Found the BSD unixmessage bug. recvmsg(..., MSG_WAITALL) blocks

until buffer full or socket shutdown on BSD, *even if the socket is
nonblocking*. So I added an "okwaitall" sysdeps, and won't set
MSG_WAITALL on retarded systems.

2 files changed, 132 insertions, 6 deletions

diff --git a/src/libunixonacid/unixmessage_receive.c b/src/libunixonacid/unixmessage_receive.c
index a86ae10..502b523 100644
--- a/src/libunixonacid/unixmessage_receive.c
+++ b/src/libunixonacid/unixmessage_receive.c
@@ -1,6 +1,5 @@
 /* ISC license. */
 
-#define _XPG4_2
 #include <skalibs/sysdeps.h>
 #include <skalibs/nonposix.h>
 #include <errno.h>
@@ -15,6 +14,20 @@
 #include <skalibs/siovec.h>
 #include <skalibs/unixmessage.h>
 
+#ifdef SKALIBS_HASOKWAITALL
+# ifdef SKALIBS_HASCMSGCLOEXEC
+#  define RECV(fd, hdr) recvmsg(fd, (hdr), MSG_WAITALL | MSG_CMSG_CLOEXEC)
+# else
+#  define RECV(fd, hdr) recvmsg(fd, (hdr), MSG_WAITALL)
+# endif
+#else
+# ifdef SKALIBS_HASCMSGCLOEXEC
+#  define RECV(fd, hdr) recvmsg(fd, (hdr), MSG_CMSG_CLOEXEC)
+# else
+#  define RECV(fd, hdr) recvmsg(fd, (hdr), 0)
+# endif
+#endif
+
 static int unixmessage_receiver_fill (unixmessage_receiver_t *b)
 {
   char ancilbuf[CMSG_SPACE(b->auxb.a - 1)] ;
@@ -38,11 +51,7 @@ static int unixmessage_receiver_fill (unixmessage_receiver_t *b)
     cbuffer_wpeek(&b->mainb, v) ;
     iovec_from_siovec(iov, v, 2) ;
   }
-#ifdef SKALIBS_HASCMSGCLOEXEC
-  r = recvmsg(b->fd, &msghdr, MSG_WAITALL | MSG_CMSG_CLOEXEC) ;
-#else
-  r = recvmsg(b->fd, &msghdr, MSG_WAITALL) ;
-#endif
+  r = RECV(b->fd, &msghdr) ;
   if (r <= 0) return r ;
   {
     struct cmsghdr *c = CMSG_FIRSTHDR(&msghdr) ;
diff --git a/src/sysdeps/tryokwaitall.c b/src/sysdeps/tryokwaitall.c
new file mode 100644
index 0000000..15721b5
--- /dev/null
+++ b/src/sysdeps/tryokwaitall.c
@@ -0,0 +1,117 @@
+/* ISC license. */
+
+#undef _POSIX_C_SOURCE
+#undef _XOPEN_SOURCE
+
+#ifndef _XPG4_2
+# define _XPG4_2
+#endif
+
+#ifndef _BSD_SOURCE
+#define _BSD_SOURCE
+#endif
+
+#include <sys/types.h>
+#include <fcntl.h>
+#include <unistd.h>
+#include <errno.h>
+#include <signal.h>
+#include <sys/socket.h>
+#include <sys/select.h>
+#include <sys/uio.h>
+#if defined(__FreeBSD__)
+#include <sys/param.h>
+#endif
+
+#ifndef MSG_NOSIGNAL
+#define MSG_NOSIGNAL 0
+#endif
+
+static int ndelay_on (int fd)
+{
+  register int got = fcntl(fd, F_GETFL) ;
+  return (got == -1) ? -1 : fcntl(fd, F_SETFL, got | O_NONBLOCK) ;
+}
+
+static int child (int p, int fd)
+{
+  char c ;
+  struct iovec v = { .iov_base = &c, .iov_len = 1 } ;
+  struct msghdr msg =
+  {
+    .msg_name = 0,
+    .msg_namelen = 0,
+    .msg_iov = &v,
+    .msg_iovlen = 1,
+    .msg_control = 0,
+    .msg_controllen = 0
+  } ;
+  fd_set rfds ;
+  if (ndelay_on(fd) < 0) return 111 ;
+  FD_ZERO(&rfds) ;
+  FD_SET(fd, &rfds) ;
+  if (write(p, "", 1) < 1) return 111 ; /* sync with the parent */
+  if (select(fd+1, &rfds, 0, 0, 0) < 1) return 111 ;
+
+  /* On buggy systems, the following recvmsg will block, despite
+     setting the fd non-blocking */
+
+  if (recvmsg(fd, &msg, MSG_WAITALL) < 1) return 111 ;
+  if (write(p, "", 1) < 1) return 111 ;
+  return 0 ;
+}
+
+static int parent (pid_t pid, int p, int fd)
+{
+  char c ;
+  struct iovec v = { .iov_base = &c, .iov_len = 1 } ;
+  struct msghdr msg =
+  {
+    .msg_name = 0,
+    .msg_namelen = 0,
+    .msg_iov = &v,
+    .msg_iovlen = 1,
+    .msg_control = 0,
+    .msg_controllen = 0
+  } ;
+  struct timeval tv = { .tv_sec = 2, .tv_usec = 0 } ;
+  fd_set rfds ;
+  unsigned int n = p > fd ? p : fd ;
+  int r ;
+  FD_ZERO(&rfds) ;
+  FD_SET(p, &rfds) ;
+  if (read(p, &c, 1) < 1) return 111 ;
+  if (sendmsg(fd, &msg, MSG_NOSIGNAL) < 1) return 111 ;
+  for (;;)
+  {
+    r = select(n+1, &rfds, 0, 0, &tv) ;
+    if (r >= 0 || errno != EINTR) break ;
+  }
+  if (!r) /* child is still blocking on recvmsg() after 2 seconds */
+  {
+    kill(pid, SIGTERM) ;
+    return 1 ;
+  }
+  if (read(p, &c, 1) < 1) return 111 ;
+  return 0 ;
+}
+
+int main (void)
+{
+  pid_t pid ;
+  int fd[2] ;
+  int p[2] ;
+  if (pipe(p) < 0) return 111 ;
+  if (socketpair(AF_UNIX, SOCK_STREAM, 0, fd) < 0) return 111 ;
+  pid = fork() ;
+  if (pid < 0) return 111 ;
+  if (!pid)
+  {
+    close(p[0]) ;
+    close(fd[0]) ;
+    return child(p[1], fd[1]) ;
+  }
+  close(p[1]) ;
+  close(fd[1]) ;
+  return parent(pid, p[0], fd[0]) ;
+}