Bug Report: Unaligned Access in BGP Code on ARMv7 Platforms

[nick]

Dear all,

Apologies for the earlier confusion—my toolchain was misconfigured, and 
I mistakenly compiled the unpatched BIRD sources. Upon correcting this, 
I believe the first patch already addresses the issue. However, I 
noticed that the two patches result in different alignments. Below is a 
summary of the observed alignments:

*No Patch:*
|ALIGN: 8 4 4 4 8|

*First Patch (fix-net-addr-alignment.patch):*
|ALIGN: 2 4 4 4 4|

*Last Patch (net-addr-u32-align.patch):*
|ALIGN: 4 4 4 4 4|

Bests,
Nick

On 12/12/24 2:31 AM, Ondrej Zajicek wrote:
> On Thu, Dec 12, 2024 at 12:50:31AM +0100, nick wrote:
>> Can you give me a quick guide on how to enable the debug output?
>> I tried for example setting in the config file:
>>
>>> log "/tmp/birddebug.log" all;
>>> debug protocols all;
>> I am not able to see any output from ALIGN.
> As the log message is generated even before the config file is parsed, it
> uses the default log output, which is stderr and syslog (if available):
>
> # ./bird -l
> bird: ALIGN: 4 4 4 8 8
>
> # grep ALIGN /var/log/syslog
> 2024-12-12T02:11:59.513900+01:00 feanor bird: ALIGN: 4 4 4 8 8
>
> (also, you could use option -D <debugfile>, in which case the default log
> output would be directed there.)
>
>
> Alternatively, if logging to the usual configured logfile would be more
> practical to you, the test call could be moved from bgp_build() to
> bgp_init() (attached log-alignment-v2.patch), which is called after
> everything is initialized and log is directed to the configured file.
>
>
>
>> On 12/11/24 4:40 AM, Ondrej Zajicek wrote:
>>> On Wed, Dec 11, 2024 at 02:14:39AM +0100, nick wrote:
>>>> Thank you! I just ran a quick test and encountered the same crash at the
>>>> same line. I’ll have more time to investigate tomorrow and can provide
>>>> additional details then. Do you have any other ideas I could try in the
>>>> meantime?
>>> A part of my previous explanation was wrong, the issue is not related to
>>> net_copy(), as the crash happened even before.
>>>
>>> For some reason, struct bgp_prefix is u64-aligned. That makes sense in
>>> the original code, where u64 alignment is forced by struct net_addr, but
>>> i have no idea why this is true even after the patch removing the forced
>>> alignment.
>>>
>>> The slab allocator in sl_alloc() just returns u32-aligned memory, because
>>> it is configured for sizeof(struct bgp_prefix) + sizeof(net_addr_ip6) and
>>> the second is 20 as net_addr_ip6 is u32-aligned. The sl_alloc() internally
>>> only enforces word-size alignment, wich is u32 in your case.
>>>
>>> Please disregard the previous patch. Can you try the first attached patch
>>> (log-alignment.patch) to log alignment of different structures with or
>>> without the second patch (net-addr-u32-align.patch) and report me results?
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://trubka.network.cz/pipermail/bird-users/attachments/20241216/37dfd265/attachment.htm>