On 31.01.18 05:23, wei.guo.simon@xxxxxxxxx wrote:
> From: Simon Guo <wei.guo.simon@xxxxxxxxx>
>
> commit 40fdd8c88c4a ("KVM: PPC: Book3S: PR: Make svcpu -> vcpu store
> preempt savvy") and commit 3d3319b45eea ("KVM: PPC: Book3S: PR: Enable
> interrupts earlier") is trying to turns on preemption early when
> return into highmem guest exit handler.
>
> However there is a race window in following example at
> arch/powerpc/kvm/book3s_interrupts.S:
>
> highmem guest exit handler:
> ...
> 195 GET_SHADOW_VCPU(r4)
> 196 bl FUNC(kvmppc_copy_from_svcpu)
> ...
> 239 bl FUNC(kvmppc_handle_exit_pr)
>
> If there comes a preemption between line 195 and 196, line 196
> may hold an invalid SVCPU reference with following sequence:
> 1) Qemu task T1 runs at GET_SHADOW_VCPU(r4) at line 195, on CPU A.
> 2) T1 is preempted and switch out CPU A. As a result, it checks
> CPU A's svcpu->in_use (=1 at present) and flush cpu A's svcpu to
> T1's vcpu.
> 3) Another task T2 switches into CPU A and it may update CPU A's
> svcpu->in_use into 1.
> 4) T1 is scheduled into CPU B. But it still holds CPU A's svcpu
> reference as R4. Then it executes kvmppc_copy_from_svcpu() with
> R4 and it will corrupt T1's VCPU with T2's content. T2's VCPU
> will also be impacted.
>
> This patch moves the svcpu->in_use into VCPU so that the vcpus
> sharing the same svcpu can work properly and fix the above case.
>
> Signed-off-by: Simon Guo <wei.guo.simon@xxxxxxxxx>
Sorry, the previous version would only compile on 32bit PPC ;). Please
find the fixed one which just uses svcpu_get() and _put() here:
https://github.com/agraf/linux-2.6/commit/f9e3ca44c9a9d4930d6dccaacb518734746059c3
Alex
