Re: [PATCH RESEND net-next 0/5] Improve the taprio qdisc's relationship with its children

[Jamal Hadi Salim <jhs@mojatatu.com>]

On Tue, Jun 6, 2023 at 12:32 PM Vladimir Oltean <vladimir.oltean@nxp.com> wrote:
>
> On Tue, Jun 06, 2023 at 11:39:32AM -0400, Jamal Hadi Salim wrote:
> > 1)Just some details become confusing in regards to offload vs not; F.e
> > class grafting (taprio_graft()) is de/activating the device but that
> > seems only needed for offload. Would it not be better to have those
> > separate and call graft_offload vs graft_software, etc? We really need
> > to create a generic document on how someone would write code for
> > qdiscs for consistency (I started working on one but never completed
> > it - if there is a volunteer i would be happy to work with one to
> > complete it).
>
> I would be a happy reader of that document. I haven't studied whether
> dev_deactivate() and dev_activate() are necessary for the pure software
> data path, where the root taprio is also directly attached to the netdev
> TXQs and that fact doesn't change across its lifetime.

I didnt go that far in the doc but was intending to. It was supposed
to be a tutorial somewhere and i ended not doing it.
something like htb will be a good example to compare with (it is a
classful qdisc which is also capable of offloading). It is not the
same as mqprio which can only be root.

> > 2) It seems like in mqprio this qdisc can only be root qdisc (like
> > mqprio)
>
> so far so good
>
> > and you dont want to replace the children with other types of
> > qdiscs i.e the children are always pfifo? i.e is it possible or
> > intended for example to replace 8001:x with bfifo etc? or even change
> > the pfifo queue size, etc?
>
> no, this is not true, why do you say this?

I am just asking questions trying to understand;-> So if can i
replace, for example, with a tbf would it make sense even in s/w?

> > 3) Offload intention seems really to be bypassing enqueue() and going
> > straigth to the driver xmit() for a specific DMA ring that the skb is
> > mapped to. Except for the case where the driver says it's busy and
> > refuses to stash the skb in ring in which case you have to requeue to
> > the appropriate child qdisc/class. I am not sure how that would work
> > here - mqprio gets away with it by not defining any of the
> > en/de/requeue() callbacks
>
> wait, there is a requeue() callback? where?

Hrm, someone removed that ops i guess at some point - not sure when,
need to look at git history.
But short answer is yes it used to be there; git will probably reveal
from the commit its disappearance.

>
> > - but likely it will be the lack of requeue that makes it work.
>
> Looking at dev_requeue_skb(), isn't it always going to be requeued to
> the same qdisc it was originally dequeued from? I don't see what is the
> problem.

In the basic case that approach is sufficient. For pfifo you want it
to the first skb dequeued the next opportunity to send to h/w.
Basically the idea is/was: if the hardware is busy you may need to run
some algorithm (present in requeue but not in enqueu) to decide if you
should put the skb at the head, at the tail, somewhere else, drop it,
check if some time limit has exceeded and do something funky etc etc.

> My understanding of the offload intention is not really to bypass dequeue()
> in general and as a matter of principle, but rather to bypass the root's
> taprio_dequeue() specifically, as that could do unrelated work, and jump
> right to the specific child's dequeue().
>
> The child could have its own complex enqueue() and dequeue() and that is
> perfectly fine - for example cbs_dequeue_soft() is a valid child dequeue
> procedure - as long as the process isn't blocked in the sendmsg() call
> by __qdisc_run() processing packets belonging to unrelated traffic
> classes.

Does it matter what type the child enqueue/dequeue? eg can i attach htb, etc?

cheers,
jamal