4 byte mem alignment

4 byte mem alignment

[Arvind Gopalan]

I wished someone to confirm the following: Is it a
requirement for the user space buffer that is provided
to the recvfrom() socket call, be 4 byte aligned?.
Note that sockets are raw sockets (PF_PACKET type). I
am running linux 2.4.18-10 on a dual xeon P4 (x86)
box. At the lowest level, i think this translates to
how strong the requirements are for copy_to_user().
does it fault to byte-by-byte mode gracefully when
given a non-4byte aligned buffer?.

Please include me in all replies, since i'm not
subscribed to the list

Thanks
-Arvind

__________________________________________________
Do you Yahoo!?
New DSL Internet Access from SBC & Yahoo!
http://sbc.yahoo.com

Re: 4 byte mem alignment

[David S. Miller]

   From: Arvind Gopalan <arvind_gopalan@yahoo.com>
   Date: Sat, 28 Sep 2002 23:26:46 -0700 (PDT)

   how strong the requirements are for copy_to_user().
   does it fault to byte-by-byte mode gracefully when
   given a non-4byte aligned buffer?.

The x86 processor handles unaligned memory accesses in hw.

On any platform, copy_to_user() must handle any user and kernel buffer
alignment.

Re: 4 byte mem alignment

[Arvind Gopalan]

Thanks Dave, for confirming my doubt. I have some user
level code that uses raw sockets on
linux. The code ran on several dual PIII x86 machines
without any issues. But since the time we started
running (the same code with no changes) on the dual
x86 xeons, i have seen some wierd corruption issues
with the packets. This kernel already has the patch
for the xeon TLB flush bug. On some of the eth packets
that I dumped, i found that 8 bytes of data being
"copied and inserted"!. As in after a certain offset
into the packet, there would be a sequence of 8 bytes
that would be an exact copy of the 8 bytes that
appeared just before. All other following bytes were
literally shifted down by 8. How can i proceed to
debug this sort of a problem. if its a xeon specific
issue, how can i isolate it?. 

Thanks
-Arvind 

--- "David S. Miller" <davem@redhat.com> wrote:
>    From: Arvind Gopalan <arvind_gopalan@yahoo.com>
>    Date: Sat, 28 Sep 2002 23:26:46 -0700 (PDT)
> 
>    how strong the requirements are for
> copy_to_user().
>    does it fault to byte-by-byte mode gracefully
> when
>    given a non-4byte aligned buffer?.
> 
> The x86 processor handles unaligned memory accesses
> in hw.
> 
> On any platform, copy_to_user() must handle any user
> and kernel buffer
> alignment.


__________________________________________________
Do you Yahoo!?
New DSL Internet Access from SBC & Yahoo!
http://sbc.yahoo.com