rethrow unwind exception

[manuel-schiller]

On Linux with pthread library spdlog causes an SIGABORT and crashes
the application in case it catches a thread specific cancellation
exception in a critical execution phase while in a try/catch block
in spdlog/detail/logger_impl.h

The exception is caught by some general catch(...) clause where
it is NOT rethrown.

However rethrowing these kind of exception is mandatory, otherwise
an abort will be caused by the glibc.

[gabime]

Intersting. Why GNUC? is it gcc specific?

[manuel-schiller]

This is not GCC specific, it is specific for usage of the "Native POSIX Thread Library" (NPTL), see also https://udrepper.livejournal.com/21541.html.

I guess we should rather use __linux__ instead of __GNUG__.

[gabime]

Will it compile under clang? also, could you please move this “#define” to spdlog/common.h?

[manuel-schiller]

moved to common.h

yes this will compile on clang as long as cxxabi.h is present

[manuel-schiller]

the error handler will be called for all exceptions except for abi::__forced_unwind&. Do you want the error handler to be called in this case as well and then rethrow?

[gabime]

Yes please.

[manuel-schiller]

done

[alurie]

@gabime we encountered the same issue described in this PR (using spdlog v1.6.1, on centos7).
We do pthread_cancel of a thread that happens to be in the middle of spdlog sink/flush, exception is caught and not re-thrown, causing the crash.
Looking at the spdlog code changes, the fix introduced in this PR was refactored out.
(e.g. 8ca1d84 removed catch(...),
then fd170b0 added it back with re-throw, then 78ea362 removed the re-throw).
If possible, could you please shed some light on those changes (I could not locate comments associated with those commits). Most interesting question is whether the re-throw was removed because it was causing some issue?
Thank you.

[gabime]

I think there was some problem with it, but I do not recall exactly..

[gabime]

@alurie This was probably the cause: #549

[gabime]

@alurie Fixed again

[Axas1986]

Why is this issue associated with v1.9.0? Is the release note of v1.9.0 incorrect?