DMARC-Filter: OpenDMARC Filter v1.3.2 mx1.redhat.com 0089E40431F
Subject: Re: [PATCH v2 5/8] KVM: arm/arm64: vgic: Handle mapped level
 sensitive SPIs
To: Marc Zyngier <marc.zyngier@arm.com>,
        Christoffer Dall <christoffer.dall@linaro.org>
References: <1497531160-29162-1-git-send-email-eric.auger@redhat.com>
 <1497531160-29162-6-git-send-email-eric.auger@redhat.com>
 <f21e17b7-9437-ca96-6847-91e7ecee3af4@arm.com>
 <6a177b08-1871-c5b1-6ece-404017dfd23c@redhat.com>
 <20170721130331.GG16350@cbox> <b396c691-13d6-d17f-570e-7787970e759f@arm.com>
 <20170725144835.GD1588@lvm> <c2b481d0-3094-8a19-6257-da4927af014e@arm.com>
Cc: Christoffer Dall <cdall@linaro.org>, eric.auger.pro@gmail.com,
        linux-kernel@vger.kernel.org, kvm@vger.kernel.org,
        kvmarm@lists.cs.columbia.edu, alex.williamson@redhat.com,
        b.reynal@virtualopensystems.com, pbonzini@redhat.com,
        drjones@redhat.com, wei@redhat.com
From: Auger Eric <eric.auger@redhat.com>
Message-ID: <ea3d3821-a3ec-dae4-9633-0cbf65afe057@redhat.com>
Date: Tue, 22 Aug 2017 16:35:26 +0200
User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:45.0) Gecko/20100101
 Thunderbird/45.4.0
MIME-Version: 1.0
In-Reply-To: <c2b481d0-3094-8a19-6257-da4927af014e@arm.com>
Content-Type: text/plain; charset=windows-1252
Content-Transfer-Encoding: 7bit
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 7996
Lines: 192

Hi,

On 25/07/2017 17:41, Marc Zyngier wrote:
> On 25/07/17 15:48, Christoffer Dall wrote:
>> On Tue, Jul 25, 2017 at 02:47:55PM +0100, Marc Zyngier wrote:
>>> On 21/07/17 14:03, Christoffer Dall wrote:
>>>> On Fri, Jul 07, 2017 at 09:41:42AM +0200, Auger Eric wrote:
>>>>> Hi Marc,
>>>>>
>>>>> On 04/07/2017 14:15, Marc Zyngier wrote:
>>>>>> Hi Eric,
>>>>>>
>>>>>> On 15/06/17 13:52, Eric Auger wrote:
>>>>>>> Currently, the line level of unmapped level sensitive SPIs is
>>>>>>> toggled down by the maintenance IRQ handler/resamplefd mechanism.
>>>>>>>
>>>>>>> As mapped SPI completion is not trapped, we cannot rely on this
>>>>>>> mechanism and the line level needs to be observed at distributor
>>>>>>> level instead.
>>>>>>>
>>>>>>> This patch handles the physical IRQ case in vgic_validate_injection
>>>>>>> and get the line level of a mapped SPI at distributor level.
>>>>>>>
>>>>>>> Signed-off-by: Eric Auger <eric.auger@redhat.com>
>>>>>>>
>>>>>>> ---
>>>>>>>
>>>>>>> v1 -> v2:
>>>>>>> - renamed is_unshared_mapped into is_mapped_spi
>>>>>>> - changes to kvm_vgic_map_phys_irq moved in the previous patch
>>>>>>> - make vgic_validate_injection more readable
>>>>>>> - reword the commit message
>>>>>>> ---
>>>>>>>  virt/kvm/arm/vgic/vgic.c | 16 ++++++++++++++--
>>>>>>>  virt/kvm/arm/vgic/vgic.h |  7 ++++++-
>>>>>>>  2 files changed, 20 insertions(+), 3 deletions(-)
>>>>>>>
>>>>>>> diff --git a/virt/kvm/arm/vgic/vgic.c b/virt/kvm/arm/vgic/vgic.c
>>>>>>> index 075f073..2e35ac7 100644
>>>>>>> --- a/virt/kvm/arm/vgic/vgic.c
>>>>>>> +++ b/virt/kvm/arm/vgic/vgic.c
>>>>>>> @@ -139,6 +139,17 @@ void vgic_put_irq(struct kvm *kvm, struct vgic_irq *irq)
>>>>>>>  	kfree(irq);
>>>>>>>  }
>>>>>>>  
>>>>>>> +bool irq_line_level(struct vgic_irq *irq)
>>>>>>> +{
>>>>>>> +	bool line_level = irq->line_level;
>>>>>>> +
>>>>>>> +	if (unlikely(is_mapped_spi(irq)))
>>>>>>> +		WARN_ON(irq_get_irqchip_state(irq->host_irq,
>>>>>>> +					      IRQCHIP_STATE_PENDING,
>>>>>>> +					      &line_level));
>>>>>>> +	return line_level;
>>>>>>> +}
>>>>>>> +
>>>>>>>  /**
>>>>>>>   * kvm_vgic_target_oracle - compute the target vcpu for an irq
>>>>>>>   *
>>>>>>> @@ -236,13 +247,14 @@ static void vgic_sort_ap_list(struct kvm_vcpu *vcpu)
>>>>>>>  
>>>>>>>  /*
>>>>>>>   * Only valid injection if changing level for level-triggered IRQs or for a
>>>>>>> - * rising edge.
>>>>>>> + * rising edge. Injection of virtual interrupts associated to physical
>>>>>>> + * interrupts always is valid.
>>>>>>>   */
>>>>>>>  static bool vgic_validate_injection(struct vgic_irq *irq, bool level)
>>>>>>>  {
>>>>>>>  	switch (irq->config) {
>>>>>>>  	case VGIC_CONFIG_LEVEL:
>>>>>>> -		return irq->line_level != level;
>>>>>>> +		return (irq->line_level != level || unlikely(is_mapped_spi(irq)));
>>>>>>>  	case VGIC_CONFIG_EDGE:
>>>>>>>  		return level;
>>>>>>>  	}
>>>>>>> diff --git a/virt/kvm/arm/vgic/vgic.h b/virt/kvm/arm/vgic/vgic.h
>>>>>>> index bba7fa2..da254ae 100644
>>>>>>> --- a/virt/kvm/arm/vgic/vgic.h
>>>>>>> +++ b/virt/kvm/arm/vgic/vgic.h
>>>>>>> @@ -96,14 +96,19 @@
>>>>>>>  /* we only support 64 kB translation table page size */
>>>>>>>  #define KVM_ITS_L1E_ADDR_MASK		GENMASK_ULL(51, 16)
>>>>>>>  
>>>>>>> +bool irq_line_level(struct vgic_irq *irq);
>>>>>>> +
>>>>>>>  static inline bool irq_is_pending(struct vgic_irq *irq)
>>>>>>>  {
>>>>>>>  	if (irq->config == VGIC_CONFIG_EDGE)
>>>>>>>  		return irq->pending_latch;
>>>>>>>  	else
>>>>>>> -		return irq->pending_latch || irq->line_level;
>>>>>>> +		return irq->pending_latch || irq_line_level(irq);
>>>>>>
>>>>>> I'm a bit concerned that an edge interrupt doesn't take the distributor
>>>>>> state into account here. Why is that so? Once an SPI is forwarded to a
>>>>>> guest, a large part of the edge vs level differences move into the HW,
>>>>>> and are not that different anymore from a SW PoV.
>>>>>
>>>>> As pointed out by Christoffer in https://lkml.org/lkml/2017/6/8/322,
>>>>> isn't it a bit risky in general to poke the physical state instead of
>>>>> the virtual state. For level sensitive, to me we don't really have many
>>>>> other alternatives. For edge, we are not obliged to.
>>>>
>>>> I think we need to be clear on the fundamental question of whether or
>>>> not we consider pending_latch and/or line_level for mapped interrupts.
>>>>
>>>> I can definitely see the argument that the pending state is kept in
>>>> hardware, so if you want to know that for a mapped interrupt, ask the
>>>> hardware.
>>>>
>>>> The upside of this appraoch is a clean separation of state and we avoid
>>>> any logic to synchronize a virtual state with the physical state.
>>>>
>>>> The downside is that it's slower to peek into the physical GIC than to
>>>> read a variable from memory, and we need to special case the validate
>>>> path (which I now understand).
>>>>
>>>> If we move to keeping the state in HW, how do we deal with GICD_SPENDR ?
>>>> Does that mean we will forward a from the VM handled by the VGIC to the
>>>> physical GIC?
>>>
>>> Sounds like it to me. Otherwise, we start loosing some state. 
>>
>> How do we loose state?  Is it not more a question of complexity to make
>> sure the 'cached' vgic state is up to date with the real state?  (like
>> what we currently do for the timer mapped interrupt).
> 
> Sorry, I was very imprecise here. It is not so much that we'd loose
> state, but that we'd have some state at the wrong location. If we have a
> guest playing with the pending state, we need to make sure the physical
> side is up to date. Otherwise, we can end-up in situations where we'd
> inject an interrupt for the guest based on a pending state that only
> exists in the virtual distributor, and yet the virtual CPUIF is going to
> try and deactivate it in the physical side on EOI.

At the moment the write to GICD_ISPENDR (vgic_mmio_write_spending) sets
the pending_latch virtual state.

and the read uses irq_is_pending() new wrapper which looks at both the
physical state and the pending_latch virtual state in case of physically
mapped irq (irq_is_pending).

But As Marc mentioned, as I don't propagate the pending_latch to the
physical distributor, effectively we can have a guest "DIR" that tries
to complete a physical IRQ that is not active and that's wrong.

> 
>> On GICv2 this is likely going to make injecting timer interrupts slower,
>> because we'll check the pending state of whatever's in the AP list on
>> entry to the guest and peek into the physical GIC again.
aren't timer interrupts PPIs? My series only affect the state machine
for physically mapped SPIs so this shouldn't have any impact, no?

Thanks

Eric
> 
> That's a very valid concern. Though there is a slight distinction with
> the timer, in that we entirely control the injection of the interrupt,
> while an SPI can fire an any particular moment.
> 
>>
>>> Note that
>>> this is what my GICv4 patches are also doing, by forwarding the INT and
>>> CLEAR commands to the physical ITS.
>>>
>>>>> Don't we have situations, due to the lazy disable approach, where the
>>>>> physical IRQ hits, enters the genirq handler and the actual handler is
>>>>> not called, ie. the virtual IRQ is not injected?
>>>>>
>>>>
>>>> I'm not sure I remember what these situations were, specifically, but
>>>> certainly if we ever have a situation where a mapped irq's pending state
>>>> should be different from that of the physical one, then it doesn't work.
>>>
>>> There is a very simple way around the issue Eric mentions, which is to
>>> use the "IRQ_DISABLE_UNLAZY" flag, which disable the lazy disabling of
>>> interrupts.  That's also something the GICv4 patches use, as when we
>>> mask an interrupt, we want to be sure that it is immediately done (the
>>> host side will never see the interrupt firing, and thus can never
>>> disabled it).
>>>
>> If we don't care about the potential performance hit mentioned above, it
>> sounds like a good solution to me.
> 
> I think we need to measure the damage this would cause.
> 
> Thanks,
> 
> 	M.
>