Vladimir Ozerov created IGNITE-2876:
---------------------------------------
Summary: IGFS: System pool starvation is possible during data
block write.
Key: IGNITE-2876
URL: https://issues.apache.org/jira/browse/IGNITE-2876
Project: Ignite
Issue Type: Bug
Components: IGFS
Affects Versions: 1.5.0.final
Reporter: Vladimir Ozerov
Assignee: Ivan Veselovsky
Priority: Critical
Fix For: 1.6
*Problem*
IGFS has a set of messages to exchange data and signal events between nodes.
These are:
- {{IgfsAckMessage}}
- {{IgfsBlocksMessage}}
- {{IgfsDeleteMessage}}
- {{IgfsFragmentizerRequest}}
- {{IgfsFragmentizerResponse}}
Currently these messages are processed in a system pool which is wrong and may
lead to starvation, deadlocks and incorrect behavior.
Several examples:
1) {{IgfsBlocksMessage}} handling logic performs "Cache.putAsync" operation.
This operation involves acquiring of semaphore permit. This semaphore, in turn,
can only be released from another thread in the same system pool. As such, all
system pool threads could hang on permit acquire forever.
2) In case file system size is exceeded, the same message waits for some time
in hope that free space in cache will appear. However, if all system pool
threads waits for this point, concurrent block removal cannot proceed, so these
threads are doomed to receive {{IgfsOutOfSpaceException}} irrespective of
whether they wait or not.
*Solution*
1) Introduce new IO policy for IGFS (see {{GridIoPolicy}}).
2) Force all IGFS message to be processed with this policy. No backward
compatibility is needed.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
