[PATCH 03/11] PCI: aardvark: Add support for DLLSC and hotplug interrupt

[Marek Behún]

On Mon, 26 Sep 2022 16:00:13 +0200
Lorenzo Pieralisi <lpieralisi at kernel.org> wrote:

> On Mon, Sep 26, 2022 at 08:35:51AM -0400, Marc Zyngier wrote:
> > On Mon, 26 Sep 2022 07:49:55 -0400,
> > Lorenzo Pieralisi <lpieralisi at kernel.org> wrote:  
> > > 
> > > On Sat, Sep 17, 2022 at 10:05:59AM +0100, Marc Zyngier wrote:  
> > > > Hi Lorenzo,
> > > > 
> > > > On Fri, 09 Sep 2022 15:57:11 +0100,
> > > > Lorenzo Pieralisi <lpieralisi at kernel.org> wrote:  
> > > > > 
> > > > > [+Marc, Thomas - I can't merge this code without them reviewing it,
> > > > > I am not sure at all you can mix the timer/IRQ code the way you do]
> > > > > 
> > > > > On Thu, Aug 18, 2022 at 03:51:32PM +0200, Marek Behún wrote:  
> > > > > > From: Pali Rohár <pali at kernel.org>
> > > > > > 
> > > > > > Add support for Data Link Layer State Change in the emulated slot
> > > > > > registers and hotplug interrupt via the emulated root bridge.
> > > > > > 
> > > > > > This is mainly useful for when an error causes link down event. With
> > > > > > this change, drivers can try recovery.
> > > > > > 
> > > > > > Link down state change can be implemented because Aardvark supports Link
> > > > > > Down event interrupt. Use it for signaling that Data Link Layer Link is
> > > > > > not active anymore via Hot-Plug Interrupt on emulated root bridge.
> > > > > > 
> > > > > > Link up interrupt is not available on Aardvark, but we check for whether
> > > > > > link is up in the advk_pcie_link_up() function. By triggering Hot-Plug
> > > > > > Interrupt from this function we achieve Link up event, so long as the
> > > > > > function is called (which it is after probe and when rescanning).
> > > > > > Although it is not ideal, it is better than nothing.  
> > > > > 
> > > > > So before even coming to the code review: this patch does two things.
> > > > > 
> > > > > 1) It adds support for handling the Link down state
> > > > > 2) It adds some code to emulate a Link-up event
> > > > > 
> > > > > Now, for (2). IIUC you are adding code to make sure that an HP
> > > > > event is triggered if advk_pcie_link_up() is called and it
> > > > > detects a Link-down->Link-up transition, that has to be notified
> > > > > through an HP event.
> > > > > 
> > > > > If that's correct, you have to explain to me please what this is
> > > > > actually achieving and a specific scenario where we want this to be
> > > > > implemented, in fine details; then we add it to the commit log.
> > > > > 
> > > > > That aside, the interaction of the timer and the IRQ domain code
> > > > > must be reviewed by Marc and Thomas to make sure this is not
> > > > > a gross violation of the respective subsystems usage.  
> > > > 
> > > > I don't see anything being a "gross violation" here, at least from an
> > > > interrupt subsystem perspective. In a way, this is synthesising an
> > > > interrupt on the back of some other event, and as long as the context
> > > > is somehow appropriate (something that looks like an interrupt when
> > > > pretending there is one), this should be OK. Other subsystems such as
> > > > i2c GPIO expanders do similar things.  
> > > 
> > > Right, thanks.
> > >   
> > > > The one thing I'm dubious about is the frequency of the timer. Asking
> > > > for a poll of the link every jiffy is bound to be expensive, and it
> > > > would be good to relax this as much as possible, specially on low-end
> > > > HW such as this, where every cycle counts. It is always going to be a
> > > > "best effort" thing, and the commit message doesn't say what's the
> > > > actual grace period to handle this (the spec probably has one).  
> > > 
> > > AFAICS, the code does not poll the link. It sets a timer only if
> > > the link is checked (eg upon PCI bus forced rescan or config access)
> > > the link is up and it was down, to emulate a HP IRQ.  
> > 
> > I still find the timer frequency pretty high, but surely the authors
> > of the code have worked out that this wasn't a problem.  
> 
> Please correct me if I am wrong but with mod_timer() all they want to do
> is emulating/firing an (hotplug) IRQ as soon as possible - a one-off.
> 
> It is not a periodic timer - at least that's what I understand from the
> code and that's the reason why such a short interval was chosen but
> it should not be me who comment on that :)

This is true, it is not a periodic timer. We are just using an IRQSAFE
timer to call generic_handle_domain_irq() from it's handler, because we
can't do it during the config space access.

Marek