Re: [lwip-users] out of pbuf's

[Jonathan Larmour]

Mark Six wrote:
> 
> For each incoming ethernet packet, one or more pbuf's are allocated. The
> number of pbuf's available is defined by the PBUF_POOL_SIZE.
> If there are no more pbuf's available (e.g. all pbuf's are allocated) the
> incoming ethernet packet is dropped. The TCP/IP stack handles the incoming
> packets and will free them. If the packet is a data-packet, the
> application layer is responsible for freeing the packet.
> However if the packets come out of sequence, the pbuf is *NOT* free'd, but
> placed on the oos queue. If the out-of-sequence packets are equal to the
> PBUF_POOL_SIZE the lwip-stack is in a deadlock for some time (until some
> timer times out and frees the oos queue).
> To limit the number of out-of-sequence packets one could adjust the
> receive window size (TCP_WND), however it is possible that one packet can
> have only 1 byte of data, so the value for the receive window can be no
> more than the available pbuf's divided by the maximum number of TCP
> connections.

Strictly yes, although in practice you can usually raise it a little above
that, as statistically you don't always get the worst case!

If you usually expect smaller packets, then you should use a smaller
PBUF_POOL_BUFSIZE, and then fewer will be used, and they will be chained if
more total space is needed for a larger packet. What's most efficient
depends on your expected traffic pattern.

But if your hypothesis is correct, then turning off TCP_QUEUE_OOSEQ may be
a good idea. You can't expect to be able to use every nice feature when you
have limited memory.

> This is a problem, because we need a lot of memory or a very small receive
> window.

Indeed.

> Any comments on the above are welcome..

It might be nice to have an lwIP option tuning guide at some point to cover
all this.

Jifl
-- 
eCosCentric Limited      http://www.eCosCentric.com/     The eCos experts
Barnwell House, Barnwell Drive, Cambridge, UK.       Tel: +44 1223 245571
Registered in England and Wales: Reg No 4422071.
------["Si fractum non sit, noli id reficere"]------       Opinions==mine