[email protected] (Rick Fochtman) writes: > ISTR a similar problem with the 360/67, when a BAL or its target were > split across a page boundary. The dialectric material in the ROS needed > replacement and the CE replaced it with three layers of SARAN Wrap and > the machine worked perfectly after that. Of course the capacitive ROS > had to be tightened to a different torque spec, but Burlington (or > whereever) came through with the right specs for that as well.
sounds like hardware bug in specific machine ... since it would have been very evident on in lots of applications. re: http://www.garlic.com/~lynn/2009n.html#74 Best IEFACTRT (off topic) however, charlie (invented compare&swap instruction ... CAS was chosen because they are charlie's initials) found a implementation flaw in 360/67s ... that I don't believe ever got fixed. he was trying to squeeze a couple more cycles out of cp67 kernel ... and the interrupt handlers had a LCTL CR0,CR0 ... loading segment table pointer control register (this was moved to CR1 in 370 architecture) ... so he no-op'ed the LCTL (since it was just reloading the value that was already there. System started failing. After lots of diagnoses ... turned up hardware design flaw. A page fault interrupt resulted in the look-aside buffer (DAT, associative array) having all the real page numbers set to zero ... but valid/not-valid indicators for the entries weren't reset. This resulted in any virtual page numbers already in the associative array being mapped to real page zero. The "LCTL CRO" (even if reloading the same segment table pointer) would reset the associative array (setting all entries to invalid ... which had masked the associative array hardware problem on page fault. The LCTL CRO" went back in (I think cheaper than trying to correct the problem. misc. past posts mentioning multiprocessor and/or compare&swap instruction http://www.garlic.com/~lynn/subtopic.html#smp for other drift ... trying to get compare&swap into 370 was initially rebuffed ... we were told that the favorite son operating system didn't feel that it was necessary ... that TEST&SET (from 360 multiprocessor support was more than adequate) ... if compare&swap was ever going to be justified for 370, a non-multiprocessor specific use for the instruction would be needed. thus was born the programming notes in the principles of operations ... on how to use compare&swap in multitasking (multithreaded) applications (whether or not they were running on multiprocessor machine). -- 40+yrs virtualization experience (since Jan68), online at home since Mar1970 ---------------------------------------------------------------------- For IBM-MAIN subscribe / signoff / archive access instructions, send email to [email protected] with the message: GET IBM-MAIN INFO Search the archives at http://bama.ua.edu/archives/ibm-main.html
