Skip to contentSkip to conversation

fork: extend clone3() to support CLONE_SET_TID

PATCH Superseded Adrian Reber 2 participants 7 messages Aug 8, 2019 – Aug 11, 2019

Subsystems (5)

v4 of 5 Superseded Latest is v5 →

lore
Share & tools
lore
#0 lore raw

PATCH v4 1/2 fork: extend clone3() to support CLONE_SET_TID

The main motivation to add set_tid to clone3() is CRIU.

To restore a process with the same PID/TID CRIU currently uses /proc/sys/kernel/ns_last_pid. It writes the desired (PID - 1) to ns_last_pid and then (quickly) does a clone(). This works most of the time, but it is racy. It is also slow as it requires multiple syscalls.

Extending clone3() to support set_tid makes it possible restore a process using CRIU without accessing /proc/sys/kernel/ns_last_pid and race free (as long as the desired PID/TID is available).

This clone3() extension places the same restrictions (CAP_SYS_ADMIN) on clone3() with set_tid as they are currently in place for ns_last_pid.

---
v2:
- Removed (size < sizeof(struct clone_args)) as discussed with Christian and Dmitry
- Added comment to ((set_tid != 1) && idr_get_cursor() <= 1) (Oleg)
- Use idr_alloc() instead of idr_alloc_cyclic() (Oleg)

v3:
- Return EEXIST if PID is already in use (Christian)
- Drop CLONE_SET_TID (Christian and Oleg)
- Use idr_is_empty() instead of idr_get_cursor() (Oleg)
- Handle different struct clone_args sizes (Dmitry)

v4:
- Rework struct size check with defines (Christian)
- Reduce number of set_tid checks (Oleg)
- Less parentheses and more robust code (Oleg)
- Do ns_capable() on correct user_ns (Oleg, Christian)

---
 include/linux/pid.h        |  2 +-
 include/linux/sched/task.h |  1 +
 include/uapi/linux/sched.h |  1 +
 kernel/fork.c              | 25 +++++++++++++++++++++++--
 kernel/pid.c               | 34 +++++++++++++++++++++++++++-------
 5 files changed, 53 insertions(+), 10 deletions(-)
include/linux/pid.h +1 -1
index 2a83e434db9d..052000db0ced 100644
@@ -116,7 +116,7 @@ extern struct pid *find_vpid(int nr);
 extern struct pid *find_get_pid(int nr);
 extern struct pid *find_ge_pid(int nr, struct pid_namespace *);
 
-extern struct pid *alloc_pid(struct pid_namespace *ns);
+extern struct pid *alloc_pid(struct pid_namespace *ns, pid_t set_tid);
 extern void free_pid(struct pid *pid);
 extern void disable_pid_allocation(struct pid_namespace *ns);
 
include/linux/sched/task.h +1 -0
index 0497091e40c1..4f2a80564332 100644
@@ -26,6 +26,7 @@ struct kernel_clone_args {
 	unsigned long stack;
 	unsigned long stack_size;
 	unsigned long tls;
+	pid_t set_tid;
 };
 
 /*
include/uapi/linux/sched.h +1 -0
index b3105ac1381a..e1ce103a2c47 100644
@@ -45,6 +45,7 @@ struct clone_args {
 	__aligned_u64 stack;
 	__aligned_u64 stack_size;
 	__aligned_u64 tls;
+	__aligned_u64 set_tid;
 };
 
 /*
kernel/fork.c +23 -2
index 2852d0e76ea3..2a03f0e201e9 100644
@@ -117,6 +117,13 @@
  */
 #define MAX_THREADS FUTEX_TID_MASK
 
+/*
+ * Different sizes of struct clone_args
+ */
+#define CLONE3_ARGS_SIZE_V0 64
+/* V1 includes set_tid */
+#define CLONE3_ARGS_SIZE_V1 72
+
 /*
  * Protected counters by write_lock_irq(&tasklist_lock)
  */
@@ -2031,7 +2038,13 @@ static __latent_entropy struct task_struct *copy_process(
 	stackleak_task_init(p);
 
 	if (pid != &init_struct_pid) {
-		pid = alloc_pid(p->nsproxy->pid_ns_for_children);
+		if (args->set_tid && !ns_capable(
+				p->nsproxy->pid_ns_for_children->user_ns,
+				CAP_SYS_ADMIN)) {
+			retval = -EPERM;
+			goto bad_fork_cleanup_thread;
+		}
+		pid = alloc_pid(p->nsproxy->pid_ns_for_children, args->set_tid);
 		if (IS_ERR(pid)) {
 			retval = PTR_ERR(pid);
 			goto bad_fork_cleanup_thread;
@@ -2535,9 +2548,14 @@ noinline static int copy_clone_args_from_user(struct kernel_clone_args *kargs,
 	if (unlikely(size > PAGE_SIZE))
 		return -E2BIG;
 
-	if (unlikely(size < sizeof(struct clone_args)))
+	/* The struct needs to be at least the size of the original struct. */
+	if (unlikely(size < CLONE3_ARGS_SIZE_V0))
 		return -EINVAL;
 
+	if (size < sizeof(struct clone_args))
+		memset((void *)&args + size, 0,
+				sizeof(struct clone_args) - size);
+
 	if (unlikely(!access_ok(uargs, size)))
 		return -EFAULT;
 
@@ -2573,6 +2591,9 @@ noinline static int copy_clone_args_from_user(struct kernel_clone_args *kargs,
 		.tls		= args.tls,
 	};
 
+	if (size >= CLONE3_ARGS_SIZE_V1)
+		kargs->set_tid = args.set_tid;
+
 	return 0;
 }
 
kernel/pid.c +27 -7
index 0a9f2e437217..9ce89c35c5be 100644
@@ -157,7 +157,7 @@ void free_pid(struct pid *pid)
 	call_rcu(&pid->rcu, delayed_put_pid);
 }
 
-struct pid *alloc_pid(struct pid_namespace *ns)
+struct pid *alloc_pid(struct pid_namespace *ns, int set_tid)
 {
 	struct pid *pid;
 	enum pid_type type;
@@ -186,12 +186,32 @@ struct pid *alloc_pid(struct pid_namespace *ns)
 		if (idr_get_cursor(&tmp->idr) > RESERVED_PIDS)
 			pid_min = RESERVED_PIDS;
 
-		/*
-		 * Store a null pointer so find_pid_ns does not find
-		 * a partially initialized PID (see below).
-		 */
-		nr = idr_alloc_cyclic(&tmp->idr, NULL, pid_min,
-				      pid_max, GFP_ATOMIC);
+		if (set_tid) {
+			/*
+			 * Also fail if a PID != 1 is requested
+			 * and no PID 1 exists.
+			 */
+			nr = -EINVAL;
+			if (set_tid < pid_max && set_tid > 0 &&
+			    (set_tid == 1 || !idr_is_empty(&tmp->idr)))
+				nr = idr_alloc(&tmp->idr, NULL, set_tid,
+					       set_tid + 1, GFP_ATOMIC);
+			/*
+			 * If ENOSPC is returned it means that the PID is
+			 * alreay in use. Return EEXIST in that case.
+			 */
+			if (nr == -ENOSPC)
+				nr = -EEXIST;
+			/* Only use set_tid for one PID namespace. */
+			set_tid = 0;
+		} else {
+			/*
+			 * Store a null pointer so find_pid_ns does not find
+			 * a partially initialized PID (see below).
+			 */
+			nr = idr_alloc_cyclic(&tmp->idr, NULL, pid_min,
+					      pid_max, GFP_ATOMIC);
+		}
 		spin_unlock_irq(&pidmap_lock);
 		idr_preload_end();
 
Signature · 1 line
2.21.0
#1 lore raw

PATCH v4 2/2 selftests: add tests for clone3()

This tests clone3() with and without set_tid to see if all desired PIDs are working as expected. The test tries to clone3() with a set_tid of -1, 1, pid_max, a PID which is already in use and an unused PID. The same tests are also running in PID namespace.

---
 tools/testing/selftests/clone3/.gitignore     |   2 +
 tools/testing/selftests/clone3/Makefile       |  11 ++
 tools/testing/selftests/clone3/clone3.c       | 141 +++++++++++++++
 .../testing/selftests/clone3/clone3_set_tid.c | 161 ++++++++++++++++++
 4 files changed, 315 insertions(+)
 create mode 100644 tools/testing/selftests/clone3/.gitignore
 create mode 100644 tools/testing/selftests/clone3/Makefile
 create mode 100644 tools/testing/selftests/clone3/clone3.c
 create mode 100644 tools/testing/selftests/clone3/clone3_set_tid.c
tools/testing/selftests/clone3/.gitignore new file +2 -0
new file mode 100644
index 000000000000..c63c64a78ddf
@@ -0,0 +1,2 @@
+clone3_set_tid
+clone3
tools/testing/selftests/clone3/Makefile new file +11 -0
new file mode 100644
index 000000000000..4efcf45b995b
@@ -0,0 +1,11 @@
+# SPDX-License-Identifier: GPL-2.0
+uname_M := $(shell uname -m 2>/dev/null || echo not)
+ARCH ?= $(shell echo $(uname_M) | sed -e s/i.86/i386/)
+
+CFLAGS += -I../../../../usr/include/
+
+ifeq ($(ARCH),x86_64)
+	TEST_GEN_PROGS := clone3 clone3_set_tid
+endif
+
+include ../lib.mk
tools/testing/selftests/clone3/clone3.c new file +141 -0

141 lines not on this page · raw message

tools/testing/selftests/clone3/clone3_set_tid.c new file +161 -0

161 lines not on this page · raw message

Signature · 1 line
2.21.0
#2 lore raw
In reply to Adrian Reber
Quoted 112 lines · Adrian Reber

The main motivation to add set_tid to clone3() is CRIU.

To restore a process with the same PID/TID CRIU currently uses /proc/sys/kernel/ns_last_pid. It writes the desired (PID - 1) to ns_last_pid and then (quickly) does a clone(). This works most of the time, but it is racy. It is also slow as it requires multiple syscalls.

Extending clone3() to support set_tid makes it possible restore a process using CRIU without accessing /proc/sys/kernel/ns_last_pid and race free (as long as the desired PID/TID is available).

This clone3() extension places the same restrictions (CAP_SYS_ADMIN) on clone3() with set_tid as they are currently in place for ns_last_pid.

---
v2:
- Removed (size < sizeof(struct clone_args)) as discussed with Christian and Dmitry
- Added comment to ((set_tid != 1) && idr_get_cursor() <= 1) (Oleg)
- Use idr_alloc() instead of idr_alloc_cyclic() (Oleg)

v3:
- Return EEXIST if PID is already in use (Christian)
- Drop CLONE_SET_TID (Christian and Oleg)
- Use idr_is_empty() instead of idr_get_cursor() (Oleg)
- Handle different struct clone_args sizes (Dmitry)

v4:
- Rework struct size check with defines (Christian)
- Reduce number of set_tid checks (Oleg)
- Less parentheses and more robust code (Oleg)
- Do ns_capable() on correct user_ns (Oleg, Christian)

---
 include/linux/pid.h        |  2 +-
 include/linux/sched/task.h |  1 +
 include/uapi/linux/sched.h |  1 +
 kernel/fork.c              | 25 +++++++++++++++++++++++--
 kernel/pid.c               | 34 +++++++++++++++++++++++++++-------
 5 files changed, 53 insertions(+), 10 deletions(-)
include/linux/pid.h +1 -1
index 2a83e434db9d..052000db0ced 100644
@@ -116,7 +116,7 @@ extern struct pid *find_vpid(int nr);
 extern struct pid *find_get_pid(int nr);
 extern struct pid *find_ge_pid(int nr, struct pid_namespace *);
 
-extern struct pid *alloc_pid(struct pid_namespace *ns);
+extern struct pid *alloc_pid(struct pid_namespace *ns, pid_t set_tid);
 extern void free_pid(struct pid *pid);
 extern void disable_pid_allocation(struct pid_namespace *ns);
 
include/linux/sched/task.h +1 -0
index 0497091e40c1..4f2a80564332 100644
@@ -26,6 +26,7 @@ struct kernel_clone_args {
 	unsigned long stack;
 	unsigned long stack_size;
 	unsigned long tls;
+	pid_t set_tid;
 };
 
 /*
include/uapi/linux/sched.h +1 -0
index b3105ac1381a..e1ce103a2c47 100644
@@ -45,6 +45,7 @@ struct clone_args {
 	__aligned_u64 stack;
 	__aligned_u64 stack_size;
 	__aligned_u64 tls;
+	__aligned_u64 set_tid;
 };
 
 /*
kernel/fork.c +14 -1
index 2852d0e76ea3..2a03f0e201e9 100644
@@ -117,6 +117,13 @@
  */
 #define MAX_THREADS FUTEX_TID_MASK
 
+/*
+ * Different sizes of struct clone_args
+ */
+#define CLONE3_ARGS_SIZE_V0 64
+/* V1 includes set_tid */
+#define CLONE3_ARGS_SIZE_V1 72
+
 /*
  * Protected counters by write_lock_irq(&tasklist_lock)
  */
@@ -2031,7 +2038,13 @@ static __latent_entropy struct task_struct *copy_process(
 	stackleak_task_init(p);
 
 	if (pid != &init_struct_pid) {
-		pid = alloc_pid(p->nsproxy->pid_ns_for_children);
+		if (args->set_tid && !ns_capable(
+				p->nsproxy->pid_ns_for_children->user_ns,
+				CAP_SYS_ADMIN)) {
+			retval = -EPERM;
+			goto bad_fork_cleanup_thread;
+		}
+		pid = alloc_pid(p->nsproxy->pid_ns_for_children, args->set_tid);
 		if (IS_ERR(pid)) {
 			retval = PTR_ERR(pid);
 			goto bad_fork_cleanup_thread;

kernel/fork.c noinline static int copy_clone_args_from_user(struct kernel_clone_args *kargs, · line 2551

Quoted 5 lines
kernel/fork.c +1 -1
@@ -2535,9 +2548,14 @@ noinline static int copy_clone_args_from_user(struct kernel_clone_args *kargs,
 	if (unlikely(size > PAGE_SIZE))
 		return -E2BIG;
 
-	if (unlikely(size < sizeof(struct clone_args)))
+	/* The struct needs to be at least the size of the original struct. */

I don't think you need that comment. I think the macro is pretty self-explanatory. If you want it to be even clearer you could even make it CLONE3_ARGS_SIZE_MIN but V0 is good enough. :)

Quoted 11 lines
kernel/fork.c +5 -0
+	if (unlikely(size < CLONE3_ARGS_SIZE_V0))
 		return -EINVAL;
 
+	if (size < sizeof(struct clone_args))
+		memset((void *)&args + size, 0,
+				sizeof(struct clone_args) - size);
+
 	if (unlikely(!access_ok(uargs, size)))
 		return -EFAULT;
 

kernel/fork.c noinline static int copy_clone_args_from_user(struct kernel_clone_args *kargs, · line 2595

Quoted 5 lines
kernel/fork.c +2 -0
@@ -2573,6 +2591,9 @@ noinline static int copy_clone_args_from_user(struct kernel_clone_args *kargs,
 		.tls		= args.tls,
 	};
 
+	if (size >= CLONE3_ARGS_SIZE_V1)
+		kargs->set_tid = args.set_tid;

Hm, the if-condition is not needed though, right? At this point we will have already copied from struct clone_args __user *uargs into struct clone_args args. If we hit that codepath that means the kernel definitely has a field for set_tid in its struct clone_args. :) So this could probably just be:

.tls = args.tls,

		.set_tid	= args.set_tid,
	}

	?
Quoted 24 lines
kernel/fork.c +1 -0
+
 	return 0;
 }
 
kernel/pid.c +1 -4
index 0a9f2e437217..9ce89c35c5be 100644
@@ -157,7 +157,7 @@ void free_pid(struct pid *pid)
 	call_rcu(&pid->rcu, delayed_put_pid);
 }
 
-struct pid *alloc_pid(struct pid_namespace *ns)
+struct pid *alloc_pid(struct pid_namespace *ns, int set_tid)
 {
 	struct pid *pid;
 	enum pid_type type;
@@ -186,12 +186,32 @@ struct pid *alloc_pid(struct pid_namespace *ns)
 		if (idr_get_cursor(&tmp->idr) > RESERVED_PIDS)
 			pid_min = RESERVED_PIDS;
 
-		/*
-		 * Store a null pointer so find_pid_ns does not find
-		 * a partially initialized PID (see below).

kernel/pid.c struct pid *alloc_pid(struct pid_namespace *ns) · line 195

Quoted 10 lines
kernel/pid.c +7 -3
-		 */
-		nr = idr_alloc_cyclic(&tmp->idr, NULL, pid_min,
-				      pid_max, GFP_ATOMIC);
+		if (set_tid) {
+			/*
+			 * Also fail if a PID != 1 is requested
+			 * and no PID 1 exists.
+			 */
+			nr = -EINVAL;
+			if (set_tid < pid_max && set_tid > 0 &&

Hm, you're already in the if-branch hat verified if (set_tid) so the set_tid > 0 conjunct seems redundant. :)

kernel/pid.c struct pid *alloc_pid(struct pid_namespace *ns) · line 198

Quoted 3 lines
kernel/pid.c +3 -0
+			    (set_tid == 1 || !idr_is_empty(&tmp->idr)))
+				nr = idr_alloc(&tmp->idr, NULL, set_tid,
+					       set_tid + 1, GFP_ATOMIC);

I'm confused, shouldn't this be

	if (set_tid < pid_max || (set_tid == 1 && !idr_is_emtpy(&tmp->idf)))
Quoted 22 lines
kernel/pid.c +16 -0
+			/*
+			 * If ENOSPC is returned it means that the PID is
+			 * alreay in use. Return EEXIST in that case.
+			 */
+			if (nr == -ENOSPC)
+				nr = -EEXIST;
+			/* Only use set_tid for one PID namespace. */
+			set_tid = 0;
+		} else {
+			/*
+			 * Store a null pointer so find_pid_ns does not find
+			 * a partially initialized PID (see below).
+			 */
+			nr = idr_alloc_cyclic(&tmp->idr, NULL, pid_min,
+					      pid_max, GFP_ATOMIC);
+		}
 		spin_unlock_irq(&pidmap_lock);
 		idr_preload_end();
 
Signature · 1 line
2.21.0
#3 lore raw
Quoted 116 lines · Christian Brauner
Quoted 115 lines · Adrian Reber

The main motivation to add set_tid to clone3() is CRIU.

To restore a process with the same PID/TID CRIU currently uses /proc/sys/kernel/ns_last_pid. It writes the desired (PID - 1) to ns_last_pid and then (quickly) does a clone(). This works most of the time, but it is racy. It is also slow as it requires multiple syscalls.

Extending clone3() to support set_tid makes it possible restore a process using CRIU without accessing /proc/sys/kernel/ns_last_pid and race free (as long as the desired PID/TID is available).

This clone3() extension places the same restrictions (CAP_SYS_ADMIN) on clone3() with set_tid as they are currently in place for ns_last_pid.

---
v2:
- Removed (size < sizeof(struct clone_args)) as discussed with Christian and Dmitry
- Added comment to ((set_tid != 1) && idr_get_cursor() <= 1) (Oleg)
- Use idr_alloc() instead of idr_alloc_cyclic() (Oleg)

v3:
- Return EEXIST if PID is already in use (Christian)
- Drop CLONE_SET_TID (Christian and Oleg)
- Use idr_is_empty() instead of idr_get_cursor() (Oleg)
- Handle different struct clone_args sizes (Dmitry)

v4:
- Rework struct size check with defines (Christian)
- Reduce number of set_tid checks (Oleg)
- Less parentheses and more robust code (Oleg)
- Do ns_capable() on correct user_ns (Oleg, Christian)

---
 include/linux/pid.h        |  2 +-
 include/linux/sched/task.h |  1 +
 include/uapi/linux/sched.h |  1 +
 kernel/fork.c              | 25 +++++++++++++++++++++++--
 kernel/pid.c               | 34 +++++++++++++++++++++++++++-------
 5 files changed, 53 insertions(+), 10 deletions(-)
include/linux/pid.h +1 -1
index 2a83e434db9d..052000db0ced 100644
@@ -116,7 +116,7 @@ extern struct pid *find_vpid(int nr);
 extern struct pid *find_get_pid(int nr);
 extern struct pid *find_ge_pid(int nr, struct pid_namespace *);
 
-extern struct pid *alloc_pid(struct pid_namespace *ns);
+extern struct pid *alloc_pid(struct pid_namespace *ns, pid_t set_tid);
 extern void free_pid(struct pid *pid);
 extern void disable_pid_allocation(struct pid_namespace *ns);
 
include/linux/sched/task.h +1 -0
index 0497091e40c1..4f2a80564332 100644
@@ -26,6 +26,7 @@ struct kernel_clone_args {
 	unsigned long stack;
 	unsigned long stack_size;
 	unsigned long tls;
+	pid_t set_tid;
 };
 
 /*
include/uapi/linux/sched.h +1 -0
index b3105ac1381a..e1ce103a2c47 100644
@@ -45,6 +45,7 @@ struct clone_args {
 	__aligned_u64 stack;
 	__aligned_u64 stack_size;
 	__aligned_u64 tls;
+	__aligned_u64 set_tid;
 };
 
 /*
kernel/fork.c +14 -1
index 2852d0e76ea3..2a03f0e201e9 100644
@@ -117,6 +117,13 @@
  */
 #define MAX_THREADS FUTEX_TID_MASK
 
+/*
+ * Different sizes of struct clone_args
+ */
+#define CLONE3_ARGS_SIZE_V0 64
+/* V1 includes set_tid */
+#define CLONE3_ARGS_SIZE_V1 72
+
 /*
  * Protected counters by write_lock_irq(&tasklist_lock)
  */
@@ -2031,7 +2038,13 @@ static __latent_entropy struct task_struct *copy_process(
 	stackleak_task_init(p);
 
 	if (pid != &init_struct_pid) {
-		pid = alloc_pid(p->nsproxy->pid_ns_for_children);
+		if (args->set_tid && !ns_capable(
+				p->nsproxy->pid_ns_for_children->user_ns,
+				CAP_SYS_ADMIN)) {
+			retval = -EPERM;
+			goto bad_fork_cleanup_thread;
+		}
+		pid = alloc_pid(p->nsproxy->pid_ns_for_children, args->set_tid);
 		if (IS_ERR(pid)) {
 			retval = PTR_ERR(pid);
 			goto bad_fork_cleanup_thread;
@@ -2535,9 +2548,14 @@ noinline static int copy_clone_args_from_user(struct kernel_clone_args *kargs,
 	if (unlikely(size > PAGE_SIZE))
 		return -E2BIG;
 
Quoted 6 lines

kernel/fork.c noinline static int copy_clone_args_from_user(struct kernel_clone_args *kargs, · line 2551

Quoted 2 lines
kernel/fork.c +1 -1
-	if (unlikely(size < sizeof(struct clone_args)))
+	/* The struct needs to be at least the size of the original struct. */

I don't think you need that comment. I think the macro is pretty self-explanatory. If you want it to be even clearer you could even make it CLONE3_ARGS_SIZE_MIN but V0 is good enough. :)

Will remove the comment.

Quoted 23 lines
Quoted 11 lines
kernel/fork.c +5 -0
+	if (unlikely(size < CLONE3_ARGS_SIZE_V0))
 		return -EINVAL;
 
+	if (size < sizeof(struct clone_args))
+		memset((void *)&args + size, 0,
+				sizeof(struct clone_args) - size);
+
 	if (unlikely(!access_ok(uargs, size)))
 		return -EFAULT;
 

kernel/fork.c noinline static int copy_clone_args_from_user(struct kernel_clone_args *kargs, · line 2595

Quoted 5 lines
kernel/fork.c +2 -0
@@ -2573,6 +2591,9 @@ noinline static int copy_clone_args_from_user(struct kernel_clone_args *kargs,
 		.tls		= args.tls,
 	};
 
+	if (size >= CLONE3_ARGS_SIZE_V1)
+		kargs->set_tid = args.set_tid;

Hm, the if-condition is not needed though, right? At this point we will have already copied from struct clone_args __user *uargs into struct clone_args args. If we hit that codepath that means the kernel definitely has a field for set_tid in its struct clone_args. :) So this could probably just be:

Quoted 5 lines
   		.tls		= args.tls,
		.set_tid	= args.set_tid,
	}

	?

Right.

Quoted 21 lines
Quoted 21 lines
kernel/fork.c +1 -0
+
 	return 0;
 }
 
kernel/pid.c +1 -1
index 0a9f2e437217..9ce89c35c5be 100644
@@ -157,7 +157,7 @@ void free_pid(struct pid *pid)
 	call_rcu(&pid->rcu, delayed_put_pid);
 }
 
-struct pid *alloc_pid(struct pid_namespace *ns)
+struct pid *alloc_pid(struct pid_namespace *ns, int set_tid)
 {
 	struct pid *pid;
 	enum pid_type type;
@@ -186,12 +186,32 @@ struct pid *alloc_pid(struct pid_namespace *ns)
 		if (idr_get_cursor(&tmp->idr) > RESERVED_PIDS)
 			pid_min = RESERVED_PIDS;
 
Quoted 16 lines

kernel/pid.c struct pid *alloc_pid(struct pid_namespace *ns) · line 195

Quoted 13 lines
kernel/pid.c +7 -6
-		/*
-		 * Store a null pointer so find_pid_ns does not find
-		 * a partially initialized PID (see below).
-		 */
-		nr = idr_alloc_cyclic(&tmp->idr, NULL, pid_min,
-				      pid_max, GFP_ATOMIC);
+		if (set_tid) {
+			/*
+			 * Also fail if a PID != 1 is requested
+			 * and no PID 1 exists.
+			 */
+			nr = -EINVAL;
+			if (set_tid < pid_max && set_tid > 0 &&

Hm, you're already in the if-branch hat verified if (set_tid) so the set_tid > 0 conjunct seems redundant. :)

Yes, but I dropped all checks to see if set_tid is negative as suggested by Oleg and moved it here.

Quoted 7 lines

kernel/pid.c struct pid *alloc_pid(struct pid_namespace *ns) · line 198

Quoted 3 lines
kernel/pid.c +3 -0
+			    (set_tid == 1 || !idr_is_empty(&tmp->idr)))
+				nr = idr_alloc(&tmp->idr, NULL, set_tid,
+					       set_tid + 1, GFP_ATOMIC);

I'm confused, shouldn't this be

	if (set_tid < pid_max || (set_tid == 1 && !idr_is_emtpy(&tmp->idf)))

Now I am also confused ;). This does not work. This will always return true if set_tid is less than pid_max. So pid_max needs to be something like 1 for the check after || to make sense, right? But you really got me confused here right now. Right now I still think what I did is correct.

Quoted 22 lines
Quoted 22 lines
kernel/pid.c +16 -0
+			/*
+			 * If ENOSPC is returned it means that the PID is
+			 * alreay in use. Return EEXIST in that case.
+			 */
+			if (nr == -ENOSPC)
+				nr = -EEXIST;
+			/* Only use set_tid for one PID namespace. */
+			set_tid = 0;
+		} else {
+			/*
+			 * Store a null pointer so find_pid_ns does not find
+			 * a partially initialized PID (see below).
+			 */
+			nr = idr_alloc_cyclic(&tmp->idr, NULL, pid_min,
+					      pid_max, GFP_ATOMIC);
+		}
 		spin_unlock_irq(&pidmap_lock);
 		idr_preload_end();
 
Signature · 1 line
2.21.0
#4 lore raw
Quoted 204 lines · Adrian Reber
Quoted 116 lines · Christian Brauner
Quoted 115 lines · Adrian Reber

The main motivation to add set_tid to clone3() is CRIU.

To restore a process with the same PID/TID CRIU currently uses /proc/sys/kernel/ns_last_pid. It writes the desired (PID - 1) to ns_last_pid and then (quickly) does a clone(). This works most of the time, but it is racy. It is also slow as it requires multiple syscalls.

Extending clone3() to support set_tid makes it possible restore a process using CRIU without accessing /proc/sys/kernel/ns_last_pid and race free (as long as the desired PID/TID is available).

This clone3() extension places the same restrictions (CAP_SYS_ADMIN) on clone3() with set_tid as they are currently in place for ns_last_pid.

---
v2:
- Removed (size < sizeof(struct clone_args)) as discussed with Christian and Dmitry
- Added comment to ((set_tid != 1) && idr_get_cursor() <= 1) (Oleg)
- Use idr_alloc() instead of idr_alloc_cyclic() (Oleg)

v3:
- Return EEXIST if PID is already in use (Christian)
- Drop CLONE_SET_TID (Christian and Oleg)
- Use idr_is_empty() instead of idr_get_cursor() (Oleg)
- Handle different struct clone_args sizes (Dmitry)

v4:
- Rework struct size check with defines (Christian)
- Reduce number of set_tid checks (Oleg)
- Less parentheses and more robust code (Oleg)
- Do ns_capable() on correct user_ns (Oleg, Christian)

---
 include/linux/pid.h        |  2 +-
 include/linux/sched/task.h |  1 +
 include/uapi/linux/sched.h |  1 +
 kernel/fork.c              | 25 +++++++++++++++++++++++--
 kernel/pid.c               | 34 +++++++++++++++++++++++++++-------
 5 files changed, 53 insertions(+), 10 deletions(-)
include/linux/pid.h +1 -1
index 2a83e434db9d..052000db0ced 100644
@@ -116,7 +116,7 @@ extern struct pid *find_vpid(int nr);
 extern struct pid *find_get_pid(int nr);
 extern struct pid *find_ge_pid(int nr, struct pid_namespace *);
 
-extern struct pid *alloc_pid(struct pid_namespace *ns);
+extern struct pid *alloc_pid(struct pid_namespace *ns, pid_t set_tid);
 extern void free_pid(struct pid *pid);
 extern void disable_pid_allocation(struct pid_namespace *ns);
 
include/linux/sched/task.h +1 -0
index 0497091e40c1..4f2a80564332 100644
@@ -26,6 +26,7 @@ struct kernel_clone_args {
 	unsigned long stack;
 	unsigned long stack_size;
 	unsigned long tls;
+	pid_t set_tid;
 };
 
 /*
include/uapi/linux/sched.h +1 -0
index b3105ac1381a..e1ce103a2c47 100644
@@ -45,6 +45,7 @@ struct clone_args {
 	__aligned_u64 stack;
 	__aligned_u64 stack_size;
 	__aligned_u64 tls;
+	__aligned_u64 set_tid;
 };
 
 /*
kernel/fork.c +14 -1
index 2852d0e76ea3..2a03f0e201e9 100644
@@ -117,6 +117,13 @@
  */
 #define MAX_THREADS FUTEX_TID_MASK
 
+/*
+ * Different sizes of struct clone_args
+ */
+#define CLONE3_ARGS_SIZE_V0 64
+/* V1 includes set_tid */
+#define CLONE3_ARGS_SIZE_V1 72
+
 /*
  * Protected counters by write_lock_irq(&tasklist_lock)
  */
@@ -2031,7 +2038,13 @@ static __latent_entropy struct task_struct *copy_process(
 	stackleak_task_init(p);
 
 	if (pid != &init_struct_pid) {
-		pid = alloc_pid(p->nsproxy->pid_ns_for_children);
+		if (args->set_tid && !ns_capable(
+				p->nsproxy->pid_ns_for_children->user_ns,
+				CAP_SYS_ADMIN)) {
+			retval = -EPERM;
+			goto bad_fork_cleanup_thread;
+		}
+		pid = alloc_pid(p->nsproxy->pid_ns_for_children, args->set_tid);
 		if (IS_ERR(pid)) {
 			retval = PTR_ERR(pid);
 			goto bad_fork_cleanup_thread;
@@ -2535,9 +2548,14 @@ noinline static int copy_clone_args_from_user(struct kernel_clone_args *kargs,
 	if (unlikely(size > PAGE_SIZE))
 		return -E2BIG;
 
Quoted 6 lines

kernel/fork.c noinline static int copy_clone_args_from_user(struct kernel_clone_args *kargs, · line 2551

Quoted 2 lines
kernel/fork.c +1 -1
-	if (unlikely(size < sizeof(struct clone_args)))
+	/* The struct needs to be at least the size of the original struct. */

I don't think you need that comment. I think the macro is pretty self-explanatory. If you want it to be even clearer you could even make it CLONE3_ARGS_SIZE_MIN but V0 is good enough. :)

Will remove the comment.

Quoted 23 lines
Quoted 11 lines
kernel/fork.c +5 -0
+	if (unlikely(size < CLONE3_ARGS_SIZE_V0))
 		return -EINVAL;
 
+	if (size < sizeof(struct clone_args))
+		memset((void *)&args + size, 0,
+				sizeof(struct clone_args) - size);
+
 	if (unlikely(!access_ok(uargs, size)))
 		return -EFAULT;
 

kernel/fork.c noinline static int copy_clone_args_from_user(struct kernel_clone_args *kargs, · line 2595

Quoted 5 lines
kernel/fork.c +2 -0
@@ -2573,6 +2591,9 @@ noinline static int copy_clone_args_from_user(struct kernel_clone_args *kargs,
 		.tls		= args.tls,
 	};
 
+	if (size >= CLONE3_ARGS_SIZE_V1)
+		kargs->set_tid = args.set_tid;

Hm, the if-condition is not needed though, right? At this point we will have already copied from struct clone_args __user *uargs into struct clone_args args. If we hit that codepath that means the kernel definitely has a field for set_tid in its struct clone_args. :) So this could probably just be:

Quoted 5 lines
   		.tls		= args.tls,
		.set_tid	= args.set_tid,
	}

	?

Right.

Quoted 21 lines
Quoted 21 lines
kernel/fork.c +1 -0
+
 	return 0;
 }
 
kernel/pid.c +1 -1
index 0a9f2e437217..9ce89c35c5be 100644
@@ -157,7 +157,7 @@ void free_pid(struct pid *pid)
 	call_rcu(&pid->rcu, delayed_put_pid);
 }
 
-struct pid *alloc_pid(struct pid_namespace *ns)
+struct pid *alloc_pid(struct pid_namespace *ns, int set_tid)
 {
 	struct pid *pid;
 	enum pid_type type;
@@ -186,12 +186,32 @@ struct pid *alloc_pid(struct pid_namespace *ns)
 		if (idr_get_cursor(&tmp->idr) > RESERVED_PIDS)
 			pid_min = RESERVED_PIDS;
 
Quoted 16 lines

kernel/pid.c struct pid *alloc_pid(struct pid_namespace *ns) · line 195

Quoted 13 lines
kernel/pid.c +7 -6
-		/*
-		 * Store a null pointer so find_pid_ns does not find
-		 * a partially initialized PID (see below).
-		 */
-		nr = idr_alloc_cyclic(&tmp->idr, NULL, pid_min,
-				      pid_max, GFP_ATOMIC);
+		if (set_tid) {
+			/*
+			 * Also fail if a PID != 1 is requested
+			 * and no PID 1 exists.
+			 */
+			nr = -EINVAL;
+			if (set_tid < pid_max && set_tid > 0 &&

Hm, you're already in the if-branch hat verified if (set_tid) so the set_tid > 0 conjunct seems redundant. :)

Yes, but I dropped all checks to see if set_tid is negative as suggested by Oleg and moved it here.

Quoted 6 lines

kernel/pid.c struct pid *alloc_pid(struct pid_namespace *ns) · line 198

Quoted 3 lines
kernel/pid.c +3 -0
+			    (set_tid == 1 || !idr_is_empty(&tmp->idr)))
+				nr = idr_alloc(&tmp->idr, NULL, set_tid,
+					       set_tid + 1, GFP_ATOMIC);

I'm confused, shouldn't this be

Quoted 7 lines
Quoted 1 line
	if (set_tid < pid_max || (set_tid == 1 && !idr_is_emtpy(&tmp->idf)))

Now I am also confused ;). This does not work. This will always return true if set_tid is less than pid_max. So pid_max needs to be something like 1 for the check after || to make sense, right? But you really got me confused here right now. Right now I still think what I did is correct.

I missed the part where you reset set_tid to 0 below and mis-parsed the rightmost conjunct in the if statement.

One thing I dislike is that you do

if (set_tid) {
	if ([...] && set_tid > 0 && [...])
}

For consistency this should rather be:

if (set_tid) {
	if ([...] && set_tid && [...])
}

Overall I'd rather write the second if as:

if (set_tid && set_tid < pid_max && (set_tid == 1 || !idr_is_empty(&tmp->idr)))

to place the most basic condition (i.e. set_tid is non-zero) leftmost.

The right-most conjunct (set_tid == 1 || !idr_is_empty(&tmp->idr)) is non-trivial to parse. Someone not familiar with that code would need time to figure out why and how exactly 1 is special-cased.
I think overall this is begging for alloc_pid() to be documented now if you don't mind. :)

Christian

#5 lore raw
Quoted 213 lines · Christian Brauner
Quoted 204 lines · Adrian Reber
Quoted 116 lines · Christian Brauner
Quoted 115 lines · Adrian Reber

The main motivation to add set_tid to clone3() is CRIU.

To restore a process with the same PID/TID CRIU currently uses /proc/sys/kernel/ns_last_pid. It writes the desired (PID - 1) to ns_last_pid and then (quickly) does a clone(). This works most of the time, but it is racy. It is also slow as it requires multiple syscalls.

Extending clone3() to support set_tid makes it possible restore a process using CRIU without accessing /proc/sys/kernel/ns_last_pid and race free (as long as the desired PID/TID is available).

This clone3() extension places the same restrictions (CAP_SYS_ADMIN) on clone3() with set_tid as they are currently in place for ns_last_pid.

---
v2:
- Removed (size < sizeof(struct clone_args)) as discussed with Christian and Dmitry
- Added comment to ((set_tid != 1) && idr_get_cursor() <= 1) (Oleg)
- Use idr_alloc() instead of idr_alloc_cyclic() (Oleg)

v3:
- Return EEXIST if PID is already in use (Christian)
- Drop CLONE_SET_TID (Christian and Oleg)
- Use idr_is_empty() instead of idr_get_cursor() (Oleg)
- Handle different struct clone_args sizes (Dmitry)

v4:
- Rework struct size check with defines (Christian)
- Reduce number of set_tid checks (Oleg)
- Less parentheses and more robust code (Oleg)
- Do ns_capable() on correct user_ns (Oleg, Christian)

---
 include/linux/pid.h        |  2 +-
 include/linux/sched/task.h |  1 +
 include/uapi/linux/sched.h |  1 +
 kernel/fork.c              | 25 +++++++++++++++++++++++--
 kernel/pid.c               | 34 +++++++++++++++++++++++++++-------
 5 files changed, 53 insertions(+), 10 deletions(-)
include/linux/pid.h +1 -1
index 2a83e434db9d..052000db0ced 100644
@@ -116,7 +116,7 @@ extern struct pid *find_vpid(int nr);
 extern struct pid *find_get_pid(int nr);
 extern struct pid *find_ge_pid(int nr, struct pid_namespace *);
 
-extern struct pid *alloc_pid(struct pid_namespace *ns);
+extern struct pid *alloc_pid(struct pid_namespace *ns, pid_t set_tid);
 extern void free_pid(struct pid *pid);
 extern void disable_pid_allocation(struct pid_namespace *ns);
 
include/linux/sched/task.h +1 -0
index 0497091e40c1..4f2a80564332 100644
@@ -26,6 +26,7 @@ struct kernel_clone_args {
 	unsigned long stack;
 	unsigned long stack_size;
 	unsigned long tls;
+	pid_t set_tid;
 };
 
 /*
include/uapi/linux/sched.h +1 -0
index b3105ac1381a..e1ce103a2c47 100644
@@ -45,6 +45,7 @@ struct clone_args {
 	__aligned_u64 stack;
 	__aligned_u64 stack_size;
 	__aligned_u64 tls;
+	__aligned_u64 set_tid;
 };
 
 /*
kernel/fork.c +14 -1
index 2852d0e76ea3..2a03f0e201e9 100644
@@ -117,6 +117,13 @@
  */
 #define MAX_THREADS FUTEX_TID_MASK
 
+/*
+ * Different sizes of struct clone_args
+ */
+#define CLONE3_ARGS_SIZE_V0 64
+/* V1 includes set_tid */
+#define CLONE3_ARGS_SIZE_V1 72
+
 /*
  * Protected counters by write_lock_irq(&tasklist_lock)
  */
@@ -2031,7 +2038,13 @@ static __latent_entropy struct task_struct *copy_process(
 	stackleak_task_init(p);
 
 	if (pid != &init_struct_pid) {
-		pid = alloc_pid(p->nsproxy->pid_ns_for_children);
+		if (args->set_tid && !ns_capable(
+				p->nsproxy->pid_ns_for_children->user_ns,
+				CAP_SYS_ADMIN)) {
+			retval = -EPERM;
+			goto bad_fork_cleanup_thread;
+		}
+		pid = alloc_pid(p->nsproxy->pid_ns_for_children, args->set_tid);
 		if (IS_ERR(pid)) {
 			retval = PTR_ERR(pid);
 			goto bad_fork_cleanup_thread;
@@ -2535,9 +2548,14 @@ noinline static int copy_clone_args_from_user(struct kernel_clone_args *kargs,
 	if (unlikely(size > PAGE_SIZE))
 		return -E2BIG;
 
Quoted 6 lines

kernel/fork.c noinline static int copy_clone_args_from_user(struct kernel_clone_args *kargs, · line 2551

Quoted 2 lines
kernel/fork.c +1 -1
-	if (unlikely(size < sizeof(struct clone_args)))
+	/* The struct needs to be at least the size of the original struct. */

I don't think you need that comment. I think the macro is pretty self-explanatory. If you want it to be even clearer you could even make it CLONE3_ARGS_SIZE_MIN but V0 is good enough. :)

Will remove the comment.

Quoted 23 lines
Quoted 11 lines
kernel/fork.c +5 -0
+	if (unlikely(size < CLONE3_ARGS_SIZE_V0))
 		return -EINVAL;
 
+	if (size < sizeof(struct clone_args))
+		memset((void *)&args + size, 0,
+				sizeof(struct clone_args) - size);
+
 	if (unlikely(!access_ok(uargs, size)))
 		return -EFAULT;
 

kernel/fork.c noinline static int copy_clone_args_from_user(struct kernel_clone_args *kargs, · line 2595

Quoted 5 lines
kernel/fork.c +2 -0
@@ -2573,6 +2591,9 @@ noinline static int copy_clone_args_from_user(struct kernel_clone_args *kargs,
 		.tls		= args.tls,
 	};
 
+	if (size >= CLONE3_ARGS_SIZE_V1)
+		kargs->set_tid = args.set_tid;

Hm, the if-condition is not needed though, right? At this point we will have already copied from struct clone_args __user *uargs into struct clone_args args. If we hit that codepath that means the kernel definitely has a field for set_tid in its struct clone_args. :) So this could probably just be:

Quoted 5 lines
   		.tls		= args.tls,
		.set_tid	= args.set_tid,
	}

	?

Right.

Quoted 21 lines
Quoted 21 lines
kernel/fork.c +1 -0
+
 	return 0;
 }
 
kernel/pid.c +1 -1
index 0a9f2e437217..9ce89c35c5be 100644
@@ -157,7 +157,7 @@ void free_pid(struct pid *pid)
 	call_rcu(&pid->rcu, delayed_put_pid);
 }
 
-struct pid *alloc_pid(struct pid_namespace *ns)
+struct pid *alloc_pid(struct pid_namespace *ns, int set_tid)
 {
 	struct pid *pid;
 	enum pid_type type;
@@ -186,12 +186,32 @@ struct pid *alloc_pid(struct pid_namespace *ns)
 		if (idr_get_cursor(&tmp->idr) > RESERVED_PIDS)
 			pid_min = RESERVED_PIDS;
 
Quoted 16 lines

kernel/pid.c struct pid *alloc_pid(struct pid_namespace *ns) · line 195

Quoted 13 lines
kernel/pid.c +7 -6
-		/*
-		 * Store a null pointer so find_pid_ns does not find
-		 * a partially initialized PID (see below).
-		 */
-		nr = idr_alloc_cyclic(&tmp->idr, NULL, pid_min,
-				      pid_max, GFP_ATOMIC);
+		if (set_tid) {
+			/*
+			 * Also fail if a PID != 1 is requested
+			 * and no PID 1 exists.
+			 */
+			nr = -EINVAL;
+			if (set_tid < pid_max && set_tid > 0 &&

Hm, you're already in the if-branch hat verified if (set_tid) so the set_tid > 0 conjunct seems redundant. :)

Yes, but I dropped all checks to see if set_tid is negative as suggested by Oleg and moved it here.

Quoted 6 lines

kernel/pid.c struct pid *alloc_pid(struct pid_namespace *ns) · line 198

Quoted 3 lines
kernel/pid.c +3 -0
+			    (set_tid == 1 || !idr_is_empty(&tmp->idr)))
+				nr = idr_alloc(&tmp->idr, NULL, set_tid,
+					       set_tid + 1, GFP_ATOMIC);

I'm confused, shouldn't this be

Quoted 7 lines
Quoted 1 line
	if (set_tid < pid_max || (set_tid == 1 && !idr_is_emtpy(&tmp->idf)))

Now I am also confused ;). This does not work. This will always return true if set_tid is less than pid_max. So pid_max needs to be something like 1 for the check after || to make sense, right? But you really got me confused here right now. Right now I still think what I did is correct.

Quoted 8 lines

I missed the part where you reset set_tid to 0 below and mis-parsed the rightmost conjunct in the if statement.

One thing I dislike is that you do

if (set_tid) {
	if ([...] && set_tid > 0 && [...])
}

Thinking about this a bit more you probably did the explicit set_tid > 0 to catch the case where it is negative? But also, I don't understand why we do any work at all before having verified that set_tid is sensible, i.e. why do we call kmem_cache_calloc(), do idr_preload(), and take a spinlock, before even verifying that our parameters are sane? If there's no specific reason for this I suggest to patch alloc_pid() like this:

-struct pid *alloc_pid(struct pid_namespace *ns) +struct pid *alloc_pid(struct pid_namespace *ns, int set_tid)

 {
        struct pid *pid;
        enum pid_type type;
@@ -166,6 +166,9 @@ struct pid *alloc_pid(struct pid_namespace *ns)
        struct upid *upid;
        int retval = -ENOMEM;

+ if (set_tid < 0 || set_tid >= pid_max)
+ return ERR_PTR(-EINVAL);
+

        pid = kmem_cache_alloc(ns->pid_cachep, GFP_KERNEL);
        if (!pid)
                return ERR_PTR(retval);
@@ -186,12 +189,31 @@ struct pid *alloc_pid(struct pid_namespace *ns)
                if (idr_get_cursor(&tmp->idr) > RESERVED_PIDS)
                        pid_min = RESERVED_PIDS;

- /*

-                * Store a null pointer so find_pid_ns does not find
-                * a partially initialized PID (see below).
-                */

- nr = idr_alloc_cyclic(&tmp->idr, NULL, pid_min,
- pid_max, GFP_ATOMIC);
+ if (set_tid) {
+ /*

+                        * Also fail if a PID != 1 is requested
+                        * and no PID 1 exists.
+                        */

+ nr = -EINVAL;
+ if (set_tid == 1 || !idr_is_empty(&tmp->idr))
+ nr = idr_alloc(&tmp->idr, NULL, set_tid,
+ set_tid + 1, GFP_ATOMIC);
+ /*

+                        * If ENOSPC is returned it means that the PID is
+                        * alreay in use. Return EEXIST in that case.
+                        */

+ if (nr == -ENOSPC)
+ nr = -EEXIST;
+ /* Only use set_tid for one PID namespace. */
+ set_tid = 0;
+ } else {
+ /*

+                        * Store a null pointer so find_pid_ns does not find
+                        * a partially initialized PID (see below).
+                        */
+                       nr = idr_alloc_cyclic(&tmp->idr, NULL, pid_min,
+                                             pid_max, GFP_ATOMIC);
+               }
                spin_unlock_irq(&pidmap_lock);
                idr_preload_end()

This makes things a lot more clearer in my opinion. First, verify that the pre-conditions are met. Second, verify that the conditions are met which depend on the state of the pid namespace, i.e. there's either already a pid 1 or pid 1 is requested.
We should also do this since alloc_pid() is exported in a header file and so we can't and shouldn't rely on the fact that all callers will pass in something sensible for set_tid.

Quoted 20 lines

For consistency this should rather be:

if (set_tid) {
	if ([...] && set_tid && [...])
}

Overall I'd rather write the second if as:

if (set_tid && set_tid < pid_max && (set_tid == 1 || !idr_is_empty(&tmp->idr)))

to place the most basic condition (i.e. set_tid is non-zero) leftmost.

The right-most conjunct (set_tid == 1 || !idr_is_empty(&tmp->idr)) is non-trivial to parse. Someone not familiar with that code would need time to figure out why and how exactly 1 is special-cased.
I think overall this is begging for alloc_pid() to be documented now if you don't mind. :)

Christian

#6 lore raw
Quoted 283 lines · Christian Brauner
Quoted 213 lines · Christian Brauner
Quoted 204 lines · Adrian Reber
Quoted 116 lines · Christian Brauner
Quoted 115 lines · Adrian Reber

The main motivation to add set_tid to clone3() is CRIU.

To restore a process with the same PID/TID CRIU currently uses /proc/sys/kernel/ns_last_pid. It writes the desired (PID - 1) to ns_last_pid and then (quickly) does a clone(). This works most of the time, but it is racy. It is also slow as it requires multiple syscalls.

Extending clone3() to support set_tid makes it possible restore a process using CRIU without accessing /proc/sys/kernel/ns_last_pid and race free (as long as the desired PID/TID is available).

This clone3() extension places the same restrictions (CAP_SYS_ADMIN) on clone3() with set_tid as they are currently in place for ns_last_pid.

---
v2:
- Removed (size < sizeof(struct clone_args)) as discussed with Christian and Dmitry
- Added comment to ((set_tid != 1) && idr_get_cursor() <= 1) (Oleg)
- Use idr_alloc() instead of idr_alloc_cyclic() (Oleg)

v3:
- Return EEXIST if PID is already in use (Christian)
- Drop CLONE_SET_TID (Christian and Oleg)
- Use idr_is_empty() instead of idr_get_cursor() (Oleg)
- Handle different struct clone_args sizes (Dmitry)

v4:
- Rework struct size check with defines (Christian)
- Reduce number of set_tid checks (Oleg)
- Less parentheses and more robust code (Oleg)
- Do ns_capable() on correct user_ns (Oleg, Christian)

---
 include/linux/pid.h        |  2 +-
 include/linux/sched/task.h |  1 +
 include/uapi/linux/sched.h |  1 +
 kernel/fork.c              | 25 +++++++++++++++++++++++--
 kernel/pid.c               | 34 +++++++++++++++++++++++++++-------
 5 files changed, 53 insertions(+), 10 deletions(-)
include/linux/pid.h +1 -1
index 2a83e434db9d..052000db0ced 100644
@@ -116,7 +116,7 @@ extern struct pid *find_vpid(int nr);
 extern struct pid *find_get_pid(int nr);
 extern struct pid *find_ge_pid(int nr, struct pid_namespace *);
 
-extern struct pid *alloc_pid(struct pid_namespace *ns);
+extern struct pid *alloc_pid(struct pid_namespace *ns, pid_t set_tid);
 extern void free_pid(struct pid *pid);
 extern void disable_pid_allocation(struct pid_namespace *ns);
 
include/linux/sched/task.h +1 -0
index 0497091e40c1..4f2a80564332 100644
@@ -26,6 +26,7 @@ struct kernel_clone_args {
 	unsigned long stack;
 	unsigned long stack_size;
 	unsigned long tls;
+	pid_t set_tid;
 };
 
 /*
include/uapi/linux/sched.h +1 -0
index b3105ac1381a..e1ce103a2c47 100644
@@ -45,6 +45,7 @@ struct clone_args {
 	__aligned_u64 stack;
 	__aligned_u64 stack_size;
 	__aligned_u64 tls;
+	__aligned_u64 set_tid;
 };
 
 /*
kernel/fork.c +14 -1
index 2852d0e76ea3..2a03f0e201e9 100644
@@ -117,6 +117,13 @@
  */
 #define MAX_THREADS FUTEX_TID_MASK
 
+/*
+ * Different sizes of struct clone_args
+ */
+#define CLONE3_ARGS_SIZE_V0 64
+/* V1 includes set_tid */
+#define CLONE3_ARGS_SIZE_V1 72
+
 /*
  * Protected counters by write_lock_irq(&tasklist_lock)
  */
@@ -2031,7 +2038,13 @@ static __latent_entropy struct task_struct *copy_process(
 	stackleak_task_init(p);
 
 	if (pid != &init_struct_pid) {
-		pid = alloc_pid(p->nsproxy->pid_ns_for_children);
+		if (args->set_tid && !ns_capable(
+				p->nsproxy->pid_ns_for_children->user_ns,
+				CAP_SYS_ADMIN)) {
+			retval = -EPERM;
+			goto bad_fork_cleanup_thread;
+		}
+		pid = alloc_pid(p->nsproxy->pid_ns_for_children, args->set_tid);
 		if (IS_ERR(pid)) {
 			retval = PTR_ERR(pid);
 			goto bad_fork_cleanup_thread;
@@ -2535,9 +2548,14 @@ noinline static int copy_clone_args_from_user(struct kernel_clone_args *kargs,
 	if (unlikely(size > PAGE_SIZE))
 		return -E2BIG;
 
Quoted 6 lines

kernel/fork.c noinline static int copy_clone_args_from_user(struct kernel_clone_args *kargs, · line 2551

Quoted 2 lines
kernel/fork.c +1 -1
-	if (unlikely(size < sizeof(struct clone_args)))
+	/* The struct needs to be at least the size of the original struct. */

I don't think you need that comment. I think the macro is pretty self-explanatory. If you want it to be even clearer you could even make it CLONE3_ARGS_SIZE_MIN but V0 is good enough. :)

Will remove the comment.

Quoted 23 lines
Quoted 11 lines
kernel/fork.c +5 -0
+	if (unlikely(size < CLONE3_ARGS_SIZE_V0))
 		return -EINVAL;
 
+	if (size < sizeof(struct clone_args))
+		memset((void *)&args + size, 0,
+				sizeof(struct clone_args) - size);
+
 	if (unlikely(!access_ok(uargs, size)))
 		return -EFAULT;
 

kernel/fork.c noinline static int copy_clone_args_from_user(struct kernel_clone_args *kargs, · line 2595

Quoted 5 lines
kernel/fork.c +2 -0
@@ -2573,6 +2591,9 @@ noinline static int copy_clone_args_from_user(struct kernel_clone_args *kargs,
 		.tls		= args.tls,
 	};
 
+	if (size >= CLONE3_ARGS_SIZE_V1)
+		kargs->set_tid = args.set_tid;

Hm, the if-condition is not needed though, right? At this point we will have already copied from struct clone_args __user *uargs into struct clone_args args. If we hit that codepath that means the kernel definitely has a field for set_tid in its struct clone_args. :) So this could probably just be:

Quoted 5 lines
   		.tls		= args.tls,
		.set_tid	= args.set_tid,
	}

	?

Right.

Quoted 21 lines
Quoted 21 lines
kernel/fork.c +1 -0
+
 	return 0;
 }
 
kernel/pid.c +1 -1
index 0a9f2e437217..9ce89c35c5be 100644
@@ -157,7 +157,7 @@ void free_pid(struct pid *pid)
 	call_rcu(&pid->rcu, delayed_put_pid);
 }
 
-struct pid *alloc_pid(struct pid_namespace *ns)
+struct pid *alloc_pid(struct pid_namespace *ns, int set_tid)
 {
 	struct pid *pid;
 	enum pid_type type;
@@ -186,12 +186,32 @@ struct pid *alloc_pid(struct pid_namespace *ns)
 		if (idr_get_cursor(&tmp->idr) > RESERVED_PIDS)
 			pid_min = RESERVED_PIDS;
 
Quoted 16 lines

kernel/pid.c struct pid *alloc_pid(struct pid_namespace *ns) · line 195

Quoted 13 lines
kernel/pid.c +7 -6
-		/*
-		 * Store a null pointer so find_pid_ns does not find
-		 * a partially initialized PID (see below).
-		 */
-		nr = idr_alloc_cyclic(&tmp->idr, NULL, pid_min,
-				      pid_max, GFP_ATOMIC);
+		if (set_tid) {
+			/*
+			 * Also fail if a PID != 1 is requested
+			 * and no PID 1 exists.
+			 */
+			nr = -EINVAL;
+			if (set_tid < pid_max && set_tid > 0 &&

Hm, you're already in the if-branch hat verified if (set_tid) so the set_tid > 0 conjunct seems redundant. :)

Yes, but I dropped all checks to see if set_tid is negative as suggested by Oleg and moved it here.

Quoted 6 lines

kernel/pid.c struct pid *alloc_pid(struct pid_namespace *ns) · line 198

Quoted 3 lines
kernel/pid.c +3 -0
+			    (set_tid == 1 || !idr_is_empty(&tmp->idr)))
+				nr = idr_alloc(&tmp->idr, NULL, set_tid,
+					       set_tid + 1, GFP_ATOMIC);

I'm confused, shouldn't this be

Quoted 7 lines
Quoted 1 line
	if (set_tid < pid_max || (set_tid == 1 && !idr_is_emtpy(&tmp->idf)))

Now I am also confused ;). This does not work. This will always return true if set_tid is less than pid_max. So pid_max needs to be something like 1 for the check after || to make sense, right? But you really got me confused here right now. Right now I still think what I did is correct.

Quoted 8 lines

I missed the part where you reset set_tid to 0 below and mis-parsed the rightmost conjunct in the if statement.

One thing I dislike is that you do

if (set_tid) {
	if ([...] && set_tid > 0 && [...])
}

Thinking about this a bit more you probably did the explicit set_tid > 0 to catch the case where it is negative? But also, I don't understand why we do any work at all before having verified that set_tid is sensible, i.e. why do we call kmem_cache_calloc(), do idr_preload(), and take a spinlock, before even verifying that our parameters are sane? If there's no specific reason for this I suggest to patch alloc_pid() like this:

kernel/pid.c +29 -7
-struct pid *alloc_pid(struct pid_namespace *ns)
+struct pid *alloc_pid(struct pid_namespace *ns, int set_tid)
 {
        struct pid *pid;
        enum pid_type type;
@@ -166,6 +166,9 @@ struct pid *alloc_pid(struct pid_namespace *ns)
        struct upid *upid;
        int retval = -ENOMEM;
 
+       if (set_tid < 0 || set_tid >= pid_max)
+               return ERR_PTR(-EINVAL);
+
        pid = kmem_cache_alloc(ns->pid_cachep, GFP_KERNEL);
        if (!pid)
                return ERR_PTR(retval);
@@ -186,12 +189,31 @@ struct pid *alloc_pid(struct pid_namespace *ns)
                if (idr_get_cursor(&tmp->idr) > RESERVED_PIDS)
                        pid_min = RESERVED_PIDS;
 
-               /*
-                * Store a null pointer so find_pid_ns does not find
-                * a partially initialized PID (see below).
-                */
-               nr = idr_alloc_cyclic(&tmp->idr, NULL, pid_min,
-                                     pid_max, GFP_ATOMIC);
+               if (set_tid) {
+                       /*
+                        * Also fail if a PID != 1 is requested
+                        * and no PID 1 exists.
+                        */
+                       nr = -EINVAL;
+                       if (set_tid == 1 || !idr_is_empty(&tmp->idr))
+                               nr = idr_alloc(&tmp->idr, NULL, set_tid,
+                                              set_tid + 1, GFP_ATOMIC);
+                       /*
+                        * If ENOSPC is returned it means that the PID is
+                        * alreay in use. Return EEXIST in that case.
+                        */
+                       if (nr == -ENOSPC)
+                               nr = -EEXIST;
+                       /* Only use set_tid for one PID namespace. */
+                       set_tid = 0;
+               } else {
+                       /*
+                        * Store a null pointer so find_pid_ns does not find
+                        * a partially initialized PID (see below).
+                        */
+                       nr = idr_alloc_cyclic(&tmp->idr, NULL, pid_min,
+                                             pid_max, GFP_ATOMIC);
+               }
                spin_unlock_irq(&pidmap_lock);
                idr_preload_end()
 
Quoted 7 lines

This makes things a lot more clearer in my opinion. First, verify that the pre-conditions are met. Second, verify that the conditions are met which depend on the state of the pid namespace, i.e. there's either already a pid 1 or pid 1 is requested.
We should also do this since alloc_pid() is exported in a header file and so we can't and shouldn't rely on the fact that all callers will pass in something sensible for set_tid.

I like that. I will update the patch and send it out to match this suggestion.

Adrian