On Mon, Sep 28, 2026 at 10:36:18PM +0200, Neil Armstrong wrote:
> On 9/28/26 21:48, Benjamin Tissoires wrote:
> > On Sep 28 2026, Neil Armstrong wrote:
> > > On 9/28/26 19:24, Benjamin Tissoires wrote:
> > > > On Sep 28 2026, Neil Armstrong wrote:
> > > > > Hi,
> > > > > 
> > > > > On 9/28/26 18:22, Maxime Ripard wrote:
> > > > > > Hi,
> > > > > > 
> > > > > > Panels in general, and MIPI-DSI panels in particular, are pretty
> > > > > > difficult to support and require pretty much a panel driver for each
> > > > > > panel produced. Most of them are pretty simple, and require an opaque
> > > > > > initialization sequence that is usually poorly documented.
> > > > > > 
> > > > > > This creates a tension between OEMs and distros because OEMs will
> > > > > > typically get a new panel to react to a sourcing issue during
> > > > > > production, and thus need some swift turnaround between getting their
> > > > > > new panel and it being operational in the OS. Distributions on the other
> > > > > > hand can take years to ship a kernel with that new panel driver.
> > > > > > 
> > > > > > To solve this, I followed the example of HID-BPF and wrote a panel
> > > > > > driver that will rely on BPF programs to perform the panel
> > > > > > initialization. That way, we can ship the programs separately from the
> > > > > > kernel, and with a different lifecycle. If this driver is accepted, the
> > > > > > plan is to have a userspace component started by udev to identify and
> > > > > > load the right BPF program for the panels found on the device.
> > > > > 
> > > > > This is kind of late for serious applications except if we manage to
> > > > > solve the bootloader to Linux display engine transition.
> > > > > 
> > > > > > 
> > > > > > This driver is fully functional and works with both 5" and 7" Touch
> > > > > > Display 2 panels for the RaspberryPi. However, it breaks away from the
> > > > > > typical panel driver in multiple ways:
> > > > > > 
> > > > > > - BPF programs can only be loaded by userspace. This leaves us with two
> > > > > >      choices:
> > > > > > 
> > > > > >      * We prevent the driver from loading until the script itself is
> > > > > >        loaded. This has the side effect of preventing any other output to
> > > > > >        be used until the initramfs is ran at the earliest, and possibly
> > > > > >        ever if the loader isn't installed for example.
> > > > > 
> > > > > This adds a dependency on user-space behavior and if somehow the
> > > > > initramfs doesn't load for a reason we won't have a way to display
> > > > > an error.
> > > > > 
> > > > > > 
> > > > > >      * Or we probe the driver all the time, but only report it as connected
> > > > > >        once a program has been registered. This is somewhat unconventional,
> > > > > >        but allows the other outputs to be functional, *and* allows the user
> > > > > >        to force the output if their panel doesn't require any
> > > > > >        initialization or during debugging. I chose this solution.
> > > > > 
> > > > > Both options are not really great...
> > > > > 
> > > > > > 
> > > > > > - It's not a panel driver, but a bridge one, which is also pretty
> > > > > >      unconventional. This is required because panel drivers don't have
> > > > > >      access to a detect callback that is required for the above, but I also
> > > > > >      think that the recent work from Luca blurs the line from panels and
> > > > > >      bridges and we'll end up going that road anyway.
> > > > > 
> > > > > On this point, DDIC _are_ bridges, but in the current panel API we blur the line between
> > > > > the panel and the DDIC. So being a bridge is fine, but in a general way we lack
> > > > > a proper way to describe the display/panel/monitor independently of the DDIC.
> > > > > 
> > > > > At first glance it's a nice driver, but moving the timings into a blob moves something
> > > > > into possible proprietary binaries with possible closed licence and distribution
> > > > > restriction so it's a downgrade for the same of bringing up a panel faster.
> > > > 
> > > > Quick answer on this, because I had the very same questions regarding
> > > > HID-BPF:
> > > > - in BPF, you can require (and by default it does) that only GPL
> > > > 	compatible BPF programs are loaded, closing the argument of "closed
> > > > 	licence and distribution restriction"
> > > > - also, a BPF program can be disassembled much easier than a binary
> > > > 	blob, and I remember Alexei showing me an example where you get almost
> > > > 	the source code from the BPF object in just one pass.
> > > 
> > > Right, it "solve" one of my question, but doesn't really solve the issue
> > > of vendors providing "GPL" bpf programs with source available "somewhere".
> > > 
> > > Another big issue is the API, I don't want to keep the current API as-is,
> > > we plan to support more advanced panel features and use try to use the
> > > atomic states to support rate switching for example, and I'm not confident
> > > it's a good idea since there's no "simple" and "forever valid" API
> > > to initialize panels...
> > > 
> > 
> > That's exactly where BPF shines. The simple rule of thumb is: there is
> > no API stability guaranteed. Basically, thanks to the verifier and CORE
> > (Compile Once Run Everywhere), you don't need to keep the API stable and
> > available forever. There are multiple ways of dealing with it, but the
> > gist is that if a BPF is trying to load an "old" API, it will be
> > rejected by the verifier.
> > 
> > Then it's a matter of being nice enough in the kernel and provide ways
> > to deal with those.
> > 
> > To give you a few examples:
> > 
> > - in HID-BPF, in userspace, we keep old versions of APIs in separate
> > 	compiled objects. Each is incremented (by 10 so we have a little bit
> > 	of room in the middle). Then the loader tries first
> > 	0020-device-with-new-api.bpf.o, and if it fails, it tries
> > 	0010-device-with-old-api.bpf.o
> > 
> > 	I'm not saying we should do the same here, but that's one idea
> > 
> > - recently, a BPF commit broke the ABI of a function while removing
> > 	implicits: bpf_wq_set_callback_impl() was replaced by
> > 	bpf_wq_set_callback() with a different number of arguments. I simply
> > 	had to add a new function in my header which basically does:
> > 
> > 	static inline int
> > 	hid_bpf_wq_set_callback(struct bpf_wq *wq,
> > 	                        int (*callback_fn)(void *, int *, void *),
> > 	                        unsigned int flags)
> > 	{
> >        if (bpf_ksym_exists(bpf_wq_set_callback))
> >                return bpf_wq_set_callback(wq, callback_fn, flags);
> >        if (bpf_ksym_exists(bpf_wq_set_callback_impl))
> >                return bpf_wq_set_callback_impl(wq, callback_fn, flags, NULL);
> >    }
> > 
> >    And then I changed the bpf.c to use hid_bpf_wq_set_callback() and the
> >    bpf is compatible with both APIs
> > 
> > - with CORE, struct fields are relocated on the fly when you load the BPF
> >    So for instance, if my BPF only accesses fields .name, .phys and .id
> >    in the BPF, I can define:
> >    struct hid_device {
> >      char                       name[128];
> >      char                       phys[64];
> >      unsigned int               id;
> >    }
> > 
> >    Then when loading the bpf, the verifier replaces all offset to the
> >    ones actually used by the running kernel, and the program loads
> >    transparently, even if you add fields before/after or change the
> >    fields order in the kernel.
> > 
> > Changing the mindset is the hardest part of it. But once you are making
> > the shift, it's actually much better to work with. It doesn't mean you
> > can go yolo. You still need to be careful in your choices knowing the
> > impact on your users. But the API at version 0 is not frozen and you
> > don't need to maintain it forever, especially if you control the loader
> > and the headers used to compile the BPFs.
> 
> It's really impressive, but there's no way this would become the de-facto
> way to program the panels

Nobody said it would? And like I said in my previous mail, I actually
expect us to keep merging panel drivers. Do I think it should become the
go-to solution for simple panels? yes. But we can always make
exceptions, and for more complex panels we should totally do a more
complex driver. Like HID has been doing.

> and if the motivation is to make it simpler this implementation
> requires adding bindings and bpf programs which that are the same as
> native panel drivers, but won't be able to work until user-space
> starts and will require to be in initramfs or rootfs to have
> functional display.
> 
> Not sure to understand what are the positive features here except
> isolating the panel code in a safe bpf program (is this really needed
> ?).

From the cover letter:

"""
Panels in general, and MIPI-DSI panels in particular, are pretty
difficult to support and require pretty much a panel driver for each
panel produced. Most of them are pretty simple, and require an opaque
initialization sequence that is usually poorly documented.

This creates a tension between OEMs and distros because OEMs will
typically get a new panel to react to a sourcing issue during
production, and thus need some swift turnaround between getting their
new panel and it being operational in the OS. Distributions on the other
hand can take years to ship a kernel with that new panel driver.

To solve this, I followed the example of HID-BPF and wrote a panel
driver that will rely on BPF programs to perform the panel
initialization. That way, we can ship the programs separately from the
kernel, and with a different lifecycle.
"""

You can disagree with the solution, that's fair, we can also discuss on
how to solve this problem, but I'd appreciate it if you weren't claiming
it's all useless.

Maxime