Skip to contentSkip to conversation

Build failure with Vivante and QT5 w/ eglfs

Buildroot Antoine Tenart 4 participants 12 messages Mar 4, 2016 – Mar 7, 2016

lore
Share & tools
lore
#0 lore raw

Hi all,

I've run into a build failure while compiling an image, when building QT5. I made some build tests to narrow down the issue and found the configuration leading to this: when compiling an image with both BR2_PACKAGE_QT5BASE_EGLFS and BR2_PACKAGE_IMX_GPU_VIV_OUTPUT_X11 the build fails with the following error:

http://code.bulix.org/qxj1bn-92391?raw

I understand this configuration can't work, but as this leads to a build issue we might need to think about a solution. We could disallow such a configuration tuple, or fix the build.

Any idea?

Thanks,

Antoine

Signature · 10 lines
Antoine T?nart, Free Electrons
Embedded Linux, Kernel and Android engineering
http://free-electrons.com
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 819 bytes
Desc: Digital signature
URL: <http://lists.busybox.net/pipermail/buildroot/attachments/20160304/3840cc84/attachment.asc>
#1 lore raw

Antoine, All,

Adding J?r?me in cc since he's the one behind the x11/fb selection as well as qt5 graphics support for i.MX.

Quoted 9 lines · Antoine Tenart

Hi all,

I've run into a build failure while compiling an image, when building QT5. I made some build tests to narrow down the issue and found the configuration leading to this: when compiling an image with both BR2_PACKAGE_QT5BASE_EGLFS and BR2_PACKAGE_IMX_GPU_VIV_OUTPUT_X11 the build fails with the following error:

http://code.bulix.org/qxj1bn-92391?raw

Indeed that's a problem. It looks like the question has been raised in the Yocto community not so long ago:
https://github.com/Freescale/meta-fsl-arm/commit/da8f520a

It looks like their patch is to define EGL_API_FB although X11 is in use. I imagine we could do the same using J?r?me's approach but depending on EGLFS define:
https://git.buildroot.net/buildroot/tree/package/freescale-imx/imx-gpu-viv/imx-gpu-viv.mk#n55

However I guess that would only fix the build but would cause issues at runtime which might be confusing for users.

Quoted 5 lines

I understand this configuration can't work, but as this leads to a build issue we might need to think about a solution. We could disallow such a configuration tuple, or fix the build.

Any idea?

I'm afraid forcing the imx-gpu-viv package output format depending on qt5 graphical back end would bring many "recursive dependency" issues.

J?r?me, what are your thoughts?

Thanks,
Gary

#2 lore raw

Gary,

Quoted 8 lines · Gary Bisson
Quoted 5 lines

I understand this configuration can't work, but as this leads to a build issue we might need to think about a solution. We could disallow such a configuration tuple, or fix the build.

Any idea?

I'm afraid forcing the imx-gpu-viv package output format depending on qt5 graphical back end would bring many "recursive dependency" issues.

I don't see why it would create recursive dependency issues.

However, it would violate the principle of virtual packages, since Qt5 is not supposed to have some dependencies that are specific to certain OpenGL implementations. It should normally just use dependencies on the OpenGL ES / EGL packages, and nothing else.

However, we could probably add some exceptions to this rule. Or rework the virtual package logic so that consumers of OpenGL can now whether the FB or X11 support is available.

Thomas

Signature · 3 lines
Thomas Petazzoni, CTO, Free Electrons
Embedded Linux, Kernel and Android engineering
http://free-electrons.com
#3 lore raw

Thomas, All,

Quoted 8 lines · Thomas Petazzoni

Gary,

Quoted 4 lines · Gary Bisson
Quoted 4 lines

I understand this configuration can't work, but as this leads to a build issue we might need to think about a solution. We could disallow such a configuration tuple, or fix the build.

Quoted 6 lines
Quoted 4 lines
Quoted 1 line

Any idea?

I'm afraid forcing the imx-gpu-viv package output format depending on qt5 graphical back end would bring many "recursive dependency" issues.

I don't see why it would create recursive dependency issues.

Well my thought process was that it was kind of forbidden to modify the qtbase package for a i.MX-specific corner case.

So I tried to force it in imx-gpu-viv to use fb depending on BR2_PACKAGE_QT5BASE_EGLFS but it only brought those recursive issues.

Quoted 8 lines

However, it would violate the principle of virtual packages, since Qt5 is not supposed to have some dependencies that are specific to certain OpenGL implementations. It should normally just use dependencies on the OpenGL ES / EGL packages, and nothing else.

However, we could probably add some exceptions to this rule. Or rework the virtual package logic so that consumers of OpenGL can now whether the FB or X11 support is available.

Well if we are allowed to modify qtbase I guess it would be way easier. Would something like this be acceptable? index 64a7f65..aa54b3e 100644

package/qt5/qt5base/Config.in +4 -0
@@ -201,10 +201,14 @@ config BR2_PACKAGE_QT5BASE_EGLFS
        select BR2_PACKAGE_QT5BASE_OPENGL
        depends on BR2_PACKAGE_HAS_LIBEGL
        depends on BR2_PACKAGE_QT5_GL_AVAILABLE
+       depends on !BR2_PACKAGE_IMX_GPU_VIV_OUTPUT_X11
 
 comment "eglfs backend available if OpenGL and EGL are enabled"
        depends on !BR2_PACKAGE_HAS_LIBEGL || !BR2_PACKAGE_QT5_GL_AVAILABLE
 
+comment "eglfs backend not available for X11 version of i.MX 3D libraries"
+       depends on BR2_PACKAGE_IMX_GPU_VIV_OUTPUT_X11
+
 config BR2_PACKAGE_QT5BASE_DEFAULT_QPA
        string "Default graphical platform"
        help

Regards,
Gary

#4 lore raw

Hello,

Quoted 5 lines · Gary Bisson

Well my thought process was that it was kind of forbidden to modify the qtbase package for a i.MX-specific corner case.

So I tried to force it in imx-gpu-viv to use fb depending on BR2_PACKAGE_QT5BASE_EGLFS but it only brought those recursive issues.

You indeed can't do this, as it creates a circular dependency.

Quoted 7 lines

Well if we are allowed to modify qtbase I guess it would be way easier. Would something like this be acceptable? index 64a7f65..aa54b3e 100644

package/qt5/qt5base/Config.in +0 -0
@@ -201,10 +201,14 @@ config BR2_PACKAGE_QT5BASE_EGLFS
        select BR2_PACKAGE_QT5BASE_OPENGL

package/qt5/qt5base/Config.in config BR2_PACKAGE_QT5BASE_EGLFS · line 211

Quoted 10 lines
package/qt5/qt5base/Config.in +4 -0
        depends on BR2_PACKAGE_HAS_LIBEGL
        depends on BR2_PACKAGE_QT5_GL_AVAILABLE
+       depends on !BR2_PACKAGE_IMX_GPU_VIV_OUTPUT_X11
 
 comment "eglfs backend available if OpenGL and EGL are enabled"
        depends on !BR2_PACKAGE_HAS_LIBEGL || !BR2_PACKAGE_QT5_GL_AVAILABLE
 
+comment "eglfs backend not available for X11 version of i.MX 3D libraries"
+       depends on BR2_PACKAGE_IMX_GPU_VIV_OUTPUT_X11
+

No this is not good, because there are other OpenGL implementations than imx-gpu-viv. So at the very least, you would need something like the below. I also believe it makes more sense to use positive condition, i.e express a dependency on the FB output, rather than a dependency on not having the X11 output.

	# If the imx-gpu-viv implementation is used, we need its FB variant
	depends on !BR2_PACKAGE_IMX_GPU_VIV || BR2_PACKAGE_IMX_GPU_VIV_OUTPUT_FB

comment "eglfs backend needs the FB output of i.MX GPU libraries"

	depends on BR2_PACKAGE_IMX_GPU_VIV && !BR2_PACKAGE_IMX_GPU_VIV_OUTPUT_FB

Best regards,

Thomas

Signature · 3 lines
Thomas Petazzoni, CTO, Free Electrons
Embedded Linux, Kernel and Android engineering
http://free-electrons.com
#5 lore raw

Thomas, All,

Quoted 36 lines · Thomas Petazzoni

Hello,

Quoted 5 lines · Gary Bisson

Well my thought process was that it was kind of forbidden to modify the qtbase package for a i.MX-specific corner case.

So I tried to force it in imx-gpu-viv to use fb depending on BR2_PACKAGE_QT5BASE_EGLFS but it only brought those recursive issues.

You indeed can't do this, as it creates a circular dependency.

Quoted 7 lines

Well if we are allowed to modify qtbase I guess it would be way easier. Would something like this be acceptable? index 64a7f65..aa54b3e 100644

package/qt5/qt5base/Config.in +0 -0
@@ -201,10 +201,14 @@ config BR2_PACKAGE_QT5BASE_EGLFS
        select BR2_PACKAGE_QT5BASE_OPENGL

package/qt5/qt5base/Config.in config BR2_PACKAGE_QT5BASE_EGLFS · line 211

Quoted 10 lines
package/qt5/qt5base/Config.in +4 -0
        depends on BR2_PACKAGE_HAS_LIBEGL
        depends on BR2_PACKAGE_QT5_GL_AVAILABLE
+       depends on !BR2_PACKAGE_IMX_GPU_VIV_OUTPUT_X11
 
 comment "eglfs backend available if OpenGL and EGL are enabled"
        depends on !BR2_PACKAGE_HAS_LIBEGL || !BR2_PACKAGE_QT5_GL_AVAILABLE
 
+comment "eglfs backend not available for X11 version of i.MX 3D libraries"
+       depends on BR2_PACKAGE_IMX_GPU_VIV_OUTPUT_X11
+

No this is not good, because there are other OpenGL implementations than imx-gpu-viv. So at the very least, you would need something like the below. I also believe it makes more sense to use positive condition, i.e express a dependency on the FB output, rather than a dependency on not having the X11 output.

Quoted 5 lines

# If the imx-gpu-viv implementation is used, we need its FB variant depends on !BR2_PACKAGE_IMX_GPU_VIV || BR2_PACKAGE_IMX_GPU_VIV_OUTPUT_FB

comment "eglfs backend needs the FB output of i.MX GPU libraries"

        depends on BR2_PACKAGE_IMX_GPU_VIV && !BR2_PACKAGE_IMX_GPU_VIV_OUTPUT_FB

I agree using the positive condition is better. However I'm not sure to understand why !BR2_PACKAGE_IMX_GPU_VIV_OUTPUT_X11 wouldn't work with other OpenGL implementations. This variable is only set when selecting the vivante package + x11 output. So for any other implementation it wouldn't be set. It's just out of curiosity that I'd like to understand the difference.

Regards,
Gary

#6 lore raw

Gary,

Quoted 30 lines · Gary Bisson
Quoted 24 lines
Quoted 7 lines

Well if we are allowed to modify qtbase I guess it would be way easier. Would something like this be acceptable? index 64a7f65..aa54b3e 100644

package/qt5/qt5base/Config.in +0 -0
@@ -201,10 +201,14 @@ config BR2_PACKAGE_QT5BASE_EGLFS
        select BR2_PACKAGE_QT5BASE_OPENGL

package/qt5/qt5base/Config.in config BR2_PACKAGE_QT5BASE_EGLFS · line 211

Quoted 10 lines
package/qt5/qt5base/Config.in +4 -0
        depends on BR2_PACKAGE_HAS_LIBEGL
        depends on BR2_PACKAGE_QT5_GL_AVAILABLE
+       depends on !BR2_PACKAGE_IMX_GPU_VIV_OUTPUT_X11
 
 comment "eglfs backend available if OpenGL and EGL are enabled"
        depends on !BR2_PACKAGE_HAS_LIBEGL || !BR2_PACKAGE_QT5_GL_AVAILABLE
 
+comment "eglfs backend not available for X11 version of i.MX 3D libraries"
+       depends on BR2_PACKAGE_IMX_GPU_VIV_OUTPUT_X11
+

No this is not good, because there are other OpenGL implementations than imx-gpu-viv. So at the very least, you would need something like the below. I also believe it makes more sense to use positive condition, i.e express a dependency on the FB output, rather than a dependency on not having the X11 output.

Quoted 5 lines

# If the imx-gpu-viv implementation is used, we need its FB variant depends on !BR2_PACKAGE_IMX_GPU_VIV || BR2_PACKAGE_IMX_GPU_VIV_OUTPUT_FB

comment "eglfs backend needs the FB output of i.MX GPU libraries"

        depends on BR2_PACKAGE_IMX_GPU_VIV && !BR2_PACKAGE_IMX_GPU_VIV_OUTPUT_FB
Quoted 6 lines

I agree using the positive condition is better. However I'm not sure to understand why !BR2_PACKAGE_IMX_GPU_VIV_OUTPUT_X11 wouldn't work with other OpenGL implementations. This variable is only set when selecting the vivante package + x11 output. So for any other implementation it wouldn't be set. It's just out of curiosity that I'd like to understand the difference.

Aah, yes, you're right. Your version is functionally equivalent to mine.

Thomas

Signature · 3 lines
Thomas Petazzoni, CTO, Free Electrons
Embedded Linux, Kernel and Android engineering
http://free-electrons.com
#7 lore raw
In reply to Gary Bisson
Quoted 15 lines · Gary Bisson
Quoted 8 lines · Antoine Tenart

I've run into a build failure while compiling an image, when building QT5. I made some build tests to narrow down the issue and found the configuration leading to this: when compiling an image with both BR2_PACKAGE_QT5BASE_EGLFS and BR2_PACKAGE_IMX_GPU_VIV_OUTPUT_X11 the build fails with the following error:

http://code.bulix.org/qxj1bn-92391?raw

Indeed that's a problem. It looks like the question has been raised in the Yocto community not so long ago:
https://github.com/Freescale/meta-fsl-arm/commit/da8f520a

Quoted 7 lines

It looks like their patch is to define EGL_API_FB although X11 is in use. I imagine we could do the same using J?r?me's approach but depending on EGLFS define:
https://git.buildroot.net/buildroot/tree/package/freescale-imx/imx-gpu-viv/imx-gpu-viv.mk#n55

However I guess that would only fix the build but would cause issues at runtime which might be confusing for users.

It's not possible to ensure every possible configuration will be working. A simple example would be to compile a bootloader for a board A and a kernel for a completely different board B.

Here, the configuration leading to this build error does not make sense. So if we only fix the build, leading to an unusable image I guess that would be OK. (And fixing the build is important, for build bots for example).

Another solution, fixing both the build and the generated image could be to define other virtual packages for the EGL implementation. We could end up with the following for imx-gpu-viv:

select BR2_PACKAGE_HAS_LIBEGL
select BR2_PACKAGE_HAS_LIBEGL_X11

and

select BR2_PACKAGE_HAS_LIBEGL
select BR2_PACKAGE_HAS_LIBEGL_FB

This would still allow packages to depend on BR2_PACKAGE_HAS_LIBEGL, while allowing a fine dependency check given the EGL output.

Exact implementation of these new virtual packages can be debated if the solution is approved. (Like do we make them select BR2_PACKAGE_HAS_LIBEGL or do we still select it manually in all the packages providing an EGL implementation).

The drawback is we need to patch all EGL implementations, to fix a build issue introduced by only one of them.

Antoine

Signature · 10 lines
Antoine T?nart, Free Electrons
Embedded Linux, Kernel and Android engineering
http://free-electrons.com
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 819 bytes
Desc: Digital signature
URL: <http://lists.busybox.net/pipermail/buildroot/attachments/20160304/032c60cb/attachment.asc>
#8 lore raw

Antoine, All,

Quoted 24 lines · Antoine Tenart
Quoted 15 lines · Gary Bisson
Quoted 8 lines · Antoine Tenart

I've run into a build failure while compiling an image, when building QT5. I made some build tests to narrow down the issue and found the configuration leading to this: when compiling an image with both BR2_PACKAGE_QT5BASE_EGLFS and BR2_PACKAGE_IMX_GPU_VIV_OUTPUT_X11 the build fails with the following error:

http://code.bulix.org/qxj1bn-92391?raw

Indeed that's a problem. It looks like the question has been raised in the Yocto community not so long ago:
https://github.com/Freescale/meta-fsl-arm/commit/da8f520a

Quoted 7 lines

It looks like their patch is to define EGL_API_FB although X11 is in use. I imagine we could do the same using J?r?me's approach but depending on EGLFS define:
https://git.buildroot.net/buildroot/tree/package/freescale-imx/imx-gpu-viv/imx-gpu-viv.mk#n55

However I guess that would only fix the build but would cause issues at runtime which might be confusing for users.

Quoted 8 lines

It's not possible to ensure every possible configuration will be working. A simple example would be to compile a bootloader for a board A and a kernel for a completely different board B.

Here, the configuration leading to this build error does not make sense. So if we only fix the build, leading to an unusable image I guess that would be OK. (And fixing the build is important, for build bots for example).

I agree but here Thomas gave a green light on modifying the qtbase package directly for this corner case which fixes both the build and possible runtime issues.

BTW, who's making/sending the patch now?

Regards,
Gary

#11 lore raw
In reply to Gary Bisson

Hello Gary,

Quoted 36 lines · On Friday 04 March 2016 12:02:31 Gary Bisson wrote:

Antoine, All,

Adding J?r?me in cc since he's the one behind the x11/fb selection as well as qt5 graphics support for i.MX.

On Fri, Mar 4, 2016 at 10:55 AM, Antoine Tenart

Quoted 10 lines · <[email protected]> wrote:

Hi all,

I've run into a build failure while compiling an image, when building
QT5. I made some build tests to narrow down the issue and found the configuration leading to this: when compiling an image with both BR2_PACKAGE_QT5BASE_EGLFS and BR2_PACKAGE_IMX_GPU_VIV_OUTPUT_X11 the build fails with the following error:

http://code.bulix.org/qxj1bn-92391?raw

Indeed that's a problem. It looks like the question has been raised in the Yocto community not so long ago:
https://github.com/Freescale/meta-fsl-arm/commit/da8f520a

It looks like their patch is to define EGL_API_FB although X11 is in use. I imagine we could do the same using J?r?me's approach but depending on EGLFS define:
https://git.buildroot.net/buildroot/tree/package/freescale-imx/imx-gpu
-viv/imx-gpu-viv.mk#n55

However I guess that would only fix the build but would cause issues at runtime which might be confusing for users.

Quoted 4 lines

I understand this configuration can't work, but as this leads to a build issue we might need to think about a solution. We could disallow such a configuration tuple, or fix the build.

Quoted 6 lines
Quoted 1 line

Any idea?

I'm afraid forcing the imx-gpu-viv package output format depending on qt5 graphical back end would bring many "recursive dependency" issues.

J?r?me, what are your thoughts?

An alternative solution would be to rename EGL_API_FB into MESA_EGL_NO_X11_HEADERS. Indeed, since this variable is used by Mesa headers, it is the de facto standard way to disable X11 headers inclusion.
As far as I remember, Qt5 correctly set MESA_EGL_NO_X11_HEADERS when it compile EGLFS.

IMHO, it is the most elegant way to solve compilation problem. It would allow to compile Qt5 with both X11 and EGLFS enabled (sure, only one backend will work at runtime). However, I afraid to do such deep change in Vivante headers.

BR,

Signature · 3 lines
J?r?me Pouiller, Sysmic
Embedded Linux specialist
http://www.sysmic.fr
#9 lore raw
In reply to Gary Bisson
Quoted 30 lines · Gary Bisson
Quoted 28 lines · Antoine Tenart
Quoted 15 lines · Gary Bisson
Quoted 8 lines · Antoine Tenart

I've run into a build failure while compiling an image, when building QT5. I made some build tests to narrow down the issue and found the configuration leading to this: when compiling an image with both BR2_PACKAGE_QT5BASE_EGLFS and BR2_PACKAGE_IMX_GPU_VIV_OUTPUT_X11 the build fails with the following error:

http://code.bulix.org/qxj1bn-92391?raw

Indeed that's a problem. It looks like the question has been raised in the Yocto community not so long ago:
https://github.com/Freescale/meta-fsl-arm/commit/da8f520a

Quoted 7 lines

It looks like their patch is to define EGL_API_FB although X11 is in use. I imagine we could do the same using J?r?me's approach but depending on EGLFS define:
https://git.buildroot.net/buildroot/tree/package/freescale-imx/imx-gpu-viv/imx-gpu-viv.mk#n55

However I guess that would only fix the build but would cause issues at runtime which might be confusing for users.

It's not possible to ensure every possible configuration will be working. A simple example would be to compile a bootloader for a board A and a kernel for a completely different board B.

Quoted 8 lines
Quoted 4 lines

Here, the configuration leading to this build error does not make sense. So if we only fix the build, leading to an unusable image I guess that would be OK. (And fixing the build is important, for build bots for example).

I agree but here Thomas gave a green light on modifying the qtbase package directly for this corner case which fixes both the build and possible runtime issues.

Reading Thomas' mail, I don't think he gave a green light on this.

So we still have to choose a solution to fix this (which may be the one directly modifying the qt5base package, but we first need to agree on the solution).

Antoine

Signature · 10 lines
Antoine T?nart, Free Electrons
Embedded Linux, Kernel and Android engineering
http://free-electrons.com
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 819 bytes
Desc: Digital signature
URL: <http://lists.busybox.net/pipermail/buildroot/attachments/20160307/c534fe30/attachment.asc>
#10 lore raw

Antoine, All,

Quoted 37 lines · Antoine Tenart
Quoted 35 lines · Gary Bisson
Quoted 32 lines · Antoine Tenart
Quoted 15 lines · Gary Bisson
Quoted 8 lines · Antoine Tenart

I've run into a build failure while compiling an image, when building QT5. I made some build tests to narrow down the issue and found the configuration leading to this: when compiling an image with both BR2_PACKAGE_QT5BASE_EGLFS and BR2_PACKAGE_IMX_GPU_VIV_OUTPUT_X11 the build fails with the following error:

http://code.bulix.org/qxj1bn-92391?raw

Indeed that's a problem. It looks like the question has been raised in the Yocto community not so long ago:
https://github.com/Freescale/meta-fsl-arm/commit/da8f520a

Quoted 7 lines

It looks like their patch is to define EGL_API_FB although X11 is in use. I imagine we could do the same using J?r?me's approach but depending on EGLFS define:
https://git.buildroot.net/buildroot/tree/package/freescale-imx/imx-gpu-viv/imx-gpu-viv.mk#n55

However I guess that would only fix the build but would cause issues at runtime which might be confusing for users.

It's not possible to ensure every possible configuration will be working. A simple example would be to compile a bootloader for a board A and a kernel for a completely different board B.

Here, the configuration leading to this build error does not make sense. So if we only fix the build, leading to an unusable image I guess that would be OK. (And fixing the build is important, for build bots for example).

Quoted 5 lines
Quoted 3 lines

I agree but here Thomas gave a green light on modifying the qtbase package directly for this corner case which fixes both the build and possible runtime issues.

Reading Thomas' mail, I don't think he gave a green light on this.

Green light might be an overstatement indeed, but let's say he's not against this kind of solution.

Quoted 3 lines

So we still have to choose a solution to fix this (which may be the one directly modifying the qt5base package, but we first need to agree on the solution).

I agree, I think we have 2 viable solutions so far:
1) modifying the qt5base for this corner which fixes both the build and runtime issues
2) modifying the eglvivante.h which is a more elegant solution but only fixes the build, not the runtime issues.

I'm in favor of #1 because it's easy to implement, fixes both runtime and build plus I personally don't have much time to test #2.

Regards,
Gary