[ipxe-devel] alloc_iob failing

[Michael Brown]

On 08/03/14 04:23, Pfaff, Mark wrote:
> I am working on a propriety driver which is very similar to the legacy.c code in the ipxe codebase.

That's not a good starting point.  As the name suggests, legacy.c is 
just a wrapper around the legacy drivers which have not been converted 
to the iPXE driver model.  It does not make sense to create new code 
following that model.

> In the same fashion I initialize a new iobuf ptr on every receive iteration and then send it to netdev_rx.  I don't have the need for a circular private iobuffer list at this time as I have my own backend dma buffer which memcpy's the packet to my iobuf for receives.

The general structure of an iPXE driver RX datapath is that you allocate 
a ring of RX buffers which you pass to the hardware, and then when the 
hardware reports a received packet you just mark the buffer as filled 
with the appropriate length (using iob_put()) and hand the iobuf to 
netdev_rx().  See e.g. intel.c for a good example of how this works.

>   I'd like for my poll to call receive, send my iobuf to netdev_rx and have it sent to the stack and cleaned up.
>
> My code base works perfectly for pings and well as small transfers.  After loading about 40% of an 800k kernel/image alloc_iob is failing.  Am I missing a step to release iobuf's at some point, or do you have other suggestions?

Please publish your driver (as required by the GPL), and we can then 
review the code.

Michael