On 10/17/2020 3:17 PM, Vladimir Oltean wrote:
On Sat, Oct 17, 2020 at 03:11:52PM -0700, Florian Fainelli wrote:
slave_dev->needed_headroom += master->needed_headroom;
slave_dev->needed_tailroom += master->needed_tailroom;
Not positive you need that because you may be account for more head or tail
room than necessary.
For instance with tag_brcm.c and systemport.c we need 4 bytes of head room
for the Broadcom tag and an additional 8 bytes for pushing the transmit
status block descriptor in front of the Ethernet frame about to be
transmitted. These additional 8 bytes are a requirement of the DSA master
here and exist regardless of DSA being used, but we should not be
propagating them to the DSA slave.
And that's exactly what I'm trying to do here, do you see any problem
with it? Basically I'm telling the network stack to allocate skbs with
large enough headroom and tailroom so that reallocations will not be
necessary for its entire TX journey. Not in DSA and not in the
systemport either. That's the exact reason why the VLAN driver does this
too, as far as I understand. Doing this trick also has the benefit that
it works with stacked DSA devices too. The real master has a headroom
of, say, 16 bytes, the first-level switch has 16 bytes, and the
second-level switch has 16 more bytes. So when you inject an skb into
the second-level switch (the one with the user ports that applications
will use), the skb will be reallocated only once, with a new headroom of
16 * 3 bytes, instead of potentially 3 times (incrementally, first for
16, then for 32, then for 48). Am I missing something?
That is fine with me, given that we can resolve most of the TX path
ahead of time, I suppose we should indeed take advantage of that
knowledge. Thanks!
--
Florian
