lkml.org 
[lkml]   [2006]   [May]   [24]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH 3/4] myri10ge - Driver core
Anton Blanchard wrote:
>> +/*
>> + * Set of routunes to get a new receive buffer. Any buffer which
>> + * crosses a 4KB boundary must start on a 4KB boundary due to PCIe
>> + * wdma restrictions. We also try to align any smaller allocation to
>> + * at least a 16 byte boundary for efficiency. We assume the linux
>> + * memory allocator works by powers of 2, and will not return memory
>> + * smaller than 2KB which crosses a 4KB boundary. If it does, we fall
>> + * back to allocating 2x as much space as required.
>> + *
>> + * We intend to replace large (>4KB) skb allocations by using
>> + * pages directly and building a fraglist in the near future.
>> + */
>>
>
> You go to a lot of trouble to align things. One thing on ppc64 is that
> we really want to start all DMA writes on a cacheline boundary. We
> enforce that in network drivers by making NET_IP_ALIGN = 0 and having
> the drivers do:
>
> skb_reserve(skb, NET_IP_ALIGN);
>
> It sounds like your small allocations will be only aligned to 16 bytes.
>

We didn't get any ppc64 with PCI-E to run Linux so far. What performance
drop should we expect with our current code ?

> Id suggest using the dma API instead of the pci API. We have seen
> machines in the field that have failed large pci_alloc_consistent calls
> because it always asks for GFP_ATOMIC memory (it presumes the worst).
> The dma API allows you to pass a GFP_ flag in which will have a much
> better chance of succeeding when you dont need GFP_ATOMIC memory.
>

Good idea, thanks,

>> +#ifdef CONFIG_MTRR
>> + mgp->mtrr = mtrr_add(mgp->iomem_base, mgp->board_span,
>> + MTRR_TYPE_WRCOMB, 1);
>> +#endif
>>
> ...
>
>> + mgp->sram = ioremap(mgp->iomem_base, mgp->board_span);
>>
>
> Not sure how we are meant to specify write through in drivers. Any ideas Ben?
>

I am not sure what you mean.
The only ppc64 with PCI-E that we have seen so far (a G5) couldn't do
write combining according to Apple.

Thanks,
Brice

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

\
 
 \ /
  Last update: 2006-05-24 10:08    [W:0.077 / U:6.724 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site