[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <20221206124342.7f429399@kernel.org>
Date: Tue, 6 Dec 2022 12:43:42 -0800
From: Jakub Kicinski <kuba@...nel.org>
To: Justin Iurman <justin.iurman@...ege.be>
Cc: Eric Dumazet <edumazet@...gle.com>, netdev@...r.kernel.org,
davem@...emloft.net, yoshfuji@...ux-ipv6.org, dsahern@...nel.org,
pabeni@...hat.com, stable@...r.kernel.org
Subject: Re: [RFC net] Fixes: b63c5478e9cb ("ipv6: ioam: Support for Queue
depth data field")
On Mon, 5 Dec 2022 21:44:09 +0100 Justin Iurman wrote:
> > Please revert this patch.
> >
> > Many people use FQ qdisc, where packets are waiting for their Earliest
> > Departure Time to be released.
>
> The IOAM queue depth is a very important value and is already used.
Can you say more about the use? What signal do you derive from it?
I do track qlen on Meta's servers but haven't found a strong use
for it yet (I did for backlog drops but not the qlen itself).
> > Also, the draft says:
> >
> > 5.4.2.7. queue depth
> >
> > The "queue depth" field is a 4-octet unsigned integer field. This
> > field indicates the current length of the egress interface queue of
> > the interface from where the packet is forwarded out. The queue
> > depth is expressed as the current amount of memory buffers used by
> > the queue (a packet could consume one or more memory buffers,
> > depending on its size).
> >
> > 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1
> > +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
> > | queue depth |
> > +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
> >
> >
> > It is relatively clear that the egress interface is the aggregate
> > egress interface,
> > not a subset of the interface.
>
> Correct, even though the definition of an interface in RFC 9197 is quite
> abstract (see the end of section 4.4.2.2: "[...] could represent a
> physical interface, a virtual or logical interface, or even a queue").
>
> > If you have 32 TX queues on a NIC, all of them being backlogged (line rate),
> > sensing the queue length of one of the queues would give a 97% error
> > on the measure.
>
> Why would it? Not sure I get your idea based on that example.
Because it measures the length of a single queue not the device.
Powered by blists - more mailing lists