[PATCH] Wrong calculation in as_path_getlen, and check_aspa improvement suggestion

[Evann DREUMONT]

Hello!

Yes, I can confirm that, especially with the new RFC 9774 standard, see https://datatracker.ietf.org/doc/html/rfc9774#section-4.
That something we address in the second patch.

Originally, the segfault was handled by returning ASPA_INVALID on AS_PATH that contained an AS\_SET, but we discovered that this was not the real problem, but rather a bad allocation due to an incorrect calculation of the AS path length. We mitigated this issue with our first patch.

Best,
--
Evann & Alarig

On 8/31/25 8:02 PM, Alexander Zubkov via Bird-users wrote:

> Hi,
>
> AFAIK, ASPA RFC forbid AS sets and considers such announces invalid:
>
> https://www.ietf.org/archive/id/draft-ietf-sidrops-aspa-verification-22.html#section-6.2-3.3.1
>
>> If the AS_PATH has an AS_SET, then the procedure halts with the outcome "Invalid".
>
> Regards,
> Alexander
>
> On Sun, Aug 31, 2025, 18:11 Alarig Le Lay via Bird-users <bird-users at network.cz> wrote:
>
>> Hello,
>>
>> We (Evann and I) found a bug related to as_path_getlen() when used by
>> aspa_check(). When a route contains an AS_SET segment type, the length
>> returned by as_path_getlen() is incorrect. The function assumes that the
>> length of an AS_PATH_SET is a single AS (1), while in reality an
>> AS_PATH_SET is an unordered set of ASN (as described here
>> https://www.rfc-editor.org/rfc/rfc4271#section-9.2.2.1).
>>
>> See the following update:
>> 2025-08-31 15:35:15.134 <INFO> Checking prefix 76.165.0.0/16 (path 208627 29075 174 32440 {2055 10349 22985 23207 23294 23366 26002 26303 26333 30564 54529 396992 401290}) IN from bgp_alarig_ipv4
>>
>> Using gdb we can inspect the path object:
>> (gdb) p path->length
>> $101 = 72
>> (gdb) x /72xb path->data
>> 0x555555725e54: 0x02 0x04 0x00 0x03 0x2e 0xf3 0x00 0x00
>> 0x555555725e5c: 0x71 0x93 0x00 0x00 0x00 0xae 0x00 0x00
>> 0x555555725e64: 0x7e 0xb8 0x01 0x0d 0x00 0x00 0x08 0x07
>> 0x555555725e6c: 0x00 0x00 0x28 0x6d 0x00 0x00 0x59 0xc9
>> 0x555555725e74: 0x00 0x00 0x5a 0xa7 0x00 0x00 0x5a 0xfe
>> 0x555555725e7c: 0x00 0x00 0x5b 0x46 0x00 0x00 0x65 0x92
>> 0x555555725e84: 0x00 0x00 0x66 0xbf 0x00 0x00 0x66 0xdd
>> 0x555555725e8c: 0x00 0x00 0x77 0x64 0x00 0x00 0xd5 0x01
>> 0x555555725e94: 0x00 0x06 0x0e 0xc0 0x00 0x06 0x1f 0x8a
>>
>> In this example, we have a route with an AS_PATH that contain:
>> - an AS_PATH_SEQUENCE (0x02) with a length of 4 (0x04): (208627
>> 29075 174 32440);
>> - an AS_PATH_SET (0x01) with a length of 13 (0x0d): {2055 10349
>> 22985 23207 23294 23366 26002 26303 26333 30564 54529 396992
>> 401290}.
>> The total length of this update is then 17, but if we dump the function
>> result, we can see that the actual computed length is 5 (4 + 1 for the
>> AS_PATH_SET).
>> (gdb) p len
>> $103 = 5
>>
>> This leads to a too small memory allocation, when normalizing the AS
>> Path in aspa_check():
>> /* Normalize the AS Path: drop stuffings */
>> u32 *asns = alloca(sizeof(u32) * len);
>> Causing a SEGFAULT during the as path walk. Since as_path_walk()
>> considers each element of the AS_PATH_SET as a step. In the while
>> (as_path_walk(path, &ppos, &asns[nsz])), the asns object should have a
>> size of 17 and not 5 resulting in overwriting memory and finally
>> triggering a SEGFAULT. (However we've seen this working when the AS_SET
>> is small, for example, it's working for the following route, but this is
>> mostly luck and could lead to weird behaviors):
>> Checking prefix 104.141.0.0/16 (path 208627 29075 174 32440 {400943}) IN from bgp_alarig_ipv4
>>
>> Here is the gdb output showing this behaviour:
>> 2025-08-31 15:35:15.134 <TRACE> bgp_alarig_ipv4: Got UPDATE
>> 2025-08-31 15:35:15.134 <INFO> Checking prefix 76.165.0.0/16 (path 208627 29075 174 32440 {2055 10349 22985 23207 23294 23366 26002 26303 26333 30564 54529 396992 401290}) IN from bgp_alarig_ipv4
>>
>> Thread 1 "bird" received signal SIGSEGV, Segmentation fault.
>> 0x00005555555d68ac in as_path_walk (path=0x5555000066dd, pos=0x7fffffffd15c,
>> val=0x7fffffffd144) at nest/a-path.c:702
>> 702 const u8 *q = p + path->length;
>> (gdb) p path->data
>> $1 = 0x5555000066e1 <error: Cannot access memory at address 0x5555000066e1>
>>
>> And here is a dump of asns just before the segfault :
>> (gdb) p *asns at nsz+1
>> $57 = {208627, 29075, 174, 32440, 2055, 10349, 22985, 23207, 23294, 23366, 26002, 26303,
>> 26333}
>>
>> We propose to set the AS_PATH_SET length to the announced length in the
>> AS_PATH data instead of 1, see
>> 0001-NEST-correct-as_path-len-calculation.patch.
>>
>> Furthermore, as per
>> https://datatracker.ietf.org/doc/html/rfc9774#name-updates-to-the-requirements
>> (BGP speakers MUST use the "treat-as-withdraw" error handling behavior
>> per [RFC7606] upon reception of BGP UPDATE messages containing AS_SETs
>> or AS_CONFED_SETs in the AS_PATH or AS4_PATH [RFC6793]) and even if
>> https://datatracker.ietf.org/doc/html/draft-ietf-sidrops-aspa-verification-22#name-as_path-verification
>> changes it to a SHOULD, another improvement we propose is to check for
>> AS_PATH_SET the same way it’s already done for AS_PATH_CONFED_SEQUENCE
>> and AS_PATH_CONFED_SET at the beginning of the aspa_check() (see
>> 0002-NEST-return-ASPA_INVALID-for-path-containing-AS_SET.patch). The
>> proposed patch is only for ASPA, not for ROV, in order to avoid dropping
>> routes for too much people, and the patch only drop a few amounts of
>> routes (including a few routes dropped for invalid ROV) :
>> Routes: 1031692 imported, 212 filtered, 0 exported, 1031692 preferred
>>
>> Don’t hesitate to discuss the patch if needed,
>> --
>> Alarig and Evann
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://trubka.network.cz/pipermail/bird-users/attachments/20250831/eae3fb18/attachment.htm>