render_instruction_count_tables

Function render_instruction_count_tables 

Source
pub fn render_instruction_count_tables(
    records: &[VerifierCellRecord],
) -> Option<String>
Expand description

Render one verified_insns table per declared scheduler. Within each scheduler’s section: rows = BPF program, columns = kernel version, and each cell is that program’s verified_insns for the (scheduler, kernel) summarized ACROSS the topologies that ran it — a single number when topology-invariant, lo..hi when it varies (- when that program reported no stats on that kernel).

verified_insns is the verifier’s PROCESSED-instruction count (env->insn_processed) — fixed per load, but NOT topology-invariant (a scheduler whose verification path depends on topology-derived .rodata, e.g. nr_cpus, processes a different count per topology). So topology is folded into the cell as a range rather than shown as its own (usually all-identical) axis; the axes it genuinely varies on — BPF program (y) and kernel version (x) — are the table axes, sectioned per declared scheduler. Program is the ROW axis because a scheduler declares many programs (scx_lavd has ~26) and the count is unbounded, so rows scale down the page without wrapping; kernel is the COLUMN axis because a sweep runs only a handful of kernels, so the columns stay narrow enough to read in a CI log. Identical-binary declarations are sectioned separately on purpose (they are run separately). Returns None when no record carries any per-program stats (the caller prints nothing).

A kernel that ran a scheduler but produced NO stats at all — e.g. every cell on it died during BPF load, so no program existed to introspect — still gets a column (all cells -) as long as the scheduler has at least one program row from some OTHER kernel. Without it the stats-less kernel would silently vanish from the table, hiding that it ran and failed to load; the explicit all-- column makes that absence visible. Kernel columns therefore come from every kernel that ran the scheduler, not only those that produced stats.

Schedulers, kernels, and programs are BTree-sorted so the same run renders the same output (shell-pipeline stable). The range drops which topology produced which count; a per-topology breakdown is a separate detailed view, not this summary.