>>>> Hi,
>>>>
>>>> I work on a driver (drivers/net/ethernet/mellanox/mlxsw) that is made up
>>>> of three modules: mlxsw_pci, mlxsw_core and a third module that is
>>>> loaded by mlxsw_core according to the probed PCI device ID via
>>>> request_module(). However, this function fails with rawhide kernels
>>>> during boot.
>>>>
>>>> While debugging this, I found out that if I build the kernel myself with
>>>> the exact same config, but set CONFIG_MODULE_SIG=n, then everything is
>>>> fine. In addition, when modprobing mlxsw_pci myself all the modules are
>>>> successfully loaded.
>>>>
>>>> Any ideas how this can be solved? Did anyone else bump into this
>>>> problem?
>>>>
>>>> Thanks.
>>>>
>>>
>>>Can you share kernel logs showing the error?
>>
>>The only error in the log is the one from the driver:
>>
>>mlxsw_pci 0000:03:00.0: cannot register bus device
>>mlxsw_pci: probe of 0000:03:00.0 failed with error -22
>
> Laura, note that this is because module mlxsw_spectrum failed to load in
> this function:
>
> static struct mlxsw_driver *mlxsw_core_driver_get(const char *kind)
> {
> struct mlxsw_driver *mlxsw_driver;
>
> spin_lock(&mlxsw_core_driver_list_lock);
> mlxsw_driver = __driver_find(kind);
> if (!mlxsw_driver) {
> spin_unlock(&mlxsw_core_driver_list_lock);
> request_module(MLXSW_MODULE_ALIAS_PREFIX "%s", kind);
> spin_lock(&mlxsw_core_driver_list_lock);
> mlxsw_driver = __driver_find(kind);
> }
> if (mlxsw_driver) {
> if (!try_module_get(mlxsw_driver->owner))
> mlxsw_driver = NULL;
> }
>
> spin_unlock(&mlxsw_core_driver_list_lock);
> return mlxsw_driver;
> }
>
> Here, "request_module" won't load the mlxsw_spectrum module. But this
issue
> happens only during the boot time.
>
> If you try to "modprobe mlxsw_pci" by hand later on, all works fine.
Is mlxsw_pci actually in the initramfs, and is it signed in it?
I was about to ask the same. for the generic initrd dracut matches
modules against a particular setup call/funciton [1] for a class of
devices to work out what to pull in and often these sort of modules
don't match as the sub modules. I have a similar issue with the
dw_mmc-* where the core dw_mmc.ko gets pulled in but not the HW
specific variants. I'm not worked out a good clean way to deal with
that in dracut as yet.
Peter
[1] storage uses these
'ahci_platform_get_resources|ata_scsi_ioctl|scsi_add_host|blk_cleanup_queue|register_mtd_blktrans|scsi_esp_register|register_virtio_device|usb_stor_disconnect|mmc_add_host|sdhci_add_host|scsi_add_host_with_dma'