Re: [RFC PATCH] KVM: arm/arm64: vgic: change condition for level interrupt resampling

From: Marc Zyngier
Date: Fri Mar 09 2018 - 04:12:41 EST


On 08/03/18 18:12, Auger Eric wrote:
> Hi Marc, Christoffer,
>
> On 08/03/18 18:28, Marc Zyngier wrote:
>> On Thu, 08 Mar 2018 16:19:00 +0000,
>> Christoffer Dall wrote:
>>>
>>> On Thu, Mar 08, 2018 at 11:54:27AM +0000, Marc Zyngier wrote:
>>>> On 08/03/18 09:49, Marc Zyngier wrote:
>>>>> [updated Christoffer's email address]
>>>>>
>>>>> Hi Shunyong,
>>>>>
>>>>> On 08/03/18 07:01, Shunyong Yang wrote:
>>>>>> When resampling irqfds is enabled, level interrupt should be
>>>>>> de-asserted when resampling happens. On page 4-47 of GIC v3
>>>>>> specification IHI0069D, it said,
>>>>>> "When the PE acknowledges an SGI, a PPI, or an SPI at the CPU
>>>>>> interface, the IRI changes the status of the interrupt to active
>>>>>> and pending if:
>>>>>> â It is an edge-triggered interrupt, and another edge has been
>>>>>> detected since the interrupt was acknowledged.
>>>>>> â It is a level-sensitive interrupt, and the level has not been
>>>>>> deasserted since the interrupt was acknowledged."
>>>>>>
>>>>>> GIC v2 specification IHI0048B.b has similar description on page
>>>>>> 3-42 for state machine transition.
>>>>>>
>>>>>> When some VFIO device, like mtty(8250 VFIO mdev emulation driver
>>>>>> in samples/vfio-mdev) triggers a level interrupt, the status
>>>>>> transition in LR is pending-->active-->active and pending.
>>>>>> Then it will wait resampling to de-assert the interrupt.
>>>>>>
>>>>>> Current design of lr_signals_eoi_mi() will return false if state
>>>>>> in LR is not invalid(Inactive). It causes resampling will not happen
>>>>>> in mtty case.
>>>>>
>>>>> Let me rephrase this, and tell me if I understood it correctly:
>>>>>
>>>>> - A level interrupt is injected, activated by the guest (LR state=active)
>>>>> - guest exits, re-enters, (LR state=pending+active)
>>>>> - guest EOIs the interrupt (LR state=pending)
>>>>> - maintenance interrupt
>>>>> - we don't signal the resampling because we're not in an invalid state
>>>>>
>>>>> Is that correct?
>>>>>
>>>>> That's an interesting case, because it seems to invalidate some of the
>>>>> optimization that went in over a year ago.
>>>>>
>>>>> 096f31c4360f KVM: arm/arm64: vgic: Get rid of MISR and EISR fields
>>>>> b6095b084d87 KVM: arm/arm64: vgic: Get rid of unnecessary save_maint_int_state
>>>>> af0614991ab6 KVM: arm/arm64: vgic: Get rid of unnecessary process_maintenance operation
>>>>>
>>>>> We could compare the value of the LR before the guest entry with
>>>>> the value at exit time, but we still could miss it if we have a
>>>>> transition such as P+A -> P -> A and assume a long enough propagation
>>>>> delay for the maintenance interrupt (which is very likely).
>>>>>
>>>>> In essence, we have lost the benefit of EISR, which was to give us a
>>>>> way to deal with asynchronous signalling.
>>>>>
>>>>>>
>>>>>> This will cause interrupt fired continuously to guest even 8250 IIR
>>>>>> has no interrupt. When 8250's interrupt is configured in shared mode,
>>>>>> it will pass interrupt to other drivers to handle. However, there
>>>>>> is no other driver involved. Then, a "nobody cared" kernel complaint
>>>>>> occurs.
>>>>>>
>>>>>> / # cat /dev/ttyS0
>>>>>> [ 4.826836] random: crng init done
>>>>>> [ 6.373620] irq 41: nobody cared (try booting with the "irqpoll"
>>>>>> option)
>>>>>> [ 6.376414] CPU: 0 PID: 1307 Comm: cat Not tainted 4.16.0-rc4 #4
>>>>>> [ 6.378927] Hardware name: linux,dummy-virt (DT)
>>>>>> [ 6.380876] Call trace:
>>>>>> [ 6.381937] dump_backtrace+0x0/0x180
>>>>>> [ 6.383495] show_stack+0x14/0x1c
>>>>>> [ 6.384902] dump_stack+0x90/0xb4
>>>>>> [ 6.386312] __report_bad_irq+0x38/0xe0
>>>>>> [ 6.387944] note_interrupt+0x1f4/0x2b8
>>>>>> [ 6.389568] handle_irq_event_percpu+0x54/0x7c
>>>>>> [ 6.391433] handle_irq_event+0x44/0x74
>>>>>> [ 6.393056] handle_fasteoi_irq+0x9c/0x154
>>>>>> [ 6.394784] generic_handle_irq+0x24/0x38
>>>>>> [ 6.396483] __handle_domain_irq+0x60/0xb4
>>>>>> [ 6.398207] gic_handle_irq+0x98/0x1b0
>>>>>> [ 6.399796] el1_irq+0xb0/0x128
>>>>>> [ 6.401138] _raw_spin_unlock_irqrestore+0x18/0x40
>>>>>> [ 6.403149] __setup_irq+0x41c/0x678
>>>>>> [ 6.404669] request_threaded_irq+0xe0/0x190
>>>>>> [ 6.406474] univ8250_setup_irq+0x208/0x234
>>>>>> [ 6.408250] serial8250_do_startup+0x1b4/0x754
>>>>>> [ 6.410123] serial8250_startup+0x20/0x28
>>>>>> [ 6.411826] uart_startup.part.21+0x78/0x144
>>>>>> [ 6.413633] uart_port_activate+0x50/0x68
>>>>>> [ 6.415328] tty_port_open+0x84/0xd4
>>>>>> [ 6.416851] uart_open+0x34/0x44
>>>>>> [ 6.418229] tty_open+0xec/0x3c8
>>>>>> [ 6.419610] chrdev_open+0xb0/0x198
>>>>>> [ 6.421093] do_dentry_open+0x200/0x310
>>>>>> [ 6.422714] vfs_open+0x54/0x84
>>>>>> [ 6.424054] path_openat+0x2dc/0xf04
>>>>>> [ 6.425569] do_filp_open+0x68/0xd8
>>>>>> [ 6.427044] do_sys_open+0x16c/0x224
>>>>>> [ 6.428563] SyS_openat+0x10/0x18
>>>>>> [ 6.429972] el0_svc_naked+0x30/0x34
>>>>>> [ 6.431494] handlers:
>>>>>> [ 6.432479] [<000000000e9fb4bb>] serial8250_interrupt
>>>>>> [ 6.434597] Disabling IRQ #41
>>>>>>
>>>>>> This patch changes the lr state condition in lr_signals_eoi_mi() from
>>>>>> invalid(Inactive) to active and pending to avoid this.
>>>>>>
>>>>>> I am not sure about the original design of the condition of
>>>>>> invalid(active). So, This RFC is sent out for comments.
>>>>>>
>>>>>> Cc: Joey Zheng <yu.zheng@xxxxxxxxxxxxxxxx>
>>>>>> Signed-off-by: Shunyong Yang <shunyong.yang@xxxxxxxxxxxxxxxx>
>>>>>> ---
>>>>>> virt/kvm/arm/vgic/vgic-v2.c | 4 ++--
>>>>>> virt/kvm/arm/vgic/vgic-v3.c | 4 ++--
>>>>>> 2 files changed, 4 insertions(+), 4 deletions(-)
>>>>>>
>>>>>> diff --git a/virt/kvm/arm/vgic/vgic-v2.c b/virt/kvm/arm/vgic/vgic-v2.c
>>>>>> index e9d840a75e7b..740ee9a5f551 100644
>>>>>> --- a/virt/kvm/arm/vgic/vgic-v2.c
>>>>>> +++ b/virt/kvm/arm/vgic/vgic-v2.c
>>>>>> @@ -46,8 +46,8 @@ void vgic_v2_set_underflow(struct kvm_vcpu *vcpu)
>>>>>>
>>>>>> static bool lr_signals_eoi_mi(u32 lr_val)
>>>>>> {
>>>>>> - return !(lr_val & GICH_LR_STATE) && (lr_val & GICH_LR_EOI) &&
>>>>>> - !(lr_val & GICH_LR_HW);
>>>>>> + return !((lr_val & GICH_LR_STATE) ^ GICH_LR_STATE) &&
>>>>>
>>>>> That feels very wrong. You're now signalling the resampling in both
>>>>> invalid and pending+active, and the latter state doesn't mean you've
>>>>> EOIed anything. You're now over-signalling, and signalling the
>>>>> wrong event.
>>>>>
>>>>>> + (lr_val & GICH_LR_EOI) && !(lr_val & GICH_LR_HW);
>>>>>> }
>>>>>>
>>>>>> /*
>>>>>> diff --git a/virt/kvm/arm/vgic/vgic-v3.c b/virt/kvm/arm/vgic/vgic-v3.c
>>>>>> index 6b329414e57a..43111bba7af9 100644
>>>>>> --- a/virt/kvm/arm/vgic/vgic-v3.c
>>>>>> +++ b/virt/kvm/arm/vgic/vgic-v3.c
>>>>>> @@ -35,8 +35,8 @@ void vgic_v3_set_underflow(struct kvm_vcpu *vcpu)
>>>>>>
>>>>>> static bool lr_signals_eoi_mi(u64 lr_val)
>>>>>> {
>>>>>> - return !(lr_val & ICH_LR_STATE) && (lr_val & ICH_LR_EOI) &&
>>>>>> - !(lr_val & ICH_LR_HW);
>>>>>> + return !((lr_val & ICH_LR_STATE) ^ ICH_LR_STATE) &&
>>>>>> + (lr_val & ICH_LR_EOI) && !(lr_val & ICH_LR_HW);
>>>>>> }
>>>>>>
>>>>>> void vgic_v3_fold_lr_state(struct kvm_vcpu *vcpu)
>>>>>>
>>>>>
>>>>> Assuming I understand the issue correctly, I cannot really see how
>>>>> to solve this without reintroducing EISR, which sucks majorly.
>>>>>
>>>>> I'll try to cook something shortly and we can all have a good
>>>>> fight about how crap this is.
>>>>
>>>> Here's what I came up with. I don't really like it, but that's
>>>> the least invasive this I could come up with. Please let me
>>>> know if that helps with your test case. Note that I have only
>>>> boot-tested this on a sample of 1 machine, so I don't expect this
>>>> to be perfect.
>>>>
>>>> Also, any guideline on how to reproduce this would be much appreciated.
>>>> I never used this mdev/mtty thing, so please bear with me.
>>>>
>>>> Thanks,
>>>>
>>>> M.
>>>>
>>>> From 66a7c4cfc1029b0169dd771e196e2876ba3f17b1 Mon Sep 17 00:00:00 2001
>>>> From: Marc Zyngier <marc.zyngier@xxxxxxx>
>>>> Date: Thu, 8 Mar 2018 11:14:06 +0000
>>>> Subject: [PATCH] KVM: arm/arm64: Do not rely on LR state to guess EOI MI
>>>> status
>>>>
>>>> We so far rely on the LR state to decide whether the guest has
>>>> EOI'd a level interrupt or not. While this looks like a good
>>>> idea on the surface, it leads to a couple of annoying corner
>>>> cases:
>>>>
>>>> Example 1: (P = Pending, A = Active, MI = Maintenance Interrupt)
>>>> P -> guest IAR -> A -> exit/entry -> P+A -> guest EOI -> P -> MI
>>>
>>> Do we really get an EOI maintenance interrupt here? Reading the MISR
>>> and EISR descriptions make me thing this is not the case...
>
> Hum yes in EISR it is said that ICH_LR.State = 0b00!
>>
>> Yeah, it looks like I always want EISR to do what I want, and not to
>> do what it does. Man, this thing is such a piece of crap.
>>
>> OK, scratch that. We need to do it without the help of the HW.
>>
>>>> The state is now pending, we've really EOI'd the interrupt, and
>>>> yet lr_signals_eoi_mi() returns false, since the state is not 0.
>>>> The result is that we won't signal anything on the corresponding
>>>> irqfd, which people complain about. Meh.
>>>
>>> So the core of the problem is that when we've entered the guest with
>>> PENDING+ACTIVE and when we exit (for some reason) we don't signal the
>>> resamplefd, right? The solution seems to me that we don't ever do
>>> PENDING+ACTIVE if you need to resample after each deactivate. What
>>> would be the point of appending a pending state that you only know to be
>>> valid after a resample anyway?
>>
>> The question is then to identify that a given source needs to be
>> signalled back to VFIO. Calling into the eventfd code on the hot path
>> is pretty horrid (I'm not sure if we can really call into this with
>> interrupts disabled, for example).
>>
>>>
>>>>
>>>> Example 2:
>>>> P+A -> guest EOI -> P -> delayed MI -> guest IAR -> A -> MI fires
>>>
>>> We could be more clever and do the following calculation on every exit:
>>>
>>> If you enter with P, and exit with either A or 0, then signal.
>>>
>>> If you enter with P+A, and you exit with either P, A, or 0, then signal.
>>>
>>> Wouldn't that also solve it? (Although I have a feeling you'd miss some
>>> exits in this case).
>>
>> I'd be more confident if we did forbid P+A for such interrupts
>> altogether, as they really feel like another kind of HW interrupt.
>
> the LR P+A looks strange to me too. all the more so it may cause the
> same IRQ to be acked twice?

If the pending bit isn't dropped by the time we get to EOI the first
one, probably. But that's pretty much expected with a level interrupt
isn't it?

> P -> A -> 0 (resample). Doesn't our issue come from the fact we reinject
> the P in LR until the line level is deasserted?

Which is consistent with the life cycle of a level interrupt. What
usually happens is (for a non HW interrupt):

P -> IAR -> A -> lower the line in the device -> 0

If you generate an exit at the right spot, and yet don't lower the line,
you end up with:

P -> IAR -> A -> exit/enter -> P+A

>From there, if you lower the line, it is likely to cause an exit:

P+A -> MMIO trap lowering the line -> A

>>
>> Eric: Is there any way to get a callback from the eventfd code to flag
>> a given irq as requiring a notification on EOI?
>
> bool kvm_irq_has_notifier(struct kvm *kvm, unsigned irqchip, unsigned
> pin) was used in the past. I think it does what you want.
>

Not exactly. I'm very reluctant to call this on the hot path (I'd need
the info on hw_flush), and I'd rather have a callback from the eventfd
subsystem to tell me when a pin is being associated with a notifier
(because this is likely to be very rare).

If that doesn't exit, never mind. We can see if that solves Shunyong
issue and optimize later.

M.
--
Jazz is not dead. It just smells funny...