vfs: allow tail call optimisation in vops in the common case

[mjg]

Summary

Most frequently used vops boil down to checking SDT probes, doing the call and checking again. There is no vop_post/pre in their case. The check after the call prevents tail call optimisation from taking place. Instead, check once upfront. Kernels with debug or vops with non-empty vop_post still don't short circuit.

Generated code:

int
VOP_LOCK1_APV(struct vop_vector *vop, struct vop_lock1_args *a)
{
        int rc;

        VNASSERT(a->a_gen.a_desc == &vop_lock1_desc, a->a_vp,
            ("Wrong a_desc in vop_lock1(%p, %p)", a->a_vp, a));
        VNASSERT(vop != NULL, a->a_vp, ("No vop_lock1(%p, %p)", a->a_vp, a));
        KTR_START4(KTR_VOP, "VOP", "VOP_LOCK1", (uintptr_t)a,
            "vp:0x%jX", (uintptr_t)a->a_vp, "flags:0x%jX", a->a_flags, "file:0x%jX", a->a_file, "line:0x%jX", a->a_line);
        vop_lock_pre(a);
        if (__predict_true(!SDT_PROBES_ENABLED() && vop->vop_lock1 != NULL)) {
                rc = vop->vop_lock1(a);
        } else {
                SDT_PROBE2(vfs, vop, vop_lock1, entry, a->a_vp, a);

                if (vop->vop_lock1 != NULL)
                        rc = vop->vop_lock1(a);
                else
                        rc = vop->vop_bypass(&a->a_gen);
                SDT_PROBE3(vfs, vop, vop_lock1, return, a->a_vp, a, rc);
        }
        if (rc == 0) {
        } else {
        }
        vop_lock_post(a, rc);
        KTR_STOP4(KTR_VOP, "VOP", "VOP_LOCK1", (uintptr_t)a,
            "vp:0x%jX", (uintptr_t)a->a_vp, "flags:0x%jX", a->a_flags, "file:0x%jX", a->a_file, "line:0x%jX", a->a_line);
        return (rc);
}

I verified that at least on amd64 tail call is indeed used.

Test Plan

will ask pho

[jeff]

It would be nice to flatten the vop_pre/vop_post stuff. We could define them in a file that is included in the interface generated source so they can be inlined and removed when necessary.

This would also look nicer when we don't have to check for NULL vops.

[mjg]

The vop_*_{pre,post} stuff is per operation and on kernels without debug they typically expand to nothing. I even noted vop->vop_lock1(a); gets tail-called which would not be possible if vop_lock_post had anything to do. The few cases which do anything in production kernels require fixing by hand (as in their internals are iffy).

[mjg]

• drop spurious newline