Merge branch 'prevent-permanently-closed-tc-taprio-gates-from-blocking-a-felix-dsa...

Merge branch 'prevent-permanently-closed-tc-taprio-gates-from-blocking-a-felix-dsa-switch-port'

Vladimir Oltean says:

====================
Prevent permanently closed tc-taprio gates from blocking a Felix DSA switch port

Richie Pearn reports that if we install a tc-taprio schedule on a Felix
switch port, and that schedule has at least one gate that never opens
(for example TC0 below):

tc qdisc add dev swp1 root taprio num_tc 8 map 0 1 2 3 4 5 6 7 \
	queues 1@0 1@1 1@2 1@3 1@4 1@5 1@6 1@7 \
	base-time 0 sched-entry S fe 1000000 flags 0x2

then packets classified to the permanently closed traffic class will not
be dequeued by the egress port. They will just remain in the queue
system, to consume resources. Frame aging does not trigger either,
because in order for that to happen, the packets need to be eligible for
egress scheduling in the first place, which they aren't. If that port is
allowed to consume the entire shared buffer of the switch (as we
configure things by default using devlink-sb), then eventually, by
sending enough packets, the entire switch will hang.

If we think enough about the problem, we realize that this is only a
special case of a more general issue, and can also be reproduced with
gates that aren't permanently closed, but are not large enough to send
an entire frame. In that sense, a permanently closed gate is simply a
case where all frames are oversized.

The ENETC has logic to reject transmitted packets that would overrun the
time window - see commit 285e8dedb4bd ("net: enetc: count the tc-taprio
window drops").

The Felix switch has no such thing on a per-packet basis, but it has a
register replicated per {egress port, TC} which essentially limits the
max MTU. A packet which exceeds the per-port-TC MTU is immediately
discarded and therefore will not hang the port anymore (albeit, sadly,
this only bumps a generic drop hardware counter and we cannot really
infer the reason such as to offer a dedicated counter for these events).

This patch set calculates the max MTU per {port, TC} when the tc-taprio
config, or link speed, or port-global MTU values change. This solves the
larger "gate too small for packet" problem, but also the original issue
with the gate permanently closed that was reported by Richie.
====================

Link: https://lore.kernel.org/r/20220628145238.3247853-1-vladimir.oltean@nxp.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>