Re: [lwip-users] low_level_output question

[address@hidden]

Bill Auerbach wrote:
> Incrementing the reference count is the way to move the freeing of the pbuf
> from the stack to the Ethernet driver.  It's been done by several
> implementers successfully.
I can tell you that won't work if
- sending data over UDP using the sockets API and
- having a fast device and
- modifying the buffer passed to lwip_send(to) after it returns

The reason is that (like I said in my previous post), incrementing the 
ref-count will do for _RAM and _POOL, but not for _REF and _ROM pbufs; 
udp sockets use _REF. Here, incrementing the ref-count prevents the pbuf 
from being freed, but it does not prevent the actual from getting 
modified before sending. The only two solutions are a) not returning 
from lwip_send(to) until the data is sent by the hardware and b) copying 
the data before sending (using PBUF_RAM, not PBUF_REF). Both solutions 
have their problems, thus task #7896 to discuss it.

> A comment in low_level_output to increment the
> ref count if the packet isn't being physically sent before the function
> returns and to free it at a later time when it has is sufficient.
I'm afraid it's not. See above.
> What hasn't been clear to me is the safety of handling pbufs in interrupt 
> service
> routines.  Can a TX dma complete ISR call pbuf_free?
>   
Yes, it can, as long as LWIP_ALLOW_MEM_FREE_FROM_OTHER_CONTEXT is set to 
1. Note this has been fixed after 1.3.0, so unless you are using a CVS 
version from after that, you will have to wait for 1.3.1 to use it.

Simon