Re: [PATCH 01/12] soc: qcom: add firmware name helper

[Bjorn Andersson]

On Mon, May 27, 2024 at 02:42:44PM GMT, Dmitry Baryshkov wrote:
> On Thu, 23 May 2024 at 01:48, Bjorn Andersson <quic_bjorande@xxxxxxxxxxx> wrote:
> >
> > On Tue, May 21, 2024 at 03:08:31PM +0200, Dmitry Baryshkov wrote:
> > > On Tue, 21 May 2024 at 13:20, Kalle Valo <kvalo@xxxxxxxxxx> wrote:
> > > >
> > > > Dmitry Baryshkov <dmitry.baryshkov@xxxxxxxxxx> writes:
> > > >
> > > > > On Tue, 21 May 2024 at 12:52, <neil.armstrong@xxxxxxxxxx> wrote:
> > > > >>
> > > > >> On 21/05/2024 11:45, Dmitry Baryshkov wrote:
> > > > >> > Qualcomm platforms have different sets of the firmware files, which
> > > > >> > differ from platform to platform (and from board to board, due to the
> > > > >> > embedded signatures). Rather than listing all the firmware files,
> > > > >> > including full paths, in the DT, provide a way to determine firmware
> > > > >> > path based on the root DT node compatible.
> > > > >>
> > > > >> Ok this looks quite over-engineered but necessary to handle the legacy,
> > > > >> but I really think we should add a way to look for a board-specific path
> > > > >> first and fallback to those SoC specific paths.
> > > > >
> > > > > Again, CONFIG_FW_LOADER_USER_HELPER => delays.
> > > >
> > > > To me this also looks like very over-engineered, can you elaborate more
> > > > why this is needed? Concrete examples would help to understand better.
> > >
> > > Sure. During the meeting last week Arnd suggested evaluating if we can
> > > drop firmware-name from the board DT files. Several reasons for that:
> > > - DT should describe the hardware, not the Linux-firmware locations
> > > - having firmware name in DT complicates updating the tree to use
> > > different firmware API (think of mbn vs mdt vs any other format)
> > > - If the DT gets supplied by the vendor (e.g. for
> > > SystemReady-certified devices), there should be a sync between the
> > > vendor's DT, linux kernel and the rootfs. Dropping firmware names from
> > > DT solves that by removing one piece of the equation
> > >
> > > Now for the complexity of the solution. Each SoC family has their own
> > > firmware set. This includes firmware for the DSPs, for modem, WiFi
> > > bits, GPU shader, etc.
> > > For the development boards these devices are signed by the testing key
> > > and the actual signature is not validated against the root of trust
> > > certificate.
> > > For the end-user devices the signature is actually validated against
> > > the bits fused to the SoC during manufacturing process. CA certificate
> > > (and thus the fuses) differ from vendor to vendor (and from the device
> > > to device)
> > >
> > > Not all of the firmware files are a part of the public linux-firmware
> > > tree. However we need to support the rootfs bundled with the firmware
> > > for different platforms (both public and vendor). The non-signed files
> > > come from the Adreno GPU and can be shared between platforms. All
> > > other files are SoC-specific and in some cases device-specific.
> > >
> > > So for example the SDM845 db845c (open device) loads following firmware files:
> > > Not signed:
> > > - qcom/a630_sqe.fw
> > > - qcom/a630_gmu.bin
> > >
> > > Signed, will work for any non-secured sdm845 device:
> > > - qcom/sdm845/a630_zap.mbn
> > > - qcom/sdm845/adsp.mbn
> > > - qcom/sdm845/cdsp.mbn
> > > - qcom/sdm485/mba.mbn
> > > - qcom/sdm845/modem.mbn
> > > - qcom/sdm845/wlanmdsp.mbn (loaded via TQFTP)
> > > - qcom/venus-5.2/venus.mbn
> > >
> > > Signed, works only for DB845c.
> > > - qcom/sdm845/Thundercomm/db845c/slpi.mbn
> > >
> > > In comparison, the SDM845 Pixel-3 phone (aka blueline) should load the
> > > following firmware files:
> > > - qcom/a630_sqe.fw (the same, non-signed file)
> > > - qcom/a630_gmu.bin (the same, non-signed file)
> > > - qcom/sdm845/Google/blueline/a630_zap.mbn
> >
> > How do you get from "a630_zap.mbn" to this? By extending the lookup
> > table for every target, or what am I missing?
> 
> More or less so. Matching the root OF node gives us the firmware
> location, then it gets prepended to all firmware targets. Not an ideal
> solution, as there is no fallback support, but at least it gives us
> some points to discuss (and to decide whether to move to some
> particular direction or to abandon the idea completely, making Arnd
> unhappy again).
> 

I understand the desire to not put linux-firmware-specific paths in the
DeviceTree, but I think I'm less keen on having a big lookup table in
the kernel...

Regards,
Bjorn