Implement emulator netns support for TCP and UDP

5 files changed, 133 insertions, 15 deletions

diff --git a/erts/configure.in b/erts/configure.in
index b056ba44e2..236c7075f2 100644
--- a/erts/configure.in
+++ b/erts/configure.in
@@ -1663,6 +1663,10 @@ if test x"$ac_cv_header_netinet_sctp_h" = x"yes"; then
         ])
 fi
 
+dnl Check for setns
+AC_CHECK_HEADERS(sched.h setns.h)
+AC_CHECK_FUNCS([setns])
+
 HAVE_VALGRIND=no
 AC_CHECK_HEADER(valgrind/valgrind.h, HAVE_VALGRIND=yes)
 AC_SUBST(HAVE_VALGRIND)
diff --git a/erts/emulator/drivers/common/inet_drv.c b/erts/emulator/drivers/common/inet_drv.c
index 301ce2d0e2..9ef2a1cfc0 100644
--- a/erts/emulator/drivers/common/inet_drv.c
+++ b/erts/emulator/drivers/common/inet_drv.c
@@ -282,7 +282,7 @@ static BOOL (WINAPI *fpSetHandleInformation)(HANDLE,DWORD,DWORD);
 static unsigned long zero_value = 0;
 static unsigned long one_value = 1;
 
-#else
+#else /* #ifdef __WIN32__ */
 
 #include <sys/time.h>
 #ifdef NETDB_H_NEEDS_IN_H
@@ -315,9 +315,17 @@ static unsigned long one_value = 1;
 
 #include <net/if.h>
 
+#ifdef HAVE_SCHED_H
+#include <sched.h>
+#endif
+
+#ifdef HAVE_SETNS_H
+#include <setns.h>
+#endif
+
 /* SCTP support -- currently for UNIX platforms only: */
 #undef HAVE_SCTP
-#if (!defined(__WIN32__) && defined(HAVE_SCTP_H))
+#if defined(HAVE_SCTP_H)
 
 #include <netinet/sctp.h>
 
@@ -418,7 +426,7 @@ static int (*p_sctp_bindx)(int sd, struct sockaddr *addrs,
 static int (*p_sctp_peeloff)(int sd, sctp_assoc_t assoc_id) = NULL;
 #endif
 
-#endif /* SCTP supported */
+#endif /* #if defined(HAVE_SCTP_H) */
 
 #ifndef WANT_NONBLOCKING
 #define WANT_NONBLOCKING
@@ -512,7 +520,7 @@ static int my_strncasecmp(const char *s1, const char *s2, size_t n)
    } while(0)
 
 
-#endif /* __WIN32__ */
+#endif /* #ifdef __WIN32__    #else */
 
 #ifdef HAVE_SOCKLEN_T
 #  define SOCKLEN_T socklen_t
@@ -680,6 +688,7 @@ static int my_strncasecmp(const char *s1, const char *s2, size_t n)
 #define INET_LOPT_TCP_SEND_TIMEOUT_CLOSE 35  /* auto-close on send timeout or not */
 #define INET_LOPT_MSGQ_HIWTRMRK     36  /* set local msgq high watermark */
 #define INET_LOPT_MSGQ_LOWTRMRK     37  /* set local msgq low watermark */
+#define INET_LOPT_NETNS             38  /* Network namespace pathname */
 /* SCTP options: a separate range, from 100: */
 #define SCTP_OPT_RTOINFO		100
 #define SCTP_OPT_ASSOCINFO		101
@@ -955,6 +964,10 @@ typedef struct {
     int is_ignored;             /* if a fd is ignored by the inet_drv.
 				   This flag should be set to true when
 				   the fd is used outside of inet_drv. */
+#ifdef HAVE_SETNS
+    char *netns;                /* Socket network namespace name
+				   as full file path */
+#endif
 } inet_descriptor;
 
 
@@ -3908,10 +3921,57 @@ static int erl_inet_close(inet_descriptor* desc)
 static ErlDrvSSizeT inet_ctl_open(inet_descriptor* desc, int domain, int type,
 				  char** rbuf, ErlDrvSizeT rsize)
 {
+#ifdef HAVE_SETNS
+    int current_ns, new_ns, save_errno;
+    current_ns = new_ns = save_errno = 0;
+#endif
     if (desc->state != INET_STATE_CLOSED)
 	return ctl_xerror(EXBADSEQ, rbuf, rsize);
+#ifdef HAVE_SETNS
+    if (desc->netns != NULL) {
+	current_ns = open("/proc/self/ns/net", O_RDONLY);
+	if (current_ns == INVALID_SOCKET)
+	    return ctl_error(sock_errno(), rbuf, rsize);
+	new_ns = open(desc->netns, O_RDONLY);
+	if (new_ns == INVALID_SOCKET) {
+	    save_errno = sock_errno();
+	    while (close(current_ns) == INVALID_SOCKET &&
+		   sock_errno() == EINTR);
+	    return ctl_error(save_errno, rbuf, rsize);
+	}
+	if (setns(new_ns, CLONE_NEWNET) != 0) {
+	    save_errno = sock_errno();
+	    while (close(new_ns) == INVALID_SOCKET &&
+		   sock_errno() == EINTR);
+	    while (close(current_ns) == INVALID_SOCKET &&
+		   sock_errno() == EINTR);
+	    return ctl_error(save_errno, rbuf, rsize);
+	}
+	else {
+	    while (close(new_ns) == INVALID_SOCKET &&
+		   sock_errno() == EINTR);
+	}
+    }
+#endif
     if ((desc->s = sock_open(domain, type, desc->sprotocol)) == INVALID_SOCKET)
 	return ctl_error(sock_errno(), rbuf, rsize);
+#ifdef HAVE_SETNS
+    if (desc->netns != NULL) {
+	if (setns(current_ns, CLONE_NEWNET) != 0) {
+	    save_errno = sock_errno();
+	    while (close(desc->s) == INVALID_SOCKET &&
+		   sock_errno() == EINTR);
+	    desc->s = INVALID_SOCKET;
+	    while (close(current_ns) == INVALID_SOCKET &&
+		   sock_errno() == EINTR);
+	    return ctl_error(save_errno, rbuf, rsize);
+	}
+	else {
+	    while (close(current_ns) == INVALID_SOCKET &&
+		   sock_errno() == EINTR);
+	}
+    }
+#endif
     if ((desc->event = sock_create_event(desc)) == INVALID_EVENT)
 	return ctl_error(sock_errno(), rbuf, rsize);
     SET_NONBLOCKING(desc->s);
@@ -5529,6 +5589,20 @@ static int inet_set_opts(inet_descriptor* desc, char* ptr, int len)
 	    }
 	    continue;
 
+#ifdef HAVE_SETNS
+	case INET_LOPT_NETNS:
+	    /* It is annoying that ival and len are both (signed) int */
+	    if (ival < 0) return -1;
+	    if (len < ival) return -1;
+	    if (desc->netns != NULL) FREE(desc->netns);
+	    desc->netns = ALLOC(((unsigned int) ival) + 1);
+	    memcpy(desc->netns, ptr, ival);
+	    desc->netns[ival] = '\0';
+	    ptr += ival;
+	    len -= ival;
+	    continue;
+#endif
+
 	case INET_OPT_REUSEADDR: 
 #ifdef __WIN32__
 	    continue;  /* Bjorn says */
@@ -6454,6 +6528,22 @@ static ErlDrvSSizeT inet_fill_opts(inet_descriptor* desc,
 	    }
 	    continue;
 
+#ifdef HAVE_SETNS
+	case INET_LOPT_NETNS:
+	    if (desc->netns != NULL) {
+		size_t netns_len;
+		netns_len = strlen(desc->netns);
+		*ptr++ = opt;
+		put_int32(netns_len, ptr);
+		PLACE_FOR(netns_len, ptr);
+		memcpy(ptr, desc->netns, netns_len);
+		ptr += netns_len;
+	    } else {
+		TRUNCATE_TO(0,ptr);
+	    }
+	    continue;
+#endif
+
 	case INET_OPT_PRIORITY:
 #ifdef SO_PRIORITY
 	    type = SO_PRIORITY;
@@ -7458,6 +7548,10 @@ static ErlDrvSSizeT inet_subscribe(inet_descriptor* desc,
 static void inet_stop(inet_descriptor* desc)
 {
     erl_inet_close(desc);
+#ifdef HAVE_SETNS
+    if (desc->netns != NULL)
+	FREE(desc->netns);
+#endif
     FREE(desc);
 }
 
@@ -7537,6 +7631,10 @@ static ErlDrvData inet_start(ErlDrvPort port, int size, int protocol)
 
     desc->is_ignored = 0;
 
+#ifdef HAVE_SETNS
+    desc->netns = NULL;
+#endif
+
     return (ErlDrvData)desc;
 }
 
diff --git a/erts/preloaded/ebin/prim_inet.beam b/erts/preloaded/ebin/prim_inet.beam
index 8638ef677e..5b38871282 100644
--- a/erts/preloaded/ebin/prim_inet.beam
+++ b/erts/preloaded/ebin/prim_inet.beam
diff --git a/erts/preloaded/src/prim_inet.erl b/erts/preloaded/src/prim_inet.erl
index fb1269cf91..fa621681f3 100644
--- a/erts/preloaded/src/prim_inet.erl
+++ b/erts/preloaded/src/prim_inet.erl
@@ -25,7 +25,7 @@
 
 %% Primitive inet_drv interface
 
--export([open/3, fdopen/4, close/1]).
+-export([open/3, open/4, fdopen/4, close/1]).
 -export([bind/3, listen/1, listen/2, peeloff/2]).
 -export([connect/3, connect/4, async_connect/4]).
 -export([accept/1, accept/2, async_accept/2]).
@@ -64,22 +64,31 @@
 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
 
 open(Protocol, Family, Type) ->
-    open(Protocol, Family, Type, ?INET_REQ_OPEN, []).
+    open(Protocol, Family, Type, [], ?INET_REQ_OPEN, []).
+
+open(Protocol, Family, Type, Opts) ->
+    open(Protocol, Family, Type, Opts, ?INET_REQ_OPEN, []).
 
 fdopen(Protocol, Family, Type, Fd) when is_integer(Fd) ->
-    open(Protocol, Family, Type, ?INET_REQ_FDOPEN, ?int32(Fd)).
+    open(Protocol, Family, Type, [], ?INET_REQ_FDOPEN, ?int32(Fd)).
 
-open(Protocol, Family, Type, Req, Data) ->
+open(Protocol, Family, Type, Opts, Req, Data) ->
     Drv = protocol2drv(Protocol),
     AF = enc_family(Family),
     T = enc_type(Type),
     try erlang:open_port({spawn_driver,Drv}, [binary]) of
 	S ->
-	    case ctl_cmd(S, Req, [AF,T,Data]) of
-		{ok,_} -> {ok,S};
-		{error,_}=Error ->
+	    case setopts(S, Opts) of
+		ok ->
+		    case ctl_cmd(S, Req, [AF,T,Data]) of
+			{ok,_} -> {ok,S};
+			{error,_}=E1 ->
+			    close(S),
+			    E1
+		    end;
+		{error,_}=E2 ->
 		    close(S),
-		    Error
+		    E2
 	    end
     catch
 	%% The only (?) way to get here is to try to open
@@ -1108,6 +1117,7 @@ enc_opt(send_timeout_close) -> ?INET_LOPT_TCP_SEND_TIMEOUT_CLOSE;
 enc_opt(delay_send)      -> ?INET_LOPT_TCP_DELAY_SEND;
 enc_opt(packet_size)     -> ?INET_LOPT_PACKET_SIZE;
 enc_opt(read_packets)    -> ?INET_LOPT_READ_PACKETS;
+enc_opt(netns)           -> ?INET_LOPT_NETNS;
 enc_opt(raw)             -> ?INET_OPT_RAW;
 % Names of SCTP opts:
 enc_opt(sctp_rtoinfo)	 	   -> ?SCTP_OPT_RTOINFO;
@@ -1164,6 +1174,7 @@ dec_opt(?INET_LOPT_TCP_SEND_TIMEOUT_CLOSE) -> send_timeout_close;
 dec_opt(?INET_LOPT_TCP_DELAY_SEND)   -> delay_send;
 dec_opt(?INET_LOPT_PACKET_SIZE)      -> packet_size;
 dec_opt(?INET_LOPT_READ_PACKETS)     -> read_packets;
+dec_opt(?INET_LOPT_NETNS)           -> netns;
 dec_opt(?INET_OPT_RAW)              -> raw;
 dec_opt(I) when is_integer(I)     -> undefined.
 
@@ -1261,6 +1272,7 @@ type_opt_1(send_timeout_close) -> bool;
 type_opt_1(delay_send)      -> bool;
 type_opt_1(packet_size)     -> uint;
 type_opt_1(read_packets)    -> uint;
+type_opt_1(netns)           -> binary;
 %% 
 %% SCTP options (to be set). If the type is a record type, the corresponding
 %% record signature is returned, otherwise, an "elementary" type tag 
@@ -1487,9 +1499,12 @@ type_value_2({bitenumlist,List,_}, EnumList) ->
 	Ls when is_list(Ls)                         -> true;
 	false                                       -> false
     end;
-type_value_2(binary,Bin) when is_binary(Bin) -> true;
-type_value_2(binary_or_uint,Bin) when is_binary(Bin) -> true;
-type_value_2(binary_or_uint,Int) when is_integer(Int), Int >= 0 -> true;
+type_value_2(binary,Bin)
+  when is_binary(Bin), byte_size(Bin) < (1 bsl 32)  -> true;
+type_value_2(binary_or_uint,Bin)
+  when is_binary(Bin), byte_size(Bin) < (1 bsl 32)  -> true;
+type_value_2(binary_or_uint,Int)
+  when is_integer(Int), Int >= 0                    -> true;
 %% Type-checking of SCTP options
 type_value_2(sctp_assoc_id, X)
   when X band 16#ffffffff =:= X                     -> true;
diff --git a/lib/kernel/src/inet_int.hrl b/lib/kernel/src/inet_int.hrl
index 67a99913a1..18a4a61b2f 100644
--- a/lib/kernel/src/inet_int.hrl
+++ b/lib/kernel/src/inet_int.hrl
@@ -143,6 +143,7 @@
 -define(INET_LOPT_TCP_SEND_TIMEOUT_CLOSE, 35).
 -define(INET_LOPT_MSGQ_HIWTRMRK,  36).
 -define(INET_LOPT_MSGQ_LOWTRMRK,  37).
+-define(INET_LOPT_NETNS,          38).
 % Specific SCTP options: separate range:
 -define(SCTP_OPT_RTOINFO,	 	100).
 -define(SCTP_OPT_ASSOCINFO,	 	101).