Re: [V2 PATCH net-next 1/2] xdp: implement xdp_redirect_map for generic XDP

On Thu, 07 Sep 2017 16:09:56 +0200
Daniel Borkmann [off-list ref] wrote:

On 09/07/2017 02:33 PM, Jesper Dangaard Brouer wrote:

quoted

Using bpf_redirect_map is allowed for generic XDP programs, but the
appropriate map lookup was never performed in xdp_do_generic_redirect().

Instead the map-index is directly used as the ifindex.  For the
xdp_redirect_map sample in SKB-mode '-S', this resulted in trying
sending on ifindex 0 which isn't valid, resulting in getting SKB
packets dropped.  Thus, the reported performance numbers are wrong in
commit 24251c264798 ("samples/bpf: add option for native and skb mode
for redirect apps") for the 'xdp_redirect_map -S' case.

It might seem innocent this was lacking, but it can actually crash the
kernel.  The potential crash is caused by not consuming redirect_info->map.
The bpf_redirect_map helper will set this_cpu_ptr(&redirect_info)->map
pointer, which will survive even after unloading the xdp bpf_prog and
deallocating the devmap data-structure.  This leaves a dead map
pointer around.  The kernel will crash when loading the xdp_redirect
sample (in native XDP mode) as it doesn't reset map (via bpf_redirect)
and returns XDP_REDIRECT, which will cause it to dereference the map
pointer.

Fixes: 6103aa96ec07 ("net: implement XDP_REDIRECT for xdp generic")
Fixes: 24251c264798 ("samples/bpf: add option for native and skb mode for redirect apps")
Signed-off-by: Jesper Dangaard Brouer <redacted>
---
  include/trace/events/xdp.h |    4 ++--
  net/core/filter.c          |   14 +++++++++++---
  2 files changed, 13 insertions(+), 5 deletions(-)

diff --git a/include/trace/events/xdp.h b/include/trace/events/xdp.h
index 862575ac8da9..4e16c43fba10 100644
--- a/include/trace/events/xdp.h
+++ b/include/trace/events/xdp.h

@@ -138,11 +138,11 @@ DEFINE_EVENT_PRINT(xdp_redirect_template, xdp_redirect_map_err,

  #define _trace_xdp_redirect_map(dev, xdp, fwd, map, idx)		\
  	 trace_xdp_redirect_map(dev, xdp, fwd ? fwd->ifindex : 0,	\
-				0, map, idx);
+				0, map, idx)

  #define _trace_xdp_redirect_map_err(dev, xdp, fwd, map, idx, err)	\
  	 trace_xdp_redirect_map_err(dev, xdp, fwd ? fwd->ifindex : 0,	\
-				    err, map, idx);
+				    err, map, idx)

  #endif /* _TRACE_XDP_H */

diff --git a/net/core/filter.c b/net/core/filter.c
index 5912c738a7b2..3767470cab6c 100644
--- a/net/core/filter.c
+++ b/net/core/filter.c

@@ -2566,13 +2566,19 @@ int xdp_do_generic_redirect(struct net_device *dev, struct sk_buff *skb,
  			    struct bpf_prog *xdp_prog)
  {
  	struct redirect_info *ri = this_cpu_ptr(&redirect_info);
+	struct bpf_map *map = ri->map;
  	u32 index = ri->ifindex;
  	struct net_device *fwd;
  	unsigned int len;
  	int err = 0;

-	fwd = dev_get_by_index_rcu(dev_net(dev), index);
  	ri->ifindex = 0;
+	ri->map = NULL;
+
+	if (map)
+		fwd = __dev_map_lookup_elem(map, index);
+	else
+		fwd = dev_get_by_index_rcu(dev_net(dev), index);
  	if (unlikely(!fwd)) {
  		err = -EINVAL;
  		goto err;

@@ -2590,10 +2596,12 @@ int xdp_do_generic_redirect(struct net_device *dev, struct sk_buff *skb,
  	}

  	skb->dev = fwd;  

Looks much better above, thanks!

quoted

-	_trace_xdp_redirect(dev, xdp_prog, index);
+	map ? _trace_xdp_redirect_map(dev, xdp_prog, fwd, map, index)
+		: _trace_xdp_redirect(dev, xdp_prog, index);  

Could we rather make this in a way such that when the two
tracepoints are disabled and thus patched out, that we can
also omit the extra conditional which has no purpose then?

First of all I don't think it make much of a difference, I measured the
impact of the full patch to "cost" 1.62 nanosec (which is arguably
below the accuracy level of the system under test)

Secondly, I plan to optimize the map case for generic XDP later, where
I would naturally split this into two functions (as V1, and as
native-XDP), thus this extra conditional would go away.  As I've shown
offlist (to you, John and Andy) I demonstrated a 24% speedup via a
xmit_more hack for generic XDP.

Perhaps just a consolidated _trace_xdp_generic_redirect_map()
would be better to avoid this altogether given we have twice
the same anyway, here and in err path.

I do want separate tracepoints for xdp_redirect and xdp_redirect_map,
as it makes it more clear for users of the tracepoint (and attached
bpf_prog's can be faster, knowing the context).

-- 
Best regards,
  Jesper Dangaard Brouer
  MSc.CS, Principal Kernel Engineer at Red Hat
  LinkedIn: http://www.linkedin.com/in/brouer