Hi Javier,
On Thu, Sep 12, 2024 at 06:33:58PM +0200, Javier Martinez Canillas wrote:
That's a very good point. I'm actually not familiar with Coreboot and I used an educated guess (in the case of DT for example, that's the main source of truth and I didn't know if a Core table was in a similar vein).
Maybe something like the following (untested) patch then?
Julius is more familiar with the Coreboot + payload ecosystem than me, but his explanations make sense to me, as does this patch.
From de1c32017006f4671d91b695f4d6b4e99c073ab2 Mon Sep 17 00:00:00 2001 From: Javier Martinez Canillas javierm@redhat.com Date: Thu, 12 Sep 2024 18:31:55 +0200 Subject: [PATCH] firmware: coreboot: Don't register a pdev if screen_info data is available
On Coreboot platforms, a system framebuffer may be provided to the Linux kernel by filling a LB_TAG_FRAMEBUFFER entry in the Coreboot table. But a Coreboot payload (e.g: SeaBIOS) could also provide this information to the Linux kernel.
If that the case, early arch x86 boot code will fill the global struct screen_info data and that data used by the Generic System Framebuffers (sysfb) framework to add a platform device with platform data about the system framebuffer.
Normally, these sorts of "early" and "later" ordering descriptions would set alarm bells when talking about independent drivers. But I suppose the "early arch" code has better ordering guaranteeds than drivers, so this should be fine.
But later then the framebuffer_coreboot driver will try to do the same framebuffer (using the information from the Coreboot table), which will lead to an error due a simple-framebuffer.0 device already registered:
sysfs: cannot create duplicate filename '/bus/platform/devices/simple-framebuffer.0' ... coreboot: could not register framebuffer framebuffer coreboot8: probe with driver framebuffer failed with error -17To prevent the issue, make the framebuffer_core driver to not register a platform device if the global struct screen_info data has been filled.
Reported-by: Brian Norris briannorris@chromium.org Link: https://lore.kernel.org/all/ZuCG-DggNThuF4pj@b20ea791c01f/T/#ma7fb65acbc1a56... Suggested-by: Julius Werner jwerner@chromium.org Signed-off-by: Javier Martinez Canillas javierm@redhat.com
drivers/firmware/google/framebuffer-coreboot.c | 16 ++++++++++++++++ 1 file changed, 16 insertions(+)
diff --git a/drivers/firmware/google/framebuffer-coreboot.c b/drivers/firmware/google/framebuffer-coreboot.c index daadd71d8ddd..4e50da17cd7e 100644 --- a/drivers/firmware/google/framebuffer-coreboot.c +++ b/drivers/firmware/google/framebuffer-coreboot.c @@ -15,6 +15,7 @@ #include <linux/module.h> #include <linux/platform_data/simplefb.h> #include <linux/platform_device.h> +#include <linux/screen_info.h> #include "coreboot_table.h" @@ -27,6 +28,7 @@ static int framebuffer_probe(struct coreboot_device *dev) int i; u32 length; struct lb_framebuffer *fb = &dev->framebuffer;
- struct screen_info *si = &screen_info; struct platform_device *pdev; struct resource res; struct simplefb_platform_data pdata = {
@@ -36,6 +38,20 @@ static int framebuffer_probe(struct coreboot_device *dev) .format = NULL, };
- /*
* If the global screen_info data has been filled, the Generic* System Framebuffers (sysfb) will already register a platform
Did you mean 'platform_device'?
* and pass the screen_info as platform_data to a driver that* could scan-out using the system provided framebuffer.** On Coreboot systems, the advertise LB_TAG_FRAMEBUFFER entry
s/advertise/advertised/ ?
* in the Coreboot table should only be used if the payload did* not set video mode info and passed it to the Linux kernel.
s/passed/pass/
*/- if (si->orig_video_isVGA == VIDEO_TYPE_VLFB ||
si->orig_video_isVGA == VIDEO_TYPE_EFI)
This line is using spaces for indentation. It should use a tab, and then spaces for alignment. But presumably this will change based on Thomas's suggestions anyway.
return -EINVAL;
Is EINVAL right? IIUC, that will print a noisier error to the logs. I believe the "expected" sorts of return codes are ENODEV or ENXIO. (See call_driver_probe().) ENODEV seems like a fine choice, similar to several of the other return codes already used here.
Anyway, this seems along the right track. Thanks for tackling, and feel free to carry a:
Reviewed-by: Brian Norris briannorris@chromium.org
- if (!fb->physical_address) return -ENODEV;
Best regards,
Javier Martinez Canillas Core Platforms Red Hat