dump+restore: Implement membarrier() registration c/r.

Note: Silently drops MEMBARRIER_CMD_REGISTER_GLOBAL_EXPEDITED as it's
not currently detectable. This is still better than silently dropping
all membarrier() registrations.

Signed-off-by: Michał Mirosław <[email protected]>

compel/arch/arm/plugins/std/syscalls/syscall.def

@@ -122,3 +122,4 @@ pidfd_open			434	434	(pid_t pid, unsigned int flags)
 openat2				437	437	(int dirfd, char *pathname, struct open_how *how, size_t size)
 pidfd_getfd			438	438	(int pidfd, int targetfd, unsigned int flags)
 rseq				293	398	(void *rseq, uint32_t rseq_len, int flags, uint32_t sig)
+membarrier 			283	389	(int cmd, unsigned int flags, int cpu_id)

compel/arch/loongarch64/plugins/std/syscalls/syscall_64.tbl

@@ -102,6 +102,7 @@ __NR_kcmp			2	sys_kcmp		(pid_t pid1, pid_t pid2, int type, unsigned long idx1, u
 __NR_seccomp			2	sys_seccomp		(unsigned int op, unsigned int flags, const char *uargs)
 __NR_memfd_create		2	sys_memfd_create	(const char *name, unsigned int flags)
 __NR_userfaultfd		2	sys_userfaultfd		(int flags)
+__NR_membarrier			3	sys_membarrier		(int cmd, unsigned int flags, int cpu_id)
 __NR_rseq			2	sys_rseq		(void *rseq, uint32_t rseq_len, int flags, uint32_t sig)
 __NR_open_tree			4	sys_open_tree		(int dirfd, const char *pathname, unsigned int flags)
 __NR_move_mount			4	sys_move_mount		(int from_dfd, const char *from_pathname, int to_dfd, const char *to_pathname, int flags)

compel/arch/mips/plugins/std/syscalls/syscall_64.tbl

@@ -119,3 +119,4 @@ __NR_pidfd_open			5434		sys_pidfd_open		(pid_t pid, unsigned int flags)
 __NR_openat2			5437		sys_openat2		(int dirfd, char *pathname, struct open_how *how, size_t size)
 __NR_pidfd_getfd		5438		sys_pidfd_getfd		(int pidfd, int targetfd, unsigned int flags)
 __NR_rseq		        5327		sys_rseq		(void *rseq, uint32_t rseq_len, int flags, uint32_t sig)
+__NR_membarrier 		5318		sys_membarrier		(int cmd, unsigned int flags, int cpu_id)

compel/arch/ppc64/plugins/std/syscalls/syscall-ppc64.tbl

@@ -118,3 +118,4 @@ __NR_pidfd_open		434		sys_pidfd_open		(pid_t pid, unsigned int flags)
 __NR_openat2		437		sys_openat2		(int dirfd, char *pathname, struct open_how *how, size_t size)
 __NR_pidfd_getfd	438		sys_pidfd_getfd		(int pidfd, int targetfd, unsigned int flags)
 __NR_rseq       	387		sys_rseq		(void *rseq, uint32_t rseq_len, int flags, uint32_t sig)
+__NR_membarrier 	365		sys_membarrier		(int cmd, unsigned int flags, int cpu_id)

compel/arch/s390/plugins/std/syscalls/syscall-s390.tbl

@@ -118,3 +118,4 @@ __NR_pidfd_open		434		sys_pidfd_open		(pid_t pid, unsigned int flags)
 __NR_openat2		437		sys_openat2		(int dirfd, char *pathname, struct open_how *how, size_t size)
 __NR_pidfd_getfd	438		sys_pidfd_getfd		(int pidfd, int targetfd, unsigned int flags)
 __NR_rseq       	383		sys_rseq		(void *rseq, uint32_t rseq_len, int flags, uint32_t sig)
+__NR_membarrier 	356		sys_membarrier		(int cmd, unsigned int flags, int cpu_id)

compel/arch/x86/plugins/std/syscalls/syscall_32.tbl

@@ -106,3 +106,4 @@ __NR_pidfd_open		434		sys_pidfd_open		(pid_t pid, unsigned int flags)
 __NR_openat2		437		sys_openat2		(int dirfd, char *pathname, struct open_how *how, size_t size)
 __NR_pidfd_getfd	438		sys_pidfd_getfd		(int pidfd, int targetfd, unsigned int flags)
 __NR_rseq       	386		sys_rseq		(void *rseq, uint32_t rseq_len, int flags, uint32_t sig)
+__NR_membarrier 	375		sys_membarrier		(int cmd, unsigned int flags, int cpu_id)

compel/arch/x86/plugins/std/syscalls/syscall_64.tbl

@@ -117,3 +117,4 @@ __NR_pidfd_open			434		sys_pidfd_open		(pid_t pid, unsigned int flags)
 __NR_openat2		437		sys_openat2		(int dirfd, char *pathname, struct open_how *how, size_t size)
 __NR_pidfd_getfd		438		sys_pidfd_getfd		(int pidfd, int targetfd, unsigned int flags)
 __NR_rseq       		334		sys_rseq		(void *rseq, uint32_t rseq_len, int flags, uint32_t sig)
+__NR_membarrier 		324		sys_membarrier		(int cmd, unsigned int flags, int cpu_id)

criu/cr-dump.c

@@ -770,6 +770,11 @@ static int dump_task_core_all(struct parasite_ctl *ctl, struct pstree_item *item
 	core->tc->child_subreaper = misc->child_subreaper;
 	core->tc->has_child_subreaper = true;
 
+	if (misc->membarrier_registration_mask) {
+		core->tc->membarrier_registration_mask = misc->membarrier_registration_mask;
+		core->tc->has_membarrier_registration_mask = true;
+	}
+
 	ret = get_task_personality(pid, &core->tc->personality);
 	if (ret < 0)
 		goto err;

criu/cr-restore.c

@@ -863,6 +863,9 @@ static int prepare_proc_misc(pid_t pid, TaskCoreEntry *tc, struct task_restore_a
 	if (tc->has_child_subreaper)
 		args->child_subreaper = tc->child_subreaper;
 
+	if (tc->has_membarrier_registration_mask)
+		args->membarrier_registration_mask = tc->membarrier_registration_mask;
+
 	/* loginuid value is critical to restore */
 	if (kdat.luid == LUID_FULL && tc->has_loginuid && tc->loginuid != INVALID_UID) {
 		ret = prepare_loginuid(tc->loginuid);

criu/include/parasite.h

@@ -128,6 +128,7 @@ struct parasite_dump_misc {
 	int dumpable;
 	int thp_disabled;
 	int child_subreaper;
+	int membarrier_registration_mask;
 };
 
 /*

criu/include/restorer.h

@@ -229,6 +229,7 @@ struct task_restore_args {
 #endif
 	int lsm_type;
 	int child_subreaper;
+	int membarrier_registration_mask;
 	bool has_clone3_set_tid;
 
 	/*

criu/pie/parasite.c

@@ -211,6 +211,42 @@ static int dump_thread_common(struct parasite_dump_thread *ti)
 	return ret;
 }
 
+/*
+ * Returns a membarrier() registration command (it is a bitmask) if the process
+ * was registered for specified (as a bit index) membarrier()-issuing command;
+ * returns zero otherwise.
+ */
+static int get_membarrier_registration_mask(int cmd_bit)
+{
+	unsigned cmd = 1 << cmd_bit;
+	int ret;
+
+	/*
+	 * Issuing a barrier will be successful only if the process was registered
+	 * for this type of membarrier. All errors are a sign that the type issued
+	 * was not registered (EPERM) or not supported by kernel (EINVAL or ENOSYS).
+	 */
+	ret = sys_membarrier(cmd, 0, 0);
+	if (ret && ret != -EPERM && ret != -EINVAL && ret != -ENOSYS) {
+		pr_err("membarrier(1 << %d) returned %d\n", cmd_bit, ret);
+		return -1;
+	}
+	pr_debug("membarrier(1 << %d) returned %d\n", cmd_bit, ret);
+	/*
+	 * For supported registrations, MEMBARRIER_CMD_REGISTER_xxx = MEMBARRIER_CMD_xxx << 1.
+	 * See: enum membarrier_cmd in include/uapi/linux/membarrier.h in kernel sources.
+	 */
+	return ret ? 0 : cmd << 1;
+}
+
+/*
+ * It would be better to check the following with BUILD_BUG_ON, but we might
+ * have an old linux/membarrier.h header without necessary enum values.
+ */
+#define MEMBARRIER_CMDBIT_PRIVATE_EXPEDITED	      3
+#define MEMBARRIER_CMDBIT_PRIVATE_EXPEDITED_SYNC_CORE 5
+#define MEMBARRIER_CMDBIT_PRIVATE_EXPEDITED_RSEQ      7
+
 static int dump_misc(struct parasite_dump_misc *args)
 {
 	int ret;
@@ -225,6 +261,20 @@ static int dump_misc(struct parasite_dump_misc *args)
 	args->dumpable = sys_prctl(PR_GET_DUMPABLE, 0, 0, 0, 0);
 	args->thp_disabled = sys_prctl(PR_GET_THP_DISABLE, 0, 0, 0, 0);
 
+	args->membarrier_registration_mask = 0;
+	ret = get_membarrier_registration_mask(MEMBARRIER_CMDBIT_PRIVATE_EXPEDITED);
+	if (ret < 0)
+		return -1;
+	args->membarrier_registration_mask |= ret;
+	ret = get_membarrier_registration_mask(MEMBARRIER_CMDBIT_PRIVATE_EXPEDITED_SYNC_CORE);
+	if (ret < 0)
+		return -1;
+	args->membarrier_registration_mask |= ret;
+	ret = get_membarrier_registration_mask(MEMBARRIER_CMDBIT_PRIVATE_EXPEDITED_RSEQ);
+	if (ret < 0)
+		return -1;
+	args->membarrier_registration_mask |= ret;
+
 	ret = sys_prctl(PR_GET_CHILD_SUBREAPER, (unsigned long)&args->child_subreaper, 0, 0, 0);
 	if (ret)
 		pr_err("PR_GET_CHILD_SUBREAPER failed (%d)\n", ret);

criu/pie/restorer.c

@@ -1537,6 +1537,30 @@ int cleanup_current_inotify_events(struct task_restore_args *task_args)
 	return 0;
 }
 
+/*
+ * Restore membarrier() registrations.
+ */
+static int restore_membarrier_registrations(int mask)
+{
+	unsigned long bitmap[1] = { mask };
+	int i, err, ret = 0;
+
+	if (!mask)
+		return 0;
+
+	pr_info("Restoring membarrier() registrations %x\n", mask);
+
+	for_each_bit(i, bitmap) {
+		err = sys_membarrier(1 << i, 0, 0);
+		if (!err)
+			continue;
+		pr_err("Can't restore membarrier(1 << %d) registration: %d\n", i, err);
+		ret = -1;
+	}
+
+	return ret;
+}
+
 /*
  * The main routine to restore task via sigreturn.
  * This one is very special, we never return there
@@ -2023,6 +2047,9 @@ long __export_restore_task(struct task_restore_args *args)
 		goto core_restore_end;
 	}
 
+	if (restore_membarrier_registrations(args->membarrier_registration_mask) < 0)
+		goto core_restore_end;
+
 	pr_info("%ld: Restored\n", sys_getpid());
 
 	restore_finish_stage(task_entries_local, CR_STATE_RESTORE);

images/core.proto

@@ -64,6 +64,8 @@ message task_core_entry {
 	optional uint64			blk_sigset_extended = 20[(criu).hex = true];
 
 	optional uint32			stop_signo = 21;
+
+	optional uint32			membarrier_registration_mask	= 22 [(criu).hex = true];
 }
 
 message task_kobj_ids_entry {