kfunc-gen: Don't list sleepable kfuncs for non-sleepable program types

[NAME-ASHWANIYADAV]

Fixes #138

What

kfunc-gen builds the kfunc <-> program type references purely from the sets in data/kfuncs.yaml, without looking at whether a kfunc may sleep. But a kfunc flagged KF_SLEEPABLE can only be called from a sleepable program:

sleepable = is_kfunc_sleepable(&meta);
if (sleepable && !in_sleepable(env)) {
    verbose(env, "program must be sleepable to call sleepable kfunc %s\n", func_name);
    return -EACCES;
}

So listing one under a program type that can never be sleepable is wrong in both directions: the kfunc page offers a program type that cannot call it, and the program type page lists a kfunc the verifier will always reject.

Which program types can be sleepable

check_attach_btf_id() decides this. Every type opts in with BPF_F_SLEEPABLE and has to pass can_be_sleepable(), except BPF_PROG_TYPE_SYSCALL, which is handled just above it and is always sleepable. That gives:

Program type	Note
BPF_PROG_TYPE_TRACING	fentry, fexit, fmod_ret, iter
BPF_PROG_TYPE_LSM
BPF_PROG_TYPE_KPROBE	uprobes only
BPF_PROG_TYPE_STRUCT_OPS
BPF_PROG_TYPE_SYSCALL	always sleepable

The set has only ever grown: v5.15 allowed TRACING and LSM, v6.1 added KPROBE, v6.4 added STRUCT_OPS and factored the check into can_be_sleepable(), and it has been unchanged since. So filtering against the current set cannot drop a
combination that is valid on an older kernel.

Note the filter is per program type, so KPROBE and TRACING stay in the lists even though only some of their attach types can be sleepable. Going finer would need attach-type granularity, which the data file does not carry.

Where the fix goes

One filter on the merged progTypes, next to the existing Except handling. That covers both directions at once, since the KFUNC_PROG_REF blocks and the progToKfunc map behind PROG_KFUNC_REF are both derived from it.

Impact

301 kfunc <-> program type combinations removed: 22 kfunc pages and 18 program type pages, 583 deletions and no additions.

BPF_PROG_TYPE_PERF_EVENT and BPF_PROG_TYPE_TRACEPOINT lose 23 entries each; the 15 program types reachable through BPF_PROG_TYPE_UNSPEC expansion (XDP, SCHED_CLS, the CGROUP_* family, LWT_*, NETFILTER, ...) lose 15 each.

Four KF_SLEEPABLE kfuncs are unaffected because they were already correct: bpf_crypto_ctx_create (SYSCALL), bpf_io_uring_submit_sqes (STRUCT_OPS), scx_bpf_create_dsq (STRUCT_OPS + SYSCALL), and cgroup_rstat_flush, which is
in removeKfuncs.

Testing

• gofmt clean, go vet ./tools/kfunc-gen/ clean
• The generator is idempotent: a second run produces no further diff, which is
  what the generate-docs workflow asserts
• No kfunc page is left with an empty program type list
• Only KF_SLEEPABLE kfunc pages changed, nothing else was touched
• Cross-checked the output against an independent script that replays the merge
  logic over data/kfuncs.yaml. The only mismatch was
  BPF_PROG_TYPE_LWT_SEG6LOCAL, whose page has no PROG_KFUNC_REF markers, so
  the generator skips it. Pre-existing, left alone.

[NAME-ASHWANIYADAV]

@dylandreimerink, could you please review this PR when you have a moment?

[dylandreimerink]

This is great, thank you