Re: [PATCH v3 2/6] PCI: iproc: Add INTx support with better modeling

[Ray Jui]

On 12/4/19 8:07 AM, Andrew Murray wrote:
> On Wed, Dec 04, 2019 at 10:29:51AM +0200, Andy Shevchenko wrote:
> > On Wed, Dec 4, 2019 at 12:09 AM Ray Jui <ray.jui@xxxxxxxxxxxx> wrote:
> > > On 12/3/19 11:27 AM, Andy Shevchenko wrote:
> > > > On Tue, Dec 3, 2019 at 5:55 PM Andrew Murray <andrew.murray@xxxxxxx> wrote:
> > > > > On Tue, Dec 03, 2019 at 10:27:02AM +0530, Srinath Mannam wrote:
> > > >
> > > > > > +     /* go through INTx A, B, C, D until all interrupts are handled */
> > > > > > +     do {
> > > > > > +             status = iproc_pcie_read_reg(pcie, IPROC_PCIE_INTX_CSR);
> > > > >
> > > > > By performing this read once and outside of the do/while loop you may improve
> > > > > performance. I wonder how probable it is to get another INTx whilst handling
> > > > > one?
> > > >
> > > > May I ask how it can be improved?
> > > > One read will be needed any way, and so does this code.
> > >
> > > I guess the current code will cause the IPROC_PCIE_INTX_CSR register to
> > > be read TWICE, if it's ever set to start with.
> > >
> > > But then if we do it outside of the while loop, if we ever receive an
> > > interrupt while servicing one, the interrupt will still need to be
> > > serviced, and in this case, it will cause additional context switch
> > > overhead by going out and back in the interrupt context.
>
> Yes it's a trade off - if you dropped the do/while loop and thus had a single
> read you'd reduce the overhead on interrupt handling in every case except
> where another INTx is received whilst in this function. But as you point out
> each time that does happen you'll pay the penalty of a context switch.

Exactly, it's a tradeoff between: 1) saving one register read (which is 
likely in the 10th of nanosecond range) in all INTx handling; and 2) 
saving context switches (which is likely in 10th of microsecond range) 
in cases when we have multiple INTx when servicing it.

The current implementation takes 2), which I thought it makes sense.

> I don't have any knowledge of this platform so I have no idea if such a change
> would be good/bad or material. However I thought I'd point it out. Looking at
> the other controller drivers, some handle in a loop and some don't.
>
>
> > > My take is that it's probably more ideal to leave this portion of code
> > > as it is.
> >
> > Can't we simple drop a do-while completely and leave only
> > for_each_set_bit() loop?

Like both Andrew and I pointed out. There's a tradeoff here. Could you 
please help to justify why you favor 1) than 2)?

> I'm happy either way.
>
> Thanks,
>
> Andrew Murray
>
> > > > > > +             for_each_set_bit(bit, &status, PCI_NUM_INTX) {
> > > > > > +                     virq = irq_find_mapping(pcie->irq_domain, bit);
> > > > > > +                     if (virq)
> > > > > > +                             generic_handle_irq(virq);
> > > > > > +                     else
> > > > > > +                             dev_err(dev, "unexpected INTx%u\n", bit);
> > > > > > +             }
> > > > > > +     } while ((status & SYS_RC_INTX_MASK) != 0);
> >
> >
> >
> > --
> > With Best Regards,
> > Andy Shevchenko