Re: [PATCH v1 5/5] driver core: Set fw_devlink=on by default

[Marc Zyngier]

On 2021-01-18 19:16, Geert Uytterhoeven wrote:
> Hi Marc,
>
> On Mon, Jan 18, 2021 at 6:59 PM Marc Zyngier <maz@xxxxxxxxxx> wrote:
> > On 2021-01-18 17:39, Geert Uytterhoeven wrote:
> > > On Fri, Dec 18, 2020 at 4:34 AM Saravana Kannan <saravanak@xxxxxxxxxx>
> > > wrote:
> > >> Cyclic dependencies in some firmware was one of the last remaining
> > >> reasons fw_devlink=on couldn't be set by default. Now that cyclic
> > >> dependencies don't block probing, set fw_devlink=on by default.
> > >>
> > >> Setting fw_devlink=on by default brings a bunch of benefits
> > >> (currently,
> > >> only for systems with device tree firmware):
> > >> * Significantly cuts down deferred probes.
> > >> * Device probe is effectively attempted in graph order.
> > >> * Makes it much easier to load drivers as modules without having to
> > >>   worry about functional dependencies between modules (depmod is still
> > >>   needed for symbol dependencies).
> > >>
> > >> If this patch prevents some devices from probing, it's very likely due
> > >> to the system having one or more device drivers that "probe"/set up a
> > >> device (DT node with compatible property) without creating a struct
> > >> device for it.  If we hit such cases, the device drivers need to be
> > >> fixed so that they populate struct devices and probe them like normal
> > >> device drivers so that the driver core is aware of the devices and
> > >> their
> > >> status. See [1] for an example of such a case.
> > >>
> > >> [1] -
> > >> https://lore.kernel.org/lkml/CAGETcx9PiX==mLxB9PO8Myyk6u2vhPVwTMsA5NkD-ywH5xhusw@xxxxxxxxxxxxxx/
> > >> Signed-off-by: Saravana Kannan <saravanak@xxxxxxxxxx>
> > >
> > > Shimoda-san reported that next-20210111 and later fail to boot
> > > on Renesas R-Car Gen3 platforms. No output is seen, unless earlycon
> > > is enabled.
> > >
> > > I have bisected this to commit e590474768f1cc04 ("driver core: Set
> > > fw_devlink=on by default").
> >
> > There is a tentative patch from Saravana here[1], which works around
> > some issues on my RK3399 platform, and it'd be interesting to find
> > out whether that helps on your system.
> >
> > Thanks,
> >
> >          M.
> >
> > [1]
> > https://lore.kernel.org/r/20210116011412.3211292-1-saravanak@xxxxxxxxxx
>
> Thanks for the suggestion, but given no devices probe (incl. GPIO
> providers), I'm afraid it won't help. [testing] Indeed.
>
> With the debug prints in device_links_check_suppliers enabled, and
> some postprocessing, I get:
>
>     255 supplier e6180000.system-controller not ready
>       9 supplier fe990000.iommu not ready
>       9 supplier fe980000.iommu not ready
>       6 supplier febd0000.iommu not ready
>       6 supplier ec670000.iommu not ready
>       3 supplier febe0000.iommu not ready
>       3 supplier e7740000.iommu not ready
>       3 supplier e6740000.iommu not ready
>       3 supplier e65ee000.usb-phy not ready
>       3 supplier e6570000.iommu not ready
>       3 supplier e6054000.gpio not ready
>       3 supplier e6053000.gpio not ready
>
> As everything is part of a PM Domain, the (lack of the) system 
> controller
> must be the culprit. What's wrong with it? It is registered very early 
> in
> the boot:
>
> [    0.142096] rcar_sysc_pd_init:442: of_genpd_add_provider_onecell() 
> returned 0

Yeah, this looks like the exact same problem. The devlink stuff assumes
that because there is a "compatible" property, there will be a driver
directly associated with the node containing this property.

If any other node has a reference to that first node, the dependency
will only get resolved if/when that first node is bound to a driver.
Trouble is, there are *tons* of code in the tree that invalidate
this heuristic, and for each occurrence of this we get another failure.

The patch I referred to papers over it by registering a dummy driver,
but that doesn't scale easily...

        M.
--
Jazz is not dead. It just smells funny...