On Mon, Sep 21, 2020 at 9:10 PM Peter Collingbourne pcc@google.com wrote:
On Mon, Sep 21, 2020 at 2:32 PM Linus Walleij linus.walleij@linaro.org wrote:
On Tue, Sep 15, 2020 at 3:10 AM Peter Collingbourne pcc@google.com wrote:
On Tue, Jun 9, 2020 at 1:08 PM Linus Walleij linus.walleij@linaro.org wrote:
All the functionality in this driver has been reimplemented in the new DRM driver in drivers/gpu/drm/pl111/* and all the boards using it have been migrated to use the DRM driver with all configuration coming from the device tree.
Android's FVP configuration still uses FBDEV.
But all DRM drivers support frame buffer emulation so I don't see how this can be a problem? The new DRM driver provides a framebuffer.
Okay, I was unaware of that feature and it is disabled in Android kernels. I am now testing the DRM driver by applying this diff to FVP's kernel config:
diff --git a/fvp.fragment b/fvp.fragment index b12be37278a4..481767c06ac9 100644 --- a/fvp.fragment +++ b/fvp.fragment @@ -1,5 +1,6 @@ -CONFIG_FB=m -CONFIG_FB_ARMCLCD=m +CONFIG_DRM_PANEL_SIMPLE=m +CONFIG_DRM_PL111=m +CONFIG_DRM_FBDEV_EMULATION=y CONFIG_MOUSE_PS2=m CONFIG_SERIO_AMBAKMI=m CONFIG_MMC_ARMMMCI=m
I'm also confused about how this can be in use. fvp-base-revc.dts includes rtsm_ve-motherboard.dtsi where the PL111 is defined.
commit f1fe12c8bf33241e04c57a0fc5b25b16ba77ba2d "ARM: dts: Modernize the Vexpress PL111 integration" changes the DTS for the FVP such that only the new DRM driver really works with it: it removes the panel-dpi encoded panel and defines the panel "arm,rtsm-display" for FVP and that is *only* supported by drivers/gpu/drm/panel/panel-simple.c from DRM.
That commit predates commit fa083b99eb284186ae65193ae856ef2801a1646d which added the DTS for FVP, along with a panel-dpi compatible panel which is still present: https://github.com/torvalds/linux/blob/98477740630f270aecf648f1d6a9dbc6027d4...
The upstream FVP has been using the new DRM driver by default since commit 37ad688497785c9ad1471236e28efda1e2f39741 "arm64: defconfig: Switch to PL11x DRM driver" in january 2019.
That only affected defconfig though which Android doesn't use.
Are you using old or outoftree devicetrees with a newer kernel?
No, I'm using an up-to-date device tree (pretty sure about that otherwise I wouldn't have hit problems like https://android-review.googlesource.com/c/device/generic/goldfish/+/1394347 ).
While it would be great to see it migrated to DRM, this first requires resolving a design incompatibility between Android's composer and DRM devices that only support software rendering. I proposed a patch that would help resolve this [1], but there was disagreement about the approach and I haven't had time to get back to this.
Can this please be reverted until FVP on Android can be migrated to DRM?
We need a clear technical reason, I understand that using DRM directly might be a problem, but DRM supports full framebuffer emulation and from a userspace point of view, what the new driver provides should be equivalent.
I can think of problems like Android seeing the new fancy DRM userspace ABI but surely it can be instructed to ignore that and just use the framebuffer emulation instead?
If there is anything else making DRMs framebuffer emulation substandard I am sure the DRM developers would like to know, especially if it makes Android unhappy.
Okay, I found at least two issues so far. The first is that the driver does not come up in 32bpp mode, seemingly because of this line of code here: https://github.com/torvalds/linux/blob/98477740630f270aecf648f1d6a9dbc6027d4...
The second problem seems to be that Android's calls to ioctl(FBIOPUT_VSCREENINFO) fail. I thought that this might be related to the declared lack of 32bpp support, but even with the above line changed to .fb_bpp = 32 (probably not the right fix given the comment above it would probably break any users of the "real" Versatile Express board; the driver probably instead needs to select the bpp based on the max-memory-bandwidth like the old fbdev driver did) I still get this logcat output:
09-22 01:31:08.807 272 272 W gralloc : FBIOPUT_VSCREENINFO failed, page flipping not supported 09-22 01:31:08.807 272 272 W gralloc : page flipping not supported (yres_virtual=768, requested=1536) 09-22 01:31:08.807 272 272 I gralloc : using (fd=7) 09-22 01:31:08.807 272 272 I gralloc : id = pl111drmfb 09-22 01:31:08.807 272 272 I gralloc : xres = 1024 px 09-22 01:31:08.807 272 272 I gralloc : yres = 768 px 09-22 01:31:08.807 272 272 I gralloc : xres_virtual = 1024 px 09-22 01:31:08.807 272 272 I gralloc : yres_virtual = 768 px 09-22 01:31:08.807 272 272 I gralloc : bpp = 32 09-22 01:31:08.807 272 272 I gralloc : r = 16:8 09-22 01:31:08.807 272 272 I gralloc : g = 8:8 09-22 01:31:08.807 272 272 I gralloc : b = 0:8 09-22 01:31:08.807 272 272 I gralloc : width = 400 mm (65.024002 dpi) 09-22 01:31:08.807 272 272 I gralloc : height = 300 mm (65.023994 dpi) 09-22 01:31:08.807 272 272 I gralloc : refresh rate = 60.00 Hz
which is coming from the code starting here: https://cs.android.com/android/platform/superproject/+/master:hardware/libha... (you can ignore the part of the log message that talks about page flipping; in this context the main purpose of the ioctl is to give effect to the code on lines 166-172 that sets the pixel format, and as can be seen from the subsequent output the format remains BGRX888 rather than RGBX8888).
The output from a successful boot with the old fbdev driver looks like this:
09-22 00:54:57.750 272 272 W gralloc : page flipping not supported (yres_virtual=768, requested=1536) 09-22 00:54:57.750 272 272 I gralloc : using (fd=7) 09-22 00:54:57.750 272 272 I gralloc : id = CLCD FB 09-22 00:54:57.750 272 272 I gralloc : xres = 1024 px 09-22 00:54:57.750 272 272 I gralloc : yres = 768 px 09-22 00:54:57.750 272 272 I gralloc : xres_virtual = 1024 px 09-22 00:54:57.750 272 272 I gralloc : yres_virtual = 768 px 09-22 00:54:57.750 272 272 I gralloc : bpp = 32 09-22 00:54:57.750 272 272 I gralloc : r = 0:8 09-22 00:54:57.750 272 272 I gralloc : g = 8:8 09-22 00:54:57.750 272 272 I gralloc : b = 16:8 09-22 00:54:57.750 272 272 I gralloc : width = 163 mm (159.568100 dpi) 09-22 00:54:57.750 272 272 I gralloc : height = 122 mm (159.895081 dpi) 09-22 00:54:57.750 272 272 I gralloc : refresh rate = 65.34 Hz
which aside from the pixel format indicates some differences with the width, height and refresh rate which may be significant.
As a result of FBIOPUT_VSCREENINFO failing, the FVP window fails to resize so I don't see any graphical output (this ioctl in the fbdev driver set the magic registers that caused FVP to resize its window; my investigation into DRM earlier this year revealed the equivalent register-setting operation in DRM to be DRM_IOCTL_MODE_ATOMIC).
I'll try to look at this closer tomorrow to see whether Android is doing something wrong, but it seems plausible that DRM's FBDEV emulation is missing a critical feature, at least for pl111.
This does indeed appear to be the case. Among other reasons, Android's ioctl would fail because of this code, which forbids changing the pixel format in fbdev emulation: https://github.com/torvalds/linux/blob/eff48ddeab782e35e58ccc8853f7386bbae9d...
Android's generic fbdev userspace driver (which FVP uses) is only compatible with the RGBX8888 pixel format. So there is a fundamental incompatibility here, as long as the kernel driver comes up in a different format. We could change the driver's default format to match what Android requires, but then that could break some other application that requires a format other than RGBX8888, which would fail when changing to its required format because of the above linked code.
There is a less important issue where the new driver fails the syscall if yres_virtual exceeds a limit: https://github.com/torvalds/linux/blob/eff48ddeab782e35e58ccc8853f7386bbae9d...
whereas the old driver would reset yres_virtual based on the value of yres: https://github.com/torvalds/linux/blob/3f1f6981afed9fa21efa12ce396b35ca684b8...
With the current userspace implementation which sets yres_virtual = yres * 2 this leads to the driver not even reaching the code that rejects based on pixel format. This can be fixed on the userspace side (by reissuing the syscall with yres_virtual set to its original value) but API stability would seem to require the behavior to match here as well, and this of course does nothing for the pixel format issue.
So I think we should bring the old driver back, at least until fbdev emulation gains the ability to change the pixel format.
Peter