[PATCH] dcache: fix locking bug in __dentry_path()

[PATCH] dcache: fix locking bug in __dentry_path()

[Denys Vlasenko]

If given buffer size is zero, we forget to rcu_read_unlock()
on error path.

Cc: linux-kernel@vger.kernel.org
Cc: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Denys Vlasenko <dvlasenk@redhat.com>
---
 fs/dcache.c | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)

diff --git a/fs/dcache.c b/fs/dcache.c
index cb4a106..24a01fc 100644
--- a/fs/dcache.c
+++ b/fs/dcache.c
@@ -3127,8 +3127,11 @@ restart:
 	end = buf + buflen;
 	len = buflen;
 	prepend(&end, &len, "\0", 1);
-	if (buflen < 1)
+	if (buflen < 1) {
+		if (!(seq & 1))
+			rcu_read_unlock();
 		goto Elong;
+	}
 	/* Get '/' right */
 	retval = end-1;
 	*retval = '/';
-- 
1.8.1.4

Re: [PATCH] dcache: fix locking bug in __dentry_path()

[Al Viro]

On Fri, Jan 24, 2014 at 11:21:26AM +0100, Denys Vlasenko wrote:
> If given buffer size is zero, we forget to rcu_read_unlock()
> on error path.

Er...  Where could we ever get called with size zero?  IOW, the
real question is whether that check makes any sense.  If nothing
else, a much more obvious fix (again, assuming there is a call
chain that could have triggered that in the first place), would
be to take that
        if (buflen < 1)
                goto Elong;
on the very top of __dentry_path().  Note that buflen never
change, so this check could bloody well be handled once - and
earlier.  But again, the callers shouldn't be calling it that
way...

Re: [PATCH] dcache: fix locking bug in __dentry_path()

[Denys Vlasenko]

On 01/24/2014 01:24 PM, Al Viro wrote:
> On Fri, Jan 24, 2014 at 11:21:26AM +0100, Denys Vlasenko wrote:
>> If given buffer size is zero, we forget to rcu_read_unlock()
>> on error path.
> 
> Er...  Where could we ever get called with size zero?  IOW, the
> real question is whether that check makes any sense.  If nothing
> else, a much more obvious fix (again, assuming there is a call
> chain that could have triggered that in the first place), would
> be to take that
>         if (buflen < 1)
>                 goto Elong;
> on the very top of __dentry_path().  Note that buflen never
> change, so this check could bloody well be handled once - and
> earlier.

Yes. In fact the check is buggier than I thought:

        prepend(&end, &len, "\0", 1);
        if (buflen < 1)
                goto Elong;
        /* Get '/' right */
        retval = end-1;
        *retval = '/';   <=== !!!

It should be "buflen < 2". With buflen == 1, we touch buffer[-1].

> But again, the callers shouldn't be calling it that
> way...

I'll add a check in d_path() and post another patch.
I assume the assertion should be 'size >= 2', since
path "" should never be produced anyway, the shortest path
is "/"?

-- 
vda