On 05.11.2021 16:38, Roger Pau Monné wrote: > On Fri, Nov 05, 2021 at 01:32:18PM +0100, Jan Beulich wrote: >> --- a/xen/arch/x86/setup.c >> +++ b/xen/arch/x86/setup.c >> @@ -1699,6 +1699,13 @@ void __init noreturn __start_xen(unsigne >> >> dmi_scan_machine(); >> >> + /* >> + * IOMMU-related ACPI table parsing has to happen before APIC probing, >> for >> + * check_x2apic_preenabled() to be able to observe respective findings, >> in >> + * particular iommu_intremap having got turned off. >> + */ >> + acpi_iommu_init(); > > If we pull this out I think we should add a check for acpi_disabled > and if set turn off iommu_intremap and iommu_enable?
Hmm, I should have added a note regarding this. If we want to exactly retain prior behavior, acpi_ht would also need checking. Yet that has gone wrong long ago: We parse way too many tables when acpi_disabled && acpi_ht, and hence while correct wrt to prior behavior I'd consider it wrong to (re)add a "!acpi_ht" check. As a result I'm of the opinion that checking acpi_disabled here also isn't necessarily appropriate, and instead IOMMU disabling would better be solely under the control of "iommu=". Additionally iirc Andrew has been suggesting to drop all this "ACPI disabled / HT-only" machinery (I'm somewhat hesitant with that, but as a result I'm also not very eager to actually correct to accumulated bad behavior). The change here simply would be a tiny first step in that direction. Jan
