Re: [Xen-devel] xen/evtchn and forced threaded irq

[Julien Grall]

Hi,

On 26/02/2019 10:17, Roger Pau Monnà wrote:
> On Tue, Feb 26, 2019 at 10:03:38AM +0000, Julien Grall wrote:
> > Hi Roger,
> >
> > On 26/02/2019 09:44, Roger Pau Monnà wrote:
> > > On Tue, Feb 26, 2019 at 09:30:07AM +0000, Andrew Cooper wrote:
> > > > On 26/02/2019 09:14, Roger Pau Monnà wrote:
> > > > > On Mon, Feb 25, 2019 at 01:55:42PM +0000, Julien Grall wrote:
> > > > > > Hi Oleksandr,
> > > > > >
> > > > > > On 25/02/2019 13:24, Oleksandr Andrushchenko wrote:
> > > > > > > On 2/22/19 3:33 PM, Julien Grall wrote:
> > > > > > > > Hi,
> > > > > > > >
> > > > > > > > On 22/02/2019 12:38, Oleksandr Andrushchenko wrote:
> > > > > > > > > On 2/20/19 10:46 PM, Julien Grall wrote:
> > > > > > > > > > Discussing with my team, a solution that came up would be to
> > > > > > > > > > introduce one atomic field per event to record the number of
> > > > > > > > > > event received. I will explore that solution tomorrow.
> > > > > > > > > How will this help if events have some payload?
> > > > > > > > What payload? The event channel does not carry any payload. It only
> > > > > > > > notify you that something happen. Then this is up to the user to
> > > > > > > > decide what to you with it.
> > > > > > > Sorry, I was probably not precise enough. I mean that an event might have
> > > > > > > associated payload in the ring buffer, for example [1]. So, counting events
> > > > > > > may help somehow, but the ring's data may still be lost
> > > > > >   From my understanding of event channels are edge interrupts. By definition,
> > > > > IMO event channels are active high level interrupts.
> > > > >
> > > > > Let's take into account the following situation: you have an event
> > > > > channel masked and the event channel pending bit (akin to the line on
> > > > > bare metal) goes from low to high (0 -> 1), then you unmask the
> > > > > interrupt and you get an event injected. If it was an edge interrupt
> > > > > you wont get an event injected after unmasking, because you would
> > > > > have lost the edge. I think the problem here is that Linux treats
> > > > > event channels as edge interrupts, when they are actually level.
> > > >
> > > > Event channels are edge interrupts. There are several very subtle bugs
> > > > to be had by software which treats them as line interrupts.
> > > >
> > > > Most critically, if you fail to ack them, rebind them to a new vcpu, and
> > > > reenable interrupts, you don't get a new interrupt notification. This
> > > > was the source of a 4 month bug when XenServer was moving from
> > > > classic-xen to PVOps where using irqbalance would cause dom0 to
> > > > occasionally lose interrupts.
> > >
> > > I would argue that you need to mask them first, rebind to a new vcpu
> > > and unmask, and then you will get an interrupt notification, or this
> > > should be fixed in Xen to work as you expect: trigger an interrupt
> > > notification when moving an asserted event channel between CPUs.
> > >
> > > Is there any document that describes how such non trivial things (like
> > > moving between CPUs) work for event/level interrupts?
> > >
> > > Maybe I'm being obtuse, but from the example I gave above it's quite
> > > clear to me event channels don't get triggered based on edge changes,
> > > but rather on the line level.
> >
> > Your example above is not enough to give the semantics of level. You would
> > only use the MASK bit if your interrupt handler is threaded to avoid the
> > interrupt coming up again.
> >
> > So if you remove the mask from the equation, then the interrupt flow should be:
> >
> > 1) handle interrupt
> > 2) EOI
>
> This is bogus if you don't mask the interrupt source. You should
> instead do
>
> 1) EOI
> 2) Handle interrupt
>
> And loop over this.
So that's not a level semantics. It is a edge one :). In the level case, you 
would clear the state once you are done with the interrupt.

Also, it would be ACK and not EOI.

> > The EOI in our case would be clearing the PENDING state. In a proper level
> > interrupt, the state would stay PENDING if there were more to come. This is
> > not the case with the events and you will lose the interrupt.
> >
> > So I don't think they are proper level interrupts. They have more a
> > semantics of edge interrupts with some property of level (i.e for the
> > mask/unmask).
>
> OK, I guess it depends on how you look at it, to me event channels are
> maybe quirky level interrupts, but are definitely closer to level than
> edge interrupts, specially taking into account the interrupt injection
> that happens on unmask of a pending line, there's no such thing at all
> with edge interrupts.

Cheers,

--
Julien Grall