Skip to contentSkip to conversation

soc/tegra: fuse: Drop Kconfig dependency on TEGRA20_APB_DMA

PATCH In review Dmitry Osipenko 2 participants 6 messages Sep 23, 2020 – Nov 16, 2020 Tegra architecture support

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 soc/tegra: fuse: Drop Kconfig dependency on TEGRA20_APB_DMA

The DMA subsystem could be entirely disabled in Kconfig and then the TEGRA20_APB_DMA option isn't available too. Hence kernel configuration fails if DMADEVICES Kconfig option is disabled due to the unsatisfiable dependency.

The FUSE driver isn't a critical driver and currently it only provides NVMEM interface to userspace which isn't known to be widely used, and thus, it's fine if FUSE driver fails to load.

Let's remove the erroneous Kconfig dependency and let the FUSE driver to fail the probing if DMA is unavailable.

Fixes 19d41e5e9c68 ("soc/tegra: fuse: Add APB DMA dependency for Tegra20")
---
 drivers/soc/tegra/Kconfig | 1 -
 1 file changed, 1 deletion(-)
drivers/soc/tegra/Kconfig +0 -1
index 976dee036470..676807c5a215 100644
@@ -136,7 +136,6 @@ config SOC_TEGRA_FUSE
 	def_bool y
 	depends on ARCH_TEGRA
 	select SOC_BUS
-	select TEGRA20_APB_DMA if ARCH_TEGRA_2x_SOC
 
 config SOC_TEGRA_FLOWCTRL
 	bool
Signature · 1 line
2.27.0
#1 lore raw

23.09.2020 03:34, Dmitry Osipenko пишет:

Quoted 26 lines

The DMA subsystem could be entirely disabled in Kconfig and then the TEGRA20_APB_DMA option isn't available too. Hence kernel configuration fails if DMADEVICES Kconfig option is disabled due to the unsatisfiable dependency.

The FUSE driver isn't a critical driver and currently it only provides NVMEM interface to userspace which isn't known to be widely used, and thus, it's fine if FUSE driver fails to load.

Let's remove the erroneous Kconfig dependency and let the FUSE driver to fail the probing if DMA is unavailable.

Fixes 19d41e5e9c68 ("soc/tegra: fuse: Add APB DMA dependency for Tegra20")
---
 drivers/soc/tegra/Kconfig | 1 -
 1 file changed, 1 deletion(-)
drivers/soc/tegra/Kconfig +0 -0
index 976dee036470..676807c5a215 100644

drivers/soc/tegra/Kconfig config SOC_TEGRA_FUSE · line 141

Quoted 7 lines
drivers/soc/tegra/Kconfig +0 -1
@@ -136,7 +136,6 @@ config SOC_TEGRA_FUSE
 	def_bool y
 	depends on ARCH_TEGRA
 	select SOC_BUS
-	select TEGRA20_APB_DMA if ARCH_TEGRA_2x_SOC
 
 config SOC_TEGRA_FLOWCTRL
 	bool

Thierry, a reminder ping.

#2 lore raw
In reply to Dmitry Osipenko
Quoted 8 lines · Dmitry Osipenko

The DMA subsystem could be entirely disabled in Kconfig and then the TEGRA20_APB_DMA option isn't available too. Hence kernel configuration fails if DMADEVICES Kconfig option is disabled due to the unsatisfiable dependency.

The FUSE driver isn't a critical driver and currently it only provides NVMEM interface to userspace which isn't known to be widely used, and thus, it's fine if FUSE driver fails to load.

This isn't entirely accurate. The FUSE driver also provides SKU specific information via tegra_sku_info and exposes some of the FUSE registers to other drivers, such as the calibration data for XUSB.

The APB DMA is really only needed to work around an issue on Tegra20, so perhaps this really is okay. On the other hand, could we not just change the dependency to something like

select DMADEVICES if ARCH_TEGRA_2x_SOC select TEGRA20_APB_DMA if ARCH_TEGRA_2x_SOC

to fix the Kconfig issue but keep the explicit documentation of this dependency?

Thierry

Quoted 27 lines

Let's remove the erroneous Kconfig dependency and let the FUSE driver to fail the probing if DMA is unavailable.

Fixes 19d41e5e9c68 ("soc/tegra: fuse: Add APB DMA dependency for Tegra20")
---
 drivers/soc/tegra/Kconfig | 1 -
 1 file changed, 1 deletion(-)
drivers/soc/tegra/Kconfig +0 -1
index 976dee036470..676807c5a215 100644
@@ -136,7 +136,6 @@ config SOC_TEGRA_FUSE
 	def_bool y
 	depends on ARCH_TEGRA
 	select SOC_BUS
-	select TEGRA20_APB_DMA if ARCH_TEGRA_2x_SOC
 
 config SOC_TEGRA_FLOWCTRL
 	bool
Signature · 1 line
2.27.0
#3 lore raw

16.11.2020 16:20, Thierry Reding пишет:

Quoted 6 lines · Dmitry Osipenko
Quoted 5 lines

The DMA subsystem could be entirely disabled in Kconfig and then the TEGRA20_APB_DMA option isn't available too. Hence kernel configuration fails if DMADEVICES Kconfig option is disabled due to the unsatisfiable dependency.

Quoted 7 lines
Quoted 3 lines

The FUSE driver isn't a critical driver and currently it only provides NVMEM interface to userspace which isn't known to be widely used, and thus, it's fine if FUSE driver fails to load.

This isn't entirely accurate. The FUSE driver also provides SKU specific information via tegra_sku_info and exposes some of the FUSE registers to other drivers, such as the calibration data for XUSB.

The SKU data is read out only once during early boot and the DMA is not used for this. The FUSE platform driver exists separately from the early FUSE code.

Quoted 9 lines

The APB DMA is really only needed to work around an issue on Tegra20, so perhaps this really is okay. On the other hand, could we not just change the dependency to something like

select DMADEVICES if ARCH_TEGRA_2x_SOC select TEGRA20_APB_DMA if ARCH_TEGRA_2x_SOC

to fix the Kconfig issue but keep the explicit documentation of this dependency?

On Tegra20, the APB DMA is used only for by NVMEM which is exposed via sysfs. If DMA is disabled, then NVMEM isn't registered because tegra20_fuse_probe() returns -EPROBE_DEFER.

Hence there is no real need to enforce the extra dependencies. It's always better to allow minimizing the build dependencies whenever possible, IMO, and it's possible to do it in this case.

#4 lore raw
Quoted 11 lines · Dmitry Osipenko

16.11.2020 16:20, Thierry Reding пишет:

Quoted 10 lines · Dmitry Osipenko
Quoted 8 lines

The DMA subsystem could be entirely disabled in Kconfig and then the TEGRA20_APB_DMA option isn't available too. Hence kernel configuration fails if DMADEVICES Kconfig option is disabled due to the unsatisfiable dependency.

The FUSE driver isn't a critical driver and currently it only provides NVMEM interface to userspace which isn't known to be widely used, and thus, it's fine if FUSE driver fails to load.

Quoted 7 lines
Quoted 3 lines

This isn't entirely accurate. The FUSE driver also provides SKU specific information via tegra_sku_info and exposes some of the FUSE registers to other drivers, such as the calibration data for XUSB.

The SKU data is read out only once during early boot and the DMA is not used for this. The FUSE platform driver exists separately from the early FUSE code.

True, but the commit message isn't entirely accurate as-is, because on later Tegra SoCs the driver does a bit more than that. So if you don't mind I'll reword this to be a little more accurate if I end up deciding to apply it.

Quoted 7 lines
Quoted 7 lines

The APB DMA is really only needed to work around an issue on Tegra20, so perhaps this really is okay. On the other hand, could we not just change the dependency to something like

select DMADEVICES if ARCH_TEGRA_2x_SOC select TEGRA20_APB_DMA if ARCH_TEGRA_2x_SOC

Quoted 6 lines
Quoted 2 lines

to fix the Kconfig issue but keep the explicit documentation of this dependency?

On Tegra20, the APB DMA is used only for by NVMEM which is exposed via sysfs. If DMA is disabled, then NVMEM isn't registered because tegra20_fuse_probe() returns -EPROBE_DEFER.

Again, true. But -EPROBE_DEFER is a silent error, so if somebody indeed has DMADEVICES disabled and TEGRA20_APB_DMA is not available, then they will not get FUSE support and they are going to struggle to find out why that's not working.

Quoted 3 lines

Hence there is no real need to enforce the extra dependencies. It's always better to allow minimizing the build dependencies whenever possible, IMO, and it's possible to do it in this case.

I don't entirely agree with this. Dependencies (and especially selects) are used to pull in driver and/or features that are generally considered useful. In this particular case, TEGRA20_APB_DMA is needed for the FUSE driver to work correctly on Tegra20. Whether FUSE support on Tegra20 is useful or not isn't quite relevant at this point.

There's also a balance as to what makes sense and what doesn't. APB DMA is a useful feature in itself and disabling it is very much discouraged because there are plenty of other drivers that can make use of it. That is also the reason why we enable it in the default configuration. So I don't consider a select on a symbol that's enabled by default an extra dependency. Instead it's more of an explicit statement that you really do want this enabled if you want that driver to work.

And like I said, if we don't have this and the driver will fail to probe because of -EPROBE_DEFER, the user is going to have a very difficult time of finding out why exactly that's happening. We're not even emitting an error for this, so there's no way of knowing, even if they enable driver debugging, /why/ the FUSE driver isn't working.

Thierry

#5 lore raw

16.11.2020 19:49, Thierry Reding пишет:

Quoted 16 lines · Dmitry Osipenko
Quoted 15 lines

16.11.2020 16:20, Thierry Reding пишет:

Quoted 13 lines · Dmitry Osipenko
Quoted 8 lines

The DMA subsystem could be entirely disabled in Kconfig and then the TEGRA20_APB_DMA option isn't available too. Hence kernel configuration fails if DMADEVICES Kconfig option is disabled due to the unsatisfiable dependency.

The FUSE driver isn't a critical driver and currently it only provides NVMEM interface to userspace which isn't known to be widely used, and thus, it's fine if FUSE driver fails to load.

This isn't entirely accurate. The FUSE driver also provides SKU specific information via tegra_sku_info and exposes some of the FUSE registers to other drivers, such as the calibration data for XUSB.

Quoted 8 lines
Quoted 3 lines

The SKU data is read out only once during early boot and the DMA is not used for this. The FUSE platform driver exists separately from the early FUSE code.

True, but the commit message isn't entirely accurate as-is, because on later Tegra SoCs the driver does a bit more than that. So if you don't mind I'll reword this to be a little more accurate if I end up deciding to apply it.

Please feel free to edit it as yo wish.

Quoted 37 lines
Quoted 7 lines
Quoted 7 lines

The APB DMA is really only needed to work around an issue on Tegra20, so perhaps this really is okay. On the other hand, could we not just change the dependency to something like

select DMADEVICES if ARCH_TEGRA_2x_SOC select TEGRA20_APB_DMA if ARCH_TEGRA_2x_SOC

Quoted 6 lines
Quoted 2 lines

to fix the Kconfig issue but keep the explicit documentation of this dependency?

On Tegra20, the APB DMA is used only for by NVMEM which is exposed via sysfs. If DMA is disabled, then NVMEM isn't registered because tegra20_fuse_probe() returns -EPROBE_DEFER.

Again, true. But -EPROBE_DEFER is a silent error, so if somebody indeed has DMADEVICES disabled and TEGRA20_APB_DMA is not available, then they will not get FUSE support and they are going to struggle to find out why that's not working.

Quoted 3 lines

Hence there is no real need to enforce the extra dependencies. It's always better to allow minimizing the build dependencies whenever possible, IMO, and it's possible to do it in this case.

I don't entirely agree with this. Dependencies (and especially selects) are used to pull in driver and/or features that are generally considered useful. In this particular case, TEGRA20_APB_DMA is needed for the FUSE driver to work correctly on Tegra20. Whether FUSE support on Tegra20 is useful or not isn't quite relevant at this point.

There's also a balance as to what makes sense and what doesn't. APB DMA is a useful feature in itself and disabling it is very much discouraged because there are plenty of other drivers that can make use of it. That is also the reason why we enable it in the default configuration. So I don't consider a select on a symbol that's enabled by default an extra dependency. Instead it's more of an explicit statement that you really do want this enabled if you want that driver to work.

Quoted 5 lines

And like I said, if we don't have this and the driver will fail to probe because of -EPROBE_DEFER, the user is going to have a very difficult time of finding out why exactly that's happening. We're not even emitting an error for this, so there's no way of knowing, even if they enable driver debugging, /why/ the FUSE driver isn't working.

Recent kernels have /sys/kernel/debug/devices_deferred.