aboutsummaryrefslogtreecommitdiff
path: root/include/kvm
AgeCommit message (Collapse)Author
2015-04-07Merge tag 'kvm-arm-for-4.1' of ↵Paolo Bonzini
git://git.kernel.org/pub/scm/linux/kernel/git/kvmarm/kvmarm into 'kvm-next' KVM/ARM changes for v4.1: - fixes for live migration - irqfd support - kvm-io-bus & vgic rework to enable ioeventfd - page ageing for stage-2 translation - various cleanups
2015-03-30KVM: arm/arm64: rework MMIO abort handling to use KVM MMIO busAndre Przywara
Currently we have struct kvm_exit_mmio for encapsulating MMIO abort data to be passed on from syndrome decoding all the way down to the VGIC register handlers. Now as we switch the MMIO handling to be routed through the KVM MMIO bus, it does not make sense anymore to use that structure already from the beginning. So we keep the data in local variables until we put them into the kvm_io_bus framework. Then we fill kvm_exit_mmio in the VGIC only, making it a VGIC private structure. On that way we replace the data buffer in that structure with a pointer pointing to a single location in a local variable, so we get rid of some copying on the way. With all of the virtual GIC emulation code now being registered with the kvm_io_bus, we can remove all of the old MMIO handling code and its dispatching functionality. I didn't bother to rename kvm_exit_mmio (to vgic_mmio or something), because that touches a lot of code lines without any good reason. This is based on an original patch by Nikolay. Signed-off-by: Andre Przywara <andre.przywara@arm.com> Cc: Nikolay Nikolaev <n.nikolaev@virtualopensystems.com> Reviewed-by: Marc Zyngier <marc.zyngier@arm.com> Signed-off-by: Marc Zyngier <marc.zyngier@arm.com>
2015-03-30KVM: arm/arm64: prepare GICv3 emulation to use kvm_io_bus MMIO handlingAndre Przywara
Using the framework provided by the recent vgic.c changes, we register a kvm_io_bus device on mapping the virtual GICv3 resources. The distributor mapping is pretty straight forward, but the redistributors need some more love, since they need to be tagged with the respective redistributor (read: VCPU) they are connected with. We use the kvm_io_bus framework to register one devices per VCPU. Signed-off-by: Andre Przywara <andre.przywara@arm.com> Reviewed-by: Marc Zyngier <marc.zyngier@arm.com> Signed-off-by: Marc Zyngier <marc.zyngier@arm.com>
2015-03-26KVM: arm/arm64: prepare GICv2 emulation to be handled by kvm_io_busAndre Przywara
Using the framework provided by the recent vgic.c changes we register a kvm_io_bus device when initializing the virtual GICv2. Signed-off-by: Andre Przywara <andre.przywara@arm.com> Signed-off-by: Marc Zyngier <marc.zyngier@arm.com>
2015-03-26KVM: arm/arm64: implement kvm_io_bus MMIO handling for the VGICAndre Przywara
Currently we use a lot of VGIC specific code to do the MMIO dispatching. Use the previous reworks to add kvm_io_bus style MMIO handlers. Those are not yet called by the MMIO abort handler, also the actual VGIC emulator function do not make use of it yet, but will be enabled with the following patches. Signed-off-by: Andre Przywara <andre.przywara@arm.com> Reviewed-by: Marc Zyngier <marc.zyngier@arm.com> Signed-off-by: Marc Zyngier <marc.zyngier@arm.com>
2015-03-26KVM: move iodev.h from virt/kvm/ to include/kvmAndre Przywara
iodev.h contains definitions for the kvm_io_bus framework. This is needed both by the generic KVM code in virt/kvm as well as by architecture specific code under arch/. Putting the header file in virt/kvm and using local includes in the architecture part seems at least dodgy to me, so let's move the file into include/kvm, so that a more natural "#include <kvm/iodev.h>" can be used by all of the code. This also solves a problem later when using struct kvm_io_device in arm_vgic.h. Fixing up the FSF address in the GPL header and a wrong include path on the way. Signed-off-by: Andre Przywara <andre.przywara@arm.com> Acked-by: Christoffer Dall <christoffer.dall@linaro.org> Reviewed-by: Marc Zyngier <marc.zyngier@arm.com> Reviewed-by: Marcelo Tosatti <mtosatti@redhat.com> Signed-off-by: Marc Zyngier <marc.zyngier@arm.com>
2015-03-14arm/arm64: KVM: Fix migration race in the arch timerChristoffer Dall
When a VCPU is no longer running, we currently check to see if it has a timer scheduled in the future, and if it does, we schedule a host hrtimer to notify is in case the timer expires while the VCPU is still not running. When the hrtimer fires, we mask the guest's timer and inject the timer IRQ (still relying on the guest unmasking the time when it receives the IRQ). This is all good and fine, but when migration a VM (checkpoint/restore) this introduces a race. It is unlikely, but possible, for the following sequence of events to happen: 1. Userspace stops the VM 2. Hrtimer for VCPU is scheduled 3. Userspace checkpoints the VGIC state (no pending timer interrupts) 4. The hrtimer fires, schedules work in a workqueue 5. Workqueue function runs, masks the timer and injects timer interrupt 6. Userspace checkpoints the timer state (timer masked) At restore time, you end up with a masked timer without any timer interrupts and your guest halts never receiving timer interrupts. Fix this by only kicking the VCPU in the workqueue function, and sample the expired state of the timer when entering the guest again and inject the interrupt and mask the timer only then. Signed-off-by: Christoffer Dall <christoffer.dall@linaro.org> Signed-off-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Christoffer Dall <christoffer.dall@linaro.org>
2015-03-14arm/arm64: KVM: support for un-queuing active IRQsChristoffer Dall
Migrating active interrupts causes the active state to be lost completely. This implements some additional bitmaps to track the active state on the distributor and export this to user space. Signed-off-by: Christoffer Dall <christoffer.dall@linaro.org> Signed-off-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Christoffer Dall <christoffer.dall@linaro.org>
2015-03-14arm/arm64: KVM: Keep elrsr/aisr in sync with software modelChristoffer Dall
There is an interesting bug in the vgic code, which manifests itself when the KVM run loop has a signal pending or needs a vmid generation rollover after having disabled interrupts but before actually switching to the guest. In this case, we flush the vgic as usual, but we sync back the vgic state and exit to userspace before entering the guest. The consequence is that we will be syncing the list registers back to the software model using the GICH_ELRSR and GICH_EISR from the last execution of the guest, potentially overwriting a list register containing an interrupt. This showed up during migration testing where we would capture a state where the VM has masked the arch timer but there were no interrupts, resulting in a hung test. Cc: Marc Zyngier <marc.zyngier@arm.com> Reported-by: Alex Bennee <alex.bennee@linaro.org> Signed-off-by: Christoffer Dall <christoffer.dall@linaro.org> Signed-off-by: Alex Bennée <alex.bennee@linaro.org> Acked-by: Marc Zyngier <marc.zyngier@arm.com> Signed-off-by: Christoffer Dall <christoffer.dall@linaro.org>
2015-03-12arm/arm64: KVM: Kill CONFIG_KVM_ARM_{VGIC,TIMER}Christoffer Dall
We can definitely decide at run-time whether to use the GIC and timers or not, and the extra code and data structures that we allocate space for is really negligable with this config option, so I don't think it's worth the extra complexity of always having to define stub static inlines. The !CONFIG_KVM_ARM_VGIC/TIMER case is pretty much an untested code path anyway, so we're better off just getting rid of it. Signed-off-by: Christoffer Dall <christoffer.dall@linaro.org> Acked-by: Marc Zyngier <marc.zyngier@arm.com>
2015-01-20arm/arm64: KVM: allow userland to request a virtual GICv3Andre Przywara
With all of the GICv3 code in place now we allow userland to ask the kernel for using a virtual GICv3 in the guest. Also we provide the necessary support for guests setting the memory addresses for the virtual distributor and redistributors. This requires some userland code to make use of that feature and explicitly ask for a virtual GICv3. Document that KVM_CREATE_IRQCHIP only works for GICv2, but is considered legacy and using KVM_CREATE_DEVICE is preferred. Signed-off-by: Andre Przywara <andre.przywara@arm.com> Signed-off-by: Christoffer Dall <christoffer.dall@linaro.org>
2015-01-20arm/arm64: KVM: enable kernel side of GICv3 emulationAndre Przywara
With all the necessary GICv3 emulation code in place, we can now connect the code to the GICv3 backend in the kernel. The LR register handling is different depending on the emulated GIC model, so provide different implementations for each. Also allow non-v2-compatible GICv3 implementations (which don't provide MMIO regions for the virtual CPU interface in the DT), but restrict those hosts to support GICv3 guests only. If the device tree provides a GICv2 compatible GICV resource entry, but that one is faulty, just disable the GICv2 emulation and let the user use at least the GICv3 emulation for guests. To provide proper support for the legacy KVM_CREATE_IRQCHIP ioctl, note virtual GICv2 compatibility in struct vgic_params and use it on creating a VGICv2. Signed-off-by: Andre Przywara <andre.przywara@arm.com> Signed-off-by: Christoffer Dall <christoffer.dall@linaro.org>
2015-01-20arm64: KVM: add SGI generation register emulationAndre Przywara
While the generation of a (virtual) inter-processor interrupt (SGI) on a GICv2 works by writing to a MMIO register, GICv3 uses the system register ICC_SGI1R_EL1 to trigger them. Add a trap handler function that calls the new SGI register handler in the GICv3 code. As ICC_SRE_EL1.SRE at this point is still always 0, this will not trap yet, but will only be used later when all the data structures have been initialized properly. Signed-off-by: Andre Przywara <andre.przywara@arm.com> Reviewed-by: Christoffer Dall <christoffer.dall@linaro.org> Signed-off-by: Christoffer Dall <christoffer.dall@linaro.org>
2015-01-20arm/arm64: KVM: add virtual GICv3 distributor emulationAndre Przywara
With everything separated and prepared, we implement a model of a GICv3 distributor and redistributors by using the existing framework to provide handler functions for each register group. Currently we limit the emulation to a model enforcing a single security state, with SRE==1 (forcing system register access) and ARE==1 (allowing more than 8 VCPUs). We share some of the functions provided for GICv2 emulation, but take the different ways of addressing (v)CPUs into account. Save and restore is currently not implemented. Similar to the split-off of the GICv2 specific code, the new emulation code goes into a new file (vgic-v3-emul.c). Signed-off-by: Andre Przywara <andre.przywara@arm.com> Signed-off-by: Christoffer Dall <christoffer.dall@linaro.org>
2015-01-20arm/arm64: KVM: make the value of ICC_SRE_EL1 a per-VM variableAndre Przywara
ICC_SRE_EL1 is a system register allowing msr/mrs accesses to the GIC CPU interface for EL1 (guests). Currently we force it to 0, but for proper GICv3 support we have to allow guests to use it (depending on their selected virtual GIC model). So add ICC_SRE_EL1 to the list of saved/restored registers on a world switch, but actually disallow a guest to change it by only restoring a fixed, once-initialized value. This value depends on the GIC model userland has chosen for a guest. Signed-off-by: Andre Przywara <andre.przywara@arm.com> Reviewed-by: Christoffer Dall <christoffer.dall@linaro.org> Acked-by: Marc Zyngier <marc.zyngier@arm.com> Signed-off-by: Christoffer Dall <christoffer.dall@linaro.org>
2015-01-20arm/arm64: KVM: make the maximum number of vCPUs a per-VM valueAndre Przywara
Currently the maximum number of vCPUs supported is a global value limited by the used GIC model. GICv3 will lift this limit, but we still need to observe it for guests using GICv2. So the maximum number of vCPUs is per-VM value, depending on the GIC model the guest uses. Store and check the value in struct kvm_arch, but keep it down to 8 for now. Signed-off-by: Andre Przywara <andre.przywara@arm.com> Signed-off-by: Christoffer Dall <christoffer.dall@linaro.org>
2015-01-20arm/arm64: KVM: introduce per-VM opsAndre Przywara
Currently we only have one virtual GIC model supported, so all guests use the same emulation code. With the addition of another model we end up with different guests using potentially different vGIC models, so we have to split up some functions to be per VM. Introduce a vgic_vm_ops struct to hold function pointers for those functions that are different and provide the necessary code to initialize them. Also split up the vgic_init() function to separate out VGIC model specific functionality into a separate function, which will later be different for a GICv3 model. Signed-off-by: Andre Przywara <andre.przywara@arm.com> Reviewed-by: Christoffer Dall <christoffer.dall@linaro.org> Signed-off-by: Christoffer Dall <christoffer.dall@linaro.org>
2015-01-20arm/arm64: KVM: pass down user space provided GIC type into vGIC codeAndre Przywara
With the introduction of a second emulated GIC model we need to let userspace specify the GIC model to use for each VM. Pass the userspace provided value down into the vGIC code and store it there to differentiate later. Signed-off-by: Andre Przywara <andre.przywara@arm.com> Acked-by: Christoffer Dall <christoffer.dall@linaro.org> Acked-by: Marc Zyngier <marc.zyngier@arm.com> Signed-off-by: Christoffer Dall <christoffer.dall@linaro.org>
2014-12-15arm/arm64: KVM: Require in-kernel vgic for the arch timersChristoffer Dall
It is curently possible to run a VM with architected timers support without creating an in-kernel VGIC, which will result in interrupts from the virtual timer going nowhere. To address this issue, move the architected timers initialization to the time when we run a VCPU for the first time, and then only initialize (and enable) the architected timers if we have a properly created and initialized in-kernel VGIC. When injecting interrupts from the virtual timer to the vgic, the current setup should ensure that this never calls an on-demand init of the VGIC, which is the only call path that could return an error from kvm_vgic_inject_irq(), so capture the return value and raise a warning if there's an error there. We also change the kvm_timer_init() function from returning an int to be a void function, since the function always succeeds. Reviewed-by: Marc Zyngier <marc.zyngier@arm.com> Signed-off-by: Christoffer Dall <christoffer.dall@linaro.org>
2014-12-13arm/arm64: KVM: Add (new) vgic_initialized macroChristoffer Dall
Some code paths will need to check to see if the internal state of the vgic has been initialized (such as when creating new VCPUs), so introduce such a macro that checks the nr_cpus field which is set when the vgic has been initialized. Also set nr_cpus = 0 in kvm_vgic_destroy, because the error path in vgic_init() will call this function, and code should never errornously assume the vgic to be properly initialized after an error. Acked-by: Marc Zyngier <marc.zyngier@arm.com> Reviewed-by: Eric Auger <eric.auger@linaro.org> Signed-off-by: Christoffer Dall <christoffer.dall@linaro.org>
2014-12-13arm/arm64: KVM: Rename vgic_initialized to vgic_readyChristoffer Dall
The vgic_initialized() macro currently returns the state of the vgic->ready flag, which indicates if the vgic is ready to be used when running a VM, not specifically if its internal state has been initialized. Rename the macro accordingly in preparation for a more nuanced initialization flow. Acked-by: Marc Zyngier <marc.zyngier@arm.com> Reviewed-by: Eric Auger <eric.auger@linaro.org> Signed-off-by: Christoffer Dall <christoffer.dall@linaro.org>
2014-12-13arm/arm64: KVM: vgic: move reset initialization into vgic_init_maps()Peter Maydell
VGIC initialization currently happens in three phases: (1) kvm_vgic_create() (triggered by userspace GIC creation) (2) vgic_init_maps() (triggered by userspace GIC register read/write requests, or from kvm_vgic_init() if not already run) (3) kvm_vgic_init() (triggered by first VM run) We were doing initialization of some state to correspond with the state of a freshly-reset GIC in kvm_vgic_init(); this is too late, since it will overwrite changes made by userspace using the register access APIs before the VM is run. Move this initialization earlier, into the vgic_init_maps() phase. This fixes a bug where QEMU could successfully restore a saved VM state snapshot into a VM that had already been run, but could not restore it "from cold" using the -loadvm command line option (the symptoms being that the restored VM would run but interrupts were ignored). Finally rename vgic_init_maps to vgic_init and renamed kvm_vgic_init to kvm_vgic_map_resources. [ This patch is originally written by Peter Maydell, but I have modified it somewhat heavily, renaming various bits and moving code around. If something is broken, I am to be blamed. - Christoffer ] Acked-by: Marc Zyngier <marc.zyngier@arm.com> Reviewed-by: Eric Auger <eric.auger@linaro.org> Signed-off-by: Peter Maydell <peter.maydell@linaro.org> Signed-off-by: Christoffer Dall <christoffer.dall@linaro.org>
2014-10-16arm/arm64: KVM: Fix BE accesses to GICv2 EISR and ELRSR regsChristoffer Dall
The EIRSR and ELRSR registers are 32-bit registers on GICv2, and we store these as an array of two such registers on the vgic vcpu struct. However, we access them as a single 64-bit value or as a bitmap pointer in the generic vgic code, which breaks BE support. Instead, store them as u64 values on the vgic structure and do the word-swapping in the assembly code, which already handles the byte order for BE systems. Tested-by: Victor Kamensky <victor.kamensky@linaro.org> Acked-by: Marc Zyngier <marc.zyngier@arm.com> Signed-off-by: Christoffer Dall <christoffer.dall@linaro.org>
2014-10-07ARM: KVM: fix vgic-disabled buildArnd Bergmann
The vgic code can be disabled in Kconfig and there are dummy implementations of most of the provided API functions for the disabled case. However, the newly introduced kvm_vgic_destroy/kvm_vgic_vcpu_destroy functions are lacking those dummies, resulting in this build error: arch/arm/kvm/arm.c: In function 'kvm_arch_destroy_vm': arch/arm/kvm/arm.c:165:2: error: implicit declaration of function 'kvm_vgic_destroy' [-Werror=implicit-function-declaration] kvm_vgic_destroy(kvm); ^ arch/arm/kvm/arm.c: In function 'kvm_arch_vcpu_free': arch/arm/kvm/arm.c:248:2: error: implicit declaration of function 'kvm_vgic_vcpu_destroy' [-Werror=implicit-function-declaration] kvm_vgic_vcpu_destroy(vcpu); ^ This adds two inline helpers to get it to build again in this configuration. Signed-off-by: Arnd Bergmann <arnd@arndb.de> Fixes: c1bfb577add ("arm/arm64: KVM: vgic: switch to dynamic allocation") Signed-off-by: Christoffer Dall <christoffer.dall@linaro.org>
2014-09-18arm/arm64: KVM: vgic: delay vgic allocation until init timeMarc Zyngier
It is now quite easy to delay the allocation of the vgic tables until we actually require it to be up and running (when the first vcpu is kicking around, or someones tries to access the GIC registers). This allow us to allocate memory for the exact number of CPUs we have. As nobody configures the number of interrupts just yet, use a fallback to VGIC_NR_IRQS_LEGACY. Reviewed-by: Christoffer Dall <christoffer.dall@linaro.org> Signed-off-by: Marc Zyngier <marc.zyngier@arm.com>
2014-09-18arm/arm64: KVM: vgic: kill VGIC_NR_IRQSMarc Zyngier
Nuke VGIC_NR_IRQS entierly, now that the distributor instance contains the number of IRQ allocated to this GIC. Also add VGIC_NR_IRQS_LEGACY to preserve the current API. Reviewed-by: Christoffer Dall <christoffer.dall@linaro.org> Signed-off-by: Marc Zyngier <marc.zyngier@arm.com>
2014-09-18arm/arm64: KVM: vgic: handle out-of-range MMIO accessesMarc Zyngier
Now that we can (almost) dynamically size the number of interrupts, we're facing an interesting issue: We have to evaluate at runtime whether or not an access hits a valid register, based on the sizing of this particular instance of the distributor. Furthermore, the GIC spec says that accessing a reserved register is RAZ/WI. For this, add a new field to our range structure, indicating the number of bits a single interrupts uses. That allows us to find out whether or not the access is in range. Reviewed-by: Christoffer Dall <christoffer.dall@linaro.org> Signed-off-by: Marc Zyngier <marc.zyngier@arm.com>
2014-09-18arm/arm64: KVM: vgic: kill VGIC_MAX_CPUSMarc Zyngier
We now have the information about the number of CPU interfaces in the distributor itself. Let's get rid of VGIC_MAX_CPUS, and just rely on KVM_MAX_VCPUS where we don't have the choice. Yet. Reviewed-by: Christoffer Dall <christoffer.dall@linaro.org> Signed-off-by: Marc Zyngier <marc.zyngier@arm.com>
2014-09-18arm/arm64: KVM: vgic: Parametrize VGIC_NR_SHARED_IRQSMarc Zyngier
Having a dynamic number of supported interrupts means that we cannot relly on VGIC_NR_SHARED_IRQS being fixed anymore. Instead, make it take the distributor structure as a parameter, so it can return the right value. Reviewed-by: Christoffer Dall <christoffer.dall@linaro.org> Signed-off-by: Marc Zyngier <marc.zyngier@arm.com>
2014-09-18arm/arm64: KVM: vgic: switch to dynamic allocationMarc Zyngier
So far, all the VGIC data structures are statically defined by the *maximum* number of vcpus and interrupts it supports. It means that we always have to oversize it to cater for the worse case. Start by changing the data structures to be dynamically sizeable, and allocate them at runtime. The sizes are still very static though. Signed-off-by: Marc Zyngier <marc.zyngier@arm.com>
2014-09-18arm/arm64: KVM: vgic: Improve handling of GICD_I{CS}PENDRnChristoffer Dall
Writes to GICD_ISPENDRn and GICD_ICPENDRn are currently not handled correctly for level-triggered interrupts. The spec states that for level-triggered interrupts, writes to the GICD_ISPENDRn activate the output of a flip-flop which is in turn or'ed with the actual input interrupt signal. Correspondingly, writes to GICD_ICPENDRn simply deactivates the output of that flip-flop, but does not (of course) affect the external input signal. Reads from GICC_IAR will also deactivate the flip-flop output. This requires us to track the state of the level-input separately from the state in the flip-flop. We therefore introduce two new variables on the distributor struct to track these two states. Astute readers may notice that this is introducing more state than required (because an OR of the two states gives you the pending state), but the remaining vgic code uses the pending bitmap for optimized operations to figure out, at the end of the day, if an interrupt is pending or not on the distributor side. Refactoring the code to consider the two state variables all the places where we currently access the precomputed pending value, did not look pretty. Signed-off-by: Christoffer Dall <christoffer.dall@linaro.org>
2014-09-18arm/arm64: KVM: Rename irq_active to irq_queuedChristoffer Dall
We have a special bitmap on the distributor struct to keep track of when level-triggered interrupts are queued on the list registers. This was named irq_active, which is confusing, because the active state of an interrupt as per the GIC spec is a different thing, not specifically related to edge-triggered/level-triggered configurations but rather indicates an interrupt which has been ack'ed but not yet eoi'ed. Rename the bitmap and the corresponding accessor functions to irq_queued to clarify what this is actually used for. Signed-off-by: Christoffer Dall <christoffer.dall@linaro.org>
2014-09-18arm/arm64: KVM: Rename irq_state to irq_pendingChristoffer Dall
The irq_state field on the distributor struct is ambiguous in its meaning; the comment says it's the level of the input put, but that doesn't make much sense for edge-triggered interrupts. The code actually uses this state variable to check if the interrupt is in the pending state on the distributor so clarify the comment and rename the actual variable and accessor methods. Acked-by: Marc Zyngier <marc.zyngier@arm.com> Signed-off-by: Christoffer Dall <christoffer.dall@linaro.org>
2014-07-11KVM: ARM: vgic: add the GICv3 backendMarc Zyngier
Introduce the support code for emulating a GICv2 on top of GICv3 hardware. Acked-by: Catalin Marinas <catalin.marinas@arm.com> Signed-off-by: Marc Zyngier <marc.zyngier@arm.com>
2014-07-11arm64: KVM: split GICv2 world switch from hyp codeMarc Zyngier
Move the GICv2 world switch code into its own file, and add the necessary indirection to the arm64 switch code. Also introduce a new type field to the vgic_params structure. Acked-by: Catalin Marinas <catalin.marinas@arm.com> Reviewed-by: Christoffer Dall <christoffer.dall@linaro.org> Signed-off-by: Marc Zyngier <marc.zyngier@arm.com>
2014-07-11KVM: ARM: vgic: revisit implementation of irqchip_in_kernelMarc Zyngier
So far, irqchip_in_kernel() was implemented by testing the value of vctrl_base, which worked fine with GICv2. With GICv3, this field is useless, as we're using system registers instead of a emmory mapped interface. To solve this, add a boolean flag indicating if the we're using a vgic or not. Reviewed-by: Christoffer Dall <christoffer.dall@linaro.org> Signed-off-by: Marc Zyngier <marc.zyngier@arm.com>
2014-07-11KVM: ARM: vgic: split GICv2 backend from the main vgic codeMarc Zyngier
Brutally hack the innocent vgic code, and move the GICv2 specific code to its own file, using vgic_ops and vgic_params as a way to pass information between the two blocks. Acked-by: Catalin Marinas <catalin.marinas@arm.com> Reviewed-by: Christoffer Dall <christoffer.dall@linaro.org> Signed-off-by: Marc Zyngier <marc.zyngier@arm.com>
2014-07-11KVM: ARM: introduce vgic_params structureMarc Zyngier
Move all the data specific to a given GIC implementation into its own little structure. Acked-by: Catalin Marinas <catalin.marinas@arm.com> Reviewed-by: Christoffer Dall <christoffer.dall@linaro.org> Signed-off-by: Marc Zyngier <marc.zyngier@arm.com>
2014-07-11KVM: ARM: vgic: introduce vgic_enableMarc Zyngier
Move the code dealing with enabling the VGIC on to vgic_ops. Acked-by: Catalin Marinas <catalin.marinas@arm.com> Reviewed-by: Christoffer Dall <christoffer.dall@linaro.org> Signed-off-by: Marc Zyngier <marc.zyngier@arm.com>
2014-07-11KVM: ARM: vgic: abstract VMCR accessMarc Zyngier
Instead of directly messing with with the GICH_VMCR bits for the CPU interface save/restore code, add accessors that encode/decode the entire set of registers exposed by VMCR. Not the most efficient thing, but given that this code is only used by the save/restore code, performance is far from being critical. Reviewed-by: Christoffer Dall <christoffer.dall@linaro.org> Signed-off-by: Marc Zyngier <marc.zyngier@arm.com>
2014-07-11KVM: ARM: vgic: move underflow handling to vgic_opsMarc Zyngier
Move the code dealing with LR underflow handling to its own functions, and make them accessible through vgic_ops. Acked-by: Catalin Marinas <catalin.marinas@arm.com> Reviewed-by: Christoffer Dall <christoffer.dall@linaro.org> Signed-off-by: Marc Zyngier <marc.zyngier@arm.com>
2014-07-11KVM: ARM: vgic: abstract MISR decodingMarc Zyngier
Instead of directly dealing with the GICH_MISR bits, move the code to its own function and use a couple of public flags to represent the actual state. Acked-by: Catalin Marinas <catalin.marinas@arm.com> Reviewed-by: Christoffer Dall <christoffer.dall@linaro.org> Signed-off-by: Marc Zyngier <marc.zyngier@arm.com>
2014-07-11KVM: ARM: vgic: abstract EISR bitmap accessMarc Zyngier
Move the GICH_EISR access to its own function. Acked-by: Catalin Marinas <catalin.marinas@arm.com> Reviewed-by: Christoffer Dall <christoffer.dall@linaro.org> Signed-off-by: Marc Zyngier <marc.zyngier@arm.com>
2014-07-11KVM: ARM: vgic: abstract access to the ELRSR bitmapMarc Zyngier
Move the GICH_ELRSR access to its own functions, and add them to the vgic_ops structure. Acked-by: Catalin Marinas <catalin.marinas@arm.com> Reviewed-by: Christoffer Dall <christoffer.dall@linaro.org> Signed-off-by: Marc Zyngier <marc.zyngier@arm.com>
2014-07-11KVM: ARM: vgic: introduce vgic_ops and LR manipulation primitivesMarc Zyngier
In order to split the various register manipulation from the main vgic code, introduce a vgic_ops structure, and start by abstracting the LR manipulation code with a couple of accessors. Reviewed-by: Christoffer Dall <christoffer.dall@linaro.org> Signed-off-by: Marc Zyngier <marc.zyngier@arm.com>
2014-07-11KVM: arm/arm64: vgic: move GICv2 registers to their own structureMarc Zyngier
In order to make way for the GICv3 registers, move the v2-specific registers to their own structure. Acked-by: Catalin Marinas <catalin.marinas@arm.com> Reviewed-by: Christoffer Dall <christoffer.dall@linaro.org> Signed-off-by: Marc Zyngier <marc.zyngier@arm.com>
2014-07-11arm64: KVM: allow export and import of generic timer regsAlex Bennée
For correct guest suspend/resume behaviour we need to ensure we include the generic timer registers for 64 bit guests. As CONFIG_KVM_ARM_TIMER is always set for arm64 we don't need to worry about null implementations. However I have re-jigged the kvm_arm_timer_set/get_reg declarations to be in the common include/kvm/arm_arch_timer.h headers. Acked-by: Marc Zyngier <marc.zyngier@arm.com> Reviewed-by: Christoffer Dall <christoffer.dall@linaro.org> Signed-off-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Christoffer Dall <christoffer.dall@linaro.org>
2014-03-06ARM: KVM: fix non-VGIC compilationMarc Zyngier
Add a stub for kvm_vgic_addr when compiling without CONFIG_KVM_ARM_VGIC. The usefulness of this configurarion is extremely doubtful, but let's fix it anyway (until we decide that we'll always support a VGIC). Reported-by: Michele Paolino <m.paolino@virtualopensystems.com> Cc: Paolo Bonzini <pbonzini@redhat.com> Cc: Christoffer Dall <christoffer.dall@linaro.org> Signed-off-by: Marc Zyngier <marc.zyngier@arm.com> Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2013-12-21KVM: arm-vgic: Set base addr through device APIChristoffer Dall
Support setting the distributor and cpu interface base addresses in the VM physical address space through the KVM_{SET,GET}_DEVICE_ATTR API in addition to the ARM specific API. This has the added benefit of being able to share more code in user space and do things in a uniform manner. Also deprecate the older API at the same time, but backwards compatibility will be maintained. Acked-by: Marc Zyngier <marc.zyngier@arm.com> Signed-off-by: Christoffer Dall <christoffer.dall@linaro.org>
2013-08-30ARM: KVM: vgic: Bump VGIC_NR_IRQS to 256Christoffer Dall
The Versatile Express TC2 board, which we use as our main emulated platform in QEMU, defines 160+32 == 192 interrupts, so limiting the number of interrupts to 128 is not quite going to cut it for real board emulation. Note that this didn't use to be a problem because QEMU was buggy and only defined 128 interrupts until recently. Signed-off-by: Christoffer Dall <christoffer.dall@linaro.org> Signed-off-by: Marc Zyngier <marc.zyngier@arm.com> Signed-off-by: Gleb Natapov <gleb@redhat.com>