Re: [PATCH] filter: introduce SKF_AD_VLAN_PROTO BPF extension

[Daniel Borkmann <daniel@iogearbox.net>]

On 03/05/2015 05:52 PM, Alexei Starovoitov wrote:
> On 3/5/15 2:37 AM, Michal Sekletar wrote:
>> On Wed, Mar 04, 2015 at 01:03:50PM -0800, Alexei Starovoitov wrote:
>>> On 3/4/15 12:41 PM, Michal Sekletar wrote:
>>>> This commit introduces new BPF extension. It makes possible to load value of
>>>> skb->vlan_proto (vlan tpid) to register A.
>>>>
>>>> Currently, vlan header is removed from frame and information is available to
>>>> userspace only via tpacket interface. Hence, it is not possible to install
>>>> filter which uses value of vlan tpid field.
>>>>
>>>> AFAICT only way how to filter based on tpid value is to reconstruct original
>>>> frame encapsulation and interpret BPF filter code in userspace. Doing that is
>>>> way slower than doing filtering in kernel.
>>>>
>>>> Cc: Alexei Starovoitov <ast@plumgrid.com>
>>>> Cc: Jiri Pirko <jpirko@redhat.com>
>>>> Signed-off-by: Michal Sekletar <msekleta@redhat.com>
>>>> ---
>>>> @@ -282,6 +282,7 @@ Possible BPF extensions are shown in the following table:
>>>>     vlan_tci                              skb_vlan_tag_get(skb)
>>>>     vlan_pr                               skb_vlan_tag_present(skb)
>>>>     rand                                  prandom_u32()
>>>> +  vlan_proto                            skb->vlan_proto
>>>
>>> the patch is correct and looks clean, but I don't understand
>>> the motivation for the patch.
>>
>> Way how libpcap currently uses BPF extensions is not compatible with old
>> behavior where actual value of tpid field was checked. I wanted to address
>> that, i.e. if "vlan" keyword is used as filter expression, libpcap should
>> install a filter such that only ethernet frames having tpid value of 0x8100 or
>> 0x9100 will pass. That is not the case with current libpcap git and 4.0-rc1
>> kernel.
>>
>> Given that I broke libpcap as described above I tried to come up with the way
>> how to fix that. However I realized that with recent kernels there is no other
>> way than adding new BPF extension.
>>
>>> There is already SKF_AD_VLAN_TAG_PRESENT. If it is set then only
>>> two possible values of vlan_proto are ETH_P_8021Q or ETH_P_8021AD.
>>
>> Any reason why ETH_P_QINQ1, ETH_P_QINQ2, ETH_P_QINQ3 no longer works? If I
>> understand correctly, you are basically saying, that there is no point checking
>> for vlan tpid because PF_PACKET socket will never receive frame having other
>> tpid value than above two anyway.
>>
>> So bottom line is that I wanted to grant userspace programs more flexibility,
>> and you are saying that it is pointless because for example if outer tpid is
>> 0x9100 socket will never receive the frame. If that is the case then
>> disregard the patch.
>
> steering towards vlan device happens only for ETH_P_8021Q and
> ETH_P_8021AD. Non-standard 0x9100 and other tags won't be popped into
> skb metadata and will stay as-is in the packet body.
> If the meaning of 'vlan 100' in libpcap is to detect all possible
> vlan tpid then bpf program would need to check VLAN_TAG_PRESENT
> (that would mean vlan_proto is either 0x8100 or 0x88A8)
> and parse packet body for tpids 0x9[123]00.
> Whether we add access to skb->vlan_proto or not, the program would
> still need to do the above steps, but instead of checking for
> vlan_tag_present only, it would need to do vlan_proto==0x8100
> or vlan_proto=0x88a8 and then parse the packet for tpid=0x9[123]00
> so adding access to vlan_proto will not simplify libpcap job.

+1, libpcap would, of course, need to check for both, the offloaded
and non-offloaded case. I'm not sure if you're already doing this or
plan to fix it there, Michal?

> At this point I think it's up to Dave to decide whether we need
> this patch (after fixing the issue pointed by Denis) or not.
> imo there is a benefit of giving programs more visibility into
> skb metadata.

I'm not really a big fan of it, but given we added commit a0cdfcf39362
("packet: deliver VLAN TPID to userspace") to packet sockets ...

Michal, please make sure in v2 that you Cc related JIT people.

In case you're sending v2, please fix up bpf_asm in the following way:

I don't like the vlan_{pr,proto} ... it's confusing, so lets add a
better alias and keep compatibility with vlan_pr. Here's the chunk:

diff --git a/Documentation/networking/filter.txt b/Documentation/networking/filter.txt
index 9930ecfb..d9a2080 100644
--- a/Documentation/networking/filter.txt
+++ b/Documentation/networking/filter.txt
@@ -279,8 +279,9 @@ Possible BPF extensions are shown in the following table:
    hatype                                skb->dev->type
    rxhash                                skb->hash
    cpu                                   raw_smp_processor_id()
+  vlan_avail                            skb_vlan_tag_present(skb)
    vlan_tci                              skb_vlan_tag_get(skb)
-  vlan_pr                               skb_vlan_tag_present(skb)
+  vlan_proto                            skb->vlan_proto
    rand                                  prandom_u32()

  These extensions can also be prefixed with '#'.
diff --git a/tools/net/bpf_exp.l b/tools/net/bpf_exp.l
index 833a966..c1a1a39 100644
--- a/tools/net/bpf_exp.l
+++ b/tools/net/bpf_exp.l
@@ -90,8 +90,10 @@ extern void yyerror(const char *str);
  "#"?("hatype")	{ return K_HATYPE; }
  "#"?("rxhash")	{ return K_RXHASH; }
  "#"?("cpu")	{ return K_CPU; }
-"#"?("vlan_tci") { return K_VLANT; }
  "#"?("vlan_pr")	{ return K_VLANP; }
+"#"?("vlan_avail") { return K_VLANP; }
+"#"?("vlan_tci") { return K_VLANT; }
+"#"?("vlan_proto") { return K_VLANPR; }
  "#"?("rand")	{ return K_RAND; }

  ":"		{ return ':'; }
diff --git a/tools/net/bpf_exp.y b/tools/net/bpf_exp.y
index e6306c5..d4c749c 100644
--- a/tools/net/bpf_exp.y
+++ b/tools/net/bpf_exp.y
@@ -56,7 +56,7 @@ static void bpf_set_jmp_label(char *label, enum jmp_type type);
  %token OP_LDXI

  %token K_PKT_LEN K_PROTO K_TYPE K_NLATTR K_NLATTR_NEST K_MARK K_QUEUE K_HATYPE
-%token K_RXHASH K_CPU K_IFIDX K_VLANT K_VLANP K_POFF K_RAND
+%token K_RXHASH K_CPU K_IFIDX K_VLANT K_VLANP K_VLANPR K_POFF K_RAND

  %token ':' ',' '[' ']' '(' ')' 'x' 'a' '+' 'M' '*' '&' '#' '%'

@@ -167,6 +167,9 @@ ldb
  	| OP_LDB K_RAND {
  		bpf_set_curr_instr(BPF_LD | BPF_B | BPF_ABS, 0, 0,
  				   SKF_AD_OFF + SKF_AD_RANDOM); }
+	| OP_LDB K_VLANPR {
+		bpf_set_curr_instr(BPF_LD | BPF_B | BPF_ABS, 0, 0,
+				   SKF_AD_OFF + SKF_AD_VLAN_PROTO); }
  	;

  ldh
@@ -218,6 +221,9 @@ ldh
  	| OP_LDH K_RAND {
  		bpf_set_curr_instr(BPF_LD | BPF_H | BPF_ABS, 0, 0,
  				   SKF_AD_OFF + SKF_AD_RANDOM); }
+	| OP_LDH K_VLANPR {
+		bpf_set_curr_instr(BPF_LD | BPF_H | BPF_ABS, 0, 0,
+				   SKF_AD_OFF + SKF_AD_VLAN_PROTO); }
  	;

  ldi
@@ -274,6 +280,9 @@ ld
  	| OP_LD K_RAND {
  		bpf_set_curr_instr(BPF_LD | BPF_W | BPF_ABS, 0, 0,
  				   SKF_AD_OFF + SKF_AD_RANDOM); }
+	| OP_LD K_VLANPR {
+		bpf_set_curr_instr(BPF_LD | BPF_W | BPF_ABS, 0, 0,
+				   SKF_AD_OFF + SKF_AD_VLAN_PROTO); }
  	| OP_LD 'M' '[' number ']' {
  		bpf_set_curr_instr(BPF_LD | BPF_MEM, 0, 0, $4); }
  	| OP_LD '[' 'x' '+' number ']' {