Re: [PATCH 0/3] KVM: Make kvm_lock non-raw

[Jan Kiszka <jan.kiszka@siemens.com>]

On 2013-09-20 20:18, Paul Gortmaker wrote:
> On 13-09-20 02:04 PM, Jan Kiszka wrote:
>> On 2013-09-20 19:51, Paul Gortmaker wrote:
>>> [Re: [PATCH 0/3] KVM: Make kvm_lock non-raw] On 16/09/2013 (Mon 18:12) Paul Gortmaker wrote:
>>>
>>>> On 13-09-16 10:06 AM, Paolo Bonzini wrote:
>>>>> Paul Gortmaker reported a BUG on preempt-rt kernels, due to taking the
>>>>> mmu_lock within the raw kvm_lock in mmu_shrink_scan.  He provided a
>>>>> patch that shrunk the kvm_lock critical section so that the mmu_lock
>>>>> critical section does not nest with it, but in the end there is no reason
>>>>> for the vm_list to be protected by a raw spinlock.  Only manipulations
>>>>> of kvm_usage_count and the consequent hardware_enable/disable operations
>>>>> are not preemptable.
>>>>>
>>>>> This small series thus splits the kvm_lock in the "raw" part and the
>>>>> "non-raw" part.
>>>>>
>>>>> Paul, could you please provide your Tested-by?
>>>>
>>>> Sure, I'll go back and see if I can find what triggered it in the
>>>> original report, and give the patches a spin on 3.4.x-rt (and probably
>>>> 3.10.x-rt, since that is where rt-current is presently).
>>>
>>> Seems fine on 3.4-rt.  On 3.10.10-rt7 it looks like there are other
>>> issues, probably not explicitly related to this patchset (see below).
>>>
>>> Paul.
>>> --
>>>
>>> e1000e 0000:00:19.0 eth1: removed PHC
>>> assign device 0:0:19.0
>>> pci 0000:00:19.0: irq 43 for MSI/MSI-X
>>> pci 0000:00:19.0: irq 43 for MSI/MSI-X
>>> pci 0000:00:19.0: irq 43 for MSI/MSI-X
>>> pci 0000:00:19.0: irq 43 for MSI/MSI-X
>>> BUG: sleeping function called from invalid context at /home/paul/git/linux-rt/kernel/rtmutex.c:659
>>> in_atomic(): 1, irqs_disabled(): 1, pid: 0, name: swapper/0
>>> 2 locks held by swapper/0/0:
>>>  #0:  (rcu_read_lock){.+.+.+}, at: [<ffffffff8100998a>] kvm_set_irq_inatomic+0x2a/0x4a0
>>>  #1:  (rcu_read_lock){.+.+.+}, at: [<ffffffff81038800>] kvm_irq_delivery_to_apic_fast+0x60/0x3d0
>>> irq event stamp: 6121390
>>> hardirqs last  enabled at (6121389): [<ffffffff819f9ae0>] restore_args+0x0/0x30
>>> hardirqs last disabled at (6121390): [<ffffffff819f9a2a>] common_interrupt+0x6a/0x6f
>>> softirqs last  enabled at (0): [<          (null)>]           (null)
>>> softirqs last disabled at (0): [<          (null)>]           (null)
>>> Preemption disabled at:[<ffffffff810ebb9a>] cpu_startup_entry+0x1ba/0x430
>>>
>>> CPU: 0 PID: 0 Comm: swapper/0 Not tainted 3.10.10-rt7 #2
>>> Hardware name: Dell Inc. OptiPlex 990/0VNP2H, BIOS A17 03/14/2013
>>>  ffffffff8201c440 ffff880223603cf0 ffffffff819f177d ffff880223603d18
>>>  ffffffff810c90d3 ffff880214a50110 0000000000000001 0000000000000001
>>>  ffff880223603d38 ffffffff819f89a4 ffff880214a50110 ffff880214a50110
>>> Call Trace:
>>>  <IRQ>  [<ffffffff819f177d>] dump_stack+0x19/0x1b
>>>  [<ffffffff810c90d3>] __might_sleep+0x153/0x250
>>>  [<ffffffff819f89a4>] rt_spin_lock+0x24/0x60
>>>  [<ffffffff810ccdd6>] __wake_up+0x36/0x70
>>>  [<ffffffff81003bbb>] kvm_vcpu_kick+0x3b/0xd0
>>
>> -rt lacks an atomic waitqueue for triggering VCPU wakeups on MSIs from
>> assigned devices directly from the host IRQ handler. We need to disable
>> this fast-path in -rt or introduce such an abstraction (I did this once
>> over 2.6.33-rt).
> 
> Ah, right -- the simple wait queue support (currently -rt specific)
> would have to be used here.  It is on the todo list to get that moved
> from -rt into mainline.

Oh, it's there in -rt already - perfect! If there is a good reason for
upstream, kvm can switch of course.

Jan

-- 
Siemens AG, Corporate Technology, CT RTC ITP SES-DE
Corporate Competence Center Embedded Linux