Re: [PATCH v4] lpc32xx: Added ethernet driver: smp_wmb()

[Ben Hutchings <bhutchings@solarflare.com>]

On Tue, 2012-03-06 at 14:17 +0000, Russell King - ARM Linux wrote:
> On Tue, Mar 06, 2012 at 02:03:28PM +0000, Ben Hutchings wrote:
> > On Tue, 2012-03-06 at 11:43 +0100, Roland Stigge wrote:
> > > On 03/05/2012 11:45 PM, Ben Hutchings wrote:
> > > >> +	/* Clear and enable interrupts */
> > > >> +	writel(0xFFFF, LPC_ENET_INTCLEAR(pldat->net_base));
> > > >> +	lpc_eth_enable_int(pldat->net_base);
> > > >> +
> > > >> +	/* Get the next TX buffer output index */
> > > >> +	pldat->num_used_tx_buffs = 0;
> > > >> +	pldat->last_tx_idx =
> > > >> +		readl(LPC_ENET_TXCONSUMEINDEX(pldat->net_base));
> > > > 
> > > > Doesn't this need to be done *before* enabling interrupts?  Also, I
> > > > think you need an smp_wmb() so that the interrupt handler is guaranteed
> > > > to see all these writes.
> > > 
> > > Do you mean _one_ smp_wmb() directly after lpc_eth_enable_int() (which
> > > I'm moving behind the above code?
> > 
> > The sequence should be
> > 
> > 	pldat->state = values...;
> > 	smp_wmb();
> > 	enable_interrupts();
> 
> Is this correct?
> 
> "SMP BARRIER PAIRING" and "EXAMPLES OF MEMORY BARRIER SEQUENCES" in
> Documentation/memory-barriers.txt suggest that there should be some kind
> of pairing with smp_wmb() to ensure correctness.

The thread we're synchronising with (the interrupt handler) starts
*after* the smp_wmb().  Therefore there is no need for a second barrier.

Ben.

-- 
Ben Hutchings, Staff Engineer, Solarflare
Not speaking for my employer; that's the marketing department's job.
They asked us to note that Solarflare product names are trademarked.