[PATCH v3 1/1] xen: delay xen_hvm_init_time_ops() to xen_hvm_smp_prepare_boot_cpu()

[Dongli Zhang <dongli.zhang@xxxxxxxxxx>]

The sched_clock() can be used very early since commit 857baa87b642
("sched/clock: Enable sched clock early"). In addition, with commit
38669ba205d1 ("x86/xen/time: Output xen sched_clock time from 0"), kdump
kernel in Xen HVM guest may panic at very early stage when accessing
&__this_cpu_read(xen_vcpu)->time as in below:

setup_arch()
 -> init_hypervisor_platform()
     -> x86_init.hyper.init_platform = xen_hvm_guest_init()
         -> xen_hvm_init_time_ops()
             -> xen_clocksource_read()
                 -> src = &__this_cpu_read(xen_vcpu)->time;

This is because Xen HVM supports at most MAX_VIRT_CPUS=32 'vcpu_info'
embedded inside 'shared_info' during early stage until xen_vcpu_setup() is
used to allocate/relocate 'vcpu_info' for boot cpu at arbitrary address.

However, when Xen HVM guest panic on vcpu >= 32, since
xen_vcpu_info_reset(0) would set per_cpu(xen_vcpu, cpu) = NULL when
vcpu >= 32, xen_clocksource_read() on vcpu >= 32 would panic.

This patch always delays xen_hvm_init_time_ops() to later in
xen_hvm_smp_prepare_boot_cpu() after the 'vcpu_info' for boot vcpu is
registered.

This issue can be reproduced on purpose via below command at the guest
side when kdump/kexec is enabled:

"taskset -c 33 echo c > /proc/sysrq-trigger"

Unfortunately, the 'soft_reset' (kexec) does not work with mainline xen
version so that I can test this patch only with HVM guest on old xen
hypervisor where 'soft_reset' is working. The bugfix for PVM is not
implemented due to the lack of testing environment.

Cc: Joe Jin <joe.jin@xxxxxxxxxx>
Signed-off-by: Dongli Zhang <dongli.zhang@xxxxxxxxxx>
---
Changed since v1:
  - Add commit message to explain why xen_hvm_init_time_ops() is delayed
    for any vcpus. (Suggested by Boris Ostrovsky)
  - Add a comment in xen_hvm_smp_prepare_boot_cpu() referencing the related
    code in xen_hvm_guest_init(). (suggested by Juergen Gross)
Changed since v2:
  - Delay for all VCPUs. (Suggested by Boris Ostrovsky)
  - Add commit message that why PVM is not supported by this patch
  - Test if kexec/kdump works with mainline xen (HVM and PVM)

 arch/x86/xen/enlighten_hvm.c |  1 -
 arch/x86/xen/smp_hvm.c       | 17 +++++++++++++++++
 2 files changed, 17 insertions(+), 1 deletion(-)

diff --git a/arch/x86/xen/enlighten_hvm.c b/arch/x86/xen/enlighten_hvm.c
index 517a9d8d8f94..53f306ec1d3b 100644
--- a/arch/x86/xen/enlighten_hvm.c
+++ b/arch/x86/xen/enlighten_hvm.c
@@ -216,7 +216,6 @@ static void __init xen_hvm_guest_init(void)
        WARN_ON(xen_cpuhp_setup(xen_cpu_up_prepare_hvm, xen_cpu_dead_hvm));
        xen_unplug_emulated_devices();
        x86_init.irqs.intr_init = xen_init_IRQ;
-       xen_hvm_init_time_ops();
        xen_hvm_init_mmu_ops();
 
 #ifdef CONFIG_KEXEC_CORE
diff --git a/arch/x86/xen/smp_hvm.c b/arch/x86/xen/smp_hvm.c
index 6ff3c887e0b9..9a5efc1a1633 100644
--- a/arch/x86/xen/smp_hvm.c
+++ b/arch/x86/xen/smp_hvm.c
@@ -19,6 +19,23 @@ static void __init xen_hvm_smp_prepare_boot_cpu(void)
         */
        xen_vcpu_setup(0);
 
+       /*
+        * xen_hvm_init_time_ops() used to be called at very early stage
+        * by xen_hvm_guest_init(). While only MAX_VIRT_CPUS 'vcpu_info'
+        * are embedded inside 'shared_info', the VM would use them until
+        * xen_vcpu_setup() is used to allocate/relocate them at arbitrary
+        * address.
+        *
+        * However, when Xen HVM guest boots on vcpu >= MAX_VIRT_CPUS
+        * (e.g., kexec kernel), per_cpu(xen_vcpu, cpu) is NULL at early
+        * stage. To access per_cpu(xen_vcpu, cpu) via
+        * xen_clocksource_read() would panic the kernel.
+        *
+        * Therefore we always delay xen_hvm_init_time_ops() to
+        * xen_hvm_smp_prepare_boot_cpu() to avoid the panic.
+        */
+       xen_hvm_init_time_ops();
+
        /*
         * The alternative logic (which patches the unlock/lock) runs before
         * the smp bootup up code is activated. Hence we need to set this up
-- 
2.17.1