cr: Task CapAmb support

Signed-off-by: Liu Chao <liuchao173@huawei.com>

criu/cr-restore.c

@@ -2992,6 +2992,7 @@ static struct thread_creds_args *rst_prep_creds_args(CredsEntry *ce, unsigned lo
 	args->creds.cap_eff = NULL;
 	args->creds.cap_prm = NULL;
 	args->creds.cap_bnd = NULL;
+	args->creds.cap_amb = NULL;
 	args->creds.groups = NULL;
 	args->creds.lsm_profile = NULL;
 
@@ -2999,6 +3000,7 @@ static struct thread_creds_args *rst_prep_creds_args(CredsEntry *ce, unsigned lo
 	copy_caps(args->cap_eff, ce->cap_eff, ce->n_cap_eff);
 	copy_caps(args->cap_prm, ce->cap_prm, ce->n_cap_prm);
 	copy_caps(args->cap_bnd, ce->cap_bnd, ce->n_cap_bnd);
+	copy_caps(args->cap_amb, ce->cap_amb, ce->n_cap_amb);
 
 	if (ce->n_groups && !groups_match(ce->groups, ce->n_groups)) {
 		unsigned int *groups;

criu/include/parasite.h

@@ -148,6 +148,7 @@ struct parasite_dump_creds {
 	u32 cap_prm[CR_CAP_SIZE];
 	u32 cap_eff[CR_CAP_SIZE];
 	u32 cap_bnd[CR_CAP_SIZE];
+	u32 cap_amb[CR_CAP_SIZE];
 
 	int uids[4];
 	int gids[4];

criu/include/prctl.h

@@ -36,6 +36,15 @@
 #ifndef PR_SET_NO_NEW_PRIVS
 #define PR_SET_NO_NEW_PRIVS 38
 #endif
+#ifndef PR_CAP_AMBIENT
+#define PR_CAP_AMBIENT 47
+#endif
+#ifndef PR_CAP_AMBIENT_IS_SET
+#define PR_CAP_AMBIENT_IS_SET 1
+#endif
+#ifndef PR_CAP_AMBIENT_RAISE
+#define PR_CAP_AMBIENT_RAISE 2
+#endif
 
 #ifndef PR_SET_MM
 #define PR_SET_MM	      35

criu/include/proc_parse.h

@@ -81,6 +81,7 @@ struct proc_status_creds {
 	u32 cap_prm[PROC_CAP_SIZE];
 	u32 cap_eff[PROC_CAP_SIZE];
 	u32 cap_bnd[PROC_CAP_SIZE];
+	u32 cap_amb[PROC_CAP_SIZE];
 };
 
 #define INVALID_UID ((uid_t)-1)

criu/include/restorer.h

@@ -75,6 +75,7 @@ struct thread_creds_args {
 	u32 cap_prm[CR_CAP_SIZE];
 	u32 cap_eff[CR_CAP_SIZE];
 	u32 cap_bnd[CR_CAP_SIZE];
+	u32 cap_amb[CR_CAP_SIZE];
 
 	char *lsm_profile;
 	unsigned int *groups;

criu/parasite-syscall.c

@@ -103,16 +103,19 @@ static int alloc_groups_copy_creds(CredsEntry *ce, struct parasite_dump_creds *c
 	BUILD_BUG_ON(sizeof(ce->cap_prm[0]) != sizeof(c->cap_prm[0]));
 	BUILD_BUG_ON(sizeof(ce->cap_eff[0]) != sizeof(c->cap_eff[0]));
 	BUILD_BUG_ON(sizeof(ce->cap_bnd[0]) != sizeof(c->cap_bnd[0]));
+	BUILD_BUG_ON(sizeof(ce->cap_amb[0]) != sizeof(c->cap_amb[0]));
 
 	BUG_ON(ce->n_cap_inh != CR_CAP_SIZE);
 	BUG_ON(ce->n_cap_prm != CR_CAP_SIZE);
 	BUG_ON(ce->n_cap_eff != CR_CAP_SIZE);
 	BUG_ON(ce->n_cap_bnd != CR_CAP_SIZE);
+	BUG_ON(ce->n_cap_amb != CR_CAP_SIZE);
 
 	memcpy(ce->cap_inh, c->cap_inh, sizeof(c->cap_inh[0]) * CR_CAP_SIZE);
 	memcpy(ce->cap_prm, c->cap_prm, sizeof(c->cap_prm[0]) * CR_CAP_SIZE);
 	memcpy(ce->cap_eff, c->cap_eff, sizeof(c->cap_eff[0]) * CR_CAP_SIZE);
 	memcpy(ce->cap_bnd, c->cap_bnd, sizeof(c->cap_bnd[0]) * CR_CAP_SIZE);
+	memcpy(ce->cap_amb, c->cap_amb, sizeof(c->cap_amb[0]) * CR_CAP_SIZE);
 
 	if (c->no_new_privs > 0) {
 		ce->no_new_privs = c->no_new_privs;

criu/pie/parasite.c

@@ -324,6 +324,7 @@ static int dump_creds(struct parasite_dump_creds *args)
 		args->cap_prm[i] = data[i].prm;
 		args->cap_inh[i] = data[i].inh;
 		args->cap_bnd[i] = 0;
+		args->cap_amb[i] = 0;
 
 		for (j = 0; j < 32; j++) {
 			if (j + i * 32 > args->cap_last_cap)
@@ -336,6 +337,18 @@ static int dump_creds(struct parasite_dump_creds *args)
 			if (ret)
 				args->cap_bnd[i] |= (1 << j);
 		}
+
+		for (j = 0; j < 32; j++) {
+			if (j + i * 32 > args->cap_last_cap)
+				break;
+			ret = sys_prctl(PR_CAP_AMBIENT, PR_CAP_AMBIENT_IS_SET, j + i * 32, 0, 0);
+			if (ret < 0) {
+				pr_err("Unable to read ambient capability %d: %d\n", j + i * 32, ret);
+				return -1;
+			}
+			if (ret)
+				args->cap_amb[i] |= (1 << j);
+		}
 	}
 
 	args->no_new_privs = sys_prctl(PR_GET_NO_NEW_PRIVS, 0, 0, 0, 0);

criu/pie/restorer.c

@@ -347,6 +347,22 @@ skip_xids:
 		return -1;
 	}
 
+	for (b = 0; b < CR_CAP_SIZE; b++) {
+		for (i = 0; i < 32; i++) {
+			if (b * 32 + i > args->cap_last_cap)
+				break;
+			if ((args->cap_amb[b] & (1 << i)) == 0)
+				/* don't set */
+				continue;
+			ret = sys_prctl(PR_CAP_AMBIENT, PR_CAP_AMBIENT_RAISE, i + b * 32, 0, 0);
+			if (!ret)
+				continue;
+			pr_err("Unable to raise ambient capability %d: %d\n", i + b * 32, ret);
+			return -1;
+		}
+	}
+
+
 	if (lsm_type != LSMTYPE__SELINUX) {
 		/*
 		 * SELinux does not support setting the process context for

criu/proc_parse.c

@@ -1071,7 +1071,7 @@ int parse_pid_status(pid_t pid, struct seize_task_status *ss, void *data)
 	if (bfdopenr(&f))
 		return -1;
 
-	while (done < 13) {
+	while (done < 14) {
 		str = breadline(&f);
 		if (str == NULL)
 			break;
@@ -1155,6 +1155,13 @@ int parse_pid_status(pid_t pid, struct seize_task_status *ss, void *data)
 			continue;
 		}
 
+		if (!strncmp(str, "CapAmb:", 7)) {
+			if (cap_parse(str + 8, cr->cap_amb))
+				goto err_parse;
+			done++;
+			continue;
+		}
+
 		if (!strncmp(str, "Seccomp:", 8)) {
 			if (sscanf(str + 9, "%d", &cr->s.seccomp_mode) != 1) {
 				goto err_parse;
@@ -1198,7 +1205,7 @@ int parse_pid_status(pid_t pid, struct seize_task_status *ss, void *data)
 	}
 
 	/* seccomp and nspids are optional */
-	expected_done = (parsed_seccomp ? 12 : 11);
+	expected_done = (parsed_seccomp ? 13 : 12);
 	if (kdat.has_nspid)
 		expected_done++;
 	if (done == expected_done)

criu/pstree.c

@@ -63,6 +63,7 @@ CoreEntry *core_entry_alloc(int th, int tsk)
 		sz += CR_CAP_SIZE * sizeof(ce->cap_prm[0]);
 		sz += CR_CAP_SIZE * sizeof(ce->cap_eff[0]);
 		sz += CR_CAP_SIZE * sizeof(ce->cap_bnd[0]);
+		sz += CR_CAP_SIZE * sizeof(ce->cap_amb[0]);
 		/*
 		 * @groups are dynamic and allocated
 		 * on demand.
@@ -122,10 +123,12 @@ CoreEntry *core_entry_alloc(int th, int tsk)
 			ce->n_cap_prm = CR_CAP_SIZE;
 			ce->n_cap_eff = CR_CAP_SIZE;
 			ce->n_cap_bnd = CR_CAP_SIZE;
+			ce->n_cap_amb = CR_CAP_SIZE;
 			ce->cap_inh = xptr_pull_s(&m, CR_CAP_SIZE * sizeof(ce->cap_inh[0]));
 			ce->cap_prm = xptr_pull_s(&m, CR_CAP_SIZE * sizeof(ce->cap_prm[0]));
 			ce->cap_eff = xptr_pull_s(&m, CR_CAP_SIZE * sizeof(ce->cap_eff[0]));
 			ce->cap_bnd = xptr_pull_s(&m, CR_CAP_SIZE * sizeof(ce->cap_bnd[0]));
+			ce->cap_amb = xptr_pull_s(&m, CR_CAP_SIZE * sizeof(ce->cap_amb[0]));
 
 			if (arch_alloc_thread_info(core)) {
 				xfree(core);

images/creds.proto

@@ -25,4 +25,6 @@ message creds_entry {
 	optional string lsm_sockcreate = 16;
 	optional bytes	apparmor_data	= 17;
 	optional uint32	no_new_privs	= 18;
+
+	repeated uint32 cap_amb = 19;
 }