This continues to build on the basic Xen on KVM platform support from
https://lore.kernel.org/qemu-devel/20230110122042.1562155-1-dwmw2@infradead.org/
We're working on hooking up the PV backend devices, and the biggest
remaining noticeably missing part was PIRQ support. This allows a Xen
guest to route GSI and MSI interrupts to event channels instead of being
delivered via the emulated I/OAPIC or local APIC respectively.
It starts relatively simple, with the basic hypercalls and infrastructure
for tracking/migrating the PIRQ table (and as I type this I've just
remembered I forgot to write the post_load function to reconstitute the
data structures which explicitly *state* that they need to be rebuilt).
I'm particularly interested in opinions on the hook in gsi_handler()
which lets the Xen emulation 'eat' the event instead of passing it to
the I/OAPIC.
I did ponder replacing the qemu_irq in gsi_state->ioapic_irq[n] when
GSI#n is redirected to a PIRQ, but I figured that was worse.
I definitely need to rethink the locking a little bit to avoid the
potential for deadlock when gsi_handler calls back into the evtchn code
to translate the event channel GSI. It's non-trivial to drop the lock
before sending the IRQ; maybe just a different lock with a smaller
scope. A previous implementation of event channels was a bit more
lockless, with atomic updates of the port table (the port_info fits in a
uint64_t). But now we have all the interesting fast paths accelerated in
the kernel that didn't seem worth it, so I went with simple locking...
too simple, it seems.
There's a similar recursive locking issue when pirq_bind_port() wants to
call kvm_update_msi_routes_all(), but is already holding the lock that
we'd take again when called to redo a translation. (And I still don't
much like the way that kvm_update_msi_routes_all() has to have a list of
PCI devices and actually recalculates the routes at all, instead of just
detaching the IRQFD and letting them be recalculated on demand. But I
was trying to avoid actually fixing that this week).
David Woodhouse (5):
i386/xen: Implement HYPERVISOR_physdev_op
hw/xen: Implement emulated PIRQ hypercall support
hw/xen: Support GSI mapping to PIRQ
hw/xen: [FIXME] Avoid deadlock in xen_evtchn_set_gsi()
hw/xen: Support MSI mapping to PIRQ
hw/i386/kvm/trace-events | 4 ++
hw/i386/kvm/trace.h | 1 +
hw/i386/kvm/xen-stubs.c | 11 ++++
hw/i386/kvm/xen_evtchn.c | 461 +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++-
hw/i386/kvm/xen_evtchn.h | 22 +++++++
hw/i386/x86.c | 15 +++++
hw/pci/msi.c | 13 ++++
hw/pci/msix.c | 7 ++-
hw/pci/pci.c | 14 +++++
meson.build | 1 +
target/i386/kvm/kvm.c | 12 +++-
target/i386/kvm/kvm_i386.h | 2 +
target/i386/kvm/xen-compat.h | 19 ++++++
target/i386/kvm/xen-emu.c | 136 +++++++++++++++++++++++++++++++++++++++++-
14 files changed, 712 insertions(+), 6 deletions(-)