THANKS! Thanks to all who replied! I'm very impressed with the generosity,
expertise and quality of responses from all who replied!
I had a small problem: recompiling a program while an earlier version is
running
crashes the running process. This is a bother to me because I test enhancements
to my programs while a production version is running.
The general consensus was that OSF/1, like many other UNIX OSs, performs paging
on demand--code is read in from a.out only when needed.
It's not clear to me that this is true of every UNIX, as claimed by some
writers.
I didn't get this on AIX, for example. Two writers pointed out that other
UNIXes, including ULTRIX, had something called ETXTBSY. A signal meaning
"executable text is busy--come back later." It was suggested that this wasn't
included in OSF/1 out of deference to the recent UNIX standard.
A few respondents thought that a UNIX with demand paging should normally protect
the running process by locking the disk file, or keeping its pages in the file
system. As one writer pointed out though, things like fork() might get weird...
Another point made by some respondents is that demand paging can cause extra
trouble under NFS. If the filed a.out changes while a remote machine executes
the older a.out, the process doesn't crash. Instead, the network can get
flooded
with requests when the client demands more a.out code from the server.
One thing still bothers me. Sometimes my running version kept on running
after I
compiled. Perhaps all of the code text had been loaded into memory by then...
Here are fixes/workarounds:
* Wait till the executing process finishes before recompiling!
* Rename the old executable. (If this works though, it suggests that the OS is
smart enough to keep track of the file with the old executable--perhaps by
keeping a link to the old file. Why shouldn't it be smart enough to protect the
file , or its links, or move it by itself? Because cc/ld reuses existing links.
* Delete the old executable. This is similar to the last point.
* Put a call to plock(2) in the code to prevent it being paged while compiling,
then more code to rexec the processif it gets SIGHUP. Whew!
* Compile with -N: makes a.out neither sharable nor pageable.
All up, I think I'll just give stable running versions a new name, like
cprog.running and compile test versions to plain cprog.
CREDITS:
Ireland: Fergal McCarthy
England: Kevin J Walters, Dave Newbold
UK: Andrew Gabriel
USA: Clyde Hoover, Jon Howell, David Warren, Drew Kramer, Tom Blinn
USA (?): Tr Nguyen, Drew Gallatin, Mike Iglesias, Alan Rollow, Steve Gibbons
Germany: Rob Hooft
Austria: Peter Lanstoeger
Switzerland: Stefan Monnier
----------
Andrej Panjkov, | email: matap_at_lure.latrobe.edu.au
School of Mathematics, | ph: 61 +3 479 2595
La Trobe University, | fax: 61 +3 478 7621
Bundoora, Vic, 3083 |
Australia. | Aphorism/Film quote/Freudian Slip goes here.
Received on Mon Feb 13 1995 - 22:48:11 NZDT