On Fri, 2022-02-18 at 15:57 +0800, David Gow wrote:
Note that, while this does build again, it still segfaults on startup, so more work remains to be done.
That's probably just a lot more stuff getting included somehow?
They are:
- CONFIG_VFIO_PCI: Needs ioport_map/ioport_unmap.
- CONFIG_INFINIBAND_RDMAVT: Needs cpuinfo_x86 and __copy_user_nocache
- CONFIG_BNXT: Failing under UML with -Werror
ERROR:root:../drivers/net/ethernet/broadcom/bnxt/bnxt_ptp.c: In function ‘bnxt_ptp_enable’: ../drivers/net/ethernet/broadcom/bnxt/bnxt_ptp.c:400:43: error: array subscript 255 is above array bounds of ‘struct pps_pin[4]’ [-Werror=array-bounds] 400 | ptp->pps_info.pins[pin_id].event = BNXT_PPS_EVENT_EXTERNAL; | ~~~~~~~~~~~~~~~~~~^~~~~~~~
- CONFIG_PATA_CS5535: Needs MSR access (__tracepoint_{read,write}_msr)
- CONFIG_VDPA: Enables CONFIG_DMA_OPS, which is unimplemented. ('dma_ops' is not defined)
These are all issues which should be investigated properly and the corresponding options either fixed or disabled under UML. Having this list of broken options should act as a good to-do list here, and will allow these issues to be worked on independently, and other tests to work in the meantime.
I'm not really sure it makes sense to even do anything other than disabling these.
It looks like all of them are just exposed by now being able to build PCI drivers on UML. Surely the people writing the driver didn't expect their drivers to run over simulated PCI (which is what the UML PCI support is all about).
Now from a PCI driver point of view you can't really tell the difference (and anyway the driver won't be probed), but the issues (at least the build time ones) come from having
UML && PCI && X86_64
or
UML && PCI && X86_32
because drivers typically depend on X86_64 or X86_32, rather than on "X86 && X86_64" or "X86 && X86_32". In a sense thus, the issue is those drivers don't know that "!X86 && (X86_32 || X86_64)" can happen (with UML).
Now you could say that's the driver bug, or you could say that they should just add "depends on !UML" (though that's basically equivalent to adding "depends on X86" and the latter may be preferable in some cases).
Or actually in the three patches you have (1-3) it's in the code, but same thing, you can either add && !UML (like you did) or add && X86.
Arguably, however, building PCI drivers by default is somewhat questionable in the first place?
So maybe you should just add
# CONFIG_UML_PCI_OVER_VIRTIO is not set
to the broken_on_uml.config since it exposes all these issues, and really is not very useful since you're not going to actually run with any simulated PCI devices anyway, so drivers will not be probed.
johannes