Re: [PATCH V2 0/3] improve -overcommit cpu-pm=on|off

[Chen, Zide]

On 5/30/2024 6:54 AM, Zhao Liu wrote:
> Hi Zide,
> 
> On Wed, May 29, 2024 at 10:31:21AM -0700, Chen, Zide wrote:
>> Date: Wed, 29 May 2024 10:31:21 -0700
>> From: "Chen, Zide" <zide.chen@intel.com>
>> Subject: Re: [PATCH V2 0/3] improve -overcommit cpu-pm=on|off
>>
>>
>>
>> On 5/29/2024 5:46 AM, Igor Mammedov wrote:
>>> On Tue, 28 May 2024 11:16:59 -0700
>>> "Chen, Zide" <zide.chen@intel.com> wrote:
>>>
>>>> On 5/28/2024 2:23 AM, Igor Mammedov wrote:
>>>>> On Fri, 24 May 2024 13:00:14 -0700
>>>>> Zide Chen <zide.chen@intel.com> wrote:
>>>>>   
>>>>>> Currently, if running "-overcommit cpu-pm=on" on hosts that don't
>>>>>> have MWAIT support, the MWAIT/MONITOR feature is advertised to the
>>>>>> guest and executing MWAIT/MONITOR on the guest triggers #UD.  
>>>>>
>>>>> this is missing proper description how do you trigger issue
>>>>> with reproducer and detailed description why guest sees MWAIT
>>>>> when it's not supported by host.  
>>>>
>>>> If "overcommit cpu-pm=on" and "-cpu host" are present, as shown in the
>>> it's bette to provide full QEMU CLI and host/guest kernels used and what
>>> hardware was used if it's relevant so others can reproduce problem.
>>
>> I ever reproduced this on an older Intel Icelake machine, a
>> Sapphire Rapids and a Sierra Forest, but I believe this is a x86 generic
>> issue, not specific to particular models.
>>
>> For the CLI, I think the only command line options that matter are
>>  -overcommit cpu-pm=on: to set enable_cpu_pm
>>  -cpu host: so that cpu->max_features is set
>>
>> For QEMU version, as long as it's after this commit: 662175b91ff2
>> ("i386: reorder call to cpu_exec_realizefn")
>>
>> The guest fails to boot:
>>
>> [ 24.825568] smpboot: x86: Booting SMP configuration:
>> [ 24.826377] .... node #0, CPUs: #1 #2 #3 #4 #5 #6 #7 #8 #9 #10 #11 #12
>> #13 #14 #15 #17
>> [ 24.985799] .... node #1, CPUs: #128 #129 #130 #131 #132 #133 #134 #135
>> #136 #137 #138 #139 #140 #141 #142 #143 #145
>> [ 25.136955] invalid opcode: 0000 1 PREEMPT SMP NOPTI
>> [ 25.137790] CPU: 0 PID: 0 Comm: swapper/0 Not tainted 6.8.0 #2
>> [ 25.137790] Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS
>> rel-1.16.3-0-ga6ed6b701f0a-prebuilt.qemu.org 04/04
>> [ 25.137790] RIP: 0010:mwait_idle+0x35/0x80
>> [ 25.137790] Code: 6f f0 80 48 02 20 48 8b 10 83 e2 08 75 3e 65 48 8b 15
>> 47 d6 56 6f 48 0f ba e2 27 72 41 31 d2 48 89 d8
>> [ 25.137790] RSP: 0000:ffffffff91403e70 EFLAGS: 00010046
>> [ 25.137790] RAX: ffffffff9140a980 RBX: ffffffff9140a980 RCX:
>> 0000000000000000
>> [ 25.137790] RDX: 0000000000000000 RSI: ffff97f1ade21b20 RDI:
>> 0000000000000004
>> [ 25.137790] RBP: 0000000000000000 R08: 00000005da4709cb R09:
>> 0000000000000001
>> [ 25.137790] R10: 0000000000005da4 R11: 0000000000000009 R12:
>> 0000000000000000
>> [ 25.137790] R13: ffff98573ff90fc0 R14: ffffffff9140a038 R15:
>> 0000000000093ff0
>> [ 25.137790] FS: 0000000000000000(0000) GS:ffff97f1ade00000(0000)
>> knlGS:0000000000000000
>> [ 25.137790] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
>> [ 25.137790] CR2: ffff97d8aa801000 CR3: 00000049e9430001 CR4:
>> 0000000000770ef0
>> [ 25.137790] DR0: 0000000000000000 DR1: 0000000000000000 DR2:
>> 0000000000000000
>> [ 25.137790] DR3: 0000000000000000 DR6: 00000000ffff07f0 DR7:
>> 0000000000000400
>> [ 25.137790] PKRU: 55555554
>> [ 25.137790] Call Trace:
>> [ 25.137790] <TASK>
>> [ 25.137790] ? die+0x37/0x90
>> [ 25.137790] ? do_trap+0xe3/0x110
>> [ 25.137790] ? mwait_idle+0x35/0x80
>> [ 25.137790] ? do_error_trap+0x6a/0x90
>> [ 25.137790] ? mwait_idle+0x35/0x80
>> [ 25.137790] ? exc_invalid_op+0x52/0x70
>> [ 25.137790] ? mwait_idle+0x35/0x80
>> [ 25.137790] ? asm_exc_invalid_op+0x1a/0x20
>> [ 25.137790] ? mwait_idle+0x35/0x80
>> [ 25.137790] default_idle_call+0x30/0x100
>> [ 25.137790] cpuidle_idle_call+0x12c/0x170
>> [ 25.137790] ? tsc_verify_tsc_adjust+0x73/0xd0
>> [ 25.137790] do_idle+0x7f/0xd0
>> [ 25.137790] cpu_startup_entry+0x29/0x30
>> [ 25.137790] rest_init+0xcc/0xd0
>> [ 25.137790] start_kernel+0x396/0x5d0
>> [ 25.137790] x86_64_start_reservations+0x18/0x30
>> [ 25.137790] x86_64_start_kernel+0xe7/0xf0
>> [ 25.137790] common_startup_64+0x13e/0x148
>> [ 25.137790] </TASK>
>> [ 25.137790] Modules linked in:
>> [ 25.137790] --[ end trace 0000000000000000 ]--
>> [ 25.137790] invalid opcode: 0000 2 PREEMPT SMP NOPTI
>> [ 25.137790] RIP: 0010:mwait_idle+0x35/0x80
>> [ 25.137790] Code: 6f f0 80 48 02 20 48 8b 10 83 e2 08 75 3e 65 48 8b 15
>> 47 d6 56 6f 48 0f ba e2 27 72 41 31 d2 48 89 d8
>>
>>>
>>>> following, CPUID_EXT_MONITOR is set after x86_cpu_filter_features(), so
>>>> that it doesn't have a chance to check MWAIT against host features and
>>>> will be advertised to the guest regardless of whether it's supported by
>>>> the host or not.
>>>>
>>>> x86_cpu_realizefn()
>>>>   x86_cpu_filter_features()
>>>>   cpu_exec_realizefn()
>>>>     kvm_cpu_realizefn
>>>>       host_cpu_realizefn
>>>>         host_cpu_enable_cpu_pm
>>>>           env->features[FEAT_1_ECX] |= CPUID_EXT_MONITOR;
>>>>
>>>>
>>>> If it's not supported by the host, executing MONITOR or MWAIT
>>>> instructions from the guest triggers #UD, no matter MWAIT_EXITING
>>>> control is set or not.
>>>
>>> If I recall right, kvm was able to emulate mwait/monitor.
>>> So question is why it leads to exception instead?
>>
>> KVM can come to play only iff it can trigger MWAIT/MONITOR VM exits. I
>> didn't find explicit proof from Intel SDM that #UD exceptions take
>> precedence over MWAIT/MONITOR VM exits, but this is my speculation. For
>> example, in ancient machines which don't support MWAIT yet, the only way
>> it can do is #UD, not MWAIT VM exit?
> 
> For the Host which doesn't support MWAIT, it shouldn't have the VMX
> control bit for mwait exit either, right?
> 
> Could you pls check this on your machine? If VMX doesn't support this
> exit event, then triggering an exception will make sense.


As Sean just confirmed, #UD takes priority over MWAIT exiting VM-exit,
thus if the host doesn't support MWAIT, regardless the MWAIT exiting is
set or not, executing MWAIT instruction from the guest triggers #UD, and
the guest doesn't boot.

This is not desired and VMM should not advertise MWAIT to the guest in
this case.