Andrii Nakryiko <andrii.nakryiko@xxxxxxxxx> writes:
> On Fri, Feb 11, 2022 at 3:49 PM Toke Høiland-Jørgensen <toke@xxxxxxxxxx> wrote:
>>
>> When receiving netlink messages, libbpf was using a statically allocated
>> stack buffer of 4k bytes. This happened to work fine on systems with a 4k
>> page size, but on systems with larger page sizes it can lead to truncated
>> messages. The user-visible impact of this was that libbpf would insist no
>> XDP program was attached to some interfaces because that bit of the netlink
>> message got chopped off.
>>
>> Fix this by switching to a dynamically allocated buffer; we borrow the
>> approach from iproute2 of using recvmsg() with MSG_PEEK|MSG_TRUNC to get
>> the actual size of the pending message before receiving it, adjusting the
>> buffer as necessary. While we're at it, also add retries on interrupted
>> system calls around the recvmsg() call.
>>
>> v2:
>> - Move peek logic to libbpf_netlink_recv(), don't double free on ENOMEM.
>>
>> Reported-by: Zhiqian Guan <zhguan@xxxxxxxxxx>
>> Fixes: 8bbb77b7c7a2 ("libbpf: Add various netlink helpers")
>> Acked-by: Kumar Kartikeya Dwivedi <memxor@xxxxxxxxx>
>> Signed-off-by: Toke Høiland-Jørgensen <toke@xxxxxxxxxx>
>> ---
>
> Applied to bpf-next.
Awesome, thanks!
> One improvement would be to avoid initial malloc of 4096, especially
> if that size is enough for most cases. You could detect this through
> iov.iov_base == buf and not free(iov.iov_base) at the end. Seems
> reliable and simple enough. I'll leave it up to you to follow up, if
> you think it's a good idea.
Hmm, seems distributions tend to default the stack size limit to 8k; so
not sure if blowing half of that on a buffer just to avoid a call to
malloc() in a non-performance-sensitive is ideal to begin with? I think
I'd prefer to just keep the dynamic allocation...
-Toke
