Attachment #30332 for ticket #16713




 *                 (task_id = NO_VAL)
 */

/*
 * Get the relative path to the root of the cgroup namespace by
 * reading /proc/PID/mountinfo for the given PID.
 *
 * In the root cgroup namespace (most environments) we expect that the
 * cgroup2 filesystem is mounted on /, such as:
 * 34 25 0:29 / /sys/fs/cgroup rw,nosuid,nodev,noexec,relatime shared:9 - cgroup2 cgroup2 rw
 *
 * If the given PID is in a cgroup namespace, then it could look like:
 * 3526 3512 0:29 /../../../../cont1 /sys/fs/cgroup rw,nosuid,nodev,noexec,relatime - cgroup2 cgroup2 rw
 *
 * See man:cgroup_namespaces(7).
 *
 */
static char *_get_relative_cg_path(const char *pid)
{
	char *buf, *p, *ret = NULL, *procfile=NULL, *mount=NULL;
	size_t sz, name_len;
	int i;

	xstrfmtcat(procfile, "/proc/%s/mountinfo", pid);
	if (common_file_read_content(procfile, &buf, &sz) !=
	    SLURM_SUCCESS)
		fatal("cannot read /proc/?/mountinfo contents: %m");
	debug4("%s contents: %s", procfile, buf);

	if (!(p = xstrstr(buf, "cgroup2 cgroup2"))) {
		fatal("Did not find cgroup2 mount point in %s", procfile);
	}

	*p = '\0';
	if ((mount = xstrrchr(buf, '\n')) == NULL)
		mount = buf;
	debug4("First 8 data of cgroup2 line in %s is %s.", procfile, mount);

	p = mount;
	for (i=0; (i < 4) && (p != NULL); i++)
		p = strtok((i==0)?mount:NULL, " ");
	if (p == NULL)
		fatal("Could not parse cgroup2 mount point in %s.", procfile);
	ret = xstrdup(p);
	debug2("Relative path to root of cgroup namespace is %s.", ret);
	xfree(buf);
	xfree(procfile);
	return ret;
}

/*
 * The cgroup v2 documented way to know which is the process root in the cgroup
 * hierarchy is just to read /proc/self/cgroup. In Unified hierarchies this
- 
--
src/plugins/cgroup/v2/cgroup_v2.c | 30 +++++++++++++++++++++++++++---
1 file changed, 27 insertions(+), 3 deletions(-)

Lines 178-188 static char _get_relative_cg_path(const char pid) Link Here

(-)a/src/plugins/cgroup/v2/cgroup_v2.c (-5 / +27 lines)
178	static char *_get_self_cg_path()	178	static char *_get_self_cg_path()
179	{	179	{
180	char buf, start = NULL, p, ret = NULL;	180	char buf, start = NULL, p, ret = NULL;
181	size_t sz;	181	char *relative_cg_path = NULL;
		182	size_t sz, relative_cg_len;
182		183
183	if (common_file_read_content("/proc/self/cgroup", &buf, &sz) !=	184	if (common_file_read_content("/proc/self/cgroup", &buf, &sz) !=
184	SLURM_SUCCESS)	185	SLURM_SUCCESS)
185	fatal("cannot read /proc/self/cgroup contents: %m");	186	fatal("cannot read /proc/self/cgroup contents: %m");
		187	debug4("Self cgroup data is %s.", buf);
186		188
187	/*	189	/*
188	* In Unified mode there will be just one line containing the path	190	* In Unified mode there will be just one line containing the path
Lines 196-201 static char *_get_self_cg_path() Link Here
196	* If we have multiple slurmd, we will likely have one unit file per	198	* If we have multiple slurmd, we will likely have one unit file per
197	* node, and the path takes the name of the service file, e.g:	199	* node, and the path takes the name of the service file, e.g:
198	* /sys/fs/cgroup/system.slice/slurmd-<nodename>.service/	200	* /sys/fs/cgroup/system.slice/slurmd-<nodename>.service/
		201	*
		202	* If the cgroup path contains .. then cgroup namespaces are
		203	* in use. In this case remove the relative path to the root
		204	* of the cgroup namespace from our path.
199	*/	205	*/
200	if ((p = xstrchr(buf, ':'))) {	206	if ((p = xstrchr(buf, ':'))) {
201	if ((p + 2) < (buf + sz - 1))	207	if ((p + 2) < (buf + sz - 1))
Lines 205-212 static char *_get_self_cg_path() Link Here
205	if (start && (*start != '\0')) {	211	if (start && (*start != '\0')) {
206	if ((p = xstrchr(start, '\n')))	212	if ((p = xstrchr(start, '\n')))
207	*p = '\0';	213	*p = '\0';
208	xstrfmtcat(ret, "%s%s",	214	debug3("Found self cgroup path %s.", start);
209	slurm_cgroup_conf.cgroup_mountpoint, start);	215	if (xstrstr(start, "..")) {
		216	debug3("Relative path to cgroup namespace root found; removing the relative path prefix.");
		217	relative_cg_path = _get_relative_cg_path("self");
		218	if (xstrstr(start, relative_cg_path)) {
		219	relative_cg_len = strlen(relative_cg_path);
		220	xstrfmtcat(ret, "%s%s",
		221	slurm_cgroup_conf.cgroup_mountpoint,
		222	start + relative_cg_len);
		223	debug2("Using %s as my cgroup path.", ret);
		224	} else {
		225	error("Did not find relative path to cgroup namespace root prefix %s in self cgroup path %s.",
		226	relative_cg_path, start);
		227	}
		228	xfree(relative_cg_path);
		229	} else {
		230	debug2("Using constructed cgroup path %s + %s.", slurm_cgroup_conf.cgroup_mountpoint, start);
		231	xstrfmtcat(ret, "%s%s",
		232	slurm_cgroup_conf.cgroup_mountpoint, start);
		233	}
210	}	234	}
211		235
212	xfree(buf);	236	xfree(buf);
213	-
214	--
215	src/plugins/cgroup/v2/cgroup_v2.c \| 31 ++++++++++++++++++++++++++++---	237	src/plugins/cgroup/v2/cgroup_v2.c \| 31 ++++++++++++++++++++++++++++---
216	1 file changed, 28 insertions(+), 3 deletions(-)	238	1 file changed, 28 insertions(+), 3 deletions(-)

Lines 246-252 static char *_get_self_cg_path() Link Here

(-)a/src/plugins/cgroup/v2/cgroup_v2.c (-5 / +28 lines)
246	* But in containerized environments it could look like:	246	* But in containerized environments it could look like:
247	* "0::/docker.slice/docker-<some UUID>.scope/init.scope"	247	* "0::/docker.slice/docker-<some UUID>.scope/init.scope"
248	*	248	*
249	* This function just strips the "0::" and "init.scope" portions.	249	* If cgroup namespaces are in use it could look like:
		250	* 0::/../../../../cont1/init.scope
		251	*
		252	* This function just strips the "0::" and "init.scope" portions and,
		253	* if cgroup namespaces are in use, removes the relative path to the
		254	* root of the cgroup namespace.
250	*	255	*
251	* In normal systems the final path will look like this:	256	* In normal systems the final path will look like this:
252	* /sys/fs/cgroup[/]	257	* /sys/fs/cgroup[/]
Lines 254-268 static char *_get_self_cg_path() Link Here
254	* In containerized environments it will look like:	259	* In containerized environments it will look like:
255	* /sys/fs/cgroup[/docker.slice/docker-<some UUID>.scope]	260	* /sys/fs/cgroup[/docker.slice/docker-<some UUID>.scope]
256	*	261	*
		262	* In cgroup namespaces it will look like:
		263	* /sys/fs/cgroup/init.scope
		264	*
257	*/	265	*/
258	static char *_get_init_cg_path()	266	static char *_get_init_cg_path()
259	{	267	{
260	char buf, start = NULL, p, ret = NULL;	268	char buf, start = NULL, p, ret = NULL;
261	size_t sz;	269	char *relative_cg_path = NULL;
		270	size_t sz, relative_cg_len;
262		271
263	if (common_file_read_content("/proc/1/cgroup", &buf, &sz) !=	272	if (common_file_read_content("/proc/1/cgroup", &buf, &sz) !=
264	SLURM_SUCCESS)	273	SLURM_SUCCESS)
265	fatal("cannot read /proc/1/cgroup contents: %m");	274	fatal("cannot read /proc/1/cgroup contents: %m");
		275	debug4("Init cgroup data is %s.", buf);
266		276
267	/*	277	/*
268	* In Unified mode there will be just one line containing the path	278	* In Unified mode there will be just one line containing the path
Lines 282-288 static char *_get_init_cg_path() Link Here
282	if ((p = xstrchr(start, '\n')))	292	if ((p = xstrchr(start, '\n')))
283	*p = '\0';	293	*p = '\0';
284	p = xdirname(start);	294	p = xdirname(start);
285	if (!xstrcmp(p, "/"))	295	debug3("Found PID 1 cgroup path %s.", p);
		296	if (xstrstr(p, "..")) {
		297	debug3("Relative path to cgroup namespace root found; removing the relative path prefix.");
		298	relative_cg_path = _get_relative_cg_path("1");
		299	if (xstrstr(p, relative_cg_path)) {
		300	relative_cg_len = strlen(relative_cg_path);
		301	xstrfmtcat(ret, "%s%s",
		302	slurm_cgroup_conf.cgroup_mountpoint,
		303	p + relative_cg_len);
		304	debug2("Using %s as init cgroup path.", ret);
		305	} else {
		306	error("Did not find relative path to cgroup namespace root prefix %s in init cgroup path %s.",
		307	relative_cg_path, start);
		308	}
		309	xfree(relative_cg_path);
		310	} else if (!xstrcmp(p, "/"))
286	xstrfmtcat(ret, "%s",	311	xstrfmtcat(ret, "%s",
287	slurm_cgroup_conf.cgroup_mountpoint);	312	slurm_cgroup_conf.cgroup_mountpoint);
288	else	313	else
289	-
290	--
291	src/plugins/cgroup/v2/cgroup_v2.c \| 1 +	314	src/plugins/cgroup/v2/cgroup_v2.c \| 1 +
292	1 file changed, 1 insertion(+)	315	1 file changed, 1 insertion(+)




 *  cgroup_v2.c - Cgroup v2 plugin
 *****************************************************************************
 *  Copyright (C) 2021 SchedMD LLC
 *  Copyright (C) 2023 Urban Borstnik, ETH Zurich
 *  Written by Felip Moll <felip.moll@schedmd.com>
 *
 *  This file is part of Slurm, a resource management program.
- 
root.
--
src/plugins/cgroup/v2/cgroup_v2.c | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)

Lines 140-146 extern int cgroup_p_task_addto(cgroup_ctl_type_t ctl, stepd_step_rec_t *step, Link Here

(-)a/src/plugins/cgroup/v2/cgroup_v2.c (-2 / +1 lines)
140	static char _get_relative_cg_path(const char pid)	140	static char _get_relative_cg_path(const char pid)
141	{	141	{
142	char buf, p, ret = NULL, procfile=NULL, *mount=NULL;	142	char buf, p, ret = NULL, procfile=NULL, *mount=NULL;
143	size_t sz, name_len;	143	size_t sz;
144	int i;	144	int i;
145		145
146	xstrfmtcat(procfile, "/proc/%s/mountinfo", pid);	146	xstrfmtcat(procfile, "/proc/%s/mountinfo", pid);
147	-

Return to ticket 16713

Lines 122-127 extern int cgroup_p_task_addto(cgroup_ctl_type_t ctl, stepd_step_rec_t *step, Link Here

(-)a/src/plugins/cgroup/v2/cgroup_v2.c (-2 / +47 lines)
122	* (task_id = NO_VAL)	122	* (task_id = NO_VAL)
123	*/	123	*/
124		124
		125	/*
		126	* Get the relative path to the root of the cgroup namespace by
		127	* reading /proc/PID/mountinfo for the given PID.
		128	*
		129	* In the root cgroup namespace (most environments) we expect that the
		130	* cgroup2 filesystem is mounted on /, such as:
		131	* 34 25 0:29 / /sys/fs/cgroup rw,nosuid,nodev,noexec,relatime shared:9 - cgroup2 cgroup2 rw
		132	*
		133	* If the given PID is in a cgroup namespace, then it could look like:
		134	* 3526 3512 0:29 /../../../../cont1 /sys/fs/cgroup rw,nosuid,nodev,noexec,relatime - cgroup2 cgroup2 rw
		135	*
		136	* See man:cgroup_namespaces(7).
		137	*
		138	*/
		139	static char _get_relative_cg_path(const char pid)
		140	{
		141	char buf, p, ret = NULL, procfile=NULL, *mount=NULL;
		142	size_t sz, name_len;
		143	int i;
		144
		145	xstrfmtcat(procfile, "/proc/%s/mountinfo", pid);
		146	if (common_file_read_content(procfile, &buf, &sz) !=
		147	SLURM_SUCCESS)
		148	fatal("cannot read /proc/?/mountinfo contents: %m");
		149	debug4("%s contents: %s", procfile, buf);
		150
		151	if (!(p = xstrstr(buf, "cgroup2 cgroup2"))) {
		152	fatal("Did not find cgroup2 mount point in %s", procfile);
		153	}
		154
		155	*p = '\0';
		156	if ((mount = xstrrchr(buf, '\n')) == NULL)
		157	mount = buf;
		158	debug4("First 8 data of cgroup2 line in %s is %s.", procfile, mount);
		159
		160	p = mount;
		161	for (i=0; (i < 4) && (p != NULL); i++)
		162	p = strtok((i==0)?mount:NULL, " ");
		163	if (p == NULL)
		164	fatal("Could not parse cgroup2 mount point in %s.", procfile);
		165	ret = xstrdup(p);
		166	debug2("Relative path to root of cgroup namespace is %s.", ret);
		167	xfree(buf);
		168	xfree(procfile);
		169	return ret;
		170	}
		171
125	/*	172	/*
126	* The cgroup v2 documented way to know which is the process root in the cgroup	173	* The cgroup v2 documented way to know which is the process root in the cgroup
127	* hierarchy is just to read /proc/self/cgroup. In Unified hierarchies this	174	* hierarchy is just to read /proc/self/cgroup. In Unified hierarchies this
128	-
129	--
130	src/plugins/cgroup/v2/cgroup_v2.c \| 30 +++++++++++++++++++++++++++---	175	src/plugins/cgroup/v2/cgroup_v2.c \| 30 +++++++++++++++++++++++++++---
131	1 file changed, 27 insertions(+), 3 deletions(-)	176	1 file changed, 27 insertions(+), 3 deletions(-)

Lines 2-7 Link Here

(-)a/src/plugins/cgroup/v2/cgroup_v2.c (-2 / +1 lines)
2	* cgroup_v2.c - Cgroup v2 plugin	2	* cgroup_v2.c - Cgroup v2 plugin
3	*****************************************************************************	3	*****************************************************************************
4	* Copyright (C) 2021 SchedMD LLC	4	* Copyright (C) 2021 SchedMD LLC
		5	* Copyright (C) 2023 Urban Borstnik, ETH Zurich
5	* Written by Felip Moll <felip.moll@schedmd.com>	6	* Written by Felip Moll <felip.moll@schedmd.com>
6	*	7	*
7	* This file is part of Slurm, a resource management program.	8	* This file is part of Slurm, a resource management program.
8	-
9	root.	9	root.
10	--
11	src/plugins/cgroup/v2/cgroup_v2.c \| 2 +-	10	src/plugins/cgroup/v2/cgroup_v2.c \| 2 +-
12	1 file changed, 1 insertion(+), 1 deletion(-)	11	1 file changed, 1 insertion(+), 1 deletion(-)