Re: [PATCH] cdx: Fix missing GENERIC_MSI_IRQ on compile test

[Alex Williamson <alex.williamson@redhat.com>]

On Thu, 17 Jul 2025 13:33:58 -0600
Alex Williamson <alex.williamson@redhat.com> wrote:

> On Wed, 16 Jul 2025 08:49:04 +0200
> Krzysztof Kozlowski <krzysztof.kozlowski@linaro.org> wrote:
> 
> > CDX_BUS driver uses msi_setup_device_data() which is selected by
> > GENERIC_MSI_IRQ, thus compile testing without the latter failed:
> > 
> >   /usr/bin/ld: drivers/cdx/cdx.o: in function `cdx_probe':
> >   build/drivers/cdx/cdx.c:314: undefined reference to `msi_setup_device_data'
> > 
> > Reported-by: Randy Dunlap <rdunlap@infradead.org>
> > Closes: https://lore.kernel.org/all/b2c54a12-480c-448a-8b90-333cb03d9c14@infradead.org/
> > Fixes: 7f81907b7e3f ("cdx: Enable compile testing")
> > Signed-off-by: Krzysztof Kozlowski <krzysztof.kozlowski@linaro.org>
> > ---
> >  drivers/cdx/Kconfig | 1 +
> >  1 file changed, 1 insertion(+)
> > 
> > diff --git a/drivers/cdx/Kconfig b/drivers/cdx/Kconfig
> > index 1f1e360507d7..3af41f51cf38 100644
> > --- a/drivers/cdx/Kconfig
> > +++ b/drivers/cdx/Kconfig
> > @@ -8,6 +8,7 @@
> >  config CDX_BUS
> >  	bool "CDX Bus driver"
> >  	depends on OF && ARM64 || COMPILE_TEST
> > +	select GENERIC_MSI_IRQ
> >  	help
> >  	  Driver to enable Composable DMA Transfer(CDX) Bus. CDX bus
> >  	  exposes Fabric devices which uses composable DMA IP to the  
> 
> This looks incomplete to me.  The Makefile has:
> 
> ifdef CONFIG_GENERIC_MSI_IRQ
> obj-$(CONFIG_CDX_BUS) += cdx_msi.o
> endif
> 
> The call to msi_setup_device_data(), the conditional inclusion of
> cdx_msi.o, and cdx_msi.c itself were added in 0e439ba38e61 ("cdx: add
> MSI support for CDX bus").  That's the source of the issue, not the
> Kconfig update to build under COMPILE_TEST, but also the Makefile
> should be updated to statically include cdx_msi.o.  Thanks,

Also, we may not be philosophically aligned on this, but if CDX_BUS
selects GENERIC_MSI_IRQ and CDX_CONTROLLER depends on CDX_BUS, then the
select of the same under CDX_CONTROLLER becomes redundant.

Independent of that there are also some ifdefs in cdx_bus.h that would
become invalid if the entire CDX_BUS depends on GENERIC_MSI_IRQ.  It
really kinda looks like CDX_BUS expects a use case that doesn't depend
on GENERIC_MSI_IRQ and I wonder if the right answer isn't just:

diff --git a/drivers/cdx/cdx.c b/drivers/cdx/cdx.c
index 092306ca2541..3d50f8cd9c0b 100644
--- a/drivers/cdx/cdx.c
+++ b/drivers/cdx/cdx.c
@@ -310,7 +310,7 @@ static int cdx_probe(struct device *dev)
 	 * Setup MSI device data so that generic MSI alloc/free can
 	 * be used by the device driver.
 	 */
-	if (cdx->msi_domain) {
+	if (IS_ENABLED(CONFIG_GENERIC_MSI_IRQ) && cdx->msi_domain) {
 		error = msi_setup_device_data(&cdx_dev->dev);
 		if (error)
 			return error;
@@ -833,7 +833,7 @@ int cdx_device_add(struct cdx_dev_params *dev_params)
 		     ((cdx->id << CDX_CONTROLLER_ID_SHIFT) | (cdx_dev->bus_num & CDX_BUS_NUM_MASK)),
 		     cdx_dev->dev_num);
 
-	if (cdx->msi_domain) {
+	if (IS_ENABLED(CONFIG_GENERIC_MSI_IRQ) && cdx->msi_domain) {
 		cdx_dev->num_msi = dev_params->num_msi;
 		dev_set_msi_domain(&cdx_dev->dev, cdx->msi_domain);
 	}

Then I suppose vfio-cdx would also have the option to either
select/depend on GENERIC_MSI_IRQ or perhaps get away with stubbing
vfio_cdx_set_irqs_ioctl and vfio_cdx_irqs_cleanup and conditionally
building intr.o, depending on if there's a use case without MSI.
Thanks,

Alex