Re: [MirageOS-devel] wireshark capture of failed download from mirage-www on ARM

[David Scott <scott.dj@xxxxxxxxx>]

On Saturday, July 26, 2014, Anil Madhavapeddy <anil@xxxxxxxxxx> wrote:
> On 25 Jul 2014, at 04:54, Thomas Leonard <talex5@xxxxxxxxx> wrote:
> >
> > Here's a slightly better work-around:
> >
> > https://github.com/talex5/mirage-tcpip/commit/9f7e0a628ebd1cece7d0fb979273a96be78233ad
> >
> > It changes the TCP packet splitting code so that the second part of
> > the page is copied to a new IO page. This avoids sending the same
> > physical page to Linux twice (which is what doesn't work).
> >
> > With this, the previous hack can be reverted.
> >
> > In my testing, this increased the streaming TCP download speed from 56
> > KB/s to 3.9 MB/s on the CubieTruck.
>
> This is certainly a fix that works, but I'm loath to insert an unnecessary
> copy in the TCP stack just because of a limitation in Xen/ARM (that will go
> away in Xen 4.5, I expect).
>
> Could we keep track of the active grants in the Gnttab libary, and force a
> page duplication at that point instead? ÂThat would fix it for any other
> Cstruct consumers as well, and be a more logical place to insert Xen
> version detection in the future to remove the extra copy if not needed.

If the problem is granting the same page twice, wouldÂit be enough to maintain a map of page address to (grant index * refount), and then "grant_access" could return the existing grant id and bump the refcount? I think we just need to share the grant, rather than duplicate the page. (Is that right?)

Cheers,

Dave

Â

> -anil
> _______________________________________________
> MirageOS-devel mailing list
> MirageOS-devel@xxxxxxxxxxxxxxxxxxxx
> http://lists.xenproject.org/cgi-bin/mailman/listinfo/mirageos-devel


-- 
Dave Scott
_______________________________________________
MirageOS-devel mailing list
MirageOS-devel@xxxxxxxxxxxxxxxxxxxx
http://lists.xenproject.org/cgi-bin/mailman/listinfo/mirageos-devel