On Tue, Feb 7, 2023 at 11:57 PM Geert Uytterhoeven <geert@xxxxxxxxxxxxxx> wrote:
>
> Hi Saravana,
>
> On Wed, Feb 8, 2023 at 8:32 AM Saravana Kannan <saravanak@xxxxxxxxxx> wrote:
> > On Tue, Feb 7, 2023 at 6:08 PM Saravana Kannan <saravanak@xxxxxxxxxx> wrote:
> > > On Tue, Feb 7, 2023 at 12:57 PM Geert Uytterhoeven <geert@xxxxxxxxxxxxxx> wrote:
> > > > On Tue, Feb 7, 2023 at 2:42 AM Saravana Kannan <saravanak@xxxxxxxxxx> wrote:
> > > > > The driver core now:
> > > > > - Has the parent device of a supplier pick up the consumers if the
> > > > > supplier never has a device created for it.
> > > > > - Ignores a supplier if the supplier has no parent device and will never
> > > > > be probed by a driver
> > > > >
> > > > > And already prevents creating a device link with the consumer as a
> > > > > supplier of a parent.
> > > > >
> > > > > So, we no longer need to find the "compatible" node of the supplier or
> > > > > do any other checks in of_link_to_phandle(). We simply need to make sure
> > > > > that the supplier is available in DT.
> > > > >
> > > > > Signed-off-by: Saravana Kannan <saravanak@xxxxxxxxxx>
> > > >
> > > > Thanks for your patch!
> > > >
> > > > This patch introduces a regression when dynamically loading DT overlays.
> > > > Unfortunately this happens when using the out-of-tree OF configfs,
> > > > which is not supported upstream. Still, there may be (obscure)
> > > > in-tree users.
> > > >
> > > > When loading a DT overlay[1] to enable an SPI controller, and
> > > > instantiate a connected SPI EEPROM:
>
> [...]
>
> > > > The SPI controller and the SPI EEPROM are no longer instantiated.
>
> > > Sigh... I spent way too long trying to figure out if I caused a memory
> > > leak. I should have scrolled down further! Doesn't look like that part
> > > is related to anything I did.
> > >
> > > There are some flags set to avoid re-parsing fwnodes multiple times.
> > > My guess is that the issue you are seeing has to do with how many of
> > > the in memory structs are reused vs not when an overlay is
> > > applied/removed and some of these flags might not be getting cleared
> > > and this is having a bigger impact with this patch (because the fwnode
> > > links are no longer anchored on "compatible" nodes).
> > >
> > > With/without this patch (let's keep the series) can you look at how
> > > the following things change between each step you do above (add,
> > > remove, retry):
> > > 1) List of directories under /sys/class/devlink
> > > 2) Enable the debug logs inside __fwnode_link_add(),
> > > __fwnode_link_del(), device_link_add()
> > >
> > > My guess is that the final solution would entail clearing
> > > FWNODE_FLAG_LINKS_ADDED for some fwnodes.
> >
> > You replied just as I was about to hit send. So sending this anyway...
> >
> > Ok, I took a closer look and I think it's a bit of a mess. The fact
> > that it even worked for you without this patch is a bit of a
> > coincidence.
> >
> > Let's just take platform devices that are created by
> > driver/of/platform.c as an example.
> >
> > The main problem is that when you add/remove properties to a DT node
> > of an existing platform device, nothing is really done about it at the
> > device level. We don't even unbind and rebind the driver so the driver
> > could make use of the new properties. We don't remove and add back the
> > device so whoever might use the new property will use it. And if you
> > are adding a new node, it'll only trigger any platform device level
> > impact if it's a new node of a "simple-bus" (or similar bus) device.
> >
> > Problem 1:
> > So if you add a new child node to an existing probed device that adds
> > its children explicitly (as in, the parent is not a "simple-bus" like
> > device), nothing will happen. The newly added child device node will
> > get converted into a platform device, not will the parent device
> > notice it. So in your case of adding msiof0_pins, it's just that when
> > the consumer gets the pins, the driver doesn't get involved much and
> > it's the pinctrl framework that reads the DT and figures it out.
> >
> > With this patch, the fwnode links point to the actual resource and the
> > actual parent device inherits them if they don't get converted to a
> > struct device. But since we are adding this msiof0_pins after the
> > parent device has probed, the fwnode link isn't inherited by the
> > parent pinctrl device.
> >
> > Problem 2:
> > So if you add a property to an already bound device, nothing is done
> > by the driver. In your overlay example, if you move the status="okay"
> > line to be the first property you change in the msiof0 spi device,
> > you'll probably see that fw_devlink is no longer the one blocking the
> > probe. This is because the platform device will get added as soon as
> > the status flips from disabled to enabled and at that point fw_devlink
> > will think it has no suppliers and won't do any probe deferring. And
> > then as the new properties get added nothing will happen at the device
> > or fw_devlink level. If the msiof0's spi driver fails immediately with
> > NOT -EPROBE_DEFER when platform device is added because it couldn't
> > find any pinctrl property, then msiof0 will never probe (unless you
> > remove and add the driver). If it had failed with -EPROBE_DEFER, then
> > it might probe again if something else triggers a deferred probe
> > attempt. Clearly, things working/not working based on the order of
> > properties in DT is not a good implementation.
> >
> > Problem 3:
> > What if you enable a previously disabled supplier. There's no way to
> > handle that from a fw_devlink level without re-parsing the entire
> > device tree because existing devices might be consumers now.
> >
> > Anyway, long story short, it's sorta worked due to coincidence and
> > it's quite messy to get it to work correctly.
>
> Several subsystems register notifiers to be informed of the events
> above. E.g. drivers/spi/spi.c:
>
> if (IS_ENABLED(CONFIG_OF_DYNAMIC))
> WARN_ON(of_reconfig_notifier_register(&spi_of_notifier));
> if (IS_ENABLED(CONFIG_ACPI))
> WARN_ON(acpi_reconfig_notifier_register(&spi_acpi_notifier));
>
> So my issue might be triggered using ACPI, too.
Yeah, I did notice this before my email. Here's an ugly hack (at end
of email) to test my theory about Problem 1. I didn't compile test it
(because I should go to bed now), but you get the idea. Can you give
this a shot? It should fix your specific case. Basically for all
overlays (I hope the function is only used for overlays) we assume all
nodes are NOT devices until they actually get added as a device. Don't
review the code, it's not meant to be :)
-Saravana
--- a/drivers/of/dynamic.c
+++ b/drivers/of/dynamic.c
@@ -226,6 +226,7 @@ static void __of_attach_node(struct device_node *np)
np->sibling = np->parent->child;
np->parent->child = np;
of_node_clear_flag(np, OF_DETACHED);
+ np->fwnode.flags |= FWNODE_FLAG_NOT_DEVICE;
}
/**
diff --git a/drivers/of/platform.c b/drivers/of/platform.c
index 81c8c227ab6b..7299cd668e51 100644
--- a/drivers/of/platform.c
+++ b/drivers/of/platform.c
@@ -732,6 +732,7 @@ static int of_platform_notify(struct notifier_block *nb,
if (of_node_check_flag(rd->dn, OF_POPULATED))
return NOTIFY_OK;
+ rd->dn->fwnode.flags &= ~FWNODE_FLAG_NOT_DEVICE;
/* pdev_parent may be NULL when no bus platform device */
pdev_parent = of_find_device_by_node(rd->dn->parent);
pdev = of_platform_device_create(rd->dn, NULL,
diff --git a/drivers/spi/spi.c b/drivers/spi/spi.c
index 15f174f4e056..1de55561b25d 100644
--- a/drivers/spi/spi.c
+++ b/drivers/spi/spi.c
@@ -4436,6 +4436,7 @@ static int of_spi_notify(struct notifier_block
*nb, unsigned long action,
return NOTIFY_OK;
}
+ rd->dn->fwnode.flags &= ~FWNODE_FLAG_NOT_DEVICE;
spi = of_register_spi_device(ctlr, rd->dn);
put_device(&ctlr->dev);
