Re: barriers before {req/rsp}_cons = cons?

[Samuel Thibault <samuel.thibault@eu.citrix.com>]

Samuel Thibault, le Fri 18 Jul 2008 20:33:04 +0100, a écrit :
> Keir Fraser, le Fri 18 Jul 2008 20:31:13 +0100, a écrit :
> > > cons = netif->tx.req_cons;
> > > rmb(); /* Ensure that we see the request before we copy it. */
> > > memcpy(&txreq, RING_GET_REQUEST(&netif->tx, i), sizeof(txreq));
> > > some checks
> > > netif->tx.req_cons = ++cons;
> > 
> > Could you describe the race you believe is made possible by the absence of
> > the barrier?
> 
> It is very unlikely, but the write into req_cons could happen before
> actually doing the memcpy,

(more precisely, the read, of course).

> and so another processor could try to reuse
> the same request slot.

Samuel