Skip to contentSkip to conversation

(update): handling bad inodes in 2.4.x kernels

BUGFIX Daniel Mack 1 participant 1 message Feb 15, 2002 VFS

lore
Share & tools
lore
#0 lore raw

hi,

i already posted this to the lkml but found a new approach how to fix the bug - see below. furthermore, i forgot to cc the maintainers in my last posting.

the bug is about the handling of bad inodes in at least the 2.4.9, .16, .17 and .18-pre9 kernel releases (i suspect all 2.4 kernels are affected) and causes the names_cache to get confused.

you can easily reproduce this effect by making an inode bad using debugfs (or using a bad one if you have, of course) and open() it with the flag O_CREAT set. watching /proc/slabinfo will show you the weirdness. consider /tmp/bad is a bad inode:

# cd /tmp
# ls | grep bad
bad
# cat bad
cat: bad: Input/output error
# cat /proc/slabinfo | grep names
names_cache 0 2 4096 0 2 1 : 2 333 2 0 0
# echo foo >bad
sh: bad: Input/output error
# cat /proc/slabinfo | grep names
names_cache 4294967295 2 4096 1 2 1

boom! 4294967295 == (unsigned long) -1, the cache's length is walking backwards.

the bug is in/or around open_namei(). there is kernel memory beeing allcated from the names_cache by sys_open():783 which is freed twice, by open_namei():1184 and by sys_open():795. (linenumbers from 2.4.17) i found 2 possiblities to fix this - one is to simply break the branch in open_namei() as soon as we can know we're playing with a bad inode:

fs/namei.c Fri Feb 8 02:53:36 2002 renamed from fs/namei.c Wed Oct 17 23:46:29 2001 +0 -0
@@ -1052,6 +1052,11 @@

error = -ENOENT;
if (!dentry->d_inode)
goto exit_dput;
+
+ error = -EIO;
+ if (is_bad_inode(dentry->d_inode))
+ goto exit_dput;
+
if (dentry->d_inode->i_op && dentry->d_inode->i_op->follow_link) goto do_link;

since open() does not make sense on a bad inode, this is the simplest fix and it's also faster than the other one. but that's a matter of taste i guess ;)

the real reason for this behavior is rather a broken follow_link implementation for bad inodes: the function bad_follow_link is called in namei.c, line 1161 which does (in contrast to all other follow_link routines i found) not allocate memory with __getname() if called from within this special open_namei() loop (see __vfs_follow_link).
my alternative proposal for fixing this goes here:

fs/bad_inode.c +16 -0
@@ -17,8 +17,24 @@
  */
 static int bad_follow_link(struct dentry *dent, struct nameidata *nd)
 {
+	char *name;
+
 	dput(nd->dentry);
 	nd->dentry = dget(dent);
+        
+	/*
+	 * If it is an iterative symlinks resolution in open_namei() we
+	 * have to copy the last component. And all that crap because of
+	 * bloody create() on broken symlinks. Furrfu...
+	 */
+	if (!current->link_count) {
+		name = __getname();
+		if (!name)
+			return -ENOMEM;
+		strcpy(name, nd->last.name);
+		nd->last.name = name;				
+	}
+	
 	return 0;
 }
 

but since this is a very special workaround _only_ for open_namei(), those two patches should have the same effekt.
both patches fix the bug on all machines i tried.

thanks for listening,
daniel