Message ID | 20231023160539.1537355-2-arnd@kernel.org (mailing list archive) |
---|---|
State | New, archived |
Headers | show |
Series | [1/2] media: pci: mgb4: add COMMON_CLK dependency | expand |
On Mon, Oct 23, 2023, at 18:05, Arnd Bergmann wrote: > From: Arnd Bergmann <arnd@arndb.de> > > As this is just a regular device driver, it has no business force-enabling > other drivers in the system, it should be entirely independent of the > implementation of the spi-nor layer or the specific DMA engine. > > The IIO symbols that are selected here are library modules that > are legitimately used. > > Fixes: 0ab13674a9bd ("media: pci: mgb4: Added Digiteq Automotive MGB4 driver") > Signed-off-by: Arnd Bergmann <arnd@arndb.de> > --- > drivers/media/pci/mgb4/Kconfig | 4 ---- > 1 file changed, 4 deletions(-) > > diff --git a/drivers/media/pci/mgb4/Kconfig b/drivers/media/pci/mgb4/Kconfig > index f2a05a1c8ffa..b90347c7f19b 100644 > --- a/drivers/media/pci/mgb4/Kconfig > +++ b/drivers/media/pci/mgb4/Kconfig > @@ -6,10 +6,6 @@ config VIDEO_MGB4 > select VIDEOBUF2_DMA_SG > select IIO_BUFFER > select IIO_TRIGGERED_BUFFER > - select I2C_XILINX > - select SPI_XILINX > - select MTD_SPI_NOR > - select XILINX_XDMA Apparently, the XDMA reference was in fact needed, as MGB4 calls some exported symbols from that particular dmaengine driver: aarch64-linux-ld: drivers/media/pci/mgb4/mgb4_core.o: in function `init_i2c': mgb4_core.c:(.text+0x3ec): undefined reference to `xdma_get_user_irq' aarch64-linux-ld: mgb4_core.c:(.text+0x404): undefined reference to `xdma_enable_user_irq' I couldn't easily figure out what a 'user_irq' is here, but I wonder if this is the expected way to use the DMA engine layer. Maybe this should have been a nested irqchip instead, or it should be encoded in the DMA request specifier? Arnd
On 10/24/23 06:27, Arnd Bergmann wrote: > On Mon, Oct 23, 2023, at 18:05, Arnd Bergmann wrote: >> From: Arnd Bergmann <arnd@arndb.de> >> >> As this is just a regular device driver, it has no business force-enabling >> other drivers in the system, it should be entirely independent of the >> implementation of the spi-nor layer or the specific DMA engine. >> >> The IIO symbols that are selected here are library modules that >> are legitimately used. >> >> Fixes: 0ab13674a9bd ("media: pci: mgb4: Added Digiteq Automotive MGB4 driver") >> Signed-off-by: Arnd Bergmann <arnd@arndb.de> >> --- >> drivers/media/pci/mgb4/Kconfig | 4 ---- >> 1 file changed, 4 deletions(-) >> >> diff --git a/drivers/media/pci/mgb4/Kconfig b/drivers/media/pci/mgb4/Kconfig >> index f2a05a1c8ffa..b90347c7f19b 100644 >> --- a/drivers/media/pci/mgb4/Kconfig >> +++ b/drivers/media/pci/mgb4/Kconfig >> @@ -6,10 +6,6 @@ config VIDEO_MGB4 >> select VIDEOBUF2_DMA_SG >> select IIO_BUFFER >> select IIO_TRIGGERED_BUFFER >> - select I2C_XILINX >> - select SPI_XILINX >> - select MTD_SPI_NOR >> - select XILINX_XDMA > Apparently, the XDMA reference was in fact needed, as MGB4 > calls some exported symbols from that particular dmaengine > driver: > > aarch64-linux-ld: drivers/media/pci/mgb4/mgb4_core.o: in function `init_i2c': mgb4_core.c:(.text+0x3ec): undefined reference to `xdma_get_user_irq' > aarch64-linux-ld: mgb4_core.c:(.text+0x404): undefined reference to `xdma_enable_user_irq' > > I couldn't easily figure out what a 'user_irq' is here, > but I wonder if this is the expected way to use the DMA engine > layer. Maybe this should have been a nested irqchip instead, > or it should be encoded in the DMA request specifier? Hi Arnd, Here is a brief description of 'user_irq' and 'xdma_enable_user_irq' The XDMA subsystem is used in conjunction with the PCIe IP block. Please see https://lwn.net/Articles/911496/ for the overall information. XDMA can forward PCIe msi-x interrupt to/from user logic hardware (e.g. Digiteq device) which is connected to its user irq pin. And XDMA has a register to enable/disabe interrupt forwarding for a specific user irq pin. 'xdma_enable_user_irq' and 'xdma_disable_user_irq' are provided for hardware driver which is designed to use XDMA to enable/disable its interrupt. And based on the previous discussion with Mark, Digiteq device does not use its own register to enable/disable interrupt but relies on XDMA. Please see https://lore.kernel.org/lkml/daccee4a-ac3c-bfc1-4876-24e6ecf5bcf1@gpxsee.org/ Thanks, Lizhi > > Arnd
On Fri, Oct 27, 2023, at 16:17, Martin Tůma wrote: > > On 23. 10. 23 18:05, Arnd Bergmann wrote: >> From: Arnd Bergmann <arnd@arndb.de> >> >> As this is just a regular device driver, it has no business force-enabling >> other drivers in the system, it should be entirely independent of the >> implementation of the spi-nor layer or the specific DMA engine. >> > > The drivers are required for IP cores that are used on the card (in the > FPGA). Without I2C_XILINX and XILINX_XDMA the card won't work at all. > Without SPI_XILINX the access to the card's FLASH (used e.g. for FW > changes) won't be possible. > > A change to "depend" instead of "select" is thus possible if it makes > more sense to you, but removing it would make the module not compile or > not work at runtime (there is no symbol dependency to I2C_XILINX and > SPI_XILINX, but both need to be present and are loaded using > request_module() at runtime). Sorry for the delay at getting back to you here. I don't think there is a good answer here, though I normally try to only list the minimal dependencies that are required at build time. E.g. for on-chip devices we don't require the use of a particular clock/irq/pin/gpio/... controller even if we know exactly which of those are used on a given chip. Since this is a PCI device, it's a bit different, so maybe something like this would work to correctly document which dependencies are required at build time vs run time: --- a/drivers/media/pci/mgb4/Kconfig +++ b/drivers/media/pci/mgb4/Kconfig @@ -1,15 +1,13 @@ # SPDX-License-Identifier: GPL-2.0-only config VIDEO_MGB4 tristate "Digiteq Automotive MGB4 support" - depends on VIDEO_DEV && PCI && I2C && DMADEVICES && SPI && MTD && IIO + depends on VIDEO_DEV && PCI && I2C && SPI && MTD && IIO depends on COMMON_CLK + depends on XILINX_XDMA + depends on (I2C_XILINX && SPI_XILINX) || COMPILE_TEST select VIDEOBUF2_DMA_SG select IIO_BUFFER select IIO_TRIGGERED_BUFFER - select I2C_XILINX - select SPI_XILINX - select MTD_SPI_NOR - select XILINX_XDMA help This is a video4linux driver for Digiteq Automotive MGB4 grabber cards. Arnd
On 08. 11. 23 17:13, Arnd Bergmann wrote: > On Fri, Oct 27, 2023, at 16:17, Martin Tůma wrote: >> >> On 23. 10. 23 18:05, Arnd Bergmann wrote: >>> From: Arnd Bergmann <arnd@arndb.de> >>> >>> As this is just a regular device driver, it has no business force-enabling >>> other drivers in the system, it should be entirely independent of the >>> implementation of the spi-nor layer or the specific DMA engine. >>> >> >> The drivers are required for IP cores that are used on the card (in the >> FPGA). Without I2C_XILINX and XILINX_XDMA the card won't work at all. >> Without SPI_XILINX the access to the card's FLASH (used e.g. for FW >> changes) won't be possible. >> >> A change to "depend" instead of "select" is thus possible if it makes >> more sense to you, but removing it would make the module not compile or >> not work at runtime (there is no symbol dependency to I2C_XILINX and >> SPI_XILINX, but both need to be present and are loaded using >> request_module() at runtime). > > Sorry for the delay at getting back to you here. > > I don't think there is a good answer here, though I normally > try to only list the minimal dependencies that are required > at build time. E.g. for on-chip devices we don't require the > use of a particular clock/irq/pin/gpio/... controller even if > we know exactly which of those are used on a given chip. > On SoCs you probably get a kernel configuration that is missing some feature but still boots up when you do not select/depend on the exact controller, but in the case of the mgb4 PCIe card you get a driver that does not work at all (The SPI_XILINX dependency could theoretically be made configurable, but you would lose the ability to flash the correct FW for the current HW module and the access to the card's serial number. I2C and XDMA are crucial.). > Since this is a PCI device, it's a bit different, so maybe > something like this would work to correctly document which > dependencies are required at build time vs run time: > > --- a/drivers/media/pci/mgb4/Kconfig > +++ b/drivers/media/pci/mgb4/Kconfig > @@ -1,15 +1,13 @@ > # SPDX-License-Identifier: GPL-2.0-only > config VIDEO_MGB4 > tristate "Digiteq Automotive MGB4 support" > - depends on VIDEO_DEV && PCI && I2C && DMADEVICES && SPI && MTD && IIO > + depends on VIDEO_DEV && PCI && I2C && SPI && MTD && IIO > depends on COMMON_CLK > + depends on XILINX_XDMA > + depends on (I2C_XILINX && SPI_XILINX) || COMPILE_TEST > select VIDEOBUF2_DMA_SG > select IIO_BUFFER > select IIO_TRIGGERED_BUFFER > - select I2C_XILINX > - select SPI_XILINX > - select MTD_SPI_NOR > - select XILINX_XDMA > help > This is a video4linux driver for Digiteq Automotive MGB4 grabber > cards. > My motivation when using "select" was to help people using "make menuconfig" to get the module selected/configured as they will usually not know that there are some Xilinx IP cores used that need separate drivers and the menuconfig GUI simply hides the mgb4 option making it almost impossible just from the menus to find out what has to be selected. But when there are reasons, why to chose "depends on" (like various configurations, tests or the "readability" of the dependencies) than I'm ok with your patch proposal. M. > Arnd
On Wed, Nov 8, 2023, at 19:33, Martin Tůma wrote: > On 08. 11. 23 17:13, Arnd Bergmann wrote: >> On Fri, Oct 27, 2023, at 16:17, Martin Tůma wrote: >>> On 23. 10. 23 18:05, Arnd Bergmann wrote: >>>> From: Arnd Bergmann <arnd@arndb.de> > > On SoCs you probably get a kernel configuration that is missing some > feature but still boots up when you do not select/depend on the exact > controller, but in the case of the mgb4 PCIe card you get a driver that > does not work at all (The SPI_XILINX dependency could theoretically be > made configurable, but you would lose the ability to flash the correct > FW for the current HW module and the access to the card's serial number. > I2C and XDMA are crucial.). My point was that we do this all the time for things that are essential: if your clock controller or the irqchip have no driver, then the camera device won't work, but neither would anything else. So in a SoC setting, you really just need to enable all the drivers for devices on that chip through the .config. >> Since this is a PCI device, it's a bit different, so maybe >> something like this would work to correctly document which >> dependencies are required at build time vs run time: >> >> --- a/drivers/media/pci/mgb4/Kconfig >> +++ b/drivers/media/pci/mgb4/Kconfig >> @@ -1,15 +1,13 @@ >> # SPDX-License-Identifier: GPL-2.0-only >> config VIDEO_MGB4 >> tristate "Digiteq Automotive MGB4 support" >> - depends on VIDEO_DEV && PCI && I2C && DMADEVICES && SPI && MTD && IIO >> + depends on VIDEO_DEV && PCI && I2C && SPI && MTD && IIO >> depends on COMMON_CLK >> + depends on XILINX_XDMA >> + depends on (I2C_XILINX && SPI_XILINX) || COMPILE_TEST >> select VIDEOBUF2_DMA_SG >> select IIO_BUFFER >> select IIO_TRIGGERED_BUFFER >> - select I2C_XILINX >> - select SPI_XILINX >> - select MTD_SPI_NOR >> - select XILINX_XDMA >> help >> This is a video4linux driver for Digiteq Automotive MGB4 grabber >> cards. >> > > My motivation when using "select" was to help people using "make > menuconfig" to get the module selected/configured as they will usually > not know that there are some Xilinx IP cores used that need separate > drivers and the menuconfig GUI simply hides the mgb4 option making it > almost impossible just from the menus to find out what has to be selected. > > But when there are reasons, why to chose "depends on" (like various > configurations, tests or the "readability" of the dependencies) than I'm > ok with your patch proposal. The main reason to use 'depends on' over 'select' here is that mixing the two is a common source of dependency loops that end up breaking the build. As a rule of thumb, I would use 'select' only for symbols that others already select, or that are hidden from visibility. Arnd
On 10/11/2023 15:37, Arnd Bergmann wrote: > On Wed, Nov 8, 2023, at 19:33, Martin Tůma wrote: >> On 08. 11. 23 17:13, Arnd Bergmann wrote: >>> On Fri, Oct 27, 2023, at 16:17, Martin Tůma wrote: >>>> On 23. 10. 23 18:05, Arnd Bergmann wrote: >>>>> From: Arnd Bergmann <arnd@arndb.de> >> >> On SoCs you probably get a kernel configuration that is missing some >> feature but still boots up when you do not select/depend on the exact >> controller, but in the case of the mgb4 PCIe card you get a driver that >> does not work at all (The SPI_XILINX dependency could theoretically be >> made configurable, but you would lose the ability to flash the correct >> FW for the current HW module and the access to the card's serial number. >> I2C and XDMA are crucial.). > > My point was that we do this all the time for things that are > essential: if your clock controller or the irqchip have > no driver, then the camera device won't work, but neither > would anything else. > > So in a SoC setting, you really just need to enable all > the drivers for devices on that chip through the .config. > >>> Since this is a PCI device, it's a bit different, so maybe >>> something like this would work to correctly document which >>> dependencies are required at build time vs run time: >>> >>> --- a/drivers/media/pci/mgb4/Kconfig >>> +++ b/drivers/media/pci/mgb4/Kconfig >>> @@ -1,15 +1,13 @@ >>> # SPDX-License-Identifier: GPL-2.0-only >>> config VIDEO_MGB4 >>> tristate "Digiteq Automotive MGB4 support" >>> - depends on VIDEO_DEV && PCI && I2C && DMADEVICES && SPI && MTD && IIO >>> + depends on VIDEO_DEV && PCI && I2C && SPI && MTD && IIO >>> depends on COMMON_CLK >>> + depends on XILINX_XDMA >>> + depends on (I2C_XILINX && SPI_XILINX) || COMPILE_TEST >>> select VIDEOBUF2_DMA_SG >>> select IIO_BUFFER >>> select IIO_TRIGGERED_BUFFER >>> - select I2C_XILINX >>> - select SPI_XILINX >>> - select MTD_SPI_NOR >>> - select XILINX_XDMA >>> help >>> This is a video4linux driver for Digiteq Automotive MGB4 grabber >>> cards. >>> >> >> My motivation when using "select" was to help people using "make >> menuconfig" to get the module selected/configured as they will usually >> not know that there are some Xilinx IP cores used that need separate >> drivers and the menuconfig GUI simply hides the mgb4 option making it >> almost impossible just from the menus to find out what has to be selected. >> >> But when there are reasons, why to chose "depends on" (like various >> configurations, tests or the "readability" of the dependencies) than I'm >> ok with your patch proposal. > > The main reason to use 'depends on' over 'select' here is that > mixing the two is a common source of dependency loops that end > up breaking the build. As a rule of thumb, I would use 'select' > only for symbols that others already select, or that are hidden > from visibility. > > Arnd > Arnd, can you post a v2? I think this should go to v6.7 as a fix. Regards, Hans
Arnd, Can you post a v2 for this? Note that patch 1/1 is in v6.7 already. Regards, Hans On 13/11/2023 15:57, Hans Verkuil wrote: > On 10/11/2023 15:37, Arnd Bergmann wrote: >> On Wed, Nov 8, 2023, at 19:33, Martin Tůma wrote: >>> On 08. 11. 23 17:13, Arnd Bergmann wrote: >>>> On Fri, Oct 27, 2023, at 16:17, Martin Tůma wrote: >>>>> On 23. 10. 23 18:05, Arnd Bergmann wrote: >>>>>> From: Arnd Bergmann <arnd@arndb.de> >>> >>> On SoCs you probably get a kernel configuration that is missing some >>> feature but still boots up when you do not select/depend on the exact >>> controller, but in the case of the mgb4 PCIe card you get a driver that >>> does not work at all (The SPI_XILINX dependency could theoretically be >>> made configurable, but you would lose the ability to flash the correct >>> FW for the current HW module and the access to the card's serial number. >>> I2C and XDMA are crucial.). >> >> My point was that we do this all the time for things that are >> essential: if your clock controller or the irqchip have >> no driver, then the camera device won't work, but neither >> would anything else. >> >> So in a SoC setting, you really just need to enable all >> the drivers for devices on that chip through the .config. >> >>>> Since this is a PCI device, it's a bit different, so maybe >>>> something like this would work to correctly document which >>>> dependencies are required at build time vs run time: >>>> >>>> --- a/drivers/media/pci/mgb4/Kconfig >>>> +++ b/drivers/media/pci/mgb4/Kconfig >>>> @@ -1,15 +1,13 @@ >>>> # SPDX-License-Identifier: GPL-2.0-only >>>> config VIDEO_MGB4 >>>> tristate "Digiteq Automotive MGB4 support" >>>> - depends on VIDEO_DEV && PCI && I2C && DMADEVICES && SPI && MTD && IIO >>>> + depends on VIDEO_DEV && PCI && I2C && SPI && MTD && IIO >>>> depends on COMMON_CLK >>>> + depends on XILINX_XDMA >>>> + depends on (I2C_XILINX && SPI_XILINX) || COMPILE_TEST >>>> select VIDEOBUF2_DMA_SG >>>> select IIO_BUFFER >>>> select IIO_TRIGGERED_BUFFER >>>> - select I2C_XILINX >>>> - select SPI_XILINX >>>> - select MTD_SPI_NOR >>>> - select XILINX_XDMA >>>> help >>>> This is a video4linux driver for Digiteq Automotive MGB4 grabber >>>> cards. >>>> >>> >>> My motivation when using "select" was to help people using "make >>> menuconfig" to get the module selected/configured as they will usually >>> not know that there are some Xilinx IP cores used that need separate >>> drivers and the menuconfig GUI simply hides the mgb4 option making it >>> almost impossible just from the menus to find out what has to be selected. >>> >>> But when there are reasons, why to chose "depends on" (like various >>> configurations, tests or the "readability" of the dependencies) than I'm >>> ok with your patch proposal. >> >> The main reason to use 'depends on' over 'select' here is that >> mixing the two is a common source of dependency loops that end >> up breaking the build. As a rule of thumb, I would use 'select' >> only for symbols that others already select, or that are hidden >> from visibility. >> >> Arnd >> > > Arnd, can you post a v2? I think this should go to v6.7 as a fix. > > Regards, > > Hans >
diff --git a/drivers/media/pci/mgb4/Kconfig b/drivers/media/pci/mgb4/Kconfig index f2a05a1c8ffa..b90347c7f19b 100644 --- a/drivers/media/pci/mgb4/Kconfig +++ b/drivers/media/pci/mgb4/Kconfig @@ -6,10 +6,6 @@ config VIDEO_MGB4 select VIDEOBUF2_DMA_SG select IIO_BUFFER select IIO_TRIGGERED_BUFFER - select I2C_XILINX - select SPI_XILINX - select MTD_SPI_NOR - select XILINX_XDMA help This is a video4linux driver for Digiteq Automotive MGB4 grabber cards.