Skip to contentSkip to conversation

Fix PID namespace persistence

PATCH Applied [email protected] 4 participants 8 messages Apr 15, 2020 – Apr 30, 2020 Util-Linux package

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 Fix PID namespace persistence

From: michael-dev <[email protected]>

After unshare(...) is called, /proc/self/ns/pid does not change.
Instead, only /proc/self/ns/pid_for_children is affected. So bind-mounting /proc/self/ns/pid results in the original namespace getting bind-mounted.

Fix this by instead bind-mounting ns/pid_for_children.

---
 sys-utils/unshare.c | 66 ++++++++++++++++++++++++++-------------------
 1 file changed, 38 insertions(+), 28 deletions(-)
sys-utils/unshare.c +38 -28
index 8652ebdaf..c3ba18e32 100644
@@ -63,7 +63,7 @@ static struct namespace_file {
 	{ .type = CLONE_NEWIPC,   .name = "ns/ipc"  },
 	{ .type = CLONE_NEWUTS,   .name = "ns/uts"  },
 	{ .type = CLONE_NEWNET,   .name = "ns/net"  },
-	{ .type = CLONE_NEWPID,   .name = "ns/pid"  },
+	{ .type = CLONE_NEWPID,   .name = "ns/pid_for_children"  },
 	{ .type = CLONE_NEWNS,    .name = "ns/mnt"  },
 	{ .type = CLONE_NEWTIME,  .name = "ns/time"  },
 	{ .name = NULL }
@@ -361,6 +361,7 @@ int main(int argc, char *argv[])
 	const char *procmnt = NULL;
 	const char *newroot = NULL;
 	const char *newdir = NULL;
+	pid_t pid_bind = 0;
 	pid_t pid = 0;
 	int fds[2];
 	int status;
@@ -501,13 +502,37 @@ int main(int argc, char *argv[])
 			"unsharing of a time namespace (-t)"));
 
 	if (npersists && (unshare_flags & CLONE_NEWNS))
-		bind_ns_files_from_child(&pid, fds);
+		bind_ns_files_from_child(&pid_bind, fds);
 
 	if (-1 == unshare(unshare_flags))
 		err(EXIT_FAILURE, _("unshare failed"));
 
-	if (npersists) {
-		if (pid && (unshare_flags & CLONE_NEWNS)) {
+	if (force_boottime)
+		settime(boottime, CLOCK_BOOTTIME);
+
+	if (force_monotonic)
+		settime(monotonic, CLOCK_MONOTONIC);
+
+	if (forkit) {
+		// force child forking before mountspace binding
+		// so pid_for_children is populated
+		pid = fork();
+
+		switch(pid) {
+		case -1:
+			err(EXIT_FAILURE, _("fork failed"));
+		case 0:	/* child */
+			if (pid_bind && (unshare_flags & CLONE_NEWNS))
+				close(fds[1]);
+			break;
+		default: /* parent */
+      break;
+		}
+	}
+
+	if (npersists && (pid || !forkit)) {
+		// run in parent
+		if (pid_bind && (unshare_flags & CLONE_NEWNS)) {
 			int rc;
 			char ch = PIPE_SYNC_BYTE;
 
@@ -518,7 +543,7 @@ int main(int argc, char *argv[])
 
 			/* wait for bind_ns_files_from_child() */
 			do {
-				rc = waitpid(pid, &status, 0);
+				rc = waitpid(pid_bind, &status, 0);
 				if (rc < 0) {
 					if (errno == EINTR)
 						continue;
@@ -533,29 +558,14 @@ int main(int argc, char *argv[])
 			bind_ns_files(getpid());
 	}
 
-	if (force_boottime)
-		settime(boottime, CLOCK_BOOTTIME);
-
-	if (force_monotonic)
-		settime(monotonic, CLOCK_MONOTONIC);
-
-	if (forkit) {
-		pid = fork();
-
-		switch(pid) {
-		case -1:
-			err(EXIT_FAILURE, _("fork failed"));
-		case 0:	/* child */
-			break;
-		default: /* parent */
-			if (waitpid(pid, &status, 0) == -1)
-				err(EXIT_FAILURE, _("waitpid failed"));
-			if (WIFEXITED(status))
-				return WEXITSTATUS(status);
-			else if (WIFSIGNALED(status))
-				kill(getpid(), WTERMSIG(status));
-			err(EXIT_FAILURE, _("child exit failed"));
-		}
+	if (pid) {
+		if (waitpid(pid, &status, 0) == -1)
+			err(EXIT_FAILURE, _("waitpid failed"));
+		if (WIFEXITED(status))
+			return WEXITSTATUS(status);
+		else if (WIFSIGNALED(status))
+			kill(getpid(), WTERMSIG(status));
+		err(EXIT_FAILURE, _("child exit failed"));
 	}
 
 	if (kill_child_signo != 0 && prctl(PR_SET_PDEATHSIG, kill_child_signo) < 0)
Signature · 1 line
2.17.1
maintainer
#1 lore raw

Eric, can you review this unshare(8) patch? Please.

Karel

Quoted 124 lines · On Wed, Apr 15, 2020 at 11:16:53PM +0200, [email protected] wrote:

From: michael-dev <[email protected]>

After unshare(...) is called, /proc/self/ns/pid does not change.
Instead, only /proc/self/ns/pid_for_children is affected. So bind-mounting /proc/self/ns/pid results in the original namespace getting bind-mounted.

Fix this by instead bind-mounting ns/pid_for_children.

---
 sys-utils/unshare.c | 66 ++++++++++++++++++++++++++-------------------
 1 file changed, 38 insertions(+), 28 deletions(-)
sys-utils/unshare.c +38 -28
index 8652ebdaf..c3ba18e32 100644
@@ -63,7 +63,7 @@ static struct namespace_file {
 	{ .type = CLONE_NEWIPC,   .name = "ns/ipc"  },
 	{ .type = CLONE_NEWUTS,   .name = "ns/uts"  },
 	{ .type = CLONE_NEWNET,   .name = "ns/net"  },
-	{ .type = CLONE_NEWPID,   .name = "ns/pid"  },
+	{ .type = CLONE_NEWPID,   .name = "ns/pid_for_children"  },
 	{ .type = CLONE_NEWNS,    .name = "ns/mnt"  },
 	{ .type = CLONE_NEWTIME,  .name = "ns/time"  },
 	{ .name = NULL }
@@ -361,6 +361,7 @@ int main(int argc, char *argv[])
 	const char *procmnt = NULL;
 	const char *newroot = NULL;
 	const char *newdir = NULL;
+	pid_t pid_bind = 0;
 	pid_t pid = 0;
 	int fds[2];
 	int status;
@@ -501,13 +502,37 @@ int main(int argc, char *argv[])
 			"unsharing of a time namespace (-t)"));
 
 	if (npersists && (unshare_flags & CLONE_NEWNS))
-		bind_ns_files_from_child(&pid, fds);
+		bind_ns_files_from_child(&pid_bind, fds);
 
 	if (-1 == unshare(unshare_flags))
 		err(EXIT_FAILURE, _("unshare failed"));
 
-	if (npersists) {
-		if (pid && (unshare_flags & CLONE_NEWNS)) {
+	if (force_boottime)
+		settime(boottime, CLOCK_BOOTTIME);
+
+	if (force_monotonic)
+		settime(monotonic, CLOCK_MONOTONIC);
+
+	if (forkit) {
+		// force child forking before mountspace binding
+		// so pid_for_children is populated
+		pid = fork();
+
+		switch(pid) {
+		case -1:
+			err(EXIT_FAILURE, _("fork failed"));
+		case 0:	/* child */
+			if (pid_bind && (unshare_flags & CLONE_NEWNS))
+				close(fds[1]);
+			break;
+		default: /* parent */
+      break;
+		}
+	}
+
+	if (npersists && (pid || !forkit)) {
+		// run in parent
+		if (pid_bind && (unshare_flags & CLONE_NEWNS)) {
 			int rc;
 			char ch = PIPE_SYNC_BYTE;
 
@@ -518,7 +543,7 @@ int main(int argc, char *argv[])
 
 			/* wait for bind_ns_files_from_child() */
 			do {
-				rc = waitpid(pid, &status, 0);
+				rc = waitpid(pid_bind, &status, 0);
 				if (rc < 0) {
 					if (errno == EINTR)
 						continue;
@@ -533,29 +558,14 @@ int main(int argc, char *argv[])
 			bind_ns_files(getpid());
 	}
 
-	if (force_boottime)
-		settime(boottime, CLOCK_BOOTTIME);
-
-	if (force_monotonic)
-		settime(monotonic, CLOCK_MONOTONIC);
-
-	if (forkit) {
-		pid = fork();
-
-		switch(pid) {
-		case -1:
-			err(EXIT_FAILURE, _("fork failed"));
-		case 0:	/* child */
-			break;
-		default: /* parent */
-			if (waitpid(pid, &status, 0) == -1)
-				err(EXIT_FAILURE, _("waitpid failed"));
-			if (WIFEXITED(status))
-				return WEXITSTATUS(status);
-			else if (WIFSIGNALED(status))
-				kill(getpid(), WTERMSIG(status));
-			err(EXIT_FAILURE, _("child exit failed"));
-		}
+	if (pid) {
+		if (waitpid(pid, &status, 0) == -1)
+			err(EXIT_FAILURE, _("waitpid failed"));
+		if (WIFEXITED(status))
+			return WEXITSTATUS(status);
+		else if (WIFSIGNALED(status))
+			kill(getpid(), WTERMSIG(status));
+		err(EXIT_FAILURE, _("child exit failed"));
 	}
 
 	if (kill_child_signo != 0 && prctl(PR_SET_PDEATHSIG, kill_child_signo) < 0)
Signature · 1 line
2.17.1
Signature · 2 lines
 Karel Zak  <[email protected]>
 http://karelzak.blogspot.com
#2 lore raw
In reply to [email protected]
Quoted 6 lines · [email protected] writes:

From: michael-dev <[email protected]>

After unshare(...) is called, /proc/self/ns/pid does not change.
Instead, only /proc/self/ns/pid_for_children is affected. So bind-mounting /proc/self/ns/pid results in the original namespace getting bind-mounted.

Fix this by instead bind-mounting ns/pid_for_children.

Why all of the extra code motion and change?

Your description sounds like only the first hunk is necessary. Did something else get mixed into this change? Or are all of the hunks necessary?

Eric

Quoted 114 lines
---
 sys-utils/unshare.c | 66 ++++++++++++++++++++++++++-------------------
 1 file changed, 38 insertions(+), 28 deletions(-)
sys-utils/unshare.c +38 -28
index 8652ebdaf..c3ba18e32 100644
@@ -63,7 +63,7 @@ static struct namespace_file {
 	{ .type = CLONE_NEWIPC,   .name = "ns/ipc"  },
 	{ .type = CLONE_NEWUTS,   .name = "ns/uts"  },
 	{ .type = CLONE_NEWNET,   .name = "ns/net"  },
-	{ .type = CLONE_NEWPID,   .name = "ns/pid"  },
+	{ .type = CLONE_NEWPID,   .name = "ns/pid_for_children"  },
 	{ .type = CLONE_NEWNS,    .name = "ns/mnt"  },
 	{ .type = CLONE_NEWTIME,  .name = "ns/time"  },
 	{ .name = NULL }
@@ -361,6 +361,7 @@ int main(int argc, char *argv[])
 	const char *procmnt = NULL;
 	const char *newroot = NULL;
 	const char *newdir = NULL;
+	pid_t pid_bind = 0;
 	pid_t pid = 0;
 	int fds[2];
 	int status;
@@ -501,13 +502,37 @@ int main(int argc, char *argv[])
 			"unsharing of a time namespace (-t)"));
 
 	if (npersists && (unshare_flags & CLONE_NEWNS))
-		bind_ns_files_from_child(&pid, fds);
+		bind_ns_files_from_child(&pid_bind, fds);
 
 	if (-1 == unshare(unshare_flags))
 		err(EXIT_FAILURE, _("unshare failed"));
 
-	if (npersists) {
-		if (pid && (unshare_flags & CLONE_NEWNS)) {
+	if (force_boottime)
+		settime(boottime, CLOCK_BOOTTIME);
+
+	if (force_monotonic)
+		settime(monotonic, CLOCK_MONOTONIC);
+
+	if (forkit) {
+		// force child forking before mountspace binding
+		// so pid_for_children is populated
+		pid = fork();
+
+		switch(pid) {
+		case -1:
+			err(EXIT_FAILURE, _("fork failed"));
+		case 0:	/* child */
+			if (pid_bind && (unshare_flags & CLONE_NEWNS))
+				close(fds[1]);
+			break;
+		default: /* parent */
+      break;
+		}
+	}
+
+	if (npersists && (pid || !forkit)) {
+		// run in parent
+		if (pid_bind && (unshare_flags & CLONE_NEWNS)) {
 			int rc;
 			char ch = PIPE_SYNC_BYTE;
 
@@ -518,7 +543,7 @@ int main(int argc, char *argv[])
 
 			/* wait for bind_ns_files_from_child() */
 			do {
-				rc = waitpid(pid, &status, 0);
+				rc = waitpid(pid_bind, &status, 0);
 				if (rc < 0) {
 					if (errno == EINTR)
 						continue;
@@ -533,29 +558,14 @@ int main(int argc, char *argv[])
 			bind_ns_files(getpid());
 	}
 
-	if (force_boottime)
-		settime(boottime, CLOCK_BOOTTIME);
-
-	if (force_monotonic)
-		settime(monotonic, CLOCK_MONOTONIC);
-
-	if (forkit) {
-		pid = fork();
-
-		switch(pid) {
-		case -1:
-			err(EXIT_FAILURE, _("fork failed"));
-		case 0:	/* child */
-			break;
-		default: /* parent */
-			if (waitpid(pid, &status, 0) == -1)
-				err(EXIT_FAILURE, _("waitpid failed"));
-			if (WIFEXITED(status))
-				return WEXITSTATUS(status);
-			else if (WIFSIGNALED(status))
-				kill(getpid(), WTERMSIG(status));
-			err(EXIT_FAILURE, _("child exit failed"));
-		}
+	if (pid) {
+		if (waitpid(pid, &status, 0) == -1)
+			err(EXIT_FAILURE, _("waitpid failed"));
+		if (WIFEXITED(status))
+			return WEXITSTATUS(status);
+		else if (WIFSIGNALED(status))
+			kill(getpid(), WTERMSIG(status));
+		err(EXIT_FAILURE, _("child exit failed"));
 	}
 
 	if (kill_child_signo != 0 && prctl(PR_SET_PDEATHSIG, kill_child_signo) < 0)
#3 lore raw

Am 16.04.2020 22:49, schrieb [email protected]:

Quoted 5 lines

Why all of the extra code motion and change?

Your description sounds like only the first hunk is necessary. Did something else get mixed into this change? Or are all of the hunks necessary?

Because after unsharing, pid_for_children is a dangling symlink.
So if (forkit) needs to be before /* wait for bind_ns_files_from_child() */.
Moving this results in an overlapping use of the pid variable, so that gets renamed.
Additionally, we cannot wait for the forkit child (bash) to exit before triggering and waiting for the bind-mounting child to bind-mount the persistent namespace. So forking and waiting for the forkit child (bash) needs to become splitted.

Regards,
Michael

#4 lore raw
Quoted 8 lines · michael-dev

Am 16.04.2020 22:49, schrieb [email protected]:

Quoted 5 lines

Why all of the extra code motion and change?

Your description sounds like only the first hunk is necessary. Did something else get mixed into this change? Or are all of the hunks necessary?

Quoted 9 lines

Because after unsharing, pid_for_children is a dangling symlink. So if (forkit) needs to be before /* wait for bind_ns_files_from_child() */.
Moving this results in an overlapping use of the pid variable, so that gets renamed.
Additionally, we cannot wait for the forkit child (bash) to exit before triggering and waiting for the bind-mounting child to bind-mount the persistent namespace. So forking and waiting for the forkit child (bash) needs to become splitted.

@Michael: I confim theproblem, and thanks for this fix.

@Zak: an analogous fix is required for time namespaces.

Cheers,

Michael

Signature · 3 lines
Michael Kerrisk Linux man-pages maintainer;
http://www.kernel.org/doc/man-pages/
Author of "The Linux Programming Interface", http://blog.man7.org/
maintainer
#5 lore raw
Quoted 19 lines · Michael Kerrisk
Quoted 8 lines · michael-dev

Am 16.04.2020 22:49, schrieb [email protected]:

Quoted 5 lines

Why all of the extra code motion and change?

Your description sounds like only the first hunk is necessary. Did something else get mixed into this change? Or are all of the hunks necessary?

Quoted 9 lines

Because after unsharing, pid_for_children is a dangling symlink. So if (forkit) needs to be before /* wait for bind_ns_files_from_child() */.
Moving this results in an overlapping use of the pid variable, so that gets renamed.
Additionally, we cannot wait for the forkit child (bash) to exit before triggering and waiting for the bind-mounting child to bind-mount the persistent namespace. So forking and waiting for the forkit child (bash) needs to become splitted.

Quoted 3 lines

@Michael: I confim theproblem, and thanks for this fix.

@Zak: an analogous fix is required for time namespaces.

OK, I'll play with it later this week. Sorry for the delay, I'm busy with other things.

Karel

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

[I should have CCed Adrian, so that he is aware of the issue also, having added Time Namespaces; Adrian, for context, see
https://lore.kernel.org/util-linux/[email protected]/T/#t
]

Quoted 21 lines · Karel Zak
Quoted 19 lines · Michael Kerrisk
Quoted 17 lines · michael-dev

Am 16.04.2020 22:49, schrieb [email protected]:

Quoted 5 lines

Why all of the extra code motion and change?

Your description sounds like only the first hunk is necessary. Did something else get mixed into this change? Or are all of the hunks necessary?

Because after unsharing, pid_for_children is a dangling symlink. So if (forkit) needs to be before /* wait for bind_ns_files_from_child() */.
Moving this results in an overlapping use of the pid variable, so that gets renamed.
Additionally, we cannot wait for the forkit child (bash) to exit before triggering and waiting for the bind-mounting child to bind-mount the persistent namespace. So forking and waiting for the forkit child (bash) needs to become splitted.

Quoted 6 lines
Quoted 3 lines

@Michael: I confim theproblem, and thanks for this fix.

@Zak: an analogous fix is required for time namespaces.

OK, I'll play with it later this week. Sorry for the delay, I'm busy with other things.

Thanks, Karel!

Signature · 3 lines
Michael Kerrisk
Linux man-pages maintainer; http://www.kernel.org/doc/man-pages/
Linux/UNIX System Programming Training: http://man7.org/training/
maintainer Applied
#7 lore raw
In reply to [email protected]
Quoted 2 lines · On Wed, Apr 15, 2020 at 11:16:53PM +0200, [email protected] wrote:
 sys-utils/unshare.c | 66 ++++++++++++++++++++++++++-------------------
 1 file changed, 38 insertions(+), 28 deletions(-)

Applied. I did small changes to the patch, see:
https://github.com/karelzak/util-linux/commit/0d5260b66c5581c8a5855a5f49e298e48e8baf82

Quoted 1 line · Michael Kerrisk

@Zak: an analogous fix is required for time namespaces.

A have added ns/time_for_children to the commit.

Thanks guys!

Karel

Signature · 2 lines
 Karel Zak  <[email protected]>
 http://karelzak.blogspot.com