On 6/2/21 11:01 AM, Tianyu Lan wrote: > Hi Boris: > Thanks for your review. > > On 6/2/2021 9:16 AM, Boris Ostrovsky wrote: >> >> On 5/30/21 11:06 AM, Tianyu Lan wrote: >>> @@ -91,6 +92,6 @@ int pci_xen_swiotlb_init_late(void) >>> EXPORT_SYMBOL_GPL(pci_xen_swiotlb_init_late); >>> IOMMU_INIT_FINISH(2, >>> - NULL, >>> + hyperv_swiotlb_detect, >>> pci_xen_swiotlb_init, >>> NULL); >> >> >> Could you explain this change? > > Hyper-V allocates its own swiotlb bounce buffer and the default > swiotlb buffer should not be allocated. swiotlb_init() in pci_swiotlb_init() is to allocate default swiotlb buffer. > To achieve this, put hyperv_swiotlb_detect() as the first entry in the iommu_table_entry list. The detect loop in the pci_iommu_alloc() will exit once hyperv_swiotlb_detect() is called in Hyper-V VM and other iommu_table_entry callback will not be called. Right. But pci_xen_swiotlb_detect() will only do something for Xen PV guests, and those guests don't run on hyperV. It's either xen_pv_domain() (i.e. hypervisor_is_type(X86_HYPER_XEN_PV)) or hypervisor_is_type(X86_HYPER_MS_HYPERV) but never both. So I don't think there needs to be a dependency between the two callbacks. -boris
