dump_stack_trace() and dump_core_file() both run a script on our own pid
and copy its output to stderr a block at a time. The copy ended on any
write() which did not place the whole block:
if (write(2, buf, ret) != ret) {
// *sigh*
break;
}
write() is entitled to do that. stderr is a pipe when SITL runs under
autotest, so a full pipe shortens a write, and a timer signal can cut one
short with EINTR. Either way the copy stopped silently, part-way
through, with no "end dumpstack.sh output" line to show that anything was
missing.
Seen in CI: a backtrace ended at frame #8, in the middle of the frame
which would have named the caller - the one thing the dump exists to
provide. A truncated core dump goes the same way and is even easier to
miss.
Keep writing until the block is out, retry EINTR on both the read and the
write, and say so if a write really does fail.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Peter BarkerandClaude Opus 5
0da91dbac3