Herbert Xu wrote:

On Mon, Dec 03, 2007 at 07:14:26PM -0500, Hideo AOKI wrote:

quoted

Let me try to use sk_forward_alloc at first, while percpu storage
is an interesting idea.

Actually I don't think sk_forward_alloc would work for UDP because
it runs lockless (unlike TCP which is run under a the socket lock).

So it's either going to be the atomic op or per-cpu counters.  For
me the atomic op isn't the issue, it's the SMP cache-line bouncing
that's more important so having something that did atomic ops on a
socket counter which then feeds into the global counter would solve
my concerns.

But let's wait and see what Dave has to say about this too.

Hello,

I suppose that he also wants to have per-socket accounting to avoid
global counter access.

To achieve this, I think I have three ideas at present. I'd
appreciate if you let me know acceptable idea or any suggestions.

1. Using sk_forward_alloc and adding socket lock

   UDP already uses a socket lock to send message. However, it doesn't
   use the lock to receive message. I wonder if we can also use the
   lock when sk_forward_alloc is updated in receive processing.
   I understand performance issue might occur, but ...


2. Adding new atomic_t variable for memory accounting

   Datagram protocols will use the variable for per-socket memory
   accounting. Stream protocols continue to use sk_forward_alloc.


3. Replacing current sk_forward_alloc with union type

   Stream protocols use the union as int, and datagram protocols use
   it as atomic_t.

Best regards,
Hideo

-- 
Hitachi Computer Products (America) Inc.