On Thu, Apr 16, 2020 at 11:29:59PM -0700, Dexuan Cui wrote:
> Unlike the other CPUs, CPU0 is never offlined during hibernation. So in the
> resume path, the "new" kernel's VP assist page is not suspended (i.e.
> disabled), and later when we jump to the "old" kernel, the page is not
> properly re-enabled for CPU0 with the allocated page from the old kernel.
>
> So far, the VP assist page is only used by hv_apic_eoi_write(). When the
> page is not properly re-enabled, hvp->apic_assist is always 0, so the
> HV_X64_MSR_EOI MSR is always written. This is not ideal with respect to
> performance, but Hyper-V can still correctly handle this.
>
> The issue is: the hypervisor can corrupt the old kernel memory, and hence
> sometimes cause unexpected behaviors, e.g. when the old kernel's non-boot
> CPUs are being onlined in the resume path, the VM can hang or be killed
> due to virtual triple fault.
>
> Fix the issue by calling hv_cpu_die()/hv_cpu_init() in the syscore ops.
>
> Without the fix, hibernation can fail at a rate of 1/300 ~ 1/500.
> With the fix, hibernation can pass a long-haul test of 2000 rounds.
>
> Fixes: 05bd330a7fd8 ("x86/hyperv: Suspend/resume the hypercall page for hibernation")
> Cc: stable@xxxxxxxxxxxxxxx
> Signed-off-by: Dexuan Cui <decui@xxxxxxxxxxxxx>
> ---
> arch/x86/hyperv/hv_init.c | 12 ++++++++++--
> 1 file changed, 10 insertions(+), 2 deletions(-)
>
> diff --git a/arch/x86/hyperv/hv_init.c b/arch/x86/hyperv/hv_init.c
> index b0da5320bcff..4d3ce86331a3 100644
> --- a/arch/x86/hyperv/hv_init.c
> +++ b/arch/x86/hyperv/hv_init.c
> @@ -72,7 +72,8 @@ static int hv_cpu_init(unsigned int cpu)
> struct page *pg;
>
> input_arg = (void **)this_cpu_ptr(hyperv_pcpu_input_arg);
> - pg = alloc_page(GFP_KERNEL);
> + /* hv_cpu_init() can be called with IRQs disabled from hv_resume() */
> + pg = alloc_page(GFP_ATOMIC);
IMHO it would be better to only tap into the reserve pool if so
required, e.g.
pg = alloc_page(irqs_disabled() ? GFP_ATOMIC : GFP_KERNEL);
Wei.
