Fixed hang on race condition with UDP-RECV and fork

CHANGES

@@ -27,6 +27,13 @@ Corrections:
 	not writing complete pages.
 	Test: O_DIRECT
 
+	There was a race condition in the way Socat UDP-RECVFROM and similar
+	addresses with option fork prevents one packet from triggering
+	multiple processes. The symptom was that Socat master process seemed to
+	hang and did not process further packets. The fix makes use of
+	pselect() system call.
+	Thanks to Fulvio Scapin for reporting this issue.
+
 Porting:
 	In gcc version 10 the default changed from -fcommon to -fno-common.
 	Consequently, linking filan and procan failed with error

compat.h

@@ -626,6 +626,33 @@ typedef int sig_atomic_t;
 #  endif
 #endif
 
+/* default: long */
+#if !defined(HAVE_TYPEOF_STRUCT_TIMESPEC_TV_NSEC) || !HAVE_TYPEOF_STRUCT_TIMESPEC_TV_NSEC
+#  undef HAVE_TYPEOF_STRUCT_TIMESPEC_TV_NSEC
+#  define HAVE_TYPEOF_STRUCT_TIMESPEC_TV_NSEC 5
+#endif
+#ifndef F_tv_nsec
+#  if HAVE_TYPEOF_STRUCT_TIMESPEC_TV_NSEC==1
+#define F_tv_nsec "%09hd"
+#  elif HAVE_TYPEOF_STRUCT_TIMESPEC_TV_NSEC==2
+#define F_tv_nsec "%09hu"
+#  elif HAVE_TYPEOF_STRUCT_TIMESPEC_TV_NSEC==3
+#define F_tv_nsec "%09d"
+#  elif HAVE_TYPEOF_STRUCT_TIMESPEC_TV_NSEC==4
+#define F_tv_nsec "%09u"
+#  elif HAVE_TYPEOF_STRUCT_TIMESPEC_TV_NSEC==5
+#define F_tv_nsec "%09ld"
+#  elif HAVE_TYPEOF_STRUCT_TIMESPEC_TV_NSEC==6
+#define F_tv_nsec "%09lu"
+#  elif HAVE_TYPEOF_STRUCT_TIMESPEC_TV_NSEC==7
+#define F_tv_nsec "%09Ld"
+#  elif HAVE_TYPEOF_STRUCT_TIMESPEC_TV_NSEC==8
+#define F_tv_nsec "%09Lu"
+#  else
+#error "HAVE_TYPEOF_STRUCT_TIMESPEC_TV_NSEC is out of range:" HAVE_TYPEOF_STRUCT_TIMESPEC_TV_NSEC
+#  endif
+#endif
+
 /* default: long */
 #if !defined(HAVE_TYPEOF_RLIM_MAX) || !HAVE_TYPEOF_RLIM_MAX
 #  undef HAVE_TYPEOF_RLIM_MAX
@@ -653,6 +680,10 @@ typedef int sig_atomic_t;
 #  endif
 #endif
 
+/* sigset_t printing - not an exact solution yet */
+#define F_sigset "0x%4lx"
+typedef unsigned long T_sigset;
+
 /* default: socklen_t */
 #if !defined(HAVE_TYPEOF_STRUCT_CMSGHDR_CMSG_LEN) || !HAVE_TYPEOF_STRUCT_CMSGHDR_CMSG_LEN
 #  undef HAVE_TYPEOF_STRUCT_CMSGHDR_CMSG_LEN

config.h.in

@@ -63,6 +63,9 @@
 /* Define if you have the select function.  */
 #undef HAVE_SELECT
 
+/* Define if you have the pselect function.  */
+#undef HAVE_PSELECT
+
 /* Define if you have the poll function.  */
 #undef HAVE_POLL
 
@@ -599,6 +602,7 @@
 #undef HAVE_TYPEOF_ST64_BLOCKS
 
 #undef HAVE_TYPEOF_STRUCT_TIMEVAL_TV_USEC
+#undef HAVE_TYPEOF_STRUCT_TIMESPEC_TV_NSEC
 
 #undef HAVE_TYPEOF_RLIM_MAX
 

configure.ac

@@ -769,7 +769,7 @@ AC_PROG_GCC_TRADITIONAL
 AC_FUNC_MEMCMP
 AC_TYPE_SIGNAL
 AC_FUNC_STRFTIME
-AC_CHECK_FUNCS(putenv select poll socket strtod strtol)
+AC_CHECK_FUNCS(putenv select pselect poll socket strtod strtol)
 AC_CHECK_FUNCS(strtoul uname getpgid getsid gethostbyname getaddrinfo)
 AC_CHECK_FUNCS(setgroups inet_aton)
 AC_CHECK_FUNCS()
@@ -1851,6 +1851,8 @@ fi
 
 AC_TYPEOF_COMPONENT([#include <sys/time.h>], struct timeval, tv_usec, HAVE_TYPEOF_STRUCT_TIMEVAL_TV_USEC, sc_cv_type_struct_timeval_tv_usec)
 
+AC_TYPEOF_COMPONENT([#include <sys/time.h>], struct timespec, tv_nsec, HAVE_TYPEOF_STRUCT_TIMESPEC_TV_NSEC, sc_cv_type_struct_timespec_tv_nsec)
+
 AC_TYPEOF_COMPONENT([#include <sys/types.h>
 #include <sys/time.h>
 #include <sys/resource.h>],

sycls.c

@@ -858,6 +858,50 @@ int Select(int n, fd_set *readfds, fd_set *writefds, fd_set *exceptfds,
    return result;
 }
 
+#if HAVE_PSELECT
+/* we only show the first word of the fd_set's; hope this is enough for most
+   cases. */
+int Pselect(int n, fd_set *readfds, fd_set *writefds, fd_set *exceptfds,
+	    const struct timespec *timeout, const sigset_t *sigmask) {
+   int result, _errno;
+   if (!diag_in_handler) diag_flush();
+#if WITH_SYCLS
+#if HAVE_FDS_BITS
+   Debug8("pselect(%d, &0x%lx, &0x%lx, &0x%lx, %s%lu."F_tv_nsec", "F_sigset")",
+	  n, readfds?readfds->fds_bits[0]:0, writefds?writefds->fds_bits[0]:0,
+	  exceptfds?exceptfds->fds_bits[0]:0,
+	  timeout?"&":"NULL/", timeout?timeout->tv_sec:0,
+	  timeout?timeout->tv_nsec:0, *(T_sigset *)sigmask);
+#else
+   Debug8("pselect(%d, &0x%lx, &0x%lx, &0x%lx, %s%lu.%06u)",
+	  n, readfds?readfds->__fds_bits[0]:0, writefds?writefds->__fds_bits[0]:0,
+	  exceptfds?exceptfds->__fds_bits[0]:0,
+	  timeout?"&":"NULL/", timeout?timeout->tv_sec:0,
+	  timeout?timeout->tv_nsec:0);
+#endif
+#endif /* WITH_SYCLS */
+   result = pselect(n, readfds, writefds, exceptfds, timeout, sigmask);
+   _errno = errno;
+   if (!diag_in_handler) diag_flush();
+#if WITH_SYCLS
+#if HAVE_FDS_BITS
+   Debug5("pselect -> (, 0x%lx, 0x%lx, 0x%lx), "F_sigset", %d",
+	  readfds?readfds->fds_bits[0]:0, writefds?writefds->fds_bits[0]:0,
+	  exceptfds?exceptfds->fds_bits[0]:0, *(T_sigset *)sigmask,
+	  result);
+#else
+   Debug6("pselect -> (, 0x%lx, 0x%lx, 0x%lx), %d",
+	  readfds?readfds->__fds_bits[0]:0, writefds?writefds->__fds_bits[0]:0,
+	  exceptfds?exceptfds->__fds_bits[0]:0,
+	  result);
+#endif
+#endif /* WITH_SYCLS */
+   errno = _errno;
+
+   return result;
+}
+#endif /* HAVE_PSELECT */
+
 #if WITH_SYCLS
 
 pid_t Fork(void) {

sycls.h

@@ -85,6 +85,8 @@ int Chmod(const char *path, mode_t mode);
 int Poll(struct pollfd *ufds, unsigned int nfds, int timeout);
 int Select(int n, fd_set *readfds, fd_set *writefds, fd_set *exceptfds,
 	   struct timeval *timeout);
+int Pselect(int n, fd_set *readfds, fd_set *writefds, fd_set *exceptfds,
+	    const struct timespec *timeout, const sigset_t *sigmask);
 #if WITH_SYCLS
 pid_t Fork(void);
 #endif /* WITH_SYCLS */

xio-socket.c

@@ -1389,27 +1389,27 @@ int _xioopen_dgram_recvfrom(struct single *xfd, int xioflags,
       xfd->salen = palen;
 
       if (dofork) {
-	 sigset_t mask_sigchldusr1;
+	 sigset_t oldset, mask_sigchldusr1;
 
 	 /* we must prevent that the current packet triggers another fork;
 	    therefore we wait for a signal from the recent child: USR1
 	    indicates that is has consumed the last packet; CHLD means it has
 	    terminated */
 	 /* block SIGCHLD and SIGUSR1 until parent is ready to react */
-	 sigemptyset(&mask_sigchldusr1);
+	 Sigprocmask(SIG_BLOCK, NULL, &mask_sigchldusr1);
 	 sigaddset(&mask_sigchldusr1, SIGCHLD);
 	 sigaddset(&mask_sigchldusr1, SIGUSR1);
-	 Sigprocmask(SIG_BLOCK, &mask_sigchldusr1, NULL);
+	 Sigprocmask(SIG_SETMASK, &mask_sigchldusr1, &oldset);
 
 	 if ((pid = xio_fork(false, level)) < 0) {
 	    Close(xfd->fd);
-	    Sigprocmask(SIG_UNBLOCK, &mask_sigchldusr1, NULL);
+	    Sigprocmask(SIG_SETMASK, &oldset, NULL);
 	    return STAT_RETRYLATER;
 	 }
 
 	 if (pid == 0) {	/* child */
 	    /* no reason to block SIGCHLD in child process */
-	    Sigprocmask(SIG_UNBLOCK, &mask_sigchldusr1, NULL);
+	    Sigprocmask(SIG_SETMASK, &oldset, NULL);
 	    xfd->ppid = Getppid();	/* send parent a signal when packet has
 					   been consumed */
 
@@ -1431,12 +1431,21 @@ int _xioopen_dgram_recvfrom(struct single *xfd, int xioflags,
 
 	 /* server: continue loop with listen */
 	 xio_waitingfor = pid;
+
+#if HAVE_PSELECT
+	 {
+	    struct timespec timeout = { LONG_MAX, 0 };
+	    Pselect(0, NULL, NULL, NULL, &timeout, &oldset);
+	    Sigprocmask(SIG_SETMASK, &oldset, NULL);
+	 }
+#else /* ! HAVE_PSELECT */
 	 /* now we are ready to handle signals */
-	 Sigprocmask(SIG_UNBLOCK, &mask_sigchldusr1, NULL);
+	 Sigprocmask(SIG_SETMASK, &oldset, NULL);
 
 	 while (!xio_hashappened) {
-	    Sleep(UINT_MAX);	/* any signal lets us continue */
+	    Sleep(1);	/* any signal speeds up return */
 	 }
+#endif /* ! HAVE_PSELECT */
 	 xio_waitingfor = 0;	/* so this child will not set hashappened again */
 	 xio_hashappened = false;