Skip to contentSkip to conversation

ip: reset netns after each command in batch mode

PATCH Applied Matteo Croce 2 participants 3 messages Jun 7, 2019 – Jun 10, 2019

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

PATCH v2 1/1 ip: reset netns after each command in batch mode

When creating a new netns or executing a program into an existing one, the unshare() or setns() calls will change the current netns. In batch mode, this can run commands on the wrong interfaces, as the ifindex value is meaningful only in the current netns. For example, this command fails because veth-c doesn't exists in the init netns:

    # ip -b - <<-'EOF'
        netns add client
        link add name veth-c type veth peer veth-s netns client
        addr add 192.168.2.1/24 dev veth-c
    EOF
    Cannot find device "veth-c"
    Command failed -:7

But if there are two devices with the same name in the init and new netns, ip will build a wrong ll_map with indexes belonging to the new netns, and will execute actions in the init netns using this wrong mapping. This script will flush all eth0 addresses and bring it down, as it has the same ifindex of veth0 in the new netns:

    # ip addr
    1: lo: <LOOPBACK,UP,LOWER_UP> mtu 65536 qdisc noqueue state UNKNOWN group default qlen 1000
        link/loopback 00:00:00:00:00:00 brd 00:00:00:00:00:00
        inet 127.0.0.1/8 scope host lo
           valid_lft forever preferred_lft forever
    2: eth0: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 qdisc mq state UP group default qlen 1000
        link/ether 52:54:00:12:34:56 brd ff:ff:ff:ff:ff:ff
        inet 192.168.122.76/24 brd 192.168.122.255 scope global dynamic eth0
           valid_lft 3598sec preferred_lft 3598sec
    # ip -b - <<-'EOF'
        netns add client
        link add name veth0 type veth peer name veth1
        link add name veth-ns type veth peer name veth0 netns client
        link set veth0 down
        address flush veth0
    EOF
    # ip addr
    1: lo: <LOOPBACK,UP,LOWER_UP> mtu 65536 qdisc noqueue state UNKNOWN group default qlen 1000
        link/loopback 00:00:00:00:00:00 brd 00:00:00:00:00:00
        inet 127.0.0.1/8 scope host lo
           valid_lft forever preferred_lft forever
    2: eth0: <BROADCAST,MULTICAST> mtu 1500 qdisc mq state DOWN group default qlen 1000
        link/ether 52:54:00:12:34:56 brd ff:ff:ff:ff:ff:ff
    3: veth1@veth0: <BROADCAST,MULTICAST,M-DOWN> mtu 1500 qdisc noop state DOWN group default qlen 1000
        link/ether c2:db:d0:34:13:4a brd ff:ff:ff:ff:ff:ff
    4: veth0@veth1: <BROADCAST,MULTICAST,M-DOWN> mtu 1500 qdisc noop state DOWN group default qlen 1000
        link/ether ca:9d:6b:5f:5f:8f brd ff:ff:ff:ff:ff:ff
    5: veth-ns@if2: <BROADCAST,MULTICAST> mtu 1500 qdisc noop state DOWN group default qlen 1000
        link/ether 32:ef:22:df:51:0a brd ff:ff:ff:ff:ff:ff link-netns client

The same issue can be triggered by the netns exec subcommand with a sligthy different script:

    # ip netns add client
    # ip -b - <<-'EOF'
        netns exec client true
        link add name veth0 type veth peer name veth1
        link add name veth-ns type veth peer name veth0 netns client
        link set veth0 down
        address flush veth0
    EOF

Fix this by adding two netns_{save,reset} functions, which are used to get a file descriptor for the init netns, and restore it after each batch command.
netns_save() is called before the unshare() or setns(), while netns_restore() is called after each command.

Fixes 0dc34c7713bb ("iproute2: Add processless network namespace support")
Reviewed-and-tested-by Andrea Claudi <[email protected]>

---
v1 -> v2:
- exit on netns error
- refactor to reduce indentation

 include/namespace.h |  2 ++
 ip/ip.c             |  1 +
 ip/ipnetns.c        |  1 +
 lib/namespace.c     | 31 +++++++++++++++++++++++++++++++
 4 files changed, 35 insertions(+)
include/namespace.h +2 -0
index e47f9b5d..89cdda11 100644
@@ -49,6 +49,8 @@ static inline int setns(int fd, int nstype)
 }
 #endif /* HAVE_SETNS */
 
+void netns_save(void);
+void netns_restore(void);
 int netns_switch(char *netns);
 int netns_get_fd(const char *netns);
 int netns_foreach(int (*func)(char *nsname, void *arg), void *arg);
ip/ip.c +1 -0
index b71ae816..49b3aa49 100644
@@ -158,6 +158,7 @@ static int batch(const char *name)
 			if (!force)
 				break;
 		}
+		netns_restore();
 	}
 	if (line)
 		free(line);
ip/ipnetns.c +1 -0
index db11fdb2..8ead0c4c 100644
@@ -708,6 +708,7 @@ static int netns_add(int argc, char **argv, bool create)
 	close(fd);
 
 	if (create) {
+		netns_save();
 		if (unshare(CLONE_NEWNET) < 0) {
 			fprintf(stderr, "Failed to create a new network namespace \"%s\": %s\n",
 				name, strerror(errno));
lib/namespace.c +31 -0
index 06ae0a48..a2aea57a 100644
@@ -15,6 +15,35 @@
 #include "utils.h"
 #include "namespace.h"
 
+static int saved_netns = -1;
+
+/* Obtain a FD for the current namespace, so we can reenter it later */
+void netns_save(void)
+{
+	if (saved_netns != -1)
+		return;
+
+	saved_netns = open("/proc/self/ns/net", O_RDONLY | O_CLOEXEC);
+	if (saved_netns == -1) {
+		perror("Cannot open init namespace");
+		exit(1);
+	}
+}
+
+void netns_restore(void)
+{
+	if (saved_netns == -1)
+		return;
+
+	if (setns(saved_netns, CLONE_NEWNET)) {
+		perror("setns");
+		exit(1);
+	}
+
+	close(saved_netns);
+	saved_netns = -1;
+}
+
 static void bind_etc(const char *name)
 {
 	char etc_netns_path[sizeof(NETNS_ETC_DIR) + NAME_MAX];
@@ -61,6 +90,8 @@ int netns_switch(char *name)
 		return -1;
 	}
 
+	netns_save();
+
 	if (setns(netns, CLONE_NEWNET) < 0) {
 		fprintf(stderr, "setting the network namespace \"%s\" failed: %s\n",
 			name, strerror(errno));
Signature · 1 line
2.21.0
#1 lore raw
Quoted 70 lines · On Fri, 7 Jun 2019 22:41:22 +0200 Matteo Croce <[email protected]> wrote:

When creating a new netns or executing a program into an existing one, the unshare() or setns() calls will change the current netns. In batch mode, this can run commands on the wrong interfaces, as the ifindex value is meaningful only in the current netns. For example, this command fails because veth-c doesn't exists in the init netns:

    # ip -b - <<-'EOF'
        netns add client
        link add name veth-c type veth peer veth-s netns client
        addr add 192.168.2.1/24 dev veth-c
    EOF
    Cannot find device "veth-c"
    Command failed -:7

But if there are two devices with the same name in the init and new netns, ip will build a wrong ll_map with indexes belonging to the new netns, and will execute actions in the init netns using this wrong mapping. This script will flush all eth0 addresses and bring it down, as it has the same ifindex of veth0 in the new netns:

    # ip addr
    1: lo: <LOOPBACK,UP,LOWER_UP> mtu 65536 qdisc noqueue state UNKNOWN group default qlen 1000
        link/loopback 00:00:00:00:00:00 brd 00:00:00:00:00:00
        inet 127.0.0.1/8 scope host lo
           valid_lft forever preferred_lft forever
    2: eth0: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 qdisc mq state UP group default qlen 1000
        link/ether 52:54:00:12:34:56 brd ff:ff:ff:ff:ff:ff
        inet 192.168.122.76/24 brd 192.168.122.255 scope global dynamic eth0
           valid_lft 3598sec preferred_lft 3598sec
    # ip -b - <<-'EOF'
        netns add client
        link add name veth0 type veth peer name veth1
        link add name veth-ns type veth peer name veth0 netns client
        link set veth0 down
        address flush veth0
    EOF
    # ip addr
    1: lo: <LOOPBACK,UP,LOWER_UP> mtu 65536 qdisc noqueue state UNKNOWN group default qlen 1000
        link/loopback 00:00:00:00:00:00 brd 00:00:00:00:00:00
        inet 127.0.0.1/8 scope host lo
           valid_lft forever preferred_lft forever
    2: eth0: <BROADCAST,MULTICAST> mtu 1500 qdisc mq state DOWN group default qlen 1000
        link/ether 52:54:00:12:34:56 brd ff:ff:ff:ff:ff:ff
    3: veth1@veth0: <BROADCAST,MULTICAST,M-DOWN> mtu 1500 qdisc noop state DOWN group default qlen 1000
        link/ether c2:db:d0:34:13:4a brd ff:ff:ff:ff:ff:ff
    4: veth0@veth1: <BROADCAST,MULTICAST,M-DOWN> mtu 1500 qdisc noop state DOWN group default qlen 1000
        link/ether ca:9d:6b:5f:5f:8f brd ff:ff:ff:ff:ff:ff
    5: veth-ns@if2: <BROADCAST,MULTICAST> mtu 1500 qdisc noop state DOWN group default qlen 1000
        link/ether 32:ef:22:df:51:0a brd ff:ff:ff:ff:ff:ff link-netns client

The same issue can be triggered by the netns exec subcommand with a sligthy different script:

    # ip netns add client
    # ip -b - <<-'EOF'
        netns exec client true
        link add name veth0 type veth peer name veth1
        link add name veth-ns type veth peer name veth0 netns client
        link set veth0 down
        address flush veth0
    EOF

Fix this by adding two netns_{save,reset} functions, which are used to get a file descriptor for the init netns, and restore it after each batch command.
netns_save() is called before the unshare() or setns(), while netns_restore() is called after each command.

Quoted 3 lines
Fixes 0dc34c7713bb ("iproute2: Add processless network namespace support")
Reviewed-and-tested-by Andrea Claudi <[email protected]>

Applied, thanks.

#2 lore raw
Quoted 11 lines · Stephen Hemminger
Quoted 5 lines · On Fri, 7 Jun 2019 22:41:22 +0200 Matteo Croce <[email protected]> wrote:

When creating a new netns or executing a program into an existing one, the unshare() or setns() calls will change the current netns. In batch mode, this can run commands on the wrong interfaces, as the ifindex value is meaningful only in the current netns. For example, this command fails because veth-c doesn't exists in the init netns:

Applied, thanks.

Great!

I was thinking about the code flow when running ip netns exec in batch or doall mode:

1) save the current netns()
2) change netns with setns()
3) clear the clear vrf associations with vrf_reset()
4) fork(), exec in the child
5) restore the previous netns in the parent

With a little rework (adding a parameter to cmd_exec), we could change the netns directly in the child, so the parent doesn't need steps 1,2,3,5
This could save some errors where the previous state can't be restored, eg. netns, vrp or the like.

I have a draft patch for this which, by the way, eliminates some dead code, I will send to the ML soon.

Regards,

Signature · 2 lines
Matteo Croce
per aspera ad upstream