[GIT PULL] On-demand device probing

[Rafael J. Wysocki]

On Tuesday, October 20, 2015 09:15:01 AM Rob Herring wrote:
> On Tue, Oct 20, 2015 at 2:56 AM, Rafael J. Wysocki <rjw at rjwysocki.net> wrote:
> > On Monday, October 19, 2015 05:58:40 PM Rob Herring wrote:
> >> On Mon, Oct 19, 2015 at 4:40 PM, Rafael J. Wysocki <rjw at rjwysocki.net> wrote:
> >> > On Monday, October 19, 2015 10:58:25 AM Rob Herring wrote:
> >> >> On Mon, Oct 19, 2015 at 10:29 AM, David Woodhouse <dwmw2 at infradead.org> wrote:
> >> >> > On Mon, 2015-10-19 at 15:50 +0100, Mark Brown wrote:
> >> >> >> > But the point I'm making is that we are working towards *fixing* that,
> >> >> >> > and *not* using DT-specific code in places where we should be using the
> >> >> >> > generic APIs.
> >> >> >>
> >> >> >> What is the plan for fixing things here?  It's not obvious (at least to
> >> >> >> me) that we don't want to have the subsystems having knowledge of how
> >> >> >> they are bound to a specific firmware which is what you seem to imply
> >> >> >> here.
> >> >> >
> >> >> > I don't know that there *is* a coherent plan here to address it all.
> >> >> >
> >> >> > Certainly, we *will* need subsystems to have firmware-specific
> >> >> > knowledge in some cases. Take GPIO as an example; ACPI *has* a way to
> >> >> > describe GPIO, and properties which reference GPIO pins are intended to
> >> >> > work through that — while in DT, properties which reference GPIO pins
> >> >> > will have different contents. They'll be compatible at the driver
> >> >> > level, in the sense that there's a call to get a given GPIO given the
> >> >> > property name, but the subsystems *will* be doing different things
> >> >> > behind the scenes.
> >> >> >
> >> >> > My plan, such as it is, is to go through the leaf-node drivers which
> >> >> > almost definitely *should* be firmware-agnostic, and convert those. And
> >> >> > then take stock of what we have left, and work out what, if anything,
> >> >> > still needs to be done.
> >> >>
> >> >> Many cases are already agnostic in the drivers in terms of the *_get()
> >> >> functions. Some are DT specific, but probably because those subsystems
> >> >> are new and DT only. In any case, I don't think these 1 line changes
> >> >> do anything to make doing conversions here harder.
> >> >>
> >> >> >> It seems like we're going to have to refactor these bits of code when
> >> >> >> they get generalised anyway so I'm not sure that the additional cost
> >> >> >> here is that big.
> >> >> >
> >> >> > That's an acceptable answer — "we're adding legacy code here but we
> >> >> > know it's going to be refactored anyway". If that's true, all it takes
> >> >> > is a note in the commit comment to that effect. That's different from
> >> >> > having not thought about it :)
> >> >>
> >> >> Considering at one point we did create a fwnode based API, we did
> >> >> think about it. Plus there was little input from ACPI folks as to
> >> >> whether the change was even useful for ACPI case.
> >> >
> >> > Well, sorry, but who was asking whom, specifically?
> >>
> >> You and linux-acpi have been copied on v2 and later of the entire
> >> series I think.
> >
> > Yes, but it wasn't like a direct request, say "We need your input, so can you
> > please have a look and BTW we want this in 4.4, so please do it ASAP".  In
> > which case I'd prioritize that before other things I needed to take care of.
> 
> Fair enough. Can you please review and comment on v7 of the series? We
> can discuss at KS as well.

I'll do that.

I'm in the middle of travel right now and speaking at a conference tomorrow,
so I may not be able to get to that for a couple of days, but will do my best
to do it as soon as I possibly can.

> >> > The underlying problem is present in ACPI too and we don't really have a good
> >> > solution for it.  We might benefit from a common one if it existed.
> >>
> >> The problem for DT is we don't generically know what are the
> >> dependencies at a core level. We could know some or most dependencies
> >> if phandles (links to other nodes) were typed, but they are not. If
> >> the core had this information, we could simply control the device
> >> creation to order probing. Instead, this information is encoded into
> >> the bindings and binding parsing resides in the subsystems. That
> >> parsing happens during probe of the client side and is done by the
> >> subsystems (for common bindings). Since we already do the parsing at
> >> this point, it is a convenient place to trigger the probe of the
> >> dependency. Is ACPI going to be similar in this regard?
> >
> > It is similar in some ways.  For example, if a device's functionality depends
> > on an I2C resource (connection), the core doesn't know that at the device
> > creation time at least in some cases.  Same for GPIO, SPI, DMA engines etc.
> 
> So you will need to create devices, defer their probing and then probe
> on demand as well unless you have other ideas how you would do it.

Right.

> > There is a _DEP object in ACPI that can be used by firmware to tell the OS
> > about those dependencies, but there's no way in the driver core to use that
> > information anyway today.
> 
> I would think that the equivalent function for ACPI to of_device_probe
> could process these if they are generic and you can associate the
> dependency to a struct device.

Well, something along these lines probably.

> >> Fundamentally, it is a question of probe devices when their
> >> dependencies are present or drivers ensure their dependencies are
> >> ready. IIRC, init systems went thru a similar debate for service
> >> dependencies.
> >
> > The probe ordering is not the entire picture, though.
> >
> > Even if you get the probe ordering right, the problem is going to show up in
> > multiple other places: system suspend/resume, runtime PM, system shutdown,
> > unbinding of drivers.  In all of those cases it is necessary to handle things
> > in a specific order if there is a dependency.
> 
> My understanding was with deferred probe that it also solves suspend
> ordering problems because things are suspended in reverse order of
> probing. I suppose you could have slightly different dependencies for
> suspend, runtime PM, etc. than for probe? Perhaps we need to save the
> list of dependencies as we probe them. I don't think that would be too
> hard to add on to this series, but then if we don't need it now, why
> add it?

As Alan said, there are two problems here.  First off, the ordering of the
list used by system suspend/resume is the registration ordering, not the
probe ordering.  Moreover, though, even if we get the ordering right, it
still is not sufficient for devices with async_suspend set.

To address this, the core will have to make the involved async threads wait
for each other in accordance with the dependencies too.  That, in turn, is
very close to what's needed for runtime PM.

> >> >> In any case, we're talking about adding 1 line.
> >> >
> >> > But also about making the driver core slighly OF-centric.
> >>
> >> How so? The one line is in DT binding parsing code in subsystems, not
> >> driver core. The driver core change is we add every device (that
> >> happened to be created by DT) to the deferred probe list, so they
> >> don't probe right away.
> >
> > The "that happened to be created by DT" part is of concern here.  What is there
> > that makes DT special in that respect?  Why shouldn't that be applicable to
> > devices created by the ACPI core, for example, or by a board file or something
> > else?
> 
> DT is first. I think both examples could use this. Board files avoid
> the problem by controlling the registration order with initcall levels
> and just the call order in the code. You could come up with some way
> to define dependencies for devices in board files and reuse this
> mechanism. ACPI could use this as well if the dependencies are handled
> in a similar way and it seems like they could be.
> 
> >> > Sure, we need OF-specific code and ACPI-specific code wherever different
> >> > handling is required, but doing that at the driver core level seems to be
> >> > a bit of a stretch to me.
> >> >
> >> > Please note that we don't really have ACPI-specific calls in the driver core,
> >> > although we might have added them long ago even before the OF stuff appeared
> >> > in the kernel for the first time.  We didn't do that, (among other things)
> >> > because we didn't want that particular firmware interface to appear special
> >> > in any way and I'm not really sure why it is now OK to make OF look special
> >> > instead.
> >>
> >> I don't think DT is special and we avoid DT specific core changes as
> >> much as possible. I think the difference is DT uses platform_device
> >> and ACPI does not.
> >
> > ACPI uses platform devices too.  In fact, ACPI device objects are enumerated as
> > platform devices by default now.
> 
> Okay, I should have grepped for that:
> drivers/base/platform.c:                ACPI_COMPANION_SET(&pdev->dev, NULL);
> drivers/base/platform.c:        len = acpi_device_modalias(dev, buf,
> PAGE_SIZE -1);
> drivers/base/platform.c:        rc = acpi_device_uevent_modalias(dev, env);
> drivers/base/platform.c:        /* Then try ACPI style match */
> drivers/base/platform.c:        if (acpi_driver_match_device(dev, drv))
> 
> These are all cases which have DT version as well, so we're not really
> all that different here. There's a few more for DT, but that probably
> means you have just not hit the problems we have yet. For example,
> what happens if you have an interrupt line in which the controller is
> probed after the device connected to the interrupt line? That required
> resolving irqs in platform_get_irq rather than using static resources
> to support deferred probe.

We don't have this particular problem, because the IRQ controllers are
enumerated in a special way.

> Converting things like this to fwnode calls isn't hard to do. There
> just hasn't been a pressing need or mandate to do so.

Well, to me the problem is actually generic, so it is better to use generic
concepts to start with when trying to address it where that doesn't add too
much overhead.  Otherwise it's very easy to lose the broader context from
one's sight and then to start cutting corners.

Thanks,
Rafael