Re: [RFC PATCH 03/11] Drivers: hv: vmbus: Replace the per-CPU channel lists with a global array of channels

[Andrea Parri <parri.andrea@gmail.com>]

On Mon, Mar 30, 2020 at 02:45:54PM +0200, Vitaly Kuznetsov wrote:
> Andrea Parri <parri.andrea@gmail.com> writes:
> 
> >> Correct me if I'm wrong, but currently vmbus_chan_sched() accesses
> >> per-cpu list of channels on the same CPU so we don't need a spinlock to
> >> guarantee that during an interrupt we'll be able to see the update if it
> >> happened before the interrupt (in chronological order). With a global
> >> list of relids, who guarantees that an interrupt handler on another CPU
> >> will actually see the modified list? 
> >
> > Thanks for pointing this out!
> >
> > The offer/resume path presents implicit full memory barriers, program
> > -order after the array store which should guarantee the visibility of
> > the store to *all* CPUs before the offer/resume can complete (c.f.,
> >
> >   tools/memory-model/Documentation/explanation.txt, Sect. #13
> >
> > and assuming that the offer/resume for a channel must complete before
> > the corresponding handler, which seems to be the case considered that
> > some essential channel fields are initialized only later...)
> >
> > IIUC, the spin lock approach you suggested will work and be "simpler";
> > an obvious side effect would be, well, a global synchronization point
> > in vmbus_chan_sched()...
> >
> > Thoughts?
> 
> This is, of course, very theoretical as if we're seeing an interrupt for
> a channel at the same time we're writing its relid we're already in
> trouble. I can, however, try to suggest one tiny improvement:

Indeed.  I think the idea (still quite informal) is that:

  1) the mapping of the channel relid is propagated to (visible from)
     all CPUs before add_channel_work is queued (full barrier in
     queue_work()),

  2) add_channel_work is queued before the channel is opened (aka,
     before the channel ring buffer is allocate/initalized and the
     OPENCHANNEL msg is sent and acked from Hyper-V, cf. OPEN_STATE),

  3) the channel is opened before Hyper-V can start sending interrupts
     for the channel, and hence before vmbus_chan_sched() can find the
     channel relid in recv_int_page set,

  4) vmbus_chan_sched() finds the channel's relid in recv_int_page
     set before it search/load from the channel array (full barrier
     in sync_test_and_clear_bit()).

This is for the "normal"/not resuming from hibernation case; for the
latter, notice that:

  a) vmbus_isr() (and vmbus_chan_sched()) can not run until when
     vmbus_bus_resume() has finished (@resume_noirq callback),

  b) vmbus_bus_resume() can not complete before nr_chan_fixup_on_resume
     equals 0 in check_ready_for_resume_event().
     
(and check_ready_for_resume_event() does also provides a full barrier).

If makes sense to you, I'll try to add some of the above in comments.

Thanks,
  Andrea


> 
> vmbus_chan_sched() now clean the bit in the event page and then searches
> for a channel with this relid; in case we allow the search to
> (temporary) fail we can reverse the logic: search for the channel and
> clean the bit only if we succeed. In case we fail, next time (next IRQ)
> we'll try again and likely succeed. The only purpose is to make sure no
> interrupts are ever lost.  This may be an overkill, we may want to try
> to count how many times (if ever) this happens. 
> 
> Just a thought though.
> 
> -- 
> Vitaly
>