lkml.org 
[lkml]   [2018]   [Mar]   [8]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
SubjectRe: [RFC PATCH] KVM: arm/arm64: vgic: change condition for level interrupt resampling
From
Date
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:
>>>>> • It is an edge-triggered interrupt, and another edge has been
>>>>> detected since the interrupt was acknowledged.
>>>>> • 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@hxt-semitech.com>
>>>>> Signed-off-by: Shunyong Yang <shunyong.yang@hxt-semitech.com>
>>>>> ---
>>>>> 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@arm.com>
>>> 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?

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

Thanks

Eric
>
> Thanks,
>
> M.
>

\
 
 \ /
  Last update: 2018-03-08 19:13    [W:0.105 / U:3.068 seconds]
©2003-2018 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site