Skip to contentSkip to conversation

fs/exec: switch timens when a task gets a new mm

PATCH Applied Andrei Vagin 4 participants 10 messages Sep 21, 2022 – Oct 18, 2022

Subsystems (5)
VFS Kselftest MM core Scheduler +1 more subsystems
lore
Share & tools
lore
#0 lore raw

PATCH 1/2 fs/exec: switch timens when a task gets a new mm

From: Andrei Vagin <[email protected]>

Changing a time namespace requires remapping a vvar page, so we don't want to allow doing that if any other tasks can use the same mm.

Currently, we install a time namespace when a task is created with a new vm. exec() is another case when a task gets a new mm and so it can switch a time namespace safely, but it isn't handled now.

One more issue of the current interface is that clone() with CLONE_VM isn't allowed if the current task has unshared a time namespace (timens_for_children doesn't match the current timens).

Both these issues make some inconvenience for users. For example, Alexey and Florian reported that posix_spawn() uses vfork+exec and this pattern doesn't work with time namespaces due to the both described issues. LXC needed to workaround the exec() issue by calling setns.

In the commit 133e2d3e81de5 ("fs/exec: allow to unshare a time namespace on vfork+exec"), we tried to fix these issues with minimal impact on UAPI. But it adds extra complexity and some undesirable side effects. Eric suggested fixing the issues properly because here are all the reasons to suppose that there are no users that depend on the old behavior.

Cc Alexey Izbyshev <[email protected]>
Cc Christian Brauner <[email protected]>
Cc Dmitry Safonov <[email protected]>
Cc "Eric W. Biederman" <[email protected]>
Cc Florian Weimer <[email protected]>
Cc Kees Cook <[email protected]>

Origin-author: "Eric W. Biederman" <[email protected]>

---
 fs/exec.c               |  5 +++++
 include/linux/nsproxy.h |  1 +
 kernel/fork.c           |  9 ---------
 kernel/nsproxy.c        | 23 +++++++++++++++++++++--
 4 files changed, 27 insertions(+), 11 deletions(-)
fs/exec.c +5 -0
index d046dbb9cbd0..71284188b96d 100644
@@ -65,6 +65,7 @@
 #include <linux/io_uring.h>
 #include <linux/syscall_user_dispatch.h>
 #include <linux/coredump.h>
+#include <linux/time_namespace.h>
 
 #include <linux/uaccess.h>
 #include <asm/mmu_context.h>
@@ -1296,6 +1297,10 @@ int begin_new_exec(struct linux_binprm * bprm)
 
 	bprm->mm = NULL;
 
+	retval = exec_task_namespaces();
+	if (retval)
+		goto out_unlock;
+
 #ifdef CONFIG_POSIX_TIMERS
 	spin_lock_irq(&me->sighand->siglock);
 	posix_cpu_timers_exit(me);
include/linux/nsproxy.h +1 -0
index cdb171efc7cb..fee881cded01 100644
@@ -94,6 +94,7 @@ static inline struct cred *nsset_cred(struct nsset *set)
 int copy_namespaces(unsigned long flags, struct task_struct *tsk);
 void exit_task_namespaces(struct task_struct *tsk);
 void switch_task_namespaces(struct task_struct *tsk, struct nsproxy *new);
+int exec_task_namespaces(void);
 void free_nsproxy(struct nsproxy *ns);
 int unshare_nsproxy_namespaces(unsigned long, struct nsproxy **,
 	struct cred *, struct fs_struct *);
kernel/fork.c +0 -9
index 2b6bd511c6ed..4eb803f75225 100644
@@ -2044,15 +2044,6 @@ static __latent_entropy struct task_struct *copy_process(
 			return ERR_PTR(-EINVAL);
 	}
 
-	/*
-	 * If the new process will be in a different time namespace
-	 * do not allow it to share VM or a thread group with the forking task.
-	 */
-	if (clone_flags & (CLONE_THREAD | CLONE_VM)) {
-		if (nsp->time_ns != nsp->time_ns_for_children)
-			return ERR_PTR(-EINVAL);
-	}
-
 	if (clone_flags & CLONE_PIDFD) {
 		/*
 		 * - CLONE_DETACHED is blocked so that we can potentially
kernel/nsproxy.c +21 -2
index eec72ca962e2..a487ff24129b 100644
@@ -157,7 +157,8 @@ int copy_namespaces(unsigned long flags, struct task_struct *tsk)
 	if (likely(!(flags & (CLONE_NEWNS | CLONE_NEWUTS | CLONE_NEWIPC |
 			      CLONE_NEWPID | CLONE_NEWNET |
 			      CLONE_NEWCGROUP | CLONE_NEWTIME)))) {
-		if (likely(old_ns->time_ns_for_children == old_ns->time_ns)) {
+		if ((flags & CLONE_VM) ||
+		    likely(old_ns->time_ns_for_children == old_ns->time_ns)) {
 			get_nsproxy(old_ns);
 			return 0;
 		}
@@ -179,7 +180,8 @@ int copy_namespaces(unsigned long flags, struct task_struct *tsk)
 	if (IS_ERR(new_ns))
 		return  PTR_ERR(new_ns);
 
-	timens_on_fork(new_ns, tsk);
+	if ((flags & CLONE_VM) == 0)
+		timens_on_fork(new_ns, tsk);
 
 	tsk->nsproxy = new_ns;
 	return 0;
@@ -254,6 +256,23 @@ void exit_task_namespaces(struct task_struct *p)
 	switch_task_namespaces(p, NULL);
 }
 
+int exec_task_namespaces(void)
+{
+	struct task_struct *tsk = current;
+	struct nsproxy *new;
+
+	if (tsk->nsproxy->time_ns_for_children == tsk->nsproxy->time_ns)
+		return 0;
+
+	new = create_new_namespaces(0, tsk, current_user_ns(), tsk->fs);
+	if (IS_ERR(new))
+		return PTR_ERR(new);
+
+	timens_on_fork(new, tsk);
+	switch_task_namespaces(tsk, new);
+	return 0;
+}
+
 static int check_setns_flags(unsigned long flags)
 {
 	if (!flags || (flags & ~(CLONE_NEWNS | CLONE_NEWUTS | CLONE_NEWIPC |
Signature · 1 line
2.37.3.968.ga6b4b080e4-goog
#1 lore raw

PATCH 2/2 selftests/timens: add a test for vfork+exit

From: Andrei Vagin <[email protected]>

* check that a child process is in parent's time namespace after vfork.
* check that a child process is in the target namespace after exec.

Output on success:
1..5
ok 1 parent before vfork
ok 2 child before exec
ok 3 child after exec
ok 4 wait for child
ok 5 parent after vfork
# Totals: pass:5 fail:0 xfail:0 xpass:0 skip:0 error:0

---
 tools/testing/selftests/timens/.gitignore   |   1 +
 tools/testing/selftests/timens/Makefile     |   2 +-
 tools/testing/selftests/timens/vfork_exec.c | 132 ++++++++++++++++++++
 3 files changed, 134 insertions(+), 1 deletion(-)
 create mode 100644 tools/testing/selftests/timens/vfork_exec.c
tools/testing/selftests/timens/.gitignore +1 -0
index fe1eb8271b35..cae8dca0fbff 100644
@@ -8,3 +8,4 @@ procfs
 timens
 timer
 timerfd
+vfork_exec
tools/testing/selftests/timens/Makefile +1 -1
index 3a5936cc10ab..f0d51d4d2c87 100644
@@ -1,4 +1,4 @@
-TEST_GEN_PROGS := timens timerfd timer clock_nanosleep procfs exec futex
+TEST_GEN_PROGS := timens timerfd timer clock_nanosleep procfs exec futex vfork_exec
 TEST_GEN_PROGS_EXTENDED := gettime_perf
 
 CFLAGS := -Wall -Werror -pthread
tools/testing/selftests/timens/vfork_exec.c new file +132 -0

132 lines not on this page · raw message

Signature · 1 line
2.37.3.968.ga6b4b080e4-goog
#2 lore raw

Re: [PATCH 2/2] selftests/timens: add a test for vfork+exit

Quoted 101 lines · Andrei Vagin

From: Andrei Vagin <[email protected]>

* check that a child process is in parent's time namespace after vfork.
* check that a child process is in the target namespace after exec.

Output on success:
1..5
ok 1 parent before vfork
ok 2 child before exec
ok 3 child after exec
ok 4 wait for child
ok 5 parent after vfork
# Totals: pass:5 fail:0 xfail:0 xpass:0 skip:0 error:0

---
 tools/testing/selftests/timens/.gitignore   |   1 +
 tools/testing/selftests/timens/Makefile     |   2 +-
 tools/testing/selftests/timens/vfork_exec.c | 132 ++++++++++++++++++++
 3 files changed, 134 insertions(+), 1 deletion(-)
 create mode 100644 tools/testing/selftests/timens/vfork_exec.c
tools/testing/selftests/timens/.gitignore +1 -0
index fe1eb8271b35..cae8dca0fbff 100644
@@ -8,3 +8,4 @@ procfs
 timens
 timer
 timerfd
+vfork_exec
tools/testing/selftests/timens/Makefile +0 -1
index 3a5936cc10ab..f0d51d4d2c87 100644
@@ -1,4 +1,4 @@
-TEST_GEN_PROGS := timens timerfd timer clock_nanosleep procfs exec 

futex
+TEST_GEN_PROGS := timens timerfd timer clock_nanosleep procfs exec futex vfork_exec
TEST_GEN_PROGS_EXTENDED := gettime_perf

CFLAGS := -Wall -Werror -pthread

tools/testing/selftests/timens/vfork_exec.c new file +49 -0
new file mode 100644
index 000000000000..9fd8a64d25a9
@@ -0,0 +1,132 @@
+// SPDX-License-Identifier: GPL-2.0
+#define _GNU_SOURCE
+#include <errno.h>
+#include <fcntl.h>
+#include <sched.h>
+#include <stdio.h>
+#include <stdbool.h>
+#include <sys/stat.h>
+#include <sys/syscall.h>
+#include <sys/types.h>
+#include <sys/wait.h>
+#include <time.h>
+#include <unistd.h>
+#include <string.h>
+#include <pthread.h>
+
+#include "log.h"
+#include "timens.h"
+
+#define OFFSET (36000)
+
+static void *tcheck(void *arg)
+{
+	struct timespec *now = arg, tst;
+	int i;
+
+	for (i = 0; i < 2; i++) {
+		_gettime(CLOCK_MONOTONIC, &tst, i);
+		if (abs(tst.tv_sec - now->tv_sec) > 5) {
+			pr_fail("thread: unexpected value: %ld (%ld)\n",
+				tst.tv_sec, now->tv_sec);
+			return (void *)1UL;
+		}
+	}
+	return NULL;
+}
+
+static int check_in_thread(struct timespec *now)
+{
+	pthread_t th;
+	void *retval;
+
+	if (pthread_create(&th, NULL, tcheck, now))
+		return pr_perror("thread");
+	if (pthread_join(th, &retval))
+		return pr_perror("pthread_join");
+	return !(retval == NULL);
+}
+

tools/testing/selftests/timens/vfork_exec.c line 59

Quoted 10 lines
tools/testing/selftests/timens/vfork_exec.c new file +10 -0
+static int check(char *tst_name, struct timespec *now)
+{
+	struct timespec tst;
+	int i;
+
+	for (i = 0; i < 2; i++) {
+		_gettime(CLOCK_MONOTONIC, &tst, i);
+		if (abs(tst.tv_sec - now->tv_sec) > 5)
+			return pr_fail("%s: unexpected value: %ld (%ld)\n",
+					tst.tv_sec, now->tv_sec);

There is a missing argument for "%s" in pr_fail(). I'm actually surprised that there is no __attribute__((format)) on ksft_test_result_fail() that would allow GCC to catch this.

Quoted 41 lines
tools/testing/selftests/timens/vfork_exec.c new file +41 -0
+	}
+	if (check_in_thread(now))
+		return 1;
+	ksft_test_result_pass("%s\n", tst_name);
+	return 0;
+}
+
+int main(int argc, char *argv[])
+{
+	struct timespec now;
+	int status;
+	pid_t pid;
+
+	if (argc > 1) {
+		char *endptr;
+
+		ksft_cnt.ksft_pass = 2;
+		now.tv_sec = strtoul(argv[1], &endptr, 0);
+		if (*endptr != 0)
+			return pr_perror("strtoul");
+
+		return check("child after exec", &now);
+	}
+
+	nscheck();
+
+	ksft_set_plan(5);
+
+	clock_gettime(CLOCK_MONOTONIC, &now);
+
+	if (unshare_timens())
+		return 1;
+
+	if (_settime(CLOCK_MONOTONIC, OFFSET))
+		return 1;
+
+	if (check("parent before vfork", &now))
+		return 1;
+
+	pid = vfork();
+	if (pid < 0)

tools/testing/selftests/timens/vfork_exec.c line 110

Quoted 10 lines
tools/testing/selftests/timens/vfork_exec.c new file +10 -0
+		return pr_perror("fork");
+
+	if (pid == 0) {
+		char now_str[64];
+		char *cargv[] = {"exec", now_str, NULL};
+		char *cenv[] = {NULL};
+
+		// Check that we are still in the source timens.
+		if (check("child before exec", &now))
+			return 1;

I know this is just a test, but...

Creating threads in a vfork()-child is quite dangerous (like most other things that touch the libc state, which is shared with the parent process). Here it works probably only because pthread_create() followed by pthread_join() restores everything into more-or-less the original state before returning control to the parent, but this is something that libcs don't guarantee and that can break at any moment.

Also, returning from a vfork()-child is explicitly forbidden by the vfork() contract because the parent would then return to an invalid stack frame that could be arbitrarily clobbered by code executed in the child after main() returned. Moreover, if I'm not mistaken, on x86 with Intel CET-enabled glibc (assuming the support for CET is ever merged into the kernel) such return would cause the parent to always trap because the shadow stack will become inconsistent with the normal stack. Instead, _exit() should be used here...

tools/testing/selftests/timens/vfork_exec.c line 115

Quoted 5 lines
tools/testing/selftests/timens/vfork_exec.c new file +5 -0
+
+		/* Check for proper vvar offsets after execve. */
+		snprintf(now_str, sizeof(now_str), "%ld", now.tv_sec + OFFSET);
+		execve("/proc/self/exe", cargv, cenv);
+		return pr_perror("execve");

...and here.

Quoted 7 lines
tools/testing/selftests/timens/vfork_exec.c new file +7 -0
+	}
+
+	if (waitpid(pid, &status, 0) != pid)
+		return pr_perror("waitpid");
+
+	if (status)
+		ksft_exit_fail();

tools/testing/selftests/timens/vfork_exec.c line 132

Quoted 10 lines
tools/testing/selftests/timens/vfork_exec.c new file +10 -0
+	ksft_inc_pass_cnt();
+	ksft_test_result_pass("wait for child\n");
+
+	// Check that we are still in the source timens.
+	if (check("parent after vfork", &now))
+		return 1;
+
+	ksft_exit_pass();
+	return 0;
+}

Otherwise, both patches look good to me, thanks!

Sorry for being late,
Alexey

#3 lore raw

Re: [PATCH 2/2] selftests/timens: add a test for vfork+exit

Quoted 3 lines · Alexey Izbyshev
Quoted 1 line · Andrei Vagin

From: Andrei Vagin <[email protected]>

<snip>

Quoted 18 lines

tools/testing/selftests/timens/vfork_exec.c line 110

Quoted 8 lines
tools/testing/selftests/timens/vfork_exec.c new file +8 -0
+     if (pid == 0) {
+             char now_str[64];
+             char *cargv[] = {"exec", now_str, NULL};
+             char *cenv[] = {NULL};
+
+             // Check that we are still in the source timens.
+             if (check("child before exec", &now))
+                     return 1;

I know this is just a test, but...

Creating threads in a vfork()-child is quite dangerous (like most other things that touch the libc state, which is shared with the parent process). Here it works probably only because pthread_create() followed by pthread_join() restores everything into more-or-less the original state before returning control to the parent, but this is something that libcs don't guarantee and that can break at any moment.

Quoted 9 lines

Also, returning from a vfork()-child is explicitly forbidden by the vfork() contract because the parent would then return to an invalid stack frame that could be arbitrarily clobbered by code executed in the child after main() returned. Moreover, if I'm not mistaken, on x86 with Intel CET-enabled glibc (assuming the support for CET is ever merged into the kernel) such return would cause the parent to always trap because the shadow stack will become inconsistent with the normal stack. Instead, _exit() should be used here...

Hi Alexey,

You are right, it isn't a good idea to create threads from the vfork-ed process. Now, vfork isn't a special case in the kernel code, so I think we can just remove the check() call from here. I have sent an updated version of this patch, pls take a look at it.

Thanks,
Andrei

#4 lore raw

Re: [PATCH 2/2] selftests/timens: add a test for vfork+exit

Quoted 35 lines · Andrei Vagin
Quoted 3 lines · Alexey Izbyshev
Quoted 1 line · Andrei Vagin

From: Andrei Vagin <[email protected]>

<snip>

Quoted 18 lines

tools/testing/selftests/timens/vfork_exec.c line 110

Quoted 8 lines
tools/testing/selftests/timens/vfork_exec.c new file +8 -0
+     if (pid == 0) {
+             char now_str[64];
+             char *cargv[] = {"exec", now_str, NULL};
+             char *cenv[] = {NULL};
+
+             // Check that we are still in the source timens.
+             if (check("child before exec", &now))
+                     return 1;

I know this is just a test, but...

Creating threads in a vfork()-child is quite dangerous (like most other things that touch the libc state, which is shared with the parent process). Here it works probably only because pthread_create() followed by pthread_join() restores everything into more-or-less the original state before returning control to the parent, but this is something that libcs don't guarantee and that can break at any moment.

Quoted 9 lines

Also, returning from a vfork()-child is explicitly forbidden by the vfork() contract because the parent would then return to an invalid stack frame that could be arbitrarily clobbered by code executed in the child after main() returned. Moreover, if I'm not mistaken, on x86 with Intel CET-enabled glibc (assuming the support for CET is ever merged into the kernel) such return would cause the parent to always trap because the shadow stack will become inconsistent with the normal stack. Instead, _exit() should be used here...

Quoted 7 lines

Hi Alexey,

You are right, it isn't a good idea to create threads from the vfork-ed process. Now, vfork isn't a special case in the kernel code, so I think we can just remove the check() call from here. I have sent an updated version of this patch, pls take a look at it.

Hi, Andrei,

While I think you could just skip check_in_thread() in the vfork()-child instead of removing check() completely (the rest of the code in check() is unlikely to mess up the libc state), I agree that the test is still able to catch problems unconditionally affecting all CLONE_VM tasks thanks to check_in_thread() in the parent, so I don't see much point in holding it up further. Your v2 patch looks good enough to me, thanks!

Alexey

#5 lore raw

PATCH v2 2/2 selftests/timens: add a test for vfork+exit

In reply to Andrei Vagin

From: Andrei Vagin <[email protected]>

* check that a child process is in parent's time namespace after vfork.
* check that a child process is in the target namespace after exec.

Output on success:
1..4
ok 1 parent before vfork
ok 2 child after exec
ok 3 wait for child
ok 4 parent after vfork
# Totals: pass:4 fail:0 xfail:0 xpass:0 skip:0 error:0

---

v2: - don't create threads from a vfork-ed process.
- use _exit to exit from a vfork-ed process.

 tools/testing/selftests/timens/.gitignore   |   1 +
 tools/testing/selftests/timens/Makefile     |   2 +-
 tools/testing/selftests/timens/vfork_exec.c | 139 ++++++++++++++++++++
 3 files changed, 141 insertions(+), 1 deletion(-)
 create mode 100644 tools/testing/selftests/timens/vfork_exec.c
tools/testing/selftests/timens/.gitignore +1 -0
index fe1eb8271b35..cae8dca0fbff 100644
@@ -8,3 +8,4 @@ procfs
 timens
 timer
 timerfd
+vfork_exec
tools/testing/selftests/timens/Makefile +1 -1
index 3a5936cc10ab..f0d51d4d2c87 100644
@@ -1,4 +1,4 @@
-TEST_GEN_PROGS := timens timerfd timer clock_nanosleep procfs exec futex
+TEST_GEN_PROGS := timens timerfd timer clock_nanosleep procfs exec futex vfork_exec
 TEST_GEN_PROGS_EXTENDED := gettime_perf
 
 CFLAGS := -Wall -Werror -pthread
tools/testing/selftests/timens/vfork_exec.c new file +139 -0

139 lines not on this page · raw message

Signature · 1 line
2.38.0.413.g74048e4d9e-goog
#6 lore raw
In reply to Andrei Vagin
Quoted 31 lines · Andrei Vagin

From: Andrei Vagin <[email protected]>

Changing a time namespace requires remapping a vvar page, so we don't want to allow doing that if any other tasks can use the same mm.

Currently, we install a time namespace when a task is created with a new vm. exec() is another case when a task gets a new mm and so it can switch a time namespace safely, but it isn't handled now.

One more issue of the current interface is that clone() with CLONE_VM isn't allowed if the current task has unshared a time namespace (timens_for_children doesn't match the current timens).

Both these issues make some inconvenience for users. For example, Alexey and Florian reported that posix_spawn() uses vfork+exec and this pattern doesn't work with time namespaces due to the both described issues. LXC needed to workaround the exec() issue by calling setns.

In the commit 133e2d3e81de5 ("fs/exec: allow to unshare a time namespace on vfork+exec"), we tried to fix these issues with minimal impact on UAPI. But it adds extra complexity and some undesirable side effects. Eric suggested fixing the issues properly because here are all the reasons to suppose that there are no users that depend on the old behavior.

Cc Alexey Izbyshev <[email protected]>
Cc Christian Brauner <[email protected]>
Cc Dmitry Safonov <[email protected]>
Cc "Eric W. Biederman" <[email protected]>
Cc Florian Weimer <[email protected]>
Cc Kees Cook <[email protected]>
Quoted 2 lines

Origin-author: "Eric W. Biederman" <[email protected]>

This looks good -- my intention is for this to go into -next after the v6.1 merge window closes. Does that match everyone's expectations?

Thanks!

-Kees

Signature · 1 line
Kees Cook
#7 lore raw

On Wed, Sep 21, 2022 at 8:29 PM Kees Cook <[email protected]> wrote:
...

Quoted 4 lines

This looks good -- my intention is for this to go into -next after the v6.1 merge window closes. Does that match everyone's expectations?

This works for me.

If everything goes well, it will be merged to v6.2. Do I understand this right?

Thanks,
Andrei

#8 lore raw
Quoted 10 lines · Andrei Vagin

On Wed, Sep 21, 2022 at 8:29 PM Kees Cook <[email protected]> wrote:
...

Quoted 4 lines

This looks good -- my intention is for this to go into -next after the v6.1 merge window closes. Does that match everyone's expectations?

This works for me.

If everything goes well, it will be merged to v6.2. Do I understand this right?

Right, that's what I'm expecting. :)

Signature · 1 line
Kees Cook
Applied
#9 lore raw
In reply to Andrei Vagin
Quoted 10 lines · Andrei Vagin

From: Andrei Vagin <[email protected]>

Changing a time namespace requires remapping a vvar page, so we don't want to allow doing that if any other tasks can use the same mm.

Currently, we install a time namespace when a task is created with a new vm. exec() is another case when a task gets a new mm and so it can switch a time namespace safely, but it isn't handled now.

[...]

Applied to for-next/execve, thanks!

[1/2] fs/exec: switch timens when a task gets a new mm

      https://git.kernel.org/kees/c/f8e06046dd69
[2/2] selftests/timens: add a test for vfork+exit
      https://git.kernel.org/kees/c/de8517320d01
Signature · 1 line
Kees Cook