Skip to contentSkip to conversation

nsenter: fix ability to enter unprivileged containers

PATCH Applied James Bottomley 4 participants 12 messages Apr 18, 2016 – Apr 22, 2016 Util-Linux package

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 nsenter: fix ability to enter unprivileged containers

If you enter it first, you lose privilege for subsequent namespace enters,see issue

https://github.com/karelzak/util-linux/issues/315

The fix is to enter the user namespace last of all.

sys-utils/nsenter.c +1 -1
index d8690db..1525f15 100644
@@ -52,13 +52,13 @@ static struct namespace_file {
 	 * first.  This gives an unprivileged user the potential to
 	 * enter the other namespaces.
 	 */
-	{ .nstype = CLONE_NEWUSER,  .name = "ns/user", .fd = -1 },
 	{ .nstype = CLONE_NEWCGROUP,.name = "ns/cgroup", .fd = -1 },
 	{ .nstype = CLONE_NEWIPC,   .name = "ns/ipc",  .fd = -1 },
 	{ .nstype = CLONE_NEWUTS,   .name = "ns/uts",  .fd = -1 },
 	{ .nstype = CLONE_NEWNET,   .name = "ns/net",  .fd = -1 },
 	{ .nstype = CLONE_NEWPID,   .name = "ns/pid",  .fd = -1 },
 	{ .nstype = CLONE_NEWNS,    .name = "ns/mnt",  .fd = -1 },
+	{ .nstype = CLONE_NEWUSER,  .name = "ns/user", .fd = -1 },
 	{ .nstype = 0, .name = NULL, .fd = -1 }
 };
 
#1 lore raw

James Bottomley
<[email protected]>

Quoted 6 lines · writes:

If you enter it first, you lose privilege for subsequent namespace enters,see issue

https://github.com/karelzak/util-linux/issues/315

The fix is to enter the user namespace last of all.

I verified that with *current*/unpatched nsenter,

$ unshare -rm sleep inf &
$ nsenter -t $! -U -m --preserve

works as expected (from regular user [and with unprivileged userns enabled]).

With this patch it *won't* work [verified], of course (as you'll need root privileges in userns before joining mount-ns, and you can only obtain them by entering userns first).

Of course, you can workaround it by invoking nsenter twice:

$ nsenter -t $! -U --preserve nsenter -t $! -m

but same could be said about issue 315: you can workaround it by manually splitting entering mount-ns and user-ns, something like

# nsenter --mount=/run/build-container/aarch64 nsenter --user=/run/build-container/user

or (if /run/build-container/user is not visible inside mount-ns)

# nsenter --mount=/run/build-container/aarch64 nsenter --user=/dev/fd/3 3</run/build-container/user

(disclaimer: unverified; on my kernel mount-bind fails for mount-ns fds).

Quoted 20 lines
sys-utils/nsenter.c +1 -1
index d8690db..1525f15 100644
@@ -52,13 +52,13 @@ static struct namespace_file {
 	 * first.  This gives an unprivileged user the potential to
 	 * enter the other namespaces.
 	 */
-	{ .nstype = CLONE_NEWUSER,  .name = "ns/user", .fd = -1 },
 	{ .nstype = CLONE_NEWCGROUP,.name = "ns/cgroup", .fd = -1 },
 	{ .nstype = CLONE_NEWIPC,   .name = "ns/ipc",  .fd = -1 },
 	{ .nstype = CLONE_NEWUTS,   .name = "ns/uts",  .fd = -1 },
 	{ .nstype = CLONE_NEWNET,   .name = "ns/net",  .fd = -1 },
 	{ .nstype = CLONE_NEWPID,   .name = "ns/pid",  .fd = -1 },
 	{ .nstype = CLONE_NEWNS,    .name = "ns/mnt",  .fd = -1 },
+	{ .nstype = CLONE_NEWUSER,  .name = "ns/user", .fd = -1 },
 	{ .nstype = 0, .name = NULL, .fd = -1 }
 };
#3 lore raw
In reply to James Bottomley

OK, so if you want me to reply properly, you're going to have to keep my address in the cc list.

Quoted 13 lines

&gt; If you enter it first, you lose privilege for subsequent namespace
&gt; enters,see issue
&gt;
&gt; https://github.com/karelzak/util-linux/issues/315
&gt;
&gt; The fix is to enter the user namespace last of all.

I verified that with *current*/unpatched nsenter,

$ unshare -rm sleep inf &amp;
$ nsenter -t $! -U -m --preserve
Quoted 6 lines

works as expected (from regular user [and with unprivileged userns enabled]).

With this patch it *won't* work [verified], of course (as you'll need root privileges in userns before joining mount-ns, and you can only obtain them by entering userns first).

So we're using userns for different things. I'm using it to remove privilege (so on my userns implementation root in the host enters but on becoming root in the userns, it can do nothing other than write to its own files) and you're using it to enhance privilege. It looks like these two things will always be mutually exclusive, so perhaps we need an extra flag to nsenter to say do the userns first or last?

Quoted 11 lines

Of course, you can workaround it by invoking nsenter twice:

$ nsenter -t $! -U --preserve nsenter -t $! -m

but same could be said about issue 315: you can workaround it by manually splitting entering mount-ns and user-ns, something like

# nsenter --mount=/run/build-container/aarch64 nsenter -
-user=/run/build-container/user

or (if /run/build-container/user is not visible inside mount-ns)

That's right, it isn't

Quoted 2 lines
# nsenter --mount=/run/build-container/aarch64 nsenter -
-user=/dev/fd/3 3&lt;/run/build-container/user

It should work, but for some inexplicable reason it's giving EINVAL.

# nsenter --mount=/run/build-container/aarch64 3</run/build-container/user 
# ls -l /proc/self/fd
total 0
lrwx------ 1 root root 64 Apr 18 15:31 0 -> /dev/pts/1
lrwx------ 1 root root 64 Apr 18 15:31 1 -> /dev/pts/1
lrwx------ 1 root root 64 Apr 18 15:31 2 -> /dev/pts/1
lr-x------ 1 root root 64 Apr 18 15:31 3 -> /run/build-container/user
lr-x------ 1 root root 64 Apr 18 15:31 4 -> /proc/10304/fd
# nsenter --user=/proc/self/fd/3
nsenter: reassociate to namespace 'ns/user' failed: Invalid argument

I think it's because the fd wasn't properly opened by the shell

Quoted 2 lines

(disclaimer: unverified; on my kernel mount-bind fails for mount-ns fds).

That's probably because you're running systemd. Systemd sets all subtrees to shared and you can only bind mount a mount namespace file descriptor on to a private subtree.

James

#4 lore raw
Quoted 17 lines · James Bottomley
Quoted 2 lines
# nsenter --mount=/run/build-container/aarch64 nsenter -
-user=/dev/fd/3 3&lt;/run/build-container/user

It should work, but for some inexplicable reason it's giving EINVAL.

# nsenter --mount=/run/build-container/aarch64 3</run/build
-container/user 
# ls -l /proc/self/fd
total 0
lrwx------ 1 root root 64 Apr 18 15:31 0 -> /dev/pts/1
lrwx------ 1 root root 64 Apr 18 15:31 1 -> /dev/pts/1
lrwx------ 1 root root 64 Apr 18 15:31 2 -> /dev/pts/1
lr-x------ 1 root root 64 Apr 18 15:31 3 -> /run/build-container/user
lr-x------ 1 root root 64 Apr 18 15:31 4 -> /proc/10304/fd
# nsenter --user=/proc/self/fd/3
nsenter: reassociate to namespace 'ns/user' failed: Invalid argument
Quoted 1 line

I think it's because the fd wasn't properly opened by the shell

Actually, just to follow up, this is specifically a problem of the type of container I'm running: it's actually an architecture emulation container, so after I've done the mount namespace enter, I'm running under aarch64 emulation and, apparently, qemu has some problem with the setns system call. This does, however, mean I have to enter both namespaces together ...

James

#2 lore raw
In reply to Yuriy M. Kaminskiy
Quoted 3 lines · [email protected] (Yuriy M. Kaminskiy) writes:
# nsenter --mount=/run/build-container/aarch64 nsenter --user=/dev/fd/3 3</run/build-container/user

(disclaimer: unverified;

Just for record,
1) above workaround verified/works;
2) and this part:

Quoted 1 line

on my kernel mount-bind fails for mount-ns fds).

was *my* mistake (I somehow missed mount-ns peculiarity about incompatibility with shared propagation [it is documented in unshare(1)])

maintainer
#5 lore raw
In reply to James Bottomley
Quoted 23 lines · James Bottomley

OK, so if you want me to reply properly, you're going to have to keep my address in the cc list.

Quoted 13 lines

&gt; If you enter it first, you lose privilege for subsequent namespace
&gt; enters,see issue
&gt;
&gt; https://github.com/karelzak/util-linux/issues/315
&gt;
&gt; The fix is to enter the user namespace last of all.

I verified that with *current*/unpatched nsenter,

$ unshare -rm sleep inf &amp;
$ nsenter -t $! -U -m --preserve
Quoted 6 lines

works as expected (from regular user [and with unprivileged userns enabled]).

With this patch it *won't* work [verified], of course (as you'll need root privileges in userns before joining mount-ns, and you can only obtain them by entering userns first).

Quoted 6 lines

So we're using userns for different things. I'm using it to remove privilege (so on my userns implementation root in the host enters but on becoming root in the userns, it can do nothing other than write to its own files) and you're using it to enhance privilege. It looks like these two things will always be mutually exclusive, so perhaps we need an extra flag to nsenter to say do the userns first or last?

That's what I have talked about at github -- see Eric's comment in the code, the user NS is the first in the array for a good reason. May be it would be really better to add --user-{first,last} options to specify when you want to enter user NS.

Karel

Signature · 2 lines
 Karel Zak  <[email protected]>
 http://karelzak.blogspot.com
#6 lore raw
Quoted 31 lines · Karel Zak
Quoted 23 lines · James Bottomley

OK, so if you want me to reply properly, you're going to have to keep my address in the cc list.

Quoted 13 lines
Quoted 1 line

If you enter it first, you lose privilege for subsequent

namespace

Quoted 5 lines

enters,see issue

https://github.com/karelzak/util-linux/issues/315

The fix is to enter the user namespace last of all.

I verified that with *current*/unpatched nsenter,

$ unshare -rm sleep inf &
$ nsenter -t $! -U -m --preserve
Quoted 6 lines

works as expected (from regular user [and with unprivileged userns enabled]).

With this patch it *won't* work [verified], of course (as you'll need root privileges in userns before joining mount-ns, and you can only obtain them by entering userns first).

Quoted 6 lines

So we're using userns for different things. I'm using it to remove privilege (so on my userns implementation root in the host enters but on becoming root in the userns, it can do nothing other than write to its own files) and you're using it to enhance privilege. It looks like these two things will always be mutually exclusive, so perhaps we need an extra flag to nsenter to say do the userns first or last?

Quoted 4 lines

That's what I have talked about at github -- see Eric's comment in the code, the user NS is the first in the array for a good reason. May be it would be really better to add --user-{first,last} options to specify when you want to enter user NS.

OK, I'll code this up; hang on.

James

#7 lore raw
Quoted 32 lines · James Bottomley
Quoted 31 lines · Karel Zak
Quoted 29 lines · James Bottomley

OK, so if you want me to reply properly, you're going to have to keep my address in the cc list.

Quoted 13 lines
Quoted 1 line

If you enter it first, you lose privilege for subsequent

namespace

Quoted 5 lines

enters,see issue

https://github.com/karelzak/util-linux/issues/315

The fix is to enter the user namespace last of all.

I verified that with *current*/unpatched nsenter,

$ unshare -rm sleep inf &
$ nsenter -t $! -U -m --preserve
Quoted 6 lines

works as expected (from regular user [and with unprivileged userns enabled]).

With this patch it *won't* work [verified], of course (as you'll need root privileges in userns before joining mount-ns, and you can only obtain them by entering userns first).

So we're using userns for different things. I'm using it to remove privilege (so on my userns implementation root in the host enters but on becoming root in the userns, it can do nothing other than write to its own files) and you're using it to enhance privilege. It looks like these two things will always be mutually exclusive, so perhaps we need an extra flag to nsenter to say do the userns first or last?

Quoted 6 lines
Quoted 4 lines

That's what I have talked about at github -- see Eric's comment in the code, the user NS is the first in the array for a good reason. May be it would be really better to add --user-{first,last} options to specify when you want to enter user NS.

OK, I'll code this up; hang on.

I think we can do this even better with two passes to setns.

A first pass before the user namespace is set (that ignores failures), and a second pass that sets the user namespace first as happens today.

That should satisfy both cases without flags, and would remove the need to remember/guess which kind of container people are using.

Eric

#11 lore raw
In reply to Karel Zak
Quoted 39 lines · Karel Zak
Quoted 27 lines · James Bottomley

OK, so if you want me to reply properly, you're going to have to keep
my address in the cc list.

Quoted 17 lines

&gt; If you enter it first, you lose privilege for subsequent namespace
&gt; enters,see issue
&gt;
&gt; https://github.com/karelzak/util-linux/issues/315
&gt;
&gt; The fix is to enter the user namespace last of all.

I verified that with *current*/unpatched nsenter,

$ unshare -rm sleep inf &amp;
$ nsenter -t $! -U -m --preserve

works as expected (from regular user [and with unprivileged userns
enabled]).

Quoted 5 lines

With this patch it *won't* work [verified], of course (as you'll need
root privileges in userns before joining mount-ns, and you can only
obtain them by entering userns first).

Quoted 10 lines

So we're using userns for different things. I'm using it to remove privilege (so on my userns implementation root in the host enters but
on becoming root in the userns, it can do nothing other than write to
its own files) and you're using it to enhance privilege. It looks like
these two things will always be mutually exclusive, so perhaps we need
an extra flag to nsenter to say do the userns first or last?

Quoted 5 lines

That's what I have talked about at github -- see Eric's comment in the
code, the user NS is the first in the array for a good reason. May be it would be really better to add --user-{first,last} options to specify when you want to enter user NS.

How does this look; it's just got a --user-last switch because --user -first is current behaviour.

James

---

Quoted 1 line

From eee660d40179ed4a186f6c5a73d5596058f87473 Mon Sep 17 00:00:00 2001

From: James Bottomley <[email protected]>
Date: Fri, 15 Apr 2016 08:10:20 -0700
Subject: [PATCH] nsenter: add --user-last flag to enter the user namespace
last

We have two use cases for user namespaces, one to elevate the privilege of an unprivileged user, in which case we have to enter the user namespace before all other namespaces (otherwise there isn't enough permission to enter any other namespace). And the other one is where we're deprivileging a user and thus have to enter the user namespace last (because that's the point at which we lose the privileges).

This fixes

https://github.com/karelzak/util-linux/issues/315

sys-utils/nsenter.1 +4 -0
index ea5992e..3acf8da 100644
@@ -154,6 +154,10 @@ always sets UID for user namespaces, the default is 0.
 Don't modify UID and GID when enter user namespace. The default is to
 drops supplementary groups and sets GID and UID to 0.
 .TP
+\fB\-\-user\-last\fR
+Enter the user namespace last rather than first.  You need this option
+if the usernamespace deprivileges the current process.
+.TP
 \fB\-r\fR, \fB\-\-root\fR[=\fIdirectory\fR]
 Set the root directory.  If no directory is specified, set the root directory to
 the root directory of the target process.  If directory is specified, set the
sys-utils/nsenter.c +25 -5
index d8690db..9d82621 100644
@@ -48,9 +48,10 @@ static struct namespace_file {
 } namespace_files[] = {
 	/* Careful the order is significant in this array.
 	 *
-	 * The user namespace comes first, so that it is entered
-	 * first.  This gives an unprivileged user the potential to
-	 * enter the other namespaces.
+	 * The user namespace comes either first or last: first if
+	 * you're using it to increase your privilege and so enter the
+	 * other namespaces and last if you're using it to decrease
+	 * your privilege (see --user-last flag).
 	 */
 	{ .nstype = CLONE_NEWUSER,  .name = "ns/user", .fd = -1 },
 	{ .nstype = CLONE_NEWCGROUP,.name = "ns/cgroup", .fd = -1 },
@@ -88,6 +89,7 @@ static void usage(int status)
 	fputs(_(" -r, --root[=<dir>]     set the root directory\n"), out);
 	fputs(_(" -w, --wd[=<dir>]       set the working directory\n"), out);
 	fputs(_(" -F, --no-fork          do not fork before exec'ing <program>\n"), out);
+	fputs(_("     --user-last        enter the user namespace last intstead of first\n"), out);
 #ifdef HAVE_LIBSELINUX
 	fputs(_(" -Z, --follow-context   set SELinux context according to --target PID\n"), out);
 #endif
@@ -176,7 +178,8 @@ static void continue_as_child(void)
 int main(int argc, char *argv[])
 {
 	enum {
-		OPT_PRESERVE_CRED = CHAR_MAX + 1
+		OPT_PRESERVE_CRED = CHAR_MAX + 1,
+		OPT_USER_LAST,
 	};
 	static const struct option longopts[] = {
 		{ "help", no_argument, NULL, 'h' },
@@ -195,6 +198,7 @@ int main(int argc, char *argv[])
 		{ "wd", optional_argument, NULL, 'w' },
 		{ "no-fork", no_argument, NULL, 'F' },
 		{ "preserve-credentials", no_argument, NULL, OPT_PRESERVE_CRED },
+		{ "user-last", no_argument, NULL, OPT_USER_LAST },
 #ifdef HAVE_LIBSELINUX
 		{ "follow-context", no_argument, NULL, 'Z' },
 #endif
@@ -202,7 +206,7 @@ int main(int argc, char *argv[])
 	};
 
 	struct namespace_file *nsfile;
-	int c, namespaces = 0, setgroups_nerrs = 0, preserve_cred = 0;
+	int c, namespaces = 0, setgroups_nerrs = 0, preserve_cred = 0, user_last = 0;
 	bool do_rd = false, do_wd = false, force_uid = false, force_gid = false;
 	int do_fork = -1; /* unknown yet */
 	uid_t uid = 0;
@@ -297,6 +301,9 @@ int main(int argc, char *argv[])
 		case OPT_PRESERVE_CRED:
 			preserve_cred = 1;
 			break;
+		case OPT_USER_LAST:
+			user_last = 1;
+			break;
 #ifdef HAVE_LIBSELINUX
 		case 'Z':
 			selinux = 1;
@@ -321,6 +328,19 @@ int main(int argc, char *argv[])
 		freecon(scon);
 	}
 #endif
+
+	if (user_last) {
+		/*
+		 * swap the first and last entries in the namespace_files
+		 * array (so swap the user and mount entries)
+		 */
+		int user_new = ARRAY_SIZE(namespace_files) - 2;
+		struct namespace_file nsf = namespace_files[user_new];
+
+		namespace_files[user_new] = namespace_files[0];
+		namespace_files[0] = nsf;
+	}
+
 	/*
 	 * Open remaining namespace and directory descriptors.
 	 */
#8 lore raw
In reply to Eric W. Biederman
Quoted 54 lines · Eric W. Biederman
Quoted 41 lines · James Bottomley
Quoted 40 lines · Karel Zak
Quoted 38 lines · James Bottomley

OK, so if you want me to reply properly, you're going to have to keep
my address in the cc list.

Quoted 17 lines
Quoted 1 line

If you enter it first, you lose privilege for subsequent

namespace

Quoted 5 lines

enters,see issue

https://github.com/karelzak/util-linux/issues/315

The fix is to enter the user namespace last of all.

I verified that with *current*/unpatched nsenter,

$ unshare -rm sleep inf &
$ nsenter -t $! -U -m --preserve

works as expected (from regular user [and with unprivileged userns
enabled]).

Quoted 5 lines

With this patch it *won't* work [verified], of course (as you'll need
root privileges in userns before joining mount-ns, and you can only
obtain them by entering userns first).

So we're using userns for different things. I'm using it to remove
privilege (so on my userns implementation root in the host enters but
on becoming root in the userns, it can do nothing other than write to
its own files) and you're using it to enhance privilege. It looks like
these two things will always be mutually exclusive, so perhaps we need
an extra flag to nsenter to say do the userns first or last?

Quoted 8 lines
Quoted 6 lines

That's what I have talked about at github -- see Eric's comment in the
code, the user NS is the first in the array for a good reason. May be
it would be really better to add --user-{first,last} options to specify when you want to enter user NS.

OK, I'll code this up; hang on.

I think we can do this even better with two passes to setns.

Quoted 8 lines

A first pass before the user namespace is set (that ignores failures),
and a second pass that sets the user namespace first as happens today.

That should satisfy both cases without flags, and would remove the need
to remember/guess which kind of container people are using.

Makes sense, so like this?

James

---

Quoted 1 line

From 39cecc604a6c997c328affea52e65d6f67898bf5 Mon Sep 17 00:00:00 2001

From: James Bottomley <[email protected]>
Date: Fri, 15 Apr 2016 08:10:20 -0700
Subject: [PATCH] nsenter: enter namespaces in two passes

We have two use cases for user namespaces, one to elevate the privilege of an unprivileged user, in which case we have to enter the user namespace before all other namespaces (otherwise there isn't enough permission to enter any other namespace). And the other one is where we're deprivileging a user and thus have to enter the user namespace last (because that's the point at which we lose the privileges). On the first pass, we start at the position one after the user namespace clearing the file descriptors as we close them after calling setns(). If setns() fails on the first pass, ignore the failure assuming that it will succeed after we enter the user namespace.

This fixes

https://github.com/karelzak/util-linux/issues/315

sys-utils/nsenter.c +30 -16
index d8690db..8dbd22b 100644
@@ -48,9 +48,11 @@ static struct namespace_file {
 } namespace_files[] = {
 	/* Careful the order is significant in this array.
 	 *
-	 * The user namespace comes first, so that it is entered
-	 * first.  This gives an unprivileged user the potential to
-	 * enter the other namespaces.
+	 * The user namespace comes either first or last: first if
+	 * you're using it to increase your privilege and last if
+	 * you're using it to decrease.  We enter the namespaces in
+	 * two passes starting initially from offset 1 and then offset
+	 * 0 if that fails.
 	 */
 	{ .nstype = CLONE_NEWUSER,  .name = "ns/user", .fd = -1 },
 	{ .nstype = CLONE_NEWCGROUP,.name = "ns/cgroup", .fd = -1 },
@@ -202,7 +204,7 @@ int main(int argc, char *argv[])
 	};
 
 	struct namespace_file *nsfile;
-	int c, namespaces = 0, setgroups_nerrs = 0, preserve_cred = 0;
+	int c, pass, namespaces = 0, setgroups_nerrs = 0, preserve_cred = 0;
 	bool do_rd = false, do_wd = false, force_uid = false, force_gid = false;
 	int do_fork = -1; /* unknown yet */
 	uid_t uid = 0;
@@ -353,19 +355,31 @@ int main(int argc, char *argv[])
 	}
 
 	/*
-	 * Now that we know which namespaces we want to enter, enter them.
+	 * Now that we know which namespaces we want to enter, enter
+	 * them.  Do this in two passes, not entering the user
+	 * namespace on the first pass.  So if we're deprivileging the
+	 * container we'll enter the user namespace last and if we're
+	 * privileging it then we enter the usernamespace first
+	 * (because the initial setns will fail).
 	 */
-	for (nsfile = namespace_files; nsfile->nstype; nsfile++) {
-		if (nsfile->fd < 0)
-			continue;
-		if (nsfile->nstype == CLONE_NEWPID && do_fork == -1)
-			do_fork = 1;
-		if (setns(nsfile->fd, nsfile->nstype))
-			err(EXIT_FAILURE,
-			    _("reassociate to namespace '%s' failed"),
-			    nsfile->name);
-		close(nsfile->fd);
-		nsfile->fd = -1;
+	for (pass = 0; pass < 2; pass ++) {
+		for (nsfile = namespace_files + 1 - pass; nsfile->nstype; nsfile++) {
+			if (nsfile->fd < 0)
+				continue;
+			if (nsfile->nstype == CLONE_NEWPID && do_fork == -1)
+				do_fork = 1;
+			if (setns(nsfile->fd, nsfile->nstype)) {
+				if (pass != 0)
+					err(EXIT_FAILURE,
+					    _("reassociate to namespace '%s' failed"),
+					    nsfile->name);
+				else
+					continue;
+			}
+
+			close(nsfile->fd);
+			nsfile->fd = -1;
+		}
 	}
 
 	/* Remember the current working directory if I'm not changing it */
#9 lore raw
Quoted 8 lines · James Bottomley
Quoted 7 lines · Eric W. Biederman

I think we can do this even better with two passes to setns.

A first pass before the user namespace is set (that ignores failures),
and a second pass that sets the user namespace first as happens today.

Quoted 5 lines
Quoted 3 lines

That should satisfy both cases without flags, and would remove the need
to remember/guess which kind of container people are using.

Makes sense, so like this?

That looks like it will work, and for the right reasons.

Quoted 18 lines

---
From 39cecc604a6c997c328affea52e65d6f67898bf5 Mon Sep 17 00:00:00 2001
From: James Bottomley <[email protected]>
Date: Fri, 15 Apr 2016 08:10:20 -0700
Subject: [PATCH] nsenter: enter namespaces in two passes

We have two use cases for user namespaces, one to elevate the privilege of an unprivileged user, in which case we have to enter the user namespace before all other namespaces (otherwise there isn't enough permission to enter any other namespace). And the other one is where we're deprivileging a user and thus have to enter the user namespace last (because that's the point at which we lose the privileges). On the first pass, we start at the position one after the user namespace clearing the file descriptors as we close them after calling setns(). If setns() fails on the first pass, ignore the failure assuming that it will succeed after we enter the user namespace.

Quoted 5 lines

This fixes

https://github.com/karelzak/util-linux/issues/315

Quoted 73 lines
sys-utils/nsenter.c +30 -16
index d8690db..8dbd22b 100644
@@ -48,9 +48,11 @@ static struct namespace_file {
 } namespace_files[] = {
 	/* Careful the order is significant in this array.
 	 *
-	 * The user namespace comes first, so that it is entered
-	 * first.  This gives an unprivileged user the potential to
-	 * enter the other namespaces.
+	 * The user namespace comes either first or last: first if
+	 * you're using it to increase your privilege and last if
+	 * you're using it to decrease.  We enter the namespaces in
+	 * two passes starting initially from offset 1 and then offset
+	 * 0 if that fails.
 	 */
 	{ .nstype = CLONE_NEWUSER,  .name = "ns/user", .fd = -1 },
 	{ .nstype = CLONE_NEWCGROUP,.name = "ns/cgroup", .fd = -1 },
@@ -202,7 +204,7 @@ int main(int argc, char *argv[])
 	};
 
 	struct namespace_file *nsfile;
-	int c, namespaces = 0, setgroups_nerrs = 0, preserve_cred = 0;
+	int c, pass, namespaces = 0, setgroups_nerrs = 0, preserve_cred = 0;
 	bool do_rd = false, do_wd = false, force_uid = false, force_gid = false;
 	int do_fork = -1; /* unknown yet */
 	uid_t uid = 0;
@@ -353,19 +355,31 @@ int main(int argc, char *argv[])
 	}
 
 	/*
-	 * Now that we know which namespaces we want to enter, enter them.
+	 * Now that we know which namespaces we want to enter, enter
+	 * them.  Do this in two passes, not entering the user
+	 * namespace on the first pass.  So if we're deprivileging the
+	 * container we'll enter the user namespace last and if we're
+	 * privileging it then we enter the usernamespace first
+	 * (because the initial setns will fail).
 	 */
-	for (nsfile = namespace_files; nsfile->nstype; nsfile++) {
-		if (nsfile->fd < 0)
-			continue;
-		if (nsfile->nstype == CLONE_NEWPID && do_fork == -1)
-			do_fork = 1;
-		if (setns(nsfile->fd, nsfile->nstype))
-			err(EXIT_FAILURE,
-			    _("reassociate to namespace '%s' failed"),
-			    nsfile->name);
-		close(nsfile->fd);
-		nsfile->fd = -1;
+	for (pass = 0; pass < 2; pass ++) {
+		for (nsfile = namespace_files + 1 - pass; nsfile->nstype; nsfile++) {
+			if (nsfile->fd < 0)
+				continue;
+			if (nsfile->nstype == CLONE_NEWPID && do_fork == -1)
+				do_fork = 1;
+			if (setns(nsfile->fd, nsfile->nstype)) {
+				if (pass != 0)
+					err(EXIT_FAILURE,
+					    _("reassociate to namespace '%s' failed"),
+					    nsfile->name);
+				else
+					continue;
+			}
+
+			close(nsfile->fd);
+			nsfile->fd = -1;
+		}
 	}
 
 	/* Remember the current working directory if I'm not changing it */