Re: [PATCH v4 net-next 6/9] net/sched: mqprio: allow per-TC user input of FP adminStatus

[Jamal Hadi Salim]

On Fri, Apr 7, 2023 at 3:31 PM Vladimir Oltean <olteanv@xxxxxxxxx> wrote:
>
> On Fri, Apr 07, 2023 at 02:49:01PM -0400, Jamal Hadi Salim wrote:
> > On Fri, Apr 7, 2023 at 12:41 PM Vladimir Oltean <vladimir.oltean@xxxxxxx> wrote:
> > > On Fri, Apr 07, 2023 at 12:22:26PM -0400, Jamal Hadi Salim wrote:
> > > > > +enum {
> > > > > +       TC_FP_EXPRESS = 1,
> > > > > +       TC_FP_PREEMPTIBLE = 2,
> > > > > +};
> > > >
> > > > Suggestion: Add a MAX to this enum (as is traditionally done)..
> > >
> > > Max what? This doesn't count anything, it just expresses whether the
> > > quality of one traffic class, from the Frame Preemption standard's
> > > perspective, is express or preemptible...
> > >
> > > > > @@ -145,13 +149,94 @@ static int mqprio_parse_opt(struct net_device *dev, struct tc_mqprio_qopt *qopt,
> > > > >         return 0;
> > > > >  }
> > > > >
> > > > > +static const struct
> > > > > +nla_policy mqprio_tc_entry_policy[TCA_MQPRIO_TC_ENTRY_MAX + 1] = {
> > > > > +       [TCA_MQPRIO_TC_ENTRY_INDEX]     = NLA_POLICY_MAX(NLA_U32,
> > > > > +                                                        TC_QOPT_MAX_QUEUE),
> > > >
> > > > And use it here...
> > >
> > > Where? Above or below the comment? I think you mean below (for the
> > > policy of TCA_MQPRIO_TC_ENTRY_FP)?
> >
> > That was what I meant. I misread that code thinking it was a nested
> > TLV range check. If it is only going to be those two specific values,
> > I understand - but then wondering why you need a u32; wouldnt a u8 be
> > sufficient?
>
> I believe netlink isn't exactly optimized for passing small values; the
> netlink attributes are going to be aligned to NLA_ALIGNTO (4) anyway,
> so it's not like this is going to save space or something. Also, there's
> a policy restricting the maximum, so arbitrarily large values cannot be
> passed now, but could be passed later if needed. I did not see any good
> enough reason to close that door.
>
> > The only reason you would need a MAX is if it is possible that new
> > values greater than TC_FP_PREEMPTIBLE showing up in the future.
>
> Even if someone wants to add TC_FP_KINDA_PREEMPTIBLE = 3 and
> TC_FP_PREEMPTIBLE_WITH_STRIPES = 4 in the future, I'm still not sure how
> a MAX definition exported by the kernel is going to help them?
>
> I mean, about the only thing that it would avoid is that I wouldn't be
> changing the policy definition, but that's rather minor and doesn't
> justify exporting something to UAPI?

Yes, it is minor (and usually minor things generate the most emails;->).
I may be misunderstanding what you mean by "doesnt justify exporting
something to UAPI"  - those definitions are part of uapi and are
already
being exported.

> The changed MAX value is only a
> property of the kernel headers that the application is compiled with -
> it doesn't give the capability of the running kernel.
>

Maybe I am missing something or not communicating effectively. What i
am suggesting is something very trivial:

enum {
TC_FP_EXPRESS = 1,
TC_FP_PREEMPTIBLE = 2,
TC_FP_MAX
};

 [TCA_MQPRIO_TC_ENTRY_FP] = NLA_POLICY_RANGE(NLA_U32,
    TC_FP_EXPRESS,
    TC_FP_MAX),

And in a newer revision with TC_FP_PREEMPTIBLE_WITH_STRIPES:

enum {
TC_FP_EXPRESS = 1,
TC_FP_PREEMPTIBLE = 2,
TC_FP_PREEMPTIBLE_WITH_STRIPES = 3,
TC_FP_MAX
};
etc.

Saves you one line in a patch for when TC_FP_PREEMPTIBLE_WITH_STRIPES shows up.

> To see whether TC_FP_PREEMPTIBLE_WITH_STRIPES is supported, the
> application would have to try it and see if it fails. Which is also the
> case right now with TC_FP_PREEMPTIBLE.

You may be referring to the combination of  iproute2/kernel.
In all cases, NLA_POLICY_RANGE will take care of rejecting something
out of bound.

> > > > Lead up question: if the max is 16 then can preemptible_tcs for example be u32?
> > >
> > > I don't understand this question, sorry. preemptible_tcs is declared as
> > > "unsigned long", which IIUC is at least 32-bit.
> >
> > I meant: if you only had 16 possible values, meaning 16 bits are
> > sufficient, (although i may be misunderstanding the goal of those
> > bits) why not be explicit and use the proper type/size?
>
> If you think it's valuable to change the type of preemptible_tcs from
> unsigned long to u16 and that's a more "proper" type, I can do so.

No, no, it is a matter of taste and opinion. You may have noticed,
trivial stuff like this gets the most comments and reviews normally(we
just spent like 4-5 emails on this?). Poteto/potato: IOW, if i was to
do it i would have used a u16 or u32 because i feel it would be more
readable. I would have used NLA_U8 because i felt it is more fitting
and i would have used a max value because it would save me one line in
a patch in the future. I think weve spent enough electrons on this - I
defer to you.

cheers,
jamal