Maximum number of open sockets

[tommikas]

It seems that currently the maximum total number of simultaneously available sockets is 4, in any combination of TCPServer, TCPSocket and UDPSocket instances. UDPSocket is further limited to a maximum of 2 (4 UDP sockets total, but one seems to be reserved for DNS and one for DHCP, leaving 2 for the application).

Are these limits documented anywhere? Are they configurable?

[infinnovation-dev]

I think this is the relevant definition in features/net/FEATURE_IPV4/lwip-interface/lwip/include/lwip/opt.h

/**
 * MEMP_NUM_NETCONN: the number of struct netconns.
 * (only needed if you use the sequential API, like api_lib.c)
 */
#ifndef MEMP_NUM_NETCONN
#define MEMP_NUM_NETCONN                4
#endif

[0xc0170]

cc @geky

[geky]

Looking at the referenced issue, you can only make 4 sequentially? That sounds like a memory-leak that should be looked into, or is it in a different issue?

@infinnovation is right, the MEMP_NUM_NETCONN can be overridden for additional sockets (at a memory cost). However, reserving a socket for DNS/DHCP sounds wrong, that may be a loss from the above memory-leak.

[tommikas]

Looking at the referenced issue, you can only make 4 sequentially? That sounds like a memory-leak that should be looked into, or is it in a different issue?

I don't think it's a memory leak, since I can free instances and recreate them in any order as many times as I wish. I.e. it's not a limitation on how many I can make, only on how many can exist simultaneously.

@infinnovation is right, the MEMP_NUM_NETCONN can be overridden for additional sockets (at a memory cost).

Alright. This might be worth documenting somewhere. It takes quite a bit of digging to find as it is.

However, reserving a socket for DNS/DHCP sounds wrong, that may be a loss from the above memory-leak.

To be precise, during EthernetInterface::connect() udp_new() gets called twice, once from dhcp_start() and once from dns_init(). Thus 2 of the available 4 udp PCBs are reserved. udp_remove() doesn't get called at all, so I guess that could be a memory leak? Although I could see DNS needing a guaranteed available socket...

Calling EthernetInterface::disconnect() releases the PCB reserved by dhcp. Since it's currently possible to create sockets on a disconnect interface, that makes it possible to open a 3rd UDPSocket instance (which then of course makes a new EthernetInterface::connect() fail because it can't get a new PCB).

(As a side note, I wonder what happens/should happen to existing sockets when the network interface is disconnected... )

[sg-]

This needs to be defined as the minimum guaranteed number of sockets per interface and documented.

[ciarmcom]

ARM Internal Ref: IOTMORF-446

[geky]

Does #3078 resolve this?

[tommikas]

Mostly this was just about these restrictions being documented somewhere, so yes, this helps. It's still not very visible, but at least it's something.