Please report new issues athttps://github.com/DOCGroup
When using the RW wait startegy, the Transport_Cache_Manager::find_transport() temporarily removes the selected handler from the reactor read list. This is supposed to be undone in the wait strategy TAO_Wait_On_Read::wait(). However since in one-way invocations, the wait isn't done the handle gets lost. This is demonstrated in the attached test (a modification of the $TAO_ROOT/tests/CallbackTest). A server using RW wait makes a oneway callback to the client(). The client starts upon a fixed endpoint. Since the server handle wasn't put back into the reactor read list, the client connection closure goes undetected. When a second client comes online, the server reuses the half-closed connection to send the one-way callback. This deadlocks. ./run_test.pl // Tests the default wait startegy (LF). ./run_test.pl default // Tests using the RW wait startegy. Ciju
Created attachment 823 [details] Demostrate handle loss when making oneway invocation using RW wait startegy.
feel free to commit the regression test into the repo. Do you want this be handled as blocker for x.5.10 (you did put severity on blocker)?
> ------- Comment #2 from jwillemsen@remedy.nl 2007-07-23 12:45 ------- > feel free to commit the regression test into the repo. Ok. > Do you want this be > handled as blocker for x.5.10 (you did put severity on blocker)? No. I don't think the release should be blocked due to this (this is a corner case that occurs rarely). I will change it to major. A solution for this should refactoring out this code (removing handle from the reactor handle list) from the Transport_Cache and move it completely into the wait strategy. Ciju
do you also plan to handle this yourself, if so, please assign the bug to yourself
I wasn't planning on it. I just wanted to document this problem (interestingly I found this while debugging a DDS issue). I will accept the bug, but can't say when I will be able to work on it. Ciju
still failing with x.5.10
Reassign to john_c@ociweb.com, he indicated he was willing to accept
revert change below when fixed Tue Jul 29 18:11:32 UTC 2008 Johnny Willemsen <jwillemsen@remedy.nl> * bin/tao_orb_tests.lst: TAO/tests/Bug_3000_Regression fails because that issue remains unresolved, added !FIXED_BUGS_ONLY.
Activated this test on all builds and it runs perfectly as far as I can tell. It seems recent changes to RW strategy (and/or other changes since the submit of this bug) have resolved this issue.