test_sigwaitinfo fails on NetBSD due to POSIX violation

[furkanonder]

Bug report

Bug description:

The SignalEINTRTest.test_sigwaitinfo test in test_eintr fails on NetBSD. According to the official POSIX.1-2024 specification for sigwaitinfo:

"The sigtimedwait() and sigwaitinfo() functions may fail if: [EINTR] The wait was interrupted by an unblocked, caught signal."

NetBSD's sigwaitinfo() implementation violates POSIX by returning errno = ECANCELED (87) instead of errno = EINTR (4) when interrupted by a signal not in the wait set.

Configuration

./configure --with-pydebug

Test

╰─$ ./python -m unittest test._test_eintr.SignalEINTRTest.test_sigwaitinfo -v
test_sigwaitinfo (test._test_eintr.SignalEINTRTest.test_sigwaitinfo) ... ERROR

======================================================================
ERROR: test_sigwaitinfo (test._test_eintr.SignalEINTRTest.test_sigwaitinfo)
----------------------------------------------------------------------
Traceback (most recent call last):
  File "/home/blue/Desktop/cpython/Lib/test/_test_eintr.py", line 440, in test_sigwaitinfo
    self.check_sigwait(wait_func)
    ~~~~~~~~~~~~~~~~~~^^^^^^^^^^^
  File "/home/blue/Desktop/cpython/Lib/test/_test_eintr.py", line 430, in check_sigwait
    wait_func(signum)
    ~~~~~~~~~^^^^^^^^
  File "/home/blue/Desktop/cpython/Lib/test/_test_eintr.py", line 438, in wait_func
    signal.sigwaitinfo([signum])
    ~~~~~~~~~~~~~~~~~~^^^^^^^^^^
OSError: [Errno 87] Operation canceled

----------------------------------------------------------------------
Ran 1 test in 0.112s

FAILED (errors=1)

Reproduction

I created a minimal C program that reproduces the same issue.

#include <signal.h>
#include <pthread.h>
#include <unistd.h>
#include <sys/wait.h>
#include <stdio.h>
#include <stdlib.h>
#include <errno.h>
#include <string.h>

volatile sig_atomic_t got_sigalrm = 0;

void sigalrm_handler(int sig) {
    got_sigalrm = 1;
    printf("SIGALRM handler called\n");
}

int main() {
    printf("sigwaitinfo interruption test\n");
    
    sigset_t sigset;
    siginfo_t info;
    int target_signal = SIGUSR1;
    
    // Set SIGALRM handler
    struct sigaction sa = {0};
    sa.sa_handler = sigalrm_handler;
    sigaction(SIGALRM, &sa, NULL);
    
    // Block SIGUSR1 and save old mask
    sigemptyset(&sigset);
    sigaddset(&sigset, target_signal);
    pthread_sigmask(SIG_BLOCK, &sigset, NULL);
    printf("Blocked SIGUSR1, waiting with sigwaitinfo...\n");
    
    pid_t pid = fork();
    if (pid == 0) {
        sleep(1);
        printf("Child: sending SIGALRM to interrupt parent\n");
        kill(getppid(), SIGALRM);
        exit(0);
    }
    
    printf("Parent: calling sigwaitinfo([SIGUSR1])...\n");
    int result = sigwaitinfo(&sigset, &info);
    
    if (result == -1) {
        printf("sigwaitinfo failed: errno = %d (%s)\n", errno, strerror(errno));
    }
    
    printf("SIGALRM handler was called: %s\n", got_sigalrm ? "yes" : "no");
    
    int status;
    waitpid(pid, &status, 0);
    printf("Child exited with status: %d\n", WEXITSTATUS(status));
    
    return 0;
}

Output on NetBSD 10.0:

sigwaitinfo interruption test
Blocked SIGUSR1, waiting with sigwaitinfo...
Parent: calling sigwaitinfo([SIGUSR1])...
Child: sending SIGALRM to interrupt parent
SIGALRM handler called
sigwaitinfo failed: errno = 87 (Operation canceled)
SIGALRM handler was called: yes
Child exited with status: 0

Expected POSIX-compliant output (Linux):

sigwaitinfo interruption test
Blocked SIGUSR1, waiting with sigwaitinfo...
Parent: calling sigwaitinfo([SIGUSR1])...
Child: sending SIGALRM to interrupt parent
SIGALRM handler called
sigwaitinfo failed: errno = 4 (Interrupted system call)
SIGALRM handler was called: yes
Child exited with status: 0

CPython versions tested on:

CPython main branch, 3.15, 3.14, 3.13

Operating systems tested on:

Other

Linked PRs

• gh-137490: Skip test_sigwaitinfo on NetBSD due to POSIX violation #137491
• gh-137490: Fix signal.sigwaitinfo() on NetBSD #137523
• [3.14] gh-137490: Fix signal.sigwaitinfo() on NetBSD (GH-137523) #138935
• [3.13] gh-137490: Fix signal.sigwaitinfo() on NetBSD (GH-137523) #138936

[serhiy-storchaka]

Is returning ECANCELED on NetBSD documented anywhere?

It seems that sigtimedwait() only fails if timeout is not NULL.

Should not we handle ECANCELED the same way as EINTR for sigwaitinfo() and sigtimedwait() on NetBSD?

[serhiy-storchaka]

#137523 is an alternative solution. I am not aware of any drawbacks of it.

[furkanonder]

Is returning ECANCELED on NetBSD documented anywhere?

According to Google: site:man.netbsd.org ECANCELED, ECANCELED is mentioned for: aio_cancel, strtou, errno, aio_write, timerfd_create, strtoi.

However, NetBSD's sigwaitinfo() returns errno = ECANCELED (87) when interrupted by signals, but this behavior is completely undocumented and contradicts NetBSD's own manual page.

Comparison with FreeBSD

According to FreeBSD's sigwaitinfo():

The sigtimedwait() and sigwaitinfo() system calls fail if:
[EINTR] The wait was interrupted by an unblocked, caught signal.

According to NetBSD's sigwaitinfo():

sigwaitinfo() and sigwait() always succeed.
sigtimedwait() may also fail if:
[EAGAIN] No signal specified in set was generated in the specified timeout.
[EINVAL] The specified timeout was invalid.

Issues:

• NetBSD doesn't document that sigwaitinfo can be interrupted
• False claim: Says sigwaitinfo "always succeed" (factually incorrect)
• Missing EINTR documentation that exists in FreeBSD/POSIX

[furkanonder]

Should not we handle ECANCELED the same way as EINTR for sigwaitinfo() and sigtimedwait() on NetBSD?

Since I believe this is a NetBSD kernel and documentation error, I decided to skip the test for NetBSD instead of handling this situation.

[serhiy-storchaka]

It's not about the tests. This may be a bug in NetBSD's implementation, but what the users of programs implemented in Python should do? It is possible to write a workaround -- an explicit loop around signal.sigwaitinfo() if it raises OSError with errno=ECANCELED. But since this is NetBSD specific behavior, it will likely not be handled, unless the author tested their code on NetBSD. This leaves users with programs not working on NetBSD.

If there is other legitimate way to make sigwaitinfo() failing with errno=ECANCELED, then ECANCELED should not be ignored. Other possible workarounds are to use sigtimedwait() with very large timeout or repeat sigtimedwait() if it fails with errno=EAGAIN.

For now, I think that handling ECANCELED the same way as EINTR is safe. There is no setting ECANCELED in the NetBSD's libc, it is set in the kernel, and there is a code that turns EINTR into ECANCELED. I wonder what other sites that handle EINTR need to handle ECANCELED on NetBSD.

[serhiy-storchaka]

@furkanonder, could you please report this issue to NetBSD? Even if they not fix it, but just document as a NetBSD peculiarity, maybe they can say if repeating the sigwaitinfo() after getting ECANCELED is safe, and what other functions can set ECANCELED instead of EINTR?

[furkanonder]

@furkanonder, could you please report this issue to NetBSD? Even if they not fix it, but just document as a NetBSD peculiarity, maybe they can say if repeating the sigwaitinfo() after getting ECANCELED is safe, and what other functions can set ECANCELED instead of EINTR?

The issue has been reported to NetBSD.