On Mon, Jun 21, 2021 at 07:55:30AM -0700, Matthias Kaehlcke wrote:

On Sat, Jun 19, 2021 at 10:30:22AM +0900, Masahiro Yamada wrote:

quoted

On Sat, Jun 19, 2021 at 2:05 AM Matthias Kaehlcke [off-list ref] wrote:

quoted

Hi,

I'm adding a new driver and have an issue with Kconfig dependencies
that I coulnd't sort out so far.

Patch https://lore.kernel.org/patchwork/patch/1444212/ adds the new
onboard_usb_hub driver which exports two functions,
onboard_hub_create_pdevs() and onboard_hub_destroy_pdevs(). It also
provides stubs for these functions which are used when the driver
is not selected (CONFIG_USB_ONBOARD_HUB=n).

The new exported functions are called by the xhci-plat driver
(https://lore.kernel.org/patchwork/patch/1444215/). Since xhci-plat
now depends on symbols from the onboard_hub_driver the following
dependency was added to its Kconfig entry:

  config USB_XHCI_PLATFORM
    tristate "Generic xHCI driver for a platform device"
    select USB_XHCI_RCAR if ARCH_RENESAS
 +  depends on USB_ONBOARD_HUB || !USB_ONBOARD_HUB

This generally seems to work, however when USB_XHCI_PLATFORM is
forced to be builtin by another driver that depends on it (e.g.
USB_DWC3) it is still possible to build the onboard_hub driver
as a module, which results in unresolved symbols:

aarch64-linux-gnu-ld: drivers/usb/host/xhci-plat.o: in function
`xhci_plat_remove':
drivers/usb/host/xhci-plat.c:427: undefined reference to
`onboard_hub_destroy_pdevs'
drivers/usb/host/xhci-plat.c:427:(.text+0x82c): relocation truncated
to fit: R_AARCH64_CALL26 against undefined symbol
`onboard_hub_destroy_pdevs'
aarch64-linux-gnu-ld: drivers/usb/host/xhci-plat.o: in function
`xhci_plat_probe':
drivers/usb/host/xhci-plat.c:379: undefined reference to
`onboard_hub_create_pdevs'
drivers/usb/host/xhci-plat.c:379:(.text+0x131c): relocation truncated
to fit: R_AARCH64_CALL26 against undefined symbol
`onboard_hub_create_pdevs'

Kconfig generates the following warning with this configuration:

WARNING: unmet direct dependencies detected for USB_XHCI_PLATFORM
  Depends on [m]: USB_SUPPORT [=y] && USB [=y] && USB_XHCI_HCD [=y] && (USB_ONBOARD_HUB [=m] || !USB_ONBOARD_HUB [=m])
  Selected by [y]:
  - USB_DWC3 [=y] && USB_SUPPORT [=y] && (USB [=y] || USB_GADGET [=y]) && HAS_DMA [=y] && USB_XHCI_HCD [=y]
  Selected by [m]:
  - USB_CDNS_SUPPORT [=m] && USB_SUPPORT [=y] && (USB [=y] || USB_GADGET [=y]) && HAS_DMA [=y] && USB_XHCI_HCD [=y]
  - USB_BRCMSTB [=m] && USB_SUPPORT [=y] && USB [=y] && (ARCH_BRCMSTB [=y] && PHY_BRCM_USB [=m] || COMPILE_TEST [=y]) && USB_XHCI_HCD [=y]
  - USB_XHCI_MVEBU [=m] && USB_SUPPORT [=y] && USB [=y] && USB_XHCI_HCD [=y] && HAS_IOMEM [=y] && (ARCH_MVEBU [=y] || COMPILE_TEST [=y])

I read through kconfig-language.rst and experimented a fair bit,
but haven't found a working solution. Any advice would be
appreciated.

Thanks

Matthias

This issue should be discussed in the USB ML,

That's where it was initially brought up, but it didn't get the attention
of anyone in the position to give advice. Since the issue is more about
kbuild dependencies than USB specifically I brought it up here. The driver
already landed in the USB tree but was reverted due to this issue, I'm
stuck on this problem and really don't want the driver to die on the
finish line.

A workaround could be to make the driver 'bool' rather than 'tristate',
but I'm not sure if that would be acceptable.

quoted

but probably 'depends on USB_XHCI_PLATFORM' should be used everywhere instead of
'depends on USB_XHCI_PLATFORM'.

Did you mean 'select USB_XHCI_PLATFORM' rather than 'depends on
USB_XHCI_PLATFORM'? In general that sounds reasonable, since the drivers don't
actually depend on USB_XHCI_PLATFORM from a build perspective, and it's what
some drivers actually do. However it doesn't fix the problem, apparently a
'select X' from CONFIG_Y still results in CONFIG_X being 'y' if CONFIG_Y is
'y' (see the USB_DWC3 case above).

After some more experimentation it looks like the opposite fixes the conflict,
i.e. changing all instances of 'select USB_XHCI_PLATFORM' to 'depends on
USB_XHCI_PLATFORM'.

That's also in line with the recommendation to limit the use of 'select' to
certain use cases:

  select should be used with care. select will force a symbol to a value without
  visiting the dependencies. By abusing select you are able to select a symbol
  FOO even if FOO depends on BAR that is not set. In general use select only
  for non-visible symbols (no prompts anywhere) and for symbols with no
  dependencies. That will limit the usefulness but on the other hand avoid the
  illegal configurations all over.

  https://www.kernel.org/doc/html/latest/kbuild/kconfig-language.html