Re: [lwip-devel] again... Zero Copy Ethernet interface

[Jakob Stoklund Olesen]

Piero 74 wrote:

> I cannot imagine a implementation for sending operation using zero copy...
> i can see TWO problems:
> - first, as already someone said above, i have to solve problem of lwip
> which wants to free pbuf after low_level_output... a solution could be
> (if a RTOS is used) waiting on semaphore until emac isr unlocks it,
> after DMA sending packet. In this way CPU will be avalaible for other
> tasks (threads) during DMA sending

I have implemented zero-copy DMA transmission like this:

err_t
low_level_output(struct pbuf *pb)
{
  if (can_direct_dma(pb)) {
    if (queue_dma_transfer(pb)) {
      pbuf_ref(pb);
      return ERR_OK;
    }
    else {
      return ERR_IF;
    }
  }
  else {
    pb = flat_copy_pbuf(pb);
    if (queue_dma_transfer(pb)) {
      return ERR_OK;
    }
    else {
      pbuf_free(pb);
      return ERR_IF;
    }
  }
}

dma_done(dma_transfer *tr)
{
  pbuf_free(tr->pb);
}

This seems to work. If anybody knows why it shouldn't, please let me know.

I can DMA from any memory region, but I can only do limited
scatter-gather DMA. A pbuf that is too fragmented must be copied to
continuous memory.

Anyway, calling pbuf_ref() in low_level_output prevents the pbuf from
being deleted before DMA is completed.

> - second (correlated to similar problem for rx), when lwip send a
> packet, this is allocated in pbuf RAM, in lwip heap, (if i'm right...),
> but, again, my DMA wants packet in specific region (emac RAM)... so,
> pbuf ram needs relocation in fixed memory region too, but i think is
> difficoult to change lwip code for this...

If you want all sent pbufs to be in DMA-able memory, it seems to me that
you must place the entire lwip memory pool in DMA-able memory. This can
be done in mem.c, and possibly memp.c

Even if you do that, a pbuf with the PBUF_ROM flag set can have payload
anywhere.


/stoklund