bpf: add own ctx rewriter on ifindex for clsact progs
authorDaniel Borkmann <daniel@iogearbox.net>
Fri, 9 Sep 2016 00:45:30 +0000 (02:45 +0200)
committerDavid S. Miller <davem@davemloft.net>
Sat, 10 Sep 2016 02:36:04 +0000 (19:36 -0700)
When fetching ifindex, we don't need to test dev for being NULL since
we're always guaranteed to have a valid dev for clsact programs. Thus,
avoid this test in fast path.

Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
net/core/filter.c

index 120c813..d6d9bb8 100644 (file)
@@ -2634,10 +2634,10 @@ void bpf_warn_invalid_xdp_action(u32 act)
 }
 EXPORT_SYMBOL_GPL(bpf_warn_invalid_xdp_action);
 
-static u32 bpf_net_convert_ctx_access(enum bpf_access_type type, int dst_reg,
-                                     int src_reg, int ctx_off,
-                                     struct bpf_insn *insn_buf,
-                                     struct bpf_prog *prog)
+static u32 sk_filter_convert_ctx_access(enum bpf_access_type type, int dst_reg,
+                                       int src_reg, int ctx_off,
+                                       struct bpf_insn *insn_buf,
+                                       struct bpf_prog *prog)
 {
        struct bpf_insn *insn = insn_buf;
 
@@ -2785,6 +2785,31 @@ static u32 bpf_net_convert_ctx_access(enum bpf_access_type type, int dst_reg,
        return insn - insn_buf;
 }
 
+static u32 tc_cls_act_convert_ctx_access(enum bpf_access_type type, int dst_reg,
+                                        int src_reg, int ctx_off,
+                                        struct bpf_insn *insn_buf,
+                                        struct bpf_prog *prog)
+{
+       struct bpf_insn *insn = insn_buf;
+
+       switch (ctx_off) {
+       case offsetof(struct __sk_buff, ifindex):
+               BUILD_BUG_ON(FIELD_SIZEOF(struct net_device, ifindex) != 4);
+
+               *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct sk_buff, dev),
+                                     dst_reg, src_reg,
+                                     offsetof(struct sk_buff, dev));
+               *insn++ = BPF_LDX_MEM(BPF_W, dst_reg, dst_reg,
+                                     offsetof(struct net_device, ifindex));
+               break;
+       default:
+               return sk_filter_convert_ctx_access(type, dst_reg, src_reg,
+                                                   ctx_off, insn_buf, prog);
+       }
+
+       return insn - insn_buf;
+}
+
 static u32 xdp_convert_ctx_access(enum bpf_access_type type, int dst_reg,
                                  int src_reg, int ctx_off,
                                  struct bpf_insn *insn_buf,
@@ -2811,13 +2836,13 @@ static u32 xdp_convert_ctx_access(enum bpf_access_type type, int dst_reg,
 static const struct bpf_verifier_ops sk_filter_ops = {
        .get_func_proto         = sk_filter_func_proto,
        .is_valid_access        = sk_filter_is_valid_access,
-       .convert_ctx_access     = bpf_net_convert_ctx_access,
+       .convert_ctx_access     = sk_filter_convert_ctx_access,
 };
 
 static const struct bpf_verifier_ops tc_cls_act_ops = {
        .get_func_proto         = tc_cls_act_func_proto,
        .is_valid_access        = tc_cls_act_is_valid_access,
-       .convert_ctx_access     = bpf_net_convert_ctx_access,
+       .convert_ctx_access     = tc_cls_act_convert_ctx_access,
 };
 
 static const struct bpf_verifier_ops xdp_ops = {