On Tue, 19 Dec 2017, Michael Schmitz wrote:
> >
> > > Contrary to the Mac driver, esp->command_block and
> > > esp->command_block_dma are not identical addresses on Amiga.
> >
> > Why not make them identical, depending on the length of the tranfer?
> > (Then choose PIO or DMA by testing for the same threshold.)
>
> esp->command_block_dma is mapped at driver init time so we don't get to
> chose at command submission time.
You could set esp->command_block_dma to an address like 0xffffffff. Then
when zorro_esp's ops->send_dma_cmd method is to do a DMA to or from that
address, check the transfer length. If the length is small, use
esp->command_block as the address for a PIO transfer. Otherwise, use the
mapped esp->command_block (which you saved somewhere at driver init time)
for a DMA transfer.
This is a hack, of course. Perhaps we can avoid the hack if we pass the
virtual address as an argument to ops->send_dma_cmd.
(Or *sg could be passed, if we cook up a couple of scattergather structs
to replace ent->sense_ptr and esp->command_block. This might allow for the
elimination of ops->map_single and ops->unmap_single altogether...)
> >
> > > (I can use esp->command_block in the reconnect message special case
> > > but not otherwise ...)
> >
> > Maybe something like this...
> >
> > struct esp_cmd_entry *ent = esp->active_cmd;
> > struct esp_cmd_priv *spriv = ESP_CMD_PRIV(ent->cmd);
> > struct scatterlist *sg = spriv->cur_sg;
> > unsigned long addr = sg_page(sg) + sg->offset;
> >
> > but that doesn't work for the esp_autosense() case, which doesn't
> > involve esp->ops->map_sg...
> >
> > HTH
>
> I'll give that a try to see if this works for regular transfers.
>
> Will look at ways to identify autosense as well (do sense data fit into
> the ESP FIFO usually?)
You can identify the autosense case easily enough,
struct esp_cmd_entry *ent = esp->active_cmd;
struct esp_cmd_priv *spriv = ESP_CMD_PRIV(ent->cmd);
unsigned long addr;
if (ent->flags & ESP_CMD_FLAG_AUTOSENSE)
addr = spriv->sense_ptr;
else {
struct scatterlist *sg = spriv->cur_sg;
addr = sg_page(sg) + sg->offset;
}
But that still leaves the command_block case unhandled (I don't think we
can use spriv->cur_sg for that, so we'd need a hack like the one above).
All this complexity doesn't really belong in the ops->send_dma_cmd method,
so I think we are back to passing in the virtual pointer as a parameter.
That change would touch every esp driver. But it could improve mac_esp,
since mac_esp would no longer have to do any fake dma mapping. Maybe I
should do a patch for this?
--
>
> Cheers,
>
> Michael
>
>
--
To unsubscribe from this list: send the line "unsubscribe linux-m68k" in
the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
