Date:	Sun, 7 Oct 2012 14:44:39 +0200
From:	Daniel Borkmann <danborkmann@...earbox.net>
To:	pchavent <Paul.Chavent@...ra.fr>
Cc:	davem@...emloft.net, edumazet@...gle.com, xemul@...allels.com,
	herbert@...dor.hengli.com.au, netdev@...r.kernel.org,
	johann.baudy@...-log.net, uaca@...mni.uv.es
Subject: Re: [PATCH] Packet mmap : allow the user to choose the offset of the
 tx payload.

On Sun, Oct 7, 2012 at 12:50 PM, Daniel Borkmann
<danborkmann@...earbox.net> wrote:
> On Sat, Oct 6, 2012 at 9:43 AM, pchavent <Paul.Chavent@...ra.fr> wrote:
>> On Fri, 5 Oct 2012 21:37:58 +0200, Daniel Borkmann wrote:
>>> On Fri, Oct 5, 2012 at 9:21 PM, pchavent <Paul.Chavent@...ra.fr> wrote:
>>>> On Fri, 5 Oct 2012 16:17:12 +0200, Daniel Borkmann wrote:
>>>>> On Fri, Oct 5, 2012 at 3:10 PM, Paul Chavent <Paul.Chavent@...ra.fr>
>>>>> wrote:
>>>>>>
>>>>>>
>>>>>> The tx offset of packet mmap tx ring used to be :
>>>>>> (TPACKET2_HDRLEN - sizeof(struct sockaddr_ll))
>>>>>>
>>>>>> The problem is that depending on the usage of SOCK_DGRAM or
>>>>>> SOCK_RAW, the payload could be aligned or not.
>>>>>>
>>>>>> This patch allow to let the user give an offset for it's tx
>>>>>> payload if he desires.
>>>>>>
>>>>>> Signed-off-by: Paul Chavent <paul.chavent@...ra.fr>
>>>>>
>>>>>
>>>>> Can you provide an example when it doesn't hit TPACKET_ALIGNMENT?
>>>>
>>>>
>>>> When we use tx ring, the user have to write at (TPACKET_HDRLEN -
>>>> sizeof(struct sockaddr_ll))
>>>>
>>>> This adress is aligned on TPACKET_ALIGNMENT since
>>>> TPACKET_HDRLEN = (TPACKET_ALIGN(sizeof(struct tpacket_hdr)) +
>>>> sizeof(struct
>>>> sockaddr_ll))
>>>>
>>>> When we use the tx ring with SOCK_RAW option, the mac header is aligned
>>>> on
>>>> TPACKET_ALIGNMENT, but not the payload (14 bytes away).
>>>
>>>
>>> Okay, I'm confused about your intentions, maybe I'm missing something.
>>> The man-page of packet(7) clearly says:
>>>
>>> The socket_type is either SOCK_RAW for raw packets *including* the
>>> link level header or SOCK_DGRAM for cooked packets with  the link
>>> level header *removed*.
>>>
>>> So this is perfectly intended behavior of PF_PACKET.
>>>
>>> Cheers,
>>>
>>> Daniel
>>
>>
>> Yes, i also expect to be able to include the link level header when i use
>> SOCK_RAW.
>>
>> My intention is to send a frame with this payload (for example) :
>> typedef struct
>> {
>>   double   ts;
>>   uint64_t foo;
>> } test_t;
>>
>> So i get a pointer to the raw packet :
>> void * raw_packet = frame_base + (TPACKET_HDRLEN - sizeof(struct
>> sockaddr_ll));
>>
>> I cook the header :
>> memcpy(raw_packet +  0, dst_addr, sizeof(dst_addr));
>> memcpy(raw_packet +  6, src_addr, sizeof(src_addr));
>> memcpy(raw_packet + 12, type    , sizeof(type));
>>
>> Then i get a pointer to the beginning of payload :
>> test_t * payload = raw_packet + 14;
>>
>> Here payload is at 58 bytes from the beginning of the frame.
>>
>> Then i fill the payload :
>> payload->ts = 1.0;
>> payload->foo = 2;
>> ...
>>
>> These are misaligned accesses.
>>
>> I don't care to fill the cooked header if it's misaligned, but i would like
>> to be able to fill the frame directly in the ring buffer being on aligned
>> boundaries.

Just a minor remark: speaking about your patch, in the case of
SOCK_RAW I think you rather might want to care whether your *mac
header* starts at an aligned offset or not, since the kernel doesn't
care about your particular payload, but about the frame as a whole
that it needs to process.

> Okay.
>
> Maybe what you could do in a new version of your patch is to introduce
> a TP_STATUS flag, e.g. TP_STATUS_SEND_HAS_OFF that you pass along
> binary or'ed with the commonly used flags, and then you can fill
> tp_mac resp. tp_net with offsets. By that, you won't break legacy
> stuff.

Also, note that you should wait with your submission until net-next is reopened.
--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Powered by blists - more mailing lists