On Mon, Oct 30, 2017 at 07:13:13PM +0800, Leo Yan wrote:
> Hi Mark,
>
> On Mon, Oct 30, 2017 at 10:19:40AM +0000, Mark Rutland wrote:
> > Hi,
> >
> > On Mon, Oct 30, 2017 at 12:45:06PM +0800, Leo Yan wrote:
> > > On Fri, Oct 27, 2017 at 11:46:00AM +0100, Mark Rutland wrote:
> > > > On Fri, Oct 27, 2017 at 02:15:03PM +0800, Kaihua Zhong wrote:
> > > > > +static int hi3660_mbox_check_state(struct mbox_chan *chan)
> > > > > +{
> >
> > > > > + /* Ensure channel is released */
> > > > > + writel_relaxed(0xffffffff, base + MBOX_IMASK_REG);
> > > > > + writel_relaxed(BIT(mdev->ack_irq), base + MBOX_SRC_REG);
> > > > > + __asm__ volatile ("sev");
> > > > > + return 0;
> > > > > +}
> > > >
> > > > Drivers really shouldn't be using SEV directly (even if via the
> > > > sev() macro)...
> > > >
> > > > This SEV isn't ordered w.r.t. anything, and it's unclear what
> > > > ordering you need, so this simply does not work.
> > >
> > > I will leave your questions for Hisilicon colleagues, essentially your
> > > questions are related with mailbox mechanism.
> > >
> > > But I'd like to firstly get clear your question for "This SEV isn't
> > > ordered w.r.t. anything". From my understanding, ARMv8 architecture
> > > natually adds DMB before SEV so all previous register writing
> > > opreations should be ensured to endpoint before SEV?
> >
> > This is not the case; SEV does not add any implicit memory barrier, and
> > is not ordered w.r.t. memory accesses.
> >
> > See ARM DDI 0487B.b, page D1-1905, "The Send Event instructions":
> >
> > The PE is not required to guarantee the ordering of this event with
> > respect to the completion of memory accesses by instructions before
> > the SEV instruction. Therefore, ARM recommends that software
> > includes a DSB instruction before any SEV instruction.
>
> My fault and thanks for explanation.
>
> > Note that a DMB is not sufficient, as SEV is not a memory access.
>
> Understood now, so below code should be safe?
>
> wmb(); -> dsb(st);
> sev();
Whether that is safe depends on what you are trying to ensure is
ordered, and what the other side (with the WFE) is doing.
For example, my understanding is that in general, WFE is also not
ordered w.r.t. memory accesses.
This is a very subtle part of the architecture, and I'm very much not
keen on using WFE and SEV outside of architecture code implementing
locking primitives.
Thanks,
Mark.
--
To unsubscribe from this list: send the line "unsubscribe devicetree" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
