[PATCH v3 1/2] drivers: bus: simple-pm-bus: Add support for probing simple bus only devices

[Saravana Kannan]

On Thu, Sep 9, 2021 at 7:02 AM Rob Herring <robh+dt at kernel.org> wrote:
>
> On Wed, Sep 8, 2021 at 7:58 PM Saravana Kannan <saravanak at google.com> wrote:
> >
> > On Wed, Sep 8, 2021 at 5:16 PM Rob Herring <robh+dt at kernel.org> wrote:
> > >
> > > On Tue, Sep 7, 2021 at 2:01 AM Saravana Kannan <saravanak at google.com> wrote:
> > > >
> > > > On Mon, Sep 6, 2021 at 12:54 AM Geert Uytterhoeven <geert at linux-m68k.org> wrote:
> > > > >
> > > > > Hi Saravana,
> > > > >
> > > > > Thanks for your patch!
> > > > >
> > > > > CC linux-pm, Lee (mfd)
> > > > >
> > > > > On Sat, Sep 4, 2021 at 2:05 AM Saravana Kannan <saravanak at google.com> wrote:
> > > > > > fw_devlink could end up creating device links for bus only devices.
> > > > > > However, bus only devices don't get probed and can block probe() or
> > > > > > sync_state() [1] call backs of other devices. To avoid this, probe these
> > > > > > devices using the simple-pm-bus driver.
> > > > > >
> > > > > > However, there are instances of devices that are not simple buses (they
> > > > > > get probed by their specific drivers) that also list the "simple-bus"
> > > > > > (or other bus only compatible strings) in their compatible property to
> > > > > > automatically populate their child devices. We still want these devices
> > > > > > to get probed by their specific drivers. So, we make sure this driver
> > > > > > only probes devices that are only buses.
> > > > >
> > > > > Note that this can also be the case for buses declaring compatibility
> > > > > with "simple-pm-bus".  However, at the moment, none of such device
> > > > > nodes in upstream DTS files have device-specific drivers.
> > > >
> > > > Not sure about mfd, but I want to make sure we don't confuse busses
> > > > (which are typically added to a class) with these "simple bus" devices
> > > > that are added to platform_bus. Also if these other buses are actually
> > > > causing an issue, then then should implement their own stub driver or
> > > > use try patch[2] if they are added to classes (devices on classes
> > > > don't probe)
> > > >
> > > > [2] - https://lore.kernel.org/lkml/20210831224510.703253-1-saravanak@google.com/
> > > >
> > > > >
> > > > > > [1] - https://lore.kernel.org/lkml/CAPDyKFo9Bxremkb1dDrr4OcXSpE0keVze94Cm=zrkOVxHHxBmQ@mail.gmail.com/
> > > > > > Signed-off-by: Saravana Kannan <saravanak at google.com>
> > > > > > Tested-by: Saravana Kannan <saravanak at google.com>
> > > > >
> > > > > > --- a/drivers/bus/simple-pm-bus.c
> > > > > > +++ b/drivers/bus/simple-pm-bus.c
> > > > > > @@ -13,11 +13,26 @@
> > > > > >  #include <linux/platform_device.h>
> > > > > >  #include <linux/pm_runtime.h>
> > > > > >
> > > > > > -
> > > > > >  static int simple_pm_bus_probe(struct platform_device *pdev)
> > > > > >  {
> > > > > > -       const struct of_dev_auxdata *lookup = dev_get_platdata(&pdev->dev);
> > > > > > -       struct device_node *np = pdev->dev.of_node;
> > > > > > +       const struct device *dev = &pdev->dev;
> > > > > > +       const struct of_dev_auxdata *lookup = dev_get_platdata(dev);
> > > > > > +       struct device_node *np = dev->of_node;
> > > > > > +       const struct of_device_id *match;
> > > > > > +
> > > > > > +       match = of_match_device(dev->driver->of_match_table, dev);
> > > > > > +
> > > > > > +       /*
> > > > > > +        * These are transparent bus devices (not simple-pm-bus matches) that
> > > > > > +        * have their child nodes populated automatically.  So, don't need to
> > > > > > +        * do anything more.
> > > > > > +        */
> > > > > > +       if (match && match->data) {
> > > > > > +               if (of_property_match_string(np, "compatible", match->compatible) == 0)
> > > > >
> > > > > Does this work as expected? Having multiple compatible values in a
> > > > > device node does not guarantee there exist a separate driver for any
> > > > > of the device-specific compatible values.
> > > >
> > > > Right, and if they are platform devices that are equivalent to
> > > > simple-bus (meaning, they don't do anything in Linux and just have
> > > > their devices populated) we can add those to this list too.
> > >
> > > I think this needs to be a list of compatibles we have drivers for
> > > instead.
> >
> > I don't think a "denylist" (devices we shouldn't probe with this
> > driver) would be a short list. As of today, literally any device that
> > has children could add a "simple-bus" to the compatible property and
> > get its child devices populated for free. If we use a denylist, we'll
> > have to update it every time someone adds "simple-bus" as a general
> > match to a DT node (new or otherwise) that isn't in the denylist. The
> > list would blow up and be a maintenance headache.
> >
> > Also, a denylist won't capture any DT that isn't part of the kernel
> > repo but depends on "simple-bus" to populate the device's child nodes.
> > Keep in mind this could be true even for completely upstream drivers
> > today. And on top of that, this will also break for downstream drivers
> > and platforms in the development stage.
>
> You've got this all backwards.

I don't think so. I got what you are saying.

> The list would be compatibles for which
> they have their own driver. If they have their own driver, we know
> that because there is a driver in the kernel. If you have an out of
> tree bus driver, then I guess you shouldn't be claiming compatibility
> with 'simple-bus'.

But it isn't just out of tree drivers though. Take for example (a
random compatible string with a driver in upstream)
"qcom,sdm845-gpucc" that has DT nodes in a couple of DTS files. It's
perfectly valid today today to add child nodes under it and have the
child nodes get populated by making the compatible property:
"qcom,sdm845-gpucc", "simple-bus"

The only thing not upstreamed in that case is the DT file. And I'm
fairly certain our position is not "if your DT isn't in the kernel
repo we won't support it". In this situation we have no way of knowing
we need to add "qcom,sdm845-gpucc" to the "it has a driver, don't
probe it" list. And even if they upstream it, this list is going to
blow up. And we are breaking backward compatibility that can't be
fixed without a kernel update (no runtime fixes possible).

> > The allowlist is much smaller and manageable.
>
> I count 140 cases of 'simple-bus' with another compatible. I find
> roughly 24 of those under drivers/ that have a driver (and look,
> there's at91 pinctrl/gpio). There's some more under arch/, but I'm not
> sure if they are drivers. This is what I ran:
>
> git grep -ho '".*", "simple-bus"' -- arch/ | cut -d' ' -f1 | grep -oE
> '".+"' | grep -v '"syscon"' | sort -u > buses.txt
> git grep -f buses.txt -- drivers/
>
> I haven't looked at 'simple-mfd', but that was supposed to always mean
> 'no driver'.
>
> To put it another way, let's look at 3 possibilities:
>
> 'simple-bus'
> 'foo,has-no-driver', 'simple-bus'
> 'foo,has-a-driver', 'simple-bus'
>
> The first case is easy. The last 2 cases are not. We have no way to
> handle them differently without a list.
>
> > > A more specific compatible that the OS doesn't understand
> > > shouldn't cause a change in behavior and adding one would.
>
> Again, if a dt is modified from case 1 to case 2, there should not be
> a change in behavior.

My point is that if DT is modified from 1 to 2, we'll need to add
"foo,has-no-driver" to the list we'll maintain. That preserves
backward compatibility with existing DTs and if someone modifies the
DT it's more likely that they can modify the kernel too. See more
below.

> Presumably if Ulf changed his test in this way,
> it would again fail, right?
>
>
> > I think the amount of specific compatible strings that'll be added,
> > but won't have drivers added to Linux AND would want to boot with
> > Linux is much less likely than the amount of times we'd have to update
> > a denylist.
> >
> > Also, if we do hit the cases you mention and we want those devices to
> > get probed anyway, with my current allowlist approach, we could use
> > "driver_override" to force this driver to match them. If you use a
> > denylist like you said, there's no way you can get the simple-pm-bus
> > to unbind and let the more specific driver to bind.
>
> Where would we set "driver_override"? Aren't we going to end up with a
> driver_override list?

I'm not saying we maintain a list in the kernel. I'm saying if someone
changed their DT to add 'foo,has-no-driver', but hit the unlikely case
where they can update the DT but not the kernel, they can still get
the 'foo,has-no-driver' to probe by using the driver_override file for
that device in sysfs. TLDR: they can still get it to work without
having to modify the kernel.

>
> [...]
>
> > > > > [*] Especially if "simple-pm-bus" and "simple-bus" would be treated
> > > > >     the same.
> > > >
> > > > They are not treated the same way.
> > >
> > > I think it would be better if they were. IOW, the core code stops
> > > descending into simple-bus, etc. nodes and they are populated here.
> > > Then we just get rid of of_default_bus_match_table.
> >
> > Right, I would if we could. But we can't simply stop descending the
> > simple-bus nodes in the core code because all the specific drivers
> > that used to have their child devices populated automatically would
> > stop working and would need to be updated to populate their child
> > devices. And I'm sure there are a ton more downstream kernels and
> > downstream DTs (that use upstream kernels) that we would break.
> >
> > If you really want to do that go for it, but I'd rather not do all
> > this as part of trying to fix the issue Ulf reported that needs
> > simple-bus only devices probed.
>
> Agreed.

I'm confused. So you are okay with this patch to merge now?

> > > That could cause some issues with init ordering. As I recall the at91
> > > gpio and pinctrl drivers are sensitive to this. The default call to
> > > of_platform_populate doesn't work on those systems because the devices
> > > get created later than when their machine specific call happens. It
> > > may have been a case of a parent probe assuming a child probe
> > > completed after of_platform_populate returns (also a problem for Qcom
> > > with DWC3). There's a fix for at91 somewhere in the git history after
> > > I broke it. I started trying to untangle things with at91, but never
> > > finished that.
> >
> > I think it'll cause a lot of issues if we stop descending simple-bus
> > nodes in the core code. We're just scratching the surface here.
>
> Perhaps, but if there's cases which are that fragile, we should fix them.

Sure, but I'm not sure I want to open that can of works,

-Saravana