cgroup-v2: Dump cgroup controllers of every threads in a process

Currently, we assume all threads in process are in the same cgroup controllers.
However, with threaded controllers, threads in a process may be in different
controllers. So we need to dump cgroup controllers of every threads in process
and fixup the procfs cgroup parsing to parse from self/task/<tid>/cgroup.

Signed-off-by: Bui Quang Minh <minhquangbui99@gmail.com>

criu/cgroup.c

@@ -174,6 +174,7 @@ struct cg_controller *new_controller(const char *name)
 	nc->n_controllers = 1;
 
 	nc->n_heads = 0;
+	nc->is_threaded = false;
 	INIT_LIST_HEAD(&nc->heads);
 
 	return nc;
@@ -371,7 +372,8 @@ static void free_all_cgroup_props(struct cgroup_dir *ncd)
 	ncd->n_properties = 0;
 }
 
-static int dump_cg_props_array(const char *fpath, struct cgroup_dir *ncd, const cgp_t *cgp)
+static int dump_cg_props_array(const char *fpath, struct cgroup_dir *ncd, const cgp_t *cgp,
+			       struct cg_controller *controller)
 {
 	int j;
 	char buf[PATH_MAX];
@@ -422,6 +424,13 @@ static int dump_cg_props_array(const char *fpath, struct cgroup_dir *ncd, const
 			prop->value = new;
 		}
 
+		/*
+		 * Set the is_threaded flag if cgroup.type's value is threaded,
+		 * ignore all other values.
+		 */
+		if (!strcmp("cgroup.type", prop->name) && !strcmp("threaded", prop->value))
+			controller->is_threaded = true;
+
 		pr_info("Dumping value %s from %s/%s\n", prop->value, fpath, prop->name);
 		list_add_tail(&prop->list, &ncd->properties);
 		ncd->n_properties++;
@@ -437,7 +446,7 @@ static int add_cgroup_properties(const char *fpath, struct cgroup_dir *ncd, stru
 	for (i = 0; i < controller->n_controllers; ++i) {
 		const cgp_t *cgp = cgp_get_props(controller->controllers[i]);
 
-		if (dump_cg_props_array(fpath, ncd, cgp) < 0) {
+		if (dump_cg_props_array(fpath, ncd, cgp, controller) < 0) {
 			pr_err("dumping known properties failed\n");
 			return -1;
 		}
@@ -445,12 +454,12 @@ static int add_cgroup_properties(const char *fpath, struct cgroup_dir *ncd, stru
 
 	/* cgroup v2 */
 	if (controller->controllers[0][0] == 0) {
-		if (dump_cg_props_array(fpath, ncd, &cgp_global_v2) < 0) {
+		if (dump_cg_props_array(fpath, ncd, &cgp_global_v2, controller) < 0) {
 			pr_err("dumping global properties v2 failed\n");
 			return -1;
 		}
 	} else {
-		if (dump_cg_props_array(fpath, ncd, &cgp_global) < 0) {
+		if (dump_cg_props_array(fpath, ncd, &cgp_global, controller) < 0) {
 			pr_err("dumping global properties failed\n");
 			return -1;
 		}
@@ -735,9 +744,9 @@ static int collect_cgroups(struct list_head *ctls)
 	return 0;
 }
 
-int dump_task_cgroup(struct pstree_item *item, u32 *cg_id, struct parasite_dump_cgroup_args *args)
+int dump_thread_cgroup(const struct pstree_item *item, u32 *cg_id, struct parasite_dump_cgroup_args *args, int id)
 {
-	int pid;
+	int pid, tid;
 	LIST_HEAD(ctls);
 	unsigned int n_ctls = 0;
 	struct cg_set *cs;
@@ -750,8 +759,13 @@ int dump_task_cgroup(struct pstree_item *item, u32 *cg_id, struct parasite_dump_
 	else
 		pid = getpid();
 
-	pr_info("Dumping cgroups for %d\n", pid);
-	if (parse_task_cgroup(pid, args, &ctls, &n_ctls))
+	if (id < 0)
+		tid = pid;
+	else
+		tid = item->threads[id].real;
+
+	pr_info("Dumping cgroups for thread %d\n", tid);
+	if (parse_thread_cgroup(pid, tid, args, &ctls, &n_ctls))
 		return -1;
 
 	cs = get_cg_set(&ctls, n_ctls, item);
@@ -764,9 +778,10 @@ int dump_task_cgroup(struct pstree_item *item, u32 *cg_id, struct parasite_dump_
 		pr_info("Set %d is criu one\n", cs->id);
 	} else {
 		if (item == root_item) {
-			BUG_ON(root_cgset);
-			root_cgset = cs;
-			pr_info("Set %d is root one\n", cs->id);
+			if (!root_cgset) {
+				root_cgset = cs;
+				pr_info("Set %d is root one\n", cs->id);
+			}
 		} else {
 			struct cg_ctl *root, *stray;
 
@@ -913,6 +928,7 @@ static int dump_controllers(CgroupEntry *cg)
 	list_for_each_entry(cur, &cgroups, l) {
 		cg_controller_entry__init(ce);
 
+		ce->is_threaded = cur->is_threaded;
 		ce->cnames = cur->controllers;
 		ce->n_cnames = cur->n_controllers;
 		ce->n_dirs = cur->n_heads;

criu/cr-dump.c

@@ -759,6 +759,7 @@ static int dump_task_core_all(struct parasite_ctl *ctl, struct pstree_item *item
 	pid_t pid = item->pid->real;
 	int ret = -1;
 	struct parasite_dump_cgroup_args cgroup_args, *info = NULL;
+	u32 *cg_set;
 
 	BUILD_BUG_ON(sizeof(cgroup_args) < PARASITE_ARG_SIZE_MIN);
 
@@ -804,13 +805,23 @@ static int dump_task_core_all(struct parasite_ctl *ctl, struct pstree_item *item
 	 */
 	if (item->ids->has_cgroup_ns_id && !item->parent) {
 		info = &cgroup_args;
+		strcpy(cgroup_args.thread_cgrp, "self/cgroup");
 		ret = parasite_dump_cgroup(ctl, &cgroup_args);
 		if (ret)
 			goto err;
 	}
 
-	core->tc->has_cg_set = true;
-	ret = dump_task_cgroup(item, &core->tc->cg_set, info);
+	/*
+	 * We don't support multithreads zombie tasks so there is
+	 * no thread_core in zombie tasks, store the cg_set in
+	 * task_core in these cases.
+	 */
+	cg_set = &core->thread_core->cg_set;
+	if (item->pid->state == TASK_THREAD) {
+		core->tc->has_cg_set = true;
+		cg_set = &core->tc->cg_set;
+	}
+	ret = dump_thread_cgroup(item, cg_set, info, -1);
 	if (ret)
 		goto err;
 
@@ -1409,6 +1420,38 @@ err:
 	return ret;
 }
 
+static int dump_task_cgroup(struct parasite_ctl *parasite_ctl, const struct pstree_item *item)
+{
+	struct parasite_dump_cgroup_args cgroup_args, *info;
+	int i;
+
+	BUILD_BUG_ON(sizeof(cgroup_args) < PARASITE_ARG_SIZE_MIN);
+	for (i = 0; i < item->nr_threads; i++) {
+		CoreEntry *core = item->core[i];
+
+		/* Leader is already dumped */
+		if (item->pid->real == item->threads[i].real)
+			continue;
+
+		/* For now, we only need to dump the root task's cgroup ns, because we
+		 * know all the tasks are in the same cgroup namespace because we don't
+		 * allow nesting.
+		 */
+		info = NULL;
+		if (item->ids->has_cgroup_ns_id && !item->parent) {
+			info = &cgroup_args;
+			sprintf(cgroup_args.thread_cgrp, "self/task/%d/cgroup", item->threads[i].ns[0].virt);
+			if (parasite_dump_cgroup(parasite_ctl, &cgroup_args))
+				return -1;
+		}
+
+		if (dump_thread_cgroup(item, &core->thread_core->cg_set, info, i))
+			return -1;
+	}
+
+	return 0;
+}
+
 static int pre_dump_one_task(struct pstree_item *item, InventoryEntry *parent_ie)
 {
 	pid_t pid = item->pid->real;
@@ -1681,6 +1724,12 @@ static int dump_one_task(struct pstree_item *item, InventoryEntry *parent_ie)
 		goto err_cure;
 	}
 
+	ret = dump_task_cgroup(parasite_ctl, item);
+	if (ret) {
+		pr_err("Dump cgroup of threads in process (pid: %d) failed with %d\n", pid, ret);
+		goto err_cure;
+	}
+
 	ret = compel_stop_daemon(parasite_ctl);
 	if (ret) {
 		pr_err("Can't stop daemon in parasite (pid: %d)\n", pid);

criu/image.c

@@ -228,7 +228,7 @@ int prepare_inventory(InventoryEntry *he)
 
 	if (!opts.unprivileged)
 		he->has_root_cg_set = true;
-	if (dump_task_cgroup(NULL, &he->root_cg_set, NULL))
+	if (dump_thread_cgroup(NULL, &he->root_cg_set, NULL, -1))
 		return -1;
 
 	he->root_ids = crt.i.ids;

criu/include/cgroup.h

@@ -7,7 +7,7 @@
 struct pstree_item;
 struct parasite_dump_cgroup_args;
 extern u32 root_cg_set;
-int dump_task_cgroup(struct pstree_item *, u32 *, struct parasite_dump_cgroup_args *args);
+int dump_thread_cgroup(const struct pstree_item *, u32 *, struct parasite_dump_cgroup_args *args, int id);
 int dump_cgroups(void);
 int prepare_task_cgroup(struct pstree_item *);
 int prepare_cgroup(void);
@@ -60,6 +60,9 @@ struct cg_controller {
 
 	/* for cgroup list in cgroup.c */
 	struct list_head l;
+
+	/* controller is a threaded cgroup or not */
+	int is_threaded;
 };
 struct cg_controller *new_controller(const char *name);
 
@@ -87,7 +90,8 @@ struct cg_ctl {
  */
 struct list_head;
 struct parasite_dump_cgroup_args;
-extern int parse_task_cgroup(int pid, struct parasite_dump_cgroup_args *args, struct list_head *l, unsigned int *n);
+extern int parse_thread_cgroup(int pid, int tid, struct parasite_dump_cgroup_args *args, struct list_head *l,
+			       unsigned int *n);
 extern void put_ctls(struct list_head *);
 
 int collect_controllers(struct list_head *cgroups, unsigned int *n_cgroups);

criu/include/parasite.h

@@ -241,7 +241,12 @@ struct parasite_dump_cgroup_args {
 	 *
 	 * The string is null terminated.
 	 */
-	char contents[1 << 12];
+	char contents[(1 << 12) - 32];
+	/*
+	 * Contains the path to thread cgroup procfs.
+	 * "self/task/<tid>/cgroup"
+	 */
+	char thread_cgrp[32];
 };
 
 #endif /* !__ASSEMBLY__ */

criu/parasite-syscall.c

@@ -513,6 +513,7 @@ int parasite_dump_cgroup(struct parasite_ctl *ctl, struct parasite_dump_cgroup_a
 	struct parasite_dump_cgroup_args *ca;
 
 	ca = compel_parasite_args(ctl, struct parasite_dump_cgroup_args);
+	memcpy(ca->thread_cgrp, cgroup->thread_cgrp, sizeof(ca->thread_cgrp));
 	ret = compel_rpc_call_sync(PARASITE_CMD_DUMP_CGROUP, ctl);
 	if (ret) {
 		pr_err("Parasite failed to dump /proc/self/cgroup\n");

criu/pie/parasite.c

@@ -745,7 +745,7 @@ static int parasite_dump_cgroup(struct parasite_dump_cgroup_args *args)
 		return -1;
 	}
 
-	cgroup = sys_openat(proc, "self/cgroup", O_RDONLY, 0);
+	cgroup = sys_openat(proc, args->thread_cgrp, O_RDONLY, 0);
 	sys_close(proc);
 	if (cgroup < 0) {
 		pr_err("can't get /proc/self/cgroup fd\n");

criu/proc_parse.c

@@ -2549,7 +2549,8 @@ err:
 	return -1;
 }
 
-int parse_task_cgroup(int pid, struct parasite_dump_cgroup_args *args, struct list_head *retl, unsigned int *n)
+int parse_thread_cgroup(int pid, int tid, struct parasite_dump_cgroup_args *args, struct list_head *retl,
+			unsigned int *n)
 {
 	FILE *f;
 	int ret;
@@ -2557,7 +2558,7 @@ int parse_task_cgroup(int pid, struct parasite_dump_cgroup_args *args, struct li
 	unsigned int n_internal = 0;
 	struct cg_ctl *intern, *ext;
 
-	f = fopen_proc(pid, "cgroup");
+	f = fopen_proc(pid, "task/%d/cgroup", tid);
 	if (!f)
 		return -1;
 

images/cgroup.proto

@@ -24,6 +24,7 @@ message cgroup_dir_entry {
 message cg_controller_entry {
 	repeated string			cnames		= 1;
 	repeated cgroup_dir_entry	dirs		= 2;
+	required bool			is_threaded	= 3;
 }
 
 message cg_member_entry {

images/core.proto

@@ -105,6 +105,7 @@ message thread_core_entry {
 	optional string			comm		= 13;
 	optional uint64			blk_sigset_extended	= 14;
 	optional rseq_entry		rseq_entry	= 15;
+	required uint32			cg_set		= 16;
 }
 
 message task_rlimits_entry {