Re: Re: Issue with fw_devlink=on on 5.14

[Saravana Kannan]

On Wed, Sep 15, 2021 at 11:59 AM Jernej Škrabec
<jernej.skrabec@xxxxxxxxx> wrote:
>
> /cc linux-kernel@xxxxxxxxxxxxxxx
>
> Dne sreda, 15. september 2021 ob 20:36:11 CEST je Saravana Kannan napisal(a):
> > On Wed, Sep 15, 2021 at 11:23 AM Saravana Kannan <saravanak@xxxxxxxxxx>
> wrote:
> > >
> > > On Wed, Sep 15, 2021 at 11:07 AM Jernej Škrabec
> > > <jernej.skrabec@xxxxxxxxx> wrote:
> > > >
> > > > Hi Saravana!
> > > >
> > > > I noticed that in kernel 5.14, at least on Allwinner H3, deinterlace
> driver
> > > > doesn't load anymore if fw_devlink is set to on. If I set it to off,
> driver
> > > > loads and works without problems. If I remove interconnects property in
> DT
> > > > node and use driver hack instead, it also loads and work ok.
> > > >
> > > > For reference, I speak about this node:
> > > > https://elixir.bootlin.com/linux/v5.14/source/arch/arm/boot/dts/sun8i-h3.dtsi#L153
> >
> > Oh the dts file please. Not dtsi. I need the full picture of dependencies.
>
> I just dumped DTB from the board I use for testing this:
> http://ix.io/3yZB

I was basically asking for this:
arch/arm/boot/dts/sun8i-h3-orangepi-plus2e.dts

>
> >
> > -Saravana
> >
> > > >
> > > > Do you have any clue why it doesn't work?
> > >
> > > Kinda busy today and I'll take a close look later, but if you delete
> > > "interconnect" but it works, then it seems like you have no driver for
> > > your "mbus" interconnect? Is that expected?
>
> Ah, you're completely right. There is no MBUS driver (yet), since current
> purpose of that node is to determine DMA memory offsets and for that you don't
> need a driver.

Can you explain more how you get the DMA memory offsets in the code
from this node?

You might be able to add status="disabled" to the mbus node and this
would go away as fw_devlink doesn't cause a consumer to wait on a
disabled supplier. See more below -- you have a bunch of options.

>
> There is already WIP driver for it, though:
> https://github.com/crust-firmware/linux/commit/
> db34316a0194dbf0fcb8eba677ef11085a06a340
>
> > >
> > > If you have a driver, then make it use the proper driver model (probe
> > > a platform device).
> > >
> > > If you don't have a driver for it, I'm assuming you have
> > > CONFIG_MODULES enabled. Otherwise, fw_devlink=on is smart enough to
> > > know not to block on devices that'll never be probed.
>
> Correct, modules are enabled.
>
> > >
> > > If you do need CONFIG_MODULES enabled, then use
> > > deferred_probe_timeout=xxx so that fw_devlink doesn't wait for more
> > > modules to be loaded past xxx seconds.
>
> It's not what I want, but more that we broke default configuration and
> CONFIG_MODULES=y should really work.

Well, fw_devlink=on can't tell if you don't have a driver or if you
are going to load it soon (since you have CONFIG_MODULES enabled). So
if it's a board you are actively bringing up, one of the other options
below might be good to use.

> Is there any other way to fix this besides
> providing dummy MBUS driver? I doubt it would be accepted for backporting.

You could use fw_devlink=permissive in the kernel commandline, but it
will disable a bunch of benefits of fw_devlink=on (no pointless
deferred probes, no suspend/resume ordering based on dependencies
listed in DT, etc).

Another option would be to use deferred_probe_timeout=1 in the kernel
commandline, and this will cause this to be unblocked 1 second after
late_initcall() and you'll get the benefit of fw_devlink=on for all
the devices that have a driver registered by then (and some more
devices too -- see commit text of
d46f3e3ed5276e756caf40f760d4902d15c12dcb).

-Saravana

>
> > >
> > > Also I like to look up emails on lore, so when emailing me about
> > > fw_devlink (or any Linux stuff in general I suppose), it'd be nice if
> > > you can cc LKML.
>
> Noted.
>
> Best regards,
> Jernej
>
> > >
> > > Thanks,
> > > Saravana
> >
>
>