Hi Zdenek,
Zdenek Bouska wrote:
Hi Jiri,
in mainline changeset 1358 (Eliminate packet_t from sending direction of
NIC interface.) you added frame copying to E1000 driver.
Line 2290 of e1k.c
http://trac.helenos.org/browser/mainline/uspace/drv/nic/e1k/e1k.c?rev=mainline%2C1358#L2290
Is this only temporary solution? Or is this necessary after elimination
of packet_t?
In my networking stack rewrite I try to make things as simple as possible (at
least at the beginning). I am eliminating the packet sharing mechanism -
packets are transferred from one network task to another simply using IPC
read/write.
The fact that the frame data is memcpied() to the DMA buffer after being
IPC-written to e1k is necessary under the current simple definition of the
send_frame handler. It could be eliminated if the interface was made more
clever (i.e. allowed the driver to register DMA buffers with NIC framework and
write data there directly). I might do this (and other similar optimization) in
the future. But it's just that - an optimization.
I believe that for <2kB size buffers the cost of copying is relatively small
compared to the cost of context switch. Therefore the cost should be bearable.
When the stack is working reliably I might do some benchmarks and optimize based
on need.
I am basically taking the opposite approach from Lukas Mejdrech - make simple
first, optimize later.
Cheers,
Jiri
_______________________________________________
HelenOS-devel mailing list
[email protected]
http://lists.modry.cz/cgi-bin/listinfo/helenos-devel
