Re: [PATCH] SIW: iWARP Protocol headers

David Dillow [off-list ref] wrote on 10/06/2010 09:31:44 PM:

On Wed, 2010-10-06 at 12:37 -0600, Jason Gunthorpe wrote:

quoted

On Wed, Oct 06, 2010 at 02:22:55PM -0400, David Dillow wrote:

quoted

On Wed, 2010-10-06 at 11:25 -0600, Jason Gunthorpe wrote:

quoted

On Wed, Oct 06, 2010 at 02:52:46PM +0200, Bernard Metzler wrote:
It is actually a little more complicated than just this. I assume

you

quoted

quoted

quoted

are casting the structures over packet payloads? In this case you
have to guarentee alignment (or used packed everywhere). Does iwarp
have provisions for alignment? If so you can construct your

bitfields

quoted

quoted

quoted

using the alignment type, ie if iWarp guarantees 4 bytes then the
biggest type you can use is u32 - then you can avoid using packed.

Mind you, I'm not sure how to guarentee alignment when you consider
all the possible sources of unalignment in the stack: TCP, IP,L2

stack ?

quoted

quoted

You don't have to. The TCP stack, IIRC, requires the architecture to

fix

quoted

quoted

up misaligned accesses, or at least it used to. It will try to keep
things aligned if possible -- see the comment above NET_IP_ALIGN in
include/linux/skbuff.h.

I was under the impression that this was just to align the IP
header.

It is, and it assumes an ethernet header of 14 bytes. As you note, VLANs
and everything else can mess with that, which is why the TCP stack
requires the architecture to handle unaligned accesses. It's a bonus if
iWarp can guarantee an alignment WRT the IP or TCP header, but not
required.

quoted

quoted

The structures in Bernard's header have all fields naturally
aligned, so no need for the packed attribute and its baggage.

No, they arent:

quoted

+struct iwarp_rdma_write {
+   struct iwarp_ctrl   ctrl;
+   __u32         sink_stag;
+   __u64         sink_to;
+} __attribute__((__packed__));

For instance has sink_to unaligned if the maximum aligment of the
payload is only guarenteed to be 4. The proper thing to do in this
case is probably __attribute__((__packed__, __aligned__(4))) if that
works the way I think.. :)

struct iwarp_ctrl consists of 2 u16s, so sink_stag is aligned to 4 bytes
from the start of the struct iwarp_rdma_write, and sink_to is at 8 bytes
-- so they are naturally aligned with respect to the structure. The
compiler will not add padding, so no need for the packed attribute.

Note I'm not talking about alignment in memory -- that'll depend on
where the data starts, and as I mentioned, the stack doesn't make any
guarantees about that, relying on the architecture to fixup any
misaligned accesses.

agreed. all n-bytes are starting on a multipe of n-byte offset
relative to the start of the structure.

many thanks,
bernard.

--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html