On 20 July 2018 at 16:39, Andre Przywara <andre.przywara@xxxxxxx> wrote: > I found this in one my branches: this is an updated version of what I sent > end of 2016 [1]. I tried to address all comments that Drew and Eric had at > the time. > Please have a look whether this makes sense. > > Changelog v1..v2: > - made many functions void > - use symbolic name for first SPI being number 32 > - add test runs with one and three vCPUs > - use gic_version() directly > - factor out test_byte_access() > - drop redundant "filling priorities" test > - dropped GICv3 test > > [1] https://lists.cs.columbia.edu/pipermail/kvmarm/2016-November/022352.html > > Original cover letter: > ====================== > The GIC spec mandates certain constraints on how to acccess the MMIO > mapped registers, both in terms of which registers are available and also > in terms of which bits within a register should be masked, for instance. > Since we went through some lengths in the KVM emulation to implement this, > it's about time to give this actually a test beyond what the kernel as a > GIC user actually implements - for instance we ignore priorities in Linux. > > This series tries to attack some constraints, on a low-hanging-fruit base. > It focusses on some generic registers and the PRIORITY and TARGETS registers > of GICv2. GICv3 is not covered yet. > > This actually revealed genuine bugs in the KVM emulation in the past. KVM > passes these tests now, but QEMU fails some UP and 3-way-SMP tests. Would be interesting to see if we've fixed some of the QEMU emulation bugs, either already in master or in Luc's on-list patchset that adds support for emulation of the GICv2 virt extensions... thanks -- PMM
