Use __kvm_guest_{enter|exit} instead of kvm_guest_{enter|exit}
where interrupts are disabled.
Signed-off-by: NChristian Borntraeger <borntraeger@de.ibm.com>
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

When userland injects a SPI via the KVM_IRQ_LINE ioctl we currently
only check it against a fixed limit, which historically is set
to 127. With the new dynamic IRQ allocation the effective limit may
actually be smaller (64).
So when now a malicious or buggy userland injects a SPI in that
range, we spill over on our VGIC bitmaps and bytemaps memory.
I could trigger a host kernel NULL pointer dereference with current
mainline by injecting some bogus IRQ number from a hacked kvmtool:
-----------------
....
DEBUG: kvm_vgic_inject_irq(kvm, cpu=0, irq=114, level=1)
DEBUG: vgic_update_irq_pending(kvm, cpu=0, irq=114, level=1)
DEBUG: IRQ #114 still in the game, writing to bytemap now...
Unable to handle kernel NULL pointer dereference at virtual address 00000000
pgd = ffffffc07652e000
[00000000] *pgd=00000000f658b003, *pud=00000000f658b003, *pmd=0000000000000000
Internal error: Oops: 96000006 [#1] PREEMPT SMP
Modules linked in:
CPU: 1 PID: 1053 Comm: lkvm-msi-irqinj Not tainted 4.0.0-rc7+ #3027
Hardware name: FVP Base (DT)
task: ffffffc0774e9680 ti: ffffffc0765a8000 task.ti: ffffffc0765a8000
PC is at kvm_vgic_inject_irq+0x234/0x310
LR is at kvm_vgic_inject_irq+0x30c/0x310
pc : [<ffffffc0000ae0a8>] lr : [<ffffffc0000ae180>] pstate: 80000145
.....

So this patch fixes this by checking the SPI number against the
actual limit. Also we remove the former legacy hard limit of
127 in the ioctl code.
Signed-off-by: NAndre Przywara <andre.przywara@arm.com>
Reviewed-by: NChristoffer Dall <christoffer.dall@linaro.org>
CC: <stable@vger.kernel.org> # 4.0, 3.19, 3.18
[maz: wrap KVM_ARM_IRQ_GIC_MAX with #ifndef __KERNEL__,
as suggested by Christopher Covington]
Signed-off-by: NMarc Zyngier <marc.zyngier@arm.com>

As the infrastructure for eventfd has now been merged, report the
ioeventfd capability as being supported.
Signed-off-by: NNikolay Nikolaev <n.nikolaev@virtualopensystems.com>
[maz: grouped the case entry with the others, fixed commit log]
Signed-off-by: NMarc 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: NChristoffer Dall <christoffer.dall@linaro.org>
Signed-off-by: NAlex Bennée <alex.bennee@linaro.org>
Signed-off-by: NChristoffer Dall <christoffer.dall@linaro.org>

To cleanly restore an SMP VM we need to ensure that the current pause
state of each vcpu is correctly recorded. Things could get confused if
the CPU starts running after migration restore completes when it was
paused before it state was captured.

We use the existing KVM_GET/SET_MP_STATE ioctl to do this. The arm/arm64
interface is a lot simpler as the only valid states are
KVM_MP_STATE_RUNNABLE and KVM_MP_STATE_STOPPED.
Signed-off-by: NAlex Bennée <alex.bennee@linaro.org>
Signed-off-by: NChristoffer Dall <christoffer.dall@linaro.org>

This patch enables irqfd on arm/arm64.

Both irqfd and resamplefd are supported. Injection is implemented
in vgic.c without routing.

This patch enables CONFIG_HAVE_KVM_EVENTFD and CONFIG_HAVE_KVM_IRQFD.

KVM_CAP_IRQFD is now advertised. KVM_CAP_IRQFD_RESAMPLE capability
automatically is advertised as soon as CONFIG_HAVE_KVM_IRQFD is set.

Irqfd injection is restricted to SPI. The rationale behind not
supporting PPI irqfd injection is that any device using a PPI would
be a private-to-the-CPU device (timer for instance), so its state
would have to be context-switched along with the VCPU and would
require in-kernel wiring anyhow. It is not a relevant use case for
irqfds.
Signed-off-by: NEric Auger <eric.auger@linaro.org>
Reviewed-by: NChristoffer Dall <christoffer.dall@linaro.org>
Acked-by: NMarc Zyngier <marc.zyngier@arm.com>
Signed-off-by: NChristoffer Dall <christoffer.dall@linaro.org>

On arm/arm64 the VGIC is dynamically instantiated and it is useful
to expose its state, especially for irqfd setup.

This patch defines __KVM_HAVE_ARCH_INTC_INITIALIZED and
implements kvm_arch_intc_initialized.
Signed-off-by: NEric Auger <eric.auger@linaro.org>
Acked-by: NChristoffer Dall <christoffer.dall@linaro.org>
Reviewed-by: NAndre Przywara <andre.przywara@arm.com>
Acked-by: NMarc Zyngier <marc.zyngier@arm.com>
Signed-off-by: NChristoffer 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: NChristoffer Dall <christoffer.dall@linaro.org>
Acked-by: NMarc Zyngier <marc.zyngier@arm.com>

IS_ENABLED gives compile-time checking and keeps the code clearer.

The one exception is inside kvm_vm_ioctl_check_extension, where
the established idiom is to wrap the case labels with an #ifdef.
Signed-off-by: NChristoffer Dall <christoffer.dall@linaro.org>
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

This patch extends trace_kvm_exit() to include KVM exit reasons
(i.e. EC of HSR). The tracing function then dumps both exit reason
and PC of vCPU, shown as the following. Tracing tools can use this
new exit_reason field to better understand the behavior of guest VMs.

886.301252: kvm_exit:             HSR_EC: 0x0024, PC: 0xfffffe0000506b28
Signed-off-by: NWei Huang <wei@redhat.com>
Signed-off-by: NChristoffer Dall <christoffer.dall@linaro.org>

Trying to emulate the behaviour of set/way cache ops is fairly
pointless, as there are too many ways we can end-up missing stuff.
Also, there is some system caches out there that simply ignore
set/way operations.

So instead of trying to implement them, let's convert it to VA ops,
and use them as a way to re-enable the trapping of VM ops. That way,
we can detect the point when the MMU/caches are turned off, and do
a full VM flush (which is what the guest was trying to do anyway).

This allows a 32bit zImage to boot on the APM thingy, and will
probably help bootloaders in general.
Signed-off-by: NMarc Zyngier <marc.zyngier@arm.com>
Signed-off-by: NChristoffer Dall <christoffer.dall@linaro.org>

The return value of kvm_arch_vcpu_postcreate is not checked in its
caller.  This is okay, because only x86 provides vcpu_postcreate right
now and it could only fail if vcpu_load failed.  But that is not
possible during KVM_CREATE_VCPU (kvm_arch_vcpu_load is void, too), so
just get rid of the unchecked return value.
Signed-off-by: NDominik Dingel <dingel@linux.vnet.ibm.com>
Acked-by: NCornelia Huck <cornelia.huck@de.ibm.com>
Signed-off-by: NChristian Borntraeger <borntraeger@de.ibm.com>

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: NAndre Przywara <andre.przywara@arm.com>
Signed-off-by: NChristoffer Dall <christoffer.dall@linaro.org>

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: NAndre Przywara <andre.przywara@arm.com>
Acked-by: NChristoffer Dall <christoffer.dall@linaro.org>
Acked-by: NMarc Zyngier <marc.zyngier@arm.com>
Signed-off-by: NChristoffer Dall <christoffer.dall@linaro.org>

The virtual MPIDR registers (containing topology information) for the
guest are currently mapped linearily to the vcpu_id. Improve this
mapping for arm64 by using three levels to not artificially limit the
number of vCPUs.
To help this, change and rename the kvm_vcpu_get_mpidr() function to
mask off the non-affinity bits in the MPIDR register.
Also add an accessor to later allow easier access to a vCPU with a
given MPIDR. Use this new accessor in the PSCI emulation.
Signed-off-by: NAndre Przywara <andre.przywara@arm.com>
Reviewed-by: NChristoffer Dall <christoffer.dall@linaro.org>
Reviewed-by: NMarc Zyngier <marc.zyngier@arm.com>
Signed-off-by: NChristoffer Dall <christoffer.dall@linaro.org>

This patch enables ARMv8 ditry page logging support. Plugs ARMv8 into generic
layer through Kconfig symbol, and drops earlier ARM64 constraints to enable
logging at architecture layer.
Reviewed-by: NChristoffer Dall <christoffer.dall@linaro.org>
Signed-off-by: NMario Smarduch <m.smarduch@samsung.com>

Add support to track dirty pages between user space KVM_GET_DIRTY_LOG ioctl
calls. We call kvm_get_dirty_log_protect() function to do most of the work.
Reviewed-by: NChristoffer Dall <christoffer.dall@linaro.org>
Reviewed-by: NMarc Zyngier <marc.zyngier@arm.com>
Signed-off-by: NMario Smarduch <m.smarduch@samsung.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: NMarc Zyngier <marc.zyngier@arm.com>
Signed-off-by: NChristoffer Dall <christoffer.dall@linaro.org>

When the vgic initializes its internal state it does so based on the
number of VCPUs available at the time.  If we allow KVM to create more
VCPUs after the VGIC has been initialized, we are likely to error out in
unfortunate ways later, perform buffer overflows etc.
Acked-by: NMarc Zyngier <marc.zyngier@arm.com>
Reviewed-by: NEric Auger <eric.auger@linaro.org>
Signed-off-by: NChristoffer Dall <christoffer.dall@linaro.org>

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: NMarc Zyngier <marc.zyngier@arm.com>
Reviewed-by: NEric Auger <eric.auger@linaro.org>
Signed-off-by: NChristoffer Dall <christoffer.dall@linaro.org>

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: NMarc Zyngier <marc.zyngier@arm.com>
Reviewed-by: NEric Auger <eric.auger@linaro.org>
Signed-off-by: NPeter Maydell <peter.maydell@linaro.org>
Signed-off-by: NChristoffer Dall <christoffer.dall@linaro.org>

Introduce a new function to unmap user RAM regions in the stage2 page
tables.  This is needed on reboot (or when the guest turns off the MMU)
to ensure we fault in pages again and make the dcache, RAM, and icache
coherent.

Using unmap_stage2_range for the whole guest physical range does not
work, because that unmaps IO regions (such as the GIC) which will not be
recreated or in the best case faulted in on a page-by-page basis.

Call this function on secondary and subsequent calls to the
KVM_ARM_VCPU_INIT ioctl so that a reset VCPU will detect the guest
Stage-1 MMU is off when faulting in pages and make the caches coherent.
Acked-by: NMarc Zyngier <marc.zyngier@arm.com>
Signed-off-by: NChristoffer Dall <christoffer.dall@linaro.org>

It is not clear that this ioctl can be called multiple times for a given
vcpu.  Userspace already does this, so clarify the ABI.

Also specify that userspace is expected to always make secondary and
subsequent calls to the ioctl with the same parameters for the VCPU as
the initial call (which userspace also already does).

Add code to check that userspace doesn't violate that ABI in the future,
and move the kvm_vcpu_set_target() function which is currently
duplicated between the 32-bit and 64-bit versions in guest.c to a common
static function in arm.c, shared between both architectures.
Acked-by: NMarc Zyngier <marc.zyngier@arm.com>
Signed-off-by: NChristoffer Dall <christoffer.dall@linaro.org>

When userspace resets the vcpu using KVM_ARM_VCPU_INIT, we should also
reset the HCR, because we now modify the HCR dynamically to
enable/disable trapping of guest accesses to the VM registers.

This is crucial for reboot of VMs working since otherwise we will not be
doing the necessary cache maintenance operations when faulting in pages
with the guest MMU off.
Acked-by: NMarc Zyngier <marc.zyngier@arm.com>
Signed-off-by: NChristoffer Dall <christoffer.dall@linaro.org>

The implementation of KVM_ARM_VCPU_INIT is currently not doing what
userspace expects, namely making sure that a vcpu which may have been
turned off using PSCI is returned to its initial state, which would be
powered on if userspace does not set the KVM_ARM_VCPU_POWER_OFF flag.

Implement the expected functionality and clarify the ABI.
Acked-by: NMarc Zyngier <marc.zyngier@arm.com>
Signed-off-by: NChristoffer Dall <christoffer.dall@linaro.org>

If a VCPU was originally started with power off (typically to be brought
up by PSCI in SMP configurations), there is no need to clear the
POWER_OFF flag in the kernel, as this flag is only tested during the
init ioctl itself.
Acked-by: NMarc Zyngier <marc.zyngier@arm.com>
Signed-off-by: NChristoffer Dall <christoffer.dall@linaro.org>

This patch adds the necessary support for all host kernel PGSIZE and
VA_SPACE configuration options for both EL2 and the Stage-2 page tables.

However, for 40bit and 42bit PARange systems, the architecture mandates
that VTCR_EL2.SL0 is maximum 1, resulting in fewer levels of stage-2
pagge tables than levels of host kernel page tables.  At the same time,
systems with a PARange > 42bit, we limit the IPA range by always setting
VTCR_EL2.T0SZ to 24.

To solve the situation with different levels of page tables for Stage-2
translation than the host kernel page tables, we allocate a dummy PGD
with pointers to our actual inital level Stage-2 page table, in order
for us to reuse the kernel pgtable manipulation primitives.  Reproducing
all these in KVM does not look pretty and unnecessarily complicates the
32-bit side.

Systems with a PARange < 40bits are not yet supported.

 [ I have reworked this patch from its original form submitted by
   Jungseok to take the architecture constraints into consideration.
   There were too many changes from the original patch for me to
   preserve the authorship.  Thanks to Catalin Marinas for his help in
   figuring out a good solution to this challenge.  I have also fixed
   various bugs and missing error code handling from the original
   patch. - Christoffer ]
Reviewed-by: NCatalin Marinas <catalin.marinas@arm.com>
Acked-by: NMarc Zyngier <marc.zyngier@arm.com>
Signed-off-by: NJungseok Lee <jungseoklee85@gmail.com>
Signed-off-by: NChristoffer Dall <christoffer.dall@linaro.org>

On some platforms with no power management capabilities, the hotplug
implementation is allowed to return from a smp_ops.cpu_die() call as a
function return. Upon a CPU onlining event, the KVM CPU notifier tries
to reinstall the hyp stub, which fails on platform where no reset took
place following a hotplug event, with the message:

CPU1: smp_ops.cpu_die() returned, trying to resuscitate
CPU1: Booted secondary processor
Kernel panic - not syncing: unexpected prefetch abort in Hyp mode at: 0x80409540
unexpected data abort in Hyp mode at: 0x80401fe8
unexpected HVC/SVC trap in Hyp mode at: 0x805c6170

since KVM code is trying to reinstall the stub on a system where it is
already configured.

To prevent this issue, this patch adds a check in the KVM hotplug
notifier that detects if the HYP stub really needs re-installing when a
CPU is onlined and skips the installation call if the stub is already in
place, which means that the CPU has not been reset.
Signed-off-by: NVladimir Murzin <vladimir.murzin@arm.com>
Acked-by: NLorenzo Pieralisi <lorenzo.pieralisi@arm.com>
Acked-by: NMarc Zyngier <marc.zyngier@arm.com>
Signed-off-by: NChristoffer Dall <christoffer.dall@linaro.org>

The current aarch64 calculation for VTTBR_BADDR_MASK masks only 39 bits
and not all the bits in the PA range. This is clearly a bug that
manifests itself on systems that allocate memory in the higher address
space range.

 [ Modified from Joel's original patch to be based on PHYS_MASK_SHIFT
   instead of a hard-coded value and to move the alignment check of the
   allocation to mmu.c.  Also added a comment explaining why we hardcode
   the IPA range and changed the stage-2 pgd allocation to be based on
   the 40 bit IPA range instead of the maximum possible 48 bit PA range.
   - Christoffer ]
Reviewed-by: NCatalin Marinas <catalin.marinas@arm.com>
Signed-off-by: NJoel Schopp <joel.schopp@amd.com>
Signed-off-by: NChristoffer Dall <christoffer.dall@linaro.org>

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: NChristoffer Dall <christoffer.dall@linaro.org>
Signed-off-by: NMarc Zyngier <marc.zyngier@arm.com>

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: NMarc Zyngier <marc.zyngier@arm.com>

In the beggining was on_each_cpu(), which required an unused argument to
kvm_arch_ops.hardware_{en,dis}able, but this was soon forgotten.

Remove unnecessary arguments that stem from this.
Signed-off-by: NRadim Krčmář <rkrcmar@redhat.com>
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

Using static inline is going to save few bytes and cycles.
For example on powerpc, the difference is 700 B after stripping.
(5 kB before)

This patch also deals with two overlooked empty functions:
kvm_arch_flush_shadow was not removed from arch/mips/kvm/mips.c
  2df72e9b KVM: split kvm_arch_flush_shadow
and kvm_arch_sched_in never made it into arch/ia64/kvm/kvm-ia64.c.
  e790d9ef KVM: add kvm_arch_sched_in
Signed-off-by: NRadim Krčmář <rkrcmar@redhat.com>
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

Running sparse results in a bunch of noisy address space mismatches
thanks to the broken __percpu annotation on kvm_get_running_vcpus.

This function returns a pcpu pointer to a pointer, not a pointer to a
pcpu pointer. This patch fixes the annotation, which kills the warnings
from sparse.

Cc: Christoffer Dall <christoffer.dall@linaro.org>
Cc: Marc Zyngier <marc.zyngier@arm.com>
Signed-off-by: NWill Deacon <will.deacon@arm.com>
Signed-off-by: NChristoffer Dall <christoffer.dall@linaro.org>

When userspace loads code and data in a read-only memory regions, KVM
needs to be able to handle this on arm and arm64.  Specifically this is
used when running code directly from a read-only flash device; the
common scenario is a UEFI blob loaded with the -bios option in QEMU.

Note that the MMIO exit on writes to a read-only memory is ABI and can
be used to emulate block-erase style flash devices.
Acked-by: NMarc Zyngier <marc.zyngier@arm.com>
Signed-off-by: NChristoffer Dall <christoffer.dall@linaro.org>

Introduce preempt notifiers for architecture specific code.
Advantage over creating a new notifier in every arch is slightly simpler
code and guaranteed call order with respect to kvm_sched_in.
Signed-off-by: NRadim Krčmář <rkrcmar@redhat.com>
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

In preparation to make the check_extension function available to VM scope
we add a struct kvm * argument to the function header and rename the function
accordingly. It will still be called from the /dev/kvm fd, but with a NULL
argument for struct kvm *.
Signed-off-by: NAlexander Graf <agraf@suse.de>
Acked-by: NPaolo Bonzini <pbonzini@redhat.com>

Currently when a KVM region is deleted or moved after
KVM_SET_USER_MEMORY_REGION ioctl, the corresponding
intermediate physical memory is not unmapped.

This patch corrects this and unmaps the region's IPA range
in kvm_arch_commit_memory_region using unmap_stage2_range.
Signed-off-by: NEric Auger <eric.auger@linaro.org>
Signed-off-by: NChristoffer Dall <christoffer.dall@linaro.org>

We have PSCI v0.2 emulation available in KVM ARM/ARM64
hence advertise this to user space (i.e. QEMU or KVMTOOL)
via KVM_CHECK_EXTENSION ioctl.
Signed-off-by: NAnup Patel <anup.patel@linaro.org>
Signed-off-by: NPranavkumar Sawargaonkar <pranavkumar@linaro.org>
Acked-by: NChristoffer Dall <christoffer.dall@linaro.org>
Acked-by: NMarc Zyngier <marc.zyngier@arm.com>
Signed-off-by: NChristoffer Dall <christoffer.dall@linaro.org>

On 03/15/2014 12:40 AM, Christoffer Dall wrote:
> On Fri, Mar 14, 2014 at 11:13:29AM +0530, Srivatsa S. Bhat wrote:
>> On 03/13/2014 04:51 AM, Christoffer Dall wrote:
>>> On Tue, Mar 11, 2014 at 02:05:38AM +0530, Srivatsa S. Bhat wrote:
>>>> Subsystems that want to register CPU hotplug callbacks, as well as perform
>>>> initialization for the CPUs that are already online, often do it as shown
>>>> below:
>>>>
[...]
>>> Just so we're clear, the existing code was simply racy as not prone to
>>> deadlocks, right?

> >>> This makes it clear that the test above for compatible CPUs can be quite
> >>> easily evaded by using CPU hotplug, but we don't really have a good
> >>> solution for handling that yet...  Hmmm, grumble grumble, I guess if you
> >>> hotplug unsupported CPUs on a KVM/ARM system for now, stuff will break.

>>
>> In this particular case, there was no deadlock possibility, rather the
>> existing code had insufficient synchronization against CPU hotplug.
>>
>> init_hyp_mode() would invoke cpu_init_hyp_mode() on currently online CPUs
>> using on_each_cpu(). If a CPU came online after this point and before calling
>> register_cpu_notifier(), that CPU would remain uninitialized because this
>> subsystem would miss the hot-online event. This patch fixes this bug and
>> also uses the new synchronization method (instead of get/put_online_cpus())
>> to ensure that we don't deadlock with CPU hotplug.
>>
>
> Yes, that was my conclusion as well.  Thanks for clarifying.  (It could
> be noted in the commit message as well if you should feel so inclined).
>

Please find the patch with updated changelog (and your Ack) below.
(No changes in code).

From: Srivatsa S. Bhat <srivatsa.bhat@linux.vnet.ibm.com>
Subject: [PATCH] arm, kvm: Fix CPU hotplug callback registration

Subsystems that want to register CPU hotplug callbacks, as well as perform
initialization for the CPUs that are already online, often do it as shown
below:

	get_online_cpus();

	for_each_online_cpu(cpu)
		init_cpu(cpu);

	register_cpu_notifier(&foobar_cpu_notifier);

	put_online_cpus();

This is wrong, since it is prone to ABBA deadlocks involving the
cpu_add_remove_lock and the cpu_hotplug.lock (when running concurrently
with CPU hotplug operations).

Instead, the correct and race-free way of performing the callback
registration is:

	cpu_notifier_register_begin();

	for_each_online_cpu(cpu)
		init_cpu(cpu);

	/* Note the use of the double underscored version of the API */
	__register_cpu_notifier(&foobar_cpu_notifier);

	cpu_notifier_register_done();

In the existing arm kvm code, there is no synchronization with CPU hotplug
to avoid missing the hotplug events that might occur after invoking
init_hyp_mode() and before calling register_cpu_notifier(). Fix this bug
and also use the new synchronization method (instead of get/put_online_cpus())
to ensure that we don't deadlock with CPU hotplug.

Cc: Gleb Natapov <gleb@kernel.org>
Cc: Russell King <linux@arm.linux.org.uk>
Cc: Ingo Molnar <mingo@kernel.org>
Acked-by: NPaolo Bonzini <pbonzini@redhat.com>
Acked-by: NChristoffer Dall <christoffer.dall@linaro.org>
Signed-off-by: NSrivatsa S. Bhat <srivatsa.bhat@linux.vnet.ibm.com>
Signed-off-by: NRafael J. Wysocki <rafael.j.wysocki@intel.com>