Re: [PATCH v6 0/3] vfio-ccw: Fix interrupt handling for HALT/CLEAR

[Eric Farman <farman@linux.ibm.com>]

On Thu, 2021-05-13 at 03:05 +0200, Halil Pasic wrote:
> On Tue, 11 May 2021 21:56:28 +0200
> Eric Farman <farman@linux.ibm.com> wrote:
> 
> > Hi Conny, Matt, Halil,
> > 
> > Here's one (last?) update to my proposal for handling the collision
> > between interrupts for START SUBCHANNEL and HALT/CLEAR SUBCHANNEL.
> > 
> > Only change here is to include Conny's suggestions on patch 3.
> > 
> > Thanks,
> 
> I believe these changes are beneficial, although I don't understand
> everything about them. In that sense I'm happy with the these getting
> merged.
> 
> Let me also spend some words answering the unasked question, what I'm
> not understanding about these.
> 
> Not understanding how the problem stated in the cover letter of v4 is
> actually resolved is certainly the most important one. 

Per our phone call last week, one of Conny's suggestions from that
particular version was related to vfio_ccw_sch_io_todo() and was giving
me some difficulties. We all agreed that I should send what I had, and
leave the other corner case(s) to be addressed later along with the
broader serialization topic throughout the driver. That is still my
intention, but I suspect that's where you are going here...

(I realize I said "last?" at the top here. Poor decision on my part.)

> Let me cite
> the relevant part of it (your cover letter already contains a link to
> the full version).
> 
> """
> 
> 	CPU 1			CPU 2
>  1	CLEAR SUBCHANNEL
>  2	fsm_irq()
>  3				START SUBCHANNEL
>  4	vfio_ccw_sch_io_todo()
>  5				fsm_irq()
>  6				vfio_ccw_sch_io_todo()
> 
> From the channel subsystem's point of view the CLEAR SUBCHANNEL (step
> 1)
> is complete once step 2 is called, as the Interrupt Response Block
> (IRB)
> has been presented and the TEST SUBCHANNEL was driven by the cio
> layer.
> Thus, the START SUBCHANNEL (step 3) is submitted [1] and gets a cc=0
> to
> indicate the I/O was accepted. However, step 2 stacks the bulk of the
> actual work onto a workqueue for when the subchannel lock is NOT
> held,
> and is unqueued at step 4. That code misidentifies the data in the
> IRB
> as being associated with the newly active I/O, and may release memory
> that is actively in use by the channel subsystem and/or device. Eww.
> """
> 
> The last sentence clearly states "may release memory that is actively
> used by ... the device", and I understood it refers to the invocation
> of cp_free() from vfio_ccw_sch_io_todo(). Patch 3 of this series does
> not change the conditions under which cp_free() is called.

Correct.

> 
> Looking at the cited diagram, since patch 3 changes things in
> vfio_ccw_sch_io_todo() it probably ain't affecting steps 1-3 and
> I understood the description so that bad free happens in step 4.

You are correct that patch 3 touches vfio_ccw_sch_io_todo(), but it is
not addressing the possibility of a bad free described in the old cover
letter. The commit message for patch 3 describes pretty clearly the
scenario in question.

> 
> My guess is that your change from patch 3 somehow via the fsm
> prevents
> the SSCH on CPU 2 (using the diagram) from being executed  if it
> actually
> happens to be after vfio_ccw_sch_io_todo(). 

That's an incorrect guess. The code in vfio_ccw_sch_io_todo() today
says "If another CPU is building an I/O (FSM is CP_PROCESSING), or
there is no CPU building an I/O (FSM is IDLE), then skip the cp_free()
call." The change in patch 3 says that in that situation, it should
also not adjust the FSM state because the interrupt being handled on
CPU1 was unrelated (maybe it was for a HALT/CLEAR, maybe it was an
unsolicited interrupt). The SSCH on CPU2 will still go on as expected.

> And patch 1 is supposed to
> prevent the SSCH on CPU2 from being executed in the depicted case
> because
> if there is a cp to free, then we would bail out form if we see it
> while processing the new IO request.

Not really. It's the FSM's job to prevent a second SSCH, and route to
fsm_io_retry() or fsm_io_busy() as appropriate. But the scenario
described by patch 3 in this series would leave the cp initialized,
while also resetting the FSM back to IDLE. As such, the FSM was free to
allow another SSCH in, which would then re-initialize the cp and orphan
the existing (active) cp resources.

With the application of patch 3, that concern isn't present, so the
change in patch 1 is really a NOP. But it allows for consistency in how
the cp_*() functions are working, and a safety valve should this
situation show up another way. (We'll get trace data that says
cp_init() bailed out, rather than going on as if nothing were wrong.)

> 
> In any case, I don't want to hold this up any further.
> 

Thanks for that. You are correct that there's still a potential issue
here, in the handoff between fsm_irq() and vfio_ccw_sch_io_todo(), and
another fsm_io_request() that would arrive between those points. But
it's not anything that we haven't already discussed, and will hopefully
begin discussing in the next couple of weeks.

Thanks,
Eric

> Regards,
> Halil