On Fri, Mar 25, 2016 at 02:04:23AM +0000, Andre Przywara wrote:

This series is a joint effort to re-implement KVM's GIC emulation.

While the current implementation is centered around providing
efficient MMIO emulation, the hot path for most guests is actually
the guest entry and exit, which currently is rather costly.
Also the existing emulation has a global distributor lock, which
quickly becomes a bottleneck once the number of VCPUs increases.
Additionally the emulation was originally designed for GICv2, adding
GICv3 ITS emulation support to this proved to be rather painful.
Last, but not least the existing code became less and less
maintainable, with many special cases handled explicitly.

The new implementation is build around a struct vgic_irq data data
structure, which holds all information about a virtual interrupt.
Interruts which should be injected are hold in a per-VCPU list, this
make the entry/exit path much more efficient. Also the new structure
allows to have more fine grained locking - per IRQ and per VCPU -
getting rid of the global distributor lock.
As a result of the new design ITS emulation fits in more nicely, the
respective code will be provided as a follow-up series.

This series implements the same feature set as the existing emulation,
as a goodie we now implement priorities correctly.
To allow an easy transition with good test coverage, but still maintain
stability, both implementations live side by side, selectable via a
Kconfig option. The default is the new implementation.
If this code proves to be reliable, we will later remove the current
implementation with an extra patch set.

Please have a look at the series, review it and give the code some
serious testing (and possibly debugging). All feedback is appreciated.

Hmph, starting a guest a couple of times and running hackbench inside
the guest actually gave me (twice) the following error:

NMI watchdog: BUG: soft lockup - CPU#0 stuck for 23s! [kworker/0:0:4]

(that is using your branch on Mustang).

-Christoffer