Why is this sigfillset needed ?

Perhaps add volatile somewhere ? Or use C11 atomic_store relaxed.

I wonder if we want to be more formal there, and add a fence before accesses to last_sig and last_code.

Don't you need to truncate shmfd before access ?

I see, you did not extended shmfd at all. I think it would be more interesting to ensure that at least one page is allowed.

To prevent the signal handler being entered again between siglongjmp and sigsetjmp.

This is not a very strong reason, but I need to initialize sa.sa_mask in any case.

Not sure if that's needed, given that mmap() and munmap() are not magical the way malloc() and free() are, but I'm adding it anyway.

last_sig and last_code are volatile, which should suffice for communication with a signal handler in the same thread.