ANALYSE/SYSTEM - PRF and EXC, how do you find out what the exception is?
Folks,
Just a quick question regarding the exception information available from **bleep**/SYS. I have a system consisting of a number of processes, the majority written in Pascal with some modules in C and C++.
We're using OpenVMS 8-4, patched to the hilt running on RX2800s.
One of the processes was taking an inordinate amount of time to perform its processing, most of the CPU time was spent outside of user mode. **bleep**/SYS revealed that the process seemed to be doing nothing but exception processing. Finding a user mode PC address was quite a challenge (used PCS, EXC and FLT under **bleep**/SYS) but seemed to point to an old function using LIB$MOVC3 to perform string conversions. The code was reorganised to reduce the reliance on the problematic function, performance was improved by 95%. The process spent 1 second in COM rather than 20 seconds.
The process is still showing a large number of exceptions. The questions I have are:
1) What is an acceptable exception rate?
2) How do I find out the root cause of the exception?
I've attached some samples output from **bleep**/SYS . If the output from the PRF tool is to believed the process spends most of its time processing exceptions.
Re: ANALYSE/SYSTEM - PRF and EXC, how do you find out what the exception is?
>>> The process is still showing a large number of exceptions. The questions I have are:
You are lucky that you run on 8.4. Exception handling on VAX was cheap, it was not expensive on Alpha, it was very, very expensive on I64. For 8.4 one "very" could be optimized away.
>>> 1) What is an acceptable exception rate?
An exception should be an exception. Exceptions should not be used as a programming technique. If that's the case in either your code or in a library it should be changed.
>>> 2) How do I find out the root cause of the exception?
Didn't you enable EXC tracing? Its output should show more than you ever wanted to know including the root cause. EXC tracing was written for monitoring exception handling, the output is not (very) user friendly.
Re: ANALYSE/SYSTEM - PRF and EXC, how do you find out what the exception is?
>>> The output from EXC was less than helpful to be honest. How do I map an exception to a given process ID? Where is the root cause?
No surprise, as I said, it wasn't written for the general use. From bottom to top, look for the string '[$SRCHANDLER: Begin]', go up until you find a line with '[CHF$SEARCH begin'. Appended is the pcb, sig, ..., which is probably all you need/want (and then, for fun, watch what exection is tracing until it is done with that exception).
Re: ANALYSE/SYSTEM - PRF and EXC, how do you find out what the exception is?
OK, that helps a lot. In the space of 1 second (or less) for the process I'm interested I got 2294 exceptions all with the same basic arguments.
All I need to do now is map the sig (address I assume) into the process address space and then find it in the source. Running the offending exe in debug and examining the value that came back gives me an address in a comman library, presumably data as the debugger didn't show anything resembling source, just an offsett into one of our libraries. The offset being SHARE$NMCS2_APP_RWDATA0+0E477C which looks to be a linker construct.
So the question now becomes, how do I find out which module caused the aggravation.
Re: ANALYSE/SYSTEM - PRF and EXC, how do you find out what the exception is?
OK, so a bit of digging around seems to point the blame at the PASCAL builtin READV. The function calling READV is being called incorrectly anyway so its kind of blind luck that this has worked at all, at this point in the application the READV function is being passed a binary array (its expecting a deliniated numeric string), and hence causing the PASCAL exception handler to fire. We're using ERROR:=CONTINUE which forces contination after the event.
Re: ANALYSE/SYSTEM - PRF and EXC, how do you find out what the exception is?
>>> All I need to do now is map the sig (address I assume) into the process address space and then find it in the source.
No, pcb is the PCB address, sig is the condition value, aka chf$is_sig_name, arg1, ..., arg3 are an extract of the reminder of the signal array starting with arg1 as chf$is_sig_arg1, etc. (Write a simple C function like *(int*)47=11; and you will see:
>>> Running the offending exe in debug and examining the value that came back gives me an address in a comman library, presumably data as the debugger didn't show anything resembling source, just an offsett into one of our libraries. The offset being SHARE$NMCS2_APP_RWDATA0+0E477C which looks to be a linker construct.
No, a debugger construct used when the debugger has no (debug) symbols available for that VA. It takes whatever symbol is "close" to the VA and adds an offset. The output looks like a PSECT name to me. SDA with show image/all should help to find out what the image and the program segment for the VA really is.
>>> So the question now becomes, how do I find out which module caused the aggravation.
The usual way. You need a linker map for the image. With the SDA output you should be able to locate the module in the map. And with a machine code listing you will be able to find the offset within the module and so the source line.
Re: ANALYSE/SYSTEM - PRF and EXC, how do you find out what the exception is?
Hi,
Thanks for the clarifications and your help it has been much appreciated.
Anyway I've corrected he calls around the READV.Uultimately they weren't required any more, binary data was being passed in which caused an exception in the Pascal READV. This binary data had already been correctly converted earlier.
The problem had probably been in existance for over a decade, possibily two. We had noticed on the Alphas that this process spent too much time in COM but assumed it was normal behaviour. When we ported to Itanium the process spent even more time in COM which set the alarms bell ringing. We have plans for various extensions to the system which this problem would have badly affected. However, now we the processing down to well under a second and we're in a position now to consider the extensions.
