[Xen-devel] use of struct hvm_mirq_dpci_mapping.gmsi vs. HVM_IRQ_DPCI_*_MSI flags

["Jan Beulich" <JBeulich@xxxxxxxxxx>]

pt_irq_create_bind_vtd() initializes this substructure only when setting
.flags to HVM_IRQ_DPCI_MACH_MSI|HVM_IRQ_DPCI_GUEST_MSI (the
PT_IRQ_TYPE_MSI case), while the other path will not set
HVM_IRQ_DPCI_GUEST_MSI but may also set HVM_IRQ_DPCI_MACH_MSI.
Yet hvm_dpci_msi_eoi() and hvm_migrate_pirqs() check for
HVM_IRQ_DPCI_MACH_MSI, i.e. may run into an uninitialized
.gmsi.* field. What am I missing here?

I'm largely asking because I think struct hvm_mirq_dpci_mapping.dom
and .digl_list could actually overlay .gmsi, as much as struct
hvm_irq_dpci.hvm_timer could actually rather be folded into struct
hvm_mirq_dpci_mapping (and then also overlay .gmsi). The overlay
distinction bit would, based on initialization, be HVM_IRQ_DPCI_GUEST_MSI,
but according to use it wouldn't be clear which of the two
HVM_IRQ_DPCI_*_MSI bits is actually the correct one.

Having a single structure only would make it a lot easier to
convert struct hvm_mirq_dpci_mapping * in struct hvm_irq_dpci to
a sparse struct hvm_mirq_dpci_mapping ** (populating slots only
as they get used), thus shrinking the currently two d->nr_pirqs
sized array allocations in pt_irq_create_bind_vtd() to a single one
with only pointer size array elements (allowing up to about 512
domain pirqs rather than currently slightly above 80 without
exceeding PAGE_SIZE on allocation).

Also I'm wondering why the PT_IRQ_TYPE_MSI path of
pt_irq_create_bind_vtd() checks that on re-use of an IRQ the
flags are indicating the same kind of interrupt, while the other
path doesn't bother doing so.

Thanks, Jan


_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxxxxxxxx
http://lists.xensource.com/xen-devel