Re: [PATCH] x86/hyperv: Disable preemption while setting reenlightenment vector

[Vitaly Kuznetsov]

Peter Zijlstra <peterz@xxxxxxxxxxxxx> writes:

> On Tue, Jun 11, 2019 at 10:20:03PM +0100, Dmitry Safonov wrote:
>> KVM support may be compiled as dynamic module, which triggers the
>> following splat on modprobe:
>> 
>>  KVM: vmx: using Hyper-V Enlightened VMCS
>>  BUG: using smp_processor_id() in preemptible [00000000] code: modprobe/466 caller is debug_smp_processor_id+0x17/0x19
>>  CPU: 0 PID: 466 Comm: modprobe Kdump: loaded Not tainted 4.19.43 #1
>>  Hardware name: Microsoft Corporation Virtual Machine/Virtual Machine, BIOS 090007  06/02/2017
>>  Call Trace:
>>   dump_stack+0x61/0x7e
>>   check_preemption_disabled+0xd4/0xe6
>>   debug_smp_processor_id+0x17/0x19
>>   set_hv_tscchange_cb+0x1b/0x89
>>   kvm_arch_init+0x14a/0x163 [kvm]
>>   kvm_init+0x30/0x259 [kvm]
>>   vmx_init+0xed/0x3db [kvm_intel]
>>   do_one_initcall+0x89/0x1bc
>>   do_init_module+0x5f/0x207
>>   load_module+0x1b34/0x209b
>>   __ia32_sys_init_module+0x17/0x19
>>   do_fast_syscall_32+0x121/0x1fa
>>   entry_SYSENTER_compat+0x7f/0x91
>> 
>> The easiest solution seems to be disabling preemption while setting up
>> reenlightment MSRs. While at it, fix hv_cpu_*() callbacks.
>> 
>> Fixes: 93286261de1b4 ("x86/hyperv: Reenlightenment notifications
>> support")
>> 
>> Cc: Andy Lutomirski <luto@xxxxxxxxxx>
>> Cc: Borislav Petkov <bp@xxxxxxxxx>
>> Cc: Cathy Avery <cavery@xxxxxxxxxx>
>> Cc: Haiyang Zhang <haiyangz@xxxxxxxxxxxxx>
>> Cc: "H. Peter Anvin" <hpa@xxxxxxxxx>
>> Cc: Ingo Molnar <mingo@xxxxxxxxxx>
>> Cc: "K. Y. Srinivasan" <kys@xxxxxxxxxxxxx>
>> Cc: "Michael Kelley (EOSG)" <Michael.H.Kelley@xxxxxxxxxxxxx>
>> Cc: Mohammed Gamal <mmorsy@xxxxxxxxxx>
>> Cc: Paolo Bonzini <pbonzini@xxxxxxxxxx>
>> Cc: Radim KrÄmÃÅ <rkrcmar@xxxxxxxxxx>
>> Cc: Roman Kagan <rkagan@xxxxxxxxxxxxx>
>> Cc: Sasha Levin <sashal@xxxxxxxxxx>
>> Cc: Stephen Hemminger <sthemmin@xxxxxxxxxxxxx>
>> Cc: Thomas Gleixner <tglx@xxxxxxxxxxxxx>
>> Cc: Vitaly Kuznetsov <vkuznets@xxxxxxxxxx>
>> 
>> Cc: devel@xxxxxxxxxxxxxxxxxxxxxx
>> Cc: kvm@xxxxxxxxxxxxxxx
>> Cc: linux-hyperv@xxxxxxxxxxxxxxx
>> Cc: x86@xxxxxxxxxx
>> Reported-by: Prasanna Panchamukhi <panchamukhi@xxxxxxxxxx>
>> Signed-off-by: Dmitry Safonov <dima@xxxxxxxxxx>
>> ---
>>  arch/x86/hyperv/hv_init.c | 9 ++++++---
>>  1 file changed, 6 insertions(+), 3 deletions(-)
>> 
>> diff --git a/arch/x86/hyperv/hv_init.c b/arch/x86/hyperv/hv_init.c
>> index 1608050e9df9..0bdd79ecbff8 100644
>> --- a/arch/x86/hyperv/hv_init.c
>> +++ b/arch/x86/hyperv/hv_init.c
>> @@ -91,7 +91,7 @@ EXPORT_SYMBOL_GPL(hv_max_vp_index);
>>  static int hv_cpu_init(unsigned int cpu)
>>  {
>>  	u64 msr_vp_index;
>> -	struct hv_vp_assist_page **hvp = &hv_vp_assist_page[smp_processor_id()];
>> +	struct hv_vp_assist_page **hvp = &hv_vp_assist_page[cpu];
>>  	void **input_arg;
>>  	struct page *pg;
>>  
>> @@ -103,7 +103,7 @@ static int hv_cpu_init(unsigned int cpu)
>>  
>>  	hv_get_vp_index(msr_vp_index);
>>  
>> -	hv_vp_index[smp_processor_id()] = msr_vp_index;
>> +	hv_vp_index[cpu] = msr_vp_index;
>>  
>>  	if (msr_vp_index > hv_max_vp_index)
>>  		hv_max_vp_index = msr_vp_index;
>> @@ -182,7 +182,6 @@ void set_hv_tscchange_cb(void (*cb)(void))
>>  	struct hv_reenlightenment_control re_ctrl = {
>>  		.vector = HYPERV_REENLIGHTENMENT_VECTOR,
>>  		.enabled = 1,
>> -		.target_vp = hv_vp_index[smp_processor_id()]
>>  	};
>>  	struct hv_tsc_emulation_control emu_ctrl = {.enabled = 1};
>>  
>> @@ -196,7 +195,11 @@ void set_hv_tscchange_cb(void (*cb)(void))
>>  	/* Make sure callback is registered before we write to MSRs */
>>  	wmb();
>>  
>> +	preempt_disable();
>> +	re_ctrl.target_vp = hv_vp_index[smp_processor_id()];
>>  	wrmsrl(HV_X64_MSR_REENLIGHTENMENT_CONTROL, *((u64 *)&re_ctrl));
>> +	preempt_enable();
>> +
>>  	wrmsrl(HV_X64_MSR_TSC_EMULATION_CONTROL, *((u64 *)&emu_ctrl));
>>  }
>>  EXPORT_SYMBOL_GPL(set_hv_tscchange_cb);
>
> This looks bogus, MSRs are a per-cpu resource, you had better know what
> CPUs you're on and be stuck to it when you do wrmsr. This just fudges
> the code to make the warning go away and doesn't fix the actual problem
> afaict.

Actually, we don't care which CPU will receive the reenlightenment
notification and TSC Emulation in Hyper-V is, of course, global. We have
code which re-assignes the notification to some other CPU in case the
one it's currently assigned to goes away (see hv_cpu_die()).

-- 
Vitaly