[Bug]: Occasional segfault when throwing an error from inside a coforall #26652
Comments
|
Note that switching to a For example This always seems to give the correct result. Even with |
Where the difference is... the lack of For users bumping into this (and Arkouda), what would be an acceptable workaround? Could we have a racy flag that'll be set to |
The difference is the lack of an
This sounds possible? In essence making the body of the on statement one giant Option A proc bar() throws {
var err: owned Error? = nil;
coforall l in Locales with (ref err) do on l {
try {
var x = new C(l.id);
if l.id == 1 then throw new Error("exception");
writeln("printing ", l.id, ": ", x);
} catch e {
err = e;
}
}
if err != nil then throw err;
}Option B Option C These all still segfault every other run or so. |
|
Have you tried it with |
|
I see the same issue with CHPL_COMM=ofi (using mpirun on a Mac as well as slurm on an EX) It sporadically hits a segfault. |
|
Noting that the segfault occurs during a call to Using the option B code, I'm able to reproduce the segfault. If I remove the writeln() in the coforall body in bar(), no segfault. If I instead remove the definition of If I instead remove the |
Summary of Problem
Description:
When throwing an error from inside a coforall, I am seeing it sometimes crashes with a segfault at a few places in our runtime. Sometimes the code runs to completion and sometimes it hits a segfault. The likelihood of a segfault increases with the number of locales, I find that with 4 locales I get one about every other run.
Is this issue currently blocking your progress?
yes
Steps to Reproduce
Source Code:
Compile command:
chpl foo.chplExecution command:
./foo -nl 4Example output
Example output with GASNET_BACKTRACE=1
Expected correct output
Configuration Information
chpl --version: 2.4 prerelease$CHPL_HOME/util/printchplenv --anonymize:gcc --versionorclang --version: LLVM 19, clang 14The text was updated successfully, but these errors were encountered: