In theory, walking the stack without decoding and demangling should be async
signal safe.
In practice, it is not:
-
Looks like a page fault while dumping the trace on a containerized/virtualized
Windows system has a chance to deadlock. Page fault could happen easily
as we have to write the dump either to memory or to a file.
-
On POSIX systems a deadlock could happen if a signal is received when throwing
an exception #131.
Theoretically this could be worked around by bypassing the mutex locking
in C++-runtime at exception throw (sample
implementation in the 🐙 userver framework), or by using a very
modern runtime (glibc-2.35+ with modern
libgcc or modern
LLVM's libunwind).
-
-fomit-frame-pointer
like flags add additional complexity to the stack walking implementation,
which may also negatively affect the signal safety.
As a rule of thumb: do not capture stack traces
in signal handlers unless you are absolutely sure in your environment and inspected
all of its source codes.
