is each frag of a skb always less than 1 page?

is each frag of a skb always less than 1 page?

[Ronghua Zhang]

>From the code of tcp_sendmsg, it seems to me that each frag is at most
one page. Is it a guaranteed property or the driver should not assume
it? Thanks

Ronghua

Re: is each frag of a skb always less than 1 page?

[David Miller]

From: "Ronghua Zhang" <ronghuazhang@gmail.com>
Date: Tue, 10 Oct 2006 15:37:04 -0700

> >From the code of tcp_sendmsg, it seems to me that each frag is at most
> one page. Is it a guaranteed property or the driver should not assume
> it? Thanks

This assumption basically already exists everywhere since
all drivers that create DMA mappings of the frags call
pci_map_page().

Re: is each frag of a skb always less than 1 page?

[Ronghua Zhang]

Thank you for your explanation, David.

But pci_map_page should work fine if we have multiple contiguous pages
in one frag that are allocated using alloc_pages, ie:

  frag->page = alloc_pages(GFP_KERNEL, 2); // get 4 pages
  frag->page_offset = 0;
  frag->size = 4 * PAGE_SIZE;

  dma_addr = pci_map_page(frag->page, frag->page_offset, frag->size, ..);

The reason I asked this is that I saw the following code in forthdeth drvier:

#define NV_TX2_TSO_MAX_SHIFT) 14
/* add fragments to entries count */
for (i = 0; i < fragments; i++) {
    entries += (skb_shinfo(skb)->frags[i].size >> NV_TX2_TSO_MAX_SHIFT) +
                    ((skb_shinfo(skb)->frags[i].size &
(NV_TX2_TSO_MAX_SIZE-1)) ? 1 : 0);
}

This looks unnecessary if each frag is guaranteed not to span pages.

Ronghua

On 10/10/06, David Miller <davem@davemloft.net> wrote:
> From: "Ronghua Zhang" <ronghuazhang@gmail.com>
> Date: Tue, 10 Oct 2006 15:37:04 -0700
>
> > >From the code of tcp_sendmsg, it seems to me that each frag is at most
> > one page. Is it a guaranteed property or the driver should not assume
> > it? Thanks
>
> This assumption basically already exists everywhere since
> all drivers that create DMA mappings of the frags call
> pci_map_page().
>

Re: is each frag of a skb always less than 1 page?

[David Miller]

From: "Ronghua Zhang" <ronghuazhang@gmail.com>
Date: Tue, 10 Oct 2006 16:26:58 -0700

> Thank you for your explanation, David.
> 
> But pci_map_page should work fine if we have multiple contiguous pages
> in one frag that are allocated using alloc_pages, ie:
> 
>   frag->page = alloc_pages(GFP_KERNEL, 2); // get 4 pages
>   frag->page_offset = 0;
>   frag->size = 4 * PAGE_SIZE;
> 
>   dma_addr = pci_map_page(frag->page, frag->page_offset, frag->size, ..);

You should look at the implementation of pci_map_page() on
some interesting architectures such as sparc64 and powerpc
where an IOMMU is used to map the pages into DMA space on the
PCI bus.

Those implementations (rightly) assume only a single page needs
to be mapped.

If you assume that multiple pages would get mapped properly by
such a call, things would fail.  You might even get a PCI Master
ABORT after crossing past the end of the first page in that
multi-page chunk since the IOMMU will not have a DMA translation.

Re: is each frag of a skb always less than 1 page?

[Jarek Poplawski]

On 11-10-2006 01:26, Ronghua Zhang wrote:
...
> The reason I asked this is that I saw the following code in forthdeth 
> drvier:
> 
> #define NV_TX2_TSO_MAX_SHIFT) 14
> /* add fragments to entries count */
> for (i = 0; i < fragments; i++) {
>    entries += (skb_shinfo(skb)->frags[i].size >> NV_TX2_TSO_MAX_SHIFT) +
>                    ((skb_shinfo(skb)->frags[i].size &
> (NV_TX2_TSO_MAX_SIZE-1)) ? 1 : 0);
> }
> 
> This looks unnecessary if each frag is guaranteed not to span pages.

Even if PAGE_SIZE > NV_TX2_TSO_MAX_SIZE?

Jarek P.