| Message ID | 20230426130420.19942-1-tzimmermann@suse.de |
|---|---|
| Headers | show |
| Series | fbdev: Move framebuffer I/O helpers to <asm/fb.h> | expand |
Hi Thomas. On Wed, Apr 26, 2023 at 03:04:15PM +0200, Thomas Zimmermann wrote: > Fbdev provides helpers for framebuffer I/O, such as fb_readl(), > fb_writel() or fb_memcpy_to_fb(). The implementation of each helper > depends on the architecture. It's still all located in fbdev's main > header file <linux/fb.h>. Move all of it into each archtecture's > <asm/fb.h>, with shared code in <asm-generic/fb.h>. For once I think this cleanup is moving things in the wrong direction. The fb_* helpers predates the generic io.h support and try to add a generic layer for read read / write operations. The right fix would be to migrate fb_* to use the io helpers we have today - so we use the existing way to handle the architecture specific details. From a quick look there seems to be some challenges but the current helpers that re-do part of io.h is not the way forward and hiding them in arch/include/asm/fb.h seems counter productive. Sam
Hi Sam Am 26.04.23 um 21:21 schrieb Sam Ravnborg: > Hi Thomas. > > On Wed, Apr 26, 2023 at 03:04:15PM +0200, Thomas Zimmermann wrote: >> Fbdev provides helpers for framebuffer I/O, such as fb_readl(), >> fb_writel() or fb_memcpy_to_fb(). The implementation of each helper >> depends on the architecture. It's still all located in fbdev's main >> header file <linux/fb.h>. Move all of it into each archtecture's >> <asm/fb.h>, with shared code in <asm-generic/fb.h>. > > For once I think this cleanup is moving things in the wrong direction. > > The fb_* helpers predates the generic io.h support and try to > add a generic layer for read read / write operations. > > The right fix would be to migrate fb_* to use the io helpers > we have today - so we use the existing way to handle the architecture > specific details. I looked through the existing versions of the fb_() I/O helpers. They can apparently be implemented with the regular helpers of similar names. I'm not sure, but even Sparc looks compatible. At least these sbus_ functions seem to be equivalent to the __raw_() I/O helpers of similar names. Do you still have that Sparc emulator? > > From a quick look there seems to be some challenges but the current > helpers that re-do part of io.h is not the way forward and hiding them > in arch/include/asm/fb.h seems counter productive. Which challenges did you see? Best regards Thomas > > Sam
On Thu, Apr 27, 2023, at 08:22, Thomas Zimmermann wrote: > Am 26.04.23 um 21:21 schrieb Sam Ravnborg: >> On Wed, Apr 26, 2023 at 03:04:15PM +0200, Thomas Zimmermann wrote: >>> Fbdev provides helpers for framebuffer I/O, such as fb_readl(), >>> fb_writel() or fb_memcpy_to_fb(). The implementation of each helper >>> depends on the architecture. It's still all located in fbdev's main >>> header file <linux/fb.h>. Move all of it into each archtecture's >>> <asm/fb.h>, with shared code in <asm-generic/fb.h>. >> >> For once I think this cleanup is moving things in the wrong direction. >> >> The fb_* helpers predates the generic io.h support and try to >> add a generic layer for read read / write operations. >> >> The right fix would be to migrate fb_* to use the io helpers >> we have today - so we use the existing way to handle the architecture >> specific details. > > I looked through the existing versions of the fb_() I/O helpers. They > can apparently be implemented with the regular helpers of similar names. > > I'm not sure, but even Sparc looks compatible. At least these sbus_ > functions seem to be equivalent to the __raw_() I/O helpers of similar > names. Do you still have that Sparc emulator? I looked at the current code and came to the same conclusion: all architectures we support today do the same thing in __raw_readl() and fb_readl() etc, so we can completely remove the latter without changing semantics. I think the original list was necessary since not all architectures supported the __raw_ accessors in the past, so they were open-coded here for the rest. I thought there were also architectures on which __raw_readl() does a byteswap to reverse the swap done in a PCI host bridge, but it apears that none of those remain now, if we ever had them. Arnd
Hi Thomas, On Thu, Apr 27, 2023 at 09:22:47AM +0200, Thomas Zimmermann wrote: > Hi Sam > > Am 26.04.23 um 21:21 schrieb Sam Ravnborg: > > Hi Thomas. > > > > On Wed, Apr 26, 2023 at 03:04:15PM +0200, Thomas Zimmermann wrote: > > > Fbdev provides helpers for framebuffer I/O, such as fb_readl(), > > > fb_writel() or fb_memcpy_to_fb(). The implementation of each helper > > > depends on the architecture. It's still all located in fbdev's main > > > header file <linux/fb.h>. Move all of it into each archtecture's > > > <asm/fb.h>, with shared code in <asm-generic/fb.h>. > > > > For once I think this cleanup is moving things in the wrong direction. > > > > The fb_* helpers predates the generic io.h support and try to > > add a generic layer for read read / write operations. > > > > The right fix would be to migrate fb_* to use the io helpers > > we have today - so we use the existing way to handle the architecture > > specific details. > > I looked through the existing versions of the fb_() I/O helpers. They can > apparently be implemented with the regular helpers of similar names. > > I'm not sure, but even Sparc looks compatible. At least these sbus_ > functions seem to be equivalent to the __raw_() I/O helpers of similar > names. > Do you still have that Sparc emulator? I used qemu the last time I played with sparc and saved the instructions somewhere how to redo it - but that would use to bohcs driver only I think. I have saprc machines, but none of these are easy to get operational. We can always ask on sparclinux to get some testing feedback. > > > > > From a quick look there seems to be some challenges but the current > > helpers that re-do part of io.h is not the way forward and hiding them > > in arch/include/asm/fb.h seems counter productive. > > Which challenges did you see? sparc was the main thing - but maybe I did not look close enough. And then I tried to map the macros to some of the more highlevel ones from io.h, but as Arnd says the __raw* is the way to go here. Sam