svn commit: r226967 - head/sys/ufs/ufs

[Peter Holm]

On Fri, Mar 02, 2012 at 11:53:06AM -0500, John Baldwin wrote:
> On Friday, March 02, 2012 8:29:21 am Peter Holm wrote:
> > On Thu, Mar 01, 2012 at 04:47:41PM -0500, John Baldwin wrote:
> > > On Monday, October 31, 2011 11:01:47 am Peter Holm wrote:
> > > > Author: pho
> > > > Date: Mon Oct 31 15:01:47 2011
> > > > New Revision: 226967
> > > > URL: http://svn.freebsd.org/changeset/base/226967
> > > > 
> > > > Log:
> > > >   The kern_renameat() looks up the fvp using the DELETE flag, which causes
> > > >   the removal of the name cache entry for fvp.
> > > >   
> > > >   Reported by:	Anton Yuzhaninov <citrin citrin ru>
> > > >   In collaboration with:	kib
> > > >   MFC after:	1 week
> > > > 
> > > > Modified:
> > > >   head/sys/ufs/ufs/ufs_vnops.c
> > > 
> > > So I ran into this at work recently, and even this fix applied I was still 
> > > seeing rename()'s that were seemingly not taking effect.  After getting some 
> > > extra KTR traces, I figured out that the same purge needs to be applied to the 
> > > destination vnode.  Specifically, the issue I ran into was that was renaming 
> > > 'foo' to 'bar', but lookups for 'bar' were still returning the old file.  The 
> > > reason was that a lookup after the namei(RENAME) of the destination while 
> > > ufs_rename() had its locks dropped was readding the name cache entry for 
> > > 'bar', and then a cache_lookup() of 'bar' would return the old vnode as long 
> > > as that vnode was valid (e.g. if it had a link in another location, or other 
> > > processes had an open file descriptor for it).  I'm currently testing the 
> > > patch below:
> > > 
> > 
> > I now have a scenario that fails, but not quite the same way you
> > describe.
> > 
> > It looks like this:
> > 
> > touch file1
> > echo xxx > file2
> > rename(file1, file2)
> > 
> > A different process performs stat() on both files in a tight loop.
> > 
> > Once in a while I observe that a stat() of file2, after the rename,
> > returns a link count of zero. Size is zero as expected, but the inode
> > number of file2 is unchanged.
> 
> Hmm, that is surprising.  I would not expect inconsistent stat info.  I
> have no explanation for why that would happen.  I do not have a simplified
> test program, just a specific workload at work.  In this case it's workflow
> is more like this:
> 
> 	fd = flopen(file1, O_CREAT);
> 	fstat(fd);
> 	if (st.st_size == 0) {
> 		fd2 = open(file1.temp, O_CREAT | O_EXLOCK);
> 		fd3 = open(someotherfile);
> 		copy_data(fd3, fdf2);
> 		close(fd3);
> 		rename(file1.temp, file1);
> 		close(fd);
> 		fd = fd2;
> 	}
> 	link(file1, uniquedir/file1);
> 	close(fd);
> 
> 	/* Use uniquedir/file1, and unlink it when done. */
> 
> What I observed was that sometimes uniquedir/file1 would end up referencing
> the empty file created by flopen() after the rename() rather than linking
> to the file created when file1.temp was created.
> 
> > I've been running the same test with your patch and not observed this
> > problem. This on UFS2 with SU enabled.
> 
> Hmm, I wish I could explain explain your odd result above in terms of this
> bug, but the results from stat() should always be consistent (VOP_GETATTR()
> can't switch vnodes mid-stream as it were).
> 
> BTW, note that in my case where I had multiple processes all doing the same
> loop, in the edge case, another process always had the file open (and was
> blocked in flock() in flopen()) when the rename() happened, so that prevented
> the vnode from going away.  This is important as otherwise the use count would
> drop to zero and marked inactive which removes all references to it from the
> name cache.  In my case the flock() in flopen() and the fact that the "first"
> process held the flock until after the rename and call to link() made the
> race more likely to trigger.
> 
> Hmm, perhaps one way to do this would be:
> 
> 	touch file.always (save its i-node)
> 	fork worker process that just continually does a 'stat file1' in a loop
> 	main process:
> 		ln file.always file1
> 		touch file2
> 		rename file2 file1
> 		stat file1
> 		complain if file1 has the saved i-node
> 
> Hmm, I just whipped up something to do this and it fails early and often on
> an unpatched kernel, but does not with my patch.
> 
> #include <sys/types.h>
> #include <sys/stat.h>
> #include <err.h>
> #include <errno.h>
> #include <fcntl.h>
> #include <stdio.h>
> #include <stdlib.h>
> #include <unistd.h>
> 
> static char *always, *file1, *file2;
> static ino_t always_ino;
> 
> static void
> usage(void)
> {
> 	fprintf(stderr, "Usage: rename_race <dir>\n");
> 	exit(1);
> }
> 
> static void
> child(void)
> {
> 	struct stat sb;
> 
> 	/* Exit as soon as our parent exits. */
> 	while (getppid() != 1) {
> 		stat(file1, &sb);
> 	}
> 	exit(0);
> }
> 
> static void
> create_file(const char *path)
> {
> 	int fd;
> 
> 	fd = open(path, O_CREAT, 0666);
> 	if (fd < 0)
> 		err(1, "open(%s)", path);
> 	close(fd);
> }
> 
> int
> main(int ac, char **av)
> {
> 	struct stat sb, sb2;
> 	pid_t pid;
> 
> 	if (ac != 2)
> 		usage();
> 	if (stat(av[1], &sb) != 0)
> 		err(1, "stat(%s)", av[1]);
> 	if (!S_ISDIR(sb.st_mode))
> 		errx(1, "%s not a directory", av[1]);
> 
> 	asprintf(&always, "%s/file.always", av[1]);
> 	asprintf(&file1, "%s/file1", av[1]);
> 	asprintf(&file2, "%s/file2", av[1]);
> 
> 	create_file(always);
> 	if (stat(always, &sb) != 0)
> 		err(1, "stat(%s)", always);
> 	always_ino = sb.st_ino;
> 
> 	pid = fork();
> 	if (pid < 0)
> 		err(1, "fork");
> 	if (pid == 0)
> 		child();
> 	for (;;) {
> 		if (unlink(file1) < 0 && errno != ENOENT)
> 			err(1, "unlink(%s)", file1);
> 		if (link(always, file1) < 0)
> 			err(1, "link(%s, %s)", always, file1);
> 		create_file(file2);
> 		if (stat(file2, &sb2) < 0)
> 			err(1, "stat(%s)", file2);
> 		if (rename(file2, file1) < 0)
> 			err(1, "rename(%s, %s)", file2, file1);
> 		if (stat(file1, &sb) < 0)
> 			err(1, "stat(%s)", file1);
> 		if (sb.st_ino != sb2.st_ino ||
> 		    sb.st_ino == always_ino)
> 			printf("Bad stat: always: %d file1: %d (should be %d)\n",
> 			    always_ino, sb.st_ino, sb2.st_ino);
> 	}
> 	return (0);
> }
> 

Excellent test scenario!

For the tmpfs case I do not see any problems using both our
scenarios on a pristine HEAD.
-- 
Peter