The system GSIs are not designed for sharing. One device might assert a
shared interrupt with qemu_set_irq() and another might deassert it, and
the level from the first device is lost.
This could be solved by refactoring the x86 GSI code to use an OrIrq
device, but that still wouldn't be ideal.
The best answer would be to have a 'resample' callback which is invoked
when the interrupt is acked at the interrupt controller, and causes the
devices to re-trigger the interrupt if it should still be pending. This
is the model that VFIO in Linux uses, with a 'resampler' eventfd that
actually unmasks the interrupt on the hardware device and thus triggers
a new interrupt from it if needed.
As things stand, QEMU currently doesn't use that VFIO interface
correctly, and just bashes on the resampler for every MMIO access to the
device "just in case". Which requires unmapping and trapping the MMIO
while an interrupt is pending!
For the Xen callback GSI, QEMU does something similar — a flag is set
which triggers a poll on *every* vmexst to see if the GSI should be
deasserted.
Proper resampler support would solve all of that, but is a task for
later which has already been on the TODO list for a while.
Since the Xen event channel GSI support *already* has hooks into the PC
gsi_handler() code for routing GSIs to PIRQs, we can use that for a
simpler bug fix.
So... remember the externally-driven state of the line (from e.g. PCI
INTx) and set the logical OR of that with the GSI. As a bonus, we now
only need to enable the polling of vcpu_info on vmexit if the Xen
callback GSI is the *only* reason the corresponding line is asserted.
Closes: https://gitlab.com/qemu-project/qemu/-/issues/2731
Fixes: ddf0fd9ae1 ("hw/xen: Support HVM_PARAM_CALLBACK_TYPE_GSI callback")
Reported-by: Thomas Huth <thuth@redhat.com>
Signed-off-by: David Woodhouse <dwmw@amazon.co.uk>
Acked-by: Michael S. Tsirkin <mst@redhat.com>
981780cdda