Robbie Gemmell created DISPATCH-1496:
----------------------------------------
Summary: router can grant extra credit during drain a link-routed
receiver
Key: DISPATCH-1496
URL: https://issues.apache.org/jira/browse/DISPATCH-1496
Project: Qpid Dispatch
Issue Type: Task
Reporter: Robbie Gemmell
The router can grant extra credit during drain of a link-routed receiver,
leading to the router having more outstanding credit to the peer than the
receiver has actually granted.
The following shows a link-routed receiver draining credit from the link, after
a single message having been received from the broker it was link-routed to.
The router relays the flow from client to broker as expected, with
"delivery-count=1, link-credit=999, drain=true", and relays the brokers
response with "delivery-count=1000, link-credit=0, drain=true":
{noformat}
[0x27bb800]:1 <- @flow(19) [next-incoming-id=1, incoming-window=2047,
next-outgoing-id=5, outgoing-window=2147483647, handle=2, delivery-count=1,
link-credit=999, drain=true]
[0x2763c90]:0 -> @flow(19) [next-incoming-id=2, incoming-window=2147483647,
next-outgoing-id=4, outgoing-window=2147483647, handle=2, delivery-count=1,
link-credit=999, drain=true]
[0x2763c90]:0 <- @flow(19) [next-incoming-id=4, incoming-window=2147483647,
next-outgoing-id=2, outgoing-window=2147483647, handle=2, delivery-count=1000,
link-credit=0, drain=true]
[0x27bb800]:1 -> @flow(19) [next-incoming-id=5, incoming-window=2147483647,
next-outgoing-id=1, outgoing-window=2147483647, handle=2, delivery-count=1000,
link-credit=0, drain=true]
{noformat}
The receiver then supplies more credit, which the router again relays as
expected:
{noformat}
[0x27bb800]:1 <- @flow(19) [next-incoming-id=1, incoming-window=2047,
next-outgoing-id=5, outgoing-window=2147483647, handle=2, delivery-count=1000,
link-credit=1000]
[0x2763c90]:0 -> @flow(19) [next-incoming-id=2, incoming-window=2147483647,
next-outgoing-id=4, outgoing-window=2147483647, handle=2, delivery-count=1000,
link-credit=1000, drain=false]
{noformat}
The receiver then drained this credit without having received any messages. The
router relays the drain, but _incorrectly grants additional credit to the
broker at the same time_, despite the receiver not having done so itself. The
credit is drained, and the router updates the receiver to note it is also
drained. Note however that the delivery counts have now diverged between
client->router and router->broker, which seems unexpected for a link route.
{noformat}
[0x27bb800]:1 <- @flow(19) [next-incoming-id=1, incoming-window=2047,
next-outgoing-id=7, outgoing-window=2147483647, handle=2, delivery-count=1000,
link-credit=1000, drain=true]
[0x2763c90]:0 -> @flow(19) [next-incoming-id=2, incoming-window=2147483647,
next-outgoing-id=6, outgoing-window=2147483647, handle=2, delivery-count=1000,
link-credit=2000, drain=true]
[0x2763c90]:0 <- @flow(19) [next-incoming-id=6, incoming-window=2147483647,
next-outgoing-id=2, outgoing-window=2147483647, handle=2, delivery-count=3000,
link-credit=0, drain=true]
[0x27bb800]:1 -> @flow(19) [next-incoming-id=7, incoming-window=2147483647,
next-outgoing-id=1, outgoing-window=2147483647, handle=2, delivery-count=2000,
link-credit=0, drain=true]
{noformat}
Had the broker been able to send messages using the mistakenly granted extra
credit, the router would not be able to pass these all on to the client until
it granted more credit. It isnt clear what would happen to them, but it seems
like they could get stuck in the router until more credit is granted by the
receiver, or the receiver detaches, etc. Again, unexpected for a link-route.
The system continued on in this case since since this receiver grants (and
drains) new credit:
{noformat}
[0x27bb800]:1 <- @flow(19) [next-incoming-id=1, incoming-window=2047,
next-outgoing-id=7, outgoing-window=2147483647, handle=2, delivery-count=2000,
link-credit=1000]
[0x2763c90]:0 -> @flow(19) [next-incoming-id=2, incoming-window=2147483647,
next-outgoing-id=6, outgoing-window=2147483647, handle=2, delivery-count=3000,
link-credit=1000, drain=false]
etc etc
{noformat}
however the delivery-count divergence grows over time as the situation repeats.
Noticed while trying out changes for DISPATCH-1488.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
