Please report new issues athttps://github.com/DOCGroup
On Win32 platforms, if a oneway request is made on a server that has crashed, the client remains unaware of it. This is because, as far as we know, the MS implementation of TCP does not have the write fail (and return with errno==EPIPE) as it does on many other platforms. A read must be performed to find this out, either by explicitly reading or by setting a read mask on a select(). Evidently, this is legal as far as TCP is concerned, just different from other implementations. A workaround has been found at the application level, by Chris Cleeland <cleeland_c@ociweb.com>, where perform_work() is called with a zero time value between each oneway request. However, it would be nice to have a workaround at the ORB level. The TAO Crash_On_Write test demonstrates this condition. On Win32 platforms, no CORBA::TRANSIENT exceptions are thrown after the server is deliberately crashed. The oneway requests just get buffered by TCP.
Not a high priority at present, but this entry will keep us from forgetting it.
This still happens with 1.4.1. The Crash_On_Write test on Windows gives the result below (client timeout is increased to 120 seconds). The oneway operations are really slowed down by the fact the server is down, before the server crash each line is printed about once a second, after the server crash once per 10 seconds. C:\ACE\latest\ACE_wrappers\TAO\tests\Crash_On_Write>perl run_test.pl (2556|128) Starting crash task Finished iteration 0 Finished iteration 10 Finished iteration 20 Finished iteration 30 Finished iteration 40 Finished iteration 50 Finished iteration 60 Finished iteration 70 Finished iteration 80 Finished iteration 90 Finished iteration 100 Finished iteration 110 Finished iteration 120 Finished iteration 130 Finished iteration 140 Finished iteration 150 Finished iteration 160 Finished iteration 170 Finished iteration 180 Finished iteration 190 Finished iteration 200 Finished iteration 210 Finished iteration 220 Finished iteration 230 Finished iteration 240 Finished iteration 250 Finished iteration 260 Finished iteration 270 Finished iteration 280 Finished iteration 290 Finished iteration 300 Finished iteration 310 Finished iteration 320 Finished iteration 330 Finished iteration 340 Finished iteration 350 Finished iteration 360 Finished iteration 370 Finished iteration 380 Finished iteration 390 Finished iteration 400 Finished iteration 410 Finished iteration 420 Finished iteration 430 Finished iteration 440 Finished iteration 450 Finished iteration 460 Finished iteration 470 (2556|128) Sleep done, crashing the server Aborting... Abnormal program termination Finished iteration 480 Finished iteration 490 Finished iteration 500 Finished iteration 510 Finished iteration 520 Finished iteration 530
Jeff, because we know this test fails on Windows, shouldn't we disable the running of it on windows in our regression tests, this saves us from skipping this test in the log and we could add a remark then here when this is fixed, we must enable the test again?
Seems like a good idea to me, the Windows TCP situation is not likely to change any time soon.
Disabled the Crash_On_Write tests on Win32. When this is fixed, enable this test again.