"Fernando P. Schapachnik" wrote: > I'm trying to do Async I/O using O_ASYNC on sockets and handling > SIGIO. My testing shows that even if I unblock SIGIO at the begining of the > handler the kernel only delivers one level of nested signals. Ie: while the > first SIGIO is being handled a second might arrive, but a third delivered > signal does not reach the process.
Yes. Signals are persistant conditions, not events. > The same happens if I catch the signals with sigaction and specify > SA_NODEFER. Same program on Linux can handle up to 23 nested signals. Depending on Linux's behaviour, which is contrary to conformance to the POSIX standard, will make your code non-portable. > Is this a known behavior? Yes. Even better, it's standard. > Is there any way to change it? When you get the first SIGIO, set a flag (volatile) in the signal handler. In your main loop, check for the flag, and if it is present, use poll/select to verify that there is data pending, and while there is data pending, retrieve it. At most, you will recheck one extra time for a run of N events (N+1 times), and since you are concerned about N >= 23, this overhead is practically non-existant. If you want to avoid the "extra" system call, then switch to using non-blocking I/O, and when the number of bytes processed by the call is 0, then you are done with that iteration. This approach is portable between all versions of UNIX that support select or poll (you can conditionally compile the code based on the presence or absences of FD_SET_SIZE, or other manifest constants associated solely with the poll or select system calls, so the resulting code will be 100% portable source code). See also the source for the "init" program, which has to be able to reap child processes, and so has to do SIGCHLD or SIGCLD (depending on your flavor of UNIX) handling the same way. -- Terry To Unsubscribe: send mail to [EMAIL PROTECTED] with "unsubscribe freebsd-hackers" in the body of the message
