src/lib/cgroup.c

#include <sys/types.h>
#include <stdint.h>
#include <stdio.h>
#include <stdlib.h>
#include <errno.h>
#include <string.h>
#include <strings.h>
#include <signal.h>
#include <unistd.h>
#include <sys/stat.h>
#include <fcntl.h>

#include "types.h"
#include "vzerror.h"
#include "cgroup.h"
#include "cpu.h"
#include "bitmap.h"
#include "logger.h"
#include "util.h"
#include "env.h"

#include <assert.h>

#define MEMLIMIT	"memory.limit_in_bytes"
#define SWAPLIMIT	"memory.memsw.limit_in_bytes"
#define KMEMLIMIT	"memory.kmem.limit_in_bytes"
#define TCPLIMIT	"memory.kmem.tcp.limit_in_bytes"

static int copy_string_from_parent(struct cgroup_controller *controller,
				   struct cgroup_controller *pcont, const char *file)
{
	char *ptr = NULL;
	int ret;

	ret = cgroup_get_value_string(pcont, file, &ptr);
	if (ret)
		goto out;
	ret = cgroup_set_value_string(controller, file, ptr);
out:
	free(ptr);
	return ret;
}

static int controller_apply_config(struct cgroup *ct, struct cgroup *parent,
				   struct cgroup_controller *controller,
				   const char *name)
{
	int ret;
	if (!strcmp("cpuset", name)) {
		struct cgroup_controller *pcont = cgroup_get_controller(parent, name);
		if (!pcont)
			return 0;

		if ((ret = copy_string_from_parent(controller, pcont, "cpuset.cpus")))
			return ret;

		if ((ret = copy_string_from_parent(controller, pcont, "cpuset.mems")))
			return ret;
	} else if (!strcmp("memory", name)) {
		if ((ret = cgroup_set_value_string(controller, "memory.use_hierarchy", "1")))
			return ret;
#if 0		/* BUG: this code assumes page size to be 4096 byte. */
		/*
		 * The kernel memory controller cannot flip states from
		 * unlimited to limited if there are already tasks in it.
		 * Therefore, we always have to run with *some* value of kmem
		 * enabled. If we don't do it, we can't start unlimited and
		 * then use the set command to set any beancounters. We write
		 * The maximum amount minus two pages, which should effectively
		 * mean "accounting turned on, but unlimited". This will fail
		 * if the kmem controller is not present, but that is okay.
		 */
		cgroup_set_value_string(controller,
				"memory.kmem.limit_in_bytes",
				"9223372036854767712");
#endif
	} else if (!strcmp("devices", name)) {
		if ((ret = cgroup_set_value_string(controller, "devices.deny", "a")))
			return ret;
	}
	return 0;
}

static char *conf_names[] = {
	"Memory",
	"Kernel Memory",
	"Swap",
	"TCPbuffer",
	"CPU limits",
	"CPU mask",
	"CPU shares",
	"Allowed Devices",
	"Denied Devices",
};

int container_apply_config(envid_t veid, enum conf_files c, void *_val)
{
	struct cgroup *ct;
	char cgrp[CT_MAX_STR_SIZE];
	struct cgroup_controller *mem, *cpu, *cpuset;
	int ret = -EINVAL;
	unsigned long *val = _val;

	logger(2, 0, "function: container_apply_config(%u, %u, %p)", veid, c, _val);

	veid_to_name(cgrp, veid);

	ct = cgroup_new_cgroup(cgrp);
	/*
	 * We should really be doing some thing like:
	 *
	 *	ret = cgroup_get_cgroup(ct);
	 *
	 * and then doing cgroup_get_controller. However, libcgroup has
	 * a very nasty bug that make it sometimes fail. adding a controller
	 * to a newly "created" cgroup structure and then setting the value
	 * is a workaround that seems to work on various versions of the
	 * library
	 */
	switch (c) {
	case MEMORY:
		if ((mem = cgroup_add_controller(ct, "memory")))
			ret = cgroup_set_value_uint64(mem, MEMLIMIT, *val);
		break;
	case SWAP:
		/* Unlike kmem, this must always be greater than mem */
		if ((mem = cgroup_add_controller(ct, "memory"))) {
			uint64_t mval;
			if (!cgroup_get_value_uint64(mem, MEMLIMIT, &mval))
				ret = cgroup_set_value_uint64(mem, SWAPLIMIT,
							      mval + *val);
		}
		break;
	case KMEMORY:
		if ((mem = cgroup_add_controller(ct, "memory")))
			ret = cgroup_set_value_uint64(mem, KMEMLIMIT, *val);
		break;
	case TCP:
		if ((mem = cgroup_add_controller(ct, "memory")))
			ret = cgroup_set_value_uint64(mem, TCPLIMIT, *val);
		break;
	case CPULIMIT: {
		uint64_t period;
		uint64_t quota;
		if ((cpu = cgroup_add_controller(ct, "cpu")) == NULL)
			break;

		/* Should be 100000, but be safe. It may fail on some versions
		 * of libcgroup, so if it fails, just assume the default */
		ret = cgroup_get_value_uint64(cpu, "cpu.cfs_period_us", &period);
		if (ret)
			period = 100000;
		/* val will contain an integer percentage, like 223% */
		quota = (period * (*val)) / 100;
		ret = cgroup_set_value_uint64(cpu, "cpu.cfs_quota_us", quota);
		break;
	}
	case CPUSHARES:
		if ((cpu = cgroup_add_controller(ct, "cpu")) == NULL)
			break;
		ret = cgroup_set_value_uint64(cpu, "cpu.shares", *val);
		break;
	case CPUMASK: {
		struct cgroup_controller *pcont;
		struct cgroup *parent;
		char *ptr = NULL;
		char cpusetstr[2 * CPUMASK_NBITS];
		unsigned int i;

		if ((cpuset = cgroup_add_controller(ct, "cpuset")) == NULL)
			break;
		/*
		 * Having all bits set is a bit different, bitmap_snprintf will
		 * return a bad string. (From the PoV of the cpuset cgroup). We
		 * actually need to copy the parent's mask in that case.
		 */
		for (i = 0; i < CPUMASK_NBYTES; i++) {
			if (val[i] != (~0UL)) {
				bitmap_snprintf(cpusetstr, CPUMASK_NBITS * 2,
						val, CPUMASK_NBITS);
				goto string_ok;
			}
		}

		parent = cgroup_new_cgroup(CT_BASE_STRING);
		cgroup_get_cgroup(parent);
		pcont = cgroup_get_controller(parent, "cpuset");
		ret = cgroup_get_value_string(pcont, "cpuset.cpus", &ptr);
		if (ptr) {
			strncpy(cpusetstr, ptr, CPUMASK_NBITS *2);
			free(ptr);
		}
		cgroup_free(&parent);
string_ok:
		ret = cgroup_set_value_string(cpuset, "cpuset.cpus", cpusetstr);
		break;
	}
	case DEVICES_DENY: {
		struct cgroup_controller *dev;

		if ((dev = cgroup_add_controller(ct, "devices")) == NULL)
			break;

		ret = cgroup_set_value_string(dev, "devices.deny", (char *)_val);
		break;
	}
	case DEVICES_ALLOW: {
		struct cgroup_controller *dev;

		if ((dev = cgroup_add_controller(ct, "devices")) == NULL)
			break;

		ret = cgroup_set_value_string(dev, "devices.allow", (char *)_val);
		break;
	}
	default:
		ret = -EINVAL;
		break;
	}

	if (ret)
		goto out;

	if ((ret = cgroup_modify_cgroup(ct)))
		logger(-1, 0, "Failed to set limits for %s (%s)", conf_names[c],
		       cgroup_strerror(ret));
out:
	cgroup_free(&ct);
	return ret;
}

static int do_create_container(struct cgroup *ct, struct cgroup *parent)
{
	struct cgroup_mount_point mnt;
	struct cgroup_controller *controller;
	void *handle;
	int ret;

	ret = cgroup_get_controller_begin(&handle, &mnt);

	cgroup_get_cgroup(parent);

	do {
		logger(2, 0, "Adding group to controller %s", mnt.name);
		controller = cgroup_add_controller(ct, mnt.name);
		ret = controller_apply_config(ct, parent, controller, mnt.name);
		if (!ret)
			ret = cgroup_get_controller_next(&handle, &mnt);
	} while (!ret);

	cgroup_get_controller_end(&handle);

	if (ret == ECGEOF)
		ret = cgroup_create_cgroup(ct, 0);

	return ret;

}

int create_container(envid_t veid)
{
	char cgrp[CT_MAX_STR_SIZE];
	struct cgroup *ct, *parent;
	int ret;
	unsigned int i;
	const char *devices[] = { "c *:* m", /* everyone can mknod */
				  "b *:* m", /* block devices too */
				  "c 1:3 rmw", /* null */
				  "c 1:5 rmw", /* zero */
				  "c 1:7 rmw", /* full */
				  "c 1:8 rmw", /* random */
				  "c 1:9 rmw", /* urandom */
				  "c 5:2 rmw", /* ptmx */
				  "c 136:* rmw", /* various pts */
				};

	veid_to_name(cgrp, veid);
	ct = cgroup_new_cgroup(cgrp);
	parent = cgroup_new_cgroup("/");

	ret = do_create_container(ct, parent);
	cgroup_free(&ct);
	cgroup_free(&parent);


	/*
	 * FIXME: This is yet another hack required by libcgroup. At some point
	 * in time, this MUST go away.
	 *
	 * Problem is that libcgroup works with buffered writes. If we write to
	 * a cgroup file and want it to be seen in the filesystem, we need to
	 * call cgroup_modify_cgroup().
	 *
	 * However, all versions up to 0.38 will fail that operation for already
	 * existent cgroups, due to a bug in the way they handle modifications
	 * in the presence of read-only files (whether or not that specific file
	 * was being modified). Because of that, we need to come up with a new
	 * cgroup all the time, and free it afterwards.
	 */
	for (i = 0; i < ARRAY_SIZE(devices); i++) {
		struct cgroup_controller *dev;

		veid_to_name(cgrp, veid);
		ct = cgroup_new_cgroup(cgrp);

		if ((dev = cgroup_add_controller(ct, "devices"))) {
			cgroup_set_value_string(dev, "devices.allow", devices[i]);
			if ((ret = cgroup_modify_cgroup(ct))) {
				logger(-1, 0, "Failed to set device permissions for %s (%s)",
					devices[i], cgroup_strerror(ret));
			}
		} else {
			logger(-1, 0, "Failed to attach device controller (%s)",
			       cgroup_strerror(ret));
		}
		cgroup_free(&ct);
	}

	return ret;
}

/* libcgroup is lame. This should be done with the cgroup structure, not the
 * cgroup name
 */
static int controller_has_tasks(const char *cgrp, const char *name)
{
	int ret;
	pid_t pid;
	void *handle;

	ret = cgroup_get_task_begin(cgrp, name, &handle, &pid);
	ret = (ret != ECGEOF);
	cgroup_get_task_end(&handle);
	return ret;
}

int container_add_task(envid_t veid)
{
	char cgrp[CT_MAX_STR_SIZE];
	struct cgroup *ct;
	int ret;

	veid_to_name(cgrp, veid);
	ct = cgroup_new_cgroup(cgrp);
	ret = cgroup_get_cgroup(ct);
	if (ret)
		goto out;

	ret = cgroup_attach_task_pid(ct, getpid());
out:
	cgroup_free(&ct);
	return ret;
}

int destroy_container(envid_t veid)
{
	struct cgroup *ct;
	char cgrp[CT_MAX_STR_SIZE];
	int ret;

	veid_to_name(cgrp, veid);
	ct = cgroup_new_cgroup(cgrp);
	ret = cgroup_get_cgroup(ct);

	/* Since this can also be called from initialization, this is valid */
	if (ret == ECGROUPNOTEXIST) {
		ret = 0;
		goto out;
	}

	ret = cgroup_delete_cgroup_ext(ct, 0);
out:
	cgroup_free(&ct);
	return ret;
}

int container_is_running(envid_t veid)
{
	int ret = 0;
	void *handle;
	struct cgroup_mount_point mnt;
	struct cgroup *ct;
	char cgrp[CT_MAX_STR_SIZE];

	veid_to_name(cgrp, veid);

	ct = cgroup_new_cgroup(cgrp);
	ret = cgroup_get_cgroup(ct);
	if (ret == ECGROUPNOTEXIST) {
		ret = 0;
		goto out_free;
	}

	ret = cgroup_get_controller_begin(&handle, &mnt);
	do {
		struct cgroup_controller *controller;
		controller = cgroup_get_controller(ct, mnt.name);
		if (!controller) {
			logger(0, 0, "Controller %s seems to be missing!", mnt.name);
			continue;
		}
		if ((ret = controller_has_tasks(cgrp, mnt.name)) != 0)
			goto out;
	} while ((ret = cgroup_get_controller_next(&handle, &mnt)) == 0);

	if (ret != ECGEOF)
		ret = -ret;
	else
		ret = 0;
out:
	cgroup_get_controller_end(&handle);
out_free:
	cgroup_free(&ct);
	return ret;
}

/*
 * We send a kill signal to all processes. This is racy in theory, since they
 * could spawn new processes faster than we kill. But since one of them is the
 * init process, (we don't really know which), then eventually the init process
 * will die taking away all the others, so this is fine.
 *
 * This is a big hack, and only exists because we have no way to enter a PID
 * namespace from the outside (yet). From there, we could just issue a normal
 * reboot.
 */
int hackish_empty_container(envid_t veid)
{
	char cgrp[CT_MAX_STR_SIZE];
	struct cgroup *ct;
	int ret = 0;
	void *task_handle;
	pid_t pid;
	int i;

	veid_to_name(cgrp, veid);
	ct = cgroup_new_cgroup(cgrp);

	ret = cgroup_get_cgroup(ct);
	if (ret == ECGROUPNOTEXIST) {
		ret = 0;
		goto out;
	}

	/* Any controller will do */
	ret = cgroup_get_task_begin(cgrp, "cpu", &task_handle, &pid);
	while (!ret) {
		kill(pid, SIGKILL);
		ret = cgroup_get_task_next(&task_handle, &pid);
	}
	cgroup_get_task_end(&task_handle);

	if (ret != ECGEOF) {
		logger(-1, 0, "Could not finish all tasks: %s",
				cgroup_strerror(ret));
		goto out;
	}

	ret = 0;
	for (i = 0; i < DEF_STOP_TIMEOUT; i++) {
		if (!container_is_running(veid))
			goto out;
		usleep(500000);
	}
	logger(-1, 0, "Failed to wait for CT tasks to die");
	ret = VZ_STOP_ERROR;
out:
	cgroup_free(&ct);
	return ret;
}

static int get_process_name(pid_t pid, char *buffer, size_t len) {
	char path[32];
	int path_len = snprintf(path, sizeof path, "/proc/%d/stat", pid);
	assert(path_len > 0 && (size_t)path_len < sizeof path);
	FILE *f = fopen(path, "r");
	if(!f) return -1;
	int c, i = -1;
	while((c = fgetc(f)) != EOF) {
		if(i < 0) {
			if(c == '(') i = 0;
		} else if(c == ')') break;
		else {
			buffer[i] = c;
			if(++i >= (int)len - 1) break;
		}
	}
	if(i < 1) return -1;
	buffer[i] = 0;
	return 0;
}

/*
 * This function assumes that all pids inside a cgroup
 * belong to the same namespace, that is the container namespace.
 * Therefore, from the host box, any of them will do.
 */
pid_t get_pid_from_container(envid_t veid)
{
	char cgrp[CT_MAX_STR_SIZE];
	struct cgroup *ct;
	void *task_handle;
	void *cont_handle;
	struct cgroup_mount_point mnt;
	pid_t pid = -1;
	int ret;

	veid_to_name(cgrp, veid);
	ct = cgroup_new_cgroup(cgrp);
	ret = cgroup_get_cgroup(ct);
	if (ret == ECGROUPNOTEXIST)
		goto out_free;

	ret = cgroup_get_controller_begin(&cont_handle, &mnt);
	if (ret != 0) /* no controllers, something is wrong */
		goto out_free;

	char vzctl_program_name[16];
	if(get_process_name(getpid(), vzctl_program_name, sizeof vzctl_program_name) < 0) {
		*vzctl_program_name = 0;
	}
	ret = cgroup_get_task_begin(cgrp, mnt.name, &task_handle, &pid);
	while(ret == 0) {
		if(!*vzctl_program_name) goto out_end_cont;
		char p_name[16];
		if(get_process_name(pid, p_name, sizeof p_name) < 0) goto out_end_cont;
		if(strcmp(p_name, vzctl_program_name)) goto out_end_cont;
		ret = cgroup_get_task_next(&task_handle, &pid);
	}
	pid = -1;

out_end_cont:
	cgroup_get_task_end(&task_handle);
	cgroup_get_controller_end(&cont_handle);
out_free:
	cgroup_free(&ct);
	return pid;
}
int container_init(void)
{
	return cgroup_init();
}