Use clone3() with set_tid to create processes

With the in Linux Kernel 5.4 introduced clone3() with set_tid it is no
longer necessary to write to to /proc/../ns_last_pid to influence the
next PID number. clone3() can directly select a PID for the newly
created process/thread.

After checking for the availability of clone3() with set_tid and adding
the assembler wrapper for clone3() in previous patches, this extends
criu/pie/restorer.c and criu/clone-noasan.c to use the newly added
assembler clone3() wrapper to create processes with a certain PID.

This is a RFC and WIP, but I wanted to share it and run it through CI
for feedback. As the CI will probably not use a 5.4 based kernel it
should just keep on working as before.

Signed-off-by: Adrian Reber <areber@redhat.com>

criu/clone-noasan.c

@@ -1,4 +1,10 @@
+#include <stdlib.h>
 #include <sched.h>
+#include <unistd.h>
+
+#include <compel/plugins/std/syscall-codes.h>
+
+#include "sched.h"
 #include "common/compiler.h"
 #include "log.h"
 #include "common/bug.h"
@@ -31,6 +37,7 @@
 int clone_noasan(int (*fn)(void *), int flags, void *arg)
 {
 	void *stack_ptr = (void *)round_down((unsigned long)&stack_ptr - 1024, 16);
+
 	BUG_ON((flags & CLONE_VM) && !(flags & CLONE_VFORK));
 	/*
 	 * Reserve some bytes for clone() internal needs
@@ -38,3 +45,28 @@ int clone_noasan(int (*fn)(void *), int flags, void *arg)
 	 */
 	return clone(fn, stack_ptr, flags, arg);
 }
+
+int clone3_with_pid_noasan(int (*fn)(void *), void *arg, int flags,
+			   int exit_signal, pid_t pid)
+{
+	struct _clone_args c_args = {};
+
+	BUG_ON(flags & CLONE_VM);
+
+	/*
+	 * Make sure no child signals are requested. clone3() uses
+	 * exit_signal for that.
+	 */
+	BUG_ON(flags & 0xff);
+
+	pr_debug("Creating process using clone3()\n");
+
+	c_args.exit_signal = exit_signal;
+	c_args.flags = flags;
+	c_args.set_tid = ptr_to_u64(&pid);
+	c_args.set_tid_size = 1;
+	pid = syscall(__NR_clone3, &c_args, sizeof(c_args));
+	if (pid == 0)
+		exit(fn(arg));
+	return pid;
+}

criu/cr-restore.c

@@ -1374,32 +1374,43 @@ static inline int fork_with_pid(struct pstree_item *item)
 	if (!(ca.clone_flags & CLONE_NEWPID)) {
 		char buf[32];
 		int len;
-		int fd;
+		int fd = -1;
 
+		if (!kdat.has_clone3_set_tid) {
 			fd = open_proc_rw(PROC_GEN, LAST_PID_PATH);
 			if (fd < 0)
 				goto err;
+		}
 
 		lock_last_pid();
 
+		if (!kdat.has_clone3_set_tid) {
 			len = snprintf(buf, sizeof(buf), "%d", pid - 1);
 			if (write(fd, buf, len) != len) {
-			pr_perror("%d: Write %s to %s", pid, buf, LAST_PID_PATH);
+				pr_perror("%d: Write %s to %s", pid, buf,
+					LAST_PID_PATH);
 				close(fd);
 				goto err_unlock;
 			}
 			close(fd);
+		}
 	} else {
 		BUG_ON(pid != INIT_PID);
 	}
 
+	if (kdat.has_clone3_set_tid) {
+		ret = clone3_with_pid_noasan(restore_task_with_children,
+				&ca, (ca.clone_flags &
+					~(CLONE_NEWNET | CLONE_NEWCGROUP)),
+				SIGCHLD, pid);
+	} else {
 		/*
 		 * Some kernel modules, such as network packet generator
-	 * run kernel thread upon net-namespace creattion taking
+		 * run kernel thread upon net-namespace creation taking
-	 * the @pid we've been requeting via LAST_PID_PATH interface
+		 * the @pid we've been requesting via LAST_PID_PATH interface
 		 * so that we can't restore a take with pid needed.
 		 *
-	 * Here is an idea -- unhare net namespace in callee instead.
+		 * Here is an idea -- unshare net namespace in callee instead.
 		 */
 		/*
 		 * The cgroup namespace is also unshared explicitly in the
@@ -1407,7 +1418,11 @@ static inline int fork_with_pid(struct pstree_item *item)
 		 */
 		close_pid_proc();
 		ret = clone_noasan(restore_task_with_children,
-			(ca.clone_flags & ~(CLONE_NEWNET | CLONE_NEWCGROUP)) | SIGCHLD, &ca);
+				(ca.clone_flags &
+				 ~(CLONE_NEWNET | CLONE_NEWCGROUP)) | SIGCHLD,
+				&ca);
+	}
+
 	if (ret < 0) {
 		pr_perror("Can't fork for %d", pid);
 		goto err_unlock;
@@ -3588,6 +3603,7 @@ static int sigreturn_restore(pid_t pid, struct task_restore_args *task_args, uns
 	task_args->vdso_maps_rt = vdso_maps_rt;
 	task_args->vdso_rt_size = vdso_rt_size;
 	task_args->can_map_vdso = kdat.can_map_vdso;
+	task_args->has_clone3_set_tid = kdat.has_clone3_set_tid;
 
 	new_sp = restorer_stack(task_args->t->mz);
 

criu/include/clone-noasan.h

@@ -2,5 +2,7 @@
 #define __CR_CLONE_NOASAN_H__
 
 int clone_noasan(int (*fn)(void *), int flags, void *arg);
+int clone3_with_pid_noasan(int (*fn)(void *), void *arg, int flags,
+			   int exit_signal, pid_t pid);
 
 #endif /* __CR_CLONE_NOASAN_H__ */

criu/include/restorer.h

@@ -221,6 +221,7 @@ struct task_restore_args {
 #endif
 	int				lsm_type;
 	int				child_subreaper;
+	bool				has_clone3_set_tid;
 } __aligned(64);
 
 /*

criu/include/rst_info.h

@@ -4,6 +4,7 @@
 #include "common/lock.h"
 #include "common/list.h"
 #include "vma.h"
+#include "kerndat.h"
 
 struct task_entries {
 	int nr_threads, nr_tasks, nr_helpers;

criu/pie/restorer.c

@@ -35,6 +35,7 @@
 #include "sk-inet.h"
 #include "vma.h"
 #include "uffd.h"
+#include "sched.h"
 
 #include "common/lock.h"
 #include "common/page.h"
@@ -1771,9 +1772,11 @@ long __export_restore_task(struct task_restore_args *args)
 		long clone_flags = CLONE_VM | CLONE_FILES | CLONE_SIGHAND	|
 				   CLONE_THREAD | CLONE_SYSVSEM | CLONE_FS;
 		long last_pid_len;
+		pid_t thread_pid;
 		long parent_tid;
 		int i, fd = -1;
 
+		if (!args->has_clone3_set_tid) {
 			/* One level pid ns hierarhy */
 			fd = sys_openat(args->proc_fd, LAST_PID_PATH, O_RDWR, 0);
 			if (fd < 0) {
@@ -1781,6 +1784,7 @@ long __export_restore_task(struct task_restore_args *args)
 				goto core_restore_end;
 			}
 
+		}
 		mutex_lock(&task_entries_local->last_pid_mutex);
 
 		for (i = 0; i < args->nr_threads; i++) {
@@ -1791,6 +1795,20 @@ long __export_restore_task(struct task_restore_args *args)
 				continue;
 
 			new_sp = restorer_stack(thread_args[i].mz);
+			if (args->has_clone3_set_tid) {
+				struct _clone_args c_args = {};
+				thread_pid = thread_args[i].pid;
+				c_args.set_tid = ptr_to_u64(&thread_pid);
+				c_args.flags = clone_flags;
+				c_args.set_tid_size = 1;
+				/* The kernel does stack + stack_size. */
+				c_args.stack = new_sp - RESTORE_STACK_SIZE;
+				c_args.stack_size = RESTORE_STACK_SIZE;
+				c_args.child_tid = ptr_to_u64(&thread_args[i].pid);
+				c_args.parent_tid = ptr_to_u64(&parent_tid);
+				pr_debug("Using clone3 to restore the process\n");
+				RUN_CLONE3_RESTORE_FN(ret, c_args, sizeof(c_args), &thread_args[i], args->clone_restore_fn);
+			} else {
 				last_pid_len = std_vprint_num(last_pid_buf, sizeof(last_pid_buf), thread_args[i].pid - 1, &s);
 				sys_lseek(fd, 0, SEEK_SET);
 				ret = sys_write(fd, s, last_pid_len);
@@ -1807,8 +1825,8 @@ long __export_restore_task(struct task_restore_args *args)
 				 * thread will run with own stack and we must not
 				 * have any additional instructions... oh, dear...
 				 */
-
 				RUN_CLONE_RESTORE_FN(ret, clone_flags, new_sp, parent_tid, thread_args, args->clone_restore_fn);
+			}
 			if (ret != thread_args[i].pid) {
 				pr_err("Unable to create a thread: %ld\n", ret);
 				mutex_unlock(&task_entries_local->last_pid_mutex);