On Tue, Aug 31, 2021 at 12:54:43AM +0400, Dmitrii Banshchikov wrote:

On Mon, Aug 30, 2021 at 12:45:45PM -0700, Alexei Starovoitov wrote:

quoted

On Sun, Aug 29, 2021 at 10:36:07PM +0400, Dmitrii Banshchikov wrote:

quoted

 /*
- * # Generated by iptables-save v1.8.2 on Sat May  8 05:22:41 2021
+ *  Generated by iptables-save v1.8.2 on Sat May  8 05:22:41 2021
  * *filter

...

quoted

- * -A LOCAL -s 10.32.0.0/11 -j FROMDC
- * -A LOCAL -s 10.144.0.0/12 -j FROMDC
- * -A LOCAL -s 10.160.0.0/12 -j FROMDC
- * -A LOCAL -s 10.0.0.0/12 -j FROMDC
- * -A LOCAL -s 10.248.0.0/24 -j FROMDC
- * -A LOCAL -s 10.232.0.0/16 -j FROMDC
- * -A LOCAL -s 10.1.146.131/32 -p udp -m udp --dport 161 -j ACCEPT
- * -A LOCAL -s 10.149.118.14/32 -p udp -m udp --dport 161 -j ACCEPT
- * -A LOCAL -p icmp -j ACCEPT
+ * :INPUT ACCEPT [0:0]
+ * :FORWARD ACCEPT [0:0]
+ * :OUTPUT ACCEPT [0:0]
+ * -A INPUT -s 1.1.1.1/32 -d 2.2.2.2/32 -j DROP
+ * -A INPUT -s 2.2.0.0/16 -d 3.0.0.0/8 -j DROP
+ * -A INPUT -p udp -m udp --sport 100 --dport 500 -j DROP
  * COMMIT
  */

Patch 10 adds this test, but then patch 12 removes most of it?
Keep both?

Sorry, I missed it.
I decided that the large blob looks really ugly and switched to
the smaller one and forgot to cleanup the patchset.

quoted

Also hit this on my system with older glibc:

../net/bpfilter/codegen.c: In function ‘codegen_push_subprog’:
../net/bpfilter/codegen.c:67:4: warning: implicit declaration of function ‘reallocarray’ [-Wimplicit-function-declaration]
   67 |    reallocarray(codegen->subprogs, subprogs_max, sizeof(codegen->subprogs[0]));
      |    ^~~~~~~~~~~~
../net/bpfilter/codegen.c:66:12: warning: assignment to ‘struct codegen_subprog_desc **’ from ‘int’ makes pointer from integer without a cast [-Wint-conversion]
   66 |   subprogs =
      |            ^

In libbpf we have libbpf_reallocarray() for this reason.

Could you provide an example of generated bpf program?
And maybe add Documentation/bpf/bpfilter_design.rst ?

I will add documentation in the next iteration when
bpf_map_for_each() subprog will be introduced.

quoted

The tests don't build for me:
$ cd selftests/bpf/bpfilter; make
make: *** No rule to make target '-lelf', needed by '.../selftests/bpf/bpfilter/test_match'.  Stop.

libelf was added because libbpf depends on it.
Are you able to build libbpf?

make proceeds to build libbpf just fine, but then it stops with above message.
I manually removed -lelf from Makefile. Then run make to see it fail linking
and then manually copy pasted gcc command to build it with additional -lelf
command line.
fwiw make -v
GNU Make 4.2.1

quoted

The unit tests are great, btw. test_codegen is not end-to-end, right?
Could you add a full test with iptable command line?
or netns support is a prerequisite for it?

Yeah, as net namespaces aren't supported using iptables binary
will modify the root namespace. That is the reason why codegen
tests aren't implemented in the end-to-end fashion and rules are
represented by blobs.

I think when ifindex is no longer hardcoded the netns support
doesn't have to be gating. The generic xdp attached to veth in netns
should work to do end-to-end test. bpftiler would need to do a bit of magic
to figure out the right ifindex. Or we can extend kernel with ifindex-less
generic XDP.