replaced the select() calls by poll()

CHANGES

@@ -22,6 +22,9 @@ corrections:
 
 	improved test.sh script
 
+	replaced the select() calls by poll() to cleanly fix the problems with
+	many file descriptors already open
+
 ####################### V 1.6.0.1:
 
 new features:

VERSION

@@ -1 +1 @@
-"1.6.0.1+ip4bind+recvfromfork+x64+execstderr+execspaces+cosmetics"
+"1.6.0.1+ip4bind+recvfromfork+x64+execstderr+execspaces+cosmetics+poll"

error.h

@@ -50,6 +50,8 @@
 #define Error6(m,a1,a2,a3,a4,a5,a6) msg(E_ERROR,m,a1,a2,a3,a4,a5,a6)
 #define Error7(m,a1,a2,a3,a4,a5,a6,a7) msg(E_ERROR,m,a1,a2,a3,a4,a5,a6,a7)
 #define Error8(m,a1,a2,a3,a4,a5,a6,a7,a8) msg(E_ERROR,m,a1,a2,a3,a4,a5,a6,a7,a8)
+#define Error9(m,a1,a2,a3,a4,a5,a6,a7,a8,a9) msg(E_ERROR,m,a1,a2,a3,a4,a5,a6,a7,a8,a9)
+#define Error10(m,a1,a2,a3,a4,a5,a6,a7,a8,a9,a10) msg(E_ERROR,m,a1,a2,a3,a4,a5,a6,a7,a8,a9,a10)
 #else /* !(WITH_MSGLEVEL >= E_ERROR) */
 #define Error(m)
 #define Error1(m,a1)
@@ -60,6 +62,8 @@
 #define Error6(m,a1,a2,a3,a4,a5,a6)
 #define Error7(m,a1,a2,a3,a4,a5,a6,a7)
 #define Error8(m,a1,a2,a3,a4,a5,a6,a7,a8)
+#define Error9(m,a1,a2,a3,a4,a5,a6,a7,a8,a9)
+#define Error10(m,a1,a2,a3,a4,a5,a6,a7,a8,a9,a10)
 #endif /* !(WITH_MSGLEVEL <= E_ERROR) */
 
 #if WITH_MSGLEVEL <= E_WARN

socat.c

@@ -83,12 +83,16 @@ bool havelock;
 
 int main(int argc, const char *argv[]) {
    const char **arg1, *a;
+   char *mainwaitstring;
    char buff[10];
    double rto;
    int i, argc0, result;
    struct utsname ubuf;
    int lockrc;
 
+   if (mainwaitstring = getenv("SOCAT_MAIN_WAIT")) {
+       sleep(atoi(mainwaitstring));
+   }
    diag_set('p', strchr(argv[0], '/') ? strrchr(argv[0], '/')+1 : argv[0]);
 
    /* we must init before applying options because env settings have lower
@@ -642,37 +646,32 @@ int socat(const char *address1, const char *address2) {
    returns >0 if child died and left data
 */
 int childleftdata(xiofile_t *xfd) {
-   fd_set in, out, expt;
+   struct pollfd in;
+   int timeout = 0;	/* milliseconds */
    int retval;
+
    /* have to check if a child process died before, but left read data */
    if (XIO_READABLE(xfd) &&
        (XIO_RDSTREAM(xfd)->howtoend == END_KILL ||
 	XIO_RDSTREAM(xfd)->howtoend == END_CLOSE_KILL ||
 	XIO_RDSTREAM(xfd)->howtoend == END_SHUTDOWN_KILL) &&
        XIO_RDSTREAM(xfd)->para.exec.pid == 0) {
-      struct timeval time0 = { 0,0 };
 
-      FD_ZERO(&in); FD_ZERO(&out); FD_ZERO(&expt);
       if (XIO_READABLE(xfd) && !(XIO_RDSTREAM(xfd)->eof >= 2 && !XIO_RDSTREAM(xfd)->ignoreeof)) {
-	 FD_SET(XIO_GETRDFD(xfd), &in);
+	 in.fd = XIO_GETRDFD(xfd);
-	 /*0 FD_SET(XIO_GETRDFD(xfd), &expt);*/
+	 in.events = POLLIN/*|POLLRDBAND*/;
+	 in.revents = 0;
       }
       do {
-	 retval = Select(FD_SETSIZE, &in, &out, &expt, &time0);
+	 retval = xiopoll(&in, 1, timeout);
       } while (retval < 0 && errno == EINTR);
 
       if (retval < 0) {
-#if HAVE_FDS_BITS
+	 Error4("xiopoll({%d,%0o}, 1, %d): %s",
-	 Error5("select(%d, &0x%lx, &0x%lx, &0x%lx, {0}): %s",
+		in.fd, in.events, timeout, strerror(errno));
-		FD_SETSIZE, in.fds_bits[0], out.fds_bits[0],
-		expt.fds_bits[0], strerror(errno));
-#else
-	 Error5("select(%d, &0x%lx, &0x%lx, &0x%lx, {0}): %s",
-		FD_SETSIZE, in.__fds_bits[0], out.__fds_bits[0],
-		expt.__fds_bits[0], strerror(errno));
-#endif
 	 return -1;
-      } else if (retval == 0) {
+      }
+      if (retval == 0) {
 	 Info("terminated child did not leave data for us");
 	 XIO_RDSTREAM(xfd)->eof = 2;
 	 xfd->stream.eof = 2;
@@ -694,7 +693,11 @@ bool maywr2;		/* sock2 can be written to, according to select() */
    and their options are set/applied
    returns -1 on error or 0 on success */
 int _socat(void) {
-   fd_set in, out, expt;
+   struct pollfd fds[4],
+       *fd1in  = &fds[0],
+       *fd1out = &fds[1],
+       *fd2in  = &fds[2],
+       *fd2out = &fds[3];
    int retval;
    unsigned char *buff;
    ssize_t bytes1, bytes2;
@@ -747,7 +750,7 @@ int _socat(void) {
 	   XIO_GETRDFD(sock2), XIO_GETWRFD(sock2));
    while (XIO_RDSTREAM(sock1)->eof <= 1 ||
 	  XIO_RDSTREAM(sock2)->eof <= 1) {
-      struct timeval timeout, *to = NULL;
+      int timeout;
 
       Debug6("data loop: sock1->eof=%d, sock2->eof=%d, closing=%d, wasaction=%d, total_to={"F_tv_sec"."F_tv_usec"}",
 	     XIO_RDSTREAM(sock1)->eof, XIO_RDSTREAM(sock2)->eof,
@@ -783,59 +786,77 @@ int _socat(void) {
 
       if (polling) {
 	 /* there is a ignoreeof poll timeout, use it */
-	 timeout = socat_opts.pollintv;
+	 timeout = 1000*socat_opts.pollintv.tv_sec +
-	 to = &timeout;
+	     socat_opts.pollintv.tv_usec/1000;	/*!!! overflow?*/
       } else if (socat_opts.total_timeout.tv_sec != 0 ||
 		 socat_opts.total_timeout.tv_usec != 0) {
 	 /* there might occur a total inactivity timeout */
-	 timeout = socat_opts.total_timeout;
+	 timeout = 1000*socat_opts.total_timeout.tv_sec +
-	 to = &timeout;
+	    socat_opts.total_timeout.tv_usec/1000;
       } else {
-	 to = NULL;
+	 timeout = -1;	/* forever */
       }
 
       if (closing>=1) {
 	 /* first eof already occurred, start end timer */
-	 timeout = socat_opts.closwait;
+	 timeout = 1000*socat_opts.closwait.tv_sec +
-	 to = &timeout;
+	    socat_opts.closwait.tv_usec/1000;
 	 closing = 2;
       }
 
       do {
 	 int _errno;
-	 FD_ZERO(&in); FD_ZERO(&out); FD_ZERO(&expt);
 
 	 childleftdata(sock1);
 	 childleftdata(sock2);
 
 	 if (closing>=1) {
 	    /* first eof already occurred, start end timer */
-	    timeout = socat_opts.closwait;
+	    timeout = 1000*socat_opts.closwait.tv_sec +
-	    to = &timeout;
+	       socat_opts.closwait.tv_usec/1000;
 	    closing = 2;
 	 }
 
+	 /* now the fds are assigned */
 	 if (XIO_READABLE(sock1) &&
 	     !(XIO_RDSTREAM(sock1)->eof > 1 && !XIO_RDSTREAM(sock1)->ignoreeof) &&
 	     !socat_opts.righttoleft) {
 	    if (!mayrd1) {
-	       FD_SET(XIO_GETRDFD(sock1), &in);
+		fd1in->fd = XIO_GETRDFD(sock1);
+		fd1in->events = POLLIN;
+	    } else {
+		fd1in->fd = -1;
 	    }
 	    if (!maywr2) {
-	       FD_SET(XIO_GETWRFD(sock2), &out);
+		fd2out->fd = XIO_GETWRFD(sock2);
+		fd2out->events = POLLOUT;
+	    } else {
+		fd2out->fd = -1;
 	    }
+	 } else {
+	     fd1in->fd = -1;
+	     fd2out->fd = -1;
 	 }
 	 if (XIO_READABLE(sock2) &&
 	     !(XIO_RDSTREAM(sock2)->eof > 1 && !XIO_RDSTREAM(sock2)->ignoreeof) &&
 	     !socat_opts.lefttoright) {
 	    if (!mayrd2) {
-	       FD_SET(XIO_GETRDFD(sock2), &in);
+		fd2in->fd = XIO_GETRDFD(sock2);
+		fd2in->events = POLLIN;
+	    } else {
+		fd2in->fd = -1;
 	    }
 	    if (!maywr1) {
-	       FD_SET(XIO_GETWRFD(sock1), &out);
+		fd1out->fd = XIO_GETWRFD(sock1);
+		fd1out->events = POLLOUT;
+	    } else {
+		fd1out->fd = -1;
 	    }
+	 } else {
+	     fd1out->fd = -1;
+	     fd2in->fd = -1;
 	 }
-	 retval = Select(FD_SETSIZE, &in, &out, &expt, to);
+	 retval = xiopoll(fds, 4, timeout);
 	 _errno = errno;
 	 if (retval < 0 && errno == EINTR) {
 	    Info1("select(): %s", strerror(errno));
@@ -849,17 +870,10 @@ int _socat(void) {
 	 */
 
       if (retval < 0) {
-#if HAVE_FDS_BITS
+	 Error10("xiopoll({%d,%0o}{%d,%0o}{%d,%0o}{%d,%0o}, 4, %d): %s",
-	    Error7("select(%d, &0x%lx, &0x%lx, &0x%lx, %s%lu): %s",
+		 fds[0].fd, fds[0].events, fds[1].fd, fds[1].events,
-		   FD_SETSIZE, in.fds_bits[0], out.fds_bits[0],
+		 fds[2].fd, fds[2].events, fds[3].fd, fds[3].events,
-		   expt.fds_bits[0], to?"&":"NULL/", to?to->tv_sec:0,
+		 timeout, strerror(errno));
-		   strerror(errno));
-#else
-	    Error7("select(%d, &0x%lx, &0x%lx, &0x%lx, %s%lu): %s",
-		   FD_SETSIZE, in.__fds_bits[0], out.__fds_bits[0],
-		   expt.__fds_bits[0], to?"&":"NULL/", to?to->tv_sec:0,
-		   strerror(errno));
-#endif
 	    return -1;
       } else if (retval == 0) {
 	 Info2("select timed out (no data within %ld.%06ld seconds)",
@@ -884,17 +898,17 @@ int _socat(void) {
       }
 
       if (XIO_READABLE(sock1) && XIO_GETRDFD(sock1) >= 0 &&
-	  FD_ISSET(XIO_GETRDFD(sock1), &in)) {
+	  (fd1in->revents&(POLLIN|POLLHUP|POLLERR))) {
 	 mayrd1 = true;
       }
       if (XIO_READABLE(sock2) && XIO_GETRDFD(sock2) >= 0 &&
-	  FD_ISSET(XIO_GETRDFD(sock2), &in)) {
+	  (fd2in->revents&(POLLIN|POLLHUP|POLLERR))) {
 	 mayrd2 = true;
       }
-      if (XIO_GETWRFD(sock1) >= 0 && FD_ISSET(XIO_GETWRFD(sock1), &out)) {
+      if (XIO_GETWRFD(sock1) >= 0 && (fd1out->revents&(POLLOUT|POLLERR))) {
 	 maywr1 = true;
       }
-      if (XIO_GETWRFD(sock2) >= 0 && FD_ISSET(XIO_GETWRFD(sock2), &out)) {
+      if (XIO_GETWRFD(sock2) >= 0 && (fd2out->revents&(POLLOUT|POLLERR))) {
 	 maywr2 = true;
       }
 

sycls.c

@@ -677,9 +677,21 @@ int Chmod(const char *path, mode_t mode) {
 /* we only show the first struct pollfd; hope this is enough for most cases. */
 int Poll(struct pollfd *ufds, unsigned int nfds, int timeout) {
    int result;
-   Debug4("poll({%d, 0x%02hx, }, %u, %d)", ufds[0].fd, ufds[0].events, nfds, timeout);
+   if (nfds == 4) {
+      Debug10("poll({%d,0x%02hx,}{%d,0x%02hx,}{%d,0x%02hx,}{%d,0x%02hx,}, , %u, %d)",
+	      ufds[0].fd, ufds[0].events, ufds[1].fd, ufds[1].events,
+	      ufds[2].fd, ufds[2].events, ufds[3].fd, ufds[3].events,
+	      nfds, timeout);
+   } else {
+      Debug4("poll({%d,0x%02hx,}, , %u, %d)", ufds[0].fd, ufds[0].events, nfds, timeout);
+   }
    result = poll(ufds, nfds, timeout);
-   Debug2("poll(, {,, 0x%02hx}) -> %d", ufds[0].revents, result);
+   if (nfds == 4) {
+      Debug5("poll(, {,,0x%02hx}{,,0x%02hx}{,,0x%02hx}{,,0x%02hx}) -> %d",
+	     ufds[0].revents, ufds[1].revents, ufds[2].revents, ufds[3].revents, result);
+   } else {
+      Debug2("poll(, {,,0x%02hx}) -> %d", ufds[0].revents, result);
+   }
    return result;
 }
 #endif /* HAVE_POLL */

sysutils.c

@@ -1,5 +1,5 @@
 /* source: sysutils.c */
-/* Copyright Gerhard Rieger 2001-2007 */
+/* Copyright Gerhard Rieger 2001-2008 */
 /* Published under the GNU General Public License V.2, see file COPYING */
 
 /* translate socket addresses into human readable form */
@@ -402,6 +402,17 @@ const char *hstrerror(int err) {
 #endif /* !HAVE_HSTRERROR */
 
 
+/* this function behaves like poll(). It tries to do so even when the poll()
+   system call is not available. */
+int xiopoll(struct pollfd fds[], nfds_t nfds, int timeout) {
+#if HAVE_POLL
+    return Poll(fds, nfds, timeout);
+#else /* HAVE_POLL */
+    /*!!! wrap around Select() */
+#endif /* !HAVE_POLL */
+}
+   
+
 #if WITH_TCP || WITH_UDP
 /* returns port in network byte order */
 int parseport(const char *portname, int ipproto) {

sysutils.h

@@ -90,6 +90,8 @@ extern int getusergroups(const char *user, gid_t *list, size_t *ngroups);
 extern const char *hstrerror(int err);
 #endif
 
+extern int xiopoll(struct pollfd fds[], nfds_t nfds, int timeout);
+
 extern int parseport(const char *portname, int proto);
 
 extern int ifindexbyname(const char *ifname);

xio-socket.c

@@ -269,19 +269,21 @@ int _xioopen_connect(struct single *xfd, struct sockaddr *us, size_t uslen,
       if (errno == EINPROGRESS) {
 	 if (xfd->para.socket.connect_timeout.tv_sec  != 0 ||
 	     xfd->para.socket.connect_timeout.tv_usec != 0) {
-	    struct timeval timeout;
+	    int timeout;
-	    fd_set readfds, writefds, exceptfds;
+	    struct pollfd writefd;
 	    int result;
+
 	    Info4("connect(%d, %s, "F_Zd"): %s",
 		  xfd->fd, sockaddr_info(them, themlen, infobuff, sizeof(infobuff)),
 		  themlen, strerror(errno));
-	    timeout = xfd->para.socket.connect_timeout;
+	    timeout = 1000*xfd->para.socket.connect_timeout.tv_sec +
-	    FD_ZERO(&readfds);  FD_ZERO(&writefds);  FD_ZERO(&exceptfds);
+		xfd->para.socket.connect_timeout.tv_usec/1000;
-	    FD_SET(xfd->fd, &readfds);  FD_SET(xfd->fd, &writefds);
+	    writefd.fd = xfd->fd;
-	    result =
+	    writefd.events = (POLLIN|POLLHUP|POLLERR);
-	       Select(xfd->fd+1, &readfds, &writefds, &exceptfds, &timeout);
+	    result = Poll(&writefd, 1, timeout);
 	    if (result < 0) {
-	       Msg2(level, "select(%d,,,,): %s", xfd->fd+1, strerror(errno));
+	       Msg3(level, "poll({%d,POLLIN|POLLHUP|POLLER},,%d): %s",
+		    xfd->fd, timeout, strerror(errno));
 	       return STAT_RETRYLATER;
 	    }
 	    if (result == 0) {
@@ -290,7 +292,7 @@ int _xioopen_connect(struct single *xfd, struct sockaddr *us, size_t uslen,
 		    strerror(ETIMEDOUT));
 	       return STAT_RETRYLATER;
 	    }
-	    if (FD_ISSET(xfd->fd, &readfds)) {
+	    if (writefd.revents & POLLOUT) {
 #if 0
 	       unsigned char dummy[1];
 	       Read(xfd->fd, &dummy, 1);	/* get error message */
@@ -726,16 +728,16 @@ int _xioopen_dgram_recvfrom(struct single *xfd, int xioflags,
 
       /* loop until select() returns valid */
       do {
-	 fd_set in, out, expt;
+	 struct pollfd readfd;
 	 /*? int level = E_ERROR;*/
 	 if (us != NULL) {
 	    Notice1("receiving on %s", sockaddr_info(us, uslen, lisname, sizeof(lisname)));
 	 } else {
 	    Notice1("receiving IP protocol %u", proto);
 	 }
-	 FD_ZERO(&in); FD_ZERO(&out); FD_ZERO(&expt);
+	 readfd.fd = xfd->fd;
-	 FD_SET(xfd->fd, &in);
+	 readfd.events = POLLIN;
-	 if (Select(xfd->fd+1, &in, &out, &expt, NULL) > 0) {
+	 if (Poll(&readfd, 1, -1) > 0) {
 	    break;
 	 }
 
@@ -743,7 +745,7 @@ int _xioopen_dgram_recvfrom(struct single *xfd, int xioflags,
 	    continue;
 	 }
 
-	 Msg2(level, "select(, {%d}): %s", xfd->fd, strerror(errno));
+	 Msg2(level, "poll({%d,,},,-1): %s", xfd->fd, strerror(errno));
 	 Close(xfd->fd);
 	 return STAT_RETRYLATER;
       } while (true);

xio-udp.c

@@ -84,7 +84,7 @@ int xioopen_ipdgram_listen(int argc, const char *argv[], struct opt *opts,
    union sockaddr_union themunion;
    union sockaddr_union *them = &themunion;
    int socktype = SOCK_DGRAM;
-   fd_set in, out, expt;
+   struct pollfd readfd;
    bool dofork = false;
    pid_t pid;
    char *rangename;
@@ -205,9 +205,9 @@ int xioopen_ipdgram_listen(int argc, const char *argv[], struct opt *opts,
 
       Notice1("listening on UDP %s",
 	      sockaddr_info(&us.soa, uslen, infobuff, sizeof(infobuff)));
-      FD_ZERO(&in); FD_ZERO(&out); FD_ZERO(&expt);
+      readfd.fd = fd->stream.fd;
-      FD_SET(fd->stream.fd, &in);
+      readfd.events = POLLIN|POLLERR;
-      while (Select(fd->stream.fd+1, &in, &out, &expt, NULL) < 0) {
+      while (Poll(&readfd, 1, -1) < 0) {
 	 if (errno != EINTR)  break;
       }