Message-ID: <1395924899.22909.120.camel@kazak.uk.xensource.com>
Date:	Thu, 27 Mar 2014 12:54:59 +0000
From:	Ian Campbell <Ian.Campbell@...rix.com>
To:	Zoltan Kiss <zoltan.kiss@...rix.com>
CC:	<wei.liu2@...rix.com>, <xen-devel@...ts.xenproject.org>,
	<paul.durrant@...rix.com>, <netdev@...r.kernel.org>,
	<linux-kernel@...r.kernel.org>, <jonathan.davies@...rix.com>
Subject: Re: [PATCH net-next] xen-netback: Grant copy the header instead of
 map and memcpy

On Thu, 2014-03-27 at 12:46 +0000, Zoltan Kiss wrote:
> On 27/03/14 11:35, Ian Campbell wrote:
> > On Wed, 2014-03-26 at 21:18 +0000, Zoltan Kiss wrote:
> >> An old inefficiency of the TX path that we are grant mapping the first slot,
> >> and then copy the header part to the linear area. Instead, doing a grant copy
> >> for that header straight on is more reasonable. Especially because there are
> >> ongoing efforts to make Xen avoiding TLB flush after unmap when the page were
> >> not touched in Dom0. In the original way the memcpy ruined that.
> >> The key changes:
> >> - the vif has a tx_copy_ops array again
> >> - xenvif_tx_build_gops sets up the grant copy operations
> >> - we don't have to figure out whether the header and first frag are on the same
> >>    grant mapped page or not
> >>
> >> Unrelated, but I've made a small refactoring in xenvif_get_extras as well.
> >
> > Just a few thoughts, not really based on close review of the code.
> >
> > Do you have any measurements for this series or workloads where it is
> > particularly beneficial?
> My manual measurements showed that if I also use the Xen patch which 
> avoids TLB flush when the pages were not touched, I can easily get 9 - 
> 9.3 Gbps with iperf. But I'll run some more automated tests.

I suppose these numbers are related because avoiding the memcpy avoids
touching the page?

> >
> > You've added a second hypercall to tx_action, probably those can be
> > combined into one vm exit by using a multicall.
> Yes, good idea!
> 
> > Also you should omit
> > either if their respective nr_?ops is zero (can nr_cops ever be zero?)
> If nr_cops is zero, nr_gops should be too, but it can happen that we 
> have only copy ops, but no grant ops (all the packets processed are < 
> PKT_PROT_LEN).

That sounds like what I would expect, good.

> Actually there is a bug in tx_action, it shouldn't return when (nr_cops 
> != 0) && (nr_gops == 0)
> 
> >
> > Adding another MAX_PENDING_REQS sized array is unfortunate. Could we
> > perhaps get away with only ever doing copy for the first N requests (for
> > small N) in a frame and copying up the request, N should be chosen to be
> > large enough to cover the more common cases (which I suppose is Windows
> > which puts the network and transport headers in separate slots). This
> > might allow the copy array to be smaller, at the expense of still doing
> > the map+memcpy for some corner cases.
> N is already 1, we already copy only the first slot. But here we should 
> be prepared for the unlikely situation that a guest sends 256 packets 
> suddenly, all with one slot.

Right, I was wondering if we could just do smaller batches in that case,
but I expect that would be intolerably sucky.

> >
> > Or (and I confess this is a skanky hack): we overlay tx_copy_ops and
> > tx_map_ops in a union, and insert one set of ops from the front and the
> > other from the end, taking great care around when and where they meet.
> Unfortunately the hypercalls expect an array of _one_ of them, so we 
> can't put the 2 types into an union unless it's guaranteed that they 
> have the same size. And they are expected to be

(was there more of this last sentence?)

I was thinking 
	union {
		struct gnt_map maps[NR_...];
		struct gnt_copy copy[NR_...];
	}

Then you fill copy from 0..N and maps from M..NR_ and maintain the
invariant that
         (&maps[M] - &map[s0]) > &copy[N]

and pass to the grant ops (&copy[0], N) and (&maps[M], NR_... - M)
respectively.

Too tricksy perhaps?

> 
> >
> >
> >>   static int xenvif_tx_check_gop(struct xenvif *vif,
> >>   			       struct sk_buff *skb,
> >> -			       struct gnttab_map_grant_ref **gopp)
> >> +			       struct gnttab_map_grant_ref **gopp,
> >> +			       struct gnttab_copy **gopp_copy)
> >
> > I think a precursor patch which only does s/gopp/gopp_map/ would be
> > beneficial.
> It's 5 renaming in 4 hunks, I think it's enough to do them in the same patch

If you think it won't impact the readability of the meat of the patch
then go for it, but it seems trivial to do it first and avoid the risk
that I might think otherwise.


> >
> >>   	struct gnttab_map_grant_ref *gop = vif->tx_map_ops, *request_gop;
> >>   	struct sk_buff *skb;
> >> @@ -1267,24 +1252,37 @@ static unsigned xenvif_tx_build_gops(struct xenvif *vif, int budget)
> >>   			}
> >>   		}
> >>
> >> -		xenvif_tx_create_gop(vif, pending_idx, &txreq, gop);
> >> -
> >> -		gop++;
> >> -
> >>   		XENVIF_TX_CB(skb)->pending_idx = pending_idx;
> >>
> >>   		__skb_put(skb, data_len);
> >
> > Can't you allocate the skb with sufficient headroom? (or maybe I've
> > forgotten again how skb payload management works and __skb_put is
> > effectively free on an empty skb?)
> It effectively does:
> 
> 	skb->tail += len;
> 	skb->len  += len;
> 
> So we can copy data_len between head and tail. The skb is already 
> allocated in xenvif_alloc_skb with a buffer size NET_SKB_PAD + 
> NET_IP_ALIGN + data_len, and and the headroom is NET_SKB_PAD + NET_IP_ALIGN.
> 

Great.

> >
> >> +		vif->tx_copy_ops[*copy_ops].source.u.ref = txreq.gref;
> >> +		vif->tx_copy_ops[*copy_ops].source.domid = vif->domid;
> >> +		vif->tx_copy_ops[*copy_ops].source.offset = txreq.offset;
> >> +
> >> +		vif->tx_copy_ops[*copy_ops].dest.u.gmfn = virt_to_mfn(skb->data);
> >> +		vif->tx_copy_ops[*copy_ops].dest.domid = DOMID_SELF;
> >> +		vif->tx_copy_ops[*copy_ops].dest.offset = offset_in_page(skb->data);
> >> +
> >> +		vif->tx_copy_ops[*copy_ops].len = data_len;
> >
> > Do we want to copy the entire first frag even if it is e.g. a complete
> > page?
> >
> > I'm not sure where the tradeoff lies between doing a grant copy of more
> > than necessary and doing a map+memcpy when the map is already required
> > for the page frag anyway.
> I think we should only grant copy the parts which goes to the linear 
> part, because we would memcpy it anyway. It's expected that the stack 
> won't touch anything else while the packet passes through. data_len is 
> the size of the linear buffer here.
> >
> > What about the case where the first frag is less than PKT_PROT_LEN? I
> > think you still do map+memcpy in that case?
> Then data_len will be txreq.size, we allocate the skb for that, and 
> later on we call __pskb_pull_tail in xenvif_tx_submit to pull up for 
> that (which is essentially map+memcpy). It's not optimal, but it's like 
> that since the good old days. I agree it could be optimized, but let's 
> change it in a different patch!

OK. Can you clarify the title and/or the commit log to make it obvious
that we only copy (some portion of) the first frag and that we still map
+memcpy the remainder of PKT_PROT_LEN if the first frag is not large
enough.

Thanks,
Ian.

--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Powered by blists - more mailing lists