libmount/src/lock.c

/*
 * Copyright (C) 2009 Karel Zak <kzak@redhat.com>
 *
 * This file may be redistributed under the terms of the
 * GNU Lesser General Public License.
 */

/**
 * SECTION: lock
 * @title: Locking
 * @short_description: locking methods for /etc/mtab or another libmount files
 *
 * The mtab lock is backwards compatible with the standard linux /etc/mtab
 * locking.  Note, it's necessary to use the same locking schema in all
 * applications that access the file.
 */
#include <sys/time.h>
#include <time.h>
#include <signal.h>
#include <fcntl.h>
#include <limits.h>
#include <sys/file.h>

#include "strutils.h"
#include "closestream.h"
#include "pathnames.h"
#include "mountP.h"
#include "monotonic.h"

/*
 * lock handler
 */
struct libmnt_lock {
	char	*lockfile;	/* path to lock file (e.g. /etc/mtab~) */
	char	*linkfile;	/* path to link file (e.g. /etc/mtab~.<id>) */
	int	lockfile_fd;	/* lock file descriptor */

	unsigned int	locked :1,	/* do we own the lock? */
			sigblock :1,	/* block signals when locked */
			simplelock :1;	/* use flock rather than normal mtab lock */

	sigset_t oldsigmask;
};


/**
 * mnt_new_lock:
 * @datafile: the file that should be covered by the lock
 * @id: unique linkfile identifier or 0 (default is getpid())
 *
 * Returns: newly allocated lock handler or NULL on case of error.
 */
struct libmnt_lock *mnt_new_lock(const char *datafile, pid_t id)
{
	struct libmnt_lock *ml = NULL;
	char *lo = NULL, *ln = NULL;
	size_t losz;

	if (!datafile)
		return NULL;

	/* for flock we use "foo.lock, for mtab "foo~"
	 */
	losz = strlen(datafile) + sizeof(".lock");
	lo = malloc(losz);
	if (!lo)
		goto err;

	/* default is mtab~ lock */
	snprintf(lo, losz, "%s~", datafile);

	if (asprintf(&ln, "%s~.%d", datafile, id ? : getpid()) == -1) {
		ln = NULL;
		goto err;
	}
	ml = calloc(1, sizeof(*ml) );
	if (!ml)
		goto err;

	ml->lockfile_fd = -1;
	ml->linkfile = ln;
	ml->lockfile = lo;

	DBG(LOCKS, ul_debugobj(ml, "alloc: default linkfile=%s, lockfile=%s", ln, lo));
	return ml;
err:
	free(lo);
	free(ln);
	free(ml);
	return NULL;
}


/**
 * mnt_free_lock:
 * @ml: struct libmnt_lock handler
 *
 * Deallocates mnt_lock.
 */
void mnt_free_lock(struct libmnt_lock *ml)
{
	if (!ml)
		return;
	DBG(LOCKS, ul_debugobj(ml, "free%s", ml->locked ? " !!! LOCKED !!!" : ""));
	free(ml->lockfile);
	free(ml->linkfile);
	free(ml);
}

/**
 * mnt_lock_block_signals:
 * @ml: struct libmnt_lock handler
 * @enable: TRUE/FALSE
 *
 * Block/unblock signals when the lock is locked, the signals are not blocked
 * by default.
 *
 * Returns: <0 on error, 0 on success.
 */
int mnt_lock_block_signals(struct libmnt_lock *ml, int enable)
{
	if (!ml)
		return -EINVAL;
	DBG(LOCKS, ul_debugobj(ml, "signals: %s", enable ? "BLOCKED" : "UNBLOCKED"));
	ml->sigblock = enable ? 1 : 0;
	return 0;
}

/* don't export this to API
 */
int mnt_lock_use_simplelock(struct libmnt_lock *ml, int enable)
{
	size_t sz;

	if (!ml)
		return -EINVAL;

	assert(ml->lockfile);

	DBG(LOCKS, ul_debugobj(ml, "flock: %s", enable ? "ENABLED" : "DISABLED"));
	ml->simplelock = enable ? 1 : 0;

	sz = strlen(ml->lockfile);
	assert(sz);

	if (sz < 1)
		return -EINVAL;

	/* Change lock name:
	 *
	 *	flock:     "<name>.lock"
	 *	mtab lock: "<name>~"
	 */
	if (ml->simplelock && endswith(ml->lockfile, "~"))
		memcpy(ml->lockfile + sz - 1, ".lock", 6);

	else if (!ml->simplelock && endswith(ml->lockfile, ".lock"))
		 memcpy(ml->lockfile + sz - 5, "~", 2);

	DBG(LOCKS, ul_debugobj(ml, "new lock filename: '%s'", ml->lockfile));
	return 0;
}

/*
 * Returns path to lockfile.
 */
static const char *mnt_lock_get_lockfile(struct libmnt_lock *ml)
{
	return ml ? ml->lockfile : NULL;
}

/*
 * Note that the filename is generated by mnt_new_lock() and depends on
 * getpid() or 'id' argument of the mnt_new_lock() function.
 *
 * Returns: unique (per process/thread) path to linkfile.
 */
static const char *mnt_lock_get_linkfile(struct libmnt_lock *ml)
{
	return ml ? ml->linkfile : NULL;
}

/*
 * Simple flocking
 */
static void unlock_simplelock(struct libmnt_lock *ml)
{
	assert(ml);
	assert(ml->simplelock);

	if (ml->lockfile_fd >= 0) {
		DBG(LOCKS, ul_debugobj(ml, "%s: unflocking",
					mnt_lock_get_lockfile(ml)));
		close(ml->lockfile_fd);
	}
}

static int lock_simplelock(struct libmnt_lock *ml)
{
	const char *lfile;
	int rc;

	assert(ml);
	assert(ml->simplelock);

	lfile = mnt_lock_get_lockfile(ml);

	DBG(LOCKS, ul_debugobj(ml, "%s: locking", lfile));

	if (ml->sigblock) {
		sigset_t sigs;
		sigemptyset(&ml->oldsigmask);
		sigfillset(&sigs);
		sigprocmask(SIG_BLOCK, &sigs, &ml->oldsigmask);
	}

	ml->lockfile_fd = open(lfile, O_RDONLY|O_CREAT|O_CLOEXEC,
				      S_IWUSR|S_IRUSR|S_IRGRP|S_IROTH);
	if (ml->lockfile_fd < 0) {
		rc = -errno;
		goto err;
	}
	rc = fchmod(ml->lockfile_fd, S_IRUSR|S_IWUSR|S_IRGRP|S_IROTH);
	if (rc < 0) {
		rc = -errno;
		goto err;
	}

	while (flock(ml->lockfile_fd, LOCK_EX) < 0) {
		int errsv;
		if ((errno == EAGAIN) || (errno == EINTR))
			continue;
		errsv = errno;
		close(ml->lockfile_fd);
		ml->lockfile_fd = -1;
		rc = -errsv;
		goto err;
	}
	ml->locked = 1;
	return 0;
err:
	if (ml->sigblock)
		sigprocmask(SIG_SETMASK, &ml->oldsigmask, NULL);
	return rc;
}

/*
 * traditional mtab locking
 */

static void mnt_lockalrm_handler(int sig __attribute__((__unused__)))
{
	/* do nothing, say nothing, be nothing */
}

/*
 * Waits for F_SETLKW, unfortunately we have to use SIGALRM here to interrupt
 * fcntl() to avoid neverending waiting.
 *
 * Returns: 0 on success, 1 on timeout, -errno on error.
 */
static int mnt_wait_mtab_lock(struct libmnt_lock *ml, struct flock *fl, time_t maxtime)
{
	struct timeval now;
	struct sigaction sa, osa;
	int ret = 0;

	gettime_monotonic(&now);
	DBG(LOCKS, ul_debugobj(ml, "(%d) waiting for F_SETLKW (now=%lu, maxtime=%lu, diff=%lu)",
				getpid(),
				(unsigned long) now.tv_sec,
				(unsigned long) maxtime,
				(unsigned long) (maxtime - now.tv_sec)));

	if (now.tv_sec >= maxtime)
		return 1;		/* timeout */

	/* setup ALARM handler -- we don't want to wait forever */
	sa.sa_flags = 0;
	sa.sa_handler = mnt_lockalrm_handler;
	sigfillset (&sa.sa_mask);

	sigaction(SIGALRM, &sa, &osa);


	alarm(maxtime - now.tv_sec);
	if (fcntl(ml->lockfile_fd, F_SETLKW, fl) == -1)
		ret = errno == EINTR ? 1 : -errno;
	alarm(0);

	/* restore old sigaction */
	sigaction(SIGALRM, &osa, NULL);

	DBG(LOCKS, ul_debugobj(ml, "(%d) leaving mnt_wait_setlkw(), rc=%d",
				getpid(), ret));
	return ret;
}

/*
 * Create the mtab lock file.
 *
 * The old code here used flock on a lock file /etc/mtab~ and deleted
 * this lock file afterwards. However, as rgooch remarks, that has a
 * race: a second mount may be waiting on the lock and proceed as
 * soon as the lock file is deleted by the first mount, and immediately
 * afterwards a third mount comes, creates a new /etc/mtab~, applies
 * flock to that, and also proceeds, so that the second and third mount
 * are now both scribbling in /etc/mtab.
 *
 * The new code uses a link() instead of a creat(), where we proceed
 * only if it was us that created the lock, and hence we always have
 * to delete the lock afterwards. Now the use of flock() is in principle
 * superfluous, but avoids an arbitrary sleep().
 *
 * Where does the link point to? Obvious choices are mtab and mtab~~.
 * HJLu points out that the latter leads to races. Right now we use
 * mtab~.pid instead.
 *
 *
 * The original mount locking code has used sleep(1) between attempts and
 * maximal number of attempts has been 5.
 *
 * There was a very small number of attempts and extremely long waiting (1s)
 * that is useless on machines with large number of mount processes.
 *
 * Now we wait for a few thousand microseconds between attempts and we have a global
 * time limit (30s) rather than a limit for the number of attempts. The advantage
 * is that this method also counts time which we spend in fcntl(F_SETLKW) and
 * the number of attempts is not restricted.
 * -- kzak@redhat.com [Mar-2007]
 *
 *
 * This mtab locking code has been refactored and moved to libmount. The mtab
 * locking is really not perfect (e.g. SIGALRM), but it's stable, reliable and
 * backwardly compatible code.
 *
 * Don't forget that this code has to be compatible with 3rd party mounts
 * (/sbin/mount.foo) and has to work with NFS.
 * -- kzak@redhat.com [May-2009]
 */

/* maximum seconds between the first and the last attempt */
#define MOUNTLOCK_MAXTIME		30

/* sleep time (in microseconds, max=999999) between attempts */
#define MOUNTLOCK_WAITTIME		5000

static void unlock_mtab(struct libmnt_lock *ml)
{
	if (!ml)
		return;

	if (!ml->locked && ml->lockfile && ml->linkfile)
	{
		/* We (probably) have all the files, but we don't own the lock,
		 * Really? Check it! Maybe ml->locked wasn't set properly
		 * because the code was interrupted by a signal. Paranoia? Yes.
		 *
		 * We own the lock when linkfile == lockfile.
		 */
		struct stat lo, li;

		if (!stat(ml->lockfile, &lo) && !stat(ml->linkfile, &li) &&
		    lo.st_dev == li.st_dev && lo.st_ino == li.st_ino)
			ml->locked = 1;
	}

	if (ml->linkfile)
		unlink(ml->linkfile);
	if (ml->lockfile_fd >= 0)
		close(ml->lockfile_fd);
	if (ml->locked && ml->lockfile) {
		unlink(ml->lockfile);
		DBG(LOCKS, ul_debugobj(ml, "unlink %s", ml->lockfile));
	}
}

static int lock_mtab(struct libmnt_lock *ml)
{
	int i, rc = -1;
	struct timespec waittime;
	struct timeval maxtime;
	const char *lockfile, *linkfile;

	if (!ml)
		return -EINVAL;
	if (ml->locked)
		return 0;

	lockfile = mnt_lock_get_lockfile(ml);
	if (!lockfile)
		return -EINVAL;
	linkfile = mnt_lock_get_linkfile(ml);
	if (!linkfile)
		return -EINVAL;

	if (ml->sigblock) {
		/*
		 * Block all signals when locked, mnt_unlock_file() will
		 * restore the old mask.
		 */
		sigset_t sigs;

		sigemptyset(&ml->oldsigmask);
		sigfillset(&sigs);
		sigdelset(&sigs, SIGTRAP);
		sigdelset(&sigs, SIGALRM);
		sigprocmask(SIG_BLOCK, &sigs, &ml->oldsigmask);
	}

	i = open(linkfile, O_WRONLY|O_CREAT|O_CLOEXEC, S_IRUSR|S_IWUSR);
	if (i < 0) {
		/* linkfile does not exist (as a file) and we cannot create it.
		 * Read-only or full filesystem? Too many files open in the system?
		 */
		if (errno > 0)
			rc = -errno;
		goto failed;
	}
	close(i);

	gettime_monotonic(&maxtime);
	maxtime.tv_sec += MOUNTLOCK_MAXTIME;

	waittime.tv_sec = 0;
	waittime.tv_nsec = (1000 * MOUNTLOCK_WAITTIME);

	/* Repeat until it was us who made the link */
	while (!ml->locked) {
		struct timeval now;
		struct flock flock;
		int j;

		j = link(linkfile, lockfile);
		if (j == 0)
			ml->locked = 1;

		if (j < 0 && errno != EEXIST) {
			if (errno > 0)
				rc = -errno;
			goto failed;
		}
		ml->lockfile_fd = open(lockfile, O_WRONLY|O_CLOEXEC);

		if (ml->lockfile_fd < 0) {
			/* Strange... Maybe the file was just deleted? */
			int errsv = errno;
			gettime_monotonic(&now);
			if (errsv == ENOENT && now.tv_sec < maxtime.tv_sec) {
				ml->locked = 0;
				continue;
			}
			if (errsv > 0)
				rc = -errsv;
			goto failed;
		}

		flock.l_type = F_WRLCK;
		flock.l_whence = SEEK_SET;
		flock.l_start = 0;
		flock.l_len = 0;

		if (ml->locked) {
			/* We made the link. Now claim the lock. */
			if (fcntl (ml->lockfile_fd, F_SETLK, &flock) == -1) {
				DBG(LOCKS, ul_debugobj(ml,
					"%s: can't F_SETLK lockfile, errno=%d\n",
					lockfile, errno));
				/* proceed, since it was us who created the lockfile anyway */
			}
			break;
		} else {
			/* Someone else made the link. Wait. */
			int err = mnt_wait_mtab_lock(ml, &flock, maxtime.tv_sec);

			if (err == 1) {
				DBG(LOCKS, ul_debugobj(ml,
					"%s: can't create link: time out (perhaps "
					"there is a stale lock file?)", lockfile));
				rc = -ETIMEDOUT;
				goto failed;

			} else if (err < 0) {
				rc = err;
				goto failed;
			}
			nanosleep(&waittime, NULL);
			close(ml->lockfile_fd);
			ml->lockfile_fd = -1;
		}
	}
	DBG(LOCKS, ul_debugobj(ml, "%s: (%d) successfully locked",
					lockfile, getpid()));
	unlink(linkfile);
	return 0;

failed:
	mnt_unlock_file(ml);
	return rc;
}


/**
 * mnt_lock_file
 * @ml: pointer to struct libmnt_lock instance
 *
 * Creates a lock file (e.g. /etc/mtab~). Note that this function may
 * use alarm().
 *
 * Your application always has to call mnt_unlock_file() before exit.
 *
 * Traditional mtab locking scheme:
 *
 *   1. create linkfile (e.g. /etc/mtab~.$PID)
 *   2. link linkfile --> lockfile (e.g. /etc/mtab~.$PID --> /etc/mtab~)
 *   3. a) link() success: setups F_SETLK lock (see fcntl(2))
 *      b) link() failed:  wait (max 30s) on F_SETLKW lock, goto 2.
 *
 * Note that when the lock is used by mnt_update_table() interface then libmount
 * uses flock() for private library file /run/mount/utab. The fcntl(2) is used only
 * for backwardly compatible stuff like /etc/mtab.
 *
 * Returns: 0 on success or negative number in case of error (-ETIMEOUT is case
 * of stale lock file).
 */
int mnt_lock_file(struct libmnt_lock *ml)
{
	if (!ml)
		return -EINVAL;

	if (ml->simplelock)
		return lock_simplelock(ml);

	return lock_mtab(ml);
}

/**
 * mnt_unlock_file:
 * @ml: lock struct
 *
 * Unlocks the file. The function could be called independently of the
 * lock status (for example from exit(3)).
 */
void mnt_unlock_file(struct libmnt_lock *ml)
{
	if (!ml)
		return;

	DBG(LOCKS, ul_debugobj(ml, "(%d) %s", getpid(),
			ml->locked ? "unlocking" : "cleaning"));

	if (ml->simplelock)
		unlock_simplelock(ml);
	else
		unlock_mtab(ml);

	ml->locked = 0;
	ml->lockfile_fd = -1;

	if (ml->sigblock) {
		DBG(LOCKS, ul_debugobj(ml, "restoring sigmask"));
		sigprocmask(SIG_SETMASK, &ml->oldsigmask, NULL);
	}
}

#ifdef TEST_PROGRAM

static struct libmnt_lock *lock;

/*
 * read number from @filename, increment the number and
 * write the number back to the file
 */
static void increment_data(const char *filename, int verbose, int loopno)
{
	long num;
	FILE *f;
	char buf[256];

	if (!(f = fopen(filename, "r" UL_CLOEXECSTR)))
		err(EXIT_FAILURE, "%d: failed to open: %s", getpid(), filename);

	if (!fgets(buf, sizeof(buf), f))
		err(EXIT_FAILURE, "%d failed read: %s", getpid(), filename);

	fclose(f);
	num = atol(buf) + 1;

	if (!(f = fopen(filename, "w" UL_CLOEXECSTR)))
		err(EXIT_FAILURE, "%d: failed to open: %s", getpid(), filename);

	fprintf(f, "%ld", num);

	if (close_stream(f) != 0)
		err(EXIT_FAILURE, "write failed: %s", filename);

	if (verbose)
		fprintf(stderr, "%d: %s: %ld --> %ld (loop=%d)\n", getpid(),
				filename, num - 1, num, loopno);
}

static void clean_lock(void)
{
	if (!lock)
		return;
	mnt_unlock_file(lock);
	mnt_free_lock(lock);
}

static void __attribute__((__noreturn__)) sig_handler(int sig)
{
	errx(EXIT_FAILURE, "\n%d: catch signal: %s\n", getpid(), strsignal(sig));
}

static int test_lock(struct libmnt_test *ts, int argc, char *argv[])
{
	time_t synctime = 0;
	unsigned int usecs;
	const char *datafile = NULL;
	int verbose = 0, loops = 0, l, idx = 1;

	if (argc < 3)
		return -EINVAL;

	if (strcmp(argv[idx], "--synctime") == 0) {
		synctime = (time_t) atol(argv[idx + 1]);
		idx += 2;
	}
	if (idx < argc && strcmp(argv[idx], "--verbose") == 0) {
		verbose = 1;
		idx++;
	}

	if (idx < argc)
		datafile = argv[idx++];
	if (idx < argc)
		loops = atoi(argv[idx++]);

	if (!datafile || !loops)
		return -EINVAL;

	if (verbose)
		fprintf(stderr, "%d: start: synctime=%u, datafile=%s, loops=%d\n",
			 getpid(), (int) synctime, datafile, loops);

	atexit(clean_lock);

	/* be paranoid and call exit() (=clean_lock()) for all signals */
	{
		int sig = 0;
		struct sigaction sa;

		sa.sa_handler = sig_handler;
		sa.sa_flags = 0;
		sigfillset(&sa.sa_mask);

		while (sigismember(&sa.sa_mask, ++sig) != -1 && sig != SIGCHLD)
			sigaction (sig, &sa, (struct sigaction *) 0);
	}

	/* start the test in exactly defined time */
	if (synctime) {
		struct timeval tv;

		gettimeofday(&tv, NULL);
		if (synctime && synctime - tv.tv_sec > 1) {
			usecs = ((synctime - tv.tv_sec) * 1000000UL) -
						(1000000UL - tv.tv_usec);
			xusleep(usecs);
		}
	}

	for (l = 0; l < loops; l++) {
		lock = mnt_new_lock(datafile, 0);
		if (!lock)
			return -1;

		if (mnt_lock_file(lock) != 0) {
			fprintf(stderr, "%d: failed to lock %s file\n",
					getpid(), datafile);
			return -1;
		}

		increment_data(datafile, verbose, l);

		mnt_unlock_file(lock);
		mnt_free_lock(lock);
		lock = NULL;

		/* The mount command usually finishes after a mtab update. We
		 * simulate this via short sleep -- it's also enough to make
		 * concurrent processes happy.
		 */
		if (synctime)
			xusleep(25000);
	}

	return 0;
}

/*
 * Note that this test should be executed from a script that creates many
 * parallel processes, otherwise this test does not make sense.
 */
int main(int argc, char *argv[])
{
	struct libmnt_test tss[] = {
	{ "--lock", test_lock,  " [--synctime <time_t>] [--verbose] <datafile> <loops> "
				"increment a number in datafile" },
	{ NULL }
	};

	return mnt_run_test(tss, argc, argv);
}

#endif /* TEST_PROGRAM */