epoll_wait hangs after a large write to a dead process' non-blocking pipe returns EAGAIN

Question

Scenario:

• A subprocess is started, its stdin and stdout are attached to the parent process
• SIGPIPE is blocked to prevent erroring out when the subprocess dies (sigtimedwait is used to clear the signal)
• Soon the subprocess is killed with -9 (it's a dedicated test)
• The parent writes 1MB to the subprocess' stdin, which - in some cases - hangs. The issue becomes much more pronounces if the CPU is loaded (e.g. by stress -c 96). A race condition, yay!

It looks like the difference between the successful and the hung runs is this:

Successful runs: write attempt fails with errno=EPIPE, which is then handled

Failed runs: the first 64K are written successfully (presumably filling the buffer), next write returns -1 with errno=EAGAIN, then epoll_wait is called with EPOLLOUT without a timeout (to wait for the buffer to clear under normal circumstances) - AND it never returns

I tried:

• Adding EPOLLHUP and EPOLLERR that should trigger regardless
• Adding EPOLLRDHUP which was promising, but achieved nothing
• Checking that the process is alive right before calling epoll_wait, but it's always shown as alive (tried both kill(pid, 0) and waitpid(pid, &status, WNOHANG)). Plus, it would still be racy anyway
• Checking sigtimedwait's output - it returns -1 with errno=EAGAIN (sadly no EPIPE as I could handle that)
• Switching to level-triggered, assuming something was getting lost due to re-arming the relevant events. No change.

Note 1: This seems to be an epoll's quirk, and if so - I'd appreciate a recommended workaround

Note 2: I have a solution that works around the hang, but I'm not sure it's a great one - so will hold off sharing it to not skew the answers