On Tue, Jun 18, 2024 at 10:56:53AM +0200, Mariusz Tkaczyk wrote: > On Sat, 15 Jun 2024 12:33:45 +0200 Lukas Wunner <lukas@xxxxxxxxx> wrote: >> > On Fri, Jun 14, 2024 at 04:06:14PM -0500, stuart hayes wrote: > > > Failing pci_npem_init() if this ops->get_active_indications() fails > > > will keep this from working on most (all?) Dell servers, because the > > > _DSM get/set functions use an IPMI operation region to get/set the > > > active LEDs, and this is getting run before the IPMI drivers and > > > acpi_ipmi module (which provides ACPI access to IPMI operation > > > regions) get loaded. (GET_SUPPORTED_STATES works without IPMI.) > > > > CONFIG_ACPI_IPMI is tristate. Even if it's built-in, the > > module_initcall() becomes a device_initcall(). > > > > PCI enumeration happens from a subsys_initcall(), way earlier > > than device_initcall(). > > > > If you set CONFIG_ACPI_IPMI=y and change the module_initcall() in > > drivers/acpi/acpi_ipmi.c to arch_initcall(), does the issue go away? > > That seems to be the best option. Please test Lukas proposal and let me know. > Shouldn't I make a dependency to ACPI_IPMI in Kconfig (with optional comment > about initcall)? > > +config PCI_NPEM > + bool "Native PCIe Enclosure Management" > + depends on LEDS_CLASS=y > + depends on ACPI_IPMI=y This would effectively disallow NPEM on non-ACPI systems. I think what you want instead is to allow either ACPI_IPMI=y or ACPI_IPMI=n, but not ACPI_IPMI=m, so: depends on ACPI_IPMI!=m Thanks, Lukas
