restore: Add mutexes via futex syscall and order threads creation

Signed-off-by: Cyrill Gorcunov <gorcunov@gmail.com>

cr-restore.c

@@ -1453,7 +1453,8 @@ static void sigreturn_restore(pid_t pstree_pid, pid_t pid)
 	task_args->pid		= pid;
 	task_args->fd_core	= fd_core;
 	task_args->fd_self_vmas	= fd_self_vmas;
-	task_args->rst_lock	= 0;
+
+	rst_mutex_init(&task_args->rst_lock);
 
 	if (pstree_entry.nr_threads) {
 		int i;

include/atomic.h

@@ -0,0 +1,42 @@
+#ifndef ATOMIC_H__
+#define ATOMIC_H__
+
+#define atomic_set(mem, v)					\
+	({							\
+		asm volatile ("lock xchg %0, %1\n"		\
+				: "+r" (v), "+m" (*mem)		\
+				:				\
+				: "cc", "memory");		\
+	})
+
+#define atomic_get(mem)						\
+	({							\
+		u32 ret__ = 0;					\
+		asm volatile ("lock xadd %0, %1\n"		\
+				: "+r" (ret__),	"+m" (*mem)	\
+				:				\
+				: "cc", "memory");		\
+		ret__;						\
+	})
+
+#define atomic_inc(mem)						\
+	({							\
+		u32 ret__ = 1;					\
+		asm volatile ("lock xadd %0, %1\n"		\
+				: "+r" (ret__),	"+m" (*mem)	\
+				:				\
+				: "cc", "memory");		\
+		ret__;						\
+	})
+
+#define atomic_dec(mem)						\
+	({							\
+		u32 ret__ = -1;					\
+		asm volatile ("lock xadd %0, %1\n"		\
+				: "+r" (ret__),	"+m" (*mem)	\
+				:				\
+				: "cc", "memory");		\
+		ret__;						\
+	})
+
+#endif /* ATOMIC_H__ */

include/bitops.h

@@ -46,7 +46,6 @@ static void clear_bit(int nr, volatile unsigned long *addr)
 	asm volatile("btr %1,%0" : ADDR : "Ir" (nr));
 }
 
-
 #else /* CONFIG_X86_64 */
 # error x86-32 is not implemented yet
 #endif /* CONFIG_X86_64 */

include/restorer.h

@@ -48,13 +48,15 @@ struct restore_mem_zone {
 #define first_on_heap(ptr, heap)	((typeof(ptr))heap)
 #define next_on_heap(ptr, prev)		((typeof(ptr))((long)(prev) + sizeof(*(prev))))
 
+typedef u32 rst_mutex_t;
+
 /* Make sure it's pow2 in size */
 struct thread_restore_args {
 	struct restore_mem_zone		mem_zone;
 
 	int				pid;
 	int				fd_core;
-	long				*rst_lock;
+	rst_mutex_t			*rst_lock;
 } __aligned(sizeof(long));
 
 struct task_restore_core_args {
@@ -64,7 +66,7 @@ struct task_restore_core_args {
 	int				fd_core;		/* opened core file */
 	int				fd_self_vmas;		/* opened file with running VMAs to unmap */
 	bool				restore_threads;	/* if to restore threads */
-	long				rst_lock;
+	rst_mutex_t			rst_lock;
 
 	/* threads restoration */
 	int				nr_threads;		/* number of threads */
@@ -224,29 +226,27 @@ static void always_inline write_hex_n(unsigned long num)
 	sys_write(1, &c, 1);
 }
 
-static always_inline void rst_lock(long *v)
-{
-	while (*v) {
-		asm volatile("lfence");
-		asm volatile("pause");
-	}
-	(*v)++;
+#define FUTEX_WAIT		0
+#define FUTEX_WAKE		1
 
-	asm volatile("sfence");
+static void always_inline rst_mutex_init(rst_mutex_t *mutex)
+{
+	u32 c = 0;
+	atomic_set(mutex, c);
 }
 
-static always_inline void rst_unlock(long *v)
+static void always_inline rst_mutex_lock(rst_mutex_t *mutex)
 {
-	(*v)--;
-	asm volatile("sfence");
+	u32 c;
+	while ((c = atomic_inc(mutex)))
+		sys_futex(mutex, FUTEX_WAIT, c + 1, NULL, NULL, 0);
 }
 
-static always_inline void rst_wait_unlock(long *v)
+static void always_inline rst_mutex_unlock(rst_mutex_t *mutex)
 {
-	while (*v) {
-		asm volatile("lfence");
-		asm volatile("pause");
-	}
+	u32 c = 0;
+	atomic_set(mutex, c);
+	sys_futex(mutex, FUTEX_WAKE, 1, NULL, NULL, 0);
 }
 
 #endif /* CR_RESTORER_H__ */

include/types.h

@@ -5,6 +5,7 @@
 #include <stdbool.h>
 
 #include "bitops.h"
+#include "atomic.h"
 
 /* prctl */
 #define ARCH_SET_GS 0x1001

restorer.c

@@ -18,7 +18,8 @@
 #include "restorer.h"
 
 /*
- * Threads restoration via sigreturn.
+ * Threads restoration via sigreturn. Note it's locked
+ * routine and calls for unlock at the end.
  */
 long restore_thread(long cmd, struct thread_restore_args *args)
 {
@@ -86,7 +87,7 @@ long restore_thread(long cmd, struct thread_restore_args *args)
 			goto core_restore_end;
 		}
 
-		rst_unlock(args->rst_lock);
+		rst_mutex_unlock(args->rst_lock);
 
 		new_sp = (long)rt_sigframe + 8;
 		asm volatile(
@@ -430,7 +431,7 @@ self_len_end:
 				if (thread_args[i].pid == args->pid)
 					continue;
 
-				rst_lock(&args->rst_lock);
+				rst_mutex_lock(&args->rst_lock);
 
 				new_sp =
 					RESTORE_ALIGN_STACK((long)thread_args[i].mem_zone.stack,
@@ -482,7 +483,7 @@ self_len_end:
 						"g"(&thread_args[i])
 					: "rax", "rdi", "rsi", "rdx", "r10", "memory");
 
-				rst_wait_unlock(&args->rst_lock);
+				rst_mutex_lock(&args->rst_lock);
 			}
 		}