On Fri, Jul 29, 2005 at 11:58:57PM +0400, Evgeniy Polyakov ([EMAIL PROTECTED])
wrote:
> On Fri, Jul 29, 2005 at 12:43:34PM -0700, David S. Miller ([EMAIL PROTECTED])
> wrote:
> > From: Evgeniy Polyakov <[EMAIL PROTECTED]>
> > Date: Fri, 29 Jul 2005 20:55:06 +0400
> >
> > > Unmapping is repeatedly being done in this driver - update_address() and
> > > tlb_flush() does the thing.
> > > Current code does skb freeing when it's slot needs to be used for the
> > > next skb.
> > > So this numbers already include mapping/unmapping.
> >
> > BTW, this reminds me now that if we put this kind of change in
> > we should work to resolve the issues that prevent packet
> > mmap() from working properly on virtual cache architectures.
> >
> > In fact, I am to understand that currently only x86 even
> > tries to use packet mmap in the existing userland implementaitons.
> > I could be wrong, as I haven't looked at the pcap library stuff
> > in a long time.
> >
> > If there is cache coloring conflict, kernel stores to physical
> > locations may not show up in userspace correct, thus D-cache
> > flushing or proper coloring of userland addresses is necessary.
> > The latter is cheaper due to lack of D-cache flushing, obviously.
>
> Code has d-cache flushing, but it is nop on x86, so it's affect can not
> be measured in my environment. I wish I had more time to port 2.6 to one
> of custom 440 ppc board.
>
> > Another thing missing are proper memory barriers, and some portable
> > way for a packet_mmap userland code to know what barriers to use
> > in order to synchronize with kernel updates to the packet mmap
> > ring buffer area.
>
> Yes, it is completely missed in this module.
Writing to the shared memory only happens when control page is updated,
i.e. when new skb is mapped and new offset is written and "ps" is updated,
so this patch should be enough:
--- af_tlb.c~ 2005-07-29 21:01:56.000000000 +0400
+++ af_tlb.c 2005-08-01 16:39:16.000000000 +0400
@@ -558,6 +558,8 @@
if (++po->last == po->budget)
po->last = 0;
+ wmb();
+
{
start = po->vma->vm_start;
--
Evgeniy Polyakov
-
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to [EMAIL PROTECTED]
More majordomo info at http://vger.kernel.org/majordomo-info.html
