Programming the active state in the (re)distributor can be an
expensive operation so it makes some sense to try and reduce
the number of accesses as much as possible. So far, we
program the active state on each VM entry, but there is some
opportunity to do less.
An obvious solution is to cache the active state in memory,
and only program it in the HW when conditions change. But
because the HW can also change things under our feet (the active
state can transition from 1 to 0 when the guest does an EOI),
some precautions have to be taken, which amount to only caching
an "inactive" state, and always programing it otherwise.
With this in place, we observe a reduction of around 700 cycles
on a 2GHz GICv2 platform for a NULL hypercall.
Reviewed-by: Christoffer Dall <christoffer.dall@linaro.org>
Signed-off-by: Marc Zyngier <marc.zyngier@arm.com>
The arch timer currently uses edge-triggered semantics in the sense that
the line is never sampled by the vgic and lowering the line from the
timer to the vgic doesn't have any effect on the pending state of
virtual interrupts in the vgic. This means that we do not support a
guest with the otherwise valid behavior of (1) disable interrupts (2)
enable the timer (3) disable the timer (4) enable interrupts. Such a
guest would validly not expect to see any interrupts on real hardware,
but will see interrupts on KVM.
This patch fixes this shortcoming through the following series of
changes.
First, we change the flow of the timer/vgic sync/flush operations. Now
the timer is always flushed/synced before the vgic, because the vgic
samples the state of the timer output. This has the implication that we
move the timer operations in to non-preempible sections, but that is
fine after the previous commit getting rid of hrtimer schedules on every
entry/exit.
Second, we change the internal behavior of the timer, letting the timer
keep track of its previous output state, and only lower/raise the line
to the vgic when the state changes. Note that in theory this could have
been accomplished more simply by signalling the vgic every time the
state *potentially* changed, but we don't want to be hitting the vgic
more often than necessary.
Third, we get rid of the use of the map->active field in the vgic and
instead simply set the interrupt as active on the physical distributor
whenever the input to the GIC is asserted and conversely clear the
physical active state when the input to the GIC is deasserted.
Fourth, and finally, we now initialize the timer PPIs (and all the other
unused PPIs for now), to be level-triggered, and modify the sync code to
sample the line state on HW sync and re-inject a new interrupt if it is
still pending at that time.
Signed-off-by: Christoffer Dall <christoffer.dall@linaro.org>
We currently schedule a soft timer every time we exit the guest if the
timer did not expire while running the guest. This is really not
necessary, because the only work we do in the timer work function is to
kick the vcpu.
Kicking the vcpu does two things:
(1) If the vpcu thread is on a waitqueue, make it runnable and remove it
from the waitqueue.
(2) If the vcpu is running on a different physical CPU from the one
doing the kick, it sends a reschedule IPI.
The second case cannot happen, because the soft timer is only ever
scheduled when the vcpu is not running. The first case is only relevant
when the vcpu thread is on a waitqueue, which is only the case when the
vcpu thread has called kvm_vcpu_block().
Therefore, we only need to make sure a timer is scheduled for
kvm_vcpu_block(), which we do by encapsulating all calls to
kvm_vcpu_block() with kvm_timer_{un}schedule calls.
Additionally, we only schedule a soft timer if the timer is enabled and
unmasked, since it is useless otherwise.
Note that theoretically userspace can use the SET_ONE_REG interface to
change registers that should cause the timer to fire, even if the vcpu
is blocked without a scheduled timer, but this case was not supported
before this patch and we leave it for future work for now.
Signed-off-by: Christoffer Dall <christoffer.dall@linaro.org>
In order to remove the crude hack where we sneak the masked bit
into the timer's control register, make use of the phys_irq_map
API control the active state of the interrupt.
This causes some limited changes to allow for potential error
propagation.
Reviewed-by: Christoffer Dall <christoffer.dall@linaro.org>
Signed-off-by: Marc Zyngier <marc.zyngier@arm.com>
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>
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>
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>
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>
The arch_timer irq numbers (or PPI numbers) are implementation dependent,
so the host virtual timer irq number can be different from guest virtual
timer irq number.
This patch ensures that host virtual timer irq number is read from DTB and
guest virtual timer irq is determined based on vcpu target type.
Signed-off-by: Anup Patel <anup.patel@linaro.org>
Signed-off-by: Pranavkumar Sawargaonkar <pranavkumar@linaro.org>
Signed-off-by: Christoffer Dall <cdall@cs.columbia.edu>
As KVM/arm64 is looming on the horizon, it makes sense to move some
of the common code to a single location in order to reduce duplication.
The code could live anywhere. Actually, most of KVM is already built
with a bunch of ugly ../../.. hacks in the various Makefiles, so we're
not exactly talking about style here. But maybe it is time to start
moving into a less ugly direction.
The include files must be in a "public" location, as they are accessed
from non-KVM files (arch/arm/kernel/asm-offsets.c).
For this purpose, introduce two new locations:
- virt/kvm/arm/ : x86 and ia64 already share the ioapic code in
virt/kvm, so this could be seen as a (very ugly) precedent.
- include/kvm/ : there is already an include/xen, and while the
intent is slightly different, this seems as good a location as
any
Eventually, we should probably have independant Makefiles at every
levels (just like everywhere else in the kernel), but this is just
the first step.
Signed-off-by: Marc Zyngier <marc.zyngier@arm.com>
Signed-off-by: Gleb Natapov <gleb@redhat.com>
Marc Zyngier