Hello, Michael Kelly, le mer. 28 janv. 2026 21:07:00 +0000, a ecrit: > 4) The signal handler stack is set up using the same code as in 1) but in > this instance there is now a pre-existing ss->context within > _hurd_setup_sighandler in trampoline.c. The code however unconditionally > refetches the floating point context via thread_get_state and overwrites the > original state preserved during handling of the first signal.
Oh, nice catch! > I was able to fix the testcases and indeed the stress-ng error reports by > not refetching the floating point state at times when 'ss->context' wasn't > NULL on entry. If this seems a viable method for solving this issue then I > can tidy up what I've done and submit a patch for review. Yes, that should be fine. > The signal handling code seems very tricky to me and there's more than a > chance that I've misunderstood or misinterpreted parts of it. Signals are essentially the most complex thing in Unix. Samuel
