).

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.

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.

Under this assumption, given any A from the abstract search space, ...

... 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.

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.

G. Stating the Response-Time Bound R

Having established all necessary preliminaries, it is finally time to state the claimed response-time bound 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.

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.

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 is shown in Listing 3 in the paper.

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.

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.