Re: [PATCH v3 0/1] xen: fix HVM kexec kernel panic

From: Dongli Zhang
Date: Tue Mar 01 2022 - 00:00:05 EST


Hi Boris,

On 2/28/22 5:18 PM, Dongli Zhang wrote:
> Hi Boris,
>
> On 2/28/22 12:45 PM, Boris Ostrovsky wrote:
>>
>>
>> On 2/25/22 8:17 PM, Dongli Zhang wrote:
>>> Hi Boris,
>>>
>>> On 2/25/22 2:39 PM, Boris Ostrovsky wrote:
>>>>
>>>> On 2/24/22 4:50 PM, Dongli Zhang wrote:
>>>>> This is the v3 of the patch to fix xen kexec kernel panic issue when the
>>>>> kexec is triggered on VCPU >= 32.
>>>>>
>>>>> PANIC: early exception 0x0e IP 10:ffffffffa96679b6 error 0 cr2 0x20
>>>>> [    0.000000] CPU: 0 PID: 0 Comm: swapper Not tainted
>>>>> 5.17.0-rc4xen-00054-gf71077a4d84b-dirty #1
>>>>> [    0.000000] Hardware name: Xen HVM domU, BIOS 4.4.4OVM 12/15/2020
>>>>> [    0.000000] RIP: 0010:pvclock_clocksource_read+0x6/0xb0
>>>>> ... ...
>>>>> [    0.000000] RSP: 0000:ffffffffaae03e10 EFLAGS: 00010082 ORIG_RAX:
>>>>> 0000000000000000
>>>>> [    0.000000] RAX: 0000000000000000 RBX: 0000000000010000 RCX:
>>>>> 0000000000000002
>>>>> [    0.000000] RDX: 0000000000000003 RSI: ffffffffaac37515 RDI:
>>>>> 0000000000000020
>>>>> [    0.000000] RBP: 0000000000011000 R08: 0000000000000000 R09:
>>>>> 0000000000000001
>>>>> [    0.000000] R10: ffffffffaae03df8 R11: ffffffffaae03c68 R12:
>>>>> 0000000040000004
>>>>> [    0.000000] R13: ffffffffaae03e50 R14: 0000000000000000 R15:
>>>>> 0000000000000000
>>>>> [    0.000000] FS:  0000000000000000(0000) GS:ffffffffab588000(0000)
>>>>> knlGS:0000000000000000
>>>>> [    0.000000] CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
>>>>> [    0.000000] CR2: 0000000000000020 CR3: 00000000ea410000 CR4:
>>>>> 00000000000406a0
>>>>> [    0.000000] DR0: 0000000000000000 DR1: 0000000000000000 DR2:
>>>>> 0000000000000000
>>>>> [    0.000000] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7:
>>>>> 0000000000000400
>>>>> [    0.000000] Call Trace:
>>>>> [    0.000000]  <TASK>
>>>>> [    0.000000]  ? xen_clocksource_read+0x24/0x40
>>>>
>>>>
>>>> This is done to set xen_sched_clock_offset which I think will not be used for a
>>>> while, until sched_clock is called (and the other two uses are for
>>>> suspend/resume)
>>>>
>>>>
>>>> Can we simply defer 'xen_sched_clock_offset = xen_clocksource_read();' until
>>>> after all vcpu areas are properly set? Or are there other uses of
>>>> xen_clocksource_read() before ?
>>>>
>>>
>>> I have tested that below patch will panic kdump kernel.
>>>
>>
>>
>>
>> Oh well, so much for that then. Yes, sched_clock() is at least called from
>> printk path.
>>
>>
>> I guess we will have to go with v2 then, we don't want to start seeing time
>> going back, even if only with older hypervisors. The only thing I might ask is
>> that you roll the logic inside xen_hvm_init_time_ops(). Something like
>>
>>
>> xen_hvm_init_time_ops()
>> {
>>     /*
>>      * Wait until per_cpu(xen_vcpu, 0) is initialized which may happen
>>      * later (e.g. when kdump kernel runs on >=MAX_VIRT_CPUS vcpu)
>>      */
>>     if (__this_cpu_read(xen_vcpu_nr(0)) == NULL)
>>         return;
>>
>
> I think you meant __this_cpu_read(xen_vcpu).
>
> I will call xen_hvm_init_time_ops() at both places, and move the logic into
> xen_hvm_init_time_ops().
>
> Thank you very much!
>
> Dongli Zhang
>


How about we do not move the logic into xen_hvm_init_time_ops()?

Suppose we move the logic into xen_hvm_init_time_ops() line 573, the line line
570 might be printed twice.


559 void __init xen_hvm_init_time_ops(void)
560 {
561 /*
562 * vector callback is needed otherwise we cannot receive interrupts
563 * on cpu > 0 and at this point we don't know how many cpus are
564 * available.
565 */
566 if (!xen_have_vector_callback)
567 return;
568
569 if (!xen_feature(XENFEAT_hvm_safe_pvclock)) {
570 pr_info("Xen doesn't support pvclock on HVM, disable pv timer");
571 return;
572 }
573
574 xen_init_time_common();
575
576 x86_init.timers.setup_percpu_clockev = xen_time_init;
577 x86_cpuinit.setup_percpu_clockev = xen_hvm_setup_cpu_clockevents;
578
579 x86_platform.set_wallclock = xen_set_wallclock;
580 }

I feel the code looks better if we keep the logic at caller side. Would you mind
letting me know your feedback?

Thank you very much!

Dongli Zhang