Andy Shevchenko <andy.shevchenko@xxxxxxxxx> writes: > On Wed, Jul 22, 2020 at 12:47 PM Vitaly Kuznetsov <vkuznets@xxxxxxxxxx> wrote: >> Julia Suvorova <jusual@xxxxxxxxxx> writes: > >> > Scanning for PCI devices at boot takes a long time for KVM guests. It >> > can be reduced if KVM will handle all configuration space accesses for >> > non-existent devices without going to userspace [1]. But for this to >> > work, all accesses must go through MMCONFIG. >> > This change allows to use pci_mmcfg as raw_pci_ops for 64-bit KVM >> > guests making MMCONFIG the default access method. > > I'm not sure it won't break anything. It likely will as it's really hard to check all possible KVM configurations in existence and that's why we are converging on adding a feature bit which KVM userspace (e.g. QEMU) will set when the configuration is known to be good. > >> > [1] https://lkml.org/lkml/2020/5/14/936 > > use Link: tag and better to use lore.kernel.org. > >> This implies mmconfig access method is always functional (when present) >> for all KVM guests, regardless of hypervisor version/which KVM userspace >> is is use/... In case the assumption is true the patch looks good (to >> me) but in case it isn't or if we think that more control over this >> is needed we may want to introduce a PV feature bit for KVM. >> >> Also, I'm thinking about moving this to arch/x86/kernel/kvm.c: we can >> override x86_init.pci.arch_init and reassign raw_pci_ops after doing >> pci_arch_init(). > > % git grep -n -w x86_init.pci.arch_init -- arch/x86/ > arch/x86/hyperv/hv_init.c:400: x86_init.pci.arch_init = hv_pci_init; > arch/x86/kernel/apic/apic_numachip.c:203: x86_init.pci.arch_init > = pci_numachip_init; > arch/x86/kernel/jailhouse.c:207: x86_init.pci.arch_init > = jailhouse_pci_arch_init; > arch/x86/pci/init.c:20: if (x86_init.pci.arch_init && !x86_init.pci.arch_init()) > arch/x86/platform/intel-mid/intel-mid.c:172: x86_init.pci.arch_init > = intel_mid_pci_init; > arch/x86/platform/olpc/olpc.c:309: x86_init.pci.arch_init > = pci_olpc_init; > arch/x86/xen/enlighten_pv.c:1411: > x86_init.pci.arch_init = pci_xen_init; > > Are you going to update all these? Or how this is supposed to work (I > may be missing something)? My suggestion was to do exactly the same for KVM guests instead of switching ops in pci_mmcfg_arch_init() depending on kvm_para_available() output. Basically, keep all KVM-related tunings in one place (arch/x86/kernel/kvm.c). -- Vitaly
