Merge branch 'master' into socket

[doldaconnect.git] / daemon / net.c

diff --git a/daemon/net.c b/daemon/net.c
index a8059d8..240855a 100644 (file)
--- a/daemon/net.c
+++ b/daemon/net.c
@@ -66,8 +66,6 @@ static struct configvar myvars[] =
      * and net.visibleipv4 are unspecified the address of the hub
      * connection is used. */
     {CONF_VAR_STRING, "publicif", {.str = L""}},
-    /* Diffserv should be supported on IPv4, too, but I don't know the
-     * API to do that. */
     /** The Diffserv value to use on IPv6 connections when the
      * minimize cost TOS value is used (see the TOS VALUES
      * section). */
@@ -84,6 +82,12 @@ static struct configvar myvars[] =
      * minimize delay TOS value is used (see the TOS VALUES
      * section). */
     {CONF_VAR_INT, "diffserv-mindelay", {.num = 0}},
+    /** If enabled, the IP TOS interface will be used to set Diffserv
+     * codepoints on IPv4 sockets, by shifting the DSCP value two bits
+     * to the left (remember, the DSCP field in the IPv4 header is
+     * defined as the 6 uppermost bits of the TOS field, the lower two
+     * being left for ECN). This may only work on Linux. */
+    {CONF_VAR_BOOL, "dscp-tos", {.num = 0}},
     {CONF_VAR_END}
 };
 
@@ -211,6 +215,7 @@ static struct socket *newsock1(int dgram)
     new->refcount = 1;
     new->state = -1;
     new->dgram = dgram;
+    new->maxbuf = 65536;
     numsocks++;
     return(new);
 }
@@ -704,14 +709,16 @@ static void sockrecv(struct ufd *ufd)
     }
 }
 
-static void sockflush(struct ufd *ufd)
+static int sockflush(struct ufd *ufd)
 {
     int ret;
     struct dgrambuf *dbuf;
     int dgram;
     
-    if((dgram = ufddgram(ufd)) < 0)
-       return;
+    if((dgram = ufddgram(ufd)) < 0) {
+       errno = EBADFD;
+       return(-1);
+    }
     if(dgram) {
        dbuf = sockgetdgbuf(ufd->sk);
        sendto(ufd->fd, dbuf->data, dbuf->size, MSG_DONTWAIT | MSG_NOSIGNAL, dbuf->addr, dbuf->addrlen);
@@ -721,16 +728,14 @@ static void sockflush(struct ufd *ufd)
            ret = send(ufd->fd, ufd->sk->buf.s.buf, ufd->sk->buf.s.datasize, MSG_DONTWAIT | MSG_NOSIGNAL);
        else
            ret = write(ufd->fd, ufd->sk->buf.s.buf, ufd->sk->buf.s.datasize);
-       if(ret < 0) {
-           /* For now, assume transient error, since
-            * the socket is polled for errors */
-           return;
-       }
+       if(ret < 0)
+           return(-1);
        if(ret > 0) {
            memmove(ufd->sk->buf.s.buf, ((char *)ufd->sk->buf.s.buf) + ret, ufd->sk->buf.s.datasize -= ret);
            sockread(ufd->sk);
        }
     }
+    return(0);
 }
 
 void closesock(struct socket *sk)
@@ -756,9 +761,14 @@ size_t sockgetdatalen(struct socket *sk)
     return(ret);
 }
 
-size_t sockqueuesize(struct socket *sk)
+/* size_t sockqueuesize(struct socket *sk) */
+/* { */
+/*     return(sockgetdatalen(sk->back)); */
+/* } */
+
+ssize_t sockqueueleft(struct socket *sk)
 {
-    return(sockgetdatalen(sk->back));
+    return(sk->back->maxbuf - sockgetdatalen(sk->back));
 }
 
 /*
@@ -975,6 +985,7 @@ static void runbatches(void)
        nsc = sc->n;
        if(sc->s->conncb != NULL)
            sc->s->conncb(sc->s, 0, sc->s->data);
+       putsock(sc->s);
        free(sc);
     }
     for(sc = rbatch, rbatch = NULL; sc; sc = nsc) {
@@ -986,12 +997,14 @@ static void runbatches(void)
                sc->s->errcb(sc->s, 0, sc->s->data);
            sc->s->eos = 2;
        }
+       putsock(sc->s);
        free(sc);
     }
     for(sc = wbatch, wbatch = NULL; sc; sc = nsc) {
        nsc = sc->n;
        if(sc->s->writecb != NULL)
            sc->s->writecb(sc->s, sc->s->data);
+       putsock(sc->s);
        free(sc);
     }
 }
@@ -1002,7 +1015,7 @@ static void cleansocks(void)
     
     for(ufd = ufds; ufd != NULL; ufd = next) {
        next = ufd->next;
-       if(ufd->sk && (sockgetdatalen(ufd->sk) == 0)) {
+       if(ufd->sk && ((ufd->fd < 0) || (sockgetdatalen(ufd->sk) == 0))) {
            if(ufd->sk->eos == 1) {
                ufd->sk->eos = 2;
                closeufd(ufd);
@@ -1035,7 +1048,7 @@ int pollsocks(int timeout)
     for(maxfd = 0, ufd = ufds; ufd != NULL; ufd = ufd->next) {
        if(ufd->fd < 0)
            continue;
-       if(!ufd->ignread)
+       if(!ufd->ignread && ((ufd->sk == NULL) || (sockqueueleft(ufd->sk) > 0)))
            FD_SET(ufd->fd, &rfds);
        if(ufd->sk != NULL) {
            if(sockgetdatalen(ufd->sk) > 0)
@@ -1116,8 +1129,13 @@ int pollsocks(int timeout)
                    sockrecv(ufd);
                if(ufd->fd == -1)
                    continue;
-               if(FD_ISSET(ufd->fd, &wfds))
-                   sockflush(ufd);
+               if(FD_ISSET(ufd->fd, &wfds)) {
+                   if(sockflush(ufd)) {
+                       sockerror(ufd->sk, errno);
+                       closeufd(ufd);
+                       continue;
+                   }
+               }
            }
        }
     }
@@ -1141,6 +1159,7 @@ int socksettos(struct socket *sk, int tos)
 {
     int buf;
     struct ufd *ufd;
+    int dscp2tos;
     
     ufd = getskufd(sk);
     if(ufd->type != UFD_SOCK) {
@@ -1151,22 +1170,35 @@ int socksettos(struct socket *sk, int tos)
        return(0); /* Unix sockets are always perfect. :) */
     if(ufd->d.s.family == AF_INET)
     {
+       dscp2tos = confgetint("net", "dscp-tos");
        switch(tos)
        {
        case 0:
            buf = 0;
            break;
        case SOCK_TOS_MINCOST:
-           buf = 0x02;
+           if(dscp2tos)
+               buf = confgetint("net", "diffserv-mincost") << 2;
+           else
+               buf = 0x02;
            break;
        case SOCK_TOS_MAXREL:
-           buf = 0x04;
+           if(dscp2tos)
+               buf = confgetint("net", "diffserv-maxrel") << 2;
+           else
+               buf = 0x04;
            break;
        case SOCK_TOS_MAXTP:
-           buf = 0x08;
+           if(dscp2tos)
+               buf = confgetint("net", "diffserv-maxtp") << 2;
+           else
+               buf = 0x08;
            break;
        case SOCK_TOS_MINDELAY:
-           buf = 0x10;
+           if(dscp2tos)
+               buf = confgetint("net", "diffserv-mindelay") << 2;
+           else
+               buf = 0x10;
            break;
        default:
            flog(LOG_WARNING, "attempted to set unknown TOS value %i to IPv4 sock", tos);
@@ -1485,9 +1517,9 @@ int lstgetremotename2(struct lport *lp, struct socket *sk2, struct sockaddr **na
        errno = EOPNOTSUPP;
        return(-1);
     }
-    if(ufd1->d.s.family != ufd2->d.s.family)
+    if(ufd1->d.l.family != ufd2->d.s.family)
     {
-       flog(LOG_ERR, "using lstgetremotename2 with sockets of differing family: %i %i", ufd1->d.s.family, ufd2->d.s.family);
+       flog(LOG_ERR, "using lstgetremotename2 with sockets of differing family: %i %i", ufd1->d.l.family, ufd2->d.s.family);
        return(-1);
     }
     if(getremotename(ufd1->fd, &name1, &len1))
@@ -1521,13 +1553,13 @@ int getucred(struct socket *sk, uid_t *uid, gid_t *gid)
     return(0);
 }
 
-void sockblock(struct socket *sk, int block)
-{
-    struct ufd *ufd;
+/* void sockblock(struct socket *sk, int block) */
+/* { */
+/*     struct ufd *ufd; */
     
-    ufd = getskufd(sk);
-    ufd->ignread = block;
-}
+/*     ufd = getskufd(sk); */
+/*     ufd->ignread = block; */
+/* } */
 
 int sockfamily(struct socket *sk)
 {