Re: [Xen-devel] [PATCH 1/4] xen/netback: implements persistent grant with one page pool.

[Ian Campbell <Ian.Campbell@xxxxxxxxxx>]

On Fri, 2012-11-16 at 02:18 +0000, ANNIE LI wrote:
> In this patch,
> The maximum of memory overhead is about
> 
> (XEN_NETIF_TX_RING_SIZE+XEN_NETIF_RX_RING_SIZE)*PAGE_SIZE  (plus size of 
> grant_ref_t and handle)
> which is about 512 PAGE_SIZE. Normally, without heavy network offload, this 
> maximum can not be reached.
> 
> In next patch of splitting tx/rx pool, the maximum is about

"about" or just "is"?

>  (256+512)PAGE_SIZE.

IOW 3MB.

> >
> >> +
> >> +       return NULL;
> >> +}
> >> +
> >> @@ -1338,7 +1497,11 @@ static unsigned xen_netbk_tx_build_gops(struct 
> >> xen_netbk *netbk)
> >>                  gop->source.domid = vif->domid;
> >>                  gop->source.offset = txreq.offset;
> >>
> >> -               gop->dest.u.gmfn = virt_to_mfn(page_address(page));
> >> +               if (!vif->persistent_grant)
> >> +                       gop->dest.u.gmfn = virt_to_mfn(page_address(page));
> >> +               else
> >> +                       gop->dest.u.gmfn = (unsigned 
> >> long)page_address(page);
> > page_address doesn't return any sort of frame number, does it? This is
> > rather confusing...
> 
> Yes. I only use dest.u.gmfn element to save the page_address here for 
> future memcpy, and it does not mean to use frame number actually. To 
> avoid confusion, here I can use
> 
> gop->dest.u.gmfn = virt_to_mfn(page_address(page));
> 
> and then call mfn_to_virt when doing memcpy.

It seems a bit odd to be using the gop structure in this way when you
aren't actually doing a grant op on it. 

While investigating I noticed:
+static int
+grant_memory_copy_op(unsigned int cmd, void *vuop, unsigned int count,
+                    struct xen_netbk *netbk, bool tx_pool)
...
+       struct gnttab_copy *uop = vuop;

Why void *vuop? Why not struct gnttab_copy * in the parameter?

I also noticed your new grant_memory_copy_op() seems to have unbatched
the grant ops in the non-persistent case, which is going to suck for
performance in non-persistent mode. You need to pull the conditional and
the HYPERVISOR_grant_table_op outside the loop and pass it full array
instead of doing them one at a time.

Ian


_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
http://lists.xen.org/xen-devel