[lwip-users] Re: [lwip] Memory leak and queued packets in etharp.c

[Leon Woestenberg]

Hello,

> Pbufs of type PBUF_ROM are never supposed to be overwritten by the
> application (hence the name ROM = Read Only Memory). If there are
> problems with netbuf_ref() and pbufs, it is better to change the netbuf_
> API than the core code.
>
> So: no copying of PBUF_ROM pbuf, since this should have been done by the
> application (into a PBUF_RAM pbuf) if the application intended to
> overwrite the data.
>

Then, the question is: how can zero-copy be used between
application-managed-memory
and lwIP in the current architecture?

The PBUF_ROM type can be used to exploit the "zero-copy" feature of lwIP,
i.e.
payload is never redundantly copied between data and network interface
hardware.

This is a desirable feature. (Many UNIX stacks are not zero-copy; one of the
BSDs
recently has achieved the "zero-copy" feature, cannot remember if it was
NetBSD or
one of the others.)

AFAIK, abusing PBUF_ROM is the only way to exploit zero-copy of payload data
in
application-managed-memory.

For non-ROM non-lwIP managed memory, maybe the introduction of PBUF_REF
is desirable. It has the same properties as PBUF_ROM, only that the data
itself is
not be considered constant after it has been submitted to lwIP.

Using such an approach would allow zero-copy in case the packet can be
immediately
send, or payload-copying in case the packet is queued in lwIP.

Regards,

Leon.

[This message was sent through the lwip discussion list.]