> > Why don't we update the socket allocation when doing pskb_expand_head()=
?
quoted text
> > Sure, it could become negative, but is that so bad?
>=20
> The socket locked state at this time is variable and unknown.
>=20
> The socket must be locked in order to modify these values.
> And such locks cannot be taken, for example, from HW interrupt
> context, amongst other restrictions.

Ok, that makes sense.

quoted text
> > We need more space though. Should we then just increase the built-in
> > headroom?
>=20
> I simply don't know what to suggest at this point, that's why
> we are having this discussion :-)

I'm still not sure about the dependencies between LL_MAX_HEADER,
dev->hard_header_len and similar. Why, for example, does ipip set it to
LL_MAX_HEADER + sizeof(struct iphdr)? Because it doesn't know better
since the packets it creates could be routed anywhere?

Could mac80211 announce it needs a very long hard_header_len (say 48 (or
54) bytes)? Am I right in thinking that then we'd have to increase
LL_MAX_HEADER as well? I haven't found a check somewhere that warns you
if you set dev->hard_header_len > LL_MAX_HEADER, should there be one?

If I increase dev->hard_header_len, will that have any negative impact
on the caching since I'm still just using regular ethernet headers?


As far as I understand we have a few options:

 (a) go along with it as we do now, use pskb_expand_head, just call
     skb_orphan before. I assume this has a number of requirements just
     like sock size accounting would have, does this work from within a
     hard_start_xmit path? I haven't seen any problems with it so far
     anyway.
 (b) clone the skb and free the original. pretty much equivalent
 (c) increase hard_header_len/LL_MAX_HEADER constants to 48 (54).

Options (a) and (b) make the accounting pretty useless since that would
drop the charge to the socket quite early. (c) doesn't seem to work, I
tried =EF=BB=BFjust increasing LL_MAX_HEADER doesn't seem to help although
MAX_TCP_HEADER suggests it should be getting enough headroom then.

Ideally, we'd have enough headroom in the skb to start with, since right
now we're apparently reallocating a lot, especially encrypted frames.
Not that I understand why we don't get a truesize bug (without the
monitors) when we do that.

With smart hardware like b43 we could even think about putting the
802.11 header stuff into a separate buffer and have the hardware to
gather-dma but there are so many dumb usb devices that this won't help
much anyway.

johannes