ARM network trouble after recent mbuf changes

[Michael Tuexen]

On Aug 27, 2013, at 8:53 AM, Andre Oppermann <andre at FreeBSD.org> wrote:

> On 27.08.2013 08:30, Michael Tuexen wrote:
>> On Aug 27, 2013, at 7:53 AM, Hans Petter Selasky <hps at bitfrost.no> wrote:
>> 
>>> On 08/27/13 00:38, Michael Tuexen wrote:
>>>> I did some tests with a small program. Having in struct pkthdr 64 bit entities
>>>> results in a 64 bit alignment when used in struct mbuf. Using __packed
>>>> for struct mbuf, removes the padding.
>>> 
>>> 
>>> Hi,
>>> 
>>> Maybe you could use __aligned(8) instead, and account for the extra padding on all platforms? Packed has other disadvantages on ARM platforms when accessing the structures, like that non-aligned access is possible, and that it is sometimes slower than aligned access.
>> Isn't there a performance penalty when accessing 64-bit entities not being 64-bit
>> aligned? If that is the case, wouldn't it make sense to add a 4 byte padding to
>> struct m_hdr for ILP32? Then the problem should go away...
> 
> Either that or define MLEN and MHLEN in a way that actually reflects the true
> size of what they are representing.  The latter is the true bug.
Correct. There is the hidden assumption that there is no padding. Maybe you can
put that in a comment...
We could also have some code (maybe under INVARIANTS) where we check that
an struct mbuf has the size MSIZE and panic if not.
This would make things clearer if the problem happens again.
> 
>> We could also get rid of the 64 bit alignment by not having 64-bit entities in
>> struct pkthdr. Removing sixtyfour should be easy. However, we now have also
>> uint64_t csum_flags.
> 
> Yes, the 64 bit fields are to be used to store packet associated information
> on its way through the stack.  Reducing it to 32 bits would somewhat defeat
> their purpose.
I completely agree...
> 
> -- 
> Andre
> 
>