Andrea Parri <parri.andrea@xxxxxxxxx> writes:
> On Thu, Mar 26, 2020 at 03:16:21PM +0100, Vitaly Kuznetsov wrote:
>> "Andrea Parri (Microsoft)" <parri.andrea@xxxxxxxxx> writes:
>>
>> > The offer and rescind works are currently scheduled on the so called
>> > "connect CPU". However, this is not really needed: we can synchronize
>> > the works by relying on the usage of the offer_in_progress counter and
>> > of the channel_mutex mutex. This synchronization is already in place.
>> > So, remove this unnecessary "bind to the connect CPU" constraint and
>> > update the inline comments accordingly.
>> >
>> > Suggested-by: Dexuan Cui <decui@xxxxxxxxxxxxx>
>> > Signed-off-by: Andrea Parri (Microsoft) <parri.andrea@xxxxxxxxx>
>> > ---
>> > drivers/hv/channel_mgmt.c | 21 ++++++++++++++++-----
>> > drivers/hv/vmbus_drv.c | 39 ++++++++++++++++++++++++++++-----------
>> > 2 files changed, 44 insertions(+), 16 deletions(-)
>> >
>> > diff --git a/drivers/hv/channel_mgmt.c b/drivers/hv/channel_mgmt.c
>> > index 0370364169c4e..1191f3d76d111 100644
>> > --- a/drivers/hv/channel_mgmt.c
>> > +++ b/drivers/hv/channel_mgmt.c
>> > @@ -1025,11 +1025,22 @@ static void vmbus_onoffer_rescind(struct vmbus_channel_message_header *hdr)
>> > * offer comes in first and then the rescind.
>> > * Since we process these events in work elements,
>> > * and with preemption, we may end up processing
>> > - * the events out of order. Given that we handle these
>> > - * work elements on the same CPU, this is possible only
>> > - * in the case of preemption. In any case wait here
>> > - * until the offer processing has moved beyond the
>> > - * point where the channel is discoverable.
>> > + * the events out of order. We rely on the synchronization
>> > + * provided by offer_in_progress and by channel_mutex for
>> > + * ordering these events:
>> > + *
>> > + * { Initially: offer_in_progress = 1 }
>> > + *
>> > + * CPU1 CPU2
>> > + *
>> > + * [vmbus_process_offer()] [vmbus_onoffer_rescind()]
>> > + *
>> > + * LOCK channel_mutex WAIT_ON offer_in_progress == 0
>> > + * DECREMENT offer_in_progress LOCK channel_mutex
>> > + * INSERT chn_list SEARCH chn_list
>> > + * UNLOCK channel_mutex UNLOCK channel_mutex
>> > + *
>> > + * Forbids: CPU2's SEARCH from *not* seeing CPU1's INSERT
>>
>> WAIT_ON offer_in_progress == 0
>> LOCK channel_mutex
>>
>> seems to be racy: what happens if offer_in_progress increments after we
>> read it but before we managed to aquire channel_mutex?
>
> Remark that the RESCIND work must see the increment which is performed
> "before" queueing the work in question (and the associated OFFER work),
> cf. the comment in vmbus_on_msg_dpc() below and
>
> dbb92f88648d6 ("workqueue: Document (some) memory-ordering properties of {queue,schedule}_work()")
>
> AFAICT, this suffices to meet the intended behavior as sketched above.
> I might be missing something of course, can you elaborate on the issue
> here?
>
In case we believe that OFFER -> RESCINF sequence is always ordered
by the host AND we don't care about other offers in the queue the
suggested locking is OK: we're guaranteed to process RESCIND after we
finished processing OFFER for the same channel. However, waiting for
'offer_in_progress == 0' looks fishy so I'd suggest we at least add a
comment explaining that the wait is only needed to serialize us with
possible OFFER for the same channel - and nothing else. I'd personally
still slightly prefer the algorythm I suggested as it guarantees we take
channel_mutex with offer_in_progress == 0 -- even if there are no issues
we can think of today (not strongly though).
--
Vitaly
