path: root/api.c

/*
 * Copyright IBM Corporation. 2007
 *
 * Author:	Dhaval Giani <dhaval@linux.vnet.ibm.com>
 * Author:	Balbir Singh <balbir@linux.vnet.ibm.com>
 *
 * This program is free software; you can redistribute it and/or modify it
 * under the terms of version 2.1 of the GNU Lesser General Public License
 * as published by the Free Software Foundation.
 *
 * This program is distributed in the hope that it would be useful, but
 * WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
 *
 * TODOs:
 *	1. Convert comments to Docbook style.
 *	2. Add more APIs for the control groups.
 *	3. Handle the configuration related APIs.
 *	4. Error handling.
 *
 * Code initiated and designed by Dhaval Giani. All faults are most likely
 * his mistake.
 */

#include <errno.h>
#include <libcg.h>
#include <mntent.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/types.h>
#include <sys/stat.h>
#include <sys/syscall.h>
#include <unistd.h>
#include <fts.h>

/*
 * Remember to bump this up for major API changes.
 */
const static char cg_version[] = "0.01";

/*
 * Only one mount point is currently supported. This will be enhanced to
 * support several hierarchies in the future
 */
static char MOUNT_POINT[FILENAME_MAX];

static int cg_chown_file(FTS *fts, FTSENT *ent, uid_t owner, gid_t group)
{
	int ret = 1;
	const char *filename = fts->fts_path;
	dbg("seeing file %s\n", filename);
	switch (ent->fts_info) {
	case FTS_ERR:
		errno = ent->fts_errno;
		break;
	case FTS_D:
	case FTS_DC:
	case FTS_NSOK:
	case FTS_NS:
	case FTS_DNR:
	case FTS_DP:
	case FTS_F:
	case FTS_DEFAULT:
		ret = chown(filename, owner, group);
		break;
	}
	return ret;
}

/*
 * TODO: Need to decide a better place to put this function.
 */
static int cg_chown_recursive(char **path, uid_t owner, gid_t group)
{
	int ret = 1;
	dbg("path is %s\n", *path);
	FTS *fts = fts_open(path, FTS_PHYSICAL | FTS_NOCHDIR |
				FTS_NOSTAT, NULL);
	while (1) {
		FTSENT *ent;
		ent = fts_read(fts);
		if (!ent) {
			dbg("fts_read failed\n");
			break;
		}
		cg_chown_file(fts, ent, owner, group);
	}
	fts_close(fts);
	return ret;
}

/**
 * cg_init(), initializes the MOUNT_POINT.
 * This code is not currently thread safe (hint: getmntent is not thread safe).
 * This API is likely to change in the future to push state back to the caller
 * to achieve thread safety. The code currently supports just one mount point.
 * Complain if the cgroup filesystem controllers are bound to different mount
 * points.
 */
int cg_init()
{
	FILE *proc_mount;
	struct mntent *ent, *found_ent = NULL;
	int found_mnt = 0;
	int ret = 0;

	proc_mount = fopen("/proc/mounts", "r");

	while ((ent = getmntent(proc_mount)) != NULL) {
		if (!strncmp(ent->mnt_fsname,"cgroup", strlen("cgroup"))) {
			found_ent = ent;
			found_mnt++;
			dbg("Found cgroup option %s, count %d\n",
				found_ent->mnt_opts, found_mnt);
		}
	}

	/*
	 * Currently we require that all controllers be bound together
	 */
	if (!found_mnt)
		ret = ECGROUPNOTMOUNTED;
	if (found_mnt > 1)
		ret = ECGROUPMULTIMOUNTED;
	else {
		/*
		 * NOTE: FILENAME_MAX ensures that we don't need to worry
		 * about crossing MOUNT_POINT size. For the paranoid, yes
		 * this is a potential security hole - Balbir
		 * Dhaval - fix these things.
		 */
		strcpy(MOUNT_POINT, found_ent->mnt_dir);
		strcat(MOUNT_POINT, "/");
	}

	fclose(proc_mount);
	return ret;
}

static int cg_test_mounted_fs()
{
	FILE *proc_mount;
	struct mntent *ent;

	proc_mount = fopen("/proc/mounts", "r");
	if (proc_mount == NULL) {
		return -1;
	}
	ent = getmntent(proc_mount);

	while (strcmp(ent->mnt_fsname, "cgroup") !=0) {
		ent = getmntent(proc_mount);
		if (ent == NULL)
			return 0;
	}
	fclose(proc_mount);
	return 1;
}

static inline pid_t cg_gettid()
{
	return syscall(__NR_gettid);
}

static char* cg_build_path(char *name, char *path)
{
	strcpy(path, MOUNT_POINT);
	strcat(path, name);
	strcat(path, "/");
	return path;
}

/** cg_attach_task_pid is used to assign tasks to a cgroup.
 *  struct cgroup *cgroup: The cgroup to assign the thread to.
 *  pid_t tid: The thread to be assigned to the cgroup.
 *
 *  returns 0 on success.
 *  returns ECGROUPNOTOWNER if the caller does not have access to the cgroup.
 *  returns ECGROUPNOTALLOWED for other causes of failure.
 */
int cg_attach_task_pid(struct cgroup *cgroup, pid_t tid)
{
	char path[FILENAME_MAX];
	FILE *tasks;

	if (cgroup == NULL)
		strcpy(path, MOUNT_POINT);
	else {
		cg_build_path(cgroup->name, path);
	}

	strcat(path, "/tasks");

	tasks = fopen(path, "w");
	if (!tasks) {
		switch (errno) {
		case EPERM:
			return ECGROUPNOTOWNER;
		default:
			return ECGROUPNOTALLOWED;
		}
	}
	fprintf(tasks, "%d", tid);
	fclose(tasks);

	return 0;

}

/** cg_attach_task is used to attach the current thread to a cgroup.
 *  struct cgroup *cgroup: The cgroup to assign the current thread to.
 *
 *  See cg_attach_task_pid for return values.
 */
int cg_attach_task(struct cgroup *cgroup)
{
	pid_t tid = cg_gettid();
	int error;

	error = cg_attach_task_pid(cgroup, tid);

	return error;
}

/*
 * create_control_group()
 * This is the basic function used to create the control group. This function
 * just makes the group. It does not set any permissions, or any control values.
 * The argument path is the fully qualified path name to make it generic.
 */
static int cg_create_control_group(char *path)
{
	int error;
	if (!cg_test_mounted_fs())
		return ECGROUPNOTMOUNTED;
	error = mkdir(path, S_IRWXU | S_IRWXG | S_IROTH | S_IXOTH);
	if (error) {
		switch(errno) {
		case EPERM:
			return ECGROUPNOTOWNER;
		default:
			return ECGROUPNOTALLOWED;
		}
	}
	return error;
}

/*
 * set_control_value()
 * This is the low level function for putting in a value in a control file.
 * This function takes in the complete path and sets the value in val in that
 * file.
 *
 * TODO:
 * At this point I am not sure what all values the control file can take. So
 * I put in an int arg. But this has to be made much more robust.
 */
static int cg_set_control_value(char *path, char *val)
{
	int error;
	FILE *control_file;
	if (!cg_test_mounted_fs())
		return ECGROUPNOTMOUNTED;

	control_file = fopen(path, "a");

	if (!control_file) {
		if (errno == EPERM) {
			/*
			 * We need to set the correct error value, does the
			 * group exist but we don't have the subsystem
			 * mounted at that point, or is it that the group
			 * does not exist. So we check if the tasks file
			 * exist. Before that, we need to extract the path.
			 */
			int len = strlen(path);

			while (*(path+len) != '/')
				len--;
			*(path+len+1) = '\0';
			strcat(path, "tasks");
			control_file = fopen(path, "r");
			if (!control_file) {
				if (errno == ENOENT)
					return ECGROUPSUBSYSNOTMOUNTED;
			}
			fclose(control_file);
			return ECGROUPNOTALLOWED;
		}
		return errno;
	}

	fprintf(control_file, "%s", val);
	fclose(control_file);
	return 0;
}

/** cg_modify_cgroup modifies the cgroup control files.
 * struct cgroup *cgroup: The name will be the cgroup to be modified.
 * The values will be the values to be modified, those not mentioned
 * in the structure will not be modified.
 *
 * The uids cannot be modified yet.
 *
 * returns 0 on success.
 *
 */

int cg_modify_cgroup(struct cgroup *cgroup)
{
	char path[FILENAME_MAX], base[FILENAME_MAX];
	int i;
	int error;

	cg_build_path(cgroup->name, base);

	for (i = 0; i < CG_CONTROLLER_MAX && cgroup->controller[i];
						i++, strcpy(path, base)) {
		int j;
		for(j = 0; j < CG_NV_MAX &&
			cgroup->controller[i]->values[j];
			j++, strcpy(path, base)) {
			strcat(path, cgroup->controller[i]->values[j]->name);
			error = cg_set_control_value(path,
				cgroup->controller[i]->values[j]->value);
			if (error)
				goto err;
		}
	}
	return 0;
err:
	return error;

}

/** create_cgroup creates a new control group.
 * struct cgroup *cgroup: The control group to be created
 *
 * returns 0 on success. We recommend calling cg_delete_cgroup
 * if this routine fails. That should do the cleanup operation.
 */
int cg_create_cgroup(struct cgroup *cgroup)
{
	char *fts_path[2], base[FILENAME_MAX], *path;
	int i;
	int error;

	fts_path[0] = (char *)malloc(FILENAME_MAX);
	if (!fts_path[0])
		return ENOMEM;
	fts_path[1] = NULL;
	path = fts_path[0];

	cg_build_path(cgroup->name, path);
	error = cg_create_control_group(path);
	if (error)
		goto err;

	strcpy(base, path);

	cg_chown_recursive(fts_path, cgroup->control_uid, cgroup->control_gid);

	for (i = 0; i < CG_CONTROLLER_MAX && cgroup->controller[i];
						i++, strcpy(path, base)) {
		int j;
		for(j = 0; j < CG_NV_MAX && cgroup->controller[i]->values[j];
						j++, strcpy(path, base)) {
			strcat(path, cgroup->controller[i]->values[j]->name);
			error = cg_set_control_value(path,
				cgroup->controller[i]->values[j]->value);

			/*
 			 * Should we undo, what we've done in the loops above?
 			 */
			if (error)
				goto err;
		}
	}

	strcpy(path, base);
	strcat(path, "/tasks");
	chown(path, cgroup->tasks_uid, cgroup->tasks_gid);
err:
	free(path);
	return error;
}

/** cg_delete cgroup deletes a control group.
 *  struct cgroup *cgroup takes the group which is to be deleted.
 *
 *  returns 0 on success.
 */
int cg_delete_cgroup(struct cgroup *cgroup, int force)
{
	FILE *delete_tasks, *base_tasks;
	int tids;
	char path[FILENAME_MAX];
	int error = ECGROUPNOTALLOWED;

	strcpy(path, MOUNT_POINT);
	strcat(path,"/tasks");

	base_tasks = fopen(path, "w");
	if (!base_tasks)
		goto base_open_err;

	cg_build_path(cgroup->name, path);
	strcat(path,"/tasks");

	delete_tasks = fopen(path, "r");
	if (!delete_tasks)
		goto del_open_err;

	while (!feof(delete_tasks)) {
		fscanf(delete_tasks, "%d", &tids);
		fprintf(base_tasks, "%d", tids);
	}

	cg_build_path(cgroup->name, path);
	error = rmdir(path);

	fclose(delete_tasks);
del_open_err:
	fclose(base_tasks);
base_open_err:
	if (force) {
		cg_build_path(cgroup->name, path);
		error = rmdir(path);
	}
	return error;
}